A la m´emoire de mon p`ere, Arthur, a ` ma m`ere, Annette, et ` a Kahina. M. Bilodeau
To Rebecca and Deena. D. Brenner...

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

A la m´emoire de mon p`ere, Arthur, a ` ma m`ere, Annette, et ` a Kahina. M. Bilodeau

To Rebecca and Deena. D. Brenner

Preface

Our object in writing this book is to present the main results of the modern theory of multivariate statistics to an audience of advanced students who would appreciate a concise and mathematically rigorous treatment of that material. It is intended for use as a textbook by students taking a ﬁrst graduate course in the subject, as well as for the general reference of interested research workers who will ﬁnd, in a readable form, developments from recently published work on certain broad topics not otherwise easily accessible, as, for instance, robust inference (using adjusted likelihood ratio tests) and the use of the bootstrap in a multivariate setting. The references contains over 150 entries post-1982. The main development of the text is supplemented by over 135 problems, most of which are original with the authors. A minimum background expected of the reader would include at least two courses in mathematical statistics, and certainly some exposure to the calculus of several variables together with the descriptive geometry of linear algebra. Our book is, nevertheless, in most respects entirely self-contained, although a deﬁnite need for genuine ﬂuency in general mathematics should not be underestimated. The pace is brisk and demanding, requiring an intense level of active participation in every discussion. The emphasis is on rigorous proof and derivation. The interested reader would proﬁt greatly, of course, from previous exposure to a wide variety of statistically motivating material as well, and a solid background in statistics at the undergraduate level would obviously contribute enormously to a general sense of familiarity and provide some extra degree of comfort in dealing with the kinds of challenges and diﬃculties to be faced in the relatively advanced work

viii

Preface

of the sort with which our book deals. In this connection, a speciﬁc introduction oﬀering comprehensive overviews of the fundamental multivariate structures and techniques would be well advised. The textbook A First Course in Multivariate Statistics by Flury (1997), published by SpringerVerlag, provides such background insight and general description without getting much involved in the “nasty” details of analysis and construction. This would constitute an excellent supplementary source. Our book is in most ways thoroughly orthodox, but in several ways novel and unique. In Chapter 1 we oﬀer a brief account of the prerequisite linear algebra as it will be applied in the subsequent development. Some of the treatment is peculiar to the usages of multivariate statistics and to this extent may seem unfamiliar. Chapter 2 presents in review, the requisite concepts, structures, and devices from probability theory that will be used in the sequel. The approach taken in the following chapters rests heavily on the assumption that this basic material is well understood, particularly that which deals with equality-in-distribution and the Cram´er-Wold theorem, to be used with unprecedented vigor in the derivation of the main distributional results in Chapters 4 through 8. In this way, our approach to multivariate theory is much more structural and directly algebraic than is perhaps traditional, tied in this fashion much more immediately to the way in which the various distributions arise either in nature or may be generated in simulation. We hope that readers will ﬁnd the approach refreshing, and perhaps even a bit liberating, particularly those saturated in a lifetime of matrix derivatives and jacobians. As a textbook, the ﬁrst eight chapters should provide a more than adequate amount of material for coverage in one semester (13 weeks). These eight chapters, proceeding from a thorough discussion of the normal distribution and multivariate sampling in general, deal in random matrices, Wishart’s distribution, and Hotelling’s T 2 , to culminate in the standard theory of estimation and the testing of means and variances. The remaining six chapters treat of more specialized topics than it might perhaps be wise to attempt in a simple introduction, but would easily be accessible to those already versed in the basics. With such an audience in mind, we have included detailed chapters on multivariate regression, principal components, and canonical correlations, each of which should be of interest to anyone pursuing further study. The last three chapters, dealing, in turn, with asymptotic expansion, robustness, and the bootstrap, discuss concepts that are of current interest for active research and take the reader (gently) into territory not altogether perfectly charted. This should serve to draw one (gracefully) into the literature. The authors would like to express their most heartfelt thanks to everyone who has helped with feedback, criticism, comment, and discussion in the preparation of this manuscript. The ﬁrst author would like especially to convey his deepest respect and gratitude to his teachers, Muni Srivastava

Preface

ix

of the University of Toronto and Takeaki Kariya of Hitotsubashi University, who gave their unstinting support and encouragement during and after his graduate studies. The second author is very grateful for many discussions with Philip McDunnough of the University of Toronto. We are indebted to Nariaki Sugiura for his kind help concerning the application of Sugiura’s Lemma and to Rudy Beran for insightful comments, which helped to improve the presentation. Eric Marchand pointed out some errors in the literature about the asymptotic moments in Section 8.4.1. We would like to thank the graduate students at McGill University and Universit´e de Montr´eal, Gulhan Alpargu, Diego Clonda, Isabelle Marchand, Philippe St-Jean, Gueye N’deye Rokhaya, Thomas Tolnai and Hassan Younes, who helped improve the presentation by their careful reading and problem solving. Special thanks go to Pierre Duchesne who, as part of his Master Memoir, wrote and tested the S-Plus function for the calculation of the robust S estimate in Appendix C.

M. Bilodeau D. Brenner

Contents

Preface List of Tables List of Figures

vii xv xvii

1 Linear algebra 1.1 Introduction . . . . . . . . . . . . . . . 1.2 Vectors and matrices . . . . . . . . . . 1.3 Image space and kernel . . . . . . . . . 1.4 Nonsingular matrices and determinants 1.5 Eigenvalues and eigenvectors . . . . . . 1.6 Orthogonal projections . . . . . . . . . 1.7 Matrix decompositions . . . . . . . . . 1.8 Problems . . . . . . . . . . . . . . . . . 2 Random vectors 2.1 Introduction . . . . . . . . . . . . . 2.2 Distribution functions . . . . . . . . 2.3 Equals-in-distribution . . . . . . . . 2.4 Discrete distributions . . . . . . . . 2.5 Expected values . . . . . . . . . . . 2.6 Mean and variance . . . . . . . . . 2.7 Characteristic functions . . . . . . . 2.8 Absolutely continuous distributions 2.9 Uniform distributions . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . .

. . . . . . . . .

. . . . . . . .

. . . . . . . . .

. . . . . . . .

. . . . . . . . .

. . . . . . . .

. . . . . . . . .

. . . . . . . .

. . . . . . . . .

. . . . . . . .

. . . . . . . . .

. . . . . . . .

. . . . . . . . .

. . . . . . . .

. . . . . . . . .

. . . . . . . .

. . . . . . . . .

. . . . . . . .

1 1 1 3 4 5 9 10 11

. . . . . . . . .

14 14 14 16 16 17 18 21 22 24

xii

Contents

2.10 2.11 2.12 2.13 2.14

Joints and marginals Independence . . . . Change of variables . Jacobians . . . . . . . Problems . . . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

25 27 28 30 33

3 Gamma, Dirichlet, and F distributions 3.1 Introduction . . . . . . . . . . . . . . . 3.2 Gamma distributions . . . . . . . . . . 3.3 Dirichlet distributions . . . . . . . . . . 3.4 F distributions . . . . . . . . . . . . . 3.5 Problems . . . . . . . . . . . . . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

36 36 36 38 42 42

4 Invariance 4.1 Introduction . . . . . . . . . . . . . . . . . . 4.2 Reﬂection symmetry . . . . . . . . . . . . . 4.3 Univariate normal and related distributions 4.4 Permutation invariance . . . . . . . . . . . . 4.5 Orthogonal invariance . . . . . . . . . . . . . 4.6 Problems . . . . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

43 43 43 44 47 48 52

5 Multivariate normal 5.1 Introduction . . . . . . . . . . . . . . . 5.2 Deﬁnition and elementary properties . 5.3 Nonsingular normal . . . . . . . . . . . 5.4 Singular normal . . . . . . . . . . . . . 5.5 Conditional normal . . . . . . . . . . . 5.6 Elementary applications . . . . . . . . 5.6.1 Sampling the univariate normal 5.6.2 Linear estimation . . . . . . . . 5.6.3 Simple correlation . . . . . . . . 5.7 Problems . . . . . . . . . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

55 55 55 58 62 62 64 64 65 67 69

6 Multivariate sampling 6.1 Introduction . . . . . . . . . . . . . . . . . 6.2 Random matrices and multivariate sample 6.3 Asymptotic distributions . . . . . . . . . . 6.4 Problems . . . . . . . . . . . . . . . . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

73 73 73 78 81

7 Wishart distributions 7.1 Introduction . . . . . . . . . . . . . ¯ and S . . . . 7.2 Joint distribution of x 7.3 Properties of Wishart distributions 7.4 Box-Cox transformations . . . . . . 7.5 Problems . . . . . . . . . . . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

85 85 85 87 94 96

. . . . .

. . . . .

. . . . . . . . . .

. . . . .

. . . . .

Contents

xiii

8 Tests on mean and variance 8.1 Introduction . . . . . . . . . . . . . . . . . . 8.2 Hotelling-T 2 . . . . . . . . . . . . . . . . . . 8.3 Simultaneous conﬁdence intervals on means 8.3.1 Linear hypotheses . . . . . . . . . . . 8.3.2 Nonlinear hypotheses . . . . . . . . . 8.4 Multiple correlation . . . . . . . . . . . . . . 8.4.1 Asymptotic moments . . . . . . . . . 8.5 Partial correlation . . . . . . . . . . . . . . . 8.6 Test of sphericity . . . . . . . . . . . . . . . 8.7 Test of equality of variances . . . . . . . . . 8.8 Asymptotic distributions of eigenvalues . . . 8.8.1 The one-sample problem . . . . . . . 8.8.2 The two-sample problem . . . . . . . 8.8.3 The case of multiple eigenvalues . . . 8.9 Problems . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

98 98 98 104 104 107 109 114 116 117 121 124 124 132 133 137

9 Multivariate regression 9.1 Introduction . . . . . . . . . . . . . . . 9.2 Estimation . . . . . . . . . . . . . . . . 9.3 The general linear hypothesis . . . . . 9.3.1 Canonical form . . . . . . . . . 9.3.2 LRT for the canonical problem 9.3.3 Invariant tests . . . . . . . . . . 9.4 Random design matrix X . . . . . . . . 9.5 Predictions . . . . . . . . . . . . . . . . 9.6 One-way classiﬁcation . . . . . . . . . . 9.7 Problems . . . . . . . . . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

144 144 145 148 148 150 151 154 156 158 159

10 Principal components 10.1 Introduction . . . . . . . . . . . . . . 10.2 Deﬁnition and basic properties . . . . 10.3 Best approximating subspace . . . . . 10.4 Sample principal components from S 10.5 Sample principal components from R 10.6 A test for multivariate normality . . 10.7 Problems . . . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

161 161 162 163 164 166 169 172

11 Canonical correlations 11.1 Introduction . . . . . . . . . . . . 11.2 Deﬁnition and basic properties . . 11.3 Tests of independence . . . . . . . 11.4 Properties of U distributions . . . 11.4.1 Q-Q plot of squared radii .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

174 174 175 177 181 184

. . . . .

. . . . .

xiv

Contents

11.5 11.6

Asymptotic distributions . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . .

12 Asymptotic expansions 12.1 Introduction . . . . 12.2 General expansions 12.3 Examples . . . . . . 12.4 Problem . . . . . .

189 190

. . . .

. . . .

. . . .

. . . .

. . . .

195 195 195 200 205

13 Robustness 13.1 Introduction . . . . . . . . . . . . . . . . . . . . . 13.2 Elliptical distributions . . . . . . . . . . . . . . . 13.3 Maximum likelihood estimates . . . . . . . . . . . 13.3.1 Normal MLE . . . . . . . . . . . . . . . . 13.3.2 Elliptical MLE . . . . . . . . . . . . . . . 13.4 Robust estimates . . . . . . . . . . . . . . . . . . 13.4.1 M estimate . . . . . . . . . . . . . . . . . . 13.4.2 S estimate . . . . . . . . . . . . . . . . . . 13.4.3 Robust Hotelling-T 2 . . . . . . . . . . . . 13.5 Robust tests on scale matrices . . . . . . . . . . . 13.5.1 Adjusted likelihood ratio tests . . . . . . . 13.5.2 Weighted Nagao’s test for a given variance 13.5.3 Relative eﬃciency of adjusted LRT . . . . 13.6 Problems . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

206 206 207 213 213 213 222 222 224 226 227 228 233 236 238

14 Bootstrap conﬁdence regions and tests 14.1 Conﬁdence regions and tests for the mean 14.2 Conﬁdence regions for the variance . . . . 14.3 Tests on the variance . . . . . . . . . . . . 14.4 Problem . . . . . . . . . . . . . . . . . . .

. . . .

. . . .

. . . .

. . . .

243 243 246 249 252

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

A Inversion formulas

253

B Multivariate cumulants B.1 Deﬁnition and properties . . . . . . . . . . . . . . . . . . B.2 Application to asymptotic distributions . . . . . . . . . . B.3 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . .

256 256 259 259

C S-plus functions

261

References Author Index Subject Index

263 277 281

List of Tables

12.1 Polynomials δs and Bernoulli numbers Bs for asymptotic expansions. . . . . . . . . . . . . . . . . . . . . . . . . . . 12.2 Asymptotic expansions for U (2; 12, n) distributions. . . .

201 203

13.1 Asymptotic eﬃciency of S estimate of scatter at the normal distribution. . . . . . . . . . . . . . . . . . . . . . . . . . . 225 13.2 Asymptotic signiﬁcance level of unadjusted LRT for α = 5%. 238

This page intentionally left blank

List of Figures

2.1 3.1 5.1 5.2

5.3 8.1 8.2

Bivariate Frank density with standard normal marginals and a correlation of 0.7. . . . . . . . . . . . . . . . . . . . . . .

27

Bivariate Dirichlet density for values of the parameters p1 = p2 = 1 and p3 = 2. . . . . . . . . . . . . . . . . . . . . . .

41

Bivariate normal density for values of the parameters µ1 = µ2 = 0, σ1 = σ2 = 1, and ρ = 0.7. . . . . . . . . . . . . . . Contours of the bivariate normal density for values of the parameters µ1 = µ2 = 0, σ1 = σ2 = 1, and ρ = 0.7. Values of c = 1, 2, 3 were taken. . . . . . . . . . . . . . . . . . . A contour of a trivariate normal density. . . . . . . . . . . Power function of Hotelling-T 2 when p = 3 and n = 40 at a level of signiﬁcance α = 0.05. . . . . . . . . . . . . . . . . Power function of the likelihood ratio test for H0 : R = 0 when p = 3, and n = 20 at a level of signiﬁcance α = 0.05.

11.1 Q-Q plot for a sample of size n = 50 from a trivariate normal, N3 (0, I), distribution. . . . . . . . . . . . . . . . . . . . . 11.2 Q-Q plot for a sample of size n = 50 from a trivariate t on 1 degree of freedom, t3,1 (0, I) ≡ Cauchy3 (0, I), distribution.

59

60 61 101 113 187 188

This page intentionally left blank

1 Linear algebra

1.1 Introduction Multivariate analysis deals with issues related to the observations of many, usually correlated, variables on units of a selected random sample. These units can be of any nature such as persons, cars, cities, etc. The observations are gathered as vectors; for each selected unit corresponds a vector of observed variables. An understanding of vectors, matrices, and, more generally, linear algebra is thus fundamental to the study of multivariate analysis. Chapter 1 represents our selection of several important results on linear algebra. They will facilitate a great many of the concepts in multivariate analysis. A useful reference for linear algebra is Strang (1980).

1.2 Vectors and matrices To express the dependence of the x ∈ Rn on its coordinates, we may write any of x1 .. x = (xi , i = 1, . . . , n) = (xi ) = . . xn In this manner, x is envisaged as a “column” vector. The transpose of x is the “row” vector x ∈ Rn

x = (xi ) = (x1 , . . . , xn ) .

2

1. Linear algebra

An m × n matrix A ∈ Rm n may also be denoted in various a11 .. A = (aij , i = 1, . . . , m, j = 1, . . . , n) = (aij ) = . am1

ways:

· · · a1n .. .. . . . · · · amn

The transpose of A is the n × m matrix A ∈ Rnm : a11 · · · am1 .. . .. . A = (aij ) = (aji ) = .. . . a1n · · · amn A square matrix S ∈ Rnn satisfying S = S is termed symmetric. The product of the m × n matrix A by the n × p matrix B is the m × p matrix C = AB for which n cij = aik bkj . k=1

n

is tr A = i=1 aii and one veriﬁes that for A ∈ Rm The trace of A ∈ n n and B ∈ Rm , tr AB = tr BA. In particular, row vectors and column vectors are themselves matrices, so that for x, y ∈ Rn , we have the scalar result Rnn

x y =

n

xi yi = y x.

i=1

This provides the standard inner product, x, y = x y, in Rn with the associated “euclidian norm” (length or modulus) n 1/2 x2i . |x| = x, x1/2 = i=1

The Cauchy-Schwarz inequality is now proved. Proposition 1.1 |x, y| ≤ |x| |y|, ∀x, y ∈ Rn , with equality if and only if (iﬀ ) x = λy for some λ ∈ R. Proof. If x = λy, for some λ ∈ R, the equality clearly holds. If not, 0 < |x − λy|2 = |x|2 − 2λx, y + λ2 |y|2 , ∀λ ∈ R; thus, the discriminant of 2 the quadratic polynomial must satisfy 4x, y2 − 4|x|2 |y|2 < 0. The cosine of the angle θ between the vectors x = 0 and y = 0 is just cos(θ) =

x, y . |x| |y|

Orthogonality is another associated concept. Two vectors x and y in Rn will be said to be orthogonal iﬀ x, y = 0. In contrast, the outer (or tensor) product of x and y is an n × n matrix xy = (xi yj )

1.3. Image space and kernel

3

and this product is not commutative. The concept of orthonormal basis plays a major role in linear algebra. A set {vi } of vectors in Rn is orthonormal if 0, i = j vi vj = δij = 1, i = j. The symbol δij is referred to as the Kronecker delta. The Gram-Schmidt orthogonalization method gives a construction of an orthonormal basis from an arbitrary basis. Proposition 1.2 Let {v1 , . . . , vn } be a basis of Rn . Deﬁne u1 ui

= v1 /|v1 |, = wi /|wi |,

i−1 where wi = vi − j=1 (vi uj )uj , i = 2, . . . , n. Then, {u1 , . . . , un } is an orthonormal basis.

1.3 Image space and kernel Now, a matrix may equally well be recognized as a function either of its column vectors or its row vectors: g1 .. A = (a1 , . . . , an ) = . gm

for aj ∈ Rm , j = 1, . . . , n or gi ∈ Rn , i = 1, . . . , m. If we then write B = (b1 , . . . , bp ) with bj ∈ Rn , j = 1, . . . , p, we ﬁnd that AB = (Ab1 , . . . , Abp ) = (gi bj ) . In particular, for x ∈ Rn , we have expressly that x1 n .. = xi ai Ax = (a1 , . . . , an ) . i=1 xn or g1 x g1 .. .. Ax = . x = . . gm

(1.1)

(1.2)

gm x

The orthogonal complement of a subspace V ⊂ Rn is, by deﬁnition, the subspace V ⊥ = {y ∈ Rn : y ⊥ x, ∀x ∈ V}.

4

1. Linear algebra

Expression (1.1) identiﬁes the image space of A, Im A = {Ax : x ∈ Rn }, with the linear span of its column vectors and the expression (1.2) reveals the kernel, ker A = {x ∈ Rn : Ax = 0}, to be the orthogonal complement of the row space, equivalently ker A = (Im A )⊥ . The dimension of the subspace Im A is called the rank of A and satisﬁes rank A = rank A , whereas the dimension of ker A is called the nullity of A. They are related through the following simple relation: Proposition 1.3 For any A ∈ Rm n , n = nullity A + rank A. Proof. Let {v1 , . . . , vν } be a basis of ker A and extend it to a basis {v1 , . . . , vν , vν+1 , . . . , vn } of Rn . One can easily check {Avν+1 , . . . , Avn } is a basis of Im A. Thus, n = nullity A + rank A. 2

1.4 Nonsingular matrices and determinants We recall some basic facts about nonsingular (one-to-one) linear transformations and determinants. By writing A ∈ Rnn in terms of its column vectors A = (a1 , . . . , an ) with aj ∈ Rn , j = 1, . . . , n, it is clear that A is one-to-one ⇐⇒ a1 , . . . , an is a basis ⇐⇒ ker A = {0} and also from the simple relation n = nullity A + rank A, A is one-to-one ⇐⇒ A is one-to-one and onto. These are all equivalent ways of saying A has an inverse or that A is nonsingular. Denote by σ(1), . . . , σ(n) a permutation of 1, . . . , n and by n(σ) its parity. Let Sn be the group of all the n! permutations. The determinant is, by deﬁnition, the unique function det : Rnn → R, denoted |A| = det(A), that is, (i) multilinear: linear in each of a1 , . . . , an separately (ii) alternating: aσ(1) , . . . , aσ(n) = (−1)n(σ) |(a1 , . . . , an )| (iii) normed: |I| = 1. This produces the formula |A| =

(−1)n(σ) a1σ(1) · · · anσ(n)

σ∈Sn

by which one veriﬁes |AB| = |A| |B| and |A | = |A| .

1.5. Eigenvalues and eigenvectors

5

Determinants are usually calculated with a Laplace development along any given row or column. To this end, let A = (aij ) ∈ Rnn . Now, deﬁne the minor |m(i, j)| of aij as the determinant of the (n−1)×(n−1) “submatrix” obtained by deleting the ith row and the jth column of A and the cofactor |m(i, j)|. Then, the Laplace development of |A| of aij as c(i, j) = (−1)i+j n along the ith row is |A| = j=1 aij ·c(i, j) and a similar development along n the jth column is |A| = i=1 aij · c(i, j). By deﬁning adj(A) = (c(j, i)), the transpose of the matrix of cofactors, to be the adjoint of A, it can be shown A−1 = |A|−1 adj(A). But then Proposition 1.4 A is one-to-one ⇐⇒ |A| = 0. Proof. A is one-to-one means it has an inverse B, |A| |B| = 1 so n |A| = 0. But, conversely, if |A| = 0, suppose Ax = j=1 xj aj = 0, then substituting Ax for the ith column of A n a1 , . . . , xj aj , . . . , an = xi |A| = 0, i = 1, . . . , n j=1 so that x = 0, whereby A is one-to-one.

2

In general, for aj ∈ Rn , j = 1, . . . , k, write A = (a1 , . . . , ak ) and form the “inner product” matrix A A = (ai aj ) ∈ Rkk . We ﬁnd Proposition 1.5 For A ∈ Rnk , 1. ker A = ker A A 2. rank A = rank A A 3. a1 , . . . , ak are linearly independent in Rn ⇐⇒ |A A| = 0. Proof. If x ∈ ker A, then Ax = 0 =⇒ A Ax = 0, and, conversely, if x ∈ ker A A, then A Ax = 0 =⇒ x A Ax = 0 = |Ax|2 =⇒ Ax = 0. The second part follows from the relation k = nullity A + rank A and the 2 third part is immediate as ker A = {0} iﬀ ker A A = {0}.

1.5 Eigenvalues and eigenvectors We now brieﬂy state some concepts related to eigenvalues and eigenvectors. Consider, ﬁrst, the complex vector space Cn . The conjuguate of v = x+iy ∈ C, x, y ∈ R, is v = x − iy. The concepts deﬁned earlier are anologous in this case. The Hermitian transpose of a column vector v = (vi ) ∈ Cn is the row vector vH = (vi ) . The inner product on Cn can then be written v1 , v2 =

6

1. Linear algebra

v1H v2 for any v1 , v2 ∈ Cn . The Hermitian transpose of A = (aij ) ∈ Cm n is AH = (aji ) ∈ Cnm and satisﬁes for B ∈ Cnp , (AB)H = BH AH . The matrix A ∈ Cnn is termed Hermitian iﬀ A = AH . We now deﬁne what is meant by an eigenvalue. A scalar λ ∈ C is an eigenvalue of A ∈ Cnn if there exists a vector v = 0 in Cn such that Av = λv. Equivalently, λ ∈ C is an eigenvalue of A iﬀ |A − λI| = 0, which is a polynomial equation of degree n. Hence, there are n complex eigenvalues, some of which may be real, with possibly some repetitions (multiplicity). The vector v is then termed the eigenvector of A corresponding to the eigenvalue λ. Note that if v is an eigenvector, so is αv, ∀α = 0 in C, and, in particular, v/|v| is a normalized eigenvector. Now, before deﬁning what is meant by A is “diagonalizable” we deﬁne a matrix U ∈ Cnn to be unitary iﬀ UH U = I = UUH . This means that the columns (or rows) of U comprise an orthonormal basis of Cn . We note immediately that if {u1 , . . . , un } is an orthonormal basis of eigenvectors corresponding to eigenvalues {λ1 , . . . , λn }, then A can be diagonalized by the unitary matrix U = (u1 , . . . , un ); i.e., we can write UH AU = UH (Au1 , . . . , Aun ) = UH (λ1 u1 , . . . , λn un ) = diag(λ), where λ = (λ1 , . . . , λn ) . Another simple related property: If there exists a unitary matrix U = (u1 , . . . , un ) such that UH AU = diag(λ), then ui is an eigenvector corresponding to λi . To verify this, note that Aui = U diag(λ)UH ui = U diag(λ)ei = Uλi ei = λi ui . Two fundamental propositions concerning Hermitian matrices are the following. Proposition 1.6 If A ∈ Cnn is Hermitian, then all its eigenvalues are real. Proof. vH Av = (vH Av)H = vH AH v = vH Av, which means that vH Av is real for any v ∈ Cn . Now, if Av = λv for some v = 0 in Cn , then vH Av = λvH v = λ|v|2 . But since vH Av and |v|2 are real, so is λ. 2 Proposition 1.7 If A ∈ Cnn is Hermitian and v1 and v2 are eigenvectors corresponding to eigenvalues λ1 and λ2 , respectively, where λ1 = λ2 , then v1 ⊥ v2 . Proof. Since A is Hermitian, A = AH and λi , i = 1, 2, are real. Then, Av1 = λ1 v1 Av2 = λ2 v2

=⇒ =⇒

v1H AH = v1H A = λ1 v1H =⇒ v1H Av2 = λ1 v1H v2 , v1H Av2 = λ2 v1H v2 .

Subtracting the last two expressions, (λ1 −λ2 )v1H v2 = 0 and, thus, v1H v2 = 0. 2

1.5. Eigenvalues and eigenvectors

7

Proposition 1.7 immediately shows that if all the eigenvalues of A, Hermitian, are distinct, then there exists an orthonormal basis of eigenvectors whereby A is diagonalizable. Toward proving this is true even when the eigenvalues may be of a multiple nature, we need the following proposition. However, before stating it, deﬁne T = (tij ) ∈ Rnn to be a lower triangular matrix iﬀ tij = 0, i < j. Similarly, T ∈ Rnn is termed upper triangular iﬀ tij = 0, i > j. Proposition 1.8 Let A ∈ Cnn be any matrix. There exists a unitary matrix U ∈ Cnn such that UH AU is upper triangular. Proof. The proof is by induction on n. The result is obvious for n = 1. Next, assume the proposition holds for n and prove it is true for n + 1. Let λ1 be an eigenvalue of A and u1 , |u1 | = 1, be an eigenvector. Let U1 = (u1 , Γ) for some Γ such that U1 is unitary (such a Γ exists from the Gram-Schmidt method). Then, λ1 uH H 1 AΓ UH AU = U (λ u , AΓ) = , 1 1 1 1 1 0 B where B = ΓH AΓ ∈ Cnn . From the induction hypothesis, there exists V unitary such that VH BV = T is triangular. Deﬁne 1 0 U2 = 0 V and it is clear that U2 is also unitary. Finally, λ1 uH H H 1 AΓ (U1 U2 ) A(U1 U2 ) = U2 U2 0 B 1 0 λ1 uH 1 1 AΓ = 0 VH 0 0 B λ1 uH 1 AΓV = , 0 T

0 V

which is of the desired form. The proof is complete because U ≡ U1 U2 is unitary. 2 As a corollary we obtain that Hermitian matrices are always diagonalizable. Corollary 1.1 Let A ∈ Cnn be Hermitian. There exists a unitary matrix U such that UH AU = diag(λ). Proof. Proposition 1.8 showed there exists U, unitary, such that UH AU is triangular. However, if A is Hermitian, so is UH AU. The only matrices that are both Hermitian and triangular are the diagonal matrices. 2 In the sequel, we will always use Corollary 1.1 for S ∈ Rnn symmetric. However, ﬁrst note that when S is symmetric all its eigenvalues are real, whereby the eigenvectors can also be chosen to be real, they are the solutions of (S − λI)x = 0. When U ∈ Rnn is unitary, it is called an orthogonal

8

1. Linear algebra

matrix instead. A matrix H ∈ Rnn is said to be orthogonal iﬀ the columns (or rows) of H form an orthonormal basis of Rn , i.e., H H = I = HH . The group of orthogonal matrices in Rnn will be denoted by On = {H ∈ Rnn : HH = I}. We have proven the “spectral decomposition:” Proposition 1.9 If S ∈ Rnn is symmetric, then there exists H ∈ On such that H SH = diag(λ). The columns of H form an orthonormal basis of eigenvectors and λ is the vector of corresponding eigenvalues. Now, a symmetric matrix S ∈ Rnn is said to be positive semideﬁnite, denoted S ≥ 0 or S ∈ PSn , iﬀ v Sv ≥ 0, ∀v ∈ Rn , and it is positive Finally, the deﬁnite, denoted S > 0 or S ∈ Pn , iﬀ v Sv > 0, ∀v = 0. positive semideﬁnite and positive deﬁnite matrices can be characterized in terms of eigenvalues. Proposition 1.10 Let S ∈ Rnn symmetric with eigenvalues λ1 , . . . , λn . 1. S ≥ 0 iﬀ λi ≥ 0, i = 1, . . . , n. 2. S > 0 iﬀ λi > 0, i = 1, . . . , n. Note that if S is positive semideﬁnite, then from Proposition 1.9, we can write S = HDH = (HD1/2 )(HD1/2 ) = (HD1/2 H )2 , 1/2

where D = diag(λi ) and D1/2 = diag(λi ), so that for A = HD1/2 , S = AA , or for B = HD1/2 H , S = B2 . The positive semideﬁnite matrix B is often denoted S1/2 and is the square root of S. If S is positive deﬁnite, 2

we can also deﬁne S−1/2 = HD−1/2 H , which satisﬁes S−1/2 = S−1 . Finally, inequalities between matrices must be understood in terms of positive deﬁniteness; i.e., for matrices A and B, A ≥ B (respectively A > B) means A − B ≥ 0 (respectively A − B > 0). A related decomposition which will prove useful for canonical correlations is the singular value decomposition (SVD). Proposition 1.11 Let A ∈ Rm n of rank A = r. There exists G ∈ Om , H ∈ On such that Dρ 0 A=G H 0 0 where Dρ = diag(ρ1 , . . . , ρr ), ρi > 0, i = 1, . . . , r. Proof. Since A A ≥ 0, there exists H = (h1 , . . . , hn ) ∈ On such that A A = H diag(λ1 , . . . , λr , 0) H ,

1.6. Orthogonal projections

9

where λi > 0, i = 1, . . . , r. For j > r, |Ahj |2 = hj A Ahj = 0 which means Ahj = 0. For j ≤ r, deﬁne ρj = λj and gj = Ahj /ρj . Then, gi gj = hi A Ahj /ρi ρj = δij ; i.e., g1 , . . . , gr are orthonormal. By completing to an orthonormal basis of Rm , we can ﬁnd G = (g1 , . . . , gr , gr+1 , . . . , gm ) ∈ Om . Now, gi Ahj =

or in matrix notation,

G AH =

0, ρj δij ,

Dρ 0

j>r j ≤ r, 0 0

. 2

In the SVD ρ2j , j = 1, . . . , r, are the nonzero eigenvalues of A A and the columns of H are the eigenvectors.

1.6 Orthogonal projections Now recall some basic facts about orthogonal projections. By deﬁnition, an orthogonal projection, P, is simply a linear transformation for which x − Px ⊥ Py, ∀x, y ∈ Rn , but then, equivalently, (x − Px) (Py) = 0, ∀x, y ∈ Rn

⇐⇒ x Py = x P Py, ∀x, y ∈ Rn ⇐⇒ P P = P ⇐⇒ P = P = P2 .

A matrix P such that P = P = P2 is also called an idempotent matrix. Not surprisingly, an orthogonal projection is completely determined by its image. Proposition 1.12 If P1 and P2 are two orthogonal projections, then Im P1 = Im P2 ⇐⇒ P1 = P2 . Proof. It holds since x − P1 x ⊥ P2 y, ∀x, y ∈ Rn =⇒ P2 = P1 P2 , and, similarly, P1 = P2 P1 , whence P1 = P1 = P2 .

2

If X = (x1 , . . . , xk ) is any basis for Im P, we have explicitly P = X(X X)−1 X .

(1.3)

To see this, simply write Px = Xb, and orthogonality, X (x − Xb) = 0, determines the (unique) coeﬃcients b = (X X)−1 X x. In particular, for

10

1. Linear algebra

any orthonormal basis H, P = HH , where H H = Ik . Thus, incidentally, tr P = k and the dimension of the image space is expressed in the trace. However, by this representation we see that for any two orthogonal projections, P1 = HH and P2 = GG , P1 P2 = 0 ⇐⇒ H G = 0 ⇐⇒ G H = 0 ⇐⇒ P2 P1 = 0. Deﬁnition 1.1 P1 and P2 are said to be mutually orthogonal projections iﬀ P1 and P2 are orthogonal projections such that P1 P2 = 0. We write P1 ⊥ P2 when this is the case. Although orthogonal projection and orthogonal transformation are far from synonymous, there is, nevertheless, ﬁnally a very close connection between the two concepts. If we partition any orthogonal transformation H = (H1 , . . . , Hk ), then the brute algebraic fact HH = I = H1 H1 + · · · + Hk Hk represents a precisely corresponding partition of the identity into mutually orthogonal projections. As a last comment on othogonal projection, if P is the orthogonal projection on the subspace V ⊂ Rn , then Q = I−P, which satisﬁes Q = Q = Q2 is also an othogonal projection. In fact, since PQ = 0, then Im Q and Im P are orthogonal subspaces and, thus, Q is the orthogonal projection on V ⊥ .

1.7 Matrix decompositions Denote the groups of triangular matrices with positive diagonal elements as L+ n U+ n

= {T ∈ Rnn : T is lower triangular, tii > 0, i = 1, . . . , n}, = {T ∈ Rnn : T is upper triangular, tii > 0, i = 1, . . . , n}.

An important implication of Proposition 1.2 for matrices is the following matrix decomposition. Proposition 1.13 If A ∈ Rnn is nonsingular, then A = TH for some H ∈ On and T ∈ L+ n . Moreover, this decomposition is unique. Proof. The existence follows from the Gram-Schmidt method applied to the basis formed by the rows of A. The rows of H form the orthonormal basis obtained at the end of that procedure and the elements of T = (tij ) are the coeﬃcients needed to go from one basis to the other. By the GramSchmidt construction itself, it is clear that T ∈ L+ n . For unicity, suppose −1 TH = T1 H1 , where T1 ∈ L+ n and H1 ∈ On . Then, T1 T = H1 H is a + matrix in Ln ∩ On . But, In is the only such matrix (why?). Hence, T = T1 and H = H1 . 2

1.8. Problems

11

A slight generalization of Proposition 1.13 when A ∈ Rpn is of rank A = p is proposed in Problem 1.8.7. Another similar triangular decomposition, known in statistics as the Bartlett decomposition, for positive deﬁnite matrices can now be easily obtained. Proposition 1.14 If S ∈ Pn , then S = TT for a unique T ∈ L+ n. Proof. Since S > 0, then S = HDH , where H ∈ On and D = diag(λi ) 1/2 with λi > 0. Let D1/2 = diag(λi ) and A = HD1/2 . Then, we can write S = AA , where A is nonsingular. From Proposition 1.13, there exists T ∈ L+ n and G ∈ On such that A = TG. But, then, S = TGG T = TT . For −1 + −1 unicity, suppose TT = T1 T1 , where T1 ∈ Ln . Then, T1 TT T1 = I, + 2 which implies that T−1 1 T ∈ Ln ∩ On = {I}. Hence, T = T1 . Other notions of linear algebra such as Kronecker product and “vec” operator will be recalled when needed in the sequel.

1.8 Problems 1. Consider the partitioned matrix S = (sij ) =

S11 S21

S12 . S22

(i) If S11 is nonsingular, prove that |S| = |S11 | · |S22 − S21 S−1 11 S12 |. (ii) For S > 0, prove Hadamard’s inequality, |S| ≤ i sii . (iii) Let S and S11 be nonsingular. Prove that −1 −1 −1 −1 S11 + S−1 −S−1 11 S12 S22.1 S21 S11 11 S12 S22.1 , S−1 = −1 −S−1 S−1 22.1 S21 S11 22.1 where S22.1 = S22 − S21 S−1 11 S12 . (iv) Let S and S22 be nonsingular. Prove that −1 S−1 −S−1 11.2 11.2 S12 S22 , S−1 = −1 −1 −1 −1 −S−1 S−1 22 S21 S11.2 22 + S22 S21 S11.2 S12 S22 where S11.2 = S11 − S12 S−1 22 S21 . Hint: Deﬁne A=

I

−S21 S−1 11

0 I

and B =

I −S−1 11 S12 0 I

and consider the product ASB. 2. Establish with the partitioning (x , x ) , 1 2 S11 S12 S = S21 S22 x =

12

1. Linear algebra

that −1 −1 −1 x S−1 x = (x1 − S12 S−1 22 x2 ) S11.2 (x1 − S12 S22 x2 ) + x2 S22 x2 .

3. For any A ∈ Rpq , B ∈ Rqp , prove the following: (i) |Ip + AB| = |Iq + BA|. Hint: Ip + AB A 0 Iq Ip A 0 Iq + BA

Ip A Ip 0 = , −B Iq B Iq Ip A Ip 0 . = B Iq −B Iq

(ii) The nonzero eigenvalues of AB and BA are the same. 4. Prove Proposition 1.2. 5. Prove Proposition 1.10. 6. Show that if P deﬁnes an orthogonal projection, then the eigenvalues of P are either 0 or 1. 7. Demonstrate the slight generalizations of Proposition 1.13: (i) If A ∈ Rnp is of rank A = p, then A = HT for some T ∈ U+ p and H satisfying H H = Ip . Further, T and H are unique. Hint: For unicity, note that if A = HT = H1 T1 with T1 ∈ U+ p and H1 H1 = Ip , then Im A = Im H = Im H1 and H1 H1 is the orthogonal projection on Im H1 . (ii) If A ∈ Rnp is of rank A = n, then A = TH, where T ∈ L+ n and HH = In . Further, T and H are unique. 8. Assuming A and A + uv are nonsingular, prove (A + uv )−1 = A−1 −

A−1 uv A−1 . (1 + v A−1 u)

9. Vector diﬀerentiation. Let f (x) be a real valued function of x ∈ Rn . Deﬁne ∂f (x)/∂x = (∂f (x)/∂xi ) . Verify (i) ∂a x/∂x = a, (ii) ∂x Ax/∂x = 2Ax, if A is symmetric. 10. Matrix diﬀerentiation [Srivastava and Khatri (1979), p. 37]. Let g(S) be a real-valued function of the symmetric matrix S ∈ Rnn .

1 Deﬁne ∂f (S)/∂S = 2 (1 + δij )∂f (S)/∂sij . Verify (i) ∂tr(S−1 A)/∂S = −S−1 AS−1 , if A is symmetric, (ii) ∂ ln |S|/∂S = S−1 .

1.8. Problems

13

Hint for (ii): S−1 = |S|−1 adj(S). 11. Rayleigh’s quotient. Assume S ≥ 0 in Rnn with eigenvalues λ1 ≥ · · · ≥ λn and corresponding eigenvectors x1 , . . . , xn . Prove: (i) x Sx ≤ λ1 , ∀x = 0. x x (ii) For any ﬁxed j = 2, . . . , n, λn ≤

x Sx ≤ λj , ∀x = 0 x x such that x, x1 = · · · = x, xj−1 = 0. 12. Demonstrate that if A is symmetric and B > 0, then h Ah = λ1 (AB−1 ), |h|=1 h Bh sup

where λ1 (AB−1 ) denotes the largest eigenvalue of AB−1 . 13. Let Am > 0 in Rnn (m = 1, 2, . . .) be a sequence. For any A ∈ Rnn , 2 deﬁne ||A|| = i,j a2ij and let λ1,m ≥ · · · ≥ λn,m be the ordered eigenvalues of Am . Prove that if λ1,m → 1 and λn,m → 1, then limm→∞ ||Am − I|| = 0. 14. In Rp , prove that if |x1 | = |x2 |, then there exists H ∈ Op such that Hx1 = x2 . Hint: When x1 = 0, consider H ∈ Op with ﬁrst row x1 /|x1 |. 15. Show that for any V ∈ Rnn and any m = 1, 2, . . ., (i) if (I − tV) is nonsingular then [Srivastava and Khatri (1979), p. 33] (I − tV)−1 =

m

ti Vi + tm+1 Vm+1 (I − tV)−1 .

i=0

(ii) If V > 0 with eigenvalues λ1 ≥ · · · ≥ λp and |t| < 1/λ1 , then (I − tV)

−1

=

∞ i=0

ti V i .

2 Random vectors

2.1 Introduction A random vector is simply a vector whose components are random variables. The variables are the characteristics of interest that will be observed on each of the selected units in the sample. Questions related to probabilities of a variable to take on some values or probabilities of two or more variables to take on simultaneously values in a set are common in multivariate analysis. Chapter 2 gives a collection of important probability concepts on random vectors such as distribution functions, expected values, characteristic functions, discrete and absolutely continuous distributions, independence, etc.

2.2 Distribution functions First, some basic notations concerning “rectangles” useful to describe the ¯ = R ∪ {±∞} = distribution function of a random vector are given. Let R n ¯ [−∞, ∞]. It is convenient to deﬁne a partial order on R by x ≤ y iﬀ xi ≤ yi , ∀i = 1, . . . , n, and x < y iﬀ xi < yi , ∀i = 1, . . . , n.

2.2. Distribution functions

15

This allows us to express “n-dimensional” rectangles in Rn succinctly: ¯ n. I = (a, b] = {x ∈ Rn : a < x ≤ b} for any a, b ∈ R The interior and closure of I are respectively I ◦ = (a, b) = {x ∈ Rn : a < x < b} and I¯ = [a, b] = {x ∈ Rn : a ≤ x ≤ b} and the boundary of I is the “(n − 1)-dimensional” relative complement ∂I = I¯ − I ◦ . ¯ n ) be denoted by the cartesian Finally let the 2n “corners” of I (a subset of R product a × b = ×ni=1 {ai , bi }. Deﬁnition 2.1 For x distributed on Rn , the distribution function (d.f.) ¯ n. ¯ n → [0, 1], where F (t) = P (x ≤ t), ∀t ∈ R of x is the function F : R This is denoted x ∼ F or x ∼ Fx . A d.f. is automatically right-continuous; thus, if it is known on any dense ¯ n, subset D ⊂ Rn , it is determined everywhere. This is because for any t ∈ R a sequence dn may be chosen in D descending to t: dn ↓ t. From the d.f. may be computed the probability of any rectangle P (a < x ≤ b) = (−1)Na (t) F (t), ∀a < b, t∈a×b

n

where Na (t) = i=1 δ(ai , ti ) counts the number of ti ’s that are ai ’s. The borel subsets of Rn comprise the smallest σ-algebra containing the rectangles B n = σ ((a, b] : a, b ∈ Rn ) . unions of rectangles contains all the The class G n of all countable disjoint ∞ open subsets of Rn , and if we let G = i=1 (ai , bi ] denote a generic element in this class, it follows that P (x ∈ G) =

∞

P (ai < x ≤ bi ).

i=1

By the Caratheodory extension theorem (C.E.T.), the probability of a general borel set A ∈ Bn is then uniquely determined by the formula Px (A) ≡ P (x ∈ A) = inf P (x ∈ G). A⊂G

16

2. Random vectors

2.3 Equals-in-distribution Deﬁnition 2.2 x and y are equidistributed (identically distributed), d

denoted x = y, iﬀ Px (A) = Py (A), ∀A ∈ Bn . On the basis of the previous section, it should be clear that for any dense D ⊂ Rn : d

Proposition 2.1 (C.E.T) x = y ⇐⇒ Fx (t) = Fy (t), ∀t ∈ D. d

Although at ﬁrst glance, = looks like nothing more than a convenient shorthand symbol, there is an immediate consequence of the deﬁnition, deceptively simple to state and prove, that has powerful application in the sequel. Let g : Rn → Ω where Ω is a completely arbitrary space. d

d

Proposition 2.2 (Invariance) x = y =⇒ g(x) = g(y). Proof.

P (g(x) ∈ B) = P x ∈ g −1 (B) = P y ∈ g −1 (B) = P (g(y) ∈ B) . 2

Example 2.1 d

x=y

=⇒

d

xi = yi , i = 1, . . . , n d

=⇒ xi xj = yi yj , i, j = 1, . . . , n n n d =⇒ xri i = yiri , for any ri , i = 1, . . . , n i=1

i=1

=⇒ etc.

2.4 Discrete distributions Deﬁnition 2.3 The probability function (p.f.) of x is the function ¯ n → [0, 1] where p(t) = P (x = t), ∀t ∈ R ¯ n. p:R The p.f. may be evaluated directly from the d.f.: p(t) = lim P (sm < x ≤ t), sm ↑t

where sm ↑ t means s1 < s2 < · · · and sm → t as m → ∞. The subset D = p−1 (0)c where the p.f. is nonzero may contain at most a countable number of points. D is known as the discrete part of x, and x is said to be discrete if it is “concentrated” on D: Deﬁnition 2.4 x is discrete iﬀ P (x ∈ D) = 1.

2.5. Expected values

One may verify that x is discrete ⇐⇒ P (x ∈ A) =

17

p(t), ∀A ∈ Bn .

t∈A∩D

Thus, the distribution of x is entirely determined by its p.f. if and only if it is discrete, and in this case, we may simply write x ∼ p or x ∼ px .

2.5 Expected values For any event A, we may consider the indicator function 1, x ∈ A IA (x) = 0, x ∈ A. It is clear that IA (x) is itself a discrete random variable, referred to as a Bernoulli trial, for which P (IA (x) = 1) = Px (A) and P (IA (x) = 0) = 1 − Px (A). This is denoted IA (x) ∼ Bernoulli (Px (A)) and we deﬁne E IA (x) = Px (A). For any k mutually disjoint and exhaustive events A1 , . . . , Ak and k real numbers a1 , . . . , ak , we may form the simple function s(x) = a1 IA1 (x) + · · · + ak IAk (x). Obviously, s(x) is also discrete with P (s(x) = ai ) = Px (Ai ), i = 1, . . . , k. By requiring that E be linear, we (are forced to) deﬁne E s(x) = a1 Px (A1 ) + · · · + ak Px (Ak ). The most general function for which we need ever compute an expected value may be directly expressed as a limit of a sequence of simple functions. Such a function g(x) is said to be measurable and we may explicitly write g(x) = lim sN (x), N →∞

where convergence holds pointwise, i.e., for every ﬁxed x. If g(x) is nonnegative, it can be proven that we may always choose the sequence of simple functions to be themselves non-negative and nondecreasing as a sequence whereupon we deﬁne E g(x) = lim E sN (x) = sup E sN (x). N →∞

N

Then, in general, we write g(x) as the diﬀerence of its positive and negative parts g(x) = g + (x) − g − (x),

18

2. Random vectors

deﬁned by +

g (x) −

g (x)

= =

g(x), 0, −g(x), 0,

g(x) ≥ 0 g(x) < 0, g(x) ≤ 0 g(x) < 0,

and ﬁnish by deﬁning E g + (x) − E g − (x), if E g + (x) < ∞ or E g − (x) < ∞ E g(x) = “undeﬁned,” otherwise. We may sometimes use the Leibniz notation E g(x) = g(t)dPx (t) = g(t)dF (t). One should verify the fundamental inequality |E g(x)| ≤ E |g(x)|. Let ↑ denote convergence of a monotonically nondecreasing sequence. Something is said to happen for almost all x if it fails to happen on a set A such that Px (A) = 0. The two main theorems concerning “continuity” of E are the following: Proposition 2.3 (Monotone convergence theorem (M.C.T.)) Suppose 0 ≤ g1 (x) ≤ g2 (x) ≤ · · ·. If gN (x) ↑ g(x), for almost all x, then E gN (x) ↑ E g(x). Proposition 2.4 (Dominated convergence theorem (D.C.T.)) If gN (x) → g(x), for almost all x, and |gN (x)| ≤ h(x) with E h(x) < ∞, then E |gN (x) − g(x)| → 0 and, thus, also E gN (x) → E g(x). It should be clear by the process whereby expectation is deﬁned (in stages) that we have d

Proposition 2.5 x = y ⇐⇒ E g(x) = E g(y), ∀g measurable.

2.6 Mean and variance n n i=1 ti xi for each (ﬁxed) t ∈ R ,

n 1/2 2 and the “euclidean norm” (length) |x| = . By any of three i=1 xi equivalent ways, for p > 0 one may say that the pth moment of x is ﬁnite: Consider the “linear functional” t x =

E |t x|p < ∞, ∀t ∈ Rn

⇐⇒ E |xi |p < ∞, i = 1, . . . , n ⇐⇒ E |x|p < ∞. n To show this, one must realize that |xi | ≤ |x| ≤ i=1 |xi | and Lp = {x ∈ Rn : E |x|p < ∞} is a linear space (v. Problem 2.14.3). From the simple inequality ar ≤ 1 + ap , ∀a ≥ 0 and 0 < r ≤ p, if we let a = |x| and take expectations, we get E |x|r ≤ 1 + E |x|p . Hence, if for

2.6. Mean and variance

19

p > 0, the pth moment of x is ﬁnite, then also the rth moment is ﬁnite, for any 0 < r ≤ p. A product-moment of order p for x = (x1 , . . . , xn ) is deﬁned by E

n

xpi i , pi ≥ 0, i = 1, . . . , n,

i=1

n

pi = p.

i=1

A useful inequality to determine that a product-moment is ﬁnite is H¨ older’s inequality: Proposition 2.6 (H¨ older’s inequality) For any univariate random variables x and y, 1 1 + = 1. r s n From this inequality, if the pth moment of x ∈ R is ﬁnite, then all productmoments of order p are also ﬁnite. This can be veriﬁed for n = 2, as H¨older’s inequality gives 1/r

E |xy| ≤ (E |x|r )

p1 /p

E |xp11 xp22 | ≤ (E |x1 |p )

1/s

· (E |y|s )

p2 /p

· (E |x2 |p )

, r > 1,

, pi ≥ 0, i = 1, 2, p1 + p2 = p.

The conclusion for general n follows by induction. If the ﬁrst moment of x is ﬁnite we deﬁne the mean of x by def µ = E x = (E xi ) = (µi ). If the second moment of x is ﬁnite, we deﬁne the variance of x by def Σ = var x = (cov(xi , xj )) = (σij ) . In general, we deﬁne the expected value of any multiply indexed array of univariate random variables, ξ = (xijk··· ), componentwise by E ξ = (E xijk··· ). Vectors and matrices are thus only special cases and it is obvious that Σ = E (x − µ)(x − µ) = E xx − µµ . It is also obvious that for any A ∈ Rm n, E Ax = Aµ and var Ax = AΣA . In particular, E t x = t µ and var t x = t Σt ≥ 0, ∀t ∈ Rn . Now, the reader should verify that more generally cov(s x, t x) = s Σt and that considered as a function of s and t, the left-hand side deﬁnes a (pseudo) inner product. Thus, Σ is automatically positive semideﬁnite, Σ ≥ 0. But by this, we may immediately write Σ = HDH with H orthogonal and D = diag(λ), where the columns of H comprise an orthonormal basis of “eigenvectors” and the components of λ ≥ 0 list the corresponding

20

2. Random vectors

“eigenvalues.” Accordingly, we may always “normalize” any x with Σ > 0 by letting z = D−1/2 H (x − µ), which represents a three-stage transformation of x in which we ﬁrst relocate −1/2 independently along by µ, then rotate by H , and, ﬁnally, rescale by λi each axis. We ﬁnd, of course, that E z = 0 and var z = I. The linear transformation z = Σ−1/2 (x − µ) also satisﬁes E z = 0 and var z = I. When the vector x ∈ Rn is partitioned as x = (y , z ) , where y ∈ Rr , z ∈ Rs , and n = r + s, it is useful to deﬁne the covariance between two vectors. The covariance matrix between y and z is, by deﬁnition, cov(y, z) = (cov(yi , zj )) ∈ Rrs . Then, we may write

var(x) =

var(y) cov(z, y)

cov(y, z) var(z)

.

Sometimes, expected value of y is easier to calculate by conditioning on another random vector z. In this regard, the conditional mean theorem and conditional variance theorem are stated. A general proof of the conditional mean theorem can be found in Billingsley (1995, Section 34). Proposition 2.7 (Conditional mean formula) E[E(y|z)] = E y. An immediate consequence is the conditional variance formula. Proposition 2.8 (Conditional variance formula) var y = E[var(y|z)] + var[E(y|z)]. Example 2.2 Deﬁne a group variable I such that P (I = 1) = 1 − , P (I = 2) = . Conditionally on I, assume x|I = 1 ∼ N (µ1 , σ12 ), x|I = 2 ∼ N (µ2 , σ22 ).

Then fx (x)

=

1 1 exp − 2 (x − µ1 )2 σ1 2σ1 1 1 +(2π)−1/2 exp − 2 (x − µ2 )2 σ2 2σ2 (1 − )(2π)−1/2

2.7. Characteristic functions

21

is a mixture or -contaminated normal density. It follows from the construction of x that E x = E[E(x|I)] = (1 − )µ1 + µ2 ≡ µ, var x = E[var(x|I)] + var[E(x|I)] = (1 − )σ12 + σ22 + (1 − )(µ1 − µ)2 + (µ2 − µ)2 .

2.7 Characteristic functions We require only the most basic facts about characteristic functions. Deﬁnition 2.5 The characteristic function of x is the function c : Rn → C deﬁned by

c(t) = cx (t) = E eit x . Note: 1. c(0) = 1, |c(t)| ≤ 1 and c(−t) = c(t). 2. c(t) is uniformly continuous:

|c(t) − c(s)| = E ei(t−s) x − 1 eis x ≤ E ei(t−s) x − 1 .

Since ei(t−s) x − 1 ≤ 2, continuity follows by the D.C.T. Uniformity holds since ei(t−s) x − 1 depends only on t − s. The main result is perhaps the “inversion formula” proven in Appendix A: 1 N →∞ (2π)n

Px (a, b] = lim

2

e−it x c(t)e−t t/2N dtdx,

Rn

(a,b]

∀a, b such that Px (∂(a, b]) = 0. Thus, the C.E.T. may be applied immediately to produce the technically equivalent: d

Proposition 2.9 (Uniqueness) x = y ⇐⇒ cx (t) = cy (t), ∀t ∈ Rn . Now if we consider the linear functionals of x: t x with t ∈ Rn , it is clear that ct x (s) = cx (st), ∀s ∈ R, t ∈ Rn , so that the characteristic function of x determines all those of t x, t ∈ Rn and vice versa. Let S n−1 = {s ∈ Rn : |s| = 1} be the “unit sphere” in Rn , and we have d

d

Proposition 2.10 (Cram´ er-Wold) x = y ⇐⇒ t x = t y, ∀t ∈ S n−1 .

22

2. Random vectors

Proof. Since ct x (s) = cx (st), ∀s ∈ R, t ∈ Rn , it is clear that x = y ⇐⇒ t x = t y, ∀t ∈ Rn . d

Since t x = |t|

t |t|

d

x, ∀t = 0, it is also clear that

t x = t y, ∀t ∈ Rn ⇐⇒ t x = t y, ∀t ∈ S n−1 . d

d

2 By this result, it is clear that one may reduce a good many issues concerning random vectors to the univariate level. In the speciﬁc matter of computation, the reader should know that in the special case of a univariate random variable X: If E e±δX < ∞ for any δ > 0, the Laplace transform of X is determined in the strip |Re(z)| ≤ δ as the (absolutely convergent) power series LX (z) =

∞

E X n z n /n!,

n=0

and since such a power series is completely determined by its coeﬃcients, we ﬁnd that one may legitimately obtain the characteristic function cX (t) = LX (it), ∀t ∈ R, by merely observing the coeﬃcients in an expansion of the moment-generating function since they are necessarily the same as those of the Laplace transform: mX (t) = LX (t), ∀|t| ≤ δ. Example 2.3 Suppose fz (s) = (2π)−1/2 e−s moment-generating function (m.g.f.), ﬁnding 2

mz (t) = et

/2

2

/2

. One easily computes the

,

which has the obvious expansion for every t, whereupon 2

cz (t) = e−t

/2

.

(2.1)

2.8 Absolutely continuous distributions Lebesgue measure, λ, is the extension to all borel sets of our natural sense of volume measure in Rn . Thus, we deﬁne λ(a, b]

=

n i=1

(bi − ai ), ∀a < b in Rn ,

2.8. Absolutely continuous distributions

λ(G)

=

∞

λ(ai , bi ], ∀G =

i=1

∞

23

(ai , bi ] in G n ,

i=1

and λ(A) = inf λ(G), ∀A in Bn . A⊂G

As before, the C.E.T. guarantees that λ is a measure on B n . We will often denote Lebesgue measure explicitly as volume: λ(A) = vol(A). Incidentally, something is said to happen “almost everywhere” (a.e.) if the set where it fails to happen has zero volume. Now, the general conception of a random vector continuously distributed in space is that the probabilities of events will depend continuously on the volume of the events. Thus, Deﬁnition 2.6 x is absolutely continuous, denoted x λ, iﬀ ∀ > 0, ∃δ > 0 such that vol(A) < δ =⇒ P (x ∈ A) < . But, in that case, Proposition 2.11 x λ ⇐⇒ vol(A) = 0 =⇒ P (x ∈ A) = 0. Proof. Assume x λ. If vol(A) = 0 but P (x ∈ A) = 0 we may take = P (x ∈ A)/2 to ﬁnd the contradiction that P (x ∈ A) < . Conversely, suppose vol(A) = 0 =⇒ P (x ∈ A) = 0 but that x λ. Then, ∃0 > 0 such that ∀n, ∃An with vol(An ) < 1/2n but P (x ∈ An ) ≥ 0 . Letting A = limAn = ∩∞ n=1 ∪k≥n Ak , since ∪k≥n Ak is a monotone sequence we ﬁnd 2 the contradiction that vol(A) = 0 but P (x ∈ A) ≥ 0 . Thus, a distribution which depends continuously on volume satisﬁes the relatively simple criterion x λ ⇐⇒ vol(A) = 0 =⇒ P (x ∈ A) = 0. However, it is on this particular criterion, by the theorem of RadonNikodym, that absolute continuity is characterized ﬁnally in terms of densities: Proposition 2.12 (Radon-Nikodym) x is absolutely continuous ⇐⇒ there is a (a.e.-unique) probability density function (p.d.f.) f : Rn → [0, ∞) such that f (t)dt, ∀A ∈ Bn . P (x ∈ A) = A

But since the p.d.f. then determines such a distribution completely, we may simply write x ∼ f or x ∼ fx . It is, of course, by the extension process that deﬁnes expectation (in stages) that automatically E g(x) = g(t)f (t)dt, ∀g measurable,

24

2. Random vectors

such that E g(x) is deﬁned. Now in particular, the distribution function may itself be expressed as t F (t) = P (x ≤ t) = f (s)ds, ∀t ∈ Rn . −∞ In practice, we will often be able to invoke the fundamental theorems of calculus to obtain an explicit representation of the p.d.f. by simply diﬀerentiating the d.f.: 1. By the ﬁrst fundamental theorem of calculus, f (t) = ∂ n F (t)/∂t1 · · · ∂tn at every t where f (t) is continuous. 2. Also, by the second fundamental theorem, if f (t) = ∂ n F (t)/∂t1 · · · ∂tn exists and is continuous (a.e.) on some rectangle I, then P (x ∈ A) = f (t)dt, ∀A ⊂ I. A

Finally, in relation to the inversion formula, when the characteristic func tion c(t) is absolutely integrable, i.e., Rn |c(t)|dt < ∞, the corresponding distribution function is absolutely continuous with p.d.f. (v. Appendix A): 1 f (s) = e−it s c(t)dt. (2.2) n (2π) Rn

2.9 Uniform distributions The most fundamental absolutely continuous distribution would, of course, be conveyed by volume measure itself. Consider any event C for which 0 < vol(C) < ∞. Deﬁnition 2.7 x is uniformly distributed on C, denoted x ∼ unif(C), iﬀ P (x ∈ A) = vol(AC)/vol(C), ∀A ∈ Bn . If vol(∂C) = 0, as is often the case, we may just as well include as exclude ¯ then x, y, and z are it, so if x ∼ unif(C), y ∼ unif(C ◦ ) and z ∼ unif(C), equidistributed: d

d

x = y = z. Now, for x ∼ unif(C) we may immediately reexpress each probability as an integral: P (x ∈ A) = k · IC (t)dt, ∀A ∈ Bn , A

2.10. Joints and marginals

where

IC (t) =

25

1, t ∈ C 0, t ∈ C,

is the indicator function for C and k = vol(C)−1 . We thus have an explicit determination of “the” density for x: f (t) = k IC (t), ∀t ∈ Rn . Example 2.4 For x ∼ unif([0, 1]) on Rn , the p.d.f. may be expressed as a simple product f (t) = I[0,1] (t) =

n

I[0,1] (ti ), ∀t ∈ Rn ,

i=1

from which F (t) =

n

ti I[0,1] (ti ) + I(1,∞) (ti ) , ∀t ∈ Rn .

i=1

2.10 Joints and marginals Consider xi ∼ Fi on Rni , i = 1, . . . , k, with x1 k .. n x= ∼ F on R where n = ni . . i=1 xk x is called the joint of x1 , . . . , xk which are, in turn, called marginals of x. Since it is clear that

t1 . P (x ≤ t) = P (x1 ≤ t1 , . . . , xk ≤ tk ), ∀t = .. , tk

we will, by a slight abuse of our notation, write

t1 . F (t) = F (t1 , . . . , tk ), ∀t = .. tk

to reﬂect this “partitioning.” In this way, the distribution function is said to express the joint distribution of x1 , . . . , xk , and the marginals may be recovered on the simple substitution of ∞ in all but the ith place: Fi (s) = F (∞, . . . , s, . . . , ∞), ∀s ∈ Rni .

26

2. Random vectors

In the special case where x is absolutely continuous with p.d.f. f (t) = f (t1 , . . . , tk ), it follows that each xi is also absolutely continuous with p.d.f. fi (s) that is obtained by “integrating out” the other variables: ∞ ∞ fi (s) = dtj , ∀s ∈ Rni . ··· f (t1 , . . . , s, . . . , tk ) −∞ −∞ 1≤j≤k j=i

This is by direct application of Fubini’s theorem whereby we may interchange the order of integration in a product integral to verify that fi (s)ds, ∀A ∈ Bni P (xi ∈ A) = A

and, of course, in particular,

Fi (s) =

s

fi (u)du, ∀s ∈ Rni .

−∞ Koehler and Symanowski (1995) presented a method for constructing multivariate distributions with any speciﬁc set of univariate marginals. It provides a rich class of distributions for modeling multivariate data as well as a basis for easily simulating correlated observations. The inclusion of diﬀerent association parameters for diﬀerents subsets of variables allows for many diﬀerent patterns of associations. Their work follows those of Genest and MacKay (1986) and Marshall and Olkin (1988), among others. A tool called linkage [Li et al. (1996)] can be used for the construction of multivariate distributions with given multivariate marginals; Cuadras (1992) found related results.

Example 2.5 The bivariate parametric family of d.f.’s on [0, 1]2 of Cook and Johnson (1981) is deﬁned by −1/α 1 1 , α > 0. (2.3) F (t1 , t2 ; α) = α + α − 1 t1 t2 The case α = 0 can be deﬁned by continuity. It has marginals F1 (t1 )

= F (t1 , 1; α) = t1 ,

F2 (t2 )

= F (1, t2 ; α) = t2 ,

which are identically distributed as unif([0, 1]). Multivariate distributions on [0, 1] with uniform marginals are often referred to as copulas. The slight modiﬁcation −1/α 1 1 + − 1 , α > 0, F (t1 , t2 ; α) = F1 (t1 )α F2 (t2 )α is a bivariate distribution with arbitrary marginals F1 and F2 . The bivariate parametric family of d.f.’s on [0, 1]2 of Frank (1979) [v. also Genest (1987)] (αt1 − 1)(αt2 − 1) F (t1 , t2 ; α) = logα 1 + , α > 0, (2.4) (α − 1)

2.11. Independence

27

0.2 2 0.1 1 0 0

-2 -1

-1

0 1

-2 2

Figure 2.1. Bivariate Frank density with standard normal marginals and a correlation of 0.7.

(the case α = 1 can be deﬁned by continuity), where logα (·) denotes logarithm in base α, is also a copula. Such distributions have found applications in modeling survival data [Oakes (1982), Carri`ere (1994)]. Figure 2.1 is a graph of a bivariate Frank density with standard normal marginals. The association parameter α = 0.00296 using Nelsen (1986) corresponds to a correlation of 0.7.

2.11 Independence Deﬁnition 2.8 x1 , . . . , xk are mutually statistically independent iﬀ P (x1 ∈ A1 , . . . , xk ∈ Ak ) =

k

P (xi ∈ Ai ), ∀Ai ∈ Bni , i = 1, . . . , k. |=

i=1

x2 . By the extension Denote pairwise independence (k = 2) simply x1 process that deﬁnes expectation, we ultimately ﬁnd: Proposition 2.13 x1 , . . . , xk are independent ⇐⇒ E

k i=1

∀g1 , . . . , gk such that E |gi (xi )| < ∞.

gi (xi ) =

k i=1

E gi (xi ),

28

2. Random vectors

and also (chieﬂy by the C.E.T.) Proposition 2.14 x1 , . . . , xk are independent ⇐⇒ F (t) =

k

Fi (ti ), ∀t ∈ Rn .

i=1

In the special case where each xi is absolutely continuous with p.d.f. fi (ti ), we may conclude that x is as well, and we have: Proposition 2.15 x1 , . . . , xk are independent ⇐⇒ f (t) =

k

fi (ti ), ∀t ∈ Rn .

i=1

Finally, independence may also be characterized: Proposition 2.16 x1 , . . . , xk are independent ⇐⇒ cx (t) =

k

cxi (ti ), ∀t ∈ Rn .

i=1

Example 2.6 For x ∼ unif([0, 1]) on R , it is clear that x1 , . . . , xn are independently and identically distributed (i.i.d.) as unif([0, 1]). n

|=

Example 2.7 Let x = (x1 , x2 ) have Frank’s d.f. (2.4). Proposition 2.14 x2 iﬀ α = 1. yields, after elementary calculus, x1

2.12 Change of variables We recall some basic calculus [Spivak (1965), p. 16]. Let A ⊂ Rn be open. Deﬁnition 2.9 The derivative of φ : A → Rm , at x ∈ A, is the unique linear transformation φ (x) ∈ Rm n such that φ(x + h) − φ(x) = φ (x)h + o(h) or, equivalently, lim

h→0

φ(x + h) − φ(x) − φ (x)h |h|

= 0.

When φ (x) exists, all partial derivatives ∂φi (x)/∂xj exist. This determines the derivative componentwise as φ (x) = (∂φi (x)/∂xj ) . A condition for φ (x) to exist is that all partial derivatives ∂φi (x)/∂xj exist in an open neighborhood of x and are continuous at x. There are, of course, various notations for derivatives, all acceptable: φ (x) = Dφ(x) = ∂φ(x)/∂x.

2.12. Change of variables

29

The derivative satisﬁes the “chain rule” (φ ◦ ψ) (x) = φ (ψ(x)) ψ (x). In the very special case m = n, the jacobian of φ : Rn → Rn is, by deﬁnition, the absolute value of the determinant of φ (x) and is denoted by |φ (x)|+ . Another common notation for the jacobian of the transformation y = φ(x) is J(y → x) = |φ (x)|+ . From the chain rule, it is made clear that if z = φ(y) and y = ψ(x), then J(z → x) = J(z → y) · J(y → x), J(y → x) = [J(x → y)]−1 . At any rate, we have an important and general result, easy to state, but the proof of which is by no means trivial [Spivak (1965), p. 67]. Proposition 2.17 Let φ : A → Rn be one-to-one and continuously diﬀerentiable on A. If f : φ(A) → R is integrable, then f (x)dx = f (φ(y)) |φ (y)|+ dy. φ(A) A It is this result that is applied directly to obtain the standard “change of variables” formula for absolutely continuous random vectors. Proposition 2.18 If x ∼ f on Rn and C = {x : f (x) > 0} is open, for any φ : C → Rn one-to-one and bi-diﬀerentiable with inverse ψ : φ(C) → C, let y = φ(x). Then, y ∼ g with g(y) = f (ψ(y)) |ψ (y)|+ . Proof. P (y ∈ B)

= P (φ(x) ∈ B) = P (x ∈ ψ(B)) = f (x)dx = f (ψ(y)) |ψ (y)|+ dy. ψ (B) B 2

By an abuse of notation, y (and x) have two diﬀerent meanings in Proposition 2.18: y is a random vector in y ∼ g, whereas it is any given point of Rn in the density g(y). Now, if the function φ in question is simply a linear transformation, φ(x) = Ax, it is already its own derivative everywhere on Rn , φ (x) = A, and the formula for change of variables greatly simpliﬁes. Suppose that A : Rn → Rn is a nonsingular transformation. The group of all such nonsingular transformations is known as the general linear group and denoted by Gn = {A ∈ Rnn : A is nonsingular} = {A ∈ Rnn : |A| = 0}.

30

2. Random vectors

Two examples are as follows: Example 2.8 If x ∼ f and y = Ax, we ﬁnd y ∼ g, where g(y) = f (A−1 y) |A|−1 + . Example 2.9 x ∼ unif(C), C ⊂ Rn =⇒ Ax + b ∼ unif(AC + b) where AC + b = {Ax + b : x ∈ C}.

2.13 Jacobians The derivation of jacobians is the diﬃcult part in making transformations. It can be a daunting task. This section is directed to the derivation of more complicated jacobians. It can be skipped on a ﬁrst reading and consulted when needed in the sequel. Although jacobians are useful for densities, our approach is to derive distributions without appealing, whenever possible, to densities. Derivations of densities appear mainly in the form of problems. Proposition 2.19 The jacobian of the transformation V = AWA , W ∈ Rnn symmetric and A ∈ Rnn constant, is J(V → W) = |A|n+1 + . Proof. The transformation is linear and, thus, the jacobian is necessarily a polynomial in the elements of A, p(A) say. If W = BUB , then from the chain rule, we have J(V → U) = J(V → W) · J(W → U), i.e., p(AB) = p(A)p(B). The only polynomials in the elements of a matrix satisfying this multiplicative rule are the integer powers of the determinant [MacDuﬀy (1943, p. 50)]. Hence, p(A) = |A|k , for some integer k. We can ﬁnd k by choosing A = aI. Since V = a2 W and there are 12 n(n+1) distinct elements, then J(V → W) = an(n+1) = |aI|n+1 . We found k = n + 1. 2 James (1954) also used MacDuﬀy’s characterization of the determinant for skew-symmetric matrices. At this point, we make some comments concerning the diﬀerential of a function of several variables. Our development here closely resembles that of Srivastava and Khatri (1979, p. 26). For n a real-valued nfunction y = f (x), x ∈ R , the diﬀerential is deﬁned as dy = df = i=1 ∂f (x)/∂xi · dxi . For a vector-valued function y = f (x), x and y in Rn , the diﬀerential is deﬁned componentwise, i.e., n df1 i=1 ∂f1 (x)/∂xi · dxi . .. dy = df = .. = = ∂f (x)/∂x · dx, . n dfn i=1 ∂fn (x)/∂xi · dxi where ∂f (x)/∂x = (∂fi (x)/∂xj ) ∈ Rnn is the usual derivative of f (x). Hence, dy is a linear function of dx with jacobian J(dy → dx) = |∂f (x)/∂x|+ = J(y → x).

2.13. Jacobians

31

Note that x and y could be replaced by any “vectorized” array or matrix. For example, for F(X) = (fij (X)) ∈ Rm n , we can deﬁne the diﬀerential componentwise, i.e., dF = (dfij ). The reader can then check (v. Problem 2.14.15) F = GH =⇒ dF = G · dH + dG · H. As an example, consider the inverse transformation. Proposition 2.20 The jacobian of the transformation V = W−1 , W ∈ −(n+1) . Rnn nonsingular and symmetric, is J(V → W) = |W|+ Proof. Since VW = I, then V · dW + dV · W = 0, which implies dV = −W−1 · dW · W−1 . Hence, from Proposition 2.19, J(V → W) = J(dV → −(n+1) dW) = |W|+ . 2 The jacobian of “conditional transformations” [Srivastava and Khatri (1979), p. 29], used to prove Propositions 2.22 and 2.23, may provide simpliﬁcations in some cases. Proposition 2.21 Let xi and yi in Rpi , i = 1, . . . , r, be related through the system of “conditional transformations” y1 = f1 (x1 ), y2 = f2 (y1 , x2 ), .. . yr = fr (y1 , . . . , yr−1 , xr ), where each fi is diﬀerentiable. Then, J(y1 , . . . , yr → x1 , . . . , xr ) =

r

J(yi → xi ).

i=1

Proof. The jacobian has the triangular form 0 ··· 0 ∂y1 /∂x1 0 ∗ ∂y2 /∂x2 · · · , J = . .. .. .. . . . . . ∗ ∗ · · · ∂yr /∂xr + r and, thus, we get J = i=1 |∂yi /∂xi |+ immediately.

2

As an example of jacobian via conditional transformations, consider the Bartlett decomposition of W > 0 in Rnn as W = TT for a unique T ∈ L+ n (v. Proposition 1.14). Due to symmetry, W has eﬀectively n(n + 1)/2 elements and, thus, the decomposition gives a transformation f : Rn(n+1)/2 → Rn(n+1)/2 deﬁned by f (W) = T.

32

2. Random vectors

Proposition 2.22 The jacobian of the transformation f (W) = T is J(W → T) = 2n

n

tn−i+1 . ii

i=1

Proof. Partition W and T in conformity so that w11 w21 t11 0 t11 = w21 W22 t21 T22 0

t21 T22

.

Observe the system of conditional transformations w11

= t211 ,

w21

= w11 t21 ,

W22

1/2

= w21 w21 /w11 + T22 T22

from which 1/2

J(W → T) = (2t11 )(w11 )n−1 J(W22 → T22 ) = 2tn11 J(W22 → T22 ). 2

The conclusion follows by induction.

As another example, consider the transformation to polar coordinates on Rn , x → (r, θ1 , . . . , θn−1 ) given by x1 x2 x3 .. . xn−1 xn

= r sin(θ1 ) sin(θ2 ) · · · sin(θn−2 ) sin(θn−1 ), = r sin(θ1 ) sin(θ2 ) · · · sin(θn−2 ) cos(θn−1 ), = r sin(θ1 ) sin(θ2 ) · · · cos(θn−1 ),

= r sin(θ1 ) cos(θ2 ), = r cos(θ1 ),

where r > 0 is the “radius” and 0 < θi ≤ π, i = 1, . . . , n − 2, 0 < θn−1 ≤ 2π are the “angles”. The jacobian J(x → r, θ) is facilitated with the system of conditional transformations y1 y2 y3 .. . yn−1 yn

= x21 + · · · + x2n−2 + x2n−1 + x2n = x21 + · · · + x2n−2 + x2n−1 = x21 + · · · + x2n−2

= r2 , = y1 sin2 (θ1 ), = y2 sin2 (θ2 ),

= x21 + x22 = x21

= yn−2 sin2 (θn−2 ), = yn−1 sin2 (θn−1 ).

Proposition 2.23 The jacobian of the transformation to polar coordinates in Rn is J(x → r, θ) = rn−1 sinn−2 (θ1 ) sinn−3 (θ2 ) · · · sin(θn−2 ).

2.14. Problems

33

Proof. We give the main idea and the reader is asked in Problem 2.14.11 to complete the details. We have J(y → r, θ) = J(y → x) · J(x → r, θ). The jacobian J(y → x) is trivial and J(y → r, θ) is evaluated using Proposition 2.21 on conditional transformations. 2 Let S n−1 = {s ∈ Rn : |s| = 1} be the “unit sphere” in Rn . The superscript n − 1 refers to the dimension of this surface. At times, we would like to bypass the angles and consider directly the transformation f : Rn \{0} → (0, ∞) × S n−1 , x → (r, u) deﬁned by r = |x| and u = x/|x| ∈ S n−1 . Since [Courant (1936), p. 302] R g(|x|)dx = g(r)rn−1 drdu, 0

|x|≤R

S n−1

where du is the “area element” of S n−1 , then rn−1 is the jacobian. Proposition 2.24 The jacobian of the transformation x → (r, u) is J(x → r, u) = rn−1 . The jacobians of other transformations on k-surfaces (manifolds) in Rn are useful for sampling distributions of eigenvalues, for example, but their full understanding requires a knowledge of diﬀerential forms and integration on manifolds [Spivak (1965), James (1954)]. This will not be pursued here.

2.14 Problems 1. Show that |E g(x)| ≤ E |g(x)| for any g : Rn → R such that E |g(x)| < ∞. 2. Prove the Cr inequality: For x and y distributed on Rk , E |x + y|r ≤ Cr [E |x|r + E |y|r ] , r > 0, where

Cr =

1, 0 0, a ≥ 0, b ≥ 0. 3. For each p > 0 let Lp denote the collection of all random vectors on Rk for which the pth moment exists: E |x|p < ∞. Prove the following basic facts: (i) Lp is a vector space. (ii) For any 0 < r ≤ p, Lp ⊆ Lr .

34

2. Random vectors

(iii) E |x|p < ∞ ⇐⇒ E |xi |p < ∞, i = 1, . . . , k ⇐⇒ E |t x|p < ∞, ∀t ∈ Rk . (iv) E |a x|p = 0, for some a ∈ Rk =⇒ P (x ∈ a⊥ ) = 1. (v) For any x ∈ L1 , |E x| ≤ E |x|. Indicate also the precise circumstances under which equality occurs. 4. Prove that if the pth moment (p > 0) of x ∈ Rn is ﬁnite, then all product-moments of x of order p are ﬁnite. 5. For x distributed on Rn , consider the p.d.f. fx (x) = c|x|2 · I[0,1] (x). (i) Determine c. (ii) Determine E xand var x. n i (iii) Determine E i=1 xi . Hint: E g(x) = cE |u|2 g(u), where u ∼ unif([0, 1]). p m n n 6. Let A ∈ Rm n , B ∈ Rq , and C ∈ Rq be constant and X ∈ Rp , x ∈ R , q and y, z ∈ R be random. Check the following:

(i) E(AXB + C) = A(E X)B + C (ii) cov(Ax, By) = A cov(x, y)B (iii) cov(x, y + z) = cov(x, y) + cov(x, z). 7. Prove the conditional variance formula. 8. Pairwise versus mutual independence [Bhat (1981)]. Let x and y be i.i.d. random variables taking the values +1 and −1 with probability 1/2 each. Deﬁne z = xy.

|=

|=

|=

i) Establish x, y, z are pairwise independent, but not mutually independent. ii) Does x z and y z imply xy z? |=

9. Let x = (x1 , x2 ) have the d.f. of Cook and Johnson (1981) as in x2 iﬀ α = 0. expression (2.3). Demontrate x1 10. Given a bivariate copula d.f. C(t1 , t2 ), two measures of association are Spearman’s ρ and Kendall’s τ , t1 t2 dC(t1 , t2 ) − 3, ρ = 12 [0,1]2 C(t1 , t2 )dC(t1 , t2 ) − 1, τ = 4 [0,1]2

|ρ| ≤ 1 and |τ | ≤ 1. Now, let |α| < 1/3 in the bivariate Morgenstern copula C(t1 , t2 ) = t1 t2 [1 + 3α(1 − t1 )(1 − t2 )]. Verify this copula is parameterized by Spearman’s measure, or α = 12 t1 t2 dC(t1 , t2 ) − 3. [0,1]2

2.14. Problems

35

11. Complete the proof of Proposition 2.23. 12. Demonstrate the jacobian of the transformation T1 = AT 2 ∈ 2nfor T + i L+ n and A = (aij ) constant also in Ln is J(T1 → T2 ) = i=1 aii . 13. Demonstrate the jacobian of the transformation U1 = AU2 for U2 ∈ + U+ n and A = (aij ) constant also in Un is J(U1 → U2 ) =

n

n−i+1 aii .

i=1

14. Demonstrate the jacobian of the transformation V = AWA , where W ∈ Rnn is skew-symmetric, i.e., W = −W , is J(V → W) = |A|n−1 + . 15. Suppose F(X) = (fij (X)) ∈ Rm n and deﬁne the diﬀerential componentwise, i.e., dF = (dfij ). Demonstrate that F = GH =⇒ dF = G · dH + dG · H. 16. Let

V=

V11 V21

V12 V22

>0

and deﬁne the transformation f : (V11 , V12 , V22 ) → (V11.2 , V12 , V22 ), −1 where V11.2 = V11 − V12 V22 V21 .

(i) Prove f deﬁnes a one-to-one mapping. (ii) Obtain J(V11 , V12 , V22 → V11.2 , V12 , V22 ) = 1.

3 Gamma, Dirichlet, and F distributions

3.1 Introduction This chapter introduces some basic probability distributions useful in statistics. The gamma distribution, in particular, is the building block of many other distributions such as chi-square, F , and Dirichlet. The Dirichlet distribution, as deﬁned in Section 3.3, has the important physical interpretation of proportion of time waited in a Poisson process. However, it has other applications such as the distribution of spacing variables (v. Problem 3.5.3) and the distribution theory (v. Section 4.5) related to spherical distributions, which play an important role in robustness.

3.2 Gamma distributions Deﬁnition 3.1 Standard gamma: z ∼ gamma(p) or z ∼ G(p) on p > 0 “degrees of freedom” iﬀ Γ(p)−1 z p−1 e−z , z > 0 fz (z) = 0, z ≤ 0. The integrating constant, as it depends on p > 0, is known as the gamma function and is, in fact, deﬁned by ∞ Γ(p) = tp−1 e−t dt, p > 0. 0

3.2. Gamma distributions

One may verify some basic properties: Γ(p + 1) = pΓ(p), Γ(2) = Γ(1) = 1, Γ( 12 ) =

√

37

π

and, in particular, Γ(n) = (n − 1)!. Obviously, E z r = Γ(p + r)/Γ(p), ∀r > −p, so that E z = var z = p. A more general gamma distribution is obtained by simply rescaling the standard gamma. Deﬁnition 3.2 Scaled gamma: x ∼ gamma(p, θ) or x ∼ G(p, θ) on p > 0 “degrees of freedom” and “scale” θ > 0 iﬀ x = θz, z ∼ G(p). Obviously,

fx (x) =

Γ(p)−1 θ−p xp−1 e−x/θ , 0,

x>0 x ≤ 0,

and, thus, the characteristic function is cx (t) = (1 − iθt)−p with the “convolution” of gamma distributions as a corollary. Corollary 3.1 If xi , i = 1, . . . , n, are independent G(pi , θ), then n n xi ∼ G pi , θ . i=1

i=1

In the special case where p = 1 we have the exponential distributions. Deﬁnition 3.3 Standard exponential : Scaled exponential :

z ∼ exp(1) iﬀ z ∼ G(1). x ∼ exp(θ) iﬀ x = θz, z ∼ exp(1).

The chi-square distribution is another special case. d

Deﬁnition 3.4 Chi-square: y ∼ χ2m or y = χ2m iﬀ y = 2z, z ∼ G( 12 m). Equivalently, y ∼ χ2m iﬀ y ∼ G( 12 m, 2), and the chi-square is a special case of the scaled gamma above. Thus, the gamma distribution occurs in common statistical practice as the chi-square (2×gamma≡chi-square). The characteristic function of y ∼ χ2m is immediate: cy (t) = (1 − i2t)−m/2 . One should, however, also recall how it describes “waiting time” in a Poisson process. Recall that the Poisson process Nt arises ﬁrst on purely physical considerations as a description of the number of “successes” in what is eﬀectively an inﬁnite number of independent bernoulli trials over the ﬁxed time period t where the average number is known to be proportional to t. On these

38

3. Gamma, Dirichlet, and F distributions

assumptions, Xn ∼ binomial(n, pn ) and E Xn → λt, whereby d

Xn → Nt , where Nt ∼ Poisson(λt). One then has the (conjugate) waiting time process Tn to describe the amount of time to wait until at least n “successes.” Since Tn > t ⇐⇒ Nt < n, we ﬁnd P (Tn > t) = P (Nt < n) =

n−1

e−λt (λt)i /i!

i=0

and diﬀerentiating produces the p.d.f. fn (t) = λe−λt (λt)n−1 /(n − 1)!, t > 0, whereby we discover Tn ∼ G(n, λ−1 ) or, equivalently, zn = λTn ∼ G(n). The exponential itself is just as well predicated on a diﬀerent intuition in that one may show that it is the unique distribution that has “no memory” in the explicit sense that x ∼ exp(θ) ⇐⇒ P (x > s + t|x > t) = P (x > s) > 0, ∀s, t > 0.

3.3 Dirichlet distributions If the gamma is intuitively a waiting time, the Dirichlet, otherwise known as the multivariate beta, is simply the proportion of time waited. Deﬁnition 3.5 Dirichlet: x ∼ Dn (p; pn+1 ) or x ∼ betan (p; pn+1 ), p = (p1 , . . . , pn ) , pi > 0, i = 1, . . . , n + 1 iﬀ d

x= with zi

1 z T

n+1 indep ∼ G(pi ), i = 1, . . . , n + 1, z = (z1 , . . . , zn ) , and T = i=1 zi .

The notation

indep ∼ means “independently distributed as.”

Proposition 3.1 The joint p.d.f. of x and T can be described as n+1 n+1 1 t i=1 pi −1 e−t , t > 0 i.e., T ∼ G , fT (t) = pi n+1 Γ( i=1 pi ) i=1

3.3. Dirichlet distributions

39

|=

pn+1 −1 n+1 n n Γ( i=1 pi ) pi −1 fx (x) = n+1 xi xi , x ∈ T n, 1− i=1 Γ(pi ) i=1 i=1 n where T n = {x ∈ Rn : xi > 0, T. i=1 xi < 1}. Moreover, x Proof. Using independence, the joint p.d.f. of the zi ’s is n+1 n+1 p −1 1 · zi i · exp − zi , zi > 0, ∀i. fz,zn+1 (z, zn+1 ) = n+1 i=1 Γ(pi ) i=1 i=1 We simply transform from (z1 , . . . , zn+1 ) to (x1 , . . . , xn , t), where zi = txi , i = 1, . . . , n and

zn+1 = t 1 −

n

xi

.

i=1

The jacobian is given by ∂z1 , . . . , zn+1 ∂x1 , . . . , xn , t +

∂tx1 , . . . , txn , t(1 − ni=1 xi ) = ∂x1 , . . . , xn , t + 0 x1 t .. .. . . = t x 0 nn −t · · · −t 1 − i=1 xi 0 x1 t .. .. . . = tn . = t xn 0 0 ··· 0 1

Thus, the joint p.d.f. of (x, T ) is n n n+1 1 xipi −1 (1 − xi )pn+1 −1 t i=1 pi −1 e−t , x ∈ T n , t > 0, n+1 i=1 Γ(pi ) i=1 i=1

and the conclusions are reached.

2

Note that the Dirichlet where all the parameters are 1 is simply the uniform distribution on the triangular region T n , Dn (1; 1) ≡ unif(T n ). Also, d

the Dirichlet distribution generalizes the beta distribution, D1 (p1 ; p2 ) = beta(p1 ; p2 ), with p.d.f. fx (x) = B(p1 , p2 )−1 xp1 −1 (1 − x)p2 −1 , 0 < x < 1, where B(p1 , p2 ) = Γ(p1 )Γ(p2 )/Γ(p1 + p2 ), is the beta function. The converse of Proposition 3.1 is almost obvious (by inverse change of variables); it need only be stated.

40

3. Gamma, Dirichlet, and F distributions

n i=1

xi )

|=

Proposition 3.2 If zi = T xi , i = 1, . . . , n and zn+1 = T (1 − n+1 T , then with T ∼ G( i=1 pi ), x ∼ Dn (p; pn+1 ), and x zi

indep ∼ G(pi ), i = 1, . . . , n + 1.

Four useful corollaries are also stated and the reader is asked to prove n+1 them. For x ∼ Dn (p; pn+1 ), let p = i=1 pi denote the “grand total,” noting that, by deﬁnition, d 1 x= z T n+1 indep with zi ∼ G(pi ), i = 1, . . . , n + 1, z = (z1 , . . . , zn ) , and T = i=1 zi . We ﬁnd the following: Corollary 3.2 (Marginal Dirichlet) If x1 = (xi1 , . . . , xik ) denotes any subset of the coordinates, then x1 ∼ Dk (p1 ; q) with p1 = (pi1 , . . . , pik ) and k p = q + j=1 pij . Corollary 3.3 If x = (x1 , . . . , xm ) is “partitioned” in any manner whatever so that we may write xi ∼ Dki (pi ; qi ), i = 1, . . . , m, deﬁne y by letting yi = xi 1, i.e., the total of the components of xi , with corresponding ri = pi 1. We ﬁnd y ∼ Dm (r; pn+1 ) with r = (r1 , . . . , rm ) . n Corollary 3.4 If S = x 1 = i=1 xi and again x1 = (xi1 , . . . , xik ) , d

k < n, is any subset, let w1 = S1 x1 . We ﬁnd w1 ∼ Dk (p1 ; r) with k p1 = (pi1 , . . . , pik ) as before but this time, p − pn+1 = r + j=1 pij . Corollary 3.5 (Conditional Dirichlet) If x = (x1 , x2 ) ∼ Dn ((p1 , p2 ) ; pn+1 ) , where x1 , p1 ∈ Rr and x2 , p2 ∈ Rs , n = r + s, then x1 | x2 ∼ Dr (p1 ; pn+1 ). 1 − x2 1

|=

We easily compute the moments of a Dirichlet distribution. By the converse representation in Proposition 3.2, if T ∼ G(p), x ∼ Dn (p; pn+1 ), and T , we have x d

T x = z with zi

indep ∼ G(pi ), i = 1, . . . , n.

This gives Tr

n

d

xri i =

i=1

n i=1

ziri with r =

n

ri ,

i=1

so that E Tr E

n i=1

xri i =

n i=1

E ziri for ri > −pi , i = 1, . . . , n.

3.3. Dirichlet distributions

41

6 4

1

2

0.8 0.6

0 0 0.2

0.4 0.4 0.2

0.6 0.8 10

Figure 3.1. Bivariate Dirichlet density for values of the parameters p1 = p2 = 1 and p3 = 2.

We ﬁnd E

n

n xri i =

i=1

E ziri . E Tr

i=1

In particular, E xi E x2i and E xi xj

= pi /p, ∀i, = (pi + 1)pi /(p + 1)p, ∀i, = pi pj /(p + 1)p, ∀i = j,

and letting θ = p1 p gives E x = θ and var x =

1 diag(θ) − θθ . p+1

Figure 3.1 exhibits a bivariate Dirichlet density. Various characterizations of Dirichlet distributions can be found in the literature [Rao and Sinha (1988), Gupta and Richards (1990)].

42

3. Gamma, Dirichlet, and F distributions

3.4 F distributions The ratio of two independent gammas is described by the F distribution, intuitively a relative waiting time. d y1 /s1 y2 /s2 ,

Deﬁnition 3.6 F distribution: F ∼ F (s1 , s2 ) iﬀ F = i = 1, 2.

yi

indep 2 ∼ χsi ,

One may easily obtain the moments of an F distribution and, in particular, its mean and variance. Sometimes, the distributions are more easily expressed in terms of the canonical Fc distribution: d

Deﬁnition 3.7 Canonical Fc distribution: F ∼ Fc (s1 , s2 ) iﬀ F = y1 /y2 , indep yi ∼ χ2si , i = 1, 2. One should also verify the simple relation F ∼ Fc (s1 , s2 ) ⇐⇒ (1 + F )−1 ∼ beta( 12 s2 ; 12 s1 ). The noncentral chi-square and F distributions useful to describe the nonnull distribution of some tests are deﬁned in Section 4.3.

3.5 Problems 1. If y ∼ χ2m , then E y h = 2h Γ

1

2m

+ h /Γ 12 m , h > − 12 m.

2. Prove Corollary 3.1. 3. Assume x ∼ unif([0, 1]) in Rn . (i) Deﬁne y by y1 = x(1) = min ({x1 , . . . , xn }) and

yi = x(i) = min {x1 , . . . , xn } − {x(1) , . . . , x(i−1) } , i = 2, . . . , n. Determine the distribution of y. (ii) Deﬁne z by z1 = y1 and zi = yi − yi−1 , i = 2, . . . , n, and determine the distribution of z. (iii) Determine E x, var x, E y, and var y as well as E z and var z. 4. Prove Corollaries 3.2, 3.3, 3.4, and 3.5. 5. Show the simple equivalence F ∼ Fc (s1 , s2 ) ⇐⇒ (1 + F )−1 ∼ beta( 12 s2 ; 12 s1 ). 6. Obtain the density of F ∼ Fc (s1 , s2 ):

Γ 12 (s1 + s2 ) F s1 /2−1 , F > 0. f (F ) = 1 1 Γ 2 s1 Γ 2 s2 (1 + F )(s1 +s2 )/2

4 Invariance

4.1 Introduction Invariance is a distributional property of a random vector acted upon by a group of transformations. The simplest group of transformations {+1, −1} leads to symmetric distributions by deﬁning a random variable to be symd metric iﬀ x = −x. Groups of transformations acting on random vectors commonly encountered are the permutations and orthogonal transformations. The permutation invariance gives the “exchangeable” random vectors and the invariance by orthogonal transformations deﬁnes the spherical distributions. Of great importance is the orthogonal group, since it speciﬁes the physical basis for normality in the Maxwell-Hershell theorem. Spherical distributions will play a central role later in Chapter 13 to build the elliptical models useful in the study of robustness.

4.2 Reﬂection symmetry d

Deﬁnition 4.1 x is (reﬂection) symmetric iﬀ x = −x. d

One immediately notes that if x = −x and E |x| < ∞, then E x = 0 (why?). The distribution of a symmetric random variable x is completely determined by the distribution of its modulus |x|, as the next proposition shows.

4. Invariance d

d

Proposition 4.1 x = −x ⇐⇒ x = s|x| with s

|=

44

|x|, s ∼ unif{±1}.

Proof. (=⇒) : Let F be the d.f. of x. Then, P (s|x| ≤ t)

1 2

{P (|x| ≤ t) + P (|x| ≥ −t)} P (−t ≤ x ≤ t) + 1, t ≥ 0 1 = 2· P (|x| ≥ −t), t 2, s1 (s2 − 2) s22 (s1 + 2δ)2 + (s1 + 4δ)(s2 − 2) , s2 > 4. 2 2 s1 (s2 − 2)2 (s2 − 4)

3. Obtain the density of F ∼ Fc (s1 , s2 ; δ) using Problem 3.5.6:

∞ k Γ 12 (s1 + s2 + 2k) F (s1 +2k)/2−1 −δ δ

1 1 f (F ) = e , k! Γ 2 (s1 + 2k) Γ 2 s2 (1 + F )(s1 +s2 +2k)/2 k=0 F > 0. 4. Assume x = (x1 , x2 ) has a spherical distribution. Show that x1 also has a spherical distribution. 5. Let x ∈ Rn have a spherical distribution with a ﬁnite rth moment. Demonstrate that all product-moments of x, E(xs11 · · · xsnn ), of order n s = i=1 si ≤ r are null provided one of the si is odd. 6. Let x = (x1 , . . . , xn ) have a spherical distribution. Prove the following: (i) cx (t) = cx1 (|t|) is a function of |t|. (ii) If x is absolutely continuous, x ∼ f , then f (x) = f (|x|e1 ) depends on x only through |x|.

4.6. Problems

53

7. Assume x ∈ Rn is rotationally invariant. Prove the mixture characterization ∞ cu1 (|t|r)dF (r), cx (t) = 0

where u = (u1 , . . . , un ) ∼ unif(S n−1 ) and F is the distribution function of |x| on [0, ∞). This means any rotationally invariant distribution is a mixture of uniform distributions on spheres of varying radius r ≥ 0 [Schoenberg (1938)]. 8. Let x ∼ unif(B n ), where B n = {s : s s ≤ 1} is the “unit ball” in Rn . (i) Deﬁne y by yi = x2i , i = 1, . . . , n, and determine the distribution of y, the marginal distribution of each yi , i = 1, . . . , n, and, ﬁnally, the distribution of R2 = |x|2 = x x. (ii) Obtain vol(B n ) using (i) and indicate the special cases n = 1, 2, 3. (iii) Determine E x and var x as well as E R2 and var R2 . Hint: Realize that y is “concentrated” on T n = {y : yi ≥ 0,

n

yi ≤ 1}.

i=1

9. Assume x ∈ Rn is permutationally invariant ∀n and E |x|2 < ∞. Let S = g(x), where g : Rn → R is any (permutation) symmetric function, i.e., g(Jt) = g(t), ∀J ∈ Sn , ∀t ∈ Rn . (i) Prove ρ ≥ 0. w.p.1 = E (f (Jx) | S) , ∀J ∈ Sn . w.p.1 ¯) ≤ 0. (iii) cov(x1 , x2 | x (ii) E (f (x) | S)

10. Assume x ∈ Rn has a “rotationally invariant” distribution such that d P (x = 0) = 0, i.e., Hx = x, ∀H ∈ On . Let R = |x| and z = x/R. |=

(i) Prove that z has the same distribution as if x had been unif(B n ). z. (ii) Prove that R (iii) Determine E z and var z. (iv) Determine E x and var x in terms of E R2 . Partition x = (x1 , x2 ) , x1 ∈ Rk and x2 ∈ Rn−k and let Ri = |xi |, i = 1, 2. (v) Determine the distribution of R12 /R2 and R12 /R22 . 11. Assume u ∼ unif(S n−1 ) and u = (u1 , u2 ) , u1 ∈ Rk . (i) Prove that the density of u1 is f (u1 ) =

Γ( 12 n) (1 − u1 u1 )(n−k)/2−1 , 0 < u1 u1 < 1. π k/2 Γ[ 12 (n − k)]

54

4. Invariance

Hint: Show (u21 , . . . , u2k ) ∼ Dk ( 12 1; 12 (n − k)) and consider the one-to-many transformation u2i → ±ui . (ii) Prove |u1 |2 ∼ beta( 12 k; 12 (n − k)). 12. Assume u = (u1 , u2 , u3 ) ∼ unif(S 2 ). Show that u1 ∼ unif(−1, 1). Does this hold in other dimensions?

|=

13. Let x = (x1 , . . . , xn ) have a spherical density fx (x) = g(|x|2 ) for some function g : [0, ∞) → [0, ∞). Let x = ru, where r ≥ 0 denotes “radius” and u ∈ S n−1 represents “direction.” Prove the following using J(x → r, u) = rn−1 : u. (i) r (ii) r2 has density fr2 (s) = 12 ωn sn/2−1 g(s), s > 0, where ωn is the “area” of the unit sphere S n−1 . (iii) With the special case x1 , . . . , xn i.i.d. N (0, 1), ﬁnd the “area” ωn . (iv) What is the density of u? 14. Let x ∈ Rn have a spherical density fx (x) = g(|x|2 ) and x → r, θ1 , . . . , θn−1 be the transformation to polar coordinates as in Proposition 2.23. Prove θn−1 ∼ unif(0, 2π). What can be said about the other angles? 15. Prove the following concerning spherical distributions: (i) If g(|x|2 ) is a density on Rn for some g : [0, ∞) → [0, ∞), then ∞ n−1 r g(r2 )dr = Γ 12 n /(2π n/2 ). 0 (ii) If the kth moment of x is ﬁnite, i.e., E |x|k < ∞, then ∞ rn+k−1 g(r2 )dr < ∞. 0

(iii) If the second moment of x is ﬁnite, then var x = αI, where α = E x21 . From Problem 4.6.6, cx (t) = φ(t t) for some function φ. Prove α = −2φ (0).

5 Multivariate normal

5.1 Introduction This chapter is entirely devoted to the multivariate normal distribution. In Section 5.2, the basic properties are demonstrated. Then, Sections 5.3 and 5.4 make the distinction between the nonsingular and the singular cases. In the nonsingular case, the density is derived while we explain the geometry of the singular case. Section 5.5 contains the conditional distribution in all its generality. Finally, the last section reaps the ﬁrst beneﬁts by considering some applications in univariate sampling, regression, and elementary correlation.

5.2 Deﬁnition and elementary properties Let Σ = (σij ) ∈ Rnn be symmetric, positive semideﬁnite, and µ ∈ Rn . Deﬁnition 5.1 Multivariate normal: x ∼ Nn (µ, Σ) iﬀ t x ∼ N (t µ, t Σt), ∀t ∈ Rn . Note that x has product-moments of any order by the fact that this is true of t x, ∀t ∈ Rn . Proposition 5.1 x ∼ Nn (µ, Σ) =⇒ E x = µ and var x = Σ.

56

5. Multivariate normal

Proof. Setting t = ei = (0, . . . , 1, . . . , 0) , we ﬁnd the individual component xi = ei x: xi ∼ N (µi , σii ), where µi = E xi , σii = var xi . Similarly, setting t to be a vector with 1’s in the ith and jth components, i = j, and 0’s elsewhere, we ﬁnd xi + xj ∼ N (µi + µj , σii + σjj + 2σij ), i = j. However, since on the other hand for i = j, var (xi +xj ) = var xi +var xj + 2 2cov(xi , xj ), then cov(xi , xj ) = σij . Proposition 5.2 Let A : Rn → Rm , linear, and x ∼ Nn (µ, Σ). Then, Ax ∼ Nm (Aµ, AΣA ). Proof. Let y = Ax and merely note that s y = (A s) x ∼ N (s Aµ, s AΣA s) , ∀s ∈ Rm . 2 By specializing A to be the projection onto any particular subset of coordinates, we deduce immediately that all the marginal distributions are normal. As a simple corollary on rotational invariance, we have z ∼ Nn (0, σ 2 I) =⇒ Hz = z, ∀H ∈ On . d

The characteristic function for x ∼ Nn (µ, Σ) derives from the univariate level (4.1):

cx (t) = ct x (1) = exp − 12 t Σt + it µ . Example 5.1 Although all marginals of x have a univariate normal distribution, the vector x itself may not have a multivariate normal distribution. Consider a random vector x whose distribution is a mixture of two multivariate normal distributions, cx (t) = αcx1 (t) + (1 − α)cx2 (t), 0 < α < 1, where x1 x2

∼ Nn (0, (1 − ρ1 )I + ρ1 11 ) , ∼ Nn (0, (1 − ρ2 )I + ρ2 11 ) .

Then, cxi (ti ) = αcz (ti ) + (1 − α)cz (ti ) = cz (ti ), where z ∼ N (0, 1), d

which shows xi = z, i = 1, . . . , n, but x does not have a multivariate normal distribution. Other counterexamples can be given using copulas [v. Example 2.5].

5.2. Deﬁnition and elementary properties

57

As a special case, we have the characteristic function for z ∼ Nn (0, σ 2 I): n n

exp(− 12 t2i σ 2 ) = czi (ti ) cz (t) = exp − 12 σ 2 t t = i=1

i=1

by which An implication is that if z = (z1 , z2 ) ∼ Nn (0, I), then z1 density for z becomes

|=

z ∼ Nn (0, σ 2 I) ⇐⇒ z1 , . . . , zn i.i.d. N (0, σ 2 ). z2 , and the

n

(2π)−1/2 exp − 12 zi2 . fz (z) = (2π)−n/2 exp − 12 z z = i=1

It is also clear from the characteristic function that the family of multivariate normal is closed under translation: x ∼ Nn (µ, Σ) =⇒ x + b ∼ Nn (µ + b, Σ), ∀b ∈ Rn . Now, suppose that x ∼ Nn (µ, Σ) and write Σ = HDH with H orthogonal and D = diag(λ). We ﬁnd, of course, that y = H (x − µ) ∼ Nn (0, D), and if we then let A = HD1/2 , we deduce the representation: d

Proposition 5.3 x ∼ Nn (µ, Σ) ⇐⇒ x = Az + µ for any A such that AA = Σ, z ∼ Nn (0, I). Finally, partition x = (x1 , x2 ) , where x1 ∈ Rn1 and x2 ∈ Rn2 , n = n1 +n2 , with corresponding µ1 Σ11 Σ12 µ= and Σ = . µ2 Σ21 Σ22 Concerning independence, we have the following necessary and suﬃcient condition.

x1

|=

Proposition 5.4 Let x = (x1 , x2 ) ∼ Nn (µ, Σ). Then, x2 ⇐⇒ Σ12 = 0.

x1

|=

Proof. (=⇒): x2

=⇒ E g1 (x1 )g2 (x2 ) = E g1 (x1 )E g2 (x2 ), ∀g1 , g2 =⇒ Σ12 = 0.

(⇐=): Assume Σ12 = 0. Write Σii = Aii Aii , i = 1, 2. Then, Σ = AA , where A11 0 A= . 0 A22 Using the representation A11 z1 + µ1 x1 d = Az + µ = , x2 A22 z2 + µ2

where z ∼ Nn (0, I), it is clear that since z1

z2 , then x1

|=

5. Multivariate normal

|=

58

x2 .

2

Another simple proof based on characteristic functions is proposed in Problem 5.7.3.

5.3 Nonsingular normal n When x ∼ Nn (µ, Σ) and |Σ| = |AA | = |A|2 = |D| = i=1 λi > 0, deﬁne z = A−1 (x−µ) whereby we have an explicit density for x by simple change of variables:

fx (x) = fz A−1 (x − µ) · J(z → x) = (2π)−n/2 |Σ|−1/2 exp − 12 (x − µ) Σ−1 (x − µ) and, of course, from Proposition 4.4, then also (x − µ) Σ−1 (x − µ) = z z ∼ χ2n . The quantity [(x − µ) Σ−1 (x − µ)]1/2 is often called the Mahalanobis distance of x to µ. Example 5.2 The bivariate density function is just a special case. For µ1 σ12 ρσ1 σ2 µ= , Σ= , µ2 ρσ1 σ2 σ22 we ﬁnd −1

Σ

1 = (1 − ρ2 )

σ1−2 −ρ/σ1 σ2

−ρ/σ1 σ2 σ2−2

.

Thus, the bivariate density takes the form fx (x1 , x2 )

=

2 x1 − µ1 1 1 1 exp − 2π σ1 σ2 (1 − ρ2 )1/2 2(1 − ρ2 ) σ1 2 x1 − µ1 x2 − µ2 x2 − µ2 −2ρ . + σ1 σ2 σ2

A plot of this density is given in Figure 5.1. The contours, which consists of the set of points of equal probability density, of a multivariate normal are the points x of equal Mahalanobis distance to µ, (x − µ) Σ−1 (x − µ) = c2 , for any constant c > 0. Letting y = H x, ν = H µ, where H diagonalizes Σ, H ΣH = D, then the contours are the ellipsoids p i=1

(yi − νi )2 /di = c2

5.3. Nonsingular normal

0.2 0.15 0.1 0.05 0

59

2 1 0

-2 -1

-1

0 1

-2 2

Figure 5.1. Bivariate normal density for values of the parameters µ1 = µ2 = 0, σ1 = σ2 = 1, and ρ = 0.7. 1/2

centered at ν with principal axes of half length cdi eigenvectors in H = (h1 , . . . , hp ).

supported by the

Example 5.3 The contours of the bivariate normal density are in parametric form, and in the y coordinates, 1/2 d1 sin θ y1 ν1 , 0 ≤ θ ≤ 2π. = +c 1/2 y2 ν2 d2 cos θ Thus, the contours in the original x coordinates are just 1/2 1/2 h11 d1 sin θ + h12 d2 cos θ x1 µ1 , 0 ≤ θ ≤ 2π. = +c 1/2 1/2 x2 µ2 h21 d1 sin θ + h22 d2 cos θ A contour plot is given in Figure 5.2. Example 5.4 Using the transformation to polar coordinates on p. 32, the contours of the trivariate normal density are in parametric form, and in the y coordinates, 1/2 d1 sin θ1 sin θ2 y1 ν1 1/2 y2 = ν2 + c d2 sin θ1 cos θ2 , 0 ≤ θ1 ≤ π, 0 ≤ θ2 ≤ 2π. 1/2 y3 ν3 d cos θ1 3

60

5. Multivariate normal

x2 3 2 1 -3

-2

-1

1

2

3

x1

-1 -2 -3 Figure 5.2. Contours of the bivariate normal density for values of the parameters µ1 = µ2 = 0, σ1 = σ2 = 1, and ρ = 0.7. Values of c = 1, 2, 3 were taken.

Thus, the contours in the original x1 µ1 h11 x2 = µ2 + c h21 x3 µ3 h31

x coordinates are just 1/2 d1 sin θ1 sin θ2 h12 h13 h22 h23 d21/2 sin θ1 cos θ2 , 1/2 h32 h33 d3 cos θ1

0 ≤ θ1 ≤ π, 0 ≤ θ2 ≤ 2π. The contour plot corresponding to c = 1 is given in Figure 5.3 when µ = 0 and 13 −4 2 Σ = −4 13 −2 . 2 −2 10 The corresponding eigenvalues of d1 = 18, d2 = d3 = 9 give the typical ellipsoidal contours. Still assuming |Σ| > 0, we apply the Gram-Schmidt process to the basis formed by the row vectors of A = HD1/2 , obtaining (uniquely) A = TG (v. Proposition 1.13) with T ∈ L+ n , G ∈ On , where L+ n = {T ∈ Gn : T is lower triangular, tii > 0, i = 1, . . . , n}. Then, Σ = AA = TT for a unique T ∈ L+ n (v. Proposition 1.14). We have the “triangular” representation: Proposition 5.5 (Triangular representation) d

x ∼ Nn (µ, Σ) ⇐⇒ x = Tz + µ with T ∈ L+ n such that Σ = TT and z ∼ Nn (0, I).

5.3. Nonsingular normal

2 0 -2

2

0

-2

-2 0 2 Figure 5.3. A contour of a trivariate normal density.

61

62

5. Multivariate normal

5.4 Singular normal Now, for x ∼ N (µ, σ 2 ), we know that σ 2 = 0 ⇐⇒ x = µ w.p.1. This holds, since if σ 2 = 0, P (x = µ) = limn→∞ P (|x − µ| < 1/n), but P (|x − µ| ≥ 1/n) ≤ σ 2 n2 = 0, ∀n. Thus, the normal family includes the “trivial” (constant) random variables as special cases. By Cram´er-Wold Proposition 2.10, this also holds for random vectors x ∈ Rn with E x = µ and var x = Σ: Σ = 0 ⇐⇒ x = µ w.p.1. However, if x ∼ Nn (µ, Σ) with |Σ| = 0, we may write D1 0 H1 Σ = HDH = (H1 , H2 ) = H1 D1 H1 , H2 0 0 where D1 = diag(λ1 , . . . , λr ) comprises the nonzero eigenvalues, H1 = (h1 , . . . , hr ) gives a basis for the column space of Σ, Im Σ, and H2 = (hr+1 , . . . , hn ) gives a basis for the kernel, ker Σ. One should note that Im H2 Im H1

=

(Im H1 )⊥ = ker Σ,

=

(Im H2 )⊥ = Im Σ.

H2 (x − µ)

|=

Then it is clear that H2 ΣH2 = 0 and, thus, we ﬁnd that H2 (x − µ) = 0 w.p.1 or, equivalently, x − µ ∈ (Im H2 )⊥ w.p.1., whereas H1 (x − µ) ∼ Nr (0, D1 ) has a nonsingular normal distribution. Of course, this is yet equivalent to saying that x ∈ µ + Im Σ w.p.1 and one can then almost visualize x in this r-dimensional aﬃne subspace of Rn , r < n. A curious fact in this case is that vol(µ + Im Σ) = 0 but Px (µ + Im Σ) = 1, therefore x cannot be absolutely continuous (v. Proposition 2.11). It is worth recalling at this point that any constant random vector is automatically statistically independent of any other random vector, and so we might notice, in particular, the rather odd looking fact that x.

5.5 Conditional normal By a suitable permutation, one may rearrange an arbitrary multivariate normal x so that any subset x1 of its coordinates are brought to the fore, and the overall distribution is expressed by µ1 Σ11 Σ12 x1 ∼ Nn , , x1 ∈ Rn1 , x2 ∈ Rn2 , n = n1 + n2 . x2 µ2 Σ21 Σ22 We derive the conditional distribution of x1 given x2 . First suppose Σ22 is nonsingular and note that for any B, I 0 I −B Σ11 Σ12 −B I Σ21 Σ22 0 I

5.5. Conditional normal

=

Σ11 − Σ12 B − BΣ21 + BΣ22 B Σ21 − Σ22 B

Σ12 − BΣ22 Σ22

so by deliberately setting B = Σ12 Σ−1 22 , we ﬁnd µ1 − Bµ2 Σ11.2 x1 − Bx2 ∼ Nn , x2 µ2 0

0 Σ22

63

,

,

where Σ11.2 = Σ11 − Σ12 Σ−1 22 Σ21 . d

However, independence means x1 − Bx2 | x2 = x1 − Bx2 , so that we have x1 − Bx2 | x2 ∼ Nn1 (µ1 − Bµ2 , Σ11.2 ). Since we may legitimately treat x2 as though constant (the full justiﬁcation of this depending on the fact that we have a “regular” conditional distribution to which the Fubini theorem applies [Ash (1972)]) we may conclude that x1 | x2 ∼ Nn1 (µ1 + Σ12 Σ−1 22 (x2 − µ2 ), Σ11.2 ). In the singular case, if |Σ22 | = 0, we may always write D 0 H1 = H1 DH1 , Σ22 = (H1 , H2 ) 0 0 H2 where D ∈ Rkk is nonsingular, and we may then take what is called a “pseudo-inverse” for Σ22 : −1 D 0 H1 = H1 D−1 H1 . = (H , H ) Σ− 1 2 22 H2 0 0 w.p.1 We then have, of course, H2 (x2 − µ2 ) = 0 and also µ1 Σ11 x1 Σ12 H1 ∼ N , n1 +k H1 x2 H1 µ2 H1 Σ21 D to which the results in the nonsingular case apply, immediately showing µ1 − Bµ2 Σ11.2 0 x1 − Bx2 ∼ N , , n1 +k H1 x2 H1 µ2 0 D |=

|=

− where B = Σ12 Σ− 22 and Σ11.2 = Σ11 − Σ12 Σ22 Σ21 . But from this, H1 x2 , and thus, overall, x1 − Bx2 x2 . We arrive at the x1 − Bx2 completely general conclusion:

Proposition 5.6 x1 | x2 ∼ Nn1 (µ1.2 , Σ11.2 ), where µ1.2 Σ11.2

= µ1 + Σ12 Σ− 22 (x2 − µ2 ), = Σ11 − Σ12 Σ− 22 Σ21 .

64

5. Multivariate normal

5.6 Elementary applications 5.6.1

Sampling the univariate normal

Observe that x1 , . . . , xn i.i.d. N (µ, σ 2 ) ⇐⇒ x ∼ Nn (µ1, σ 2 I). √ Letting H = ( 1/ n, Γ ) ∈ On for some Γ and √ w = H x ∼ Nn ( nµe1 , σ 2 I), obviously w1 , . . . , wn are independent, and, of course, √ x w1 = n¯ and

d

with its trivial algebraic corollary √ d n(¯ x − µ)/sx = (n − 1)z/χn−1 , z ∼ N (0, 1), and z

s2x

|=

¯ x ¯ ∼ N (µ, σ 2 /n), (n − 1)s2x = σ 2 χ2n−1 , and x

|=

w22 + · · · + wn2 = |w|2 − w12 = |x|2 − n¯ x2 = |x − x ¯1|2 = (n − 1)s2x , n ¯)2 /(n − 1) is the sample variance. Thus, we have where s2x = i=1 (xi − x the basic statistical result

χ2n−1 .

|=

We make the following deﬁnition (W.S. Gosset, “Student,” 1908): d d √ Deﬁnition 5.2 t-Distribution: t = tp iﬀ t = pz/χp , where z ∼ N (0, 1) and z χ2p . Thus, by deﬁnition, d

√

d

n(¯ x−µ)/sx = tn−1 is a pivotal quantity for µ. Clearly,

d

t = tp ⇐⇒ t = −t and t2 ∼ F (1, p). This provides a quick way of obtaining the integral moments of tp . The Student’s t-distribution sometimes plays a role in the dependent case. The intraclass correlation model is one such example.

Example 5.5 Assume x ∼ Nn µ1, σ 2 [(1− ρ)I + ρ11 ] , where −1/(n − n n 1) ≤ ρ ≤ 1. Let x ¯ = i=1 xi /n, s2x = i=1 (xi − x ¯)2 /(n − 1), and t = √ n(¯ x − µ)/sx . We determine a constant c such that ct ∼ tn−1 . With the orthogonal transformation above, we still have w w1 w22 + · · · + wn2

= H x, √ = n¯ x, =

(n − 1)s2x .

√ Since H 1 = ( n, 0 ) , the distribution of w is √ nµ w ∼ Nn , σ 2 [(1 − ρ)I + ρ diag(n, 0, . . . , 0)] . 0

5.6. Elementary applications

65

|=

|=

Hence, w1 (w2 , . . . , wn ) , which implies x ¯ s2x . The distribution of x ¯ 2 and sx are given by √ √ n¯ x ∼ N ( nµ, σ 2 [(1 − ρ) + ρn]), (n − 1)s2x ∼ σ 2 (1 − ρ)χ2n−1 .

Finally, we can conclude that ct ∼ tn−1 by deﬁning 1/2 1−ρ c= . (1 − ρ) + ρn In fact, the Student’s t-distribution has nothing to do with normal distributions. It is more related to the concept of spherical symmetry, as in the next example [Efron (1969)].

d

the representation x = Ru, where u ∼ unif(S n−1 ) and R

|=

invariant” distribution Example 5.6 Assume x ∈ Rn has a “rotationally √ n¯ x /s ∼ tn−1 , where, as usual, and P (x = 0) = 0. We establish that x n n ¯)2 /(n − 1). Using Proposition 4.10, x ¯ = i=1 xi /n and s2x = i=1 (xi − x u, is valid.

d

Hence, (¯ x, sx ) = (R¯ u, Rsu ) and the distribution of √ x √ u ¯ d √ R¯ u ¯ n = n = n sx Rsu su √ does not depend on R. Thus, n¯ x/sx ∼ tn−1 since this is the case when x ∼ Nn (0, I).

5.6.2

Linear estimation

Consider now the problem of linear estimation in the so-called multiple regression model. Let V ⊂ Rn be any k-dimensional vector subspace and y = µ + e, E e = 0, var e = σ 2 I, and µ ∈ V. ˆ ˆ where µ ˆ= Let θ = Tµ, where T ∈ Rm n , and consider the estimate θ = Tµ, Py is the orthogonal projection of y on V (v. Section 1.6). We prove that among all possible unbiased linear estimates of θ, the regression estimate ˆ has the minimum variance. In this sense, θ ˆ is the “best” linear unbiased θ estimate (blue). ˆ = blue(θ). Proposition 5.7 (Gauss-Markov) θ ˜ = By is unbiased for θ ⇐⇒ BP = TP. But then, Proof. θ ˆ = σ 2 TPT = σ 2 BPB ≤ σ 2 BB = var θ ˜ var θ with equality iﬀ ˜ = θ, ˆ BQB = 0 ⇐⇒ BQ = 0 ⇐⇒ B = BP ⇐⇒ θ where Q = I − P.

2

66

5. Multivariate normal

ˆ = Py = blue(µ) with var µ ˆ = σ 2 P. For example, µ Now, expressing µ = Xβ with respect to any basis X = (x1 , . . . , xk ) for V and recalling the representation (1.3) for P, the coeﬃcients are uniquely ˆ = B0 µ ˆ = determined as β = B0 µ, where B0 = (X X)−1 X . But then, β 2 −1 ˆ ˆ B0 y = blue(β), where var β = σ (X X) . Obviously, β i = blue(β i ), i = 1, . . . , k. Another optimality property of the “Gauss-Markov” estimate was recently discovered [Berk and Hwang (1989), Eaton (1988), Ali and Ponnapalli (1990]: The probability of the Gauss-Markov estimate of θ falling inside any ﬁxed ellipsoid centered at θ is greater than or equal to the probability that any linear unbiased estimate of θ falls inside the same elˆ = Py lipsoid. It is interesting to remark that the Gauss-Markov estimate µ is also the least-squares estimate. This follows from a general property of orthogonal projections: Proposition 5.8 ˆ 2, min |y − µ|2 = |y − µ| µ∈V ˆ = Py is the orthogonal projection of y on V. where µ Proof. For all µ ∈ V, |y − µ|2

ˆ + (µ ˆ − µ)|2 = |(y − µ) ˆ 2 + |µ ˆ − µ|2 + 2(y − µ) ˆ (µ ˆ − µ) = |y − µ| 2 2 ˆ − µ| ˆ + |µ = |y − µ|

ˆ ∈ V ⊥ and µ ˆ − µ ∈ V. Hence, since y − µ ˆ 2 , ∀µ ∈ V, |y − µ|2 ≥ |y − µ| ˆ with equality if µ = µ.

2

Finally, since Q = I − P gives the orthogonal projection on V ⊥ , we ﬁnd ˆ = Qy = Qe =⇒ |y − µ| ˆ 2 = e Qe, y−µ so that ˆ 2 = E e Qe = E tr Qee = tr QE ee = (n − k)σ 2 . E |y − µ| Thus, we determine the unbiased estimate sˆ2 of σ 2 by ˆ 2. (n − k)ˆ s2 = |y − µ| ˆ y − µ) ˆ = cov(Py, Qy) = σ 2 PQ = 0. Before It is also clear that cov(µ, ˆ and sˆ2 , stating the joint distribution under normality of our estimates µ we prove the following lemma on quadratic forms. Lemma 5.1 Let z ∼ Nn (µ, I) and Q ∈ Rnn be an orthogonal projection of rank Q = m. Then, z Qz ∼ χ2m (δ), where δ = µ Qµ/2.

5.6. Elementary applications

67

Proof. Let H = (h1 , . . . , hm ) be an orthonormal basis for Im Q and write Q = HH , where H H = Im . Then, z Qz = (H z) (H z) = |e|2 , where 2 e = H z ∼ Nm (H µ, I). Hence, |e|2 ∼ χ2m (δ) with δ = |H µ|2 /2. If, in addition, we assume normality, y = Xβ + e, e ∼ Nn (0, σ 2 I),

we have the general result

ˆ ∼ Nk β, σ 2 (X X)−1 , (n − k)ˆ ˆ β s2 ∼ σ 2 χ2n−k , and β

ˆ y − µ,

|=

ˆ = B0 y = B0 µ, ˆ (n − k)ˆ ˆ 2 = e Qe, and µ ˆ β s2 = |y − µ|

|=

then since

s

with corollary ˆ 2 |X(β − β)| ∼ F (k, n − k). 2 kˆ s We close this section with a slight generalization of Lemma 5.1. Corollary 5.1 Assume x ∼ Nn (µ, Σ), Σ > 0, and A is symmetric such that AΣA = A and rank ΣA = m. Then, x Ax ∼ χ2m (δ), where δ = µ Aµ/2. Proof. Letting z = Σ−1/2 x and B = Σ1/2 AΣ1/2 , then x Ax = z Bz, where z ∼ Nn (Σ−1/2 µ, I), and the conclusion follows from Lemma 5.1 since B is an orthogonal projection of rank m. 2

5.6.3

Simple correlation

Let (xi , yi ) i.i.d. (x, y), i = 1, . . . , n, be any “bivariate” sample. The correlation coeﬃcient ρ

= =

cor(x, y) cov(x, y) var(x) var(y)

is usually estimated by the sample correlation coeﬃcient n ¯)(yi − y¯) i=1 (xi − x r = n 1/2 n 1/2 2 [ i=1 (xi − x ¯) ] [ i=1 (yi − y¯)2 ] (x − x ¯1) (y − y¯1) . = |x − x ¯1| |y − y¯1| Note that r is just the cosine of the angle between the residual vectors x−x ¯1 and y − y¯1. The main (nonparametric) reason for using r as an w.p.1 estimate of ρ is its (strong) consistency: r → ρ as n → ∞.

68

5. Multivariate normal

Now, suppose that ρ = 0 so that x r=

|=

If, in addition, we assume normality, a “pivotal statistic” may be derived. First, notice that since r is invariant with respect to relocation and rescaling in both x and y, we may suppose at the outset that 0 1 ρ x ∼ N2 , . 0 ρ 1 y y. Then,

(Qx) (Qy) (x − x ¯1) (y − y¯1) , = |x − x ¯1| |y − y¯1| |Qx||Qy|

where Q = In − n−1 11 is an orthogonal projection of rank Q = n − 1. We can write (v. Section 1.6) Q = HH with H H = In−1 . Then, r=

z w (H x) (H y) , = |H x||H y| |z||w|

where z = H x and w = H y are independent Nn−1 (0, I). Finally, letting |=

d

u = z/|z| and v = w/|w|, we have u v, and from Corollary 4.3, u = v ∼ unif(S n−2 ). Therefore, using Proposition 4.8, r = u v = u1 = d

z1 . |z|

Thus, √

z1 r d = 1/2 , 2 2 1−r z22 + · · · + zn−1

where the zi ’s are i.i.d. N (0, 1) and √ r d n − 2√ = tn−2 . 1 − r2 We have proved: Proposition 5.9 If (xi , yi ) , i = 1, . . . , n, are i.i.d. as a bivariate normal with ρ = 0, then √ r d n − 2√ = tn−2 . 1 − r2 However, if ρ = 0 and 2 0 σ x , ∼ N2 0 ρστ y

ρστ τ2

,

then we may apply this result to the linear transformation 0 1 − ρ2 0 x/σ − ρy/τ ∼ N2 , 0 0 1 y/τ

5.7. Problems

using

69

n

r˜ =

((xi − x ¯)/σ − ρ(yi − y¯)/τ ) (yi − y¯) . !1/2 1/2 2 n 2 ¯)/σ − ρ(yi − y¯)/τ ) [ i=1 (yi − y¯) ] i=1 ((xi − x i=1

n

We ﬁnd √

r˜ z1 d = 2 , 2 2 (z2 + · · · + zn−1 )1/2 1 − r˜

where the zi ’s are i.i.d. N (0, 1) and, by direct computation, √

r − ρc sy /τ r˜ . =√ , where c = 2 2 s 1 − r˜ 1−r x /σ

Thus, we obtain the result √

(r − ρc) d n − 2√ = tn−2 . 1 − r2

This is actually a pivotal for β = ρσ/τ . Later, the reader will be able to prove that √ d n(r − ρ) → (1 − ρ2 )z, z ∼ N (0, 1) (v. Problem 6.4.8), which can be used to obtain an approximate conﬁdence interval for ρ. The exact distribution of r is treated in Section 8.4 in the more general context of multiple correlation coeﬃcient.

5.7 Problems 1. Plot the contours of the N2 (µ, Σ) distribution when 1 µ = , 2 2 1 Σ = . 1 4 |=

y and consider 2. Let x ∼ Nn (µ1, σ 2 I), y ∼ Nn (ν1, τ 2 I), and x n ¯)(yi − y¯) i=1 (xi − x . r = n 1/2 n 1/2 2 [ i=1 (xi − x ¯) ] [ i=1 (yi − y¯)2 ] (i) Determine the distribution of r. (ii) Determine E r and var r. 3. Prove Proposition 5.4 with characteristic functions. 4. Obtain the integral moments of the tp distribution.

70

5. Multivariate normal

5. Let x be such that E x = µ and var x = Σ. Show that min E |x − c|2 = tr Σ c

and that the minimum is attained at c = µ. 6. Assume

x1 x2

∼ Nn

µ1 µ2

Σ11 , Σ21

Σ12 Σ22

.

Demonstrate that min E |x1 − (Cx2 + d)|2 = tr Σ11.2 C,d

− is attained at C = Σ12 Σ− 22 and d = µ1 − Σ12 Σ22 µ2 . n 2 7. Assume that z ∼ Nn (0, I) and let z¯ = i=1 zi /n, (n − 1)s = n 2 ¯) . i=1 (zi − z

(i) Prove z¯, s, (z1 − z¯)/s are mutually independent. (ii) Determine the distribution of (z1 − z¯)/s. √ Hint: Let H = (1/ n, (e1 − n−1 1)/ (n − 1)/n, Γ) ∈ On , for some matrix Γ, w = H z, and note that (w2 , . . . , wn ) is rotationally invariant. 8. Assume y = Xβ + e, e ∼ Nn (0, σ 2 I), where, as usual, the columns of X ∈ Rnk are linearly independent and let C ∈ Rrk be of rank r. Show that ˆ − d) [C(X X)−1 C ]−1 (Cβ ˆ − d) (Cβ ∼ F (r, n − k; δ), rˆ s2 where δ=

(Cβ − d) [C(X X)−1 C ]−1 (Cβ − d) . 2σ 2

9. Let x ∼ Nn (µ1, σ 2 I). (i) Assume y is ﬁxed, y ∈ span{1}. Find the distribution of r=

(x − x ¯1) (y − y¯1) . |x − x ¯1| |y − y¯1|

(ii) This time assume y has any distribution satisfying

and x

|=

P (y ∈ span{1}) = 1 y, and determine the distribution of r.

10. Angular gaussian distribution. The angular gaussian distribution is obtained by the projection of x ∼ Nn (0, Λ) onto the unit sphere S n−1 ; i.e., the angular gaussian density is that of u = x/|x|.

5.7. Problems

71

(i) Prove that the angular gaussian density is f (u) =

Γ( 12 n) −1/2 −1 −n/2 |Λ| (u Λ u) , u ∈ S n−1 . 2π n/2

(ii) What is the special case Λ = I? (iii) Prove that the angular gaussian distribution can also be obtained by projecting (onto S n−1 ) x with density fx (x) = |Λ|−1/2 g(x Λ−1 x). The word gaussian is misleading here; symmetry is the key. 11. Rotationally symmetric distributions on spheres [Saw (1978)]. This class of distributions will be those for which the density is constant on those points u ∈ S n−1 satisfying u θ = δ, ∀δ ∈ [−1, 1] and some ﬁxed θ ∈ S n−1 . (i) For some ﬁxed λ ≥ 0, consider the function g(λ, ·) : [−1, 1] → [0, ∞). Prove 1 (1 − t2 )(n−3)/2 ωn−1 g(λ, u θ)du = g(λ, t) 1 1 dt, B( 2 , 2 (n − 1)) −1 S n−1 where B(·, ·) denotes the beta function and ωn = 2π n/2 /Γ( 12 n) is the “area” of S n−1 . Hint: ωn−1 g(λ, u θ)du = E g(λ, u θ) = E g(λ, u1 ), S n−1

where u = (u1 , . . . , un ) ∼ unif(S n−1 ) and use Problem 4.6.11. (ii) Deduce that f (u) = ωn−1 g(λ, u θ) is a density on S n−1 if 1 (1 − t2 )(n−3)/2 g(λ, t) 1 1 dt = 1. B( 2 , 2 (n − 1)) −1 Denote this distribution u ∼ Gn (λ, θ). (iii) What are the “contours” of a Gn (λ, θ) distribution? (iv) If g(λ, t) is an increasing function of t, prove Gn (λ, θ) is unimodal. What is the mode? (v) Prove: u ∼ Gn (λ, θ) =⇒ Hu ∼ Gn (λ, Hθ), ∀H ∈ On . (vi) Obtain the ﬁrst two moments of u ∼ Gn (λ, θ), E u = ρ1 θ,

= {(1 − ρ2 )I + (nρ2 − 1)θθ }/(n − 1),

E uu where

ρi =

1

−1

ti g(λ, t)

(1 − t2 )(n−3)/2 dt < ∞, i = 1, 2. B( 12 , 12 (n − 1))

5. Multivariate normal d

Hint: Use the representation u = tθ +(1−t2 )1/2 ζ, where t = u θ and ζ is distributed uniformly on the sphere orthogonal to θ, ζ [Watson (1983), p. 44]. t (vii) Prove

fx (x) = g(λ, θ x/|x|)(2π)−n/2 exp − 12 x x |=

72

is a density on Rn by transforming to polar coordinates x → (r, u), r ≥ 0, u ∈ S n−1 . (viii) Demonstrate that the distribution Gn (λ, θ) can be obtained by projecting the distribution for x ∼ fx onto S n−1 ; i.e., if x ∼ fx , then u = x/|x| ∼ Gn (λ, θ). Remark: The very special case g(λ, t) = exp(λt) yields the Langevin distribution also known, for n = 2 and 3, as the Fisher-von Mises distribution on the circle and sphere [Fisher (1953), von Mises (1918)]. Tests for the mean direction, θ, of the Langevin distribution are discussed by Fujikoshi and Watamori (1992). Robust estimators of (λ, θ) for the Langevin distribution include the circular median [Mardia (1972)], the normalized spatial median [Ducharme and Milasevic (1987)], and the M-estimator on spheres [Ko and Chang (1993)]. Goodness-of-ﬁt for directional data using smooth tests was considered by Boulerice and Ducharme (1997). Asymptotic behavior of sample mean direction on spheres, without symmetry condition on the p.d.f., was recently derived by Hendriks at al. (1996).

6 Multivariate sampling

6.1 Introduction The basic tools for manipulating random samples from a multivariate distribution are developed in this chapter. We introduce random matrices in Section 6.2 and show the usefulness of the “vec operator” and Kronecker product in this regard. Also, the matrix variate normal distribution is deﬁned and its basic properties are explained. Section 6.3 deals with theorems in the “asymptotic world” as the sample size goes to inﬁnity. These are the central limit theorem, a general Slutsky theorem, and the so-called delta method.

6.2 Random matrices and multivariate sample For A = (aij ) = (a1 , . . . , aq ) ∈ Rpq , we may always regard A as a vector in Rpq where we deﬁne a1 .. . vec(A) = . aq This operation is obviously linear Rpq → Rpq and we may regard A and vec(A) as synonymous.

74

6. Multivariate sampling

For

x1 X = (xij ) = ...

xp

random on Rpq , we may denote the mean of X by E X = M = (µij ). However, the variance of X is a quadruply indexed array consisting of all covariances of the individual entries var X = (σijkl ) = (cov(xij , xkl )) . Since there is no inherent order to this array, we ﬁnd it convenient to impose one by equating var X = var vec(X ) = Ω = (Ωij ) = (cov(xi , xj )) . The element in position (k, l) of the block Ωij is cov(xik , xjl ). One must be very careful to remember that Ω is pq × pq. For instance, if we write X ∼ Nqp (M, Ω), we really mean that vec(X ) ∼ Npq (vec(M ), Ω). In fact, this will be the deﬁnition. Moments of a multivariate normal matrix, Nqp (M, Ω), were given by Wong and Liu (1994). Characterization of a multivariate normal matrix distribution via conditioning is discussed by Gupta and Varga (1992) and Nguyen (1997). The Kronecker product will be very handy for manipulating random matrices. The Kronecker product of A ∈ Rpq and B ∈ Rrs is a block-matrix with the block in position (i, j) being aij B, A ⊗ B = (aij B) ∈ Rpr qs . One can verify the basic properties. Lemma 6.1 The Kronecker product satisﬁes the following: (i) (aA) ⊗ (bB) = ab(A ⊗ B), a, b ∈ R (ii) (A + B) ⊗ C = (A ⊗ C) + (B ⊗ C) (iii) (A ⊗ B) ⊗ C = A ⊗ (B ⊗ C) (iv) (A ⊗ B) = A ⊗ B , (v) (AB) ⊗ (CD) = (A ⊗ C)(B ⊗ D) (vi) (A ⊗ B)−1 = A−1 ⊗ B−1 , whenever A and B are nonsingular. (vii) If v = 0 and u = 0 are eigenvectors of A and B, respectively, Av = λv, and Bu = γu, then v⊗u is an eigenvector of A⊗B corresponding to the eigenvalue λγ. (viii) tr (A ⊗ B) = (tr A)(tr B) (ix) |A ⊗ B| = |A|q |B|p , A ∈ Rpp , B ∈ Rqq

6.2. Random matrices and multivariate sample

75

(x) If A > 0 and B > 0, then A ⊗ B > 0. The following lemma will also be useful for handling random matrices. Its proof is left as an exercise. Lemma 6.2 A ∈ Rrp , X ∈ Rpq , and B ∈ Rqs =⇒ vec(AXB) = (B ⊗ A)vec(X). As a corollary useful for densities (v. Problem 6.4.4) we also have: Corollary 6.1 Let A ∈ Rpp , X ∈ Rpq , and B ∈ Rqq . If Y = AXB, then J(Y → X) = |A|q+ |B|p+ . Proof. Since vec(Y) = vec(AXB) = (B ⊗ A)vec(X), then J(Y → X) = J(vec(Y) → vec(X)) = |B ⊗ A|+ = |A|q+ |B|p+ . 2 Example 6.1 Consider a sample x1 , . . . , xn i.i.d. x, where x ∼ Np (µ, Σ) and forms the “sample matrix” x1 .. X = . . xn

Then, we see that X ∼ Npn (1µ , In ⊗ Σ). Example 6.2 As another example, suppose that z ∼ Np (0, I) and form the “outer product” matrix W = zz = (z1 z, . . . , zp z). Then, obviously, E W = var z = I, but the variance of W depends on the fourth-order moments of z. Since E zi = E zi3 = 0, E zi2 = 1, and E zi4 = 3, it follows easily that E zi z E zi zj zz

= ei , = δij I + ei ej + ej ei ,

from which cov(zi z, zj z) = δij I + ej ei . At this point it becomes useful to deﬁne the “commutation matrix” Kp , a block-matrix whose block in position (i, j) is ej ei ∈ Rpp , 2

Kp = (ej ei ) ∈ Rpp2 .

76

6. Multivariate sampling

For example, for p = 2, we have 1 0 K2 = ··· 0 0

0 0 ··· 1 0

.. . .. . ··· .. . .. .

0

0

0 ··· . 0

1 ··· 0 0

1

This enables one to write succinctly [Magnus and Neudecker (1979)] var W = (I + Kp ). To generalize slightly, suppose that x ∼ Np (0, Σ) and let W = xx . Since d

W = xx = Azz A , where z ∼ Np (0, I) and Σ = AA , the variance of W becomes var W

= =

var Azz A = var (A ⊗ A)vec(zz ) (A ⊗ A)(I + Kp )(A ⊗ A ).

However, since Kp commutes with A ⊗ A (why?) (v. Problem 6.4.2), then, ﬁnally, var W = (I + Kp )(Σ ⊗ Σ). We can also write this expression componentwise as cov(wik , wjl ) = σij σkl + σkj σil , where Σ = (σij ). Suppose that 0 1 x , ∼ N2 0 ρ y

ρ 1

(6.1)

.

We may use the above result to determine the variance of (x2 , y 2 , xy). This is needed later in obtaining the asymptotic distribution of the sample correlation coeﬃcient. For x ∼ unif(B n ), W = xx , the above method may be adapted to help determine E W and var W. The distribution for linear transformations of multivariate normal matrices is straightforward with Lemma 6.2. Proposition 6.1 If A ∈ Rrp , X ∼ Nqp (M, Ω), and B ∈ Rqs , then AXB ∼ Nsr (AMB, (A ⊗ B )Ω(A ⊗ B)) . Proof. Since vec(X ) ∼ Npq (vec(M ), Ω), then vec ((AXB) )

= (A ⊗ B )vec(X ) ∼ Nrs ((A ⊗ B )vec(M ), (A ⊗ B )Ω(A ⊗ B)) .

The proof is complete as (A ⊗ B )vec(M ) = vec ((AMB) ).

2

6.2. Random matrices and multivariate sample

77

Example 6.3 Assuming X ∼ Nqp (M, A ⊗ B), A ≥ 0 is in Rpp and B ≥ 0 is in Rqq . We evaluate E XX . Let X = (x1 , . . . , xq ) and observe, with the choice A = Ip and B = ei , that xi ∼ Np (mi , bii A), where M = (m1 , . . . , mq ). Then, E XX =

q

E xi xi =

i=1

q

(bii A + mi mi )

i=1

leads to the expression E XX = (tr B)A + MM . We now turn to considerations of convergence. For the general sample x1 , . . . , xn i.i.d. x, where E x = µ and var x = Σ, the strong law of large numbers (S.L.L.N.) provides the sample mean as a natural estimate ¯ = nj=1 xj /n for µ: x w.p.1 ¯ → µ. x Of course, Wi = xi xi , i = 1, . . . , n, i.i.d. xx , where E xx = Σ + µµ are n and the S.L.L.N. applies to W = j=1 xj xj /n so that w.p.1 W → Σ + µµ . ˆ =W−x ¯x ¯ , we ﬁnd Then, obviously, if we let Σ ˆ w.p.1 Σ → Σ. ¯x ¯ = Σ/n + µµ , so that However, E x E

n ˆ Σ=Σ n−1

and it has become customary to use this “unbiased” estimate. The reader should have no particular diﬃculty in showing that as explicit functions of the sample matrix, these (unbiased and consistent) estimates may be expressed by ¯= x

n ˆ 1 1 X 1 and S ≡ X QX, Σ= n n−1 (n − 1)

where Q = I − n−1 11 . The estimate S is the sample variance, which is often written as 1 ¯ )(xi − x ¯ ) . (xi − x (n − 1) i=1 n

S=

As an expression of “pythagorus,” we ﬁnd X = QX + PX, where P = I − Q = n−1 11

78

6. Multivariate sampling

and, thus, ¯. X X = X QX + X PX = (n − 1)S + n¯ xx

6.3 Asymptotic distributions The central limit theorem (C.L.T.) states that for any sample x1 , . . . , xn i.i.d. x, where E x = µ and var x = Σ, √ d n(¯ x − µ) → z, where z ∼ Np (0, Σ). d

Now, recall the very general fact that if xn → x on Rp and g : Rp → Rq d is any continuous (with Px probability 1)1 function, then g(xn ) → g(x) on Rq . Note that since matrices in Rpq are really only vectors in Rpq , this result is considerably more general than it might appear at ﬁrst. Thus, if Σ is nonsingular (the singular case goes through as well; v. Problem 6.4.10), x − µ) → χ2p . n(¯ x − µ) Σ−1 (¯ d

There is another very basic fact that derives from the Cram´er-Wold theorem and the (univariate) Slutsky theorem. d

d

Lemma 6.3 (Multivariate Slutsky) If Xn → X on Rpq and Yn → C on Rrs where C is any constant matrix, then d

(Xn , Yn ) → (X, C) on Rpq × Rrs . Proof. From Cram´er-Wold Proposition 2.10, for any linear combination d tij xn,ij → tij xij , i,j

i,j

skl yn,kl

d

→

k,l

skl ckl ,

k,l

and from the univariate Slutsky theorem, d tij xn,ij + skl yn,kl → tij xij + skl ckl . i,j

i,j

k,l

k,l

Using Cram´er-Wold again, the conclusion is reached.

2

A more general statement on metric spaces can be found in Billingsley (1968, p. 27). It follows, of course, that for any continuous function, d

g(Xn , Yn ) → g(X, C). 1 Let

Cg = {t ∈ Rp : g is continuous at t}. Then, g is continuous with Px probability 1 means that Px (Cg ) = P (x ∈ Cg ) = 1.

6.3. Asymptotic distributions

79

As a simple example n(¯ x − µ) S−1 (¯ x − µ) → χ2p . d

One more general proposition: √ d Proposition 6.2 (Delta method) If n(xn − c) → z on Rp and g : p q R → R is diﬀerentiable at c, then √ d n (g(xn ) − g(c)) → Dg(c) z. Proof. This is simply because by the very deﬁnition of the derivative at c, the function h(t)/|t − c|, t = c k(t) = 0, t = c, where h(t) = (g(t) − g(c)) − Dg(c) (t − c), is continuous at c, and we may, therefore, write √ √ √ n (g(xn ) − g(c)) = Dg(c) n(xn − c) + k(xn )| n(xn − c)|. d

Using Slutsky’s theorem, we may conclude that since k(xn ) → 0 and √ d | n(xn − c)| → |z|, √ d n (g(xn ) − g(c)) → Dg(c) z. 2 This, of course, applies directly to the C.L.T. to give √ d n (g(¯ x) − g(µ)) → Nq (0, Dg(µ)ΣDg(µ) ) . However, consider a more elaborate application: Let x1 , . . . , xn be i.i.d. x as before with E x = 0 and var x = Σ. Then, let Wi = xi xi , i = 1, . . . , n, and W = xx so that Wn W W1 , . . . , are i.i.d. x x1 xn with

E

and

var

W x

=

=

Σ 0

var W cov(vec(W), x) cov(x, vec(W)) Σ

By the C.L.T., √

W x

n

W−Σ ¯ x

d

→ Npp+1 (0, Ω)

≡ Ω.

80

6. Multivariate sampling

and the reader may then use Lemma 6.3 to ﬁnd that √ √ √ √ d ˆ − Σ) = n(W − Σ) − √1 ( n¯ n(Σ x)( n¯ x) → Npp (0, var W) n and, of course, √

d

n(S − Σ) → Npp (0, var W).

Note that since the function S is unchanged if x is replaced by x − µ, this result is automatically valid for the more general case where E x = µ. The expression for var W was given in Example 6.2 for the normal case, and the elliptical case is treated in the sequel in Example 13.6. Unfortunately, var W is seldom of a particular tractable form. It depends on the fourth-order multivariate cumulants of x. The relation between product-moments and multivariate cumulants is rather technical and is relegated to Appendix B. There, it is proven generally for W = (wij ) = xx that ijkl ik jl kl ij il jk cov(wik , wjl ) = µijkl 1111 − µ11 µ11 = k1111 + k11 k11 + k11 k11 ,

where the µ’s are the product-moments and the k’s are the cumulants of x. Example 6.4 For a sample of size n from a bivariate distribution with ﬁnite fourth-order moments, we ﬁnd the asymptotic distribution 2 2 σ1 s1 √ d n s12 − σ12 → N3 (0, Ω), s22 σ22 where

µ14 − (µ12 )2 Ω= · ·

12 1 µ12 31 − µ11 µ2 12 2 µ22 − (µ12 11 ) ·

1 2 µ12 22 − µ2 µ2 12 2 µ12 . 13 − µ11 µ2 2 µ4 − (µ22 )2

The product-moments are µ14 µ12

µ12 31 µ12 11 µ22 µ12 22

= E (x1 − µ1 )4 , = E (x1 − µ1 )2 = σ12 , = E (x1 − µ1 )3 (x2 − µ2 ), = E (x1 − µ1 )(x2 − µ2 ) = σ12 , = E (x2 − µ2 )2 = σ22 ,

= E (x1 − µ1 )2 (x2 − µ2 )2 , etc.

¯ and S will not be asymptotically independent unless all thirdIn general, x order product-moments of x in cov (vec(W), x) are null. But this is exactly the case when z = Σ−1/2 x has a spherical distribution since cov (vec(W), x)

=

cov (vec(xx ), x)

6.4. Problems

=

81

(Σ1/2 ⊗ Σ1/2 ) cov (vec(zz ), z) Σ1/2

and all third-order product-moments of z are null (v. Problem 4.6.5). However, if the underlying random vector x is already normal, then things reduce considerably. For p = 2, the correlation coeﬃcient, r, is a very simple function of S and, thus, it should be straightforward for the reader to obtain the asymptotic distribution of r (v. Problems 6.4.8-6.4.9). In fact, since this function is unchanged if the individual coordinates are normalized, we may assume at the outset that 0 1 ρ x , . ∼ N2 0 ρ 1 y

6.4 Problems 1. Prove Lemma 6.2: If A ∈ Rrp , X ∈ Rpq , and B ∈ Rqs , then vec(AXB) = (B ⊗ A)vec(X). 2. Let A, B ∈ Rpp and Kp be the “commutation matrix.” Show the following: p p (i) Kp = i=1 j=1 ei ej ⊗ ej ei , (ii) Kp vec(A) = vec(A ), (iii) Kp (A ⊗ B) = (B ⊗ A)Kp , (iv) tr A B = [vec(A)] vec(B), (v) If A is symmetric, tr A2 = [vec(A)] 12 (I + Kp )vec(A). |=

3. Show that if Z ∼ Nqp (0, I) and P and Q in Rpp are orthogonal QZ. projections such that PQ = 0, then PZ Hint: Obtain PZ var . QZ 4. Obtain the p.d.f. of X ∼ Nqp (M, A ⊗ B), where A > 0 is in Rpp and B > 0 is in Rqq : pq q p f (X) = (2π)− 2 |A|− 2 |B|− 2 etr − 12 A−1 (X − M)B−1 (X − M) , where etr(·) ≡ exp[tr(·)]. Hint: Let X = A1/2 ZB1/2 + M, where Z ∼ Nqp (0, Ip ⊗ Iq ), and use Corollary 6.1. 5. Assume E E = 0 and var E = In ⊗ Σ, Σ ≥ 0 is in Rpp . Show that (i) var E = Σ ⊗ In , (ii) E E AE = (tr A)Σ.

82

6. Multivariate sampling

6. Assume Σ22 is nonsingular and Σ11 n (X1 , X2 ) ∼ Np1 +p2 1(µ1 , µ2 ), In ⊗ Σ21

Σ12 Σ22

,

where Xi ∈ Rnpi , µi ∈ Rpi , and Σij ∈ Rppij , i, j = 1, 2. Prove X1 | X2 ∼ Npn1 (1µ1 + (X2 − 1µ2 )B , In ⊗ Σ11.2 ) , with B = Σ12 Σ−1 22 . 7. For W = xx , in each case determine E W and var W: 1 ρ , (i) x ∼ N2 (0, Σ) and Σ = ρ 1 1 (ii) x ∼ unif(S ), (iii) x ∼ unif(B 2 ). 8. Assume (xi , yi ) , i = 1, . . . , n, are i.i.d. 0 1 x ∼ N2 , 0 ρ y

ρ 1

and let r be the sample correlation coeﬃcient. Prove the asymptotic

√ d result n(r − ρ) → N 0, (1 − ρ2 )2 . 9. Fisher’s z-transform is z

=

ζ

=

1+r , 1−r 1+ρ tanh−1 (ρ) = 12 log . 1−ρ

tanh−1 (r) =

1 2

log

(i) Show that it is a “variance stabilizing transformation” for the √ d correlation coeﬃcient: n − 3(z − ζ) → N (0, 1). (ii) Use the fact that z is a monotone function of r to obtain an approximate (1 − α)100% conﬁdence interval for ρ, zα/2 zα/2 , tanh z + , tanh z − (n − 3)1/2 (n − 3)1/2 where P (N (0, 1) > zα/2 ) = α/2. 10. Let x1 , . . . , xn i.i.d. x, where E x = µ and var x = Σ. Prove that n(¯ x − µ) Σ− (¯ x − µ) → χ2r , d

where r = rank Σ. 11. Demonstrate the following representation of Mahalanobis distance: n h xi − n1 j=1 h xj = di , sup 2 1/2 |h|=1 n n 1 1 k=1 h xk − n j=1 h xj (n−1)

6.4. Problems

83

¯ ) S−1 (xi − x ¯ )]1/2 is the Mahalanobis distance from where di = [(xi − x ¯. xi to x Remark: This was used by Stahel (1981) and Donoho (1982) to suggest the robust estimate of location as a weighted average n w(ui )xi ˆ = i=1 µ , n i=1 w(ui ) where w(·) is a positive and strictly decreasing function and |h xi − medj (h xj )| . |h|=1 medk |h xk − medj (h xj )|

ui = sup

The notation “med” refers to the ordinary median. 12. Multivariate familial data [Konishi and Khatri (1990)]. Suppose a random sample of n families on x = (x1 , . . . , xp ) ∈ Rp with E x = µ and var x = Σ. Let x1i .. Zi = . , i = 1, . . . , n, xki ,i

denote the measurements on the ith family with ki ≥ 1 siblings, where xji = (x1ji , . . . , xpji ) , j = 1, . . . , ki , is the score of the jth child on p characteristics. It is assumed that Z1 , . . . , Zn are mutually independent and E Zi var Zi

= 1ki µ , =

(Iki ⊗ Σ) + (1ki 1ki − Iki ) ⊗ Σs .

The matrix Σs reﬂects the dependence among siblings. For the estimation of Σ, let ¯1 x ki ¯ = .. ¯ i )(xji − x ¯ i ) , = (xji − x , V X i . ¯ n x

j=1

¯ i = (x1i + · · · + xki ,i )/ki . Further, let B ∈ Rnn , B ≥ 0, such where x that B1n = 0. (i) Prove that

ˆ = (tr B)−1 Σ

¯ + ¯ BX X

n

ωi Vi

,

i=1

where the weights ω1 , . . . , ωn are non-negative constants, satisﬁes n ˆ = Σ+(tr B)−1 ωi (ki − 1) − tr[B(In − D−1 EΣ n )] (Σ−Σs ), i=1

84

6. Multivariate sampling

where Dn = diag(k1 , . . . , kn ). ˆ is unbiased for Σ. (ii) Find a condition on the weights so that Σ (iii) The corresponding estimate of Σs is given by n −1 ¯ ˆ ¯ νi V i , Σs = (tr B) X BX + i=1

where ν1 , . . . , νn are constants. Prove that for weights satisfying the condition n νi (ki − 1) + tr(BD−1 n ) = 0, i=1

ˆ s is unbiased for Σs . Σ A multivariate familial model for interclass correlation, with a “mother” for each family, was considered earlier by Srivastava et al. (1988). Principal component analysis for the model described here was developed by Konishi and Rao (1992). A general description of principal components is given in Chapter 10.

7 Wishart distributions

7.1 Introduction As before,

x1 X = ...

xn

¯ and S, denotes the sample matrix from which x n¯ x = X 1, ¯, (n − 1)S = X X − n¯ xx provide consistent unbiased estimates for µ and Σ, respectively. In Section 7.2, the maximum likelihood estimates of µ and Σ are derived assuming x1 , . . . , xn i.i.d. x with x ∼ Np (µ, Σ), Σ > 0. The fundamental ¯ and S is proved in Proposiresult about the joint distribution of x tion 7.1. The basic properties of Wishart distributions are studied in Section 7.3. Section 7.4 presents the Box-Cox transformation to enhance the multivariate normality of the data.

¯ and S 7.2 Joint distribution of x ¯ and S are “optimal” in some respects. Denote With underlying normality, x V = (n − 1)S. Using the notation exp[tr(·)] = etr(·), the p.d.f. for X can

86

7. Wishart distributions

be written in various ways: f (X)

=

− np 2

(2π)

−n 2

|Σ|

exp − 12

n

(xi − µ) Σ−1 (xi − µ)

i=1

= =

−1 n n ¯ |Σ|− 2 e− 2 µ Σ µ exp − 12 tr Σ−1 X X + nµ Σ−1 x ' & np n (7.1) x − µ)(¯ x − µ) ] Σ−1 . (2π)− 2 |Σ|− 2 etr − 12 [V + n(¯ (2π)−

np 2

By general properties of exponential families [Fraser (1976), pp. 339, 342, 406, or Casella and Berger (1990), pp. 254-255, 263], it is plain that ¯ ) (or any one-to-one function such as (S, x ¯ )) is minimal suﬃcient (X X, x and complete for (Σ, µ), so that by the Rao-Blackwell/Lehmann-Scheﬀ´e ¯ and S have minimum theorems, among all unbiased estimates of µ and Σ, x ¯ ) is the MVUE (Minimum Variance Unbiased variance. We say that (S, x Estimate) of (Σ, µ). ˆ and Furthermore, to obtain the maximum likelihood estimates (MLE) µ ˆ when n − 1 ≥ p, we minimize Σ 1 x − µ) Σ−1 (¯ x − µ) (7.2) VΣ−1 + (¯ n ˆ = x ¯ , so we need only and (since the last term is ≥ 0) it is clear that µ minimize 1 ln |nV−1 Σ| + tr VΣ−1 , n where the constant, ln |nV−1 |, was added. The condition n − 1 ≥ p ensures that V is nonsingular w.p.1. This is proved later in Corollary 7.2. But then, letting T = nV−1 Σ, we need only determine the T that minimizes ln |Σ| + tr

ln |T| + tr T−1 . However, this is accomplished when all the eigenvalues of T are 1 so that T = I and we conclude altogether ˆ = 1 V. ˆ =x ¯ and Σ µ n Remark: It is a well-known result, which can be traced back to Gauss, that the only location family, f (x − θ), of p.d.f. on R for which x ¯ is a MLE of θ originates from the normal density. This MLE characterization of normal density also holds on Rp [Stadje (1993)]. ¯ and S. It is obvious that Let us consider the exact distribution of x 1 ¯ ∼ Np µ, Σ . x n d

We begin by representing x = Az + µ, for any AA = Σ, which, for the d sample matrix, means that X = ZA + 1µ . Thus, (¯ x, Sx ) = (A¯ z + µ, ASz A ). d

7.3. Properties of Wishart distributions

87

|=

|=

|=

However, in Z = (zij ), all the components are i.i.d. N (0, 1) so that even the columns are mutually independent. Thus, with orthogonal projections P = QZ (v. Problem 6.4.3), n−1 11 and Q = I − n−1 11 , it is clear that PZ ¯ Sz , hence and since n¯ z = Z P1 and (n − 1)Sz = Z QQZ, we see that z ¯ Sx . x If next we express Q = HH , where H gives an orthonormal basis for ⊥ 1 (of dimension (n − 1)), it is made plain that (n − 1)Sz = Z HH Z = U U, where Z H = U = (u1 , . . . , un−1 ), ui i.i.d. Np (0, I). Accordingly, we make the following deﬁnition. Deﬁnition 7.1 Wishart distribution: m d zi zi , zi i.i.d. Np (0, I). W ∼ Wp (m) iﬀ W = i=1

V

∼ Wp (m, Σ) iﬀ V = AWA , Σ = AA , W ∼ Wp (m). d

Thus, we have the fundamental statistical result:

¯ ¯ ∼ Np (µ, Σ/n), (n − 1)S ∼ Wp (n − 1, Σ), and x x

|=

Proposition 7.1 For xi i.i.d. Np (µ, Σ), i = 1, . . . , n, S.

One may, of course, go to some trouble to obtain an explicit density for the Wishart. However, one needs primarily to understand some of its basic properties and the density will not really reveal very much.

7.3 Properties of Wishart distributions The distribution of the trace of W ∼ Wp (m) follows almost immediately from the deﬁnition. Proposition 7.2 W ∼ Wp (m) =⇒ tr W ∼ χ2mp . Proof. By deﬁnition of Wp (m), d

tr W = tr

m i=1

zi zi =

m

zi zi ,

i=1

where zi are i.i.d. Np (0, I). From Proposition 4.4, zi zi ∼ χ2p . Corollary 3.1 then gives tr W ∼ χ2mp . 2 Now, a useful lemma to determine when V ∼ Wp (m, Σ) is nonsingular w.p.1. is the following: Lemma 7.1 Z = (zij ) ∈ Rnn with zij i.i.d. N (0, 1) =⇒ P (|Z| = 0) = 0.

88

7. Wishart distributions

Proof. The proof proceeds by induction. The result is true for n = 1, as z11 has an absolutely continuous distribution. Next, partition z11 z12 Z= z21 Z22 n−1 and assume the result holds for Z22 ∈ Rn−1 . Then,

P (|Z| = 0)

= P (|Z| = 0, |Z22 | = 0) + P (|Z| = 0, |Z22 | = 0) = P (z11 = z12 Z−1 22 z21 , |Z22 | = 0) = E P (z11 = z12 Z−1 22 z21 , |Z22 | = 0 | z12 , z21 , Z22 ) =

0. 2

A slight generalization is contained in Corollary 7.1 Z = (zij ) ∈ Rnn with zij i.i.d. N (0, 1) =⇒ P (|Z| = t) = 0, ∀t. Proof. P (|Z| = t) = E P z11 = z12 Z−1 22 z21 +

t , |Z22 | = 0 | z12 , z21 , Z22 |Z22 |

= 0. 2 It should be observed that Lemma 7.1 and Corollary 7.1 remain valid if Z has any absolutely continuous distribution. We can now prove [Stein (1969), Dykstra (1970)]: Proposition 7.3 W ∼ Wp (m), m ≥ p =⇒ W is nonsingular w.p.1. d

Proof. The representation W = Z Z, where Z = (z1 , . . . , zm ) and zi ’s are i.i.d. Np (0, I), gives rank W = rank Z Z = rank Z ≥ rank (z1 , . . . , zp ) d

whence rank W

w.p.1 = p

w.p.1 = p.

2

Its corollary gives a condition on the sample size and the population variance for the sample variance matrix S to be nonsingular w.p.1. Corollary 7.2 V ∼ Wp (m, Σ), m ≥ p, |Σ| = 0 =⇒ |V| = 0 w.p.1. Eaton and Perlman (1973) established that the sample variance matrix S is nonsingular w.p.1 for independent observations, which are not necessarily normal or identically distributed. Concerning linear transformations of Wishart matrices, we have

7.3. Properties of Wishart distributions

89

Proposition 7.4 V ∼ Wp (m, Σ), B ∈ Rqp =⇒ BVB ∼ Wq (m, BΣB ). d

Proof. Let W ∼ Wp (m). Since V = AWA , for any AA = Σ, then BVB = (BA)W(BA) ∼ Wq (m, BAA B ). d

2 Example 7.1 Suppose W ∼ Wp (m). What is E WAW for a ﬁxed A ≥ 0? d

Since HWH = W, for all H ∈ Op , we see that WAW

= HWH AHWH , ∀H ∈ Op d

= HWDWH , d

where H was chosen to diagonalize A, H AH = D = diag(λi ). Thus, d m E WAW = H(E WDW)H . But using W = i=1 zi zi , where zi ∼ Np (0, I) are independent, we ﬁnd E zi zi Dzj zj E WDW = i,j

=

E zi zi Dzi zi +

E zi zi Dzj zj

i =j

i

= mE xx Dxx + m(m − 1)E xx Dyy , where x and y are i.i.d. Np (0, I). However, λi E x2i xx E xx Dxx = i

=

λi (I + 2ei ei )

i

= and E xx Dyy

=

(tr A)I + 2D

λi E xi yi xy

i

=

λi ei ei

i

= D. Hence, E WDW = m(tr A)I + m(m + 1)D and, ﬁnally, we obtain E WAW = m(tr A)I + m(m + 1)A. The characteristic function of Wishart distributions also follows from basic principles.

90

7. Wishart distributions

Example 7.2 The characteristic function of V ∼ Wp (m, Σ), evaluated d

at S symmetric, is deﬁned by cV (S) = E exp(i tr SV). Write V = m A j=1 zj zj A , for any AA = Σ, and diagonalize A SA = HDH to obtain m cV (S) = E exp i tr SA zj zj A j=1

= E exp i tr HDH

= E exp i tr D

m

zj zj

= E exp i tr D

m

zj zj since H zj = zj

j=1 p m

(H zj )(H zj )

j=1

= E exp i

j=1

m

d

2 zjk dk , where D = diag(d1 , . . . , dp )

j=1 k=1

=

=

m

p

j=1 k=1 p m

cχ21 (dk ) (1 − 2idk )−1/2

j=1 k=1

= |I − 2iD|−m/2 = |I − 2iSΣ|−m/2 . Hence, the characteristic function is given by cV (S) = |I − 2iΣS|−m/2 . Now, consider some results concerning the marginals of a Wishart. For this reason, partition V ∈ Rpp as V11 V12 , V= V21 V22 where V11 ∈ Rrr and V22 ∈ Rss , r + s = p. The matrix Σ is partitioned similarly. Proposition 7.5 V ∼ Wp (m, Σ) =⇒ V11 ∼ Wr (m, Σ11 ). Proof. Choose B = ( Ir

0 ) ∈ Rrp in Proposition 7.4.

Concerning independence, we have:

2

Proposition 7.6 V ∼ Wp (m, Σ) and Σ12 = 0 =⇒ V11

|=

7.3. Properties of Wishart distributions

91

V22 .

Proof. By the very deﬁnition, X X X X Y d V = U U = ( X Y ) = , Y Y X Y Y

|=

where U = (u1 , . . . , um ) and the ui ’s are i.i.d Np (0, Σ). Then, it suﬃces to recall (v. Problem 6.4.6) that for a multivariate normal, Σ12 = 0 implies Y. 2 X The previous two propositions are not surprising if we consider their statistical interpretation. First, the distribution of V is associated with the sample variance based on all p components, whereas that of V11 corresponds to a sample variance but considering only the ﬁrst r components. Second, if Σ12 = 0, then the distributions of V11 and V22 are associated with sample variances based on two independent subvectors of dimension r and s, r + s = p. The next proposition, the proof of which is left as an exercise, relates to sums of independently distributed Wishart matrices. It has to do with the way one would pool information from independent samples to estimate the population variance (v. Problem 8.9.1). indep ∼ Wp (mi , Σ), i = 1, . . . , k, then k k Vi ∼ Wp mi , Σ .

Proposition 7.7 If Vi

i=1

i=1

Lemma 7.2 Let H = (h1 , . . . , hr ), where the hi ’s are orthonormal in Rn and Z ∼ Npn (0, In ⊗ Ip ). Then, 1. H Z ∼ Npr (0, Ir ⊗ Ip ), 2. Z HH Z ∼ Wp (r). Proof. Using Proposition 6.1, H Z ∼ Npr (0, (H H) ⊗ Ip ). This proves part 1 because H H = Ir . Since Z hi , i = 1, . . . , r, arei.i.d. Np (0, I), part r 2 follows from the Wishart deﬁnition: Z HH Z = i=1 (Z hi )(Z hi ) ∼ 2 Wp (r). Proposition 7.8 Let

x1 X = ... ∼ Npn (0, In ⊗ Σ).

xn

If V ⊂ Rn is a linear subspace, dim V = r, and P is the orthogonal projection on V, then X PX ∼ Wp (r, Σ).

92

7. Wishart distributions

Proof. Choose an orthonormal basis H = (h1 , . . . , hr ) for V and observe d

that P = HH and r = rank P = dim V. Write Σ = AA and X = ZA , where Z ∼ Npn (0, In ⊗ Ip ). Therefore, using Lemma 7.2, X PX

= X HH X d = AZ HH ZA = AWA , where W ∼ Wp (r). d

Hence, X PX ∼ Wp (r, Σ).

2

General results on Wishart and chi-square distributions associated with matrix quadratic forms are available in Mathew and Nordstr¨ om (1997). A fundamental result on marginals useful in the sequel is now stated, but ﬁrst −1 recall the notation V11.2 = V11 − V12 V22 V21 , where V was partitioned as on page 90. Proposition 7.9 If V ∼ Wp (m, Σ), m ≥ p, Σ > 0, then V11.2

and V11.2

|=

V21 | V22 V22

∼ Wr (m − s, Σ11.2 ), ∼ Nrs (V22 Σ−1 22 Σ21 , V22 ⊗ Σ11.2 ), ∼ Ws (m, Σ22 ),

(V21 , V22 ).

Proof. As before, write d

V=UU=

X Y

(X Y) =

X X X Y Y X Y Y

,

where U ∼ Npm (0, Im ⊗ Σ). Thus, X | Y ∼ Nrm (YΣ−1 22 Σ21 , Im ⊗ Σ11.2 ) (v. Problem 6.4.6). Let P = Y(Y Y)−1 Y be the orthogonal projection on the column space of Y and Q = I − P, rank Q = m − s. It is clear, since Y = PY, that V11.2 V21 V22

= X [I − Y(Y Y)−1 Y ]X = (QX) (QX), = Y X = (PY) (PX), = Y Y = (PY) (PY).

|=

|=

Since Y X | Y ∼ Nrs ((Y Y)Σ−1 22 Σ21 , (Y Y)⊗Σ11.2 ) depends only on Y Y, −1 s then V21 | V22 ∼ Nr (V22 Σ22 Σ21 , V22 ⊗ Σ11.2 ). From Proposition 7.8 and QY = 0, V11.2 | Y ∼ Wr (m − s, Σ11.2 ), which does not depend on Y; hence, V11.2 Y and V11.2 ∼ Wr (m − s, Σ11.2 ), unconditionally. It is clear V22 ∼ Ws (m, Σ22 ). Only independence remains to be shown. QX. To see this, note PQ = 0 and However, conditionally on Y, PX P P var X|Y = (P , Q ) ⊗ Σ11.2 Q Q P ⊗ Σ11.2 0 = . 0 Q ⊗ Σ11.2

Hence, given Y, V11.2

|=

7.3. Properties of Wishart distributions

93

V21 . Finally, using Proposition 2.13,

E [f (V11.2 ) · g(V21 , V22 )]

= E [E f (V11.2 ) g(V21 , V22 ) | Y] = E {E [f (V11.2 ) | Y] E [g(V21 , V22 ) | Y]} = E {Ef (V11.2 ) E [g(V21 , V22 ) | Y]} = Ef (V11.2 ) · Eg(V21 , V22 ), 2

which proves independence.

Proposition 7.9 with r = 1 and s = p−1 can be used to prove inductively several results concerning Wishart distributions. Here are two corollaries: the distribution of the generalized variance, |V|, and the Wishart density. Corollary 7.3 If V ∼ Wp (m, Σ), m ≥ p, Σ > 0, then |V| ∼ |Σ|

p

χ2m−p+i ;

i=1

i.e., |V|/|Σ| is distributed as a product of p mutually independent chi-square variables.

|=

Proof. The result obviously holds for p = 1. Assume it holds for p − 1. Let r = 1 and s = p − 1 in Proposition 7.9. Then, |V| = v11.2 |V22 |, where V22 . From the v11.2 ∼ σ11.2 χ2m−p+1 , V22 ∼ Wp−1 (m, Σ22 ), and v11.2 induction hypothesis, |V22 | ∼ |Σ22 |

p−1

χ2m−p+1+i = |Σ22 | d

i=1

p

χ2m−p+i

i=2

2

and the conclusion follows. Corollary 7.4 If W ∼ Wp (m), m ≥ p, then the p.d.f. of W is

1 |W|(m−p−1)/2 etr − 12 W , W > 0, 2m p where Γp (u) = π p(p−1)/4 i=1 Γ u − 12 (i − 1) , u > 12 (p − 1). fW (W) =

1

2mp/2 Γp

(7.3)

Proof. The result holds for p = 1, as the density reduces to a chi-square density. Let r = 1 and s = p − 1 in Proposition 7.9, then w21

w11.2 | W22 W22

∼ χ2m−p+1 , ∼ Np−1 (0, W22 ), ∼ Wp−1 (m).

Thus, the joint p.d.f. of (w11.2 , w21 , W22 ) is 1 (m−p+1)/2−1 exp(− 12 w11.2 ) w 2(m−p+1)/2 Γ[ 12 (m − p + 1)] 11.2 −1 W22 w21 ) · (2π)−(p−1)/2 |W22 |−1/2 exp(− 12 w21

94

7. Wishart distributions

·

1 |W22 |(m−p)/2 etr(− 12 W22 ). m(p−1)/2 2 Γp−1 ( 12 m)

Make the change of variables (w11.2 , w21 , W22 ) → (w11 , w21 , W22 ) with jacobian J(w11.2 , w21 , W22 → w11 , w21 , W22 ) = J(w11.2 → w11 ) = 1 −1 W22 w21 while using the relations |W| = w11.2 |W22 | and w11.2 = w11 −w21 to get the result. 2 The reader should check that Γp (u) is a generalized gamma function in the sense that Γ1 (u) = Γ(u), u > 0. The density of V ∼ Wp (m, Σ), Σ > 0, m ≥ p, follows directly from the transformation V = AWA , for any AA = Σ, and the jacobian in Proposition 2.19 (v. Problem 7.5.7). James (1954) and Olkin and Roy (1954) proposed a constructive proof by jacobians of transformations on k-surfaces (manifolds). It requires a knowledge of diﬀerential forms and integration on k-surfaces which goes beyond the scope of this book. The theory of singular Wishart distributions (m < p) is available in Uhlig (1994). The function (7.3) is a density function even when the number of degrees of freedom m ∈ R, possibly noninteger, satisﬁes m > p − 1 [Muirhead (1982), p. 62].

7.4 Box-Cox transformations A method [Andrews et al. (1971)] that is an extension of the technique of Box and Cox (1964) is described for obtaining data-based transformations of multivariate observations to enhance the normality of their distribution. Speciﬁcally, power transformations of the original variables are estimated to eﬀect both marginal and joint normality. The likelihood method, used by Box and Cox (1964) for the univariate problem, is the one adopted here for the multivariate case. The simple family of power transformations deﬁned by λj (λj ) xj = (xj − 1)/λj , λj = 0, λj = 0, ln xj , j = 1, . . . , p, will be considered. Each variable xj must be non-negative, otherwise, with a known lower bound, we may add a constant suﬃciently large, aj , and consider xj + aj as the original variable. Let x1 .. X = . = (xij ) ∈ Rnp , xn

7.4. Box-Cox transformations

95

X(λ)

(λ) x1 (λj ) n . = .. = (xij ) ∈ Rp , (λ) xn

be the sample matrices of the original and transformed data, respectively, where λ = (λ1 , . . . , λp ) is the unknown vector of power transformation parameters. If λ is the vector of parameters yielding joint normality, Np (µ, Σ), the density of X(λ) is from (7.1): f (X(λ) ) = (2π)− 2 |Σ|− 2 ! ) ( ·etr − 12 V(λ) + n(¯ x(λ) − µ)(¯ x(λ) − µ) Σ−1 , np

n

where 1 (λ) x , n i=1 i n

¯ (λ) x

=

V(λ)

=

n

(λ)

(xi

(λ)

¯ (λ) )(xi −x

¯ (λ) ) . −x

i=1

The jacobian of the transformation, J(X(λ) → X), is J=

p n

λ −1

xijj

.

j=1 i=1

Hence, the density of the genuine data X is f (X) = f (X(λ) ) · J = (2π)−

np 2

! ) ( n |Σ|− 2 etr − 12 V(λ) + n(¯ x(λ) − µ)(¯ x(λ) − µ) Σ−1 · J.

The log-likelihood of (Σ, µ, λ) is, up to an additive constant, n 1 ln |Σ| − tr(V(λ) Σ−1 ) 2 2 p n n (λ) − µ) Σ−1 (¯ x(λ) − µ) + (λj − 1) ln xij . x − (¯ 2 j=1 i=1

l(Σ, µ, λ) = −

(7.4)

For a speciﬁed λ, the maximum likelihood estimate of (Σ, µ), exactly as for (7.2), is given by ¯ (λ) . V(λ) /n, x If these estimates are substituted in (7.4), the maximized log-likelihood function is, up to an additive constant, lmax (λ) = −

p n n ln |V(λ) | + (λj − 1) ln xij , 2 j=1 i=1

(7.5)

96

7. Wishart distributions

a function of p parameters which can be computed and studied. The maxˆ may be obtained by numerically maximizing imum likelihood estimate λ (7.5). Also, conﬁdence regions for λ may be obtained. One such (1−α)100% conﬁdence region for λ based on asymptotic considerations [Fraser (1976), p. 357] is ˆ − lmax (λ) ≤ 1 χ2 {λ : lmax (λ) 2 1−α,p }, where χ21−α,p is the (1 − α)-quantile of a χ2p distribution. The likelihood criterion used here speciﬁes joint normality rather than marginal normality as the goal of the transformation.

7.5 Problems ˆ were derived by ˆ and Σ 1. The maximum likelihood estimates µ minimizing 1 ln |Σ| + tr VΣ−1 + (¯ x − µ) Σ−1 (¯ x − µ). n Derive the MLE, this time by calculus, using the vector and matrix diﬀerentiation rules of Problems 1.8.9-1.8.10. 2. Prove Proposition 7.7. 3. Show that if V ∼ Wp (m, Σ), then: (i) var V = m[Σ ⊗ Σ + (σ j σ i )] = m(I + Kp )(Σ ⊗ Σ) where Kp is the “commutation matrix.” (ii) Prove Proposition 7.7 again, but using characteristic functions this time. 4. Let W ∼ Wp (m), m ≥ p. Prove: |=

(i) 1/(t W−1 t) ∼ χ2m−p+1 , for any t, |t| = 1. W and px (0) = 0, then x is independent of (ii) If x (x x) ∼ χ2m−p+1 . (x W−1 x) d

Hint: HWH = W, ∀H ∈ Op . 5. Assume W ∼ Wp (m), m ≥ p, and A ≥ 0. Prove: (i) E W = mI, (ii) E W−1 = I/(m − p − 1). 6. Moments of generalized variance. (i) Let W ∼ Wp (m), m ≥ p. Prove E |W|h = 2ph

Γp ( 12 m + h) , h > 12 (p − m − 1). Γp ( 12 m)

7.5. Problems

97

Hint: E |W|h has an integrand of the form of a Wp (m + 2h) density. Use the normalizing constant cp,m = [2mp/2 Γp ( 12 m)]−1 . (ii) If V ∼ Wp (m, Σ), m ≥ p, Σ > 0, then E |V|h = |Σ|h 2ph

Γp ( 12 m + h) , h > 12 (p − m − 1). Γp ( 12 m)

7. Wishart density. Obtain the p.d.f. of V ∼ Wp (m, Σ), m ≥ p, Σ > 0: fV (V) =

2mp/2 Γ

1

1 |V|(m−p−1)/2 etr − 12 Σ−1 V , m/2 p 2 m |Σ|

V > 0. 8. Let W ∼ Wp (m) and consider the correlation matrix R = (rij ), where wij rij = 1/2 1/2 . wii wjj Demonstrate that the density of R is f (R) =

[Γ( 12 m)]p |R|(m−p−1)/2 . Γp ( 12 m)

Hint: Use the transformation W → w11 , . . . , wpp , R. 9. Inverted Wishart distribution. Derive the p.d.f. of U = V−1 , where V ∼ Wp (m, Σ), m ≥ p, Σ > 0:

1 |U|−(m+p+1)/2 etr − 12 Σ−1 U−1 , fU (U) = mp/2 1 m/2 2 Γp 2 m |Σ| U > 0. 10. Assume W ∼ Wp (m) and deﬁne W = TT for a unique T ∈ L+ p. (i) Prove tij ∼ N (0, 1), 1 ≤ i < j ≤ p, and t2ii ∼ χ2m−i+1 , 1 ≤ i ≤ p, are all mutually independent. (ii) Using (i) prove tr W ∼ χ2pm . (iii) Using (i) again,prove that if V ∼ Wp (m, Σ), m ≥ p, Σ > 0, p then |V| ∼ |Σ| i=1 χ2m−p+i .

8 Tests on mean and variance

8.1 Introduction ¯ ) on a good footing in Chapter 7, we Having laid the distribution of (S, x now present inference problems such as the Hotelling-T 2 test on the mean vector, the simultaneous conﬁdence intervals on means, the inference about multiple and partial correlation coeﬃcients, the test of sphericity, and the test of equality of variances. In some cases, the tests are optimal in some sense. This is the case of the Hotelling-T 2 test and the test of multiple correlation, which are shown to be uniformly most powerful invariant (UMPI). The asymptotic distribution of eigenvalues, both in the one-sample and two-sample cases, is treated in Section 8.8. Tables of critical points with references to applications for most multivariate tests are available in Kres (1983). The approach adopted here rests mainly on likelihood ratio tests, although other general and valid testing procedures based on minimization of divergence measures exist in the literature [Wakaki et al. (1990)].

8.2 Hotelling-T 2 Now, assume that x1 , . . . , xn are i.i.d. x with x ∼ Np (µ, Σ) and Σ > 0. The properties of Wishart distributions in Chapter 7 provide an easy way to obtain the distribution of the Hotelling-T 2 statistic x − µ0 ) S−1 (¯ x − µ0 ), T 2 = n(¯

(8.1)

8.2. Hotelling-T 2

99

¯ and S are the unbiased estimate for µ and Σ. This is needed where x to test the hypothesis H0 : µ = µ0 against all alternatives or to build a conﬁdence ellipsoid for µ. In fact, the following proposition shows that the Hotelling-T 2 statistic is a monotone function of the likelihood ratio test (LRT) statistic. As usual, let V=

n

¯ )(xi − x ¯ ) (xi − x

i=1

be the matrix of sums of squares and cross-products. Proposition 8.1 The likelihood ratio statistic for H0 : µ = µ0 against H1 : µ = µ0 is −n/2 1 2 T . Λ= 1+ (n − 1) ˆ = 1 V. However, ˆ =x ¯ and Σ Proof. The unrestricted MLE of (Σ, µ) is µ n the MLE of Σ under H1 is obtained from (7.2) by minimizing 1 VΣ−1 + (¯ x − µ0 ) Σ−1 (¯ x − µ0 ) n 1 x − µ0 ) ]Σ−1 . = ln |Σ| + tr [V + n(¯ x − µ0 )(¯ n Using the same technique as on page 86 we ﬁnd ln |Σ| + tr

ˆ ˆ = Σ =

1 x − µ0 ) ] [V + n(¯ x − µ0 )(¯ n n 1 (xi − µ0 )(xi − µ0 ) . n i=1

Thus, with (7.1), the LRT becomes Λ = = =

ˆ ˆ µ0 ) L(Σ, ˆ µ) ˆ L(Σ, ˆ ˆ −n/2 exp(− 1 np) |Σ| 2 ˆ −n/2 exp(− 12 np) |Σ|

x − µ0 )(¯ x − µ0 ) ]|−n/2 | n1 [V + n(¯ 1 | n V|−n/2

= |I + nV−1 (¯ x − µ0 )(¯ x − µ0 ) |−n/2 −n/2 1 T2 = 1+ , (n − 1) where the last equality made use of Problem 1.8.3.

2

The distribution of T 2 is a direct consequence of the following proposition.

8. Tests on mean and variance

Proposition 8.2 If z ∼ Np (δ, I), W ∼ Wp (m), m ≥ p, and z

|=

100

W, then

z W−1 z ∼ Fc (p, m − p + 1; δ δ/2). Proof. Using an orthogonal transformation H = (z/|z|, Γ) ∈ Op , we get immediately z W−1 z

= (Hz) (HWH )−1 (Hz) = |z|2 e1 V−1 e1 , |=

where V = HWH . Since the conditional distribution V | z ∼ Wp (m) does z. Letting not depend on z, then V ∼ Wp (m), unconditionally, and V 11 v v21 , V−1 = 21 v V22 z W−1 z = |z|2 v 11 = z z/v11.2 , where the last equality made use of Problem 1.8.1. The conclusion follows since z z ∼ χ2p (δ δ/2) and, by Proposition 7.9, v11.2 ∼ χ2m−p+1 . 2 As a corollary, we obtain the distribution of Hotelling-T 2 . Corollary 8.1 The non-null distribution of T 2 for n ≥ p + 1 is T 2 /(n − 1) ∼ Fc (p, n − p; δ), with δ = n(µ − µ0 ) Σ−1 (µ − µ0 )/2. d

Proof. In terms of the sample matrix, as on page 86, X = ZA +1µ , where d Z ∼ Npn (0, In ⊗ Ip ) and Σ = AA , and, thus, (¯ x, Sx ) = (A¯ z + µ, ASz A ). Therefore, T2 (n − 1)

x − µ0 ) = n(¯ x − µ0 ) [(n − 1)S]−1 (¯

= n (A¯ z + µ − µ0 ) [(n − 1)ASz A ]−1 (A¯ z + µ − µ0 ) ¯ + A−1 (µ − µ0 ) . ¯ + A−1 (µ − µ0 ) [(n − 1)Sz ]−1 z = n z d

The proof follows from Proposition 8.2, as ¯ + A−1 (µ − µ0 ) ∼ Np n1/2 A−1 (µ − µ0 ), I n1/2 z (n − 1)Sz

∼ Wp (n − 1) 2

are independent.

Example 8.1 The power function of an α signiﬁcance level Hotelling-T 2 test may now be evaluated as a function of δ = n(µ − µ0 ) Σ−1 (µ − µ0 )/2 in the following manner: β = P (Fc (p, n − p; δ) ≥ tα ) ,

8.2. Hotelling-T 2

101

beta 1 0.8 0.6 0.4 0.2 2

4

6

8

10

delta

Figure 8.1. Power function of Hotelling-T 2 when p = 3 and n = 40 at a level of signiﬁcance α = 0.05.

where tα = [p/(n − p)]Fα (p, n − p) is the critical point. Proposition 4.7 and Problem 3.5.6 yields β

∞

δk P (Fc (p + 2k, n − p) ≥ tα ) k! k=0 ∞ −1 F (p+2k)/2−1 δk ∞ 1 = e−δ B 2 (p + 2k), 12 (n − p) dF. k! tα (1 + F )(n+2k)/2 k=0

=

e−δ

Numerical evaluation in Mathematica for p = 3, n = 40 and α = 0.05 produced the plot in Figure 8.1. The robustness of Hotelling-T 2 is easily established. Without normality, assuming x1 , . . . , xn are i.i.d. x, E x = µ0 and var x = Σ, the asymptotic d distributions in Section 6.3 gave T 2 → χ2p . This asymptotic distribution is the same regardless of the underlying distribution of x. A diﬀerent situation arises in presence of “contamination.” Assume the simple situation x1 , . . . , xn−1 are i.i.d. Np (µ0 , Σ), but there is one (or more) contaminated observation xn ∼ Np (µ0 +γ, Σ). We assume Σ known for the x − µ0 ) ∼ Np (γ/n1/2 , Σ) sake of simplicity. It is easily checked that n1/2 (¯ and, thus, from Corollary 5.1, T 2 = n(¯ x − µ0 ) Σ−1 (¯ x − µ0 ) ∼ χ2p (γ Σ−1 γ/2n). Since P (χ2p (δ) ≥ c) is monotone increasing in δ [Ghosh (1970), p. 302], all other parameters being ﬁxed, it follows that T 2 will reject H0 : µ = µ0 with probability converging to 1 as |γ| → ∞ (for ﬁxed n) even though all observations, but one, have mean µ0 . A procedure which is insensitive to

102

8. Tests on mean and variance

contamination of the data consists of building an Hotelling-T 2 test T 2 = n(µn − µ0 )Σ−1 n (µn − µ0 ) from a robust estimate (Σn , µn ) such as an M-estimate or S-estimate. These truly robust tests are studied in Chapter 13. We end this section with a discussion of invariant tests on the mean vector. Consider the canonical problem of testing H0 : µ = 0 against H1 : µ = 0. The group of transformations Gp acts on the observations as xi → Axi , where A ∈ Gp . This transformation induces the following transformations on the minimal suﬃcient statistic (¯ x, S) and parameters, ¯ → A¯ x x, S → ASA , and µ → Aµ, Σ → AΣA . Note that the hypotheses are preserved because µ = 0 iﬀ Aµ = 0, for any A ∈ Gp . We deﬁne a test function f (¯ x, S) to be invariant iﬀ it yields the same value on the original as on the transformed data, i.e., f (y, W) = f (Ay, AWA ), ∀A ∈ Gp , ∀(y, W) ∈ Rp × Pp . This has important implications. First, the choice A = S−1/2 yields ¯ , I). f (¯ x, S) = f (S−1/2 x Now, there exists an orthogonal transformation H ∈ Op (v. Problem 1.8.14) ¯ = (¯ ¯ )1/2 e1 . Choosing now A = H, we ﬁnd such that HS−1/2 x x S−1 x f (¯ x, S)

¯ , HH ) = f (HS−1/2 x ¯ )1/2 e1 , I), = f ((¯ x S−1 x

which shows that any invariant test function depends on the data only ¯. through T 2 = n¯ x S−1 x Second, selecting A = Σ−1/2 gives ¯ , Σ−1/2 SΣ−1/2 ), f (¯ x, S) = f (Σ−1/2 x where ¯ ∼ Np (Σ−1/2 µ, n−1 I) Σ−1/2 x (n − 1)Σ−1/2 SΣ−1/2

∼ Wp (n − 1).

Using the same argument, there exists an orthogonal transformation H ∈ Op such that HΣ−1/2 µ = (µ Σ−1 µ)1/2 e1 . Choosing this time A = H, we ﬁnd ¯ , HΣ−1/2 SΣ−1/2 H ), f (¯ x, S) = f (HΣ−1/2 x where ¯ ∼ Np ((µ Σ−1 µ)1/2 e1 , n−1 I) HΣ−1/2 x (n − 1)HΣ−1/2 SΣ−1/2 H

∼ Wp (n − 1),

and, thus, the power function of any invariant test depends on (µ, Σ) only through the parameter function δ = nµ Σ−1 µ/2. These results are summarized.

8.2. Hotelling-T 2

103

Proposition 8.3 For testing H0 : µ = 0 against H1 : µ = 0, any invariant test with respect to the group Gp depends on the minimal suﬃcient ¯ . Moreover, the power function x S−1 x statistic (¯ x, S) only through T 2 = n¯ of any invariant test depends on (µ, Σ) only through the parameter function δ = nµ Σ−1 µ/2. The LRT is obviously invariant. In the class of invariant tests, it is possible to show that the Hotelling-T 2 is uniformly most powerful. We say that T 2 is the UMPI test. Proposition 8.4 For testing H0 : µ = 0 against H1 : µ = 0, the ¯ , is UMPI. x S−1 x Hotelling-T 2 test, T 2 = n¯ Proof. It has already been established that T 2 is invariant and that all invariant tests, depending on (¯ x, S), are a function of T 2 . The problem thus reduces to ﬁnding the UMP test for H0 : δ = 0 based on one observation from x ≡ T 2 /(n − 1) ∼ Fc (p, n − p; δ), where δ = nµ Σ−1 µ/2. The density of x was given in Problem 4.6.3:

∞ Γ 12 (n + 2k) x(p+2k)/2−1 δk

1 1 e−δ , x > 0. f (x; δ) = k! Γ 2 (p + 2k) Γ 2 (n − p) (1 + x)(n+2k)/2 k=0 From the Neyman-Pearson lemma, the most powerful test for H0 : δ = 0 rejects H0 for large values of the ratio

1 k ∞ k x f (x; δ) −δ δ Γ 2 (n + 2k) e ≥ c2 . = c1 f (x; 0) k! Γ 12 (p + 2k) (1 + x) k=0 Since this ratio is monotone increasing in x, this is equivalent to rejecting H0 for large values of x, x ≥ c3 . This rejection region does not depend on δ and, thus, the test is uniformly most powerful. 2 An asymptotic expansion of the distribution function of T 2 was obtained whose ﬁrst term is χ2p under the elliptical distribution [Iwashita (1997)] and for general non-normality [Fujikoshi (1997), Kano (1995)]. Improvement to the chi-square approximation by monotone transformation of T 2 is also possible [Fujisawa (1997)]. For a modiﬁcation to T 2 with the same chisquare asymptotic distribution but in the case of inﬁnite second moment, refer to Sepanski (1994). Kudˆ o (1963) was the ﬁrst to propose a multivariate analogue, when Σ is known, to the one-sided t-test. The multivariate problem is to test the null hypothesis, H0 : µ = 0, against the one-sided alternative hypothesis, H1 : µ ≥ 0, where µ ≥ 0 is interpreted componentwise. It can be stated even more generally in terms of cone. The LRT for the one-sided problem, with unknown Σ, was obtained by Perlman (1969). Tang (1994, 1996) discussed unbiasedness and invariance of tests in the one-sided multivariate problem. Silvapulle (1995) derived the null distribution of a Hotelling-T 2

104

8. Tests on mean and variance

type statistic. There is a conditional test by Fraser, Guttman and Srivastava (1991); v. also Wang and McDermott (1998a, 1998b).

8.3 Simultaneous conﬁdence intervals on means Let x1 , . . . , xn be i.i.d. Np (µ, Σ). For any preassigned level β = 1−α, deﬁne the quantile Fα (p1 , p2 ) by the equation P (F (p1 , p2 ) ≥ Fα (p1 , p2 )) = α. By applying Hotelling’s result, we have exactly (n − 1)p Fα (p, n − p) = β, x − µ) ≤ P n(¯ x − µ) S−1 (¯ (n − p) whereby we see that in β × 100% of such experiments, the “true” µ lies in the random ellipsoid ¯ ) S−1 (µ − x ¯ ) ≤ cα }, {µ ∈ Rp : n(µ − x where we have simply let cα =

(n − 1)p Fα (p, n − p). (n − p)

We are β × 100% “conﬁdent” that our particular observed ellipsoid, ¯ ) S−1 (µ − x ¯ ) ≤ cα }, CR(µ; β) = {µ ∈ Rp : n(µ − x contains µ since P (µ ∈ CR(µ; β)) = β. Sequential ﬁxed-size conﬁdence regions for the mean vector were investigated by Srivastava (1967) and Datta and Mukhopadhyay (1997).

8.3.1

Linear hypotheses

In many experiments, one simply wishes to compare the various components of µ to each other. For instance, one may ask: “Is µ1 equal to µ3 ?” “Is the diﬀerence between µ2 and the average of µ1 and µ3 equal to 3.1?” Answering the ﬁrst question amounts to testing the hypothesis H0 : µ1 − µ3 = 0, while the second question is equivalent to testing the hypothesis H0 : 2µ2 − (µ1 + µ3 ) = 6.2. Questions like these are said to be linear in µ, and in the general case, there would be a certain speciﬁed vector a ∈ Rp and constant c ∈ R for which we would wish to test H0 : a µ = c.

8.3. Simultaneous conﬁdence intervals on means

105

Given x1 , . . . , xn i.i.d. Np (µ, Σ), if we let yi = a xi , i = 1, . . . , n, then clearly y1 , . . . , yn are i.i.d. N1 (a µ, a Σa). Obviously, one may apply the univariate results directly to the y data so that √ d H0 is correct iﬀ n(¯ y − c)/sy = tn−1 and, of course,

sy sy CI(a µ; β) = y¯ − √ tα/2,n−1 , y¯ + √ tα/2,n−1 . n n

¯ and s2y = a Sa. One should notice that, conveniently, y¯ = a x Realistically, one would have more than one such question to consider, and so that there would be r speciﬁed vectors ai ∈ Rp , i = 1, . . . , r, with corresponding constants ci ∈ R, i = 1, . . . , r, for which we would wish simultaneously to test H0 : a1 µ = c1 , . . . , ar µ = cr . This is clearly equivalent to testing H0 : A µ = c, where A = (a1 , . . . , ar ) and c = (c1 , . . . , cr ) . Letting yi = A xi , i = 1, . . . , n, then, clearly, y1 , . . . , yn are i.i.d. Nr (A µ, A ΣA). Under the assumption that the vectors a1 , . . . , ar are linearly independent, one may apply the multivariate results above directly to the y data so that y − c) S−1 y − c) = H0 is correct iﬀ n(¯ y (¯ d

(n − 1)r F (r, n − r) n−r

and a conﬁdence ellipsoid for ν = A µ is ¯ ) S−1 ¯ ) ≤ kα }, CR(ν; β) = {ν ∈ Rr : n(ν − y y (ν − y where kα =

(n − 1)r Fα (r, n − r). (n − r)

¯ = A x ¯ and Sy = A SA. Notice that y For obvious pragmatic reasons, one might in practice wish to have individual conﬁdence intervals for each component νi , i = 1, . . . , r. Thus, we would like to specify r intervals for these quantities in which we are simultaneously conﬁdent. The following lemma is needed. Lemma 8.1 Assume S ∈ Pp and A = (a1 , . . . , ar ) ∈ Rpr is of rank r. Then (ai x)2 ≤ x A(A SA)−1 A x ≤ x S−1 x, ∀x ∈ Rp . ai Sai

106

8. Tests on mean and variance

Proof. Using Rayleigh’s quotient (v. Problem 1.8.12),

x A(A SA)−1 A x = λ1 A(A SA)−1 A S = 1 −1 xS x x =0 sup

and the right inequality follows. Let y = A x = (y1 , y2 ) and B = A SA ∈ Pr partitioned as b11 b21 B= b21 B22 with inverse (v. Problem 1.8.1) −1 −1 b11 + b−2 −1 11 b21 B22.1 b21 B = −1 −1 −b11 B22.1 b21

−1 −b−1 11 b21 B22.1 −1 B22.1

.

Then, y B−1 y =

y12 y2 (a x)2 −1/2 −1/2 + ||y1 b11 B22.1 b21 − B22.1 y2 ||2 ≥ 1 = 1 , b11 b11 a1 Sa1 2

which is the left inequality. Since by simple algebra in Lemma 8.1, we have the inequalities (¯ yi − νi )2 ≤ (¯ y − ν) S−1 y − ν) ≤ (¯ x − µ) S−1 (¯ x − µ), y (¯ ai Sai we ﬁnd that

P

| y¯i − νi | 1/2 n ≤ kα , i = 1, . . . , r (ai Sai )1/2

≥ P n(¯ y − ν) S−1 y − ν) ≤ kα = β. y (¯ 1/2

Therefore, we are at least β × 100% conﬁdent in simultaneously presenting the r observed “Roy-Bose” intervals 1/2

1/2

kα kα (a Sai )1/2 ≤ νi ≤ y¯i + 1/2 (ai Sai )1/2 , i = 1, . . . , r. (8.2) n1/2 i n One should note that kα ≤ cα (why?), so we do somewhat better using kα . The constant cα , however, allows all possible linear combinations since y¯i −

¯ − a µ)2 (a x = (¯ x − µ) S−1 (¯ x − µ). a Sa a =0 sup

Therefore, we are at least β × 100% conﬁdent in simultaneously presenting all of the observed “Scheﬀ´e” intervals 1/2

1/2

cα cα ¯ + 1/2 (a Sa)1/2 , ∀a ∈ Rp . (a Sa)1/2 ≤ a µ ≤ a x (8.3) n1/2 n Although the “Scheﬀ´e” intervals are wider, they can be useful in making a great number of unplanned comparisons between means. ¯− a x

8.3. Simultaneous conﬁdence intervals on means

107

Actually, if the number, r, of questions that one asks is very small, we can sometimes even improve on kα . Let Ti = n1/2

(¯ y i − νi ) , i = 1, . . . , r. (ai Sai )1/2

d

Note Ti = tn−1 , i = 1, . . . , r, but they are not independent. Then, if we deﬁne the event Ai = {|Ti | ≤ t0 }, P (|Ti | ≤ t0 , i = 1, . . . , r)

= P (∩ri=1 Ai ) = 1 − P (∪ri=1 Aci ) r P (Aci ) ≥ 1−

(8.4)

i=1

=

1 − r P (|tn−1 | > t0 ) .

The inequality (8.4) is the Bonferroni inequality. If we deliberately equate the ﬁnal term to β = 1 − α and then solve for t0 , we ﬁnd that α P (tn−1 > t0 ) = , 2r or, equivalently, t0 = tα/2r,n−1 . Therefore, one can see that if we let bα = t2α/2r,n−1 , we will still be at least β × 100% conﬁdent if, instead of the “Roy-Bose” intervals (8.2), we present the r “Bonferroni” intervals 1/2

1/2

bα bα (ai Sai )1/2 ≤ νi ≤ y¯i + 1/2 (ai Sai )1/2 , i = 1, . . . , r. 1/2 n n Note that the relative length of “Roy-Bose” to “Bonferroni” is obviously bα /kα , and in a particular application, one would use the method with the shorter intervals. For non-normal data x1 , . . . , xn i.i.d. x with E x = µ and var x = Σ, large sample “Roy-Bose” and “Scheﬀ´e” simultaneous conﬁdence intervals can be constructed similarly (v. Problems 8.9.5 and 8.9.6) by appealing ﬁrst to the central limit theorem. y¯i −

8.3.2

Nonlinear hypotheses

In certain experiments, one might wish to compare the various components of µ to each other in ways that are plainly nonlinear. For instance, one may ask: “Is µ1 equal to µ23 ?” “Is the diﬀerence between µ2 and the product of µ1 and µ3 equal to 5.7?” The ﬁrst question corresponds to the hypothesis H0 : µ1 − µ23 = 0

108

8. Tests on mean and variance

and the second to H0 : µ2 − µ1 µ3 = 5.7. In each case, there is a certain function g : Rp → R and constant c ∈ R for which we are entertaining the hypothesis H0 : g(µ) = c. If we actually had r such hypotheses to consider in the same experiment, then there would, of course, be r speciﬁed real-valued functions gi , i = 1, . . . , r, with constants ci ∈ R, i = 1, . . . , r, for which we would wish simultaneously to test H0 : g1 (µ) = c1 , . . . , gr (µ) = cr . By letting g : R → Rr deﬁned by g(x) = (g1 (x), . . . , gr (x)) be continuously diﬀerentiable at µ, and c ∈ Rr , this is simply equivalent to p

H0 : g(µ) = c. The results in this section are asymptotic, so we assume possibly nonnormal data x1 , . . . , xn i.i.d. x with E x = µ and var x = Σ. One may not apply the speciﬁc results in Section 8.3.1 directly to this situation, but one may yet apply the same essential logic as formulated in that section √ d by appealing to the central limit theorem, n(¯ x − µ) → Np (0, Σ), and the √ d x) − g(µ)) → Nr (0, Σg ), where delta method in Proposition 6.2, n (g(¯ x), ν = g(µ), and Σg = [Dg(µ)]Σ[Dg(µ)] . This time, if we let y = g(¯ x)]S[Dg(¯ x)] then, as on page 79, Sg = [Dg(¯ 2 n(y − ν) S−1 g (y − ν) → χr . d

Thus, with dα = χ2α,r , the conﬁdence ellipsoid for ν, CR(ν; β) = {ν ∈ Rr : n(ν − y) S−1 g (ν − y) ≤ dα }, has an asymptotic coverage probability of β, i.e., P (ν ∈ CR(ν; β)) → β as n → ∞. To have individual conﬁdence intervals on each component νi , i = 1, . . . , r, we have the inequality in Lemma 8.1: (yi − νi )2 ≤ (y − ν) S−1 g (y − ν). Sg,ii From this purely algebraic fact, it follows that 1/2 | yi − νi | 1/2 ≤ dα , i = 1, . . . , r ≥ β. lim P n 1/2 n→∞ Sg,ii

8.4. Multiple correlation

109

Thus, asymptotically, we are at least β × 100% conﬁdent in simultaneously presenting the r observed intervals 1/2

yi −

1/2

dα dα 1/2 1/2 Sg,ii ≤ νi ≤ yi + 1/2 Sg,ii , i = 1, . . . , r. 1/2 n n

If r is quite small, one might try to improve on dα using a Bonferroni approach. The construction of simultaneous conﬁdence intervals on functions φ(Σ) is treated quite generally in D¨ umbgen (1998). Asymptotic considerations for the Wishart model show that the resulting conﬁdence bounds are substantially smaller than those obtained by inverting likelihood ratio tests.

8.4 Multiple correlation

|=

The multiple correlation coeﬃcient R is the maximum correlation possible between a variable x1 and a linear combination, t x2 , of a vector x2 . Not surprisingly, with underlying normality, the likelihood ratio test of H0 : ˆ x2 will be a function of the sample multiple correlation coeﬃcient R. x1 Assume x1 ∈ R and x2 ∈ Rp−1 have a joint normal distribution, σ11 σ 21 x1 ∼ Np 0, , x2 σ 21 Σ22 where Σ22 = A2 > 0. We have set the mean to 0 without any loss of generality. Since the simple correlation coeﬃcient is invariant to rescaling of each variable, we can assume at the outset that var t x2 = t Σ22 t = 1 and solve max cor(x1 , t x2 ).

t Σ22 t=1

For any t such that t Σ22 t = 1, cor2 (x1 , t x2 ) = (σ 21 t)2 /σ11 = A−1 σ 21 , At2 /σ11 ≤ σ 21 Σ−1 22 σ 21 /σ11 .

|=

The last inequality follows from the Cauchy-Schwarz inequality given in Proposition 1.1. It is an equality iﬀ At ∝ A−1 σ 21 , or, equivalently, 2 t ∝ Σ−1 22 σ 21 . The maximum correlation possible is R ≥ 0, where R = −1 σ 21 Σ22 σ 21 /σ11 , and is called the multiple correlation coeﬃcient between x1 and x2 . It should be noted immediately that the maximum correlation is achieved by t x2 = E (x1 | x2 ) = σ 21 Σ−1 22 x2 , i.e., by the conditional mean of x1 given x2 . In order to test H0 : R = 0 (equivalently, H0 : σ 21 = 0 or x2 ), the sample variance, based on a random sample of size n, H0 : x1 v11 v21 (n − 1)S ≡ V = , v21 V22

110

8. Tests on mean and variance

is partitioned and is distributed as V ∼ Wp (n − 1, Σ). In the obvious ˆ ≥ 0 is called ˆ 2 = v V−1 v21 /v11 and R manner, the sample version is R 21 22 the sample multiple correlation coeﬃcient. Proposition 8.5 The likelihood ratio test Λ rejects H0 for small values of ˆ 2 )n/2 . Λ = (1 − R Proof. Based on the likelihood (7.1) from x1 , . . . , xn i.i.d. Np (µ, Σ), Σ > ˆ =x ¯ . Without constraints, the MLE of Σ is 0, the MLE of µ is always µ ˆ = 1 V, but when σ 21 = 0, the constrained MLE becomes Σ n 1 v11 0 ˆ ˆ . Σ= 0 V22 n Thus, ˆ ˆ µ) ˆ L(Σ, Λ= ˆ ˆ L(Σ, µ)

−1

ˆ ˆ ˆ ) ˆ −n/2 etr(− 1 VΣ |Σ| 2 = ˆ −n/2 etr(− 1 VΣ ˆ −1 ) |Σ| 2 −n/2 exp(− 12 np) v11 |V22 | = |V| exp(− 12 np) n/2 v11.2 ˆ 2 )n/2 , = = (1 − R v11

where the last equality made use of |V| = v11.2 |V22 |.

2

ˆ 2 in which negative binomial Of greater interest is the distribution of R probabilities intervene. The reader should recall at this point that a negative binomial variable represents the number of failures, k, before the rth success in a sequence of independent bernoulli trials. Deﬁnition 8.1 Negative binomial: x ∼ nb(r, p), r > 0 and 0 ≤ p ≤ 1, iﬀ the probability function of x is given by r+k−1 r pk = P (x = k) = p (1 − p)k , k = 0, 1, . . . . k In Deﬁnition 8.1, r need not be an integer. In that case, the combination factor is calculated via the gamma function: r+k−1 Γ(r + k) (r)k = , = k k!Γ(r) k! where (r)0 = 1 and (r)k = r(r + 1) · · · (r + k − 1) for k = 1, 2, . . .. Recall that Fc (s1 , s2 ) denotes the canonical Fc distribution (v. Deﬁnition 3.7). Proposition 8.6 ˆ2 R P ≤t = ˆ2 1−R

∞ k=0

pk · P (Fc (p − 1 + 2k, n − p) ≤ t) ,

8.4. Multiple correlation

ˆ2 ≤ t P R

=

∞

111

pk · P beta 12 (p − 1 + 2k); 12 (n − p) ≤ t ,

k=0

|=

where pk are the negative binomial probabilities

1 2 (n − 1) k (1 − R2 )(n−1)/2 R2k , k = 0, 1, . . . . pk = k! ˆ 2 ), then R ˆ 2 /(1 − R ˆ 2 ) = v V−1 v21 /v11.2 . Proof. Since v11.2 = v11 (1 − R 21 22 2 (v21 , V22 ) and v11.2 ∼ σ11.2 χn−p . With the help of Proposition 7.9, v11.2 We also have v21 | V22 ∼ Np−1 (V22 Σ−1 22 σ 21 , σ11.2 V22 ) from which −1/2

−1/2

σ11.2 V22

−1/2

1/2

v21 | V22 ∼ Np−1 (σ11.2 V22 Σ−1 22 σ 21 , I)

−1 V22 v21 | V22 ∼ σ11.2 χ2p−1 (δ), where and, therefore, v21 −1 δ = σ 21 Σ−1 22 V22 Σ22 σ 21 /(2σ11.2 ).

Hence, conditional on V22 , ˆ 2 ) ∼ Fc (p − 1, n − p; δ). ˆ 2 /(1 − R R Using Proposition 4.7, ∞ ˆ2 R δk ≤ t | V22 = e−δ P (Fc (p − 1 + 2k, n − p) ≤ t) . P ˆ2 k! 1−R k=0

To obtain the unconditional distribution, take expectations on both sides with respect to the distribution of V22 . First, we need the distribution of δ. Since V22 ∼ Wp−1 (n − 1, Σ22 ), then R2 R2 d 1 2 1 δ∼ χ = G 2 (n − 1), . (1 − R2 ) 2 n−1 (1 − R2 ) The expectation computation is immediate (v. Problem 8.9.10) if we use a result well known in bayesian inference [Johnson et al. (1992), p. 204] that if K given δ is Poisson(δ) and δ ∼ G(p, θ), then the marginal of K is negative binomial, K ∼ nb(p, (1 + θ)−1 ). Hence, δk , k! completing the proof of the ﬁrst result. The second result follows with the obvious monotone transformation. 2 pk = P (K = k) = E P (K = k | δ) = E e−δ

ˆ 2 ) is distributed as a negative binomial mixture of canonical ˆ 2 /(1−R Thus, R ˆ 2 is a negative binomial mixture of beta Fc distributions, whereas that of R ˆ distributions. The moments of R (v. Problem 8.9.9) follow directly from the later characterization. The null distribution is just a special case.

112

8. Tests on mean and variance

ˆ 2 /(1 − R ˆ 2 ) ∼ Fc (p − 1, n − p). Proposition 8.7 Assuming R = 0, R The exact distribution of the simple correlation coeﬃcient, introduced earlier in Section 5.6.3, when ρ = 0 is just another special case when p = 2. The invariance of the multiple correlation coeﬃcient is discussed in Problem 8.9.13. Proposition 8.8 For testing H0 : R = 0 against H1 : R > 0, the test ˆ is UMPI. which rejects for large values of R ˆ is clearly invariant and it was established in ProbProof. The statistic R ˆ lem 8.9.13 that all invariant tests, depending on (¯ x, V), are a function of R. The problem thus reduces to ﬁnding the UMP test based on one observation ˆ The density of x ≡ R ˆ 2 follows from Proposition 8.6, from R. f (x; R2 ) =

∞ k=0

pk

B( 12 (p

1 1 1 x 2 (p−1+2k)−1 (1 − x) 2 (n−p)−1 , 1 − 1 + 2k), 2 (n − p))

0 < x < 1. From the Neyman-Pearson lemma, the most powerful test rejects H0 for large values of the ratio

∞ Γ 12 (n − 1 + 2k) k f (x; R2 ) x ≥ c2 . = c1 pk 1 f (x; 0) Γ 2 (p − 1 + 2k) k=0 Since this ratio is monotone increasing in x this is equivalent to rejecting H0 for large values of x, x ≥ c3 . This rejection region does not depend on R and, thus, the test is uniformly most powerful. 2 Example 8.2 The power function of the likelihood ratio test for H0 : R = 0 may be evaluated with Proposition 8.6: 1 ∞

−1 pk β = B 12 (p − 1 + 2k), 12 (n − p) tα k=0 (p−1+2k)/2−1

·x (1 − x)(n−p)/2−1 dx,

1 where tα = betaα 2 (p − 1); 12 (n − p) is the critical point. A numerical evaluation in Mathematica of β for p = 3 and n = 20 at the signiﬁcance level α = 0.05 gave the plot in Figure 8.2. For large samples, the asymptotic distribution provides a simpler disˆ 2 is asymptotically tribution. By the delta method in Proposition 6.2, R normal since it is a function of the sample variance S, which is itself asymptotically normal. However, rather than calculating the derivatives, it is somewhat easier to use op (n−1/2 ) asymptotic expansions. This technique is illustrated in the following proof. ˆ2, Proposition 8.9 The null and alternative asymptotic distributions of R when sampling from a multivariate normal distribution, are given by

8.4. Multiple correlation

113

beta 1 0.8 0.6 0.4 0.2

0.2

0.4

0.6

0.8

R

Figure 8.2. Power function of the likelihood ratio test for H0 : R = 0 when p = 3, and n = 20 at a level of signiﬁcance α = 0.05.

d ˆ 2 − R2 ) → (i) n1/2 (R N 0, 4R2 (1 − R2 )2 , d

ˆ 2 → χ2 . (ii) If R = 0, then nR p−1 Proof. By invariance arguments (v. Problem 8.9.13), assume without loss of generality, 1 Re1 , Σ= Re1 Ip−1 where e1 = (1, 0, . . . , 0) ∈ Rp−1 . Since n1/2 (S − Σ) → Z = d

z11 z21

z21 Z22

,

where Z ∼ Npp (0, (I + Kp )(Σ ⊗ Σ)), then we can write the op (n−1/2 ) expansions 1 + n−1/2 z11 + op (n−1/2 ),

s11

=

s21

= Re1 + n−1/2 z21 + op (n−1/2 ),

S22

= I + n−1/2 Z22 + op (n−1/2 ), p

where op (n−1/2 ) is such that n1/2 · op (n−1/2 ) → 0 [Serﬂing (1980), p. 9]. Straightforward algebra, with the aid of Problem 1.8.15, then gives s21 S−1 22 s21 s11 = [1 + n−1/2 z11 + op (n−1/2 )]−1 · [Re1 + n−1/2 z21 + op (n−1/2 )]

114

8. Tests on mean and variance

·[I + n−1/2 Z22 + op (n−1/2 )]−1 · [Re1 + n−1/2 z21 + op (n−1/2 )] =

[1 − n−1/2 z11 + op (n−1/2 )] · [Re1 + n−1/2 z21 + op (n−1/2 )] ·[I − n−1/2 Z22 + op (n−1/2 )] · [Re1 + n−1/2 z21 + op (n−1/2 )]

= R2 + 2n−1/2 Rz21 − n−1/2 R2 z22 − n−1/2 R2 z11 + op (n−1/2 ). d ˆ 2 −R2 ) → Thus, n1/2 (R 2Rz21 −R2 z22 −R2 z11 , but since (v. equation (6.1)) (z21 , z11 , z22 ) ∼ N3 (0, Ω), where 1 + R2 2R 2R 2 2R2 , Ω = 2R 2 2R 2R 2

the linear combination with a = (2R, −R2 , −R2 ) yields 2Rz21 − R2 z22 − R2 z11 ∼ N (0, a Ωa), whereby a direct evaluation provides a Ωa = 4R2 (1−R2 )2 . This proves (i). d

To prove (ii), note that when R = 0, n1/2 s21 → z21 , where z21 ∼ Np−1 (0, I). p

p

d −1/2 −1/2 S22 s21 →

However, since S22 → I and s11 → 1, then n1/2 s11 d d ˆ2 → nR |z21 |2 = χ2p−1 .

z21 and 2

ˆ and of Fisher’s As a corollary, we get the asymptotic distribution of R z-transform. ˆ and of its Fisher’s z Corollary 8.2 The asymptotic distributions of R transform, when sampling from a multivariate normal distribution, are given by

d ˆ − R) → N 0, (1 − R2 )2 , (i) n1/2 (R d ˆ − tanh−1 (R) → (ii) n1/2 tanh−1 (R) N (0, 1). Proof. It follows directly from the delta method applied to the square root 2 transformation and to the tanh−1 transformation. More general results on the asymptotic distributions of correlation coefﬁcients obtained from any asymptotically normal equivariant estimate of variance, not necessarily S, will be given in Chapter 13 for a sample from an elliptical distribution.

8.4.1

Asymptotic moments

ˆ 2 in Proposition 8.6 can be used The mixture beta characterization of R ˆ 2 in terms of those of beta to obtain immediately the exact moments of R distributions (v. Problem 8.9.9). Simple approximations for large n are, however, possible as is now shown.

8.4. Multiple correlation

115

Using Proposition 8.6, we have ∞

2 ˆ pk · P beta 12 (n − p); 12 (p − 1 + 2k) ≤ t . P 1−R ≤t = k=0

From the moments of x ∼ beta(a, b) given by E xh =

(a)h Γ(a + b)Γ(a + h) = , h = 1, 2, . . . , Γ(a)Γ(a + b + h) (a + b)h

we can write ˆ2 h

E (1 − R ) =

∞

1

2 (n

k=0

1 − 1) k 2 (n − p) h 2 (n−1)/2 2k . (1 − R ) R 1 k! 2 (n − 1 + 2k) h

The hypergeometric function 2 F1 (a, b; c; z)

≡

∞ (a)k (b)k z k (c)k k!

k=0

after some simple algebra allows to write

1 2 (n − p) h 2 h ˆ E (1 − R ) = 1 2 (n − 1) h ·(1 − R2 )(n−1)/2 2 F1 ( 12 (n − 1), 12 (n − 1); 12 (n − 1) + h; R2 ). Then, upon using Kummer’s formula [Erd´elyi et al. (1953), p. 105] 2 F1 (a, b; c; z)

we ﬁnally ﬁnd ˆ2 h

E (1 − R ) =

1

(n

21 2 (n

= (1 − z)(c−a−b) 2 F1 (c − a, c − b; c; z), − p) h (1 − R2 )h 2 F1 (h, h; 21 (n − 1) + h; R2 ). − 1) h

For h = 1, we then obtain ˆ2) E (1 − R

= =

(n − p) (1 − R2 ) 2 F1 (1, 1; 12 (n + 1); R2 ) (n − 1) 2R2 (n − p) 2 −2 (1 − R ) 1 + + O(n ) (n − 1) (n + 1)

and ˆ 2 = R2 + (p − 1) (1 − R2 ) − 2 (n − p) R2 (1 − R2 ) + O(n−2 ). ER (n − 1) (n2 − 1) ˆ 2 is biased, as it overestimates R2 . The MVUE This expression shows that R of R2 [Olkin and Pratt (1958)] is ˆ 2 ) 2 F1 (1, 1; 1 (n − p + 2); 1 − R ˆ 2 ). ˆ 2 ) = 1 − (n − 3) (1 − R U (R 2 (n − p)

116

8. Tests on mean and variance

ˆ is close to The MVUE has the drawback of taking negative values when R 0. In fact, using the relation [Erd´elyi et al. (1953), p. 61] 2 F1 (a, b; c; 1)

=

Γ(c) Γ(c − a − b) , Γ(c − a) Γ(c − b)

it is easily established that U (0) = −(p−1)/(n−p−2) and, of course, U (1) = 1. A similar expansion for h = 2 can be done to obtain an asymptotic ˆ2. expansion for var R

8.5 Partial correlation Assume two subsets of variables x1 ∈ Rp1 and x2 ∈ Rp2 have a joint normal distribution, µ1 x1 Σ11 Σ12 ∼ Np , . x2 µ2 Σ21 Σ22 The partial correlation coeﬃcient between variables xi and xj , in the subset x1 , is just the ordinary simple correlation ρ between xi and xj but with the variables in the subset x2 held ﬁxed. This will be denoted by ρij|x2 . It can be expressed in terms of Σ if one recalls the conditional normal of Section 5.5: x1 | x2 ∼ Np1 (µ1 + Σ12 Σ−1 22 (x2 − µ2 ), Σ11.2 ). Writing Σ11.2 = (σij|x2 ), where σij|x2 denotes the (i, j) element of Σ11.2 , then σij|x2 . ρij|x2 = 1/2 1/2 σii|x2 σjj|x2 Using Proposition 7.9, we already know that since (n − 1)S = V ∼ Wp (n − 1, Σ), then V11.2 ∼ Wp1 (n − 1 − p2 , Σ11.2 ), where V was partitioned in conformity as V11 V12 V= . V21 V22 ˆ of Σ, it is clear that the MLE of Since V is proportional to the MLE Σ ρij|x2 is just rij|x2 =

vij|x2 1/2 1/2 vii|x2 vjj|x2

,

where V11.2 = (vij|x2 ) and vij|x2 denotes the (i, j) element of V11.2 . Considering the distribution of V11.2 , the distribution of rij|x2 is the same as

8.6. Test of sphericity

117

for a simple correlation coeﬃcient but with n − p2 in place of n. We have proved: Proposition 8.10 2 rij|x2 P ≤t = 2 1 − rij|x 2 2 P rij|x ≤t 2

=

∞ k=0 ∞

pk · P (Fc (1 + 2k, n − p2 − 2) ≤ t) ,

pk · P beta 12 (1 + 2k); 12 (n − p2 − 2) ≤ t ,

k=0

where pk are the negative binomial probabilities

1 2 (n − p2 − 1) k (1 − ρ2ij|x2 )(n−p2 −1)/2 ρ2k pk = ij|x2 , k = 0, 1, . . . . k! For large samples, as for the simple correlation coeﬃcient, it follows from Problem 6.4.8 that

d n1/2 rij|x2 − ρij|x2 → N 0, (1 − ρ2ij|x2 )2 . A Fisher’s z-transform as for the simple correlation coeﬃcient in Problem 6.4.9 is deﬁnitely possible for a partial correlation coeﬃcient.

8.6 Test of sphericity Assume x ∼ Np (µ, Σ), Σ > 0, and consider testing the hypothesis that the p variables in x = (x1 , . . . , xp ) are independent and have the same variance: H0 : Σ = γI, γ > 0. Based on a random sample x1 , . . . , xn , regardless of the hypothesis H0 , as ˆ =x ¯ . Now, without constraint, long as Σ > 0, the MLE of µ is always µ ˆ = 1 V, where, as usual, the MLE of Σ is Σ n V=

n

¯ )(xi − x ¯ ) . (xi − x

i=1

However, under H0 , the MLE is obtained by solving

max |γI|−n/2 etr − 12 γ −1 V . γ>0

Taking logarithms, the function to maximize is − 12 np ln γ − 12 γ −1 tr V,

118

8. Tests on mean and variance

and the solution is easily calculated, γˆ = tr V/np. Therefore, the likelihood ratio, ﬁrst derived by Mauchly (1940), becomes

ˆ |ˆ γ I|−n/2 etr − 12 γˆ −1 V L(ˆ γ I, µ) = Λ = ˆ µ) ˆ −n/2 etr − 1 Σ ˆ −1 V ˆ L(Σ, |Σ| 2 n/2 exp(− 12 np) | n1 V| = . 1 |( np tr V)I| exp(− 12 np) Thus, ˜ ≡ Λ2/n Λ

|V| = 1 = ( p tr V)p

1/p p i=1 li p 1 i=1 li p

p ,

where l1 ≥ · · · ≥ lp are the ordered eigenvalues of V. The LRT compares the geometric and arithmetic means of those eigenvalues; they coincide when V has the structure as in H0 . Proposition 8.11 The LRT for testing H0 : Σ = γI, γ > 0 against ˜ = |V|/( 1 tr V)p . H1 : Σ > 0 rejects H0 for small values of Λ p At this point, we remind the reader about the general expression for the asymptotic degrees of freedom for likelihood ratio tests. In general, for testing H0 : θ ∈ Θ0 against H1 : θ ∈ Θc0 under regularity conditions, then, d

under H0 , −2 ln Λ → χ2f as the sample size n → ∞. The degrees of freedom f is the diﬀerence between the number of free parameters in Θ = Θ0 ∪ Θc0 and the number of free parameters in Θ0 . From the general theory of LRT, it is clear that the asymptotic null distribution is −2 ln Λ → χ2f , f = 12 p(p + 1) − 1. d

Lemma 8.2 When Σ = γI, γ > 0, tr V

|=

Better approximations can be obtained by calculating the moments of Λ (or ˜ as in Section 12.3. The moments are easily calculated with the following Λ) lemma. |V|/(tr V)p .

Proof. When Σ = γI, clearly the distribution of |V|/(tr V)p does not depend on γ. From the likelihood for (µ, γ), which forms an exponential family, the minimal suﬃcient and complete statistic is (¯ x, tr V). The 2 conclusion follows using Basu’s1 theorem.

|=

1 Basu’s theorem: If T is complete and suﬃcient for the family P = {P : θ ∈ Θ}, θ A, for any ancillary statistic A. By deﬁnition, a statistic A is ancillary iﬀ its then T distribution does not depend on θ.

8.6. Test of sphericity

But then, since

˜ Λ

then

1 tr V p

˜h · E EΛ

119

p = |V|,

1 tr V p

ph = E |V|h ,

from which ˜h = EΛ

E |V|h . E ( p1 tr V)ph

Proposition 8.12 When Σ = γI, γ > 0, then

1 Γp 12 (n − 1) + h h ph Γ 2 (n − 1)p ˜

E Λ =p . Γ 12 (n − 1)p + ph Γp 12 (n − 1) Proof. We have V ∼ γWp (n − 1). The proof follows directly from the above remark in conjunction with Corollary 7.3, Proposition 7.2, and the deﬁnition of Γp (·) on page 93. Moments of chi-square distributions can be obtained from Section 3.2 (v. Problem 3.5.1). 2 ˜ can be characterized as a product of Finally, the exact distribution of Λ independent beta variables [Srivastava and Khatri (1979), p. 209]. ˜ is Proposition 8.13 The exact null distribution of Λ d ˜= Λ

p−1

beta[ 12 (n − 1 − i), i( 12 + p1 )];

(8.5)

i=1

˜ is distributed as the product of p − 1 mutually independent beta i.e., Λ variables. Proof. We make use of the multiplicative formula of Gauss [Erd´elyi et al. (1953), p. 4] 1

Γ(mz) = (2π)−(m−1)/2 mmz− 2

m−1

Γ(z +

r m ),

m = 2, 3, . . . ,

r=0

with m = p and z = 12 (n − 1), 12 (n − 1) + h. We can then rewrite the moments as p−1 1 p r 1 1 r=0 Γ[ 2 (n − 1) + p ] i=1 Γ[ 2 (n − 1) + h − 2 (i − 1)] h ˜ p = EΛ 1 1 p−1 1 r i=1 Γ[ 2 (n − 1) − 2 (i − 1)] r=0 Γ[ 2 (n − 1) + h + p ] p−1 1 1 1 i i=1 Γ[ 2 (n − 1) + h − 2 i] Γ[ 2 (n − 1) + p ] . = p−1 1 1 1 i i=1 Γ[ 2 (n − 1) − 2 i] Γ[ 2 (n − 1) + h + p ]

120

8. Tests on mean and variance

It is then straightforward to check that all moments of order h > 0 on the d left and right sides of = in (8.5) are the same. Since the domain is the bounded interval [0, 1], there is a unique distribution with these moments [Serﬂing (1980), p. 46]. 2 The group Op × Rp × (R\{0}) transforms the data as xi → aHxi + b, for any H ∈ Op , b ∈ Rp , and a = 0. It preserves normality and induces transformations on the minimal suﬃcient statistic (¯ x, V) and parameters (µ, Σ) ¯ → aH¯ as x x + b, V → a2 HVH , µ → aHµ + b, and Σ → a2 HΣH . Thus, the transformation also preserves the sphericity. A test function f (¯ x, V) is said to be invariant with respect to this group of transformations when it takes the same value on the original data as on the transformed data, i.e., f (y, W) = f (aHy + b, a2 HWH ), ∀(H, b, a) ∈ Op × Rp × (R\{0}), ∀(y, W) ∈ Rp × Pp . This invariance property yields formidable simpliﬁcations. First, if we diagonalize V = HDH , where D = diag(l1 , . . . , lp ), then −1/2 ¯ , we ﬁnd choosing a = lp and b = −aH x f (¯ x, V)

¯ + b, a2 H VH) = f (aH x ˜ = f (0, D),

˜ = diag(l1 /lp , . . . , lp−1 /lp , 1) depends on the sample only through where D the ratios l1 /lp , . . . , lp−1 /lp . So, any invariant test can be written as a function of li /lp , i = 1, . . . , p − 1. Second, if we diagonalize Σ = GDλ G , where Dλ lists the eigenvalues −1/2 and b = −aG µ, λ1 ≥ · · · ≥ λp on its diagonal, then choosing a = λp we ﬁnd ˜ ), ¯ + b ∼ Np (0, n−1 D aG x λ 2 ˜ a G VG ∼ Wp (n − 1, Dλ ), ˜ where D λ = diag(λ1 /λp , . . . , λp−1 /λp , 1). Thus, the non-null distribution of any invariant test depends on (µ, Σ) only through the ratios λ1 /λp , . . . , λp−1 /λp . These invariance results are summarized in a proposition. Proposition 8.14 With respect to the above group of transformations, any invariant test depends on the minimal suﬃcient statistic (¯ x, V) only through the ratios l1 /lp , . . . , lp−1 /lp of eigenvalues of V. The power function of any invariant test depends on (µ, Σ) only through the ratios λ1 /λp , . . . , λp−1 /λp of eigenvalues of Σ. The LRT is obviously invariant. There is no uniformly most powerful invariant (UMPI) test for the sphericity hypothesis, but John (1971) showed the test based on J = tr V2 /(tr V)2 is locally most powerful (best)

8.7. Test of equality of variances

121

invariant (LBI). The null distribution of the LBI test is given in John (1972).

8.7 Test of equality of variances The data consist of a independent a samples xi1 , . . . , xini i.i.d. Np (µi , Σi ), Σi > 0, i = 1, . . . , a. Let n = i=1 ni be the total number of observations. The hypothesis in question here is the equality of variances H0 : Σ1 = · · · = Σa which is being tested against all alternatives. Since the samples are independent, the likelihood function can be built immediately from (7.1), L(Σ1 , . . . , Σa , µ1 , . . . , µa ) a ' & ni , |Σi |− 2 etr − 12 [Vi + ni (¯ xi − µi )(¯ xi − µi ) ] Σ−1 ∝ i i=1

where, as usual, ¯i x

=

Vi

=

ni 1 xij , ni j=1 ni

¯ i )(xij − x ¯ i ) , i = 1, . . . , a. (xij − x

j=1

Without the restriction speciﬁed in H0 , the parameters are unrelated and, ˆ i = 1 Vi . Under ¯ i and Σ ˆi = x thus, the unrestricted MLE is just the usual µ ni H0 , however, we have Σ1 = · · · = Σa = Σ, for some unknown Σ, and, thus, ˆ ˆˆ = x ˆ = 1 V, where V = a Vi pools all the variances together. ¯ i and Σ µ i i=1 n Thus, the LRT becomes Λ = = =

¯1, . . . , x ¯a) L( n1 V, . . . , n1 V, x ¯1, . . . , x ¯a) L( n11 V1 , . . . , n1a Va , x a | 1 V|−ni /2 exp(− 12 np) ai=1 1n −ni /2 exp(− 1 np) i=1 | ni Vi | 2 a ni /2 pn/2 n i=1 |Vi | . a pni /2 |V|n/2 n i i=1

Proposition 8.15 The LRT for testing H0 : Σ1 = · · · = Σa rejects the hypothesis for small values of a |Vi |ni /2 npn/2 Λ = i=1 n/2 . a pn /2 |V| n i i=1

i

122

8. Tests on mean and variance

The group Gp × (Rp )a transforms the observations as xij → Axij + bi , for any A ∈ Gp and bi ∈ Rp , i = 1, . . . , a. This obviously preserves the ¯ i → A¯ xi + bi , Vi → AVi A , normality and transforms the statistics as x and V → AVA , and induces the parameter transformation Σi → AΣi A . The hypothesis H0 is thus also preserved by this group of transformations. Therefore, the LRT statistic evaluated at the transformed data Axij + bi is a ni /2 npn/2 i=1 |AVi A | Λ = pni /2 a |AVA |n/2 i=1 ni a ni /2 npn/2 i=1 |Vi | = , pni /2 a |V|n/2 i=1 ni which is identical to the LRT statistic evaluated at the original data xij . We say that the LRT is invariant with respect to this group of transformations. ¯ a , V1 , . . . , Va ) is termed invariant iﬀ In general, a test function f (¯ x1 , . . . , x f (y1 , . . . , ya , W1 , . . . , Wa ) = f (Ay1 + b1 , . . . , Aya + ba , AW1 A , . . . , AWa A ), ∀(A, b1 , . . . , ba ) ∈ Gp ×(Rp )a , ∀(y1 , . . . , ya , W1 , . . . , Wa ) ∈ (Rp )a ×(Pp )a . This has important consequences. First, by deliberately choosing A = Σ−1/2 , where Σ1 = · · · = Σa = xi , it is clear that AVi A ∼ Wp (ni − 1) do Σ under H0 , and bi = −A¯ not involve any unknown parameters. Thus, the null distribution of any invariant test function ¯ a , V1 , . . . , V a ) f (¯ x1 , . . . , x = f (A¯ x1 + b1 , . . . , A¯ xa + ba , AV1 A , . . . , AVa A ) = f (0, . . . , 0, AV1 A , . . . , AVa A ) such as Λ is parameter free. Note that we need only consider test ¯ a , V1 , . . . , Va ) is suﬃcient for functions of this form since (¯ x1 , . . . , x (µ1 , . . . , µa , Σ1 , . . . , Σa ). −1/2 −1/2 Second, in the special case a = 2, diagonalize V1 V2 V1 = HDH , where H ∈ Op and D = diag(l1 , . . . , lp ) contains the eigenvalues of V1−1 V2 . −1/2 xi , we ﬁnd that This time by deliberately choosing A = H V1 , bi = −A¯ for any invariant test ¯ 2 , V1 , V2 ) f (¯ x1 , x

= f (0, 0, AV1 A , AV2 A ) = f (0, 0, I, D)

is a function of l1 , . . . , lp only. Thus, any invariant test function depends on −1/2 −1/2 = the data only through l1 , . . . , lp . Similarly, diagonalizing Σ1 Σ2 Σ1 GDλ G , where G ∈ Gp and Dλ contains the eigenvalues λ1 , . . . , λp of −1/2 , AV1 A ∼ Wp (n1 − 1) and Σ−1 1 Σ2 , we have after choosing A = G Σ1

8.7. Test of equality of variances

123

AV2 A ∼ Wp (n2 − 1, Dλ ). Thus, the non-null distribution of any invariant test function depends on (µ1 , µ2 , Σ1 , Σ2 ) only through the eigenvalues of Σ−1 1 Σ2 . Proposition 8.16 With respect to the group of transformations Gp × (Rp )2 , any invariant test for testing H0 : Σ1 = Σ2 depends on ¯ 2 , V1 , V2 ) only through the eigenvalues l1 , . . . , lp of V1−1 V2 . The (¯ x1 , x power function of any invariant test depends on (µ1 , µ2 , Σ1 , Σ2 ) only through the eigenvalues λ1 , . . . , λp of Σ−1 1 Σ2 . For example, the LRT when a = 2 can be written Λ=

npn/2 pn1 /2 pn2 /2 n2

n1

p

n /2

li 2 . (1 + li )n/2 i=1

An alternative invariant test function [Nagao (1973)] is 2 a n ni tr Vi V−1 − I . N = 12 ni i=1 Continuing now with the moments of the null distribution of the LRT, we comment ﬁrst on a result of unbiasedness. Although the LRT is a biased test, Perlman (1980) proved that the slight modiﬁcation a |Vi |mi /2 mpm/2 ∗ , Λ = i=1 m/2 a pmi /2 |V| i=1 mi where the sample sizes ni are replaced by the corresponding degrees of a freedom mi = ni − 1 and m = i=1 mi = n − a, yields an unbiased test. We will, thus, concentrate on the latter. It was Bartlett (1937) who ﬁrst proposed the use of the modiﬁed LRT, Λ∗ . For a = 2, unbiasedness of Λ∗ was established earlier by Sugiura and Nagao (1968), whereas Srivastava, Khatri, and Carter (1978) proved a monotonicity property stronger than unbiasedness. The null moments of Λ∗ is a simple consequence of invariance coupled with the normalizing constant cp,m = [2mp/2 Γp ( 12 m)]−1 of a Wp (m) p.d.f. Proposition 8.17 Under H0 , the moments of the modiﬁed LRT Λ∗ are given by a Γp ( 12 m) Γp [ 12 mi (1 + h)] . pmi h/2 Γ [ 1 m(1 + h)] Γp ( 12 mi ) p 2 i=1 i=1 mi

E Λ∗ h = a

mpmh/2

Proof. Under H0 , by invariance, we can assume Σi = I and Vi ∼ Wp (mi ) are independently distributed. Thus, from the Wp (mi ) densities, we have a ∗h cp,mi ··· |V|−mh/2 EΛ = i=1

V1 >0

Va >0

124

8. Tests on mean and variance

·

a

|Vi |[mi (1+h)−p−1]/2 etr(− 12 Vi )dV1 · · · dVa .

i=1

The integrand is seen to contain the p.d.f. of Vi ∼ Wp (mi (1 + h)) independently distributed. However, when this is the case V ∼ Wp (m(1 + h)). Thus, we ﬁnd a cp,mi ∗h E |V|−mh/2 , E Λ = a i=1 i=1 cp,mi (1+h) where V ∼ Wp (m(1 + h)). Using the moments of the generalized variance in Problem 7.5.6 and simplifying, the conclusion is reached. 2 An accurate approximation to the null distribution of the modiﬁed LRT Λ∗ by asymptotic expansion of high order is discussed in Example 12.4.

8.8 Asymptotic distributions of eigenvalues Based on a random sample x1 , . . . , xn from Np (µ, Σ), n several tests on the ¯ )(xi − x ¯ ) ∼ variance Σ are a function of the eigenvalues of V = i=1 (xi − x Wp (n−1, Σ). It was seen that an invariant test for sphericity, H0 : Σ = σ 2 I, depends only on (l1 /lp , . . . , lp−1 /lp ) where l1 ≥ · · · ≥ lp are the eigenvalues of V. Also, in the two independent samples problem, x11 , . . . , x1n1 x21 , . . . , x2n2

i.i.d. Np (µ1 , Σ1 ), i.i.d. Np (µ2 , Σ2 ),

an invariant test for the equality of variances, H0 : Σ1 = Σ2 , depends only on the eigenvalues of V1−1 V2 , where Vi =

ni

¯ i )(xij − x ¯ i ) ∼ Wp (ni − 1, Σi ), i = 1, 2. (xij − x

j=1

The distribution of eigenvalues of various random matrices thus plays an important role in testing hypotheses.

8.8.1

The one-sample problem

We investigate the asymptotic distribution of the eigenvalues l1 , . . . , lp of V ∼ Wp (m, Σ). We already know there exists H ∈ Op such that H ΣH = Λ, where Λ = diag(λ1 , . . . , λp ), and since V and H VH have the same eigenvalues, we can assume at the outset that Σ = Λ is diagonal. An eﬀective method for such problems is to write S=

V = Λ + m−1/2 V(1) , m

8.8. Asymptotic distributions of eigenvalues

125

where V(1) = m1/2 (S − Λ) is Op (1), and expand the eigenvalues of S, li /m, around λi in powers of m−1/2 . This is called the perturbation method [Bellman (1960), Kato (1982)]. We now clearly outline the steps to obtain an approximation, with remainder of the order O(m−1 ), to the distribution function of a nearly arbitrary function f (l/m) of l = (l1 , . . . , lp ) . Step 1: Perturbation method More generally, consider a diagonal matrix Λ = diag(λ1 , . . . , λp ) and assume that the perturbation of Λ can be expressed as a power series in as follows: R = Λ + V(1) + 2 V(2) + O(3 ), where V(j) , j = 1, 2, are symmetric and is a small real number. We shall discuss the case when λα is distinct from the other p − 1 eigenvalues. Let lα be the αth eigenvalue of R and cα = (c1α , . . . , cpα ) the corresponding normalized eigenvector with cαα > 0. The quantities lα and cα can be assumed of the form [Bellman (1960), p. 61] 2 (2) 3 = λα + λ(1) α + λα + O( ), p p (1) (2) = eα + aiα ei + 2 aiα ei + O(3 ),

lα cα

i=1

(8.6) (8.7)

i=1

where ei = (0, . . . , 1, . . . , 0) is the ith canonical basis vector. We determine (1) (2) (1) (2) the unknown coeﬃcients λα , λα , aiα , and aiα by substituting (8.6) and (8.7) into the equation Rcα = lα cα and equating the coeﬃcients of the powers of . This gives [Λ + V(1) + 2 V(2) + O(3 )][eα +

p

(1)

aiα ei + 2

p

i=1

2 (2) 3 = [λα + λ(1) α + λα + O( )][eα +

p

(2)

aiα ei + O(3 )]

i=1

(1)

aiα ei + 2

p

i=1

(2)

aiα ei + O(3 )],

i=1

and equating the coeﬃcients, we obtain the equations λ α eα p

(1)

(1) aiα λi ei + vα

= λ α eα , p (1) = aiα λα ei + λ(1) α eα ,

i=1

and p i=1

(2)

aiα λi ei +

(8.8) (8.9)

i=1

p i=1

=

(1) (1)

aiα vi

p i=1

(2)

+ vα

(2) aiα λα ei

+

p i=1

(1)

(2) aiα λ(1) α ei + λ α eα ,

(8.10)

126

8. Tests on mean and variance (j)

(j)

where V(j) = (v1 , . . . , vp ), j = 1, 2. The αth component of (8.9) yields (1) (1) (1) a(1) αα λα + vαα = aαα λα + λα (1)

(1)

from which λα = vαα . The component i = α of the same equation yields (1)

(1)

(1)

aiα λi + viα = aiα λα , (1)

(1)

from which aiα = −viα λiα , i = α, where λiα = 1/(λi − λα ). (1)

(1)

Note that aαα can be chosen arbitrarily and we set aαα = 0 here. The (2) (2) unknown quantities λα and aiα can be determined similarly using (8.10). The expansions (8.6)-(8.7) from the perturbation analysis thus take the ﬁnal form 2 (1) (1) (2) lα = λα + vαα + 2 vαα + λαβ vαβ + O(3 ), (8.11) β =α

ciα

(1) (1) (1) (1) (1) = −λiα viα + 2 λiα viα vαα + λαβ viβ vβα β =α

3

cαα

+O( ), i = α, 2 (1) = 1 + 2 − 12 λ2αβ vαβ + O(3 ). β =α

Returning to our one-sample problem, assuming λα is distinct, the eigenvalue lα /m of S can be expanded by setting V(2) = 0 as (1) 2 (1) lα /m = λα + m−1/2 vαα + m−1 λαβ vαβ + Op (m−3/2 ). (8.12) β =α

Step 2: Taylor series of f (l/m) Assuming f (·) is continuously diﬀerentiable in a neighborhood of λ = (λ1 , . . . , λp ) , we can write the Taylor series around λ, f (l/m) = f (λ) + Df (λ)(l/m − λ) + 12 (l/m − λ) D2 f (λ)(l/m − λ) + Op (m−3/2 ). Upon using (8.12), this becomes f (l/m)

−1/2

= f (λ) + m

p

(1) fi vii

+m

i=1

+ 12 m−1

p p i=1 j=1

−1

p i=1

(1) (1)

fi

(1) 2

λiβ viβ

β =i

fij vii vjj + Op (m−3/2 ),

(8.13)

8.8. Asymptotic distributions of eigenvalues

127

where Df (λ) D2 f (λ)

= =

(f1 , . . . , fp ) = (∂f (λ)/∂λi ) ,

(fij ) = ∂ 2 f (λ)/∂λi ∂λj .

Step 3: Expansion of the characteristic function The characteristic function of m1/2 [f (l/m) − f (λ)] thus becomes E exp{itm1/2 [f (l/m) − f (λ)]} p p it (1) (1) 2 fi vii fi λiβ viβ = E exp it exp √ m i=1 i=1 β =i p p (1) (1) +1 fij vii vjj + Op (m−1 ) 2

i=1 j=1

p it (1) (1) 2 = E exp it fi vii fi λiβ viβ 1+ √ m i=1 i=1 β =i p p (1) (1) + 12 fij vii vjj + Op (m−1 ) .

p

(8.14)

i=1 j=1

We need then to evaluate the following expectations in (8.14): p (1) fi vii , E exp it E exp it E exp it

i=1 p i=1 p

(1) fi vii

(1) 2

· viβ , β = i,

(8.15) (8.16)

(1) fi vii

(1) (1)

· vii vjj .

(8.17)

i=1

Step 4: Sugiura’s lemma [Sugiura (1973)] Let V ∼ Wp (m, Σ) and S = V/m. Lemma 8.3 Let g(S) be an analytic function at S = Σ and put T = m1/2 (S − Σ). Deﬁne a matrix of diﬀerential operators by

∂ = 12 (1 + δij )∂/∂ij applied to the function g(Γ) of a symmetric matrix Γ = (γij ). Then, for any symmetric matrix A and suﬃciently large m, 2 d2j−1 (it)2j−1 E g(S)etr(itAT) = etr[−t2 (AΣ)2 ] · 1 + m−1/2 j=1

128

8. Tests on mean and variance

+m−1

3

g2j (it)2j + O(m−3/2 ) g(Γ)|Γ=Σ ,

j=1

where each coeﬃcient is given by d1

=

d3

=

g0

=

g2

=

g4

=

g6

=

2 tr (ΣAΣ∂), 4 tr (ΣA)3 , 3 tr (Σ∂)2 , 1 4 tr (ΣA)2 Σ∂ + d21 , 2 2 tr (ΣA)4 + d1 d3 , 1 2 d . 2 3

Before presenting the proof, we comment on Taylor series and diﬀerential operators. An analytic function g(x), at x0 , of a real variable x can be written as a Taylor series g(x)

= g(x0 ) +

∞ g (j) (x0 ) j=1

(x−x0 )∂

= e =

j!

(x − x0 )j

· g(x)|x=x0

[1 + (x − x0 )∂ + 12 (x − x0 )2 ∂ 2 + · · ·]g(x)|x=x0 ,

where ∂ j g(x)|x=x0 = ∂ j g(x0 )/∂xj is the jth derivative of g evaluated at x0 . In the same way, for a function g(S) analytic at Σ, of a symmetric matrix S, we have g(S) = {etr(S − Σ)∂} g(Γ)|Γ=Σ . p

Proof. Note that S → Σ. Taylor series expansion of g(S) at Σ gives g(S) = {etr(S − Σ)∂} g(Γ)|Γ=Σ . After multiplying by etr(itAT) and taking expectations with respect to V ∼ Wp (m, Σ), we get E etr(itAT)g(S)

= |I − 2m−1/2 itAΣ − 2m−1 Σ∂|−m/2 ( ) · etr(−m1/2 itAΣ − Σ∂) g(Γ)|Γ=Σ .

The above determinant can be arranged according to powers of m as in Sugiura and Nagao (1971). 2 Evaluation of (8.15)

8.8. Asymptotic distributions of eigenvalues

129

Let A = diag(f1 , . . . , fp ), g(Γ) ≡ 1, and Σ = Λ, and note that tr AV(1)

=

p

(1)

fi vii ,

i=1

tr(AΛ)2

=

p

fi2 λ2i ≡ τ 2 /2 (say).

i=1

Since d1 g(Γ) = 0, the lemma yields p (1) E exp it fi vii = exp(− 12 t2 τ 2 )[1 + m−1/2 d3 (it)3 + O(m−1 )], i=1

p where d3 = (4/3) i=1 fi3 λ3i . Evaluation of (8.16) 2 , i = β. Note that g(Λ) = 0 and Let A = diag(f1 , . . . , fp ) and g(Γ) = mγiβ the diﬀerential operator d1 , d1 = 2

p

fk λ2k ∂kk ,

k=1

is a linear combination of ∂kk ≡ ∂/∂kk and, thus, d1 g(Γ) = 0. For similar reasons, we also have g2 g(Γ) = 0 and d3 g(Γ)|Γ=Λ = g4 g(Γ)|Γ=Λ = g6 g(Γ)|Γ=Λ = 0, which implies p (1) (1) 2 fi vii · viβ = exp(− 12 t2 τ 2 )[m−1 g0 + O(m−3/2 )]g(Γ)|Γ=Λ . E exp it i=1

However, g0 is the diﬀerential operator g0 = tr(Λ∂)2 =

p p

2 λk λl ∂kl

k=1 l=1

and, thus, g0 g(Γ)|Γ=Λ =

p p

2 2 λk λl ∂kl (mγiβ )|Γ=Λ = mλi λβ .

k=1 l=1

Hence, we get E exp it

p

(1) fi vii

(1) 2

· viβ

= exp(− 12 t2 τ 2 )[λi λβ + O(m−1/2 )].

i=1

Evaluation of (8.17) Similarly, letting A = diag(f1 , . . . , fp ) and g(Γ) = m(γii − λi )(γjj − λj ), we ﬁnd p (1) (1) (1) E exp it · vii vjj = exp(− 12 t2 τ 2 )[2λ2i δij + O(m−1/2 )]. fi vii i=1

130

8. Tests on mean and variance

We now return to the expansion of the characteristic function in (8.14). Hence, altogether, the expansion of the characteristic function becomes E exp{itm1/2 [f (l/m) − f (λ)]/τ } = exp(− 12 t2 τ 2 )[1 + m−1/2

2 a2j−1 j=1

τ 2j−1

(it)2j−1 + O(m−1 )],

where a1 a3

=

p

=

fi λiβ λi λβ +

i=1 β =i p

4 3

p

fii λ2i ,

(8.18)

i=1

fi3 λ3i .

(8.19)

i=1

Step 5: Inversion of the characteristic function Using the inversion formula (2.2), an expansion for the density function of s = m1/2 [f (l/m) − f (λ)]/τ is f (s)

=

1 2π

+m

∞

−∞ −1/2

1

2

e−its e− 2 t dt ∞ 2 1 2 a2j−1 1 e−its e− 2 t (it)2j−1 dt + O(m−1 ) 2j−1 2π τ −∞ j=1

= φ(s) − m−1/2

2 a2j−1 j=1

τ 2j−1

φ(2j−1) (s) + O(m−1 ),

and similarly for the distribution function of s, F (s) = Φ(s) − m−1/2

2 a2j−1 j=1

τ 2j−1

Φ(2j−1) (s) + O(m−1 ),

where φ and Φ are respectively the density function and distribution function of the standard normal distribution. We have proved: Proposition 8.18 Let f (·) be continuously diﬀerentiable in a neighbor2 hood p of 2λ. 2 If the population eigenvalues λα are all distinct and τ = 2 i=1 fi λi = 0, then the distribution function of s = m1/2 [f (l/m) − f (λ)]/τ can be expanded for large m as Φ(s) − m−1/2

2 a2j−1 j=1

τ 2j−1

Φ(2j−1) (s) + O(m−1 ).

8.8. Asymptotic distributions of eigenvalues

131

Corollary 8.3 Let f (·) be continuously diﬀerentiable in a neighborhood of p λ. If the population eigenvalues λα are all distinct and τ 2 = 2 i=1 fi2 λ2i = 0, then the limiting distribution is given by s = m1/2 [f (l/m) − f (λ)]/τ → N (0, 1). d

For an individual eigenvalue the expansion follows immediately. Corollary 8.4 Let lα be the αth largest eigenvalue of V ∼ Wp (m, Λ). If λα is distinct from all other p − 1 eigenvalues, the distribution function of √ s = m1/2 (lα /m − λα )/( 2λα ) can be expanded for large m as Φ(s) − m−1/2

2 a2j−1 j=1

where a1

=

τ 2j−1

Φ(2j−1) (s) + O(m−1 ),

λα λβ /(λα − λβ ),

β =α

4 3 λ . 3 α The sample eigenvalues are asymptotically independent, as the following corollary shows. a3

=

Corollary 8.5 Let l = (l1 , . . . , lp ) be the eigenvalues of V ∼ Wp (m, Λ). If the population eigenvalues λα are all distinct, then the joint limiting distribution is given by (m/2)1/2 Λ−1 (l/m − λ) → Np (0, I). d

Proof. From (8.12), we can write (1)

(1) m1/2 (l/m − λ) = (v11 , . . . , vpp ) + Op (m−1/2 ) ≡ v(1) + Op (m−1/2 ).

The asymptotic distribution of V(1) was derived in Section 6.3, and for the d marginal v(1) , we ﬁnd, using (6.1), v(1) → Np (0, 2Λ2 ). 2 The asymptotic expansion in Corollary 8.4 gives the ﬁrst two terms of a more accurate approximation, with remainder O(m−3/2 ), Φ(s) − m−1/2

2 a2j−1 j=1

τ

Φ(2j−1) (s) + m−1 2j−1

3 b2j (2j) Φ (s) + O(m−3/2 ), 2j τ j=1

where a1 and a3 are given in Corollary 8.4 and b2 = 2λ2α λβ /(λα − λβ ) − 2λ3α λβ /(λα − λβ )2 β =α

β =α

132

8. Tests on mean and variance

3 + λ2α λ2β /(λα − λβ )2 + λ2α λγ λβ /(λα − λγ )(λα − λβ ), 2 γ δ1 , then 0 ≤ αj − βj ≤ ρ2 (C)/(βj − δ1 ), j = 1, . . . , q, 0 ≤ δr−i − αp−i ≤ ρ2 (C)/(βq − δr−i ), i = 0, . . . , r − 1. A proof of Wielandt’s inequality can be found in Eaton and Tyler (1991), who used it to ﬁnd the asymptotic distribution of the eigenvalues of symmetric random matrices in the case of multiple eigenvalues. The matrix A can be viewed as a perturbation of a block-diagonal matrix, namely A = A0 + E, where B 0 0 C A0 = and E = . 0 D C 0 By Wielandt’s inequality, the eigenvalues of A0 are perturbed quadratically in E when A0 is perturbed linearly in E. Generally, eigenvalues are only perturbed linearly when the matrix is perturbed linearly. The quadratic perturbation of eigenvalues in Wielandt’s inequality is due to the special structure of E relative to A0 . Let Sp be the set of p × p real symmetric matrices. Consider a sequence of random matrices Sn ∈ Sp , and assume that Wn = n1/2 (Sn − Σ) → W, d

for some Σ ∈ Sp , and hence W ∈ Sp . Given Σ ∈ Sp , let φ(Σ) = (φ1 (Σ), . . . , φp (Σ)) be the vector of ordered eigenvalues φ1 (Σ) ≥ φ2 (Σ) ≥ · · · ≥ φp (Σ). The asymptotic distribution of n1/2 (φ(Sn ) − φ(Σ)) is studied.

(8.21)

8.8. Asymptotic distributions of eigenvalues

135

We consider in the ﬁrst place the case Σ = diag(d1 Ip1 , . . . , dk Ipk ), p = p1 + · · · + pk , where d1 > d2 > · · · > dk represent the distinct eigenvalues of Σ with the multiplicity of di being pi , i = 1, . . . , k. To reﬂect the block structure of Σ, consider the partitioned matrix S S ··· S n,11

n,12

Sn,21 Sn = .. . Sn,k1

Sn,22 .. .

n,1k

· · · Sn,2k .. .. , . . · · · Sn,kk

Sn,k2

where Sn,ij ∈ Rppij . Lemma 8.4 For k = 2,

φ(Sn ) −

φ(Sn,11 ) φ(Sn,22 )

is Op (n−1 ).

Proof. Let An = {Sn | φp1 (Sn,11 ) > φ1 (Sn,22 )}. Since φ is continuous and, p p p from (8.21), Sn,11 → d1 Ip1 and Sn,22 → d2 Ip2 , it follows that φp1 (Sn,11 ) → p d1 and φ1 (Sn,22 ) → d2 . Thus, P (An ) → 1, so attention can be restricted to An , n = 1, 2, . . .. For Sn ∈ An , Wielandt’s inequality implies for 1 ≤ i ≤ p1 , 0 ≤ φi (Sn ) − φi (Sn,11 ) ≤ ρ2 (Sn,12 )/ (φi (Sn,11 ) − φ1 (Sn,22 )) . By (8.21), Sn,12 is Op (n−1/2 ), and since ρ is continuous, it follows that ρ2 (Sn,12 ) is Op (n−1 ). Since p

φi (Sn,11 ) − φ1 (Sn,11 ) → d1 − d2 > 0, then φi (Sn ) − φi (Sn,11 ) is Op (n−1 ), i = 1, . . . , p1 . The proof of φp−j (Sn ) − 2 φp2 −j (Sn,22 ) is Op (n−1 ), j = 0, . . . , p2 − 1, is analogous. By applying Lemma 8.4, k − 1 times, the following asymptotic equivalence result is obtained. The vector 1pi ∈ Rpi is the vector of ones. Proposition 8.21 n1/2 (φ(Sn ) − φ(Σ)) = Zn + Rn , where

φ(Sn,11 ) − d1 1p1 .. Zn = n1/2 . φ(Sn,kk ) − dk 1pk

and the remainder term Rn is Op (n−1/2 ). p

Since Rn → 0, using Slutsky’s theorem the asymptotic distribution of n1/2 (φ(Sn ) − φ(Σ))

136

8. Tests on mean and variance

is that of the leading term Zn . Considering the partitioned W = (Wij ), Wij ∈ Rppij , i, j = 1, . . . , k, we have immediately from (8.21) that Sn,11 − d1 Ip1 W11 d .. .. . n1/2 → . . Sn,kk − dk Ipk Now, because the function

Wkk

φ(W11 ) .. G(W) = .

φ(Wkk ) is continuous and since φ n1/2 (Sn,11 − d1 Ip1 ) = n1/2 (φ(Sn,11 ) − d1 1p1 ) , it follows that d

Zn = G(Wn ) → G(W). We have proved: d

Proposition 8.22 If n1/2 (Sn − Σ) → W and Σ is diagonal, then n1/2 (φ(Sn ) − φ(Σ)) → G(W). d

In the general case where Σ is not diagonal, there exists H ∈ Op such that Σ = H diag (d1 Ip1 , . . . , dk Ipk ) H ≡ HDH . d

From (8.21), n1/2 (HSn H − D) → HWH . Since φ(HSn H ) = φ(Sn ) and φ(Σ) = φ(D), we obtain the general result n1/2 (φ(Sn ) − φ(Σ)) → G(HWH ). d

This general result is summarized. d

Proposition 8.23 If n1/2 (Sn − Σ) → W, then n1/2 (φ(Sn ) − φ(Σ)) → G(HWH ), d

where H diagonalizes Σ, Σ = H diag (d1 Ip1 , . . . , dk Ipk ) H . An important special case is when W in (8.21) is a multivariate normal matrix and all eigenvalues of Σ are distinct. In that case, HWH is also a multivariate normal matrix. Also, since all eigenvalues of Σ have multiplicity 1, then G(HWH ) is just a p-dimensional marginal of HWH and, hence, has a p-dimensional normal distribution.

8.9. Problems

137

Example 8.3 It was seen in Chapter 6 that when sampling from a Np (µ, Σ) distribution, the asymptotic distribution of the sample variance d

S is n1/2 (S − Σ) → W, where W ∼ Npp (0, (I + K)(Σ ⊗ Σ)). We derive the asymptotic distribution of the eigenvalues of S when all eigenvalues of Σ are distinct. Using Proposition 8.23, we have d n1/2 (φ(S) − φ(Σ)) → G(HWH ), where H diagonalizes Σ, HΣH ≡ D (say). But from Proposition 6.1, HWH ∼ Npp (0, (I + K)(D ⊗ D)). From (6.1), (HWH )ii ∼ N (0, 2d2i ), i = 1, . . . , p, and are independently distributed. Since all eigenvalues of Σ are distinct, then n1/2 (φ(S) − φ(Σ)) → Np (0, 2D2 ), d

which is the result proven previously in Corollary 8.5. Example 8.4 We now derive the asymptotic distribution of the r smallest eigenvalues of S when the smallest eigenvalue of Σ has multiplicity r, φ(Σ) = (φ1 (Σ), . . . , φp−r (Σ), λ, . . . , λ) . As in Example 8.3, HWH ∼ Npp (0, (I + Kp )(D ⊗ D)). Hence, the lower right r × r block of HWH is distributed as Nrr (0, λ2 (I + Kr )). Using Proposition 8.23, we have ﬁnally n1/2 λ−1 (φp−r+1 (S) − λ, . . . , φp (S) − λ) → w, d

where w is distributed as the eigenvalues of a Nrr (0, (I + Kr )) distribution. An application of Wielandt’s inequality to bootstrapping eigenvalues can be found in Eaton and Tyler (1991), who extend the work of Beran and Srivastava (1985, 1987). Earlier papers on the case of multiple eigenvalues include those of James (1969), Chattopadhyay and Pillai (1973), Chikuse (1976), Khatri and Srivastava (1978), and Srivastava and Carter (1980).

8.9 Problems 1. Two-sample T 2 . Let x1 , . . . , xn i.i.d. Np (µ, Σ) and y1 , . . . , ym i.i.d. Np (τ , Σ), Σ > 0, be two independent samples. Deﬁne the sample variances Sx

=

1 ¯ )(xi − x ¯ ) , (xi − x (n − 1) i=1

Sy

=

1 ¯ )(yi − y ¯ ) , (yi − y (m − 1) i=1

n

m

138

8. Tests on mean and variance

and Spool =

1 [(n − 1)Sx + (m − 1)Sy ]. (n + m − 2)

Determine (i) the distribution of Spool , (ii) the distribution of −1 1 1 ¯ ) S−1 ¯) (¯ x−y x−y T2 = + pool (¯ n m used for testing H0 : µ = τ against H1 : µ = τ . 2. Invariance of two-sample T 2 . This is a continuation of Problem 8.9.1. ¯ , Spool ) is minimal suﬃcient for (µ, τ , Σ). (i) Prove (¯ x, y (ii) Consider (A, b) in the group of transformations Gp × Rp acting ¯ → A¯ ¯ → A¯ as x x + b, y y + b, and Spool → ASpool A . Prove that this group of transformations leaves the testing problem ¯ , Spool ) invariant and that any invariant test depends on (¯ x, y only through T 2 . (iii) Prove that any invariant test has a power function depending on (µ, τ , Σ) only through (µ − τ ) Σ−1 (µ − τ ). 3. Common mean vector. For independent samples xi1 , . . . , xini , i.i.d. Np (µ, Σi ), i = 1, . . . , a, ¯ i ) be the from distributions with a common mean vector µ, let (Si , x MVUE from each sample. Consider estimating the common mean µ with respect to the weighted least-squares criterion min µ

a

ci ni (¯ xi − µ) S−1 xi − µ) i (¯

i=1

a for some constants ci > 0, i=1 ci = 1. Establish the estimate of µ is given by −1 a a −1 −1 ˜= ¯i . ci ni Si ci ni Si x µ i=1

i=1

Remark: Jordan and Krishnamoorthy (1995) built an exact β×100% ˜ conﬁdence region centered at µ. 4. Test of symmetry. Assume x1 , . . . , xn i.i.d. Np (µ, Σ), Σ > 0. Choose C ∈ Rp−1 of p rank C = p − 1 such that C1 = 0. Prove the following:

8.9. Problems

139

(i) ker C = span{1} and, therefore, H0 : Cµ = 0 ⇐⇒ H0 : µ1 = · · · = µp . (ii) Any p − 1 columns of C are linearly independent, which implies that C = A(Ip−1 , −1), for some nonsingular A ∈ Rp−1 p−1 . Conclude that, thereafter, the value of x) (CSC )−1 (C¯ x), T 2 = n(C¯ n n ¯ )(xi − x ¯ ) , ¯ = i=1 xi /n and (n − 1)S = i=1 (xi − x where x does not depend on the choice of C. (iii) The null (under H0 ) distribution of T 2 is T 2 /(n − 1) ∼ Fc (p − 1, n − p + 1). 5. Assume x1 , . . . , xn i.i.d. x ∈ Rp (possibly non-normal) with E x = µ and var x = Σ. Establish that, asymptotically, we are at least (1 − α) × 100% conﬁdent in simultaneously presenting all of the observed “Scheﬀ´e” intervals: 1/2 1/2 χ2α,p χ2α,p 1/2 ¯− ¯+ (a Sa) ≤aµ≤ax (a Sa)1/2 , ax n n ∀a ∈ Rp . 6. Assume x1 , . . . , xn i.i.d. x ∈ Rp (possibly non-normal) with E x = µ and var x = Σ. Let a1 , . . . , ar be linearly independent in Rp . Establish that, asymptotically, we are at least (1 − α) × 100% conﬁdent in simultaneously presenting the observed “Roy-Bose” intervals: 1/2 1/2 χ2α,r χ2α,r 1/2 ¯− ¯+ ai x (ai Sai ) ≤ a i µ ≤ ai x (ai Sai )1/2 , n n i = 1, . . . , r. 7. Test of proportionality. Given x1 , . . . , xn a random sample from Np (µ, Σ), Σ > 0, obtain the likelihood ratio test Λ for H0 : Σ = γΣ0 , γ > 0 where Σ0 > 0 is a known matrix: Λ with V =

n

i=1 (xi

2/n

=

|Σ−1 0 V|/

versus

H1 : Σ > 0,

1 tr Σ−1 0 V p

¯ )(xi − x ¯ ) as usual. −x

8. Test for a given variance. Let x1 , . . . , xn i.i.d. Np (µ, Σ), Σ > 0.

p ,

140

8. Tests on mean and variance

(i) Prove that the LRT for H0 : Σ = I versus H1 : Σ = I is given by Λ = (e/n)pn/2 |V|n/2 etr(− 12 V),

n ¯ )(xi − x ¯ ) . where V = i=1 (xi − x Remark: This test is biased, but Sugiura and Nagao (1968) have shown the slight modiﬁcation Λ∗ = (e/m)pm/2 |V|m/2 etr(− 12 V), where m = n − 1, gives an unbiased test. (ii) Using the Wishart density in Problem 7.5.7, prove that mph/2 Γp [ 12 m(1 + h)] 2e |Σ|mh/2 E Λ∗ h = . m |I + hΣ|m(1+h)/2 Γp ( 12 m) Hint: The integrand has the form of a Wishart density. Simply ﬁnd the normalizing constant. (iii) Under H0 , mph/2 Γp [ 12 m(1 + h)] 2e ∗h EΛ = (1 + h)−mp(1+h)/2 . m Γp ( 12 m) Remark: An accurate approximation to the null distribution of Λ∗ using those moments is given in Example 12.5. ˆ2: 9. Use Proposition 8.6 to obtain the moments of R

∞ Γ 12 (p − 1 + 2k) + h Γ 12 (n − 1 + 2k) 2h ˆ

, pk ER = 1 1 Γ (p − 1 + 2k) Γ (n − 1 + 2k) + h 2 2 k=0 where pk are the negative binomial probabilities given in Proposition 8.6. 10. Demonstrate that if K given δ is Poisson(δ) and δ ∼ G(p, θ), then the marginal of K is the negative binomial K ∼ nb(p, (1 + θ)−1 ). 11. Write Nagao’s test for the equality of two variances, H0 : Σ1 = Σ2 , as a function of the eigenvalues l1 , . . . , lp of V1−1 V2 , where as usual ni ¯ i )(xij − x ¯ i ) , i = 1, 2. Vi = j=1 (xij − x 12. Write the LBI test for sphericity as a function of l1 /lp , . . . , lp−1 /lp , n where l1 ≥ · · · ≥ lp are the ordered eigenvalues of V = i=1 (xi − ¯ ) . ¯ )(xi − x x 13. Invariance of multiple correlation. Let xi = (xi1 , xi2 ) , i = 1, . . . , n, be i.i.d. Np (µ, Σ), Σ > 0. Consider the group of transformations axi1 + b xi → , Axi2 + b for any a = 0, b ∈ R, A ∈ Gp−1 , and b ∈ Rp−1 .

8.9. Problems

141

(i) Show that this transformation induces the following transformations on the suﬃcient statistics and parameters: a¯ x1 + b ¯ → x , A¯ x2 + b 2 a v11 av21 A , V → aAv21 AV22 A aµ1 + b µ → , Aµ2 + b 2 aσ 21 A a σ11 . Σ → aAσ 21 AΣ22 A −1/2

−1/2

x1 , A = V22 , and b = −A¯ x2 to (ii) Choose a = v11 , b = −a¯ prove that any invariant test f (¯ x, V) depends on the data only −1/2 1/2 through u = V22 v21 /v11 . (iii) Prove that there exists an orthogonal transformation H ∈ Op−1 ˆ 1 (v. Problem 1.8.14). such that Hu = Re (iv) Choosing further a = 1, b = 0, A = H, and b = 0, prove that ˆ any invariant test is necessarily a function of R. (v) Prove that the non-null distribution of any invariant test depends on the parameters only through R. 14. Test of equality of means and variances. The data consists of a independent samplesxi1 , . . . , xini i.i.d. a Np (µi , Σi ), Σi > 0, i = 1, . . . , a. Let n = i=1 ni be the total number of observations. The hypothesis is the equality of the distributions H0 : µ1 = · · · = µa ; Σ1 = · · · = Σa which is being tested against all alternatives. Let ¯i x

=

ni 1 xij , ni j=1

1 ¯i, ni x n i=1 a

¯ = x

be the ith sample mean and overall mean, respectively, and Vi

=

ni

¯ i )(xij − x ¯ i ) , i = 1, . . . , a, (xij − x

j=1

B

=

a

¯ )(¯ ¯ ) ni (¯ xi − x xi − x

i=1

be the usual “within” and “between” sums of squares, respectively.

142

8. Tests on mean and variance

(i) Prove that the LRT is a |Vi |ni /2 nnp/2 Λ = ai=1 . | i=1 Vi + B|n/2 a nni p/2 i=1

i

Remark: Perlman (1980) proved this LRT yields an unbiased test. (ii) Use the group of transformations xij → Axij + a, for any A ∈ Gp , a ∈ Rp , to argue that the null distribution of Λ (or any other invariant test) can be obtained by setting µi = 0 and Σi = I without loss of generality. (iii) Establish that V1 , . . . , Va , B are mutually independent whenever Σ1 = · · · = Σa holds. Moreover, verify that, under H0 , Vi ∼ Wp (ni − 1) and B ∼ Wp (a − 1). Hint: Let xi1 X1 . Xi = ... and X = .. Xa xini be the sample matrices. For appropriately chosen orthogonal projections Qi , i = 1, . . . , a, and Q, so that Vi = Xi Qi Xi and B = X QX, write Q1 0 · · · 0 Q1 X1 0 Q2 · · · 0 .. .. .. .. .. . = X . . . . Qa Xa 0 0 · · · Qa QX Q ≡ CX. Then use Proposition 6.1 and verify that CC is block-diagonal. (iv) Obtain the null moments of Λ: Γp [ 1 ni (1 + h) − 1 ] Γp [ 12 (n − 1)] 2 2 . 1 ni ph/2 Γ [ 1 n(1 + h) − 1 ] Γ [ (n − 1)] p p i n 2 2 i=1 2 i=1 i

E Λh = a

a

nnph/2

Hint: Recall the normalizing constant cp,m of a Wp (m) p.d.f. and use a similar argument as in the proof of Proposition 8.17 to establish E Λh = a

nnph/2

a

ni ph/2 i=1 i=1 ni

cp,ni −1 cp,ni (1+h)−1

E |W|−nh/2 ,

where W ∼ Wp (n(1 + h) − 1). Remark: The null moments are invoked in Problem 12.4.1 to develop an accurate approximation to the null distribution of Λ.

8.9. Problems

143

15. Assume the population eigenvalue λα is distinct from all other p − 1 eigenvalues. Establish that the logarithmic transformation is a variance stabilizing transformation for the sample eigenvalue lα of V ∼ Wp (m, Λ).

9 Multivariate regression

9.1 Introduction Multivariate regression with p responses as opposed to p multiple regressions is getting increasingly more attention, especially in the context of prediction. In this chapter, we generalize the multiple regression model of Section 5.6.2 to the multivariate case. The estimation method of Section 9.2 relies also on orthogonal projections. The model considered is Y = XB + E, where Y ∈ Rnp , B ∈ Rkp , and X ∈ Rnk of rank X = k is ﬁxed. The error term E is such that E E = 0 and var E = In ⊗ Σ with Σ > 0 in Rpp . The observation vectors consisting of the rows of Y are thus uncorrelated. The Gauss-Markov estimate is derived ﬁrst. Then, assuming normality, the maximum likelihood estimates of B and Σ are obtained together with the fundamental result about their joint distribution. Section 9.3 derives the likelihood ratio test for the general linear hypothesis H0 : CB = 0 against all alternatives where C ∈ Rrk of rank C = r in the above model. In the last sections, we discuss the practical and more commonly encountered situation of k random (observed) predictors and the problem of prediction of p responses from the same set of k predictors. Finally, an application to the MANOVA one-way classiﬁcation model is treated as a special case.

9.2. Estimation

145

The multivariate regression model can be seen as a set of p correlated multiple regression models of Section 5.6.2. With the partition Y

=

(y1 , . . . , yp ),

B

=

(β 1 , . . . , β p ),

E =

(9.1)

(e1 , . . . , ep ),

we can rewrite Y = XB + E as yi = Xβ i + ei , i = 1, . . . , p, where ei ∼ Nn (0, σii I). However, the p multiple regression models are correlated since cov(ei , ej ) = σij I. Testing a relationship between the various β i ’s will require one to treat the p models as one multivariate regression model.

9.2 Estimation First, observe that Rnp is a linear space on which we deﬁne the usual inner product Y, Z = tr(Y Z) =

p n

yij zij , for any Y, Z ∈ Rnp .

i=1 j=1

The mean of Y is in a subspace V = {XA : A ∈ Rkp }. To deﬁne the orthogonal projection of Y on V, a basis for V is needed. Partition X into columns X = (x1 , . . . , xk ). An element XA ∈ V, where

a1 A = ...

ak

is partitioned into rows, is of the form XA

=

k

xi ai =

i=1

=

p k i=1 j=1

k

p xi aij ej

i=1

aij (xi ej );

j=1

146

9. Multivariate regression

hence, {xi ej ∈ Rnp : i = 1, . . . , k; j = 1, . . . , p} spans V. Moreover, linear independence holds since p p k k aij (xi ej ) = 0 =⇒ xi aij ej = 0 i=1 j=1

i=1

=⇒

p

j=1

aij ej = 0, ∀i

j=1

=⇒ aij = 0, ∀i, j. ˆ of Y on V, note that Next, to calculate the orthogonal projection, say XB, ⊥ ˆ ˆ since Y − XB ∈ V , then xi ej , Y − XB = 0, ∀i, j. But, this means that ! ˆ = x (Y − XB)e ˆ j = 0, ∀i, j. tr ej xi (Y − XB) i ˆ = 0, which gives Therefore, X (Y − XB) ˆ = (X X)−1 X Y. B ˆ = (βˆ1 , . . . , βˆp ), where With the partition in (9.1) we ﬁnd B βˆi = (X X)−1 X yi . For the purpose of estimation of the regression coeﬃcients, the p multiple regression models can be treated separately. ˆ = X(X X)−1 X Y ≡ The orthogonal projection of Y on V becomes XB −1 PY, where P = X(X X) X . This provides the “orthogonal direct sum” Y = PY + QY with Q = I − P as usual. The Gauss-Markov property is the subject of Proposition 9.1. Consider ˆ = D(PY)F. the parameter C = D(XB)F and its natural estimate C ˆ is the “best” (blue) in the sense Among all linear unbiased estimates, C that it has the minimum variance: ˆ = blue(C). Proposition 9.1 (Gauss-Markov) C ˜ = GYH be any linear unbiased estimate. Then, Proof. Let C ˜ = C, ∀B ∈ Rk EC p

⇐⇒ GXBH = DXBF, ∀B ∈ Rkp ⇐⇒ GPYH = DPYF, ∀Y ∈ Rnp ⇐⇒ H Y PG = F Y PD , ∀Y ∈ Rnp ⇐⇒ [(GP) ⊗ H ]vec(Y ) = [(DP) ⊗ F ]vec(Y ) ⇐⇒ (GP) ⊗ H = (DP) ⊗ F .

Now, we have ˆ = var C

[(DP) ⊗ F ](In ⊗ Σ)[(DP) ⊗ F]

9.2. Estimation

= =

147

[(GP) ⊗ H ](In ⊗ Σ)[(GP) ⊗ H] ˜ (GPG ) ⊗ (H ΣH) ≤ (GG ) ⊗ (H ΣH) = var C,

with equality iﬀ GP = G

⇐⇒ [(GP) ⊗ H ]vec(Y ) = (G ⊗ H )vec(Y ), ∀Y ∈ Rnp ⇐⇒ [(DP) ⊗ F ]vec(Y ) = (G ⊗ H )vec(Y ), ∀Y ∈ Rnp ˆ = C. ˜ ⇐⇒ C 2

ˆ = blue(B); also, In Proposition 9.1, if F = I and D = (X X)−1 X , then B −1 ˆ if F = ej and D = ei (X X) X , then bij = blue(bij ), and so on. ˆ is obviously unbiased for B; furthermore, noting that The estimate B QY = QE and using Problem 6.4.5, E Y QY = E E QE = (tr Q)Σ = (n − k)Σ. Thus, S ≡ Y QY/(n − k) is an unbiased estimate of Σ. Under normality, these estimates are optimal in the sense that they have minimum variance among all unbiased estimates. The likelihood for (B, Σ) from Y is L(B, Σ) ∝ |Σ|−n/2 etr − 12 Σ−1 (Y − XB) (Y − XB) ∝ |Σ|−n/2 etr − 12 Σ−1 B X XB & ' · exp − 12 tr Σ−1 · (Y Y) − 2Σ−1 B · (X Y) . From general properties of exponential families [Fraser (1976), pp. 339, 342, 406 or Casella and Berger (1990), pp. 254-255, 263], the statistic (Y Y, X Y) is minimal suﬃcient and complete for (B, Σ). Of course, any ˆ S) is also minimal suﬃcient and complete. one-to-one function such as (B, Thus, from Rao-Blackwell/Lehmann-Scheﬀ´e theorems, among all unbiased ˆ and S have minimum variance. estimates, B Using the decomposition Y = PY + QY, where PQ = 0, the loglikelihood can be written as ' 1 & n l(B, Σ) = cte− ln |Σ|− tr Σ−1 [Y QY + (PY − XB) (PY − XB)] . 2 2 ˆ and Σ ˆ when Thus, to obtain the maximum likelihood estimates (MLE) B n−k ≥ p (for V ≡ (n−k)S = Y QY to be nonsingular w.p.1), we minimize 1 1 VΣ−1 + tr (PY − XB)Σ−1 (PY − XB) , n n ˆ = (X X)−1 X Y, so we and since the last term is ≥ 0, it is clear that B need only minimize ln |Σ| + tr

1 ln |Σ| + tr VΣ−1 . n

148

9. Multivariate regression

However, we already solved a similar problem in Chapter 7 when we derived the maximum likelihood estimates of the mean and variance of a multivariate normal distribution. Using the same result, we ﬁnd that ˆ = 1V Σ n is the maximum likelihood estimate of Σ. Proposition 9.2 gives the joint ˆ and S. distribution of B

ˆ Moreover, B

|=

ˆ Proposition 9.2 With underlying normality, the joint distribution of B and S is

ˆ ∼ N k B, (X X)−1 ⊗ Σ , B p (n − k)S ∼ Wp (n − k, Σ). S.

ˆ QY, which implies B

|=

and, thus, PY

|=

ˆ = (X X)−1 X Y Proof. Since Y ∼ Npn (XB, In ⊗ Σ), the distribution of B follows from Proposition 6.1. Next, since (n − k)S = Y QY = E QE, the distribution of (n − k)S is a direct consequence of Proposition 7.8. Since PQ = 0, we obtain immediately that P P var Y = ⊗ Ip (In ⊗ Σ) ((P, Q) ⊗ Ip ) Q Q P 0 P⊗Σ 0 = ⊗Σ= 0 Q 0 Q⊗Σ S.

2

9.3 The general linear hypothesis Consider now the problem of testing the general linear hypothesis H0 : CB = 0 against all alternatives where C ∈ Rrk of rank C = r in the multivariate regression model Y ∼ Npn (XB, In ⊗ Σ) with X ∈ Rnk of rank X = k. The likelihood ratio test will be more easily expressed using a “canonical” form for this problem.

9.3.1

Canonical form

The canonical form is obtained by transforming the original response Y, in two steps, so that in the new model X becomes X0 and C reduces to

9.3. The general linear hypothesis

C0 , where

X0 =

Ik 0

149

and

C0 = (Ir , 0).

Step 1 : This step is to reduce X to X0 . The Gram-Schmidt method applied to the columns of X (v. Problem 1.8.7) gives X = H1 U, where U ∈ U+ k and H1 H1 = Ik . There exists Γ1 ∈ Rnn−k such that (H1 , Γ1 ) ∈ On . Let ˜ = H1 Y, Y Γ1 ˜ In ⊗ Σ) with B ˜ the ˜ ∼ N n (X0 B, ˜ = UB. But since CB = CU−1 B, then Y p ˜B ˜ = 0, where C ˜ = CU−1 . hypothesis H0 : CB = 0 becomes H0 : C ˜ to C0 . Once again, the GramStep 2 : The second step is to reduce C ˜ ˜ = LH2 , where H2 H = Schmidt method applied to the rows of C yields C 2 k−r + Ir , L ∈ Lr . There exists Γ2 ∈ Rk such that H2 ∈ Ok . Γ2 Let

˜ ˜ = Y

H2 Γ2 0

0 ˜ Y, In−k

˜ ˜ ˜ ∼ N n (E Y, ˜ In ⊗ Σ), where then Y p

H2 ˜ ˜ ˜ = Γ2 B EY 0

or

˜ ˜ ˜1 ˜1 Y B ˜ ˜ ˜2 E Y ˜2 , ≡ B ˜ ˜ 0 Y3

˜ ˜ ˜˜ = ˜ was partitioned in conformity with B ˜ ∈ Rr and B ˜ 1 = H2 B where Y 2 p ˜ ∈ Rk−r . Now, to transform the hypothesis, note that Γ2 B p ˜ ˜ ˜ = LB ˜B ˜ = LH2 B ˜ 1 = 0. ˜ 1 = 0 ⇐⇒ B C ˜ ˜˜ are all ˜ 1 = 0. Because the rows of Y Hence, the hypothesis becomes H0 : B independent, an equivalent problem in its canonical form, with the obvious change of notation, is to test H0 : M1 = 0 against H1 : M1 = 0

150

9. Multivariate regression

based on Z1 Z2 Z3

∼ Nst (M1 , It ⊗ Σ), ∼ Nsu (M2 , Iu ⊗ Σ), ∼ Nsv (0, Iv ⊗ Σ), v ≥ s,

where Z1 , Z2 , and Z3 are independent.

9.3.2

LRT for the canonical problem

We can now obtain a simple expression for the likelihood ratio test. In the canonical model, the likelihood function for (M1 , M2 , Σ) is L(M1 , M2 , Σ) ∝ |Σ|−n/2 etr − 12 Σ−1 (Z1 − M1 ) (Z1 − M1 ) ·etr − 12 Σ−1 (Z2 − M2 ) (Z2 − M2 ) ·etr − 12 Σ−1 Z3 Z3 , where n = t + u + v. Note that in this form, the minimal and suﬃcient statistic is (Z1 , Z2 , Z3 Z3 ). For maximum likelihood estimates when v ≥ s (for Z3 Z3 to be nonsingular w.p.1), we minimize 2 − l(M1 , M2 , Σ) n

1 tr Σ−1 (Z1 − M1 ) (Z1 − M1 ) n 1 1 + tr Σ−1 (Z2 − M2 ) (Z2 − M2 ) + tr Σ−1 Z3 Z3 . n n

= cte + ln |Σ| +

Since each term is ≥ 0, it follows that the maximum likelihood estimates are ˆ 2 = Z2 , and Σ ˆ = Z Z3 /n. ˆ 1 = Z1 , M M 3 Also, when M1 = 0, the maximum likelihood estimates become ˆ ˆ ˆ ˆ 1 = 0, M ˆ 2 = Z2 , and Σ ˆ = (Z Z1 + Z Z3 )/n. M 1 3 Therefore, the LRT is the test which rejects H0 for small values of Λ=

ˆ ˆ ˆ ˆ −n/2 ˆ 2 , Σ) ˆ ˆ 1, M ˆ L(M |Z3 Z3 |n/2 |Σ| = . = ˆ 1, M ˆ 2 , Σ) ˆ ˆ −n/2 |Z1 Z1 + Z3 Z3 |n/2 L(M |Σ| d

|=

Deﬁnition 9.1 U-distribution: U ∼ U (p; m, n) iﬀ U = |W1 |/|W1 + W2 |, W2 , m + n > p. where W1 ∼ Wp (n), W2 ∼ Wp (m), and W1 Properties of U -distributions are deferred to Section 11.4. Going back to the original model, the likelihood ratio test can be exˆ Σ, ˆ and X. Composing the two transformations pressed in terms of B,

9.3. The general linear hypothesis

151

˜ ˜ we obtain ˜ and Y ˜ → Y, Y → Y ˜ ˜1 Y H2 H1 Y ˜ ˜ 2 = Γ2 H1 Y , Y ˜ Γ1 Y ˜3 Y and after long but straightforward calculations, the LRT is expressed as

Λ2/n

˜ ˜ ˜ ˜ Y |Y 3 3| ˜ ˜ ˜ ˜ ˜ ˜ Y ˜ ˜ |Y 1 1 + Y3 Y3 |

=

ˆ |nΣ|

=

ˆ+ |nΣ

ˆ C [C(X X)−1 C ]−1 CB| ˆ B

.

The null distribution of the LRT statistic follows directly from the canonical form of the model and the deﬁnition of U -distributions. Proposition 9.3 The null distribution of the LRT statistic Λ for testing H0 : CB = 0 against H1 : CB = 0, where C ∈ Rrk of rank C = r, in the model Y ∼ Npn (XB, In ⊗ Σ), with X ∈ Rnk of rank X = k, is Λ2/n ∼ U (p; r, n − k). When n is large, a simple approximation can be used for the null distribution of Λ. From the LRT general theory, we can immediately write −2 ln Λ → χ2pr , n → ∞. d

9.3.3

Invariant tests

The problem in its canonical form is to test H0 : M1 = 0 against H1 : M1 = 0

(9.2)

based on Z1 Z2 Z3

∼ Nst (M1 , It ⊗ Σ), ∼ Nsu (M2 , Iu ⊗ Σ), ∼ Nsv (0, Iv ⊗ Σ), v ≥ s,

where Z1 , Z2 , and Z3 are independent. Since v ≥ s, Z3 Z3 is nonsingular w.p.1 and let m ≡ min(s, t) = rank Z1 w.p.1. The group Gs × Rus × Ot × Ou × Ov transforms the variables as Z1 → H1 Z1 A, Z2 → H2 Z2 A + B, and Z3 → H3 Z3 A for any (A, B, H1 , H2 , H3 ) ∈ Gs × Rus × Ot × Ou × Ov . This induces the parameter transformations M1 → H1 M1 A, M2 → H2 M2 A + B, and Σ → A ΣA. Thus, we will say that a test function is invariant iﬀ f (Z1 , Z2 , Z3 ) = f (H1 Z1 A, H2 Z2 A + B, H3 Z3 A),

152

9. Multivariate regression

∀(A, B, H1 , H2 , H3 ) ∈ Gs × Rus × Ot × Ou × Ov , ∀(Z1 , Z2 , Z3 ) ∈ Rts × Rus × Rvs . The choice B = −H2 Z2 A shows that any invariant test does not depend on Z2 , f (Z1 , Z2 , Z3 ) = f (H1 Z1 A, 0, H3 Z3 A). Since rank Z3 = s w.p.1., then using Problem 1.8.7, there exists U ∈ U+ s and H ∈ Rvs satisfying H H = Is such that Z3 = HU. The choice A = U−1 G, where G ∈ Gs is arbitrary for now, yields f (Z1 , Z2 , Z3 ) = f (H1 Z1 U−1 G, 0, H3 HG). From the singular value decomposition (Proposition 1.11) there exists G ∈ Os and H1 ∈ Ot such that D 0 H1 (Z1 U−1 )G = , 0 0 where D is diagonal and contains the square root of the nonzero eigenvalues of (Z1 U−1 )(Z1 U−1 ) = Z1 (Z3 Z3 )−1 Z1 . We thus have D 0 , 0, H3 HG . f (Z1 , Z2 , Z3 ) = f 0 0 Finally, since (HG) (HG) = Is , the s columns of HG are orthonormal in Rv , and by completing to an orthonormal basis of Rv , there exists Γ such that (HG) H3 ≡ ∈ Ov Γ and

H3 HG =

Thus, altogether, we ﬁnd f (Z1 , Z2 , Z3 ) = f

D 0

Is 0

0 0

.

, 0,

Is 0

,

which shows that any invariant test depends on (Z1 , Z2 , Z3 ) only through the nonzero eigenvalues of Z1 (Z3 Z3 )−1 Z1 . Invariance permits a reduction of the parameter space also. Since the transformed parameters are M1 → H1 M1 A, M2 → H2 M2 A+B, and Σ → A ΣA, the choice B = −H2 M2 A shows that the non-null distribution of any invariant test is independent of M2 . Similarly, using the singular value decomposition, there exists G ∈ Os and H1 ∈ Ot such that D 0 H1 (M1 Σ−1/2 )G = , 0 0

9.3. The general linear hypothesis

153

where D is diagonal and contains the square root of the nonzero eigenvalues of (M1 Σ−1/2 )(M1 Σ−1/2 ) = M1 Σ−1 M1 . Thus, the choice A = Σ−1/2 G shows that the non-null distribution of any invariant test depends on (M1 , M2 , Σ) only through the nonzero eigenvalues of M1 Σ−1 M1 . We have proved: Proposition 9.4 For the hypothesis testing situation (9.2) and group of transformations described above, any invariant test depends on (Z1 , Z2 , Z3 ) only through the nonzero eigenvalues of Z1 (Z3 Z3 )−1 Z1 . Moreover, the nonnull distribution of any invariant test depends on (M1 , M2 , Σ) only through the nonzero eigenvalues of M1 Σ−1 M1 . The non-null distribution of those eigenvalues is complicated except in the case m ≡ min(s, t) = 1, where there is only one such eigenvalue. From Proposition 9.4, assume without loss of generality that D 0 t Z1 ∼ Ns , It ⊗ Is , 0 0 ∼ Nsv (0, Iv ⊗ Is ),

Z3

where D contains the square root of the nonzero eigenvalues of M1 Σ−1 M1 . (i) The case t = 1. Here, we have Z1 ∼ Ns ((M1 Σ−1 M1 )1/2 e1 , Is ) and Z3 Z3 ∼ Ws (v). Hence, from Proposition 8.2 on Hotelling’s test, the conclusion is Z1 (Z3 Z3 )−1 Z1 ∼ Fc (s, v − s + 1; M1 Σ−1 M1 /2). (ii) The case s = 1. Here, the distributions are Z1 ∼ Nt ((M1 M1 /σ 2 )1/2 e1 , It ), where Σ = σ 2 was set, and Z3 Z3 ∼ χ2v . Thus, by deﬁnition, Z1 Z1 ∼ Fc (t, v; M1 M1 /2σ 2 ). Z3 Z3 Another equivalent expression for the LRT is Λ2/n

=

|Z3 Z3 | + Z3 Z3 |

|Z1 Z1

= |Is + Z1 Z1 (Z3 Z3 )−1 |−1 m = (1 + li )−1 , i=1

154

9. Multivariate regression

where l1 ≥ · · · ≥ lm are the ordered nonzero eigenvalues of Z1 (Z3 Z3 )−1 Z1 . Thus, Λ takes on small values when those eigenvalues are large. Other possible tests could be used such as the following: m Lawley-Hotelling: T02 = i=1 li = tr Z1 Z1 (Z3 Z3 )−1 Pillai: Roy:

V =

m

li i=1 1+li

= tr Z1 Z1 (Z1 Z1 + Z3 Z3 )−1

l1 = largest eigenvalue of Z1 (Z3 Z3 )−1 Z1 .

None of these tests has a power function which dominates the others over the whole parameter space or even locally [Fujikoshi (1988)]. However, it is easy to see that the asymptotic (as v → ∞, s, t, and u are ﬁxed) null distribution of the three tests −2 ln Λ, vT02 , and vV is χ2st . From the LRT d

general theory, we already know −2 ln Λ → χ2st . Under H0 , we can assume p Σ = Is without loss of generality. From the law of large numbers, Z3 Z3 /v → Is ; hence, we have from Lemma 6.3 v tr Z1 Z1 (Z3 Z3 )−1 → tr Z1 Z1 = χ2st . d

d

The same argument applies to vV . The asymptotic null distribution of Roy’s test is quite diﬀerent and is given as Problem 9.7.5 together with an interpretation as a union-intersection test. In the very special case m ≡ min(s, t) = 1, these three tests are equivalent to the LRT, which is uniformly most powerful invariant (UMPI). The proof of this UMPI property is the same as for the Hotelling-T 2 test (v. Proposition 8.4) since the non-null distribution in both cases (i) and (ii) above is a noncentral canonical Fc distribution. Kariya et al. (1987) considered hypotheses related to selection and independence under multivariate regression models. Breiman and Friedman (1997) presented several methods of predicting responses in a multivariate regression model. The likelihood ratio test for detecting a single outlier (a shift in the mean) in a multivariate regression model was obtained by Srivastava and von Rosen (1998).

9.4 Random design matrix X When the prediction variables X, just as the dependent variables Y, are observed, then it is appropriate to consider X as a random matrix. The model most commonly encountered assumes Y = XB + E, E ∼ Npn (0, In ⊗ Σ),

9.4. Random design matrix X

155

|=

where the errors are independently distributed of the prediction variables, i.e., E X. When X has an absolutely continuous distribution, the argument in the proof of Proposition 7.5 shows that X X is nonsingular w.p.1. The conditional model Y|X ∼ Npn (XB, In ⊗ Σ) is thus identical to the case of a ﬁxed X. Using Proposition 9.2, we ﬁnd the following properties of the same estimates: ˆ is unbiased, (i) B ˆ = E E(B|X) ˆ EB = E B = B. (ii) (n − k)S ∼ Wp (n − k, Σ). Indeed, (n − k)S|X ∼ Wp (n − k, Σ)

ˆ (iii) B

|=

and this conditional distribution does not depend on X. S. With Proposition 2.13, ˆ · h(S) E g(B)

ˆ = E E[g(B)h(S)|X] ˆ = E{E[g(B)|X]E[h(S)|X]} ˆ = E{E[g(B)|X]E[h(S)]} ˆ · E h(S). = E g(B)

Moreover, for testing the general linear hypothesis H0 : CB = 0, the conditional null distribution of Λ2/n does not depend on X and, thus, Λ2/n ∼ U (p; r, n − k), unconditionally. The non-null distribution of Λ2/n , however, will depend on the distribution of X, as is the case for p = 1 as exempliﬁed by Problem 5.7.8, where the noncentrality parameter of the distribution of the F -test depends on X. ˆ may be evaluated as Example 9.1 The variance of B ˆ var B

ˆ ) = var vec(B ˆ )|X] ˆ )|X] + var E[vec(B = E var[vec(B = E[(X X)−1 ⊗ Σ] + var vec(B )

= E (X X)−1 ⊗ Σ.

The last expectation may be evaluated directly in some cases. For example, if X ∼ Nkn (0, In ⊗ Ω), Ω > 0, then with Problem 7.5.5 and since X X ∼ Wk (n, Ω), E (X X)−1 = (n − k − 1)−1 Ω−1 .

156

9. Multivariate regression

9.5 Predictions

Y = XB + E, E ∼ Npn (0, In ⊗ Σ), X

|=

The problem of predicting several, possibly correlated, responses from the same set of predictors is becoming increasingly important. Applications by Breiman and Freidman (1997) include prediction of changes in the valuations of stocks in 60 industry groups by using over 100 econometric variables as predictors. Or, in chemometrics, the prediction of 6 output characteristics of the polymers produced as predicted by 22 predictor variables. Another example by Brown (1980, pp. 247-292) lists electoral results for all 71 Scottish constituencies in the Bristish general elections of February and October 1974. Data consist of total votes for each of the four parties (Conservative, Labour, Liberal, and Nationalist) in each election, together with a categorical variable listing the location of the constituency by six regions, and the size of the electorate in each constituency. The objective is to use the February and October results from part of the constituencies to predict the remaining October results from the corresponding February data. Research papers related to predictions include Stone (1974), van der Merwe and Zidek (1980), Bilodeau and Kariya (1989), and Breiman and Friedman (1997). Assume the “centered” model E.

For the sake of simplicity, we take X centered, X ∼ Nkn (0, In ⊗ Ω). For given values of the prediction variables, x = (x1 , . . . , xk ), it is desired to obtain a prediction of the dependent variables, y = (y1 , . . . , yp ). Using the ˆ an obvious prediction method is Gauss-Markov (GM) estimate B, ˆ = (x βˆ1 , . . . , x βˆp ) ˆ = x B y

|=

|=

so that prediction of the ith variable is done considering only the ith multiple regression model. Assuming the “future” observation follows the same e, model, i.e., y = x B + e , where x ∼ Nk (0, Ω), e ∼ Np (0, Σ), and x (X, E), one can evaluate the risk and is independent of the past, (x, e) of the GM prediction as y − y) E (ˆ y − y) Σ−1 (ˆ

ˆ − B) − e ]Σ−1 [(B ˆ − B) x − e] = E [x (B ! ˆ − B)Σ−1 (B ˆ − B) Ω + p = E tr (B ! ˆ − B)Σ−1 (B ˆ − B) Ω + p. = tr E (B

The SPER (sum of Squares of Prediction Error when the independent variable is Random) risk is obtained on subtracting p from the above: ˆ = E tr (B ˆ − B)Σ−1 (B ˆ − B) Ω. RSPER (B)

9.5. Predictions

157

ˆ − B)Σ−1/2 ∼ N k (0, (X X)−1 ⊗ I), Example 6.3 gives Letting U = (B p E(UU |X) = p(X X)−1 . Finally, with Example 9.1, we get ˆ = pk/(n − k − 1). RSPER (B) A closely related risk function, more tractable mathematically, is SPE deﬁned by ˆ = E tr (B ˆ − B)Σ−1 (B ˆ − B) X X. RSPE (B) ˜ = BA ˆ Smaller risk may be achieved with an estimate B for a certain ˜ = A y ˆ is seen A ∈ Rpp . The corresponding prediction for each variable in y to be a linear combination (multivariate ﬂattening) of the p prediction equations. Example 9.2 Multivariate ﬂattening. Assuming (B, Σ, Ω) is known, then an optimal multivariate ﬂattening [Breiman and Friedman (1997)] would be solution of ˆ − y) Σ−1 (A y ˆ − y). min E (A y A

Now, since ˆ − y) Σ−1 (A y ˆ − y) = (Σ−1/2 A y ˆ − Σ−1/2 y) (Σ−1/2 A y ˆ − Σ−1/2 y), (A y letting C = Σ−1/2 A and z = Σ−1/2 y, the optimization problem becomes equivalent to y |2 . min E |z − Cˆ C

With Problem 5.7.6, the solution is readily obtained: ˆ ) · [var y ˆ ]−1 . C = cov(z, y Each factor is evaluated as ˆ) cov(z, y

ˆ) = Σ−1/2 cov(y, y −1/2 = Σ E yˆ y ˆ = Σ−1/2 E (B x + e)x B −1/2 ˆ E B xx B = Σ = Σ−1/2 B ΩB

and, similarly, ˆ var y

ˆy ˆ = E y = E [B + (X X)−1 X E] xx [B + (X X)−1 X E] = B ΩB + E E X(X X)−1 xx (X X)−1 X E

= B ΩB + E x (X X)−1 x Σ (v. P roblem 6.4.5) = B ΩB + [k/(n − k − 1)]Σ (v. P roblem 7.5.4).

158

9. Multivariate regression

Hence, altogether, the optimal A is given by A = [(B ΩB) + rΣ]−1 (B ΩB), r = k/(n − k − 1). ˆ and modiﬁcations thereof are given in the above papers. Sample-based A In particular, for small r, A ≈ I − r(B ΩB)−1 Σ (v. Problem 1.8.15) and van der Merwe and Zidek (1980) established that the sample-based ˆ −1 S ˆ = I − r(n − k)(B ˆ X XB) A which they called FICYREG (FIltered Canonical Y REGression) leads to smaller SPE risk than GM for r = (k − p − 1)/(n − k + p + 1) provided n > k > p + 1. Bilodeau and Kariya (1989) proposed the modiﬁed Efron-Morris (1976) ˆ −1 S − b(n − k)S/tr(B ˆ X XB) ˆ ˆ = I − r(n − k)(B ˆ X XB) A and showed that it leads still to a smaller SPE risk than FICYREG for r = (k − p − 1)/(n − k + p + 1) and b = (p − 1)/(n − k + p + 1). Note that the choice b = 0 reduces to FICYREG. Breiman and Friedˆ built from cross-validation (CV) and generalized man (1997) considered A cross-validation (GCV). Their large-scale simulations point strongly toward the superiority of CV and GCV over other commonly used prediction techniques. The GCV in particular seems very promising since its evaluation is nearly as simple as GM. The CV, in contrast, is computationally intensive. The unbiased estimate of the SPE risk for the GCV predictions was recently obtained by Bilodeau (1998).

9.6 One-way classiﬁcation In this section, the one-factor univariate analysis of variance is generalized to test the equality of several means of multivariate normal populations. Let yi1 , . . . , yini i.i.d. Np (µi , Σ), i = 1, . . . , a, be a independent samples from multivariate normal distributions with common variance Σ > 0. In matrix notation, let y11 .. . y1n µ1 1 . .. Y = .. , X = diag(1n1 , . . . , 1na ), B = . . y µa a1 . .. yan a

9.7. Problems

159

Then, the a samples can be written as the multivariate regression model

where n =

a i=1

Y = XB + E, E ∼ Npn (0, In ⊗ Σ), ni . The hypothesis of equality of means H0 : µ1 = · · · = µa

can be translated into a general linear hypothesis. Deﬁne C = (Ia−1 , −1a−1 ); then the hypothesis becomes H0 : CB = 0, where C ∈ Raa−1 of rank C = a−1. Using the canonical formulation of this problem, the reader can verify that the LRT is |SSw | , Λ2/n = |SSw + SSb | where SSw SSb

ni a

=

¯ i )(yij − y ¯ i ) , (yij − y

i=1 j=1 a

¯ )(¯ ¯ ) ni (¯ yi − y yi − y

=

i=1

are the usual “within” and “between” sums of squares with ¯i = y

ni

¯= yij /ni and y

j=1

a

¯ i /n. ni y

i=1

The other analysis-of-variance models such as the two-way classiﬁcation model can be generalized similarly to test the eﬀect of each factor or the presence of interactions between factors. We will not pursue this any further here.

9.7 Problems 1. Show that the likelihood ratio test statistic Λ for testing the general linear hypothesis can be written Λ2/n =

ˆ |nΣ| ˆ ˆ +B ˆ C [C(X X)−1 C ]−1 CB| |nΣ

.

ˆ = PY is the orthogonal projection of Y on V = 2. The estimate XB ˆ is also the solution {XA : A ∈ Rkp }. Use this fact to prove that XB of the least-squares problem min tr Ω(Y − V) (Y − V),

V∈V

160

9. Multivariate regression

for any ﬁxed Ω ∈ Rpp , Ω > 0. |=

3. Prove that if Z1 ∼ Nst (M1 , It ⊗Σ), Z3 ∼ Nsv (0, Iv ⊗Σ), and Z1 Z3 , where Σ > 0 and v ≥ s + 2, then t 1 Is + M M1 Σ−1 . E Z1 Z1 (Z3 Z3 )−1 = v−s−1 v−s−1 1 4. Using the canonical model, prove that the LRT Λ for the hypothesis of the equality of several multivariate means is Λ2/n = |SSw |/|SSw + SSb |, as described in Section 9.6. What is the null distribution of this test? 5. The general linear hypothesis in its canonical form is to test H0 : M1 = 0 against H1 : M1 = 0 based on Z1

∼ Nst (M1 , It ⊗ Σ),

Z2

∼ Nsu (M2 , Iu ⊗ Σ),

Z3

∼ Nsv (0, Iv ⊗ Σ), v ≥ s,

where Z1 , Z2 , and Z3 are independent. (i) Prove the asymptotic result as v → ∞ concerning Roy’s test: d

If M1 = 0 then vl1 → α1 , where l1 is the largest eigenvalue of Z1 (Z3 Z3 )−1 Z1 and α1 is the largest eigenvalue of a random matrix W ∼ Ws (t). (ii) Union-intersection test. (a) For a given h ∈ Rt , |h| = 1, deﬁne Hh,0 : M1 h = 0 and Hh,1 : M1 h = 0. Prove H0 H1

= ∩h {Hh,0 : |h| = 1}, = ∪h {Hh,1 : |h| = 1}.

(b) For a given h, |h| = 1, prove that the LRT for Hh,0 against Hh,1 accepts Hh,0 for small values of Rh = h Z1 (Z3 Z3 )−1 Z1 h. Demonstrate the null distribution Rh ∼ Fc (s, v −s+1) does not depend on h. (c) The union-intersection test accepts H0 iﬀ sup|h|=1 Rh ≤ c for some constant c. Demonstrate the union-intersection test statistic sup|h|=1 Rh = l1 is, in fact, Roy’s test. Remark: For a given h, the test based on Rh is UMPI for testing Hh,0 against Hh,1 (the non-null distribution of Rh is a noncentral canonical Fc distribution just like Hotelling’s-T 2 ; v. Proposition 8.4), but Roy’s test is not generally UMPI for testing H0 against H1 .

10 Principal components

10.1 Introduction In this chapter we assume that x ∈ Rp with E x = µ and var x = Σ = (σij ). When the dimension p is large, the principal components method seeks to replace x by y ∈ Rk , where k < p (and hopefully much smaller), without losing too much “information.” This is sometimes particularly useful for a graphical description of the data since it is much easier to view vectors of low dimension. Section 10.2 deﬁnes principal components and gives their interpretation as normalized linear combinations with maximum variance. In Section 10.3, we explain an optimal property of principal components as best approximating subspace of dimension k in terms of squared prediction error. Section 10.4 introduces the sample principal components; they give the coordinates of the projected data which is closest, in terms of euclidian distance, to the original data. Section 10.5 treats the sample principal components calculated from the correlation matrix. Finally, Section 10.6 presents a simple test for multivariate normality which generalizes the univariate Shapiro and Wilk’s statistic. A book entirely devoted to principal component analysis is that of Jolliﬀe (1986).

162

10. Principal components

10.2 Deﬁnition and basic properties The total variance of x is deﬁned as p p E |x − µ|2 = var xi = σii = tr Σ. i=1

i=1

Recall that Σ ≥ 0 can be written as Σ = HDH , where H = D =

(h1 , . . . , hp ) ∈ Op , diag(λ1 , . . . , λp ),

and λ1 ≥ · · · ≥ λp are the ordered eigenvalues of Σ. Since we are only interested in var x, we will assume throughout this chapter that µ = 0. If we let h1 x y = H x = ... , hp x

p p var y = D. Then i=1 var yi = i=1 λi = tr Σ, so x and y have the same “total variance.” Moreover, the variables yi ’s are uncorrelated, cov(hi x, hj x) = hi Σhj = λj hi hj = λj δij . Deﬁnition 10.1 The variables yi = hi x, i = 1, . . . , p, are, by deﬁnition, the principal components of x. p p Since HH = I, then x = ( i=1 hi hi )x = i=1 yi hi and the principal components can be viewed as the coordinates of x withrespect to the k orthonormal basis {h1 , . . . , hp } of Rp . When the ratio i=1 λi /tr Σ is close to 1, then (y1 , . . . , yk ) can eﬀectively replace x without losing much in terms of “total variance.” The principal components can also be got sequentially as follows. First a normalized linear combination t x, |t| = 1, is sought such that var t x = t Σt is maximum. Since for all t, |t| = 1, p p p 2 2 λi (t hi ) ≤ λ1 (t hi ) = λ1 t hi hi t = λ1 |t|2 = λ1 ; t Σt = i=1

i=1

i=1

hence, maxt t=1 t Σt = λ1 , which is attained for t = h1 . So, the ﬁrst principal component y1 = h1 x is the normalized linear combination with maximum variance. Now, given yi = hi x, i = 1, . . . , k, another linear combination s x, |s| = 1, is sought which maximizes the variance s Σs and is uncorrelated with y1 , . . . , yk . Note that cov(s x, yi ) = λi s hi , i = 1, . . . , k. As above, for all s ⊥ h1 , . . . , hk , |s| = 1, we have s Σs =

p i=k+1

λi (s hi )2 ≤ λk+1

p i=k+1

(s hi )2 = λk+1

p i=1

(s hi )2 = λk+1 .

10.3. Best approximating subspace

163

Hence, max

s s=1 s⊥h1 ,...,hk

s Σs = λk+1

is attained for s = hk+1 , which means that yk+1 = hk+1 x is the normalized linear combination with maximum variance among all those uncorrelated with y1 , . . . , yk .

10.3 Best approximating subspace The orthogonal projection of x on the subspace spanned by the ﬁrst k eigenvectors, Pk x, is k k hi hi x = yi hi . Pk x = i=1

i=1

Proposition 10.1 shows that Pk x gives the best approximation to x by a subspace of dimension at most k in terms of squared prediction error. Before stating the result, we present a lemma. Denote by Pk⊥ the set of all orthogonal projections P ∈ Rpp of rank P = k. Lemma 10.1 Let Σ ≥ 0 in Rpp with eigenvalues λ1 ≥ · · · ≥ λp . Then, k

max tr ΣP =

P∈Pk⊥

min tr Σ(I − P)

P∈Pk⊥

are attained at P =

k i=1

λi ,

i=1 p

=

λi

i=k+1

hi hi , where

Σ = HDH , H = (h1 , . . . , hp ) ∈ Op , D = diag(λ1 , . . . , λp ). Proof. Take any P ∈ Pk⊥ . Let A = (a1 , . . . , ak ) whose columns form an orthonormal basis for Im P, then P = AA . Now, tr ΣP = tr HDH AA = tr D(H A)(H A) , and note that G = (g1 , . . . , gk ) ≡ H A has orthonormal columns too, i.e., G G = Ik . Therefore, tr ΣP =

tr D

k

gi gi =

i=1

≤

k i=1

max

g g=1 g⊥g1 ,...,gi−1

k

gi Dgi

i=1

g Dg =

k i=1

λi

164

10. Principal components

(when i = 1 the orthogonality condition is void) with equality if gi = ei , which means A = HG = (h1 , . . . , hk ). This shows the ﬁrst part related to the maximum. The second part is immediate. 2 Proposition 10.1 Assume x ∼ Np (0, Σ), Σ > 0, and let B ∈ Rkp of rank B = k, C ∈ Rpk . Then, min E |x − CBx|2 = B,C

p

λi

i=k+1

is attained when CB = Pk . Proof. Fix B. We have x 0 Σ ∼ Np+k , Bx 0 BΣ

ΣB BΣB

and x | Bx ∼ Np (0, Σ − ΣB (BΣB )−1 BΣ). Using Problem 5.7.6, min E |x − CBx|2 = tr Σ − ΣB (BΣB )−1 BΣ C = tr Σ I − AB (BΣB )−1 BA = tr Σ(I − P), where A = HD1/2 H and P = AB (BΣB )−1 BA, and the extremum is reached at C = (ΣB )(BΣB )−1 . Now, P is an orthogonal projection of rank k. From Lemma 10.1, tr Σ(I − P) is minimized when AB (BΣB )−1 BA =

k

hi hi

i=1

or B (BΣB )−1 B = Finally, CB = Σ[B (BΣB )−1 B] =

k

k

λ−1 i hi hi .

i=1

i=1

hi hi = Pk .

2

Obviously, if µ = 0 in Proposition 10.1, the best approximation of rank k is Pk (x − µ) + µ, which represents the orthogonal projection of x on the aﬃne subspace span{h1 , . . . , hk } + µ.

10.4 Sample principal components from S The variance Σ is usually unknown. Sample principal components can be obtained from the estimate S = V/m, m = n − 1, where, as usual, V=

n i=1

¯ )(xi − x ¯ ) . (xi − x

10.4. Sample principal components from S

165

Since S ≥ 0, write ˆ , ˆ diag(l1 /m, . . . , lp /m) H S=H where ˆ p ) ∈ Op ˆ 1, . . . , h ˆ = (h H and l1 ≥ · · · ≥ lp are the ordered eigenvalues of V. The sample principal ˆ x, i = 1, . . . , p. components of x are deﬁned as h i p Let V ⊂ R be a k-dimensional subspace and denote by V + a = {x + a : x ∈ V} the corresponding aﬃne subspace. What is the aﬃne subspace V +a of dimension k such that the orthogonal projection of the data on V + a is “closest” to the original data? First, we must specify what is meant by “closest.” As a measure of distance, take the usual euclidian distance d(V, a) =

n

ˆ i |2 , |xi − x

i=1

ˆ i − a) + a is the orthogonal projection of xi on V + a. ˆ i = P(x where x Proposition 10.2 Among all k-dimensional subspaces V and vectors a ∈ ˆ 1, . . . , h ˆ k }. ¯ and V = span{h Rp , the distance d(V, a) is minimized for a = x n ˆ the orthogonal projection on ¯ )(xi − x ¯ ) , P Proof. Deﬁne V = i=1 (xi − x ˆ ˆ V, and Q = I − P. Then, d(V, a)

=

n

ˆ i − a) − a|2 |xi − P(x

i=1

=

n

ˆ i − Qa| ˆ 2 |Qx

i=1

=

n

ˆ i − Q¯ ˆ x) + (Q¯ ˆ x − Qa)| ˆ 2 |(Qx

i=1

=

n i=1

=

ˆ i − Q¯ ˆ x|2 + |Qx

n

ˆ x − Qa| ˆ 2 |Q¯

i=1

ˆ + n(¯ ˆ x − a). tr QV x − a) Q(¯

¯ . Also, The two terms in the last expression are non-negative; hence, a = x ˆ ih ˆ . ˆ = k h 2 from Lemma 10.1 and since V ∝ S, P i i=1 p k by the The ratio f (λ) = i=1 λi / i=1 λi of total variance explained p k ﬁrst k principal components is estimated by f (l/m) = i=1 li / i=1 li . A large sample (1 − α) × 100% conﬁdence interval on this ratio f (λ), when all population eigenvalues λα are distinct, can be constructed with Proposition 8.18.

166

10. Principal components

We end this section with a word of caution: Principal components are not invariant with respect to individual rescaling of the p variables in x; that is, if w = Φx, where Φ = diag(φ1 , . . . , φp ), then ΦΣΦ does not have the same eigen-structure as Σ. This means, for example, that the interesting projections of the data found with Proposition 10.2 may look entirely diﬀerent after rescaling. Also, if the ﬁrst variable x1 has a variance much larger than the variances of all other variables, x2 , . . . , xp , then the ﬁrst principal component y1 will be approximately equivalent to x1 . Principal components are thus most meaningful when all variables are measured in the same units and have variances of the same magnitude. For this reason, principal components are often calculated from the sample correlation matrix R rather than the sample variance S.

10.5 Sample principal components from R If we let S0 Σ0

= =

diag(s11 , . . . , spp ), diag(σ11 , . . . , σpp ),

then the population and sample correlation matrices are given by −1/2

R = S0 ρ =

−1/2

SS0

,

−1/2 −1/2 Σ0 ΣΣ0 .

Then, as in the previous section, we can decompose ρ = G diag(γ1 , . . . , γp ) G , ˆ diag(f1 , . . . , fp ) G ˆ , R = G and deﬁne the sample principal components from the standardized vari−1/2 ¯ ), and R as g ˆi z, i = 1, . . . , p, where G = (g1 , . . . , gp ) ables, z = S0 (x − x ˆ and similarly for G. The ratio of total variance (of the standardized vari√ ¯i )/ sii ) explained by the ﬁrst k principal components ables zi = (xi − x becomes f (γ) =

k

γi /p.

i=1

The construction of a conﬁdence interval on this ratio f (γ) thus necessitates the asymptotic distribution of the eigenvalues fi of the sample correlation matrix R. This is now derived using the perturbation method of Section 8.8. Using the Taylor series x−1/2 = a−1/2 − 12 a−3/2 (x − a) + · · · ,

10.5. Sample principal components from R

167

we have directly −1/2

S0

−1/2

= [I − 12 Σ0

−1/2

(S0 − Σ0 )Σ0

−1/2

+ · · ·]Σ0

.

Deﬁne −1/2

V

=

(vij ) = n1/2 (Σ0

V0

=

diag(v11 , . . . , vpp ),

−1/2

SΣ0

− ρ),

and note that V is Op (1). Then, we can write R = =

−1/2

[I − 12 n−1/2 V0 + · · ·]Σ0

−1/2

[Σ + (S − Σ)]Σ0

[I − 12 n−1/2 V0 + · · ·]

[I − 12 n−1/2 V0 + · · ·](ρ + n−1/2 V)[I − 12 n−1/2 V0 + · · ·]

= ρ + n−1/2 (V − 12 ρV0 − 12 V0 ρ) + Op (n−1 ), from which G RG = Γ + n−1/2 V(1) + Op (n−1 ), where Γ = V(1)

diag(γ1 , . . . , γp ), (1)

(vij ) = G (V − 12 ρV0 − 12 V0 ρ)G.

=

(10.1)

Equation (8.11) in the perturbation method then leads, assuming γα to be a distinct eigenvalue, to the expansion (1) fα = γα + n−1/2 vαα + Op (n−1 ),

or, in vector form, assuming all eigenvalues γα to be distinct, to the expansion n1/2 (f − γ)

=

(1)

(1) (v11 , . . . , vpp ) + Op (n−1/2 )

≡ v(1) + Op (n−1/2 ). Now, since V is asymptotically normal with mean 0, so is V(1) and its marginal v(1) . We need only calculate the asymptotic variance of v(1) . From (10.1) and the relation ρG = GΓ, we have (1) vαα

= gα Vgα − γα gα V 0 gα p p p 2 = gjα gkα vjk − γα gjα vjj ; j=1 k=1

j=1

hence, (1) (1) cov(vαα , vββ )

=

p p p p j=1 k=1 i=1 l=1 p p

+γα γβ

j=1 i=1

gjα gkα giβ glβ cov(vjk , vil ) 2 2 gjα giβ cov(vjj , vii )

168

10. Principal components

−γα

p p p

2 gjα giβ glβ cov(vjj , vil )

j=1 i=1 l=1

−γβ

p p p

2 giβ gjα gkα cov(vii , vjk ).

i=1 j=1 k=1 d

Since V → Npp (0, (I + Kp )(ρ ⊗ ρ)), we ﬁnd upon using (6.1) that (1)

(1) lim cov(vαα , vββ )

n→∞

=

p p p p j=1 k=1 i=1 l=1 p p

+γα γβ

gjα gkα giβ glβ (ρkl ρji + ρjl ρki ) 2 2 gjα giβ 2ρ2ji

j=1 i=1 p p p

−γα

2 gjα giβ glβ 2ρjl ρji

j=1 i=1 l=1

−γβ

p p p

2 giβ gjα gkα 2ρik ρij .

i=1 j=1 k=1

Finally, with the simple relations p p

gkα glβ ρkl

= γα δαβ ,

k=1 l=1 p

glβ ρjl

= γβ gjβ ,

l=1

we obtain the simpliﬁcation (1)

(1) , vββ ) = 2γα γβ δαβ − (γα + γβ ) lim cov(vαα

n→∞

p

2 2 gjα gjβ

j=1

+

p p

2 2 gjα giβ ρ

j=1 i=1

We summarize the result. Proposition 10.3 Let f = (f1 , . . . , fp ) be the eigenvalues of the sample correlation matrix R. If the eigenvalues γα of the population correlation matrix ρ are all distinct, then the joint limiting distribution is n1/2 (f − γ) → Np (0, Ω), d

10.6. A test for multivariate normality

where Ω = (ωαβ ) is given by ωαβ = 2γα γβ δαβ − (γα + γβ )

p

2 2 gjα gjβ +

j=1

p p

169

2 2 2 gjα giβ ρji .

j=1 i=1

k The limiting distribution of a function such as f (f ) = i=1 fi /p for the ratio of total variance explained by the ﬁrst k principal components is easily derived by the delta method [v. Problem 10.7.5]. Problem 13.6.19 provides the asymptotic distribution of n1/2 (f − γ) when sampling from an elliptical distribution. Konishi (1979) obtained, with Sugiura’s lemma, a more accurate approximation with remainder O(n−1 ), similar to that of Proposition 8.18, for the distribution function of s = (n − 1)1/2 (f (f ) − f (γ)) , where f (·) is a continuously diﬀerentiable function in a neighborhood of γ.

10.6 A test for multivariate normality Shapiro and Wilk’s (1965) W statistic has been found to be the best omnibus test for detecting departures from univariate normality. Royston (1983) extends the application of W to testing multivariate normality, but the procedure involves a certain approximation which needs to be justiﬁed. The procedure of Srivastava and Hui (1987) does not require such an approximation and has a simple asymptotic null distribution and the calculations are straightforward. Srivastava and Hui (1987) proposed two test statistics for testing multivariate normality. These are based on principal components and may be considered as a generalization of the Shapiro-Wilk statistic. As in Section 10.4, write ˆ , m = n − 1, ˆ diag(l1 /m, . . . , lp /m) H S=H where ˆ p ) ∈ Op . ˆ 1, . . . , h ˆ = (h H The sample principal components of xj , j = 1, . . . , n, are deﬁned as ˆ xj , i = 1, . . . , p, j = 1, . . . , n. Thus, under the null hypothesis of yij = h i multivariate normality, we can treat yi1 , . . . , yin , i = 1, . . . , p, as p approximately independent samples. For sample i, the univariate Shapiro-Wilk statistic is deﬁned as 2 n m aj yi(j) , i = 1, . . . , p, W (i) = nli j=1

170

10. Principal components

where aj ’s are the constants tabulated in Shapiro and Wilk (1965) and yi(1) ≤ yi(2) ≤ · · · ≤ yi(n) are the ordered values of yi1 , . . . , yin . For n > 50, the values of aj are given by Shapiro and Francia (1972) and up to 2000 by Royston (1982). From Shapiro and Wilk (1968), we note that for each i, W (i) can be transformed to an approximate standard normal variable G(W (i)) by using Johnson’s (1949) SB system, W (i) − , G(W (i)) = γ + δ ln 1 − W (i) where γ, δ, and can be found in Table 1 of Shapiro and Wilk (1968) up to n = 50. For n > 50, values of γ, δ, and can be obtained with the help of the results in Shapiro and Francia (1972) and Royston (1982). Let M1 = −2

p

ln [Φ (G(W (i)))] ,

i=1

where Φ(·) is the distribution of a standard normal variable. Note that if U ∼ unif(0, 1), then −2 ln U ∼ χ22 . Srivastava and Hui (1987) proposed M1 as their ﬁrst test statistic for testing multivariate normality, where M1 is approximately distributed as χ22p under the hypothesis of multivariate normality. Large values of M1 will indicate non-normality. Next, they observed that small values of W (i) indicate a departure from normality for variate i. Thus, they considered the minimum of all components and proposed M2 = min W (i) 1≤i≤p

as the second test statistic. The null distribution of M2 is approximately given by p

P (M2 ≤ t) = 1 − [1 − Φ (G(t))] .

(10.2)

For p =2, 4, and 6 and n =10, 25, and 50, a simulation study [Srivastava and Hui (1987)] found that the null distribution of both M1 and M2 are well approximated by χ22p and (10.2), respectively. Examples of the use of M1 and M2 on data sets are provided by Looney (1995) with the necessary SAS procedures or FORTRAN subroutines. Most tests for multivariate normality are functions of the squared radii ¯ ), (or squared Mahalanobis distances of xi to x ¯ ) S−1 (xi − x ¯ ), i = 1, . . . , n. d2i = |zi |2 = (xi − x Some graphical procedures [Andrews et al. (1973), Cox and Small (1978), Gnanadesikan and Kettenring (1972)] are based on d2i . One such Q-Q plot is described in Section 11.4.1. Malkovich and Aﬁﬁ (1973) considered

10.6. A test for multivariate normality

171

the supremum of the standardized skewness and kurtosis over all linear combinations t x, {E (t x − t µ)3 }2 , (t Σt)3 t∈S E (t x − t µ)4 = max − 3 . 2 p−1 (t Σt) t∈S

β1M

=

β2M

max p−1

The tests are based on the sample versions M β1,n

=

M β2,n

=

max b1,n (t),

t∈S p−1

max |b2,n (t) − 3|,

t∈S p−1

respectively, where b1,n (t)

=

b2,n (t)

=

{ n1 1 n

n

n

¯ 3 2 ) } i=1 (t xi − t x , 3 (t St)

¯ 4 ) i=1 (t xi − t x . 2 (t St)

Mardia’s kurtosis test [Mardia (1970)] is a function of d2i and his skewness test is a function of the scaled residuals ¯ ), i = 1, . . . , n. zi = S−1/2 (xi − x Mardia’s measures of multivariate skewness and kurtosis are n 1 ¯ ) S−1 (xj − x ¯ )}3 , B1,n = {(xi − x n2 i,j=1 1 ¯ ) S−1 (xi − x ¯ )}2 , {(xi − x n i=1 n

B2,n

=

respectively. The tests of multivariate normality based on multivariate M and B1,n , are inconsistent against each ﬁxed non-normal skewness, β1,n elliptical distribution [Baringhaus and Henze (1991)]. However, the tests M and B2,n , are consistent. An approxibased on multivariate kurtosis, β2,n M against elliptically symmetric mation formula of the power of the test β2,n distributions was derived by Naito (1998). Cox and Small (1978) proposed tests based on linearity of regression rather than directly on normality. An omnibus test based on empirical characteristic function of the scaled residuals was also proposed [Henze and Zirkler (1990), v. also Henze and Wagner (1997)]. Goodness-of-ﬁt tests for a general multivariate distribution by the empirical characteristic function was treated by Fan (1997). A characterization of multivariate normality by hermitian polynomials was recently proposed by Kariya et al. (1997) to build an omnibus test. A comparative study of goodness-of-ﬁt tests for multivariate normality was carried out by Romeu and Ozturk (1993).

172

10. Principal components

10.7 Problems 1. In morphometric studies, it is often the case that all variables are positively correlated. Prove that if Σ has all positive covariances, σij > 0 for i = j, then all the coeﬃcients in h1 of the ﬁrst principal component may be taken non-negative. 2. For Σ ≥ 0 in Rpp with spectral decomposition Σ = HDH as in k Section 10.2, prove that Θ = i=1 λi hi hi is the matrix of rank k such that p p (σij − θij )2 ||Σ − Θ||2 = i=1 j=1

is minimum. Hint: ||Σ − Θ||2 = tr(D − E)(D − E) , where E = H ΘH. 3. Assume x ∈ Rp has density fx (x) = |Λ|−1/2 g[(x − µ1) Λ−1 (x − µ1)], where Λ = σ 2 [(1 − ρ)I + ρ11 ]. Prove there exists H ∈ Op such that y = Hx has density p y2 (y1 − p1/2 µ)2 −1/2 −(p−1)/2 fy (y) = λ1 λ2 g + i=2 i . λ1 λ2 4. Parent-child interclass correlation [Srivastava (1984)]. Assume x ∈ Rp+1 has density fx (x) = |Σ|−1/2 exp[(x − µ) Σ−1 (x − µ)], where

µ = Σ =

µm µs 1

,

2 σms 1 σm 2 σms 1 σs [(1 − ρss )I + ρss 11 ]

.

Here, “m” stands for mother and “s” means siblings. Let A = (1/p, Γ ) ∈ Rp for some Γ satisfying Γ1 = 0 and ΓΓ = Ip−1 . 2 (i) Interpret the parameters (µm , µs , σm , σms , ρss ). (ii) Prove that if 1 0 ˜ , A= 0 A

˜ A ˜ = diag(Ω, γ 2 Ip−1 ), where then AΣ s 2 σm σms Ω = , σms η 2

10.7. Problems

γs2 η2

173

= σs2 (1 − ρss ), =

[1 + (p − 1)ρss ]σs2 /p. |=

˜ is such that (y1 , y2 ) ∼ N2 ((µm , µs ) , Ω), (iii) Deduce that y = Ax 2 (y3 , . . . , yp+1 ). yi ∼ N (0, γs ) (i = 3, . . . , p + 1), and (y1 , y2 ) (iv) What are the implications for maximum likelihood estimation of the unknown parameters in i)? Remark: The yi ’s are not the principal components but are closely related to the concept. 5. Let f = (f1 , . . . , fp ) be the eigenvalues of the sample correlation matrix ρ matrix R. If the eigenvalues γα of the population correlation k are all distinct, then ﬁnd the limiting distribution of i=1 fi /p for the ratio of total variance explained by the ﬁrst k principal components.

11 Canonical correlations

11.1 Introduction The objective of canonical correlation analysis is to get a simple description of the structure of correlation between subsets of variables. Assume that two subsets of variables x1 and x2 have a joint normal distribution,

x1 x2

∼ Np

µ1 µ2

Σ11 , Σ21

Σ12 Σ22

.

The analysis searches for a pair of linear combinations t1 x1 and t2 x2 with maximum correlation. This is the ﬁrst canonical correlation. Having found such a pair, the analysis is pursued one step further by searching for a second pair of linear combinations with maximum correlation among all those uncorrelated with the ﬁrst pair. The correlation found is the second canonical correlation. The argument is repeated until all possible correlations are exhausted. This analysis is explained in detail in Section 11.2. In Section 11.3, tests of independence between x1 and x2 are derived. Not surprisingly, the tests proposed will be functions of the sample canonical correlations. Section 11.4 uses advantageously the context of testing independence to derive simple proofs of the properties of U (p; m, n) distributions introduced earlier in Section 9.3.2. As a by-product we also obtain a method of constructing Q-Q plots of squared radii for a visual inspection of multivariate normality. Asymptotic distributions of sample canonical correlations is the subject of Section 11.5.

11.2. Deﬁnition and basic properties

175

11.2 Deﬁnition and basic properties Assume Σjj > 0, Σij ∈ Rppij , i, j = 1, 2. Without any loss of generality, suppose p1 ≤ p2 . Write Σjj = A2j , where Aj > 0, j = 1, 2. Now using the SVD (v. Proposition 1.11), we have −1 A−1 1 Σ12 A2 = G(Dρ , 0)H ,

where Dρ = diag(ρ1 , . . . , ρp1 ), ρ1 ≥ · · · ≥ ρp1 ≥ 0, =

(g1 , . . . , gp1 ) ∈ Op1 ,

H =

(h1 , . . . , hp2 ) ∈ Op2 .

G

If we deﬁne u = G A−1 1 x1 = (u1 , . . . , up1 ) , −1 v = H A2 x2 = (v1 , . . . , vp2 ) ,

then

I (D , 0) p ρ 1 u . var = Dρ v Ip2 0

Obviously, var ui = var vj = 1 and cor(ui , vj ) = ρi δij , i = 1, . . . , p1 , j = 1, . . . , p2 . Deﬁnition 11.1 The variables u1 , . . . , up1 and v1 , . . . , vp2 are deﬁned to be the canonical variables. The numbers ρi ’s, 1 ≥ ρ1 ≥ ρ2 ≥ · · · ≥ ρp1 ≥ 0, are the canonical correlations. Note that the number of nonzero canonical correlations is rank Σ12 ≡ c. In a similar manner as the principal components were interpreted, the canonical variables can also be derived sequentially. First, we seek linear combinations t1 x1 and t2 x2 such that cor(t1 x1 , t2 x2 ) is maximal. But, in general, since cor(x, y) is invariant with respect to linear transformations, x → ax + b, y → cy + d, a, c > 0, we may assume at the outset that var tj xj = tj Σjj tj = 1, j = 1, 2. Introducing the ellipsoids Ej = {tj : tj Σjj tj = 1}, j = 1, 2, the problem is thus max t1 Σ12 t2 .

t1 ∈E1 t2 ∈E2

For tj ∈ Ej , |Aj tj | = 1, j = 1, 2, the Cauchy-Schwarz inequality gives (t1 Σ12 t2 )2

−1 2 = A1 t1 , A−1 1 Σ12 A2 h −1 2 ≤ |A−1 1 Σ12 A2 h| ,

176

11. Canonical correlations

−1 2 where h = A2 t2 has norm 1. Letting B = A−1 1 Σ12 A2 , then |Bh| = h B Bh, where Dρ2 0 −1 −1 −1 Σ A ) (A Σ A ) = H B B = (A−1 H . 12 2 12 2 1 1 0 0

Thus, from the method used for principal components, we ﬁnd h B Bh ≤ ρ21 with equality when h = h1 . This gives t2 x2 = h1 A−1 2 x2 = v1 . Finally, the Cauchy-Schwarz inequality is, in fact, an equality iﬀ A1 t1 ∝ Bh1 or, equivalently, t1

−1 −1 ∝ A−1 1 A1 Σ12 A2 h1 = A−1 1 G(Dρ , 0)H h1

= A−1 1 G(Dρ , 0)e1 = ρ1 A−1 1 g1 , which, in turn, gives t1 x1 = g1 A−1 1 x1 = u1 . We have proved that (u1 , v1 ) is the pair of linear combinations with maximum correlation ρ1 . Second, having found pairs of linear combinations −1 (ui , vi ) = (gi A−1 1 x1 , hi A2 x2 ), i = 1, . . . , k, k < rank Σ12 ≡ c,

another pair (t1 x1 , t2 x2 ) is sought with maximum correlation among all those uncorrelated with the preceding pairs; i.e., the restriction cov(tj xj , ui ) = cov(tj xj , vi ) = 0, i = 1, . . . , k; j = 1, 2, is imposed. This last restriction is characterized in terms of orthogonality: cov(t1 x1 , ui ) = t1 Σ11 A−1 1 gi = t1 A1 gi = 0 ⇐⇒ t1 ⊥ A1 gi .

Similarly, cov(t2 x2 , vi ) = 0 iﬀ t2 ⊥ A2 hi . We note that when t1 ⊥ A1 gi , the other condition, cov(t1 x1 , vi ) = 0, is automatically satisﬁed: cov(t1 x1 , vi )

= t1 Σ12 A−1 2 hi −1 = t1 A1 (A−1 1 Σ12 A2 )hi = t1 A1 G(Dρ , 0)H hi = ρi t1 A1 gi = 0.

Similarly, when t2 ⊥ A2 hi then cov(t2 x2 , ui ) = 0 is automatically satisﬁed. So the problem becomes max t1 Σ12 t2 ,

⊥ t1 ∈E1 ⊥ t2 ∈E2

where E1⊥ E2⊥

= {t1 ∈ E1 : t1 ⊥ A1 g1 , . . . , A1 gk }, = {t2 ∈ E2 : t2 ⊥ A2 h1 , . . . , A2 hk }.

11.3. Tests of independence

177

The Cauchy-Schwarz inequality gives for tj ∈ Ej⊥ , (t1 Σ12 t2 )2 = A1 t1 , Bh2 ≤ h B Bh, −1 and h = A2 t2 . But, using the where, as before, B = A−1 1 Σ12 A2 orthogonality restrictions, Dρ2 0 h B Bh = t2 A2 H H A2 t2 0 0 c = (t2 A2 hi )2 ρ2i ≤ ρ2k+1 , i=k+1

with equality when h = A2 t2 = hk+1 , which yields t2 x2 = hk+1 A−1 2 x2 = vk+1 . As before, the Cauchy-Schwarz inequality becomes an equality iﬀ −1 A1 t1 ∝ Bhk+1 , which implies t1 = A−1 1 gk+1 and t1 x1 = gk+1 A1 x1 = uk+1 . The solution is the pair of canonical variables (uk+1 , vk+1 ). Repeating the second stage for k = 1, . . . , c − 1, all the pairs of canonical variables (ui , vi ), i = 1, . . . , c, can be generated. Each pair of canonical variables is identiﬁed with the pair of linear combinations of x1 and x2 with maximum correlation among all those uncorrelated with the preceding pairs. Finally, the canonical correlations can be characterized as solutions of a determinant equation. In fact, the nonzero squared canonical correlations ρ2i , i = 1, . . . , c, are the nonzero eigenvalues of −1 −1 −1 −1 −1 −1 B B = (A−1 1 Σ12 A2 ) (A1 Σ12 A2 ) = A2 Σ21 Σ11 Σ12 A2 .

Hence, the nonzero ρ2i are the nonzero solutions λ of the equation −1 |Σ12 Σ−1 22 Σ21 Σ11 − λI| = 0.

11.3 Tests of independence Based on a random sample of size n from a Np1 +p2 (µ, Σ), where Σ11 Σ12 Σ= Σ21 Σ22 with Σij ∈ Rppij , we construct a test of independence reﬂected by the hypothesis, H0 : Σ12 = 0 ⇐⇒ H0 : ρ1 = · · · = ρp1 = 0, against all alternatives. The unbiased estimator S of Σ is partitioned in conformity as V11 V12 (n − 1)S ≡ V = V21 V22

178

11. Canonical correlations

ˆ = V/n is and we know already that V ∼ Wp (n − 1, Σ). The MLE Σ proportional to S. Without any restriction, the MLE of Σij , i, j = 1, 2, is ˆ ij = Vij /n. However, under H0 , the restricted MLE’s are given by Σ ˆ ˆ ˆ ˆ 11 = Σ ˆ 11 , Σ ˆ 22 = Σ ˆ 22 , Σ ˆ 12 = 0. Σ The LRT takes the form Λ = =

ˆ ˆ ˆ ˆ 22 , Σ ˆ 12 ) ˆ 11 , Σ L(¯ x, Σ ˆ 11 , Σ ˆ 22 , Σ ˆ 12 ) L(¯ x, Σ ˆ 22 |−n/2 ˆ 11 |−n/2 |Σ |Σ . ˆ −n/2 |Σ|

ˆ ∝ V and using the relation |V| = |V11 ||V22.1 |, Thus, since Σ Λ2/n

=

|V11.2 | |V22.1 | = |V22 | |V11 |

−1 −1 = |I − V12 V22 V21 V11 | p1 = (1 − ri2 ) i=1

is a function of the sample canonical correlations ri ’s, where ri2 is a solution λ of the equation −1 −1 V21 V11 − λI| = 0. |V12 V22

They satisfy w.p.1, 1 > r12 > · · · > rp21 > 0. Consider now the invariant tests. The group Gp1 × Gp2 × Rp1 × Rp2 transforms the observations as B1 0 xi1 b1 xi1 + → xi2 0 B2 xi2 b2 B1 xi1 + b1 = , i = 1, . . . , n, B2 xi2 + b2 for any (B1 , B2 , b1 , b2 ) ∈ Gp1 ×Gp2 ×Rp1 ×Rp2 . This induces the following transformations on the minimal suﬃcient statistic (¯ x, V): ¯1 ¯ 1 + b1 x B1 x → , ¯2 ¯ 2 + b2 B2 x x B1 0 V11 V12 B1 0 V11 V12 → V21 V22 V21 V22 0 B2 0 B2 B1 V11 B1 B1 V12 B2 = . B2 V21 B1 B2 V22 B2 A test function f (¯ x, V) is invariant iﬀ B1 W11 B1 B1 y1 + b1 , f (y, W) = f B2 y2 + b2 B2 W21 B1

B1 W12 B2 B2 W22 B2

,

11.3. Tests of independence

179

∀(B1 , B2 , b1 , b2 ) ∈ Gp1 × Gp2 × Rp1 × Rp2 , ∀(y, W) ∈ Rp × Pp . The choice ¯ i , i = 1, 2, immediately yields bi = −Bi x B1 V11 B1 B1 V12 B2 f (¯ x, V) = f 0, . B2 V21 B1 B2 V22 B2 Using the same arguments as in the deﬁnition of canonical correlations, let Vii = A2i , where Ai > 0, i = 1, 2, and consider the SVD −1 A−1 1 V12 A2 = G(Dr , 0)H ,

where Dr = diag(r1 , . . . , rp1 ), 1 > r1 > · · · > rp1 > 0, and we still assume and p1 ≤ p2 without loss of generality. Then, the choice B1 = G A−1 1 B2 = H A−1 2 ﬁnally gives Ip1 (Dr , 0) ; f (¯ x, V) = f 0, Dr Ip2 0 i.e., any invariant test is a function of the sample canonical correlations ri ’s. A similar argument shows that the power function of any invariant test depends only on the population canonical correlations ρi ’s. Proposition 11.1 With respect to the block-diagonal group of transformations above, any invariant test depends on the minimal suﬃcient statistic (¯ x, V) only through the sample canonical correlations ri ’s. The power function of any invariant test depends on (µ, Σ) only through the population canonical correlations ρi ’s. We now derive the null distribution of the LRT test. Proposition 11.2 Under the hypothesis of independence, H0 : Σ12 = 0 and n − 1 > min(p1 , p2 ), Λ2/n ∼ U (p2 ; p1 , n − 1 − p1 ). Proof. By invariance, assume without loss of generality that Σ = I and let m = n − 1. Write V11 V12 d X1 = (X1 , X2 ), V21 V22 X2 where (X1 , X2 ) ∼ Npm (0, Im ⊗ Ip ). The conditional distribution of X2 given X1 is X2 | X1 ∼ Npm2 (0, Im ⊗ Ip2 ). w.p.1 = p1 . Therefore, we have the SVD D G X 1 H = , 0

Now, for X1 ∈ Rm p1 , rank X1

180

11. Canonical correlations

where G ∈ Om , H ∈ Op1 , and D ∈ Rpp11 is diagonal and nonsingular. Thus, ˜1 DH X G X 1 = ≡ , 0 0 ˜ 1 ∈ Rp1 is nonsingular. Since G (a function of X1 ) is orthogonal, where X p1 G X2 | X1 ∼ Npm2 (0, Im ⊗ Ip2 ),

1 . Then, Y where Y ∈ Rpp12 and Z ∈ Rm−p p2

V22 V11 V12

|=

which does not depend on X1 and so G X2 ∼ Npm2 (0, Im ⊗ Ip2 ) unconditionally. Now, partition Y G X 2 = , Z Z and

= X2 X2 = (G X2 ) (G X2 ) = Y Y + Z Z, ˜X ˜ = X1 X1 = (G X1 ) (G X1 ) = X 1 1, ˜ = X X2 = (G X1 ) (G X2 ) = X Y. 1

1

Finally, Λ2/n

=

|V22.1 | |V22 |

=

˜X ˜ 1 (X ˜ −1 X ˜ Y| |Y Y + Z Z − Y X |Z Z| 1 1) 1 = , |Y Y + Z Z| |Y Y + Z Z|

where Z Z ∼ Wp2 (m − p1 ), and Y Y ∼ Wp2 (p1 ). By deﬁnition, Λ2/n ∼ U (p2 ; p1 , m − p1 ). 2 −1 −1 Let R = V12 V22 V21 V11 . As for multivariate regression, other invariant tests can be constructed such as

tr R =

p1

ri2 ,

i=1

tr R(I − R)−1

=

r12

=

p1 i=1

ri2 , (1 − ri2 )

max{r12 , . . . , rp21 }.

Again, none of these tests has a power function which uniformly dominates the others. It is shown in Example 14.10 how to perform a bootstrap test using the test statistics tr R or tr R(I − R)−1 .

11.4. Properties of U distributions

181

11.4 Properties of U distributions We end this chapter with some properties and characterizations useful for the tabulation and moments of U distributions. These simpliﬁed proofs are from Bilodeau (1996). Assume x1 ∈ Rp1 is ﬁxed and x2 ∼ Np2 (0, Σ22 ), p = p1 + p2 , Σ22 > 0. Based on a random sample of size n, say X ∈ Rnp , the matrix of sums of squares and cross-products V is partitioned in conformity as X1 V11 V12 V=XX= (X1 , X2 ) = . X2 V21 V22 When n > min(p1 , p2 ) and rank X1 = p1 , consider ˜= Λ

|V22.1 | |V11.2 | |V| = = . |V11 ||V22 | |V22 | |V11 |

Proposition 11.3 If n > min(p1 , p2 ) and rank X1 = p1 , then ˜ ∼ U (p2 ; p1 , n − p1 ). Λ Proof. Assume without loss of generality Σ22 = I and thus X2 ∼ Npn2 (0, I). Now, X1 ∈ Rnp1 has rank X1 = p1 . Its singular value decomposition is D G X 1 H = , 0 where G ∈ On , H ∈ Op1 , and D ∈ Rpp11 is diagonal and nonsingular. Thus, ˜1 DH X G X 1 = ≡ 0 0

1 . Then, Y where Y ∈ Rpp12 and Z ∈ Rn−p p2 ˜ ˜ ˜ V11 = X1 X1 , and V12 = X1 Y. Finally,

|=

˜ 1 ∈ Rp1 is nonsingular. Since G (a function of X1 ) is orthogonal, where X p1 G X2 ∼ Npn2 (0, I). Partition Y G X2 = , Z Z and V22 = Y Y + Z Z,

˜ ˜ ˜ −1 ˜ |Z Z| ˜ = |V22.1 | = |Y Y + Z Z − Y X1 (X1 X1 ) X1 Y| = Λ , |V22 | |Y Y + Z Z| |Y Y + Z Z|

˜ ∼ U (p2 ; p1 , n− where Z Z ∼ Wp2 (n−p1 ), Y Y ∼ Wp2 (p1 ). By deﬁnition, Λ 2 p1 ).

|=

Proposition 11.3 remains valid if X1 has any absolutely continuous distribution (and thus has rank X1 = p1 w.p.1 (v. Lemma 7.1 and the remark ˜ does X2 . It suﬃces to notice the distribution of Λ on page 88)) and X1 not depend on X1 .

182

11. Canonical correlations

˜ = |V11.2 |/|V11 |, if X1 is normal and X2 is ﬁxed, Vice versa, writing Λ ˜ ∼ U (p1 ; p2 , n − p2 ). The rank X2 = p2 , it is clear the same proof yields Λ d

duality property asserts that, in fact, U (p1 ; p2 , n − p2 ) = U (p2 ; p1 , n − p1 ). As a by-product, we show the “duality” property: d

|=

Corollary 11.1 U (p; m, n) = U (m; p, m + n − p) when m + n > p. X2 . Since X1 is Proof. Assume X1 and X2 are both normal and X1 ˜ ∼ U (p2 ; p1 , n− ˜ ∼ U (p1 ; p2 , n−p2 ), and since X2 is also normal, Λ normal, Λ d ˜ being unique, U (p1 ; p2 , n−p2 ) = p1 ). The distribution of Λ U (p2 ; p1 , n−p1 ). 2 Substitute (p, m, m + n) for (p1 , p2 , n). In order to obtain a characterization of U distributions as a product of independent beta variables, we prove the following lemma. Lemma 11.1 If n ≥ p, d

U (p; 1, n) = beta

1

2 (n

− p + 1); 12 p .

Proof. When m = 1, recalling the identity |I + AB| = |I + BA| (v. Problem 1.8.3), d

where z

|=

U (p; 1, n) =

|W| = |I + W−1 zz |−1 = (1 + z W−1 z)−1 , |W + zz |

W, z ∼ Np (0, I), and W ∼ Wp (n). Using Proposition 8.2, z W−1 z ∼ Fc (p, n − p + 1).

Finally, using Problem 3.5.5, (1 + z W−1 z)−1 ∼ beta

1

2 (n

− p + 1); 12 p . 2

Proposition 11.4 A variable distributed as U (p; m, n), n ≥ p, has the two characterizations m

d beta 12 (n − p + i); 12 p U (p; m, n) = i=1

and d

U (p; m, n) =

p

beta

1

2 (n

− p + i); 12 m ;

i=1

i.e., a U (p; m, n) variable has the same distribution as a product of independent beta variables. Proof. The second representation follows from the ﬁrst and the duality property of U distributions. We need only show the ﬁrst representation. Its proof proceeds by induction on m. From Lemma 11.1, the result is true for

11.4. Properties of U distributions

183

d

U (p; m, n) = =

|W| , W ∼ Wp (n), Z ∼ Npm (0, Im ⊗ Ip ), Z |W + Z Z| |W| |(W + z1 z1 )| ≡ U1 · U2 , · |W + z1 z1 | |(W + z1 z1 ) + Z2 Z2 |

where

Z=

z1 Z2

|=

m = 1. Assume the result is true for m − 1 and show it holds for m. By deﬁnition, W

,

|=

|=

Z2 . Consider z1 ∼ Np (0, I), Z2 ∼ Npm−1 (0, Im−1 ⊗ Ip ), and z1 now the distribution of U2 . Let W1 = W + z1 z1 and W2 = Z2 Z2 . Then, W1 ∼ Wp (n + 1), W2 ∼ Wp (m − 1), and W1 W2 . Therethe induction hypothesis gives U2 ∼ fore, U2 ∼ U (p; m − 1, n + 1) and m−1 1 1 i=1 beta 2 (n + 1 − p + i); 2 p . Translating i → i + 1, U2 ∼

m

beta

1

2 (n

− p + i); 12 p .

i=2

|=

|=

|=

|=

|=

|=

|=

The factor missing for i = 1 is U1 . The proof is complete if we prove U2 . First, note that if U1 W1 , then U1 , W1 , and Z2 are mutuU1 U2 . So, we prove U1 W1 . But, if ally independent and, therefore, U1 d V ∼ Wp (n, Σ), Σ > 0, x ∼ Np (0, Σ), and V x, then (V, x) = (Y Y, x), n x. In the model for (Y, x), Y Y + xx where Y ∼ Np (0, In ⊗ Σ), Y is complete and suﬃcient for Σ. Therefore, V + xx is complete and sufﬁcient for Σ. Using Basu’s theorem in the footnote on page 118, V + xx is independent of any ancillary statistic such as |V|/|V + xx |. This proves W1 . 2 U1 This representation is useful for ﬁnding the distribution function or quantiles of a U (p; m, n) distribution since ln U (p; m, n) can be represented as a convolution of simple distributions. Of course, it is advantageous to use the representation with min(p, m) number of factors. This number of factors can be reduced further by 12 by grouping adjacent factors by pairs [Anderson (1984), p. 304]. The following lemma allows the pairing. Lemma 11.2 For n > 1, [beta(n − 1; m)]2 = beta( 12 (n − 1); 12 m) · beta( 12 n; 12 m). d

Proof. It is straightforward to check that all moments of order h > 0 on d the left and right sides of = are the same (v. Problem 11.6.3). Since the domain is the bounded interval [0, 1], there is a unique distribution with these moments [Serﬂing (1980), p. 46]. 2 The following representation has a reduced number of factors as p is even or odd.

184

11. Canonical correlations

Corollary 11.2 For n ≥ p, a U (p; m, n) variable can be represented as r

beta( 12 (n − p + 1); 12 m) ·

i=1 r

[beta(n + 1 − 2i; m)]2 ,

if p = 2r

[beta(n + 1 − 2i; m)]2 ,

if p = 2r + 1.

i=1

Proof. The proof for p = 2r is as follows. From Proposition 11.4, we have d

U (p; m, n) =

r

beta( 12 (n − p + 2i − 1); 12 m) beta( 12 (n − p + 2i); 12 m),

i=1

and from Lemma 11.2, d

U (p; m, n) =

r

[beta(n − p + 2i − 1; m)]2 .

i=1

The conclusion follows after reversing the index i → r − i + 1. The proof for p odd is identical except for the ﬁrst isolated factor. 2 The asymptotic distribution as n → ∞ of U (p; m, n) should be clear from the asymptotic distribution of the likelihood ratio statistic in Proposition 11.2, −n ln U (p; m, n) → χ2pm . d

(11.1)

The slight modiﬁcation −[n − 12 (p − m + 1)] ln U (p; m, n) → χ2pm d

is often used as an improved approximation since it has a remainder of order O(n−2 ), whereas the remainder in (11.1) is O(n−1 ). The general asymptotic expansion of order O(n−α ) [Box (1949)] is treated in Section 12.3. As an alternative to asymptotic expansion an S-plus program in Appendix C uses the fast Fourier transform [Press (1992)] to compute the density of U (p; m, n) by convolution and thus calculates exact probabilities (up to a discretization of the beta variables) and quantiles. Srivastava and Yau (1989) presented the saddlepoint method for obtaining tail probabilities. An exact closed form solution without series representation was also recently derived [Coelho (1998)].

11.4.1

Q-Q plot of squared radii

The scaled residuals of n observations, xi , may be deﬁned as ¯ ), i = 1, . . . , n. zi = S−1/2 (xi − x ¯ ) are Then, the squared radii (or squared Mahalanobis distances of xi to x ¯ ) S−1 (xi − x ¯ ), i = 1, . . . , n. d2i = |zi |2 = (xi − x

11.4. Properties of U distributions

185

Note that if xi ∼ Np (µ, Σ), then di is an ancillary statistic; i.e., the distribution of di , say F (·), does not depend on (µ, Σ). One aspect of multivariate normality can thus be tested with a Q-Q plot of the ordered d2i against the quantiles of the distribution F (·) [Small (1978)]. Gnanadesikan and Kettenring (1972) derived the following result. Lemma 11.3 If x1 , . . . , xn are i.i.d. Np (µ, Σ), then

n d2i ∼ beta 12 p; 12 (n − p − 1) . 2 (n − 1) Proof. 1−

n d2 (n − 1)2 i

n ¯ )(xi − x ¯ ) |/|V| (xi − x (n − 1) = |W1 |/|W1 + W2 |,

= |V −

n n ¯ )(xi − x ¯ ) , W1 = V − W2 , and W2 = (n−1) where V = i=1 (xi − x (xi − ¯ ) . Assume without loss of generality that µ = 0 and Σ = I. Thus, ¯ )(xi − x x with Z ∼ Npn (0, In ⊗ Ip ), (W1 , W2 ) = (Z (Q − H)Z, Z HZ) , d

where H = Q

n (ei − n−1 1)(ei − n−1 1) , (n − 1)

= I − n−1 11 .

The following can be veriﬁed easily: (i) H is idempotent of rank 1, (ii) Q is idempotent of rank n − 1, (iii) Q(ei − n−1 1) = (ei − n−1 1) and, thus, QH = H, (iv) Q − H is idempotent of rank n − 2, (Q − H)H = 0. |=

Thus, W1 W2 , W1 ∼ Wp (n − 2), and W2 ∼ Wp (1) (v. Proposition 7.8 and Problem 6.4.3), which implies

d |W1 |/|W1 + W2 | ∼ U (p; 1, n − 2) = beta 12 (n − p − 1); 12 p . 2 Consider the ordered

d2i , d2(1) ≤ d2(2) ≤ · · · ≤ d2(n) .

Assuming d2i , i = 1, . . . , n, are i.i.d. according to the distribution in Lemma 11.3, one could evaluate the expected order statistics, E d2(i) . Then, the Q-Q plot consists of a graph of the points d2(i) , E d2(i) , i = 1, . . . , n.

186

11. Canonical correlations

To simplify matters, we can assign to d2(i) a cumulative probability of i/n and approximate E d2(i) by the quantile γi = i/n of the distribution [(n − 1)2 /n] beta

1

1 2 p; 2 (n

− p − 1) .

Blom (1958) has shown how to select α and β so that the expected order statistic E d2(i) may be well approximated by the quantile γi = (i − α)/(n − α − β + 1).

(11.2)

For beta, the distribution at hand, the indicated choice is α

=

β

=

(p − 2) , 2p (n − p − 2) . 2(n − p − 1)

(11.3)

Thus, the recommended Q-Q plot is the graph of the points

d2(i) , [(n − 1)2 /n] betaγi 12 p; 12 (n − p − 1) , i = 1, . . . , n, where betaα (a; b) denotes the quantile α of a beta(a; b) distribution and γi is given by (11.2) and (11.3). The Splus function qqbeta in Appendix C produces the Q-Q plot. One should not forget, however, that the d2i are correlated, but from Wilks (1963), cor(d2i , d2j ) = −

1 , i = j, (n − 1)

and the correlation, of the order O(n−1 ), is negligible for moderate to large sample sizes. A Q-Q plot approaching a 45◦ straight line is consistent with multivariate normality. Figure 11.1 gives the Q-Q plot for 50 observations generated from a N3 (0, I) distribution and Figure 11.2 is the Q-Q plot for 50 observations generated from a trivariate Cauchy distribution. These are easily generated with Example 13.2. The deviations from the straight line are clearly more systematic in Figure 11.2 associated with a distribution with heavier “tails” than the multivariate normal. For large n, the beta distribution can be approximated by a χ2p distribution. Gnanadesikan (1977, p. 172) remarked that in the bivariate case n = 25 may provide a suﬃciently large sample for this chi-squared approximation to be adequate. However, n = 100 does not seem large enough for p = 4, for there is a marked deviation from linearity when the ordered d2i are plotted against expected order statistics of chi-squared, and this eﬀect becomes more marked as p increases. We therefore recommend the use of the beta distribution.

11.4. Properties of U distributions

187

10

•

8

•

• 6

•

4 0

2

Beta quantiles

•

• •• ••• •• •• • • •• 0

•• •• • • ••• •• •

2

• •• • •• •••

4

•

• • •

6

8

10

12

14

Squared distances Figure 11.1. Q-Q plot for a sample of size n = 50 from a trivariate normal, N3 (0, I), distribution.

11. Canonical correlations

•

10

188

8

•

• •

6 4 0

2

Beta quantiles

•

• •• • •• •• • •• •• ••••• •• ••• •• •••• • 0

••

• •

5

•

•

10

15

20

25

Squared distances Figure 11.2. Q-Q plot for a sample of size n = 50 from a trivariate t on 1 degree of freedom, t3,1 (0, I) ≡ Cauchy3 (0, I), distribution.

11.5. Asymptotic distributions

189

11.5 Asymptotic distributions Assuming xi , i = 1, . . . , n, are i.i.d. Np (µ, Σ), we derive the asymptotic distribution of rα2 , α = 1, . . . , p1 , when ρα is distinct from all other canonical correlations. A squared sample canonical correlation, rα2 , is a value of l for which there is a nonzero solution c to the equation −1 (S−1 11 S12 S22 S21 − l I)c = 0.

(11.4)

d

Using the result n1/2 (S − Σ) → W of Section 6.3, where W ∼ Npp (0, (I + K)(Σ ⊗ Σ)), we write S11 S22

= I + n−1/2 W11 ,

S12

=

= I + n−1/2 W22 , (Dρ , 0) + n−1/2 W12 .

−1/2 W22 + Op (n−1 ) and similarly for Using Problem 1.8.15, S−1 22 = I − n −1 −1/2 , S11 . Keeping terms up to order n −1 S−1 11 S12 S22 S21

= D2ρ + n−1/2 −D2ρ W11 + (Dρ , 0)W21 Dρ Dρ −(Dρ , 0)W22 + W12 + Op (n−1 ). 0 0

We now apply the perturbation method as in Section 8.8.1 and obtain from (8.11) the expansion αα αα αα − ρ2α w22 + 2ρα w21 ] + Op (n−1 ), rα2 = ρ2α + n−1/2 [−ρ2α w11 αα is the element (α, α) of the matrix Wij . From (6.1), we have where wij αα αα αα (w11 , w22 , w21 ) → N3 (0, Ω), d

where

2 Ω = 2ρ2α 2ρα

2ρ2α 2 2ρα

2ρα 2ρα . 1 + ρ2α

Finally, deﬁning the linear combination vector a = (−ρ2α , −ρ2α , 2ρα ) , we d

obtain n1/2 (rα2 − ρ2α ) → N (0, a Ωa), whereby a direct calculation shows a Ωa = 4ρ2α (1 − ρ2α )2 . We have shown: Proposition 11.5 The asymptotic distribution of the squared sample canonical correlation rα2 , α = 1, . . . , p1 , assuming ρα is distinct from all d

other canonical correlations is n1/2 (rα2 − ρ2α ) → N (0, 4ρ2α (1 − ρ2α )2 ).

190

11. Canonical correlations

Various extensions of Proposition 11.5 to the joint distribution of sample canonical correlations can be envisaged. The simplest extension is to the joint distribution of r12 , . . . , rp21 when all canonical correlations are distinct, ρ1 > · · · > ρp1 . Corollary 11.3 The asymptotic joint distribution of r12 , . . . , rp21 when all population canonical correlations are distinct, ρ1 > · · · > ρp1 , is n1/2 (r12 − ρ21 , . . . , rp21 − ρ2p1 )

d → Np1 0, 4 diag ρ21 (1 − ρ21 )2 , . . . , ρ2p1 (1 − ρ2p1 )2 . Proof. It suﬃces to consider the asymptotic covariance of two squared sample canonical correlations, rα2 and rβ2 , when the population canonical correlations, ρα and ρβ , are of multiplicity 1. But, it is immediate from the proof of Proposition 11.5 that ββ ββ ββ αα αα αα − ρ2α w22 + 2ρα w21 , −ρ2β w11 − ρ2β w22 + 2ρβ w21 )=0 cov(−ρ2α w11 ββ αα since from (6.1), all the covariances satisfy, cov(wij , wkl ) = 0, i, j, k, l = 1, 2. 2

Hsu (1941) derived the asymptotic joint density when 1 > ρ1 > · · · > ρc > ρc+1 = · · · = ρp1 = 0. Muirhead and Waternaux (1980) obtained the asymptotic joint distribution when all population canonical correlations are distinct, as in Proposition 11.3 but for any underlying distribution with ﬁnite fourth moments. Eaton and Tyler (1994), assuming an underlying elliptical distribution or, in fact, any other distribution with ﬁnite fourth moments, derived the asymptotic joint distribution in full generality, ρ1 ≥ · · · ≥ ρc > ρc+1 = · · · = ρp1 = 0, using an extension of Wielandt’s inequality to singular values. In canonical correlation analysis, the number of nonzero population correlations is called the dimensionality. Asymptotic distributions of the dimensionality estimated by Mallow’s criterion and Akaike’s criterion were derived [Gunderson and Muirhead (1997)] for non-normal multivariate populations with ﬁnite fourth moments.

11.6 Problems 1. Obtain the hth moment, h > 0, of U ∼ U (p; m, n), n ≥ p, m Γ 12 (n − p + i) + h Γ 12 (n + i) h = EU · 1 1 Γ (n − p + i) Γ (n + i) + h 2 2 i=1

11.6. Problems

=

p Γ i=1

191

1 Γ 12 (m + n − p + i) 2(n − p + i) +h . · 1 Γ 12 (n − p + i) Γ 2 (m + n − p + i) + h

2. Establish the following exact results concerning U distributions: n[1 − U (1; m, n)] m U (1; m, n) (n − p + 1)[1 − U (p; 1, n)] p U (p; 1, n) (n − 1)[1 − U (2; m, n)1/2 ] m U (2; m, n)1/2 (n − p + 1)[1 − U (p; 2, n)1/2 ] p U (p; 2, n)1/2

∼ F (m, n), ∼ F (p, n − p + 1), ∼ F (2m, 2(n − 1)), ∼ F (2p, 2(n − p + 1)).

3. Prove that [beta(n − 1, m)]2 and beta( 12 (n − 1), 12 m) · beta( 12 n, 12 m), a product of two independent betas, have the same moments of order h > 0. 4. For x = (x1 , x2 ) = (x1 , . . . , xp1 ; xp1 +1 , . . . , xp1 +p2 ) establish that simple correlation and multiple correlation coeﬃcients are bounded above as (i) |ρxi ,xj | ≤ ρ1 , i = 1, . . . , p1 , j = p1 + 1, . . . , p1 + p2 , (ii) Rxi ,x2 ≤ ρ1 , i = 1, . . . , p1 , where ρ1 is the largest canonical correlation. 5. Let Σ12 = ρ1p1 1p2 , Σii = ρ1pi 1pi + (1 − ρ)Ipi , i = 1, 2, corresponding to the equicorrelated case. Determine the canonical variables corresponding to the nonzero canonical correlation. Hint: Σ11 1p1 = [1 + (p1 − 1)ρ]1p1 . 6. Let x1 , . . . , xn be i.i.d. Np (µ, Σ), where Σ11 Σ12 Σ= Σ21 Σ22 with Σ12 ∈ Rpp12 , p = p1 + p2 . For testing H0 : Σ12 = 0 against H1 : Σ12 = 0, consider the test statistic [Escouﬁer (1973)] E=

tr(S12 S21 ) , [tr(S211 )]1/2 [tr(S222 )]1/2

where S is the sample variance partitioned as Σ. Prove: (i) E is invariant under the group of transformations xi1 H1 0 xi1 b1 xi = → + xi2 xi2 b2 0 H2

192

11. Canonical correlations

for any (H1 , H2 , b1 , b2 ) ∈ Op1 × Op2 × Rp1 × Rp2 . (ii) If H0 holds, then the distribution of E is the same as when diag(λi ) 0 Σ= , 0 diag(γj ) where λi and γj are, respectively, the eigenvalues of Σ11 and Σ22 . d (iii) Under H0 , n1/2 S12 → Z = (zij ), where zij are independently distributed as N (0, λi γj ), i = 1, . . . , p1 , j = 1, . . . , p2 . (iv) Conclude the null distribution −1/2 p −1/2 p p1 p2 1 2 d 2 λ2i γj2 λi γj zij . nE→ i=1

j=1

i=1 j=1

Remark: Unlike for canonical correlations, the asymptotic null distribution depends on unknown parameters because of the lack of invariance of E (the group Op1 × Op2 × Rp1 × Rp2 is only a subgroup of Gp1 ×Gp2 ×Rp1 ×Rp2 ). The asymptotic distribution of E for sampling from an elliptical distribution was derived by Cl´eroux and Ducharme (1989). 7. Test of mutual independence of several subvectors. This problem given in the form of a project derives the exact null distribution of the likelihood ratio test for mutual independence. Consider a random sample of size n ≥ p + 1 from Np (µ, Σ), where Σ11 Σ12 · · · Σ1r Σ21 Σ22 · · · Σ2r Σ= .. .. .. ... . . . r

Σr1

···

Σr2

Σrr

with Σij ∈ Rppij , p = j=1 pj . We wish to test H0 : Σij = 0, 1 ≤ i < j ≤ r, versus all alternatives. (i) Prove the likelihood ratio test Λ for H0 can be written |V| , i=1 |Vii |

Λ2/n = r

n ¯ )(xi − x ¯ ) ∼ Wp (n − 1, Σ). where as usual V = i=1 (xi − x ˜ = Λ2/n , (ii) Obtain the exact null moments of Λ 1 Γpi ( 1 m) 2 ˜ h = Γp ( 2 m + h) EΛ , 1 Γp ( 2 m) i=1 Γpi ( 12 m + h) r

where m = n − 1.

11.6. Problems

193

Hint: ˜h = EΛ

cp,m

r

cp,m+2h

i=1

where cp,m = [2pm/2 Γp ( 12 m)]−1 is of a Wp (m) density and where Vii Wpi (m + 2h). (iii) Deﬁne the upper left corner of V to V11 V12 · · · V21 V22 · · · ˜ ii = . V .. .. .. . . Vi1

Vi2

···

E |Vii |−h , the normalizing constant are mutually independent be V1i V2i ∈ Rpp¯¯ii , .. . Vii

˜ ˜ where p¯i = p1 +· · ·+pi , and note that rVrr = V and V11 = V11 . Derive the equivalent form Λ2/n = i=2 Ui , where Ui =

˜ ii | |V , i = 2, . . . , r. ˜ i−1,i−1 | |Vii ||V

(iv) Use Proposition 11.2 to obtain immediately under H0 Ui ∼ U (pi ; p¯i−1 , n − 1 − p¯i−1 ) , i = 2, . . . , r. (v) When Σ=

˜ r−1,r−1 Σ 0

0 Σrr

,

|=

|=

|=

˜ r−1,r−1 , Vrr ) is suﬃcient and complete and that prove that (V Ur is ancillary. Conclude that, under H0 , Ur Vr−1,r−1 , Vrr ). U2 , . . . , Ur−1 ) under H0 . (vi) Using (iii), prove that Ur U2 , . . . , Ui−1 ), i = (vii) Repeat this argument to prove Ui 3, . . . , r, whence, altogether, U2 , . . . , Ur are mutually independent under H0 . (viii) Use Proposition 11.4 to obtain the exact null distribution d ˜= Λ

pi r

beta

1

2 (n

− p¯i−1 − j); 12 p¯i−1 .

i=2 j=1

Note that a further representation with a reduced number of factors as pi is odd or even is immediate from Corollary 11.2. n/2 (ix) Prove Ui is the likelihood ratio test for Hi : Σli = 0, l = 1, . . . , i − 1 when it is known that all the hypotheses Hi+1 , · · ·, Hr are true. Note that H0 = ∩ri=2 Hi .

194

11. Canonical correlations

(x) Use (viii) to obtain the equivalent expression for the exact null ˜ moments of Λ: pi r 1 1 Γ[ (n − p ¯ − j) + h] Γ[ (n − j)] i−1 2 2 ˜h = . EΛ Γ[ 12 (n − p¯i−1 − j)] Γ[ 12 (n − j) + h] i=2 j=1 8. Generalized squared interpoint distance [Gnanadesikan and Kettenring (1972)] Let d2ij = (xi −xj ) S−1 (xi −xj ) be the generalized squared interpoint distance (or squared Mahalanobis distance) between xi and xj , i = j. Prove that if x1 , . . . , xn are i.i.d. Np (µ, Σ), then

1 d2ij ∼ beta 12 p; 12 (n − p − 1) . 2(n − 1)

12 Asymptotic expansions

12.1 Introduction The exact distribution of likelihood ratio tests in multivariate analysis is often too complicated to be of any practical use. An asymptotic expansion due to Box (1949) is rather simple and easy to program on a computer to obtain the distribution function to any degree of accuracy. This approximation is applied on several of the testing situations previously encountered. In at least one situation where the exact distribution is known, an evaluation of the approximation is carried out for small to moderate sample sizes.

12.2 General expansions The method can be used whenever the likelihood ratio criterion Λ (or a suitable power W ) has moment of order h of the form b yj h a y j j=1 k=1 Γ[xk (1 + h) + ζk ] , (12.1) E W h = K a b xk x k=1 k j=1 Γ[yj (1 + h) + ηj ] where b j=1

yj =

a k=1

xk ,

196

12. Asymptotic expansions

and K is just a constant (not depending on h) so that E W 0 = 1. Equation (12.1) is usually obtained for real h; it is, however, generally valid on the domain where the functions are analytic. This means if we let M = −2 log W , then we can write the characteristic function of ρM , for a constant 0 < ρ ≤ 1 to be determined later, as cρM (t)

= E W −2itρ b

y y j aj=1 jxk k=1 xk

= K

−2itρ a k=1

Γ[xk (1 − 2itρ) + ζk ]

j=1

Γ[yj (1 − 2itρ) + ηj ]

b

.

Taking logarithms and deﬁning βk = (1 − ρ)xk ,

j = (1 − ρ)yj ,

(12.2)

the cumulant generating function is KρM (t) = log cρM (t) = g(t) − g(0),

(12.3)

where g(t)

=

2itρ[

a

k=1

+

a

xk log xk −

b

yj log yj ]

j=1

log Γ[ρxk (1 − 2it) + βk + ζk ]

k=1

−

b

log Γ[ρyj (1 − 2it) + j + ηj ],

j=1

with g(0) = − log K =

a

log Γ[ρxk + βk + ζk ] −

b

log Γ[ρyj + j + ηj ].

j=1

k=1

We use the asymptotic expansion in z as |z| → ∞ [Erd´elyi et al. (1953), p. 48] for bounded h, √ log Γ(z + h) = log 2π + (z + h − 12 ) log z − z (12.4) −

l

(−1)α

α=1

Bα+1 (h) −α z + O(z −(l+1) ), | arg z| < π. α(α + 1)

The terms Br (h) are the Bernoulli polynomials deﬁned to be the coeﬃcients in the Taylor series ∞

zr zehz = B (h) , |z| < 2π. r ez − 1 r=0 r! The reader can verify the ﬁrst few Bernoulli polynomials B0 (h)

=

1,

12.2. General expansions

B1 (h) B2 (h) B3 (h)

197

= h − 12 ,

= h2 − h + 16 ,

= h3 − 32 h2 + 12 h,

B4 (h)

= h4 − 2h3 + h2 −

B5 (h)

= h5 −

B6 (h)

= h6 −

1 30 , 5 4 5 3 1 2 h + 3 h − 6 h, 3h5 + 52 h4 − 12 h2

+

1 42 .

Bernoulli polynomials can be generated at will with modern symbolic computations software such as the function bernoulli(r,h); in Maple [Redfern (1996)] or BernoulliB[h,r] in Mathematica [Wolfram (1996)]. Let (z, h) = (ρxk (1 − 2it), βk + ζk ), (ρyj (1 − 2it), j + ηj ), (ρxk , βk + ζk ), and (ρyj , j + ηj ) in turn in (12.4). We assume that xk and yj are terms behaving as O(n), where n is the sample size. This will have to be checked in each application. When ρ = 1, then βk = j = 0, and h is bounded in all cases. Later, ρ will be allowed to depend on the sample size n and we will need to check that βk and j are bounded. Then substitute the four expansions for log Γ(z + h) in g(t) and g(0) of (12.3) to obtain, after long but straightforward simpliﬁcations, KρM (t) = − 12 f log(1 − 2it) +

l

ωα [(1 − 2it)−α − 1] + O(n−(l+1) ), (12.5)

α=1

where

f = −2

a

ζk −

b

ηj − 12 (a − b) ,

(12.6)

j=1

k=1

a b (−1)α+1 Bα+1 (βk + ζk ) Bα+1 (j + ηj ) − . ωα = α(α + 1) (ρxk )α (ρyj )α j=1

(12.7)

k=1

Note that ωα = O(n−α ) if xk and yj are O(n), and βk and j are O(1). The next step consists of deriving the characteristic function cρM by exponentiation of KρM and then using the inversion formula (2.2) to derive the p.d.f.: cρM (t)

= eKρM (t) =

(1 − 2it)−f /2 ·[1 + O(n

=

exp[ωα (1 − 2it)−α ]

α=1 −(l+1)

−f /2

(1 − 2it)

l

l

exp(−ωα )

α=1

)]

l ∞ α=1 k=0

l ∞ ωαk ωk −αk (1 − 2it) (−1)k α k! k! α=1 k=0

198

12. Asymptotic expansions

·[1 + O(n−(l+1) )]. The expansion of order O(n−(l+1) ) is then obtained by keeping and collecting terms of order up to O(n−l ). Let us illustrate the procedure for an expansion of order O(n−2 ) and leave the higher-order expansion to the symbolic calculators Mathematica [Wolfram (1996)] or Maple [Redfern (1996)]. For the expansion of order O(n−2 ), set l = 1 and note that cρM (t)

=

(1 − 2it)−f /2 [1 + ω1 (1 − 2it)−1 ](1 − ω1 ) + O(n−2 )

= (1 − 2it)−f /2 {1 + ω1 [(1 − 2it)−1 − 1]} + O(n−2 ) = cf (t) + ω1 [cf +2 (t) − cf (t)] + O(n−2 ), (12.8) where cf (t) = (1 − 2it)−f /2 denotes the characteristic function of χ2f on f degrees of freedom. Then, by the inversion formula (2.2), ∞ 1 fρM (s) = cρM (t)e−its dt 2π −∞ = gf (s) + ω1 [gf +2 (s) − gf (s)] + O(n−2 ),

(12.9)

where gf (t) is the p.d.f. of χ2f . Finally, by integration on (−∞, x], the d.f. takes the form P (ρM ≤ x) = Gf (x) + ω1 [Gf +2 (x) − Gf (x)] + O(n−2 ), where Gf (t) is the d.f. of χ2f . A full justiﬁcation of the last two integrations would require one to show that the remainders in (12.8) and (12.9) are O(n−2 ) uniformly in t and s, respectively; v. Box (1949) for details. The whole purpose of introducing ρ in the expansion is to reduce the number of terms. In the above example, one can choose ρ to annihilate the term of order O(n−1 ), i.e., to make ω1 = 0. Recalling (12.2) and B2 (h) = h2 − h + 16 , we have a b 1 −1 −1 ω1 = xk B2 (βk + ζk ) − yj B2 (j + ηj ) 2ρ j=1 k=1 a b 1 2 1 −(1 − ρ)f + = x−1 yj−1 (ηj2 − ηj + 16 ) . k (ζk − ζk + 6 ) − 2ρ j=1 k=1

Thus, ω1 vanishes by choosing a b 2 1 x−1 yj−1 (ηj2 − ηj + 16 ) . ρ = 1 − f −1 k (ζk − ζk + 6 ) − k=1

(12.10)

j=1

Even though ρ now depends on xk and yj , assumed to be of order O(n), the asymptotic expansion is still valid since for this choice of ρ, βk = (1 − ρ)xk and j = (1 − ρ)yj are terms of order O(1) and, thus, ωα is still O(n−α ).

12.2. General expansions

199

Proposition 12.1 If W has moments (12.1), where xk and yj are terms O(n), then with the choice of ρ in (12.10), P (ρM ≤ x) = Gf (x) + O(n−2 ).

(12.11)

The asymptotic expansions of order O(n−2 ) were proposed by Bartlett (1938). The factor ρ which annihilates the term ω1 of order O(n−1 ) is referred to as the Bartlett correction factor. We now give the general result for an expansion of order O(n−6 ) (l = 5) calculated with the help of Maple [Redfern (1996)]: P (ρM ≤ x) = Gf (x) + ω1 [Gf +2 (x) − Gf (x)] + ω2 [Gf +4 (x) − Gf (x)] + 12 ω12 [Gf +4 (x) − 2Gf +2 (x) + Gf (x)] + ω3 [Gf +6 (x) − Gf (x)] +ω2 ω1 [Gf +6 (x) − Gf +4 (x) − Gf +2 (x) + Gf (x)] + 16 ω13 [Gf +6 (x) − 3Gf +4 (x) + 3Gf +2 (x) − Gf (x)] +ω4 [Gf +8 (x) − Gf (x)] + ω3 ω1 [Gf +8 (x) − Gf +6 (x) − Gf +2 (x) + Gf (x)] + 12 ω2 ω12 [Gf +8 (x) − 2Gf +6 (x) + 2Gf +2 (x) − Gf (x)]

1 4 + 24 ω1 [Gf +8 (x) − 4Gf +6 (x) + 6Gf +4 (x) − 4Gf +2 (x)] + Gf (x)]

+ 12 ω22 [Gf +8 (x) − 2Gf +4 (x) + Gf (x)] + ω5 [Gf +10 (x) − Gf (x)] +ω4 ω1 [Gf +10 (x) − Gf +8 (x) − Gf +2 (x) + Gf (x)]

+ω3 ω2 [Gf +10 (x) − Gf +6 (x) − Gf +4 (x) + Gf (x)] + 12 ω3 ω12 [Gf +10 (x) − 2Gf +8 (x) + Gf +6 (x) −Gf +4 (x) + 2Gf +2 (x) − Gf (x)] 1 + 6 ω2 ω13 [Gf +10 (x) − 3Gf +8 (x) + 2Gf +6 (x)

+2Gf +4 (x) − 3Gf +2 (x) + Gf (x)] 1 + 120 ω15 [Gf +10 (x) − 5Gf +8 (x) + 10Gf +6 (x) −10Gf +4 (x) + 5Gf +2 (x) − Gf (x)] 1 2 + 2 ω2 ω1 [Gf +10 (x) − Gf +8 (x) − 2Gf +6 (x) +2Gf +4 (x) + Gf +2 (x) − Gf (x)]

+O(n−6 ). When ρ is chosen as in (12.10) so that ω1 = 0, then things reduce considerably. Proposition 12.2 If W has moments (12.1) where xk and yj are terms O(n), then with the choice of ρ in (12.10),

P (ρM ≤ x) = Gf (x) + ω2 [Gf +4 (x) − Gf (x)] + ω3 [Gf +6 (x) − Gf (x)] +ω4 [Gf +8 (x) − Gf (x)] + 12 ω22 [Gf +8 (x) − 2Gf +4 (x) + Gf (x)] +ω5 [Gf +10 (x) − Gf (x)] + ω3 ω2 [Gf +10 (x) − Gf +6 (x) − Gf +4 (x) + Gf (x)] +O(n−6 ).

200

12. Asymptotic expansions

Automatic correction of coverage probability of conﬁdence intervals [Martin (1990) or of error in rejection probability of tests [Beran (1987, 1988)] is now made possible by the resampling or “bootstrap” technology (v. Chapter 14). Bootstrap of a Bartlett corrected likelihood ratio test reduces level error in Proposition 12.1 from O(n−2 ) to O(n−3 ) automatically without further analytical expansions.

12.3 Examples We now present some examples. Example 12.1 Test of sphericity. The likelihood ratio test (LRT) of sphericity was derived in Proposition 8.11 and its moments are given in Proposition 8.12. Now, it is simply a matter of rewriting things in the form (12.1) to obtain the asymptotic expansion. Hence, for W = Λm/n , m = n − 1, E Wh

Γp [ 12 m + 12 mh] Γ[ 12 mp + 12 pmh] p 1 1 1 k=1 Γ[ 2 m + 2 mh − 2 (k − 1)] pmh/2 = Kp Γ[ 12 mp + 12 pmh] p Γ[ 1 m(1 + h) − 12 (k − 1)] = Kppmh/2 k=1 2 1 , Γ[ 2 mp(1 + h)] = Kppmh/2

so that we have the form (12.1) with a = p,

xk = 12 m,

ζk = − 12 (k − 1),

b = 1,

y1 = 12 mp,

η1 = 0.

p

Observe that k=1 xk = y1 is satisﬁed and xk and y1 are terms behaving as O(n). The asymptotic expansion with remainder O(n−6 ) as in Proposition 12.2 is now a simple matter of calculating with (12.6) and (12.10), f

=

1 2 (p

ρ

=

1−

+ 2)(p − 1), 2p2 + p + 2 , 6pm

βk and 1 in (12.2), and, ﬁnally, ωα , α = 2, 3, 4, 5, in (12.7). Of course, one could go to great lengths to obtain the most simpliﬁed algebraic expressions in terms of p and n. For example, Davis (1971) using properties of Bernoulli polynomials showed ωα

=

2(−1)α α(α + 1)(α + 2)ρα

12.3. Examples

s

Bs

0

1

1

− 12

2

1 6

3

0

4

1 − 30

5

0

6

1 42

7

0

201

δs − 12 p 1 4 p(p

+ 1)

1 − 16 p(2p2 + 3p − 1) 1 16 p(p

− 1)(p + 1)(p + 2)

1 − 192 p(6p4 + 15p3 − 10p2 − 30p + 3) 1 128 p(p

− 1)(p + 1)(p + 2)(2p2 + 2p − 7)

1 − 768 p(6p6 + 21p5 − 21p4 − 105p3 + 21p2 + 147p − 5) 1 768 p(p

− 1)(p + 1)(p + 2)(3p4 + 6p3 − 23p2 − 26p + 62)

Table 12.1. Polynomials δs and Bernoulli numbers Bs for asymptotic expansions.

·

α+1 s=1

α+2 Bs (1 − ρ)α+1−s δs + 12 (s + 1) s−1 ( 12 m)1−s , s+1 p

where Bs ≡ Bs (0) are the Bernoulli numbers and the δs are certain polynomials in p deﬁned by Box (1949) (v. Table 12.1). Example 12.2 Asymptotics for U (p; m, n) distributions. The LRT for the general linear hypothesis in multivariate regression was described in Proposition 9.3 as Λ2/n ∼ U (p; r, n − k). Another example is the LRT for independence between two subvectors in Proposition 11.2 where Λ2/n ∼ U (p2 ; p1 , n − 1 − p1 ). Thus, we derive the asymptotic expansion for W ∼ [U (p; m, n − c)]n/2 , where n − c ≥ p, which includes both cases. Now, the moments of U distributions were given in Problem 11.6.1. Hence, E Wh

= E [U (p; m, n − c)]nh/2 p Γ[ 1 (n − c − p + k) + 12 nh] = K p k=1 1 2 1 j=1 Γ[ 2 (m + n − c − p + j) + 2 nh] p Γ[ 1 n(1 + h) + 12 (−c − p + k)] = K p k=1 1 2 , 1 j=1 Γ[ 2 n(1 + h) + 2 (m − c − p + j)]

which is of the form (12.1) with

202

12. Asymptotic expansions

a = p,

xk = 12 n,

ζk = 12 (−c − p + k),

ηj = 21 (m − c − p + j). p p Note, again, that xk and yj are O(n) and k=1 xk = j=1 yj is satisﬁed. Using (12.6) and (12.10), we have b = p,

f ρ

yj = 12 n,

= pm, = 1 − n−1 [c − 12 (m − p − 1)].

An interesting peculiarity in this case which derives from a symmetry property of Bernoulli polynomials [Erd´elyi et al. (1953), p. 37], namely Bα (1 − h) = (−1)α Bα (h), is that ω2α−1 = 0, α = 1, 2, . . . , which means the series involves only terms of even powers of n−1 [Lee (1972)]. To see this, ﬁrst note βk = (1 − ρ)xk βk + ζk

= =

1 1 2 [c − 2 (m − p − 1)], − 12 [ 12 (m + p − 1) − k],

and, similarly, k + ηk = 12 [ 12 (m − p + 1) + k]. Therefore, we ﬁnd (note that k → p − k + 1 reverses the order of terms in the following sums) p

B2α (βk + ζk )

=

k=1

= = =

p k=1 p k=1 p k=1 p

B2α − 12 [ 12 (m + p − 1) − k]

B2α − 12 [ 12 (m + p − 1) − (p − k + 1)]

B2α 1 + 12 [ 12 (m + p − 1) − (p − k + 1)] B2α (k + ηk ),

k=1

and from (12.7), ω2α−1 = 0, α = 1, 2, . . .. Thus, the expansion in Proposition 12.2 further reduces to P (ρM ≤ x) = Gf (x) + ω2 [Gf +4 (x) − Gf (x)] + ω4 [Gf +8 (x) − Gf (x)] + 12 ω22 [Gf +8 (x) − 2Gf +4 (x) + Gf (x)] + O(n−6 ),

(12.12)

where from the same symmetry property of Bernoulli polynomials, one can easily establish that ω2α =

p

22α B2α+1 12 [ 12 (m − p + 1) + k] , α = 1, 2, . . . . 2α α(2α + 1)(ρn) k=1

12.3. Examples

O(n−2 ) O(n−4 ) O(n−6 ) exact

n=2

n=5

n = 10

n = 15

n = 20

n = 30

.9500 .8107 .7168 .4714

.9500 .8848 .8642 .8315

.9500 .9220 .9182 .9139

.9500 .9345 .9334 .9322

.9500 .9402 .9397 .9393

.9500 .9451 .9449 .9448

203

Table 12.2. Asymptotic expansions for U (2; 12, n) distributions.

A small-scale numerical evaluation of (12.12) would help to determine how large n should be for the asymptotics of U (p; m, n) distributions to be accurate. Fix p = 2 and m = 12, and vary n =2, 5, 10, 15, 20, and 30. The asymptotic distribution of −n log U (2; 12, n) is χ224 . So, we choose the critical point x = χ2.95,24 = 36.41502. The evaluation of P − n + 12 (12 − 2 − 1) log U (2; 12, n) ≤ 36.41502 using (12.12) led to Table 12.2. The exact values were obtained for p = 2 with the transformation in Problem 11.6.2: P (−ρn log U (2; 12, n) ≤ x) = P U (2; 12, n) ≥ e−x/ρn (n − 1)(1 − y 1/2 ) = P F (24, 2(n − 1)) ≤ , 12y 1/2 with y = e−x/ρn . The approximations of order O(n−6 ) can thus be used in practice for n as small as 10 in this case. They are nearly exact to four decimal places for n = 30. Example 12.3 Test of mutual independence between subvectors. This is a continuation of Problem 11.6.7, where in item (ii), we found the ˜ = Λ2/n , moments of Λ r 1 Γpj ( 12 m) ˜ h = Γp ( 2 m + h) E Λ , Γp ( 12 m) j=1 Γpj ( 12 m + h)

with m = n − 1. This can be written in the form of (12.1) for W = Λ as p Γ[ 1 n(1 + h) − 12 k] h nh/2 ˜ pj 2 1 = K r k=1 E W =E Λ 1 , j=1 l=1 Γ[ 2 n(1 + h) − 2 l] with the identiﬁcation a = p, b=

r j=1

pj = p,

xk = 12 n,

ζk = − 12 k, k = 1, . . . , a,

yjl = 12 n,

ηjl = − 12 l, j = 1, . . . , r, l = 1, . . . , pj .

204

12. Asymptotic expansions

The constants f and ρ can be veriﬁed with (12.6) and (12.10): pj p r − 12 k − − 12 l = 12 Σ2 , f = −2 j=1 l=1

k=1

1−

ρ =

2Σ3 + 9Σ2 , 6nΣ2

r where Σs ≡ ps − j=1 psj . For simpliﬁed algebraic expressions of ω2 through ω6 , the reader is referred to Box (1949). Example 12.4 Test of equality of variances. The null moments of the modiﬁed likelihood ratio test Λ∗ for the hypothesis H0 : Σ1 = · · · = Σa were obtained in Proposition 8.17. Thus, for W = Λ∗ , we can write a Γp ( 12 m) Γp [ 12 mi (1 + h)] pmi h/2 Γ [ 1 m(1 + h)] Γp ( 21 mi ) p 2 i=1 i=1 mi h a p a 1 m/2 1 1 i=1 ( 2 m) i=1 l=1 Γ[ 2 mi (1 + h) − 2 (l − 1)] = K a p 1 p 1 1 mi /2 i=1 l=1 ( 2 mi ) j=1 Γ[ 2 m(1 + h) − 2 (j − 1)]

E Wh

a

=

mpmh/2

which is of the form (12.1) with the identiﬁcation a = pa,

xkl = 12 mk ,

ζkl = − 12 (l − 1), k = 1, . . . , a, l = 1, . . . , p,

b = p,

yj = 12 m,

ηj = − 12 (j − 1), j = 1, . . . , p.

The degrees of freedom f and ρ in (12.6) and (12.10) are p p a ζkl − ηj − 1 (pa − p) f = −2 2

k=1 l=1

= ρ

=

j=1

1 2 p(p

+ 1)(a − 1), p p a 2 1 x−1 yj−1 (ηj2 − ηj + 16 ) 1 − f −1 kl (ζkl − ζkl + 6 ) − k=1 l=1

=

1−

(2p2 + 3p − 1) 6(p + 1)(a − 1)

a

k=1

1 1 − mk m

j=1

.

Values of ωα can be calculated from (12.7) in simpliﬁed algebraic form but this is unnecessary since they can be easily programmed for the computer to evaluate the expansion in Proposition 12.2. Note ﬁnally that since we require (1 − ρ)xkl and (1 − ρ)yj to remain bounded, the expansion is asymptotic as m → ∞ while mk /m → αk for some proportions 0 < αk < 1 such a that k=1 αk = 1.

12.4. Problem

205

The basic idea in asymptotic expansions was to represent the cumulant generating function in the form (12.5). This is often possible even though the moments may not be of the form (12.1). Example 12.5 An example is provided by the modiﬁed likelihood ratio test for a given variance in Problem 8.9.8, where mph/2 Γp [ 12 m(1 + h)] 2e E Λ∗ h = (1 + h)−mp(1+h)/2 , m = n − 1. m Γp ( 12 m) For W = Λ∗ , Davis (1971) showed that Proposition 12.2 holds with f

=

1 2 p(p

ρ

=

1−

ωα

=

α+1 α + 2 2(−1)α (1 − ρ)α+1−s δs ( 12 m)1−s , (ω1 = 0). α(α + 1)(α + 2)ρα s=1 s + 1

+ 1),

2p2 + 3p − 1 , 6m(p + 1)

The δs are the same as those of Table 12.1. For asymptotic expansions of the null distribution of Lawley-Hotelling and Pillai trace tests, the reader is referred to Muirhead (1970) and Fujikoshi (1970).

12.4 Problem 1. This problem develops the asymptotic expansion of the LRT for the equality of means and variances H0 : µ1 = · · · = µa ; Σ1 = · · · = Σa between a multivariate normal populations. The LRT Λ, together with its moments, are given in Problem 8.9.14. Using the same notation and W = Λ, establish the following:

(i) The moments of W have the equivalent form h p a p 1 n/2 Γ[ 1 nl (1 + h) − 12 k] j=1 ( 2 n) k=1 h p l=1 1 2 E W = K p a 1 n /2 1 l k=1 l=1 ( 2 nl ) j=1 Γ[ 2 n(1 + h) − 2 j] (ii) Perform the usual identiﬁcation to conclude the validity of Proposition 12.2 with f

=

ρ =

1 2 (a

− 1)p(p + 3),

(2p2 + 9p + 11) 1− 6(a − 1)(p + 3)

a 1 1 . − n n i=1 i

13 Robustness

13.1 Introduction Many inference methods were presented in previous chapters for multivariate normal populations. A question of theoretical and utmost practical importance is the eﬀect of non-normality on the inference. For example, what happens if the likelihood ratio test of sphericity, derived assuming normality, is performed, but, in fact, the population follows a multivariate student distribution on 10 degrees of freedom? Is the signiﬁcance level of α = 5%, say, still close to 5%? The theory of robustness gives answers as to how sensitive multivariate normal inferences are to departures from normality. Most importantly, it proposes some remedies, i.e., more robust procedures. In Section 13.2, we present some non-normal models often used in robustness, the so-called elliptical distributions. The rest of the chapter is devoted to robust estimation and adjusted likelihood ratio tests. A robust analysis of data is useful in several ways. It can validate or rebuﬀ data analysis done on classical assumptions of multivariate normality. It also comes into play in the identiﬁcation of outliers, which is a challenging task for data sets with more than two variables. Robust estimates of location vector and scale matrix serve this role admirably. They can be used to evaluate robust Mahalanobis distances from an observation vector xi to the location vector. Points with large Mahalanobis distances can then be singled out and scrutinized.

13.2. Elliptical distributions

207

13.2 Elliptical distributions Suppose that x ∈ Rp has a density fx (x) = |Λ|−1/2 g[(x − µ) Λ−1 (x − µ)], where g : [0, ∞) → [0, ∞) is a ﬁxed function independent of µ and Λ = (Λij ) and depends on x only through (x − µ) Λ−1 (x − µ). Denote this elliptical distribution by x ∼ Ep (µ, Λ). The main reference for elliptical distributions is Kelker (1970). The aﬃne linear transformation y = Bx + b with B ∈ Gp and b ∈ Rp has density fy (y) = |BΛB |−1/2 g[(y − Bµ − b) (BΛB )−1 (y − Bµ − b)]. Thus, y ∼ Ep (Bµ + b, BΛB ); i.e., the transformation x → Bx + b induces the parameter transformation µ → Bµ + b and Λ → BΛB . In particular, z = Λ−1/2 (x − µ) ∼ Ep (0, I) has a spherical or rotationally invariant distribution. Elliptical distributions are a location scale generalization of spherical distributions. Thus, for example, if z ∼ Ep (0, I) with characteristic function necessarily of the form cz (t) = φ(t t) (v. Problem 4.6.6), then x = Λ1/2 z + µ ∼ Ep (µ, Λ) has characteristic function cx (t) = exp(it µ)φ(t Λt). Moreover, if z has a ﬁnite second moment, E z = 0 and var z = αI, for some constant α, implies E x = µ and var x ≡ Σ = αΛ. An important implication is that all elliptical distributions with ﬁnite second moments have the same correlation matrix. The constant α = −2φ (0) (v. Problem 4.6.15) is easily found by diﬀerentiation of cz (t). Examples of spherical distributions commonly used in robustness are members of the normal mixture family with density ∞ (2πw)−p/2 exp(− 12 w−1 x x)dF (w), fx (x) = 0

where F (·) is the “mixing” distribution function on [0, ∞). These can be d

|=

simulated easily using the representation x = w1/2 z, where w ∼ F , z ∼ z (v. Problem 13.6.1). Np (0, I), w Example 13.1 Obviously, P (w distribution.

= σ 2 ) = 1 yields the Np (0, σ 2 I)

Example 13.2 The two-point distribution, P (w = 1) = 1 − , P (w = σ 2 ) = for some “contamination” proportion 0 < < 1, yields the symmetric contaminated normal distribution. Example 13.3 The multivariate t on ν degrees of freedom denoted tp,ν is obtained with νw−1 ∼ χ2ν . The reader is asked to show in Problem 13.6.1

208

13. Robustness

that x has density fx (x) = cp,ν (1 + x x/ν)−(ν+p)/2 , x ∈ Rp , where cp,ν = (νπ)−p/2 Γ 12 (ν + p) /Γ 12 ν . The general multivariate tp,ν (µ, Λ) is obtained by relocating and rescaling, y = Λ1/2 x + µ, and has density −(ν+p)/2 fy (y) = cp,ν |Λ|−1/2 1 + (y − µ) Λ−1 (y − µ)/ν , y ∈ Rp . The multivariate t on 1 degree of freedom is also known as the multivariate Cauchy distribution. The Kotz-type distributions form another important class of elliptical distributions [Fang et al. (1991), p. 76]. Their characteristic function was obtained recently by Kotz and Ostrovskii (1994). Elliptical distributions that can be expanded as a power series are deﬁned in Steyn (1993) and used to deﬁne other nonelliptical distributions with heterogeneous kurtosis. The following result gives the marginal and conditional distributions for an Ep (µ, Λ) distribution. Let x = (x1 , x2 ) with xi ∈ Rpi , i = 1, 2, p = p1 + p2 , and partition µ and Λ in conformity as (µ , µ ) , 1 2 Λ11 Λ12 . Λ = Λ21 Λ22 µ =

Proposition 13.1 The marginal and conditional distributions of an Ep (µ, Λ) distribution are elliptical: (i) x2 ∼ Ep2 (µ2 , Λ22 ), (ii) x1 |x2 ∼ Ep1 (µ1.2 , Λ11.2 ), where µ1.2 Λ11.2

= µ1 + Λ12 Λ−1 22 (x2 − µ2 ), = Λ11 − Λ12 Λ−1 22 Λ21 .

The conditional variance is of the form var(x1 |x2 ) = w(x2 )Λ11.2 , for some function w(x2 ) ∈ R which depends on x2 only through the quadratic form (x2 − µ2 ) Λ−1 22 (x2 − µ2 ). Proof. Letting t = (0 , t2 ) in cx (t) = exp(it µ)φ(t Λt), we ﬁnd cx2 (t2 ) = exp(it2 µ2 )φ(t2 Λ22 t2 ) and, thus, x2 ∼ Ep2 (µ2 , Λ22 ). For the conditional distribution, let z = x1 − [µ1 + Λ12 Λ−1 22 (x2 − µ2 )] with jacobian J(x → z, x2 ) = 1. Upon using Problem 1.8.2, the conditional density z|x2 is −1 |Λ|−1/2 g[z Λ−1 11.2 z + (x2 − µ2 ) Λ22 (x2 − µ2 )] , |Λ22 |−1/2 fx2 (x2 )

13.2. Elliptical distributions

209

where fx2 (x2 ) depends only on (x2 − µ2 ) Λ−1 22 (x2 − µ2 ). Thus, we have z|x2 ∼ Ep1 (0, Λ11.2 ) and, in turn, x1 |x2 ∼ Ep1 (µ1.2 , Λ11.2 ). 2 Example 13.4 The univariate power exponential distribution has p.d.f. 2α −1/2 1 x − µ exp − 2 1/2 , α > 0. (13.1) fx (x) = c1,α Λ Λ A multivariate extension seems to be ( α ) . fx (x) = cp,α |Λ|−1/2 exp − 12 (x − µ) Λ−1 (x − µ)

(13.2)

This elliptical distribution has an advantage of generating distributions with heavier and lighter tails than the multivariate normal by taking α < 1 or α > 1, whereas many other elliptical distributions including the multivariate t cannot generate lighter-tail distributions. Kuwana and Kariya (1991) used this property to derive a locally best invariant test of multivariate normality (α = 1). Taking α = 0.5 simply, in (13.2), E (x1 − µ1 )2 = 4(p + 1)Λ11 , which depends on p (v. Problem 13.6.5); the corresponding moment in (13.1) is E (x − µ)2 = 8Λ11 , with Λ = Λ11 . So, the marginal distribution of x1 in (13.2) is not that of x in (13.1). The inconsistency takes place for many other elliptical distributions. Kano (1994) characterized the consistency property of elliptical distributions: An elliptical family is consistent iﬀ it is a normal mixture family. In particular, the multivariate normal and multivariate t families are consistent. In Proposition 13.1 the marginal is elliptical but possibly of a diﬀerent functional form since the characteristic function φ may be related to p. For the estimation of (µ, Λ), it seems natural to ask that location and scatter estimates transform in exactly the same manner as the parameters; i.e., that they be “aﬃne equivariant” as described in the following deﬁnition. Formally, let x1 .. X = . ∈ Rnp xn

be the sample matrix. ˆ ˆ Deﬁnition 13.1 The location and scatter estimates µ(X) and Λ(X) are aﬃne equivariant iﬀ for all B ∈ Gp and b ∈ Rp , ˆ + 1b ) µ(XB ˆ + 1b ) Λ(XB

ˆ = Bµ(X) + b, ˆ = BΛ(X)B .

210

13. Robustness

When the underlying distribution belongs to an elliptical family, the distribution of aﬃne equivariant estimates has a special structure. In particular, the general form of the mean and variance estimates can be characterized for ﬁnite samples. To establish this general form, we need to extend the notion of rotational invariance of random vectors in Section 4.4 to symmetric random matrices. Deﬁnition 13.2 A random symmetric matrix W is rotationally invariant d iﬀ W = HWH , ∀H ∈ Op . The following lemma [Tyler (1982)] characterizes the general form of the mean and variance of any rotationally invariant random matrix. Proposition 13.2 Let W ∈ Rpp symmetric be rotationally invariant with ﬁnite second moments. Then, there exist constants η, σ1 ≥ 0, and σ2 ≥ −2σ1 /p such that E W var W

= ηI, = σ1 (I + Kp ) + σ2 vec(I)[vec(I)] .

Proof. For the mean, let E W ≡ A. By rotational invariance, A = HAH , ∀H ∈ Op . Hence, x Ax = x HAH x = y Ay, ∀x, y ∈ Rp , |x| = |y| = 1. Choosing x = hi and y = hj , the ith and jth eigenvectors of A corresponding to eigenvalues λi and λj , respectively, we get λi = λj ≡ η (say). This means A = ηI. For the variance, let Ω ≡ var W = Ωijkl ei ej ⊗ ek el , where cov(wki , wlj ) = Ωijkl . Note that {ei ej ⊗ ek el , i, j, k, l = 1, . . . , p} 2

d

d

forms a basis for Rpp2 . Since W = HWH , ∀H ∈ Op , then vec(W) = (H ⊗ H)vec(W) and, thus, Ω = (H ⊗ H)Ω(H ⊗ H ), or Ωijkl ei ej ⊗ ek el , Ωijkl hi hj ⊗ hk hl = where H = (h1 , . . . , hp ). By choosing for some m, hm = −em and hr = er , r = m, we obtain Ωijkl = 0 unless i = j = k = l, i = j and k = l, i = k and j = l, or i = l and j = k. By choosing H to give a permutation of the rows, we obtain Ωiiii = σ0 , ∀i = 1, . . . , p, Ωiikk = σ1 for i = k, Ωijij = σ2 for i = j, and Ωijji = σ3 for i = j. Thus, ei ei ⊗ ei ei + σ1 ei ei ⊗ ek ek Ω = σ0 i

+σ2

i =j

i =k

ei ej ⊗ ei ej + σ3

i =j

ei ej ⊗ ej ei

13.2. Elliptical distributions

211

= σ1 I + σ2 vec(I)[vec(I)] + σ3 Kp +(σ0 − σ1 − σ2 − σ3 ) ei ei ⊗ e i ei . i

Since ∀H ∈ Op , (H ⊗ H)I(H ⊗ H ) = I, (H ⊗ H)vec(I)[vec(I)] (H ⊗ H ) = vec(I)[vec(I)] , (H ⊗ H)Kp (H ⊗ H ) = Kp , and (H ⊗ H)

ei ei

⊗

ei ei

(H ⊗ H ) =

i

ei ei

⊗

ei ei

,

i

for some H ∈ Op , it follows that σ0 − σ1 − σ2 − σ3 = 0. Also, since W is symmetric, cov(wij , wji ) = var wij , which implies σ1 = σ3 . Therefore, Ω = σ1 (I + Kp ) + σ2 vec(I)[vec(I)] . The conditions on σ1 and σ2 follow since Ω is positive semideﬁnite.

2

The variance of W = (wij ) can be written componentwise with the Kronecker delta cov(wki , wlj ) = σ1 (δij δkl + δkj δil ) + σ2 δki δlj . The form of var W states that the oﬀ-diagonal elements of W are uncorrelated with each other and uncorrelated with the diagonal elements. Each oﬀ-diagonal element has variance σ1 . The diagonal elements all have variance 2σ1 + σ2 with the covariance between any two diagonal elements being σ2 . Example 13.5 A simple example is W ∼ Wp (m) which is rotationally invariant with var W = m(I + Kp ). Example 13.6 Assume x ∼ Ep (0, Λ) and let W = xx . Then var W = (Λ1/2 ⊗ Λ1/2 )var(zz )(Λ1/2 ⊗ Λ1/2 ), where z ∼ Ep (0, I). Using Proposition 13.2 var(zz ) is evaluated with σ1 σ2

= =

var(z1 z2 ) = E(z12 z22 ) = µ22 ,

cov(z12 , z22 ) = E(z12 z22 ) − E(z12 )E(z22 ) = µ22 − µ22 .

In terms of cumulants we have σ1 = k22 + k22 and σ2 = k22 . These cumulants are easily found with the Taylor series ln φ(t21 + t22 )

(it1 )2 (it2 )2 (it1 )4 (it2 )4 + k2 + k4 + k4 2! 2! 4! 4! (it1 )2 (it2 )2 + o(|t|4 ). +k22 2! 2!

= k2

212

13. Robustness

The reader can verify by diﬀerentiation (v. Problem 13.6.3) k2 k4 k22

= −2φ (0) = α, = =

12(φ (0) − φ (0)2 ), 4(φ (0) − φ (0)2 ).

The kurtosis of z1 is k4 (φ (0) − φ (0)2 ) = 3 ≡ 3k, k22 φ (0)2 where k represents a kurtosis parameter. Thus, k4 = 3kα2 and k22 = kα2 . Finally, we obtain σ1 = (1 + k)α2 and σ2 = kα2 from which var(zz ) = α2 (1 + k)(I + Kp ) + α2 k vec(I)[vec(I)] and var W = (1 + k)(I + Kp )(Σ ⊗ Σ) + k vec(Σ)[vec(Σ)] , where Σ = αΛ is the variance of x. ˆ ˆ Corollary 13.1 If µ(X) and Λ(X) are aﬃne equivariant with ﬁnite second moment and x1 , . . . , xn are i.i.d. Ep (µ, Λ), then there exist constants η, β ≥ 0, σ1 ≥ 0 and σ2 ≥ −2σ1 /p such that ˆ E µ(X) ˆ var µ(X) ˆ E Λ(X)

= µ,

ˆ var Λ(X)

= σ1 (I + Kp )(Λ ⊗ Λ) + σ2 vec(Λ)[vec(Λ)] .

= βΛ, = ηΛ,

d

Proof. First, X = ZΛ1/2 + 1µ , where z1 .. Z= . zn

d

1/2 ˆ ˆ ˆ and zi ’s are i.i.d. Ep (0, I). Hence, µ(X) = µ(ZΛ +1µ ) = Λ1/2 µ(Z)+µ. ˆ Obviously, µ(Z) is a rotationally invariant random vector. Using the result ˆ ˆ of Section 4.5, E µ(Z) = 0 and var µ(Z) = βI, for some β ≥ 0. Therefore, d 1/2 ˆ ˆ ˆ ˆ = Λ1/2 Λ(Z)Λ , where E µ(X) = µ and var µ(X) = βΛ. Similarly, Λ(X) ˆ Λ(Z) is a rotationally invariant matrix whose mean and variance have the ˆ general form in Proposition 13.2. Hence, E Λ(X) = ηΛ, for some η, and ! ˆ ˆ var Λ(X) = var (Λ1/2 ⊗ Λ1/2 )vec(Λ(Z)) ! ˆ = (Λ1/2 ⊗ Λ1/2 ) var Λ(Z) (Λ1/2 ⊗ Λ1/2 )

= σ1 (I + Kp )(Λ ⊗ Λ) + σ2 vec(Λ)[vec(Λ)]

13.3. Maximum likelihood estimates

for some σ1 ≥ 0, σ2 ≥ −2σ1 /p.

213

2

Complicated expressions using tensor methods for third-order and fourth-order cumulants of aﬃne equivariant estimates in elliptical families were obtained by Gr¨ ubel and Rocke (1990). ˆ Another way of writing var Λ(X) is to give the covariances between any ˆ ˆ ij ): two elements of Λ(X) = (Λ ˆ ki , Λ ˆ lj ) = σ1 (Λij Λkl + Λkj Λil ) + σ2 Λki Λlj . cov(Λ One should note that a reasonable estimate of Λ assumed positive deﬁnite ˆ should satisfy Λ(X) > 0 w.p.1, and in that case, η > 0.

13.3 Maximum likelihood estimates Assume x1 , . . . , xn i.i.d. x ∼ Ep (µ, Λ) with var x = αΛ = Σ. The simplest but ineﬃcient method to estimate (µ, Λ) would be to use the MLE under a ˆ = 1 n (xi − x ¯ and Σ ¯ )(xi − x ¯ ) . A more eﬃcient Np (µ, Σ) distribution, x i=1 n procedure would be the MLE under the “true” Ep (µ, Λ) model. These two possibilities are now investigated.

13.3.1

Normal MLE

When x has ﬁnite fourth-order moments, the general discussion of Section 6.3 showed that ˆ −Σ var W 0 d Σ p+1 n1/2 → N 0, , p 0 Σ ¯ − µ x where W = xx . From the calculation of var W in Example 13.6, it follows that

where n

|=

d ˆ − Σ, x ¯ − µ) → (N, n), n1/2 (Σ

N,

n ∼ Np (0, Σ) N ∼ Npp (0, (1 + k)(I + Kp )(Σ ⊗ Σ) + k vec(Σ)[vec(Σ)] ).

13.3.2 Elliptical MLE For x ∼ Ep (µ, Λ) deﬁned with a known function g(·) the log-likelihood for (µ, Λ) is simply ln (µ, Λ) = cte +

n i=1

ln g[(xi − µ) Λ−1 (xi − µ)] − 12 n ln |Λ|.

(13.3)

214

13. Robustness

Diﬀerention with respect to µ and Λ (v. Problems 1.8.9 and 1.8.10) leads to the equations n

ˆ −1 (xi − µ) ˆ u(si )Λ

= 0

ˆ −1 (xi − µ)(x ˆ −1 ˆ −1 − 1 nΛ ˆ i − µ) ˆ Λ u(si )Λ 2

= 0,

i=1 1 2

n i=1

ˆ −1 (xi − µ). ˆ Λ ˆ Thus, the MLE where u(s) = −2g (s)/g(s) and si = (xi − µ) ˆ estimating equations ˆ Λ)) satisﬁes the implicit (because si depends on (µ, ˆ = µ ˆ Λ =

ave [u(si )xi ] /ave [u(si )] ,

(13.4)

ˆ i − µ) ˆ ]. ave [u(si )(xi − µ)(x

(13.5)

The notation “ave” means arithmetic average over i = 1, . . . , n. Example 13.7 The multivariate Student’s tp,ν has g(s) ∝ (1+s/ν)−(ν+p)/2 and u(s) = (ν + p)/(ν + s). Note that u(s) ≥ 0 and is strictly decreasing. It acts as a weight function, giving more weight to data points with small squared Mahalanobis distances. The existence and unicity of a solution to the estimating equations is a diﬃcult problem. For the location-only problem, it is known in the univariate case [Reeds (1985)] that the estimating equation is susceptible to multiple solutions. Uniqueness of the solution in the univariate locationscale Cauchy (ν = 1) problem was established by Copas (1975) and for ν > 1 by M¨ arkel¨ ainen et al. (1981). The approach presented here is that of Kent and Tyler (1991), which works equally well in the multivariate case. The location-scale problem is very tricky, but the scale-only problem is quite simple. We will thus concentrate on the latter problem. Scale-only problem For the scale-only problem, we assume without any loss of generality that µ = 0. The log-likelihood reduces to l(A) = cte +

n

ln g(xi A−1 xi ) − 12 n ln |A|

i=1

and the estimating equation simpliﬁes to ˆ = ave [u(si )xi x ] , A i

(13.6)

ˆ −1 xi . Let ψ(s) = su(s) and assume where u(·) is as before and si = xi A that lim ψ(s) = a0 > 0.

s→∞

13.3. Maximum likelihood estimates

215

This condition is satisﬁed for the tp,ν distribution, as lims→∞ ψ(s) = ν + p. The following condition on the data is to ensure the existence of a solution to (13.6). It speciﬁes that the data points should not be too concentrated in low-dimensional linear subspaces of Rp . Let Pn (·) denote the empirical distribution of x1 , . . . , xn , i.e., for any borel set B ⊂ Rp 1 I(xi ∈ B). n i=1 n

Pn (B) =

Condition D. For all linear subspaces V ⊂ Rp with dim V ≤ p − 1, Pn (V) < 1 − [p − dim V]/a0 .

The existence of a solution under condition D is proved in Kent and Tyler (1991). Proving existence is the most diﬃcult part, but uniqueness of the solution and convergence of a numerical algorithm is much simpler. ˆ > 0 such that Proposition 13.3 Under condition D, there exists A ˆ ≤ l(A), ∀A > 0. l(A) Note that when sampling from an absolutely continuous distribution condition D is satisﬁed w.p.1 for a0 > p and sample sizes n ≥ p since for any subspace V, k = dim V ≤ p − 1, Pn (V)

w.p.1 k k (p − k) (p − k) ≤ . ≤ =1− 1. Since su(s) is strictly increasing and u(s) is nonincreasing, it follows that for x = 0, u(x A−1 x) ≤ u(λ−1 1 x x) ≤

λ−1 1 x x u(λ−1 1 x x) < λ1 u(x x), λ−1 1 x x

where the ﬁrst inequality used Rayleigh’s quotient. This implies A = ave u(xi A−1 xi )xi xi < λ1 ave [u(xi xi )xi xi ] = λ1 I.

216

13. Robustness

This gives the contradiction λ1 < λ1 , and so λ1 ≤ 1. A similar argument shows λp ≥ 1. Thus, A = I. 2 Under conditions D and M, the unique solution can be found by regarding the estimating equation as a ﬁxed-point equation. Given a starting value A0 > 0, deﬁne the iterative numerical algorithm Am+1 = ave u(xi A−1 m xi )xi xi , m = 0, 1, . . . . Proposition 13.5 Under conditions D and M, for any A0 > 0, Am converges as m → ∞ to the unique solution (the MLE) of (13.6). Proof. Conditions D and M ensure existence and uniqueness of a solution ˆ Since xi → Bxi , B ∈ Gp , induces the new solution A ˆ → BAB ˆ , A. ˆ one can assume without loss of generality that A = I is the solution. Let λ1,m ≥ · · · ≥ λp,m be the eigenvalues of Am , m = 1, 2, . . .. Step 1: The following results are established: (i) λ1,m ≤ 1 =⇒ λ1,m+1 ≤ 1, (ii) λ1,m > 1 =⇒ λ1,m+1 < λ1,m , (iii) λp,m ≥ 1 =⇒ λp,m+1 ≥ 1, and (iv) λp,m < 1 =⇒ λp,m+1 > λp,m . Note that (iii) and (iv) imply Am+1 > 0 whenever Am > 0. To prove (i), if λ1,m ≤ 1, then −1 x A−1 m x ≥ λ1,m x x ≥ x x, ˆ and since u(s) is nonincreasing, u(x A−1 m x) ≤ u(x x). Given that A = I is the solution, this implies Am+1 ≤ ave [u(xi xi )xi xi ] = I. Thus, λ1,m+1 ≤ 1. −1 The proof of (iii) is similar. To prove (ii), since x A−1 m x ≥ λ1,m x x, u(s) is nonincreasing and su(s) is strictly increasing, it follows that if λ1,m > 1, then −1 u(x A−1 m x) ≤ u(λ1,m x x) ≤ λ1,m u(x x),

with the second inequality strict for x = 0. This implies Am+1 < λ1,m ave [u(xi xi )xi xi ] = λ1,m I. Thus, λ1,m+1 < λ1,m . The proof of (iv) is similar. Step 2: We shall now show that (v) lim sup λ1,m ≤ 1, (vi) lim inf λp,m ≥ 1, from which it follows that λ1,m → 1 and λp,m → 1, so that Am → I (v. Problem 1.8.13). Given A > 0, let λ1 (A) denote the largest eigenvalue and deﬁne φ(A) = ave u(xi A−1 xi )xi xi .

13.3. Maximum likelihood estimates

217

Step 1 implies that if λ1 (A) > 1 and B = φ(A), then λ1 (B) < λ1 (A). In view of step 1, statement (v) requires proof only in the case in which λ1,m = λ1 (Am ) > 1, ∀m. Note that λ1,m is a decreasing sequence in this case. Let λ∗ = lim λ1,m ≥ 1 and suppose, if possible, that λ∗ > 1. From step 1, the eigenvalues of the sequence Am are bounded away from 0 and ∞. Thus, we can ﬁnd a convergent subsequence Amj → B0 say, where B0 > 0. Further, Amj +1 = φ(Amj ) → φ(B0 ) = B1 , say. Since λ1,m is decreasing, λ1 (B0 ) = lim λ1,mj = λ∗ and λ1 (B1 ) = lim λ1,mj +1 = λ∗ . However, step 1 implies that λ1 (B1 ) < λ1 (B0 ), giving a contradiction. Hence, (v) follows. Item (vi) is proved similarly. 2 Location-scale problem Results for location scale are derived by embedding the p-dimensional location-scale problem into a (p + 1)-dimensional scale-only problem. For given Λ ∈ Pp , µ ∈ Rp , and γ > 0, let Λ + γ −1 µµ γ −1 µ A= (13.7) ∈ Rp+1 p+1 γ −1 γ −1 µ and observe that any A ∈ Pp+1 can be written in this form. On using the inverse of a partitioned matrix (v. Problem 1.8.1), one ﬁnds Λ−1 −Λ−1 µ A−1 = . −µ Λ−1 γ + µ Λ−1 µ Now deﬁne the artiﬁcial vectors yi = (xi , 1) ∈ Rp+1 and note that yi A−1 yi = (xi − µ) Λ−1 (xi − µ) + γ.

(13.8)

Let A(1) be deﬁned as in (13.7) but with γ = 1. Upon using (13.8) and |A(1) | = |Λ|, the objective function (13.3) can be expressed as ln (µ, Λ) = l(A(1) ) = cte +

n

1 ln g(yi A−1 (1) yi − 1) − 2 n ln |A(1) |.

(13.9)

i=1

Thus, the problem of maximizing (13.3) over µ ∈ Rp and Λ ∈ Pp is equivalent to maximizing l(A(1) ) over A(1) ∈ Pp+1 with the restriction that the (p + 1, p + 1) element of A(1) be 1. Moreover, the estimating equations (13.4) and (13.5) can be rewritten in a single estimating equation as −1 ˆ ˆ ˆ ˆ µ γˆ −1 µ ˆ = Λ + γˆ µ (13.10) = ave[u(si )yi yi ], A ˆ γˆ −1 γˆ −1 µ ˆ −1 (xi − µ), ˆ Λ ˆ as in the original where γˆ −1 = ave[u(si )] with si = (xi − µ) location-scale formulation. Using (13.8), the single estimating equation can be reexpressed as ! ˆ = ave u∗ (y A ˆ −1 yi ; γˆ )yi y , (13.11) A i i

218

13. Robustness

where u∗ (s; γ) = u(s − γ), for s ≥ γ. This looks very similar to the estimating equation of a scale-only problem, the diﬀerence being that the function u∗ (·; γˆ ) depends on the data through γˆ . The next condition for existence of a solution is just the previous condition D on yi ’s recast in terms of xi ’s. Condition D1. For all translated linear subspaces (hyperplanes) H ⊂ Rp with dim H ≤ p − 1, Pn (H) < 1 − (p − dim H)/a0 .

This time if a0 > p + 1, n ≥ p + 1, then condition D1 is satisﬁed w.p.1 when sampling from an absolutely continuous distribution. Proposition 13.6 If conditions D1 and M hold, then there exists a soluˆ > 0 to (13.11). This solution is unique if (s + γˆ )u(s) is ˆ ∈ Rp and Λ tion µ strictly increasing in s ≥ 0 for γˆ −1 = ave[u(si )] deﬁned above. A diﬃculty in applying Proposition 13.6 is the strictly increasing condition ˆ the ˆ Λ) which depends on the unknown γˆ . However, given a solution (µ, condition guarantees that no other solutions exist. For the tp,ν distribution, ν ≥ 1, we prove that γˆ is independent of the data (ˆ γ = 1) and the condition is thus automatically satisﬁed for ν > 1 since (s + γˆ )u(s) = (ν + p)(s + 1)/(s + ν) is strictly increasing. Lemma 13.1 For the tp,ν distribution ν ≥ 1, γˆ = 1. Proof. If γu ≥ γ and (s + γu )u(s) is strictly increasing and condition M ˆ −1 and holds, then (s + γ)u(s) is also strictly increasing. Multiplying by Λ taking the trace of (13.5), we get ave[si u(si )] = p. Thus, ∀b > 0, p = ave[(si + b)u(si )] − bˆ γ −1 , which implies γl ≤ γˆ ≤ γu , where γu−1

=

γl−1

=

sup inf [(s + b)u(s) − p]/b, b>0 s>0

inf sup[(s + b)u(s) − p]/b.

b>0 s>0

Letting b = ν, we obtain 1 ≤ γl ≤ γˆ ≤ γu ≤ 1.

2

The Cauchy case, ν = 1, has (s + 1)u(s) = p + 1, which is not strictly increasing. It requires a special treatment, but the MLE is also unique under condition D1 [Kent and Tyler (1991)]. For the tp,ν case, since γˆ is independent of the data, this means that when condition D1 is satisﬁed, the ﬁxed-point algorithm still converges to the MLE. So, for any starting

13.3. Maximum likelihood estimates

219

values µ0 and Λ0 > 0, the iterative equations & ' ave u (xi − µm ) Λ−1 m (xi − µm ) xi & ' , µm+1 = ave u (xi − µm ) Λ−1 m (xi − µm ) ' & Λm+1 = ave u (xi − µm ) Λ−1 m (xi − µm ) (xi − µm )(xi − µm ) converge to the MLE. Asymptotics for the MLE The general theory of maximum likelihood coupled with the fact that the MLE is aﬃne equivariant tells us that for some constants β, σ1 , and σ2 , d ˆ − Λ, µ ˆ − µ) → (N, n), n1/2 (Λ

where n ∼ Np (0, βΛ) N ∼ Npp (0, σ1 (I + Kp )(Λ ⊗ Λ) + σ2 vec(Λ)[vec(Λ)] ) . |=

Using Fisher’s information, these constants can now be evaluated and we can also show that N n; i.e., they are asymptotically independent. The score function is the derivative of l(µ, Λ) = cte + ln g[(x − µ) Λ−1 (x − µ)] −

1 2

ln |Λ|

|=

with respect to (µ, Λ) and its variance is called Fisher’s information and is denoted by I(µ, Λ). It is also well known that the asymptotic variance is the inverse of Fisher’s information. Let us show that I(µ, Λ) is block-diagonal and thus N n. We have ∂l/∂µ = u(s)Λ−1 (x − µ) ∂l/∂Λ = − 12 Λ−1 + 12 u(s)Λ−1 (x − µ)(x − µ) Λ−1 , where s = (x − µ) Λ−1 (x − µ). The constants β, σ1 , and σ2 being independent of (µ, Λ), it suﬃces to evaluate the variance of the score while d

assuming (µ, Λ) = (0, I) and x = z ∼ Ep (0, I). The expectation E{∂l/∂µi · ∂l/∂Λjk } involves only ﬁrst-order and third-order product moments of z, which is spherical. Since these moments are all null, it follows that I(0, I) is blockdiagonal with blocks I1 and I2 , say. We then calculate β from I1−1 . Now, & ' I1 = E (∂l/∂µ) (∂l/∂µ) = E u2 (s)zz = E[su2 (s)]E[uu ], d

|=

where we have let z = s1/2 u, where s = |z|2 , u ∼ unif(S p−1 ), and u. Then, I1 = p−1 E[su2 (s)]I if we note that E uu = p−1 I (v. Probs lem 13.6.4). Thus, we have shown that β = p/E[su2 (s)]. We now evaluate

220

13. Robustness

σ1 and σ2 from I2 . We would like to identify var N with I2−1 , but we must ﬁrst eliminate the redundant elements of the symmetric N for var N to become nonsingular. For this reason, deﬁne Aj

(0, Ij ) : p × j, j = 1, . . . , p, diag(Ap , . . . , A1 ) : p2 × 12 p(p + 1),

= =

Mp

and verify that for any symmetric A ∈ Rpp , Mp vec(A) is the 12 p(p + 1)dimensional vector formed by stacking the columns of A after deleting the upper triangular part of A. Now, var(Mp vec(N)) = Mp var(N)Mp . It is easy to check that Mp Mp = I, Mp Kp Mp = Dp , Mp vec(Ip ) = ap , where αj

=

(1, 0, . . . , 0) ∈ Rj , j = 1, . . . , p,

ap

= =

(αp , . . . , α1 ) : 12 p(p + 1) × 1, diag(ap ).

Dp Then, we can identify

I2−1 = σ1 (I + Dp ) + σ 2 ap ap . Using the inverse of a perturbed matrix (v. Problem 1.8.8), we have with the relations (I + Dp )−1 ap = 12 ap and ap ap = p, I2

= σ1−1 (I + Dp )−1 − σ2 [4σ12 (1 + 12 pσ2 σ1−1 )]−1 ap ap = i1 (I + Dp )−1 + i2 ap ap ,

where i1 = σ1−1 ,

i2 = −σ2 [4σ12 (1 + 12 pσ2 σ1−1 )]−1 .

(13.12)

As an example for p = 2, we thus have the identiﬁcation I2

= i1 (I + Dp )−1 + i2 ap ap 1 0 i2 2 i1 + i2 = 0 i1 0 1 0 2 i1 + i2 i2 2 ∂l ∂Λ11

∂l ∂Λ11

∂l ∂l ∂l = E ∂Λ11 ∂Λ21 ∂Λ21 ∂l ∂Λ11

∂l ∂Λ22

Thus, in general for i = j, i1 i2

∂l ∂Λ21

∂l ∂Λ21 2

∂l ∂Λ22

∂l ∂Λ11 ∂l ∂Λ21

' & = E (∂l/∂Λij )2 , = E {(∂l/∂Λii )(∂l/∂Λjj )} .

∂l ∂Λ22

∂l ∂Λ22 ∂l ∂Λ22 2

.

13.3. Maximum likelihood estimates

221

These are evaluated with d

∂l/∂Λij

= u(s)zi zj = ψ(s)ui uj ,

∂l/∂Λii

= − 12 + 12 u(s)zi2 = − 12 + 12 ψ(s)u2i . d

Using Problem 13.6.4, E u2i = p−1 and E u2i u2j = [p(p + 2)]−1 , i = j, the ﬁnal result is thus i1 i2

= [p(p + 2)]−1 E ψ 2 (s), = − 14 + [p(p + 2)]−1 E ψ 2 (s)

if we note that E ψ(s) = p. The constants σ1 and σ2 are obtained by solving equation (13.12). The density of s was given in Problem 4.5.13. We have proved that under regularity conditions for the MLE [Lehmann (1983), pp. 429-430] Proposition 13.7

where N

|=

d ˆ − Λ, µ ˆ − µ) → (N, n), n1/2 (Λ

n and

n ∼ Np (0, βΛ) , N ∼ Npp (0, σ1 (I + Kp )(Λ ⊗ Λ) + σ2 vec(Λ)[vec(Λ)] ) , with β σ1 σ2

= p/E[su2 (s)], = p(p + 2)/E[ψ 2 (s)], = −2σ1 (1 − σ1 )/[2 + p(1 − σ1 )]

and s has density π p/2 1 p−1 g(s), s > 0. s2 Γ( 12 p) The parameter σ1 of the asymptotic variance will play a major role as an index of relative eﬃciency for robust tests. Example 13.8 For the tp,ν distribution, the reader can check σ1 = 1 + 2/(p + ν). The maximum likelihood estimation of the multivariate tp,ν distribution with possibly missing data and unknown degrees of freedom was treated by Liu (1997). Missing data imputation using the multivariate tp,ν distribution was also the subject of Liu (1995).

222

13. Robustness

13.4 Robust estimates An alternative approach to MLE consists of robust location and scatter estimates such as the M estimate [Maronna (1976), Huber (1981)] or the S estimate [Davies (1987), Lopuha¨a (1989)]. The theoretical proofs for existence, unicity, consistency, and asymptotic normality of these estimates go beyond the scope of this book. Of importance to us, however, is to show how √ easily these aﬃne equivariant and n-asymptotically normal estimates can serve as the building block to robust tests on location and scatter. They are succintly introduced now and invoked later to construct robust tests.

13.4.1

M estimate

Let x1 , . . . , xn i.i.d. x ∼ Ep (µ, Λ) and z ∼ Ep (0, I). The idea behind M estimate is to modify the MLE estimating equations to gain robustness. The M estimate of location and scatter are deﬁned as solution to the equations µn

=

Vn

=

ave [u1 (ti )xi ] /ave [u1 (ti )] , ave u2 (t2i )(xi − µn )(xi − µn ) ,

(13.13) (13.14)

1/2 . where ti = (xi − µn ) Vn−1 (xi − µn ) The M estimates are obviously aﬃne equivariant. Interestingly, they include, as a particular case, the MLE estimate with the functions u1 (t) = −2g (t2 )/g(t2 ) and u2 (t2 ) = u1 (t). Deﬁne ψi (s) = sui (s), i = 1, 2. The following conditions on the functions are needed and will always be assumed: M1. u1 and u2 are non-negative, nonincreasing, and continuous on [0, ∞). M2. ψ1 and ψ2 are bounded. Let Ki = sups≥0 ψi (s). M3. ψ2 is nondecreasing and is strictly increasing in the interval where ψ2 < K2 . M4. There exists s0 such that ψ2 (s20 ) > p and that u1 (s) > 0 for s ≤ s0 (and, hence, K2 > p). Example 13.9 The tp,ν MLE has ψ1 (s) = (ν + p)s/(ν + s2 ) and ψ2 (s) = (ν + p)s/(ν + s). It is easy to verify M1 through M4. Example 13.10 Huber’s ψ function is deﬁned as ψ(s, k) = max[−k, min(s, k)]. Let k > 0 be a constant and take ψ1 (s) = ψ(s, k) and ψ2 (s) = ψ(s, k2 ). A further condition on the data is needed for existence of the M estimate.

13.4. Robust estimates

223

Condition D2. There exists a > 0 such that for every hyperplane H, dim H ≤ p − 1, p Pn (H) ≤ 1 − − a. K2 When sampling from an absolutely continuous distribution, condition D2 is satisﬁed w.p.1 for n suﬃciently large. Proposition 13.8 If condition D2 is satisﬁed, there exists a solution (µn , Vn ) to (13.13) and (13.14). Moreover, µn belongs to the convex hull of {x1 , · · · , xn }. Proposition 13.9 Assume condition D2 and g is decreasing. Let (µn , Vn ) be a solution to (13.13) and (13.14), then (µn , Vn ) → (µ, V) almost surely, where V = σ −1 Λ with σ being the solution to E ψ2 (σt2 ) = p and t = |z|. The reason for the presence of σ is that Vn is consistent for a certain multiple of Λ, σ −1 Λ say, deﬁned by the implicit equation V = E u2 [(x − µ) V−1 (x − µ)](x − µ)(x − µ) . Multiplying by V−1 and taking trace yields E ψ2 (σ|z|2 ) = p. This expectation can be evaluated as a simple integral if one recalls the density of t = |z| (v. Problem 4.6.13): f (t) =

2π p/2 p−1 2 t g(t ), t ≥ 0. Γ( 12 p)

Proposition 13.10 Assume sψi (s) are bounded (i = 1, 2) and g is decreasing such that E ψ1 (σ 1/2 t) > 0. Then, n1/2 (Vn − V, µn − µ) → (N, n), where n

|=

d

N and

n ∼ Np 0, (α/β 2 )V , N ∼ Npp (0, σ1 (I + Kp )(V ⊗ V) + σ2 vec(V)[vec(V)] ) ,

with σ being the solution to E ψ2 (σt2 ) = p, where α β σ1 σ2

= p−1 E ψ12 (σ 1/2 t), !

= E 1 − p−1 u1 (σ 1/2 t) + p−1 ψ1 (σ 1/2 t) , = a1 (p + 2)2 (2a2 + p)−2 , & ' (a1 − 1) − 2a1 (a2 − 1)[p + (p + 4)a2 ](2a2 + p)−2 , = a−2 2

and a1 a2

= [p(p + 2)]−1 E ψ22 (σt2 ), = p−1 E [σt2 ψ2 (σt2 )].

224

13. Robustness

Those results are due to Maronna (1976), but Tyler (1982) found the asymptotic variance parameters σ1 and σ2 in Proposition 13.10. Asymptotic theory for robust principal components was developed by Tyler (1983b) and Boente (1987).

13.4.2

S estimate

Recently, Davies (1987) and Lopuha¨ a (1989) investigated properties of the S estimate for multivariate location and scatter. As before, consider a random sample x1 , . . . , xn i.i.d. x ∼ Ep (µ, Λ) and z ∼ Ep (0, I). Again, let t = |z|. In the context of regression, Rousseeuw and Yohai (1984) obtained an asymptotically normal and robust estimate from a function ρ assumed to satisfy the following: S1: ρ is symmetric, has a continuous derivative ψ, and ρ(0) = 0. S2: There exists a ﬁnite constant c0 > 0 such that ρ is strictly increasing on [0, c0 ] and constant on [c0 , ∞). Let a0 = sup ρ. A typical ρ function is Tukey’s biweight 2 t /2 − t4 /(2c20 ) + t6 /(6c40 ) if |t| ≤ c0 ρ(t) = c20 /6 if |t| ≥ c0 . The S estimate (µn , Vn ) is deﬁned as the solution of the optimization 1/2 problem where ti = (xi − µn ) Vn−1 (xi − µn ) : 1 ρ(ti ) = b0 n i=1 n

min |Vn | subject to

over all µn ∈ Rp and Vn > 0. The constant b0 , 0 < b0 < a0 , chosen so that 0 < b0 /a0 ≡ r ≤ (n − p)/2n, leads to a ﬁnite-sample breakdown point [Lopuha¨ a and Rousseeuw (1991)] of ∗n = nr/n. The choice r = (n−p)/2n results in the maximal breakdown point (n − p + 1)/2!/n (asymptotically 50%). Roughly speaking, the breakdown point is the minimum percentage of contaminated data necessary to bring the estimate beyond any given bound. The sample mean requires only one point and thus has a breakdown point 1/n, or asymptotically 0%. To obtain simultaneously a breakdown point of ∗n = nr/n and a consistent estimate of scale, i.e., Vn → Λ w.p.1, for a given Ep (µ, Λ) distribution the constant c0 is chosen so that E ρ(t)/a0 = r and then b0 is set to E ρ(t). A geometrical interpretation of S estimate can be given with the ellipsoidal contours of an Ep (µ, Λ). First, the volume of a p-dimensional ellipsoid z Λ−1 z ≤ 1 is |Λ|1/2 2π p/2 /[pΓ( 12 p)]; thus, minimizing |Λ| corresponds to ﬁnding a minimum volume ellipsoid [Rousseeuw (1985)]. Second, if we could allow discontinuous ρ, then ρ(t) = 1 − I[−c0 ,c0 ] (t) would count the points outside the ellipsoid. So, for r = 25%, the optimization would

13.4. Robust estimates

225

ﬁnd the minimum volume ellipsoid containing 75% of the data. An S estimate is thus a smoothed √ version of a minimum volume ellipsoid. The smoothing is done to get n-asymptotically normal estimates. Assuming further S3: ρ has a second derivative ψ , both ψ (t) and u(t) = ψ(t)/t are bounded and continuous, the asymptotic normality of the S estimate holds. Proposition 13.11 Assume S1 through S3. Let V = Λ and assume E ψ (t) > 0, E ψ (t)t2 + (p + 1)ψ(t)t > 0. Let α β

= p−1 E ψ 2 (t), = E 1 − p−1 u(t) + p−1 ψ (t) , p(p + 2)E[ψ 2 (t)t2 ] , E 2 [ψ (t)t2 + (p + 1)ψ(t)t]

σ1

=

σ2

= −2p−1 σ1 + 4

2

E [ρ(t) − b0 ] , E 2 [ψ(t)t]

then n1/2 (Vn − V, µn − µ) → (N, n), where n

|=

d

N and

n ∼ Np 0, (α/β 2 )V N ∼ Npp (0, σ1 (I + Kp )(V ⊗ V) + σ2 vec(V)[vec(V)] ) .

r = .5 r = .3 r = .1

p=1

p=2

p = 10

26.9% 40.5% 49.1%

37.7% 77.0% 98.9%

91.5% 98.0% 99.9%

Table 13.1. Asymptotic eﬃciency of S estimate of scatter at the normal distribution.

According to Lopuha¨ a (1989) the asymptotic eﬃciency for the estimation of the scatter as measured by the index σ1 (or 2σ1 + σ2 for p = 1) are as in Table 13.1 at the normal distribution. The asymptotic eﬃciency of the location estimate are even higher. For the S estimate, a high breakdown point corresponds to a low eﬃciency and vice versa. Let us mention that S estimates are able to achieve the asymptotic variance of M estimates. However, S estimates can have a

226

13. Robustness

high breakdown point in any dimension, whereas the asymptotic breakdown point of an M estimate is at most 1/(p + 1) [Tyler (1986)]. Lopuha¨ a (1991) deﬁnes τ estimates which can have the same high breakdown point as S estimates√but can attain simultaneously high eﬃciency. The τ estimates are also n-asymptotically normal. An S-plus [Statistical Sciences, (1995)] function, s.estimate, to evaluate S estimate is described in Appendix C. The implementation follows the recommendations of Ruppert (1992) to increase the speed of numerical convergence of this numerically intensive problem. The S-plus function asymp evaluates the asymptotic variance constants λ = α/β 2 , σ1 , and σ2 , at the normal distribution.

Robust Hotelling-T 2

13.4.3

Assume x1 , . . . , xn are i.i.d. x ∼ Ep (µ, Λ). Consider a test of hypothesis on the mean, H0 : µ = µ0 against H1 : µ = µ0 , using a robust version of the classical Hotelling-T 2 . Assume (Vn , µn ) is a robust aﬃne equivariant and asymptotically normal estimate (M or S estimate for example), n1/2 (Vn − V, µn − µ) → (N, n), where n

|=

d

N and

n ∼ Np 0, (α/β 2 )V , N ∼ Npp (0, σ1 (I + Kp )(V ⊗ V) + σ2 vec(V)[vec(V)] ) .

Proposition 13.12 Under the sequence of contiguous alternatives H1,n : µ = µ0 + n−1/2 γ, TR2 = n(µn − µ0 ) Vn−1 (µn − µ0 ), where (Vn , µn ) is asymptotically normal as above, satisﬁes 2 β −1 d α γV γ . TR2 → 2 · χ2p β 2α d

In particular, TR2 →

α 2 β 2 χp

under H0 .

Proof. Let X and Y be the sample matrices under H0 and H1,n , respectively. Then, we can write Y = X + n−1/2 1γ . d

Aﬃne equivariance of the estimate immediately gives µn (Y) = µn (X) + n−1/2 γ, d d

Vn (Y) = Vn (X) → V w.p.1. Since

d d n1/2 (µn (Y) − µ0 ) = n1/2 (µn (X) − µ0 ) + γ → Np γ, (α/β 2 )V ,

13.5. Robust tests on scale matrices

227

it follows from Corollary 5.1 on quadratic forms (with A = (β 2 /α)V−1 ) that 2 β −1 d α γV γ . TR2 = n(µn (Y) − µ0 ) Vn−1 (Y)(µn (Y) − µ0 ) → 2 χ2p β 2α 2 Another type of robustness found in the literature assumes an elliptical distribution on the whole data matrix X ∈ Rnp with mean 1µ and variance of the form I ⊗ Σ. Under weak assumptions on the p.d.f., the classical Hotelling-T 2 (8.1) remains UMPI and the null distribution of T 2 is the same as if xi had been i.i.d. Np (µ, Σ), i.e., T 2 ∼ Fc (p, n − p) [v. Corollary 8.1]. The main diﬀerence in the two approaches resides in that the observations under an elliptical distribution on X cannot be independent, although they are uncorrelated, unless the elliptical distribution is normal. Independence and spherical symmetry do not go together, except in the normal case, by virtue of the Maxwell-Hershell theorem [v. Proposition 4.11]. One may consult the book by Kariya and Sinha (1989) on this type of robustness for many statistical tests. Having found the asymptotic null distribution of Hotelling-T 2 , it is now a simple matter to extend the results of Section 8.3 to construct robust simultaneous conﬁdence intervals on means. For example, asymptotically, we are at least (1 − γ) × 100% conﬁdent in simultaneously presenting all of the observed “Scheﬀ´e” intervals: 1/2 1/2 α χ2γ,p α χ2γ,p 1/2 (a Vn a) ≤ a µ ≤ a µn + (a Vn a)1/2 , a µn − β2 n β2 n ∀a ∈ Rp . Realistically, the parametric family Ep (µ, Λ) is unknown. Thus, α and β will have to be replaced by consistent estimates.

13.5 Robust tests on scale matrices Assume x1 , . . . , xn are i.i.d. x ∼ Ep (µ, Λ). Consider a test of hypothesis on Λ which is of the general form h(Λ) = 0, where h(Λ) ∈ Rq is a continuously diﬀerentiable function. We will assume µ = 0. Under a Np (0, Λ) distribution, recall that a likelihoodratio test on Λ is based n uniquely on the likelihood statistic Sn = n1 i=1 xi xi . We know that nSn ∼ Wp (n, Λ). Thus, Sn has density (up to a multiplicative constant) |Λ|−n/2 etr(− 12 nΛ−1 Sn ). So, we deﬁne f (A, Λ) = |Λ|−n/2 etr(− 12 Λ−1 A), fh (A) = sup f (A, Λ), h(Λ)=0

Lh (A)

=

fh (A) . f (A, A)

228

13. Robustness

Note that Lh (Sn ) is the likelihood ratio test for H0 : h(Λ) = 0 when x ∼ Np (0, Λ). The idea to build a robust test when x ∼ Ep (0, Λ) is ˆ n ) where Λ ˆ n could be Sn [Muirhead and to use the test statistic Lh (Λ Waternaux, (1980)] or, preferably, a more robust estimate [Tyler (1983a)]. Other approaches which will not be considered here include those based on minimum discrepancy test statistics [Browne and Shapiro (1987), Shapiro and Browne (1987)].

13.5.1

Adjusted likelihood ratio tests

A general method of making a simple correction to the likelihood ratio test is possible for hypotheses satisfying the following condition H on the function h. Condition H. h(Γ) = h(γΓ), ∀γ > 0, ∀Γ > 0. Examples of hypothesis satisfying condition H are the test of sphericity and the test of covariance. Example 13.11 The test of sphericity H0 : Λ = γI for some unknown γ can be written as H0 : h(Λ) = 0 with hij (Λ) = Λij /Λpp , 1 ≤ i < j ≤ p, and hii (Λ) = Λii /Λpp − 1, i = 1, . . . , p − 1. Here, we have q = 12 (p − 1)p + p − 1. Example 13.12 The test of covariance between two subvectors H0 : Λ12 = 0, where Λ12 ∈ Rpp12 can be written as H0 : h(Λ) = 0 by choosing h(Λ) = −1/2

vec(Λ11

−1/2

Λ12 Λ22

). Obviously, q = p1 p2 .

Condition H is not dependent on the location or the spread of the elliptical contours, but concerns only the direction and relative lengths of the ˆ n is also necessary. axes of the contours. A condition E on the estimate Λ However, as we encountered in M and S estimation, we usually have an estimate Vn of a multiple V of Λ. Note that hypothesis H0 : h(Λ) = 0 is equivalent to H0 : h(V) = 0 under condition H. d

Condition E. Vn is aﬃne equivariant and n1/2 (Vn − V) → Z, where Z ∼ Npp (0, σ1 (I + Kp )(V ⊗ V) + σ2 vec(V)[vec(V)] ).

Normal and elliptical MLE, the M estimate, and the S estimate satisfy condition E under regularity conditions. An estimate of h(V) is h(Vn ), whose asymptotic distribution follows from the delta method (v. Proposition 6.2). A diﬃculty is the redundance of variables due to the symmetry of V. For this reason, the following derivative will be very useful. Deﬁne da/db = (dai /dbj ), where i varies over rows and j runs over columns. The derivative of h(V) with respect to V is deﬁned as h (V) = 12 [d h(V)/d vec(V)](I + Jp ) ∈ Rqp2 ,

13.5. Robust tests on scale matrices

229

p where Jp = i=1 ei ei ⊗ ei ei and ei ∈ Rp is a vector of zero but a 1 in position i. An example when q = 1 and p = 2 is enlightening: .. . 0 0 2 0 .. 0 . 0 0 1 1 2 (dh/ds11 , dh/ds21 , dh/ds12 , dh/ds22 ) · · · · · · · · · · · · · · · .. 0 . 1 0 0 .. . 0 2 0 0

= dh/ds11 , 12 dh/ds21 , 12 dh/ds12 , dh/ds22 is the usual gradiant of h taking into account the symmetry. Before stating the result, we need a lemma on gradiants. Lemma 13.2 Let f : Rp → R be continuously diﬀerentiable. Then, df (x)/dx, x = 0 for all x in a neighborhood of x0 iﬀ f (x) = f (αx) for all x and αx in a neighborhood of x0 . Proof. It suﬃces to notice that the contours of f are rays coming out of the origin and that the gradiant is a perpendicular vector to the contour.2 d

Proposition 13.13 Under conditions H and E, n1/2 [h(Vn ) − h(V)] → Zh , where Zh ∼ Nq (0, 2σ1 [h (V)](V ⊗ V)[h (V)] ) . Proof. As in Proposition 6.2, we can write n1/2 [h(Vn ) − h(V)] = h (V) n1/2 vec(Vn − V) + op (1). Therefore, n1/2 [h(Vn ) − h(V)] → h (V) vec(Z). d

From Lemma 13.2 and condition H, we have h (V) vec(V) = 0. Thus, the asymptotic variance is var h (V) vec(Z)

= σ1 [h (V)](I + Kp )(V ⊗ V)[h (V)] , = σ1 [h (V)](V ⊗ V)(I + Kp )[h (V)] .

Applying the identity Kp vec(A) = vec(A ) to the columns of [h (V)] gives (I + Kp )[h (V)] = 2[h (V)] and the conclusion follows. 2 An important consequence of condition H is the asymptotic variance which becomes independent of σ2 . This means that for Vn satisfying condition E, all the asymptotic distributions of h(Vn ), under condition H, such as a simple correlation coeﬃcient, a multiple correlation coeﬃcient, a ratio of eigenvalues, etc., are the same as those for the sample variance S, when

230

13. Robustness

sampling from a multivariate normal distribution, except for the factor σ1 in the asymptotic variance. For completeness, the results for correlations are now given. Denote by rij the simple correlation deﬁned from the scale estimate Vn = (vn,ij ), i.e., vn,ij rij = 1/2 1/2 , vn,ii vn,jj and let ρij =

Λij 1/2 1/2 Λii Λjj

be the correlation for the Ep (µ, Λ) distribution. Proposition 13.14 Assume condition E holds on Vn . Then, 1/2

n1/2 (rij − ρij ) → σ1 d

· N (0, (1 − ρ2ij )2 ).

From the delta method it is also clear that an arbitrary number of correlation coeﬃcients is jointly asymptotically normal. Thus, it suﬃces to consider the case of two correlation coeﬃcients rij and rkl . Proposition 13.15 Assume condition E holds on Vn . Then, (1 − ρ2ij )2 ω rij − ρij d 1/2 n1/2 → σ1 · N2 0, , rkl − ρkl ω (1 − ρ2kl )2 where the asymptotic covariance ω is given by ω = ρij ρkl + ρkj ρil − ρlj (ρij ρkj + ρil ρkl ) − ρki (ρij ρil + ρkj ρkl ) + 12 ρki ρlj (ρ2ij + ρ2il + ρ2kj + ρ2kl ). Proof. Assume V = (ρij ) without loss of generality. Write down the asymptotic distribution vn,ij ρij vn,ii 1 1 d 1/2 vn,jj n − → N6 (0, Ω) vn,kl ρkl vn,kk 1 vn,ll 1 for a certain Ω and apply the delta method.

2

ˆ ≡ R(V ˆ n ) and parSimilarly, for the multiple correlation coeﬃcient R tial correlation coeﬃcient rij|x2 ≡ rij|x2 (Vn ), obtained from Vn satisfying condition E, we can write the asymptotic distributions:

d 1/2 ˆ 2 − R2 ) → n1/2 (R σ1 · N 0, 4R2 (1 − R2 )2 ,

d 1/2 n1/2 rij|x2 − ρij|x2 → σ1 · N 0, (1 − ρ2ij|x2 )2 .

13.5. Robust tests on scale matrices

231

Higher-order asymptotic distributions for functions of the sample variance S can also be derived with the use of zonal polynomials [Iwashita and Siotani (1994)]. For the same reason, adjustment to the likelihood ratio test will take a rather simple form. The asymptotic distribution of the modiﬁed likelihood ratio test Lh (Vn ), where Vn may be a robust estimate, is obtained with the equivalent form of Wald’s test for the same hypothesis. Let un ∼ vn mean p un − vn → 0. The following result on Wald’s formulation holds regardless of condition H. Proposition 13.16 Let An > 0, n = 1, 2, . . ., be such that n1/2 (An − d

A) → (·) for a ﬁxed A > 0 satisfying h(A) = 0. If rank h (Γ) = q, ∀Γ in a neighborhood of A, then −2 ln Lh (An ) ∼ n[h(An )] [Ch (An )]−1 h(An ), where Ch (Γ) = 2[h (Γ)](Γ ⊗ Γ)[h (Γ)] . Proof. This is a generalization of Wald’s formulation for the asymptotic behavior of the likelihood ratio statistic. Refer to Tyler (1983a) for details. 2 Corollary 13.2 Assume conditions H and E. Then: d

(i) under H0 , −2 ln Lh (Vn ) → σ1 χ2q , (ii) under the sequence of contiguous alternatives Λn = Λ+n−1/2 B, where h(Λ) = 0 and B is a ﬁxed symmetric matrix, −2 ln Lh (Vn ) → σ1 χ2q (δh (Λ, B)/2σ1 ) , d

where δh (Λ, B) = [vec(B)] [h (Λ)] [Ch (Λ)]−1 h (Λ)vec(B). d

Proof. From conditions H and E, n1/2 (Vn − V) → Z and n1/2 [h(Vn ) − d

h(V)] → Zh , where Zh ∼ Nq (0, σ1 Ch (V)). Under H0 : h(V) = 0, p

d

n1/2 h(Vn ) → Zh , and since h (·) is continuous, Ch (Vn ) → Ch (V). Hence, we have [n1/2 h(Vn )] [Ch (Vn )]−1 [n1/2 h(Vn )] → Zh [Ch (V)]−1 Zh = σ1 χ2q . d

d

For contiguous alternatives, under condition H, the noncentrality parameter is invariant with respect to scalar multiplication δh (Λ, B) = 2 δh (αΛ, αB), ∀α > 0. As a particular case for Sn which has σ1 = 1 + k, we have, under H0 , d ˆ → −2 ln Lh (Sn )/(1 + k) χ2q

232

13. Robustness

for some consistent estimate kˆ of the kurtosis parameter. So, in the class of Ep (0, Λ) with ﬁnite fourth-order moments, this adjusted LRT is robust in the sense that the asymtotic distribution is the same as if x ∼ Np (0, Λ). Note that a consistent estimate kˆ can be obtained by the method of moment with the identity 1 + k = pE(s2 )/[(p + 2)E 2 (s)],

(13.15)

2

where s = |z| and z ∼ Ep (0, I) has fourth-order moments (v. Problem 13.6.12). More generally, the test statistic −2 ln Lh (Vn )/ˆ σ1 will be referred to as an adjusted LRT. The test of sphericity can serve as an example to illustrate Proposition 13.16 and Corollary 13.2. Wald’s formulation is generally obtained by a Taylor series of −2 ln Lh (Vn ) around V, satisfying H0 : h(V) = 0. For the test of sphericity, we have −2 ln Lh (Vn ) = −n ln |Vn | + pn ln(p−1 tr Vn ). = γI + n−1/2 Zn , Under H0 : V = γI and condition E, we can write Vn ∞ where Zn is bounded in probability. Since ln(1 + x) = i=1 (−1)i+1 xi /i, −1 < x < 1, it follows that for a ﬁxed symmetric A, ln |I + tA| =

∞

(−1)i+1 tr(Ai )ti /i

i=1

for all t suﬃciently small. Hence, we get the expansion −2 ln Lh (Vn )

= d

→

1 −2 [tr(Z2n ) 2γ

− p−1 (tr Zn )2 ] + Op (n−1/2 )

1 −2 [tr(Z2 ) 2γ

− p−1 (tr Z)2 ],

where Z ∼ γNpp (0, σ1 (I + Kp ) + σ2 vec(I)[vec(I)] ). From the relations (v. Problem 6.4.2) tr Z2 = [vec(Z)] 12 (I + Kp ) vec(Z) and tr Z = [vec(I)] vec(Z), it follows that 1 −2 [tr(Z2 ) 2γ

− p−1 (tr Z)2 ] = [vec(Z)] A vec(Z),

where A = 12 γ −2 { 12 (I + Kp ) − p−1 vec(I)[vec(I)] } is a quadratic form. This is Wald’s equivalent formulation for this test. The asymptotic result −2 ln Lh (Vn )/σ1 → χ2q , q = 12 (p − 1)(p + 2) d

follows from Corollary 5.1 on quadratic forms. When condition H is not satisﬁed, simple adjustments to the LRT is generally not possible, as the following corollary shows. Corollary 13.3 Under H0 : h(V) = 0 and condition E, −2 ln Lh (Vn ) → σ1 χ2q−1 + [σ1 + σ2 δh (V, V)]χ21 , d

|=

χ21 . The term δh (V, V) = 0 iﬀ for some neighborhood of V, with χ2q−1 h(Γ) = h(γΓ) for all Γ and γΓ in this neighborhood.

13.5. Robust tests on scale matrices

233

Proof. Take a closer look at the distribution of Zh in Proposition 13.13 d when condition H is not satisﬁed. Under H0 , we still have n1/2 h(Vn ) → h (V) vec(Z) but with an added term in the variance: var h (V) vec(Z)

= σ1 [h (V)](V ⊗ V)(I + Kp )[h (V)] +σ2 h (V) vec(V)[vec(V)] [h (V)] ≡ Dh (V).

Using Proposition 13.16, the equivalent Wald’s formulation is −2 ln Lh (Vn ) ∼ [n1/2 h(Vn )] [Ch (Vn )]−1 [n1/2 h(Vn )], and, thus, −2 ln Lh (Vn ) → Zh [Ch (V)]−1 Zh , d

where Zh ∼ Nq (0, Dh (V)). The result follows since [Ch (V)]−1 Dh (V) has eigenvalues σ1 of multiplicity (q − 1) and σ1 + σ2 δh (V, V). The second statement follows since h (V) vec(V) = 0 iﬀ the stated condition holds. 2

13.5.2

Weighted Nagao’s test for a given variance

In this section, we consider an example where the condition H on the hypothesis is not satisﬁed, but a simple test, robust to large kurtosis, can still be built. For testing the hypothesis, H0 : Σ = I, against H1 : Σ = I, the modiﬁed likelihood ratio test based on n i.i.d. vectors from a Np (µ, Σ) distribution is (v. Problem 8.9.8) Λ∗ = epm/2 |Sn |m/2 etr(− 12 mSn ), m = n − 1, n ¯ )(xi − x ¯ ) /m. It is invariant to orthogonal where Sn = i=1 (xi − x transformations, unbiased, and −2 ln Λ∗ is asymptotically distributed as a noncentral chi-square [Khatri and Srivastava (1974)], χ2f (δ), with f = p 1 2 i=1 di /4, under the sequence of local alternatives 2 p(p + 1) and δ = Σn = I + n−1/2 D, D = diag(d1 , . . . , dp ).

(13.16)

However, Muirhead (1982, p. 365) showed that if the sample came from an elliptical distribution, Ep (µ, Σ), with kurtosis 3k, then the asymptotic null distribution is kp d ∗ ˆ χ2 + χ2f −1 , −2 ln Λ /(1 + k) → 1 + 2(1 + k) 1 where χ21 and χ2f −1 are independently distributed and kˆ is a consistent estimate of k. A generalization to robust estimates of scale is proposed in Problem 13.6.16. Therefore, even the adjusted test statistic −2 ln Λ∗ /(1 + ˆ is not robust to non-normality of the data, especially for large values k) of k or long-tailed distribution. Moreover, the procedure of estimating k

234

13. Robustness

in the asymptotic distribution and calculating the critical points of the convolution of ˆ kp χ21 and χ2f −1 , 1+ ˆ 2(1 + k) as if kˆ were a constant, is obviously not a valid procedure. A new test statistic W is proposed, which is also invariant to orthogonal transformations and has an asymptotic null distribution χ2f for all underlying elliptical distributions with ﬁnite fourth moments. The asymptotic non-null distribution under the sequence of local alternatives (13.16) is noncentral chi-square. It is asymptotically fully eﬃcient at the normal distribution as compared to the modiﬁed likelihood ratio test. Let Sn = I + n−1/2 Un , Sn = (sij ), Un = (uij ). Then, when H0 is true, the asymptotic distribution of √ √ un = (u11 / 2, . . . , upp / 2, u12 , . . . , u1p , u23 , . . . , u2p , . . . , up−1,p ) ∈ Rf when the observations xi are drawn from an elliptical distribution with kurtosis 3k is Nf (0, Γ), where Ω 0 Γ= 0 (1 + k)If −p with Ω = (1+k)Ip + 12 k11 , 1 = (1, . . . , 1) ∈ Rp . Then, from Corollary 5.1, we have under H0 , u11 d . u2ij /(1 + k) → χ2f . un Γ−1 un = 12 (u11 , . . . , upp )Ω−1 .. + i<j upp The test statistic proposed [Bentler (1983)] is s11 − 1 n .. ˆ ˆ −1 W = (s11 − 1, . . . , spp − 1)Ω s2ij /(1 + k), + n . 2 i<j spp − 1 ˆ p + 1 k11 ˆ and kˆ is a consistent estimate of k. ˆ = (1 + k)I where Ω 2 Note that when kˆ ≡ 0, then W reduces to Nagao’s (1973) test statistic (n/2) tr(Sn − I)2 . Asymptotic expansions of Nagao’s test for elliptical distributions were derived by Purkayastha and Srivastava (1995). The test statistic W can be seen as a weighted form of Nagao’s statistic with the diagonal and oﬀ-diagonal elements of the sample variance matrix, Sn , being assigned diﬀerent weights.

13.5. Robust tests on scale matrices

235

The identity (13.15) leads by the method of moments to the consistent and orthogonally invariant estimate n ¯ |4 |xi − x (13.17) kˆ = pn ni=1 2 − 1. ¯ |2 ) ( i=1 |xi − x ˆ , it has a ˆ perturbed by a rank 1 matrix, namely 1 k11 ˆ is (1 + k)I Since Ω 2 known inverse which leads to the equivalent expression W =

kˆ n ˆ −n tr(Sn − I)2 /(1 + k) (tr Sn − p)2 , ˆ ˆ + kp] ˆ 2 2 (1 + k)[2(1 + k)

showing that W is invariant to orthogonal transformations, xi → Hxi for any orthogonal matrix H. Thus, without loss of generality, we can take for W the sequence of local alternatives (13.16) with a diagonal matrix D. The following result was given in Bilodeau (1997b). Proposition 13.17 Under the sequence of local alternatives Σn = I + n−1/2 D, D = diag(d1 , . . . , dp ), the asymptotic distribution of W is noncentral chi-square, W → χ2f (d Ω−1 d/4), d

where f

=

d = Ω =

1 2 p(p

+ 1), (d1 , . . . , dp ) , (1 + k)I + 12 k11 .

1/2

Proof. Let xi = Σn zi , where zi ∼ Ep (0, I). Also, let x1 z1 .. .. X = . and Z = . xn

zn

be the sample matrices, and Sn (X) and Sn (Z) be the sample variance matrices obtained from X and Z, respectively. Then, we have 1/2 1/2 Un (X) ≡ n1/2 [Sn (X) − I] = Σ1/2 (Σn − I), n Un (Z)Σn + n d

where Un (Z) → Npp (0, (1 + k)(I + K) + k vec(I)vec(I) ), Σn → I, and n1/2 (Σn − I) = D. Hence, the asymptotic result Un (X) → Npp (D, (1 + k)(I + K) + k vec(I)vec(I) ) d

ˆ is obtained. Since W is a continuous function of Un (X) and k, ˆ W = g(Un (X), k),

236

13. Robustness

the conclusion follows from Lemma 6.3 and classical results on quadratic forms if kˆ in (13.17) is consistent under the same sequence of local alterp 1/2 ¯ = Σn z ¯ and since z ¯ → 0, Σn → I, we natives. This is now shown. From x p ¯ → 0. Thus, the asymptotic equivalences have x 1 1 1 1 ¯ |2 ∼ ¯ |4 ∼ |xi − x |xi |2 , |xi − x |xi |4 , n i=1 n i=1 n i=1 n i=1 n

n

n

n

p

where u ∼ v means u − v → 0, hold. But now, since 1 1 1 |zi |j ≤ |xi |j ≤ (1 + n−1/2 d(p) )j/2 |zi |j , n i=1 n i=1 n i=1 n

(1 + n−1/2 d(1) )j/2

n

n

where d(1) = min{di } and d(p) = max{di }, we also have the equivalences 1 1 1 1 |xi |2 ∼ |zi |2 , |xi |4 ∼ |zi |4 . n i=1 n i=1 n i=1 n i=1 n

n

n

n

p Thus, 1 + kˆ → pE|zi |4 /E 2 |zi |2 = 1 + k, which completes the proof.

2

When k = 0, the test statistic W is asymptotically distributed, under the sequence of local alternatives (13.16), as χ2f (d d/4). Therefore, W is asymptotically fully eﬃcient at the normal distribution as compared to the modiﬁed likelihood ratio test, −2 ln Λ∗ . Sutradhar (1993) discusses the score test of the multivariate t. For testing the hypothesis H0 : µ = 0 and Σ = I against H1 : µ = 0 ¯ under the sequence of local or Σ = I, consider the test statistic W + n¯ x x alternatives µn = n−1/2 τ , Σn = I + n−1/2 D, where D = diag(d1 , . . . , dp ). Then, it can be established along the same lines ¯ → χ2f (δ), W + n¯ x x d

where f = 12 p(p + 3), δ = d Ω−1 d/4 + τ τ /2, and d and Ω are as in ¯ is thus robust in the class of elliptical Proposition 13.17. The test W +n¯ x x distributions with ﬁnite fourth moments. Its full eﬃciency at the normal distribution as compared to the likelihood ratio test follows immediately by comparing the asymptotic non-null distributions of the two tests [Khatri and Srivastava (1974)].

13.5.3

Relative eﬃciency of adjusted LRT

Under condition H, the adjusted LRT based on Sn has noncentrality parameter δh (Λ, B)/2(1 + k)

13.5. Robust tests on scale matrices

237

and for k moderately large, it is expected to have low power. A measure of eﬃciency can be derived by comparing the adjusted LRT with the exact LRT derived under a particular Ep (0, Λ) with known density deﬁned by the function g(·). The likelihood for Λ built from the sample matrix x1 .. X= . xn

is Lg (Λ) = |Λ|−n/2

n

g(xi Λ−1 xi ).

i=1

Then, ˆn Λ

=

arg min Lg (Λ),

ˆ ˆn Λ

=

arg min Lg (Λ)

Λ>0

h(Λ)=0

are respectively the restricted and unrestricted elliptical MLE of Λ. Then, the optimal procedure is the LRT derived for a given g(·): Lh,g (X) =

ˆ ˆ n) Lg (Λ . ˆ n) Lg (Λ

Then, Wald’s classical formulation for this “elliptical” LRT is ˆ n )] [Ch (Λ ˆ n )]−1 h(Λ ˆ n )/σ1,g −2 ln Lh,g (X) ∼ n[h(Λ under H0 or under the sequence of alternatives Λn = Λ + n−1/2 B. The parameter σ1,g is the value of σ1 in the asymptotic variance of the MLE given in Proposition 13.7, i.e., σ1,g = p(p + 2)/E[ψ 2 (s)]. Corollary 13.4 Assume condition H holds. Then: d

(i) under H0 , −2 ln Lh,g (X) → χ2q , (ii) under the sequence of contiguous alternatives Λn = Λ+n−1/2 B, where h(Λ) = 0 and B is a ﬁxed symmetric matrix, −2 ln Lh,g (X) → χ2q (δh (Λ, B)/2σ1,g ) , d

where, as before, δh (Λ, B) = [vec(B)] [h (Λ)] [Ch (Λ)]−1 h (Λ) vec(B). When g(·) is known, another test which is ﬁrst-order eﬃcient and asymptotically distributed as chi-square is the minimum geodesic distance test [Berkane et al. (1997)]. The proof of Corollary 13.4 is identical to that of Corollary 13.2. The ˆ to the asymptotic eﬃciency of the adjusted LRT −2 ln Lh (Sn )/(1 + k)

238

13. Robustness

q=1 q=2 q=3

ν =5

ν=6

ν=7

ν=8

ν = 30

.26 .37 .46

.17 .22 .27

.13 .17 .20

.11 .14 .16

.06 .06 .06

Table 13.2. Asymptotic signiﬁcance level of unadjusted LRT for α = 5%.

elliptical LRT can thus be measured by the ratio of the noncentrality parameters, i.e., σ1,g /(1 + k). For the tp,ν density, it was evaluated that σ1,g = 1 + 2/(p + ν), whereas 1 + k = (ν − 2)/(ν − 4) for a relative eﬃciency of (ν − 4)(ν + 2)/[(ν − 2)(ν + p)]. For p = 2 and ν = 5, this gives an eﬃciency of 33%. This is due to the poor robustness property of Sn . This adjusted LRT cannot really be thought of as a robust test because of its low eﬃciency. To obtain a truly robust adjusted LRT, one has to replace Sn by an eﬃcient robust estimate, i.e., one with a σ1 close to σ1,g . We conclude this analysis by guarding the practitioner against assuming indiscriminantly the normality of the data and using the “optimal” test for normality. If the data came from an elliptical distribution with a kurtosis parameter k and the hypothesis (satisfying condition H) was H0 : h(Λ) = 0, where h(Λ) ∈ Rq , then what was supposed to be an α = 5% signiﬁcance level test would be, in fact, for large samples, a test of signiﬁcance level: P (−2 ln Lh (Sn ) ≥ χ2.95,q )

P (−2 ln Lh (Sn )/(1 + k) ≥ χ2.95,q /(1 + k))

→ P χ2q ≥ χ2.95,q /(1 + k) . =

For a tp,ν distribution with 1 + k = (ν − 2)/(ν − 4), the signiﬁcance level may be far from 5%, as evidenced by Table 13.2. The situation worsens as ν decreases, which means the tails become heavier or q increases, which is related to the complexity of the hypothesis. For q = 3 and ν = 5, tossing a coin is nearly as reliable!

13.6 Problems 1. Demonstrate the following on normal mixture representation: d

|=

z, then (i) If x = w1/2 z, where w ∼ F , z ∼ Np (0, I), and w ∞ (2πw)−p/2 exp(− 12 w−1 x x)dF (w), fx (x) = 0

where F (·) is a distribution function on [0, ∞). (ii) If νw−1 ∼ χ2ν , then x = w1/2 z ∼ tp,ν has density

where cp,ν

fx (x) = cp,ν (1 + x x/ν)−(ν+p)/2 , x ∈ Rp ,

= (νπ)−p/2 Γ 12 (ν + p) /Γ 12 ν .

13.6. Problems

239

2. Assume x ∼ tp,ν (µ, Λ) where x is partitioned as x = (x1 , x2 ) , xi ∈ Rpi , i = 1, 2, p = p1 + p2 . Demonstrate the following: (i) E x = µ, var x = [ν/(ν − 2)]Λ, ν > 2. (ii) p−1 (x − µ) Λ−1 (x − µ) ∼ F (p, ν). (iii) The marginal distribution is x2 ∼ tp2 ,ν (µ2 , Λ22 ), where (µ , µ ) , 1 2 Λ11 Λ12 Λ = Λ21 Λ22 µ =

are partitioned in conformity. (iv) The conditional distribution is

x1 |x2 ∼ tp1 ,ν+p2 µ1 + Λ12 Λ−1 22 (x2 − µ2 ), h(x2 )Λ11.2 , where h(x2 ) = [ν/(ν + p2 )] · [1 + (x2 − µ2 ) Λ−1 22 (x2 − µ2 )/ν]. Determine E(x1 |x2 ), var(x1 |x2 ) and the condition for their existence. 3. Verify by diﬀerentiation of ln φ(t21 + t22 ) the cumulants = −2φ (0),

k2 k4

= =

k22

12(φ (0) − φ (0)2 ), 4(φ (0) − φ (0)2 ),

where φ(t21 + t22 ) is the characteristic function of a bivariate rotationally invariant vector. 4. Obtain E uu = p−1 I and var(uu ) =

1 2 (I + Kp ) − 2 vec(I)[vec(I)] , p(p + 2) p (p + 2)

where u ∼ unif(S p−1 ). 5. For the multivariate power exponential family (13.2), prove the following: (i) The normalizing constant is cp,α =

αΓ(p/2) π p/2 2p/2α Γ(p/2α)

.

(ii) The variance of x is var x =

21/α Γ[(p/2 + 1)/α] Λ. pΓ(p/2α)

(iii) For α = 1/2, verify the assertion in Example 13.4. Hint: Use the representation in polar coordinates in Proposition 4.10 together with Problems 4.6.13 and 13.6.4.

240

13. Robustness

ˆ = (Λ ˆ ij ) is aﬃne equivariant with variance 6. Check that when Λ ˆ = σ1 (I + Kp )(Λ ⊗ Λ) + σ2 vec(Λ)[vec(Λ)] , var Λ then ˆ ki , Λ ˆ lj ) = σ1 (Λij Λkl + Λkj Λil ) + σ2 Λki Λlj . cov(Λ 7. Prove if a0 > p + 1, n ≥ p + 1, then condition D1 is satisﬁed w.p.1 when sampling from an absolutely continuous distribution. 8. Verify conditions M1 through M4 for u1 and u2 corresponding to the MLE under the tp,ν distribution. 9. Verify conditions M1 through M4 for Huber’s ψ function in Example 13.10. 10. Assume z ∼ Ep (0, I) with density g(|z|2 ). Deﬁne u(s) = −2g (s)/g(s) and ψ(s) = su(s). Prove E ψ(|z|2 ) = p. Hint: Integrate by parts. 11. For the tp,ν distribution, verify that the asymptotic variance parameter σ1 of the elliptical MLE in Proposition 13.7 is σ1 = 1 + 2/(p + ν). 12. Prove 1 + k = pE(s2 )/[(p + 2)E 2 (s)], where s = |z|2 and z ∼ Ep (0, I) has fourth-order moments. 13. Deﬁne u(s) = −2g (s)/g(s) and ψ(s) = su(s). Let s = |z|2 , where z ∼ Ep (0, I). Derive an upper bound for σ1,g /(1 + k), the index of relative eﬃciency, by going through the following steps: (i) E ψ 2 (s) ≥ [E ψ(s)]2 , (ii) E(s2 ) E[ψ 2 (s)] ≥ [Esψ(s)]2 = (p + 2)2 E 2 (s), (iii) σ1,g /(1 + k) ≤ min{1, (1 + 2p−1 )(1 + k)−1 }. (iv) Interpret the bound in (iii). 14. Demonstrate that if ρ(t) = t2 and b0 = p in the deﬁnition of the S estimate, then the solution is the normal MLE. 15. Demonstrate that the S estimate (µn , Vn ) is necessarily a solution of the equations ave [u(ti )(xi − µn )] = 0 ave [pu(ti )(xi − µn )(xi − µn ) − v(ti )Vn ] = 0, 1/2 where ti = (xi − µn ) Vn−1 (xi − µn ) and v(t) = tψ(t) − ρ(t) + b0 . M and S estimates are close relatives! 16. Test for a given variance. This is a continuation of Problem 8.9.8. The LRT under the Np (0, Λ) for H0 : Λ = I versus H1 : Λ = I is given by Lh (Sn ) = epn/2 |Sn |n/2 etr(− 12 nSn ),

13.6. Problems

241

n where Sn = n1 i=1 xi xi . Suppose that, in fact, xi ∼ Ep (0, Λ) and ˆ n satisfying condition E instead one decides to use a robust estimate Λ of Sn . Then, demonstrate that under H0 ,

where χ21

|=

d ˆ n) → (σ1 + 12 σ2 p)χ21 + σ1 χ2p(p+1)/2−1 , −2 ln Lh (Λ

χ2p(p+1)/2−1 by following these steps:

ˆ n = I + n−1/2 Zn and use the Taylor series for ln |I + tA| (i) Write Λ around t = 0 to show that under H0 , d 1 ˆ n) → −2 ln Lh (Λ 2 [vec(Z)] vec(Z),

where Z ∼ Npp (0, σ1 (I + Kp ) + σ2 vec(I)[vec(I)] ). (ii) Demonstrate √ √ var(z11 / 2, . . . , zpp / 2, z12 , . . . , z1p , z23 , . . . , z2p , . . . , zp−1,p ) is given by

σ1 Ip + 12 σ2 11 0

0 σ1 Ip(p−1)/2

≡ Ω.

(iii) Verify the eigenvalues of Ω are σ1 of multiplicity 12 p(p + 1) − 1 and σ1 + 12 σ2 p of multiplicity 1. 17. Test of multiple correlation. The LRT under (x1 , x2 ) ∼ Np (0, Λ) for H0 : R2 = 0 versus H1 : R2 = 0 is given by

where Sn = the partition

1 n

n

ˆ 2 (Sn ))n/2 , Lh (Sn ) = (1 − R

i=1

ˆ 2 (Sn ) = s S−1 s21 /s11 in terms of xi xi and R 21 22 Sn =

s11 s21

s21 S22

.

Suppose, in fact, that (x1 , x2 ) ∼ Ep (0, Λ) and one decides to use ˆ n satisfying condition E instead of Sn . Then, a robust estimate Λ demonstrate that under H0 , d ˆ n) → σ1 χ2p−1 , −2 ln Lh (Λ

by following these steps: (i) Argue that one can assume Λ = I. (ii) Using a Taylor series, prove that ˆ n ) ∼ nR ˆ n ). ˆ 2 (Λ −2 ln Lh (Λ d ˆ n) → ˆ 2 (Λ z z, where z ∼ Np−1 (0, σ1 I) to (iii) Finally, prove that nR conclude.

242

13. Robustness

n 18. Let l = (l1 , . . . , lp ) be the eigenvalues of Sn = n1 i=1 xi xi , calculated from a sample of an Ep (0, Λ) distribution, Λ = diag(λ1 , . . . , λp ). If the population eigenvalues λα are all distinct, then prove that the joint limiting distribution is given by n1/2 Λ−1 (l − λ) → Np (0, 2(1 + k)I + k11 ). d

19. Suppose the sample is taken from an elliptical distribution with kurtosis 3k. Let f = (f1 , . . . , fp ) be the eigenvalues of the sample correlation matrix R = (rij ). If the eigenvalues γα of the population correlation matrix ρ = (ρij ) = G diag(γ1 , . . . , γp ) G , where G = (gij ) ∈ Op , are all distinct, then prove that the joint limiting distribution is n1/2 (f − γ) → Np (0, (1 + k)Ω), d

where Ω = (ωαβ ) is given by ωαβ = 2γα γβ δαβ − (γα + γβ )

p j=1

2 2 gjα gjβ +

p p j=1 i=1

2 2 2 gjα giβ ρji .

14 Bootstrap conﬁdence regions and tests

An important part of multivariate analysis deals with conﬁdence regions and tests of hypotheses on the mean vector and variance matrix. The classical theoretical developments for such procedures rest mainly upon the multivariate normality assumption. Without multivariate normality, the asymptotic distribution of many tests becomes more complex and often leads to untabulated limit distributions. The bootstrap conﬁdence regions and tests on the mean vector and variance matrix have the desired asymptotic levels under very mild conditions. We will present the bootstrap technique main ideas without formal proofs. The interested reader should consult the cited references. General references for the bootstrap are Efron (1982), who made the technique widely applicable by using modern computational power, Efron and Tibshirani (1993) and Hall (1992). The book by Davison and Hinkley (1997) has S-plus code which may prove useful.

14.1 Conﬁdence regions and tests for the mean Let x = (x1 , . . . , xp ) ∼ F with mean µF = (µF,i ) and variance ΣF = (σF,ij ). Let x1 , . . . , xn be i.i.d. F and ¯n x

=

1 xi , n i=1

Sn

=

1 ¯ n )(xi − x ¯ n ) , (xi − x (n − 1) i=1

n

n

244

14. Bootstrap conﬁdence regions and tests

be the sample mean and sample variance, respectively. Deﬁne the “pivot” xn − µF )|. wn = n1/2 |Sn−1/2 (¯ Then, by the central limit theorem, d

wn → |z|, z ∼ Np (0, I). The empirical distribution function of the data x1 , . . . , xn is denoted by 1 I(xi ≤ t), Fˆn (t) = n i=1 n

where I(xi ≤ t) is the indicator function. In other words, Fˆn is the discrete distribution function with equal probability 1/n at the points xi , i = 1, . . . , n. Then, for x∗ ∼ Fˆn , we have E x∗

= µFˆn =

1 xi , n i=1

var x∗

= ΣFˆn =

1 ¯ n )(xi − x ¯ n ) . (xi − x n i=1

n

n

The nonparametric bootstrap estimate of the probability distribution of wn under F , Jn (F ), is the bootstrap estimate Jn (Fˆn ), which can be interpreted ¯ ∗n and S∗n be the sample as follows. Let x∗i , i = 1, . . . , n, be i.i.d. Fˆn , and let x ∗ mean and sample variance, respectively, of the xi ’s. Then, Jn (Fˆn ) is the probability law under Fˆn of x∗n − µFˆn )|. wn∗ = n1/2 |S∗n −1/2 (¯ In practice, Jn (Fˆn ) may be approximated to any degree of accuracy with resampling by Monte Carlo methods. The consistency of the bootstrap was established by Beran (1984, example 3) using a triangular array version of the C.L.T., wn∗ → |z| w.p.1, d

which means that wn and wn∗ converge in distribution to the same limit. However, it was Singh (1981), and Bickel and Freedman (1981) who ﬁrst established the consistency of the bootstrap in the univariate situation. Let cn (α, Fˆn ) be a (1 − α)-quantile of the bootstrap distribution Jn (Fˆn ). By the consistency of the bootstrap, if Dn (α) is the bootstrap conﬁdence region for µF , (¯ xn − µF )| ≤ cn (α, Fˆn )}, Dn (α) = {µF : n1/2 |S−1/2 n then lim P (µF ∈ Dn (α)) = 1 − α.

n→∞

14.1. Conﬁdence regions and tests for the mean

245

The bootstrap conﬁdence region Dn (α) handles all norms, | · |, on Rp with equal ease. Most often though, the euclidian norm is intended, and in that case, the ellipsoidal bootstrap conﬁdence region can be written as Dn (α) = {µF : n(¯ xn − µF ) S−1 xn − µF ) ≤ c2n (α, Fˆn )}. n (¯ A (1−α)-acceptance region, A(µ0 ), for testing the hypothesis H0 : µF = µ0 against H1 : µF = µ0 may be obtained by inverting the conﬁdence region, Dn (α), in the usual way [Fraser (1976), p. 580]. Here, the test which rejects H0 : µF = µ0 iﬀ µ0 ∈ Dn (α) is a test with asymptotic type I error probability α. More generally, suppose a conﬁdence region on g(µF ) ∈ Rk , k ≤ p, is wanted where g : Rp → Rk is a continuously diﬀerentiable function and has ﬁrst derivative g˙ ∈ Rkp . Let u : Rk → R be continuous on Rk such that {z ∈ Rk : u(z) = c} has Lebesgue measure 0 for every c ∈ R. Consider the statistic ! wn,g = u n1/2 (g(¯ xn ) − g(µF )) . The central limit theorem coupled with the delta method yields d

˙ F ) zF ], zF ∼ Np (0, ΣF ). wn,g → u[g(µ The condition imposed on u ensures that the limit distribution is continuous. Using arguments as in Beran (1984), it can be established that the bootstrap estimate is consistent, i.e., ! d ∗ ˙ F ) zF ] w.p.1. wn,g = u n1/2 g(¯ x∗n ) − g(µFˆn ) → u[g(µ To construct the bootstrap conﬁdence region, let cn,g (α, Fˆn ) be a (1 − α)quantile of the bootstrap distribution. A bootstrap conﬁdence region for g(µF ) having asymptotic coverage probability 1 − α is ! Dn,g (α) = {g(µF ) : u n1/2 (g(¯ xn ) − g(µF )) ≤ cn,g (α, Fˆn )}. In the examples to be considred, the function u has the additional property, u(bz) = bu(z), ∀z ∈ Rk , ∀b > 0. Then, the factor n1/2 may be omitted and we may write equivalently Dn,g (α) = {g(µF ) : u [g(¯ xn ) − g(µF )] ≤ c∗n,g (α)}, where c∗n,g (α) is a (1 − α)-quantile of the distribution of u[g(¯ x∗n ) − g(µFˆn )] ∗ ¯ n is the bootstrap sample mean. when Fˆn is ﬁxed at its realized value and x Judicious choices of u and g give interesting conﬁdence regions, as the following examples show.

246

14. Bootstrap conﬁdence regions and tests

p Example 14.1 Let g(µF ) = µF and u(z) = |z|1 = i=1 |zi | be the l1 norm. Then, the bootstrap conﬁdence region p ∗ Dn,g (α) = µF : |¯ xn,i − µF,i | ≤ cn,g (α) , i=1

has asymptotic coverage probability 1−α, where c∗n,g (α) is a (1−α)-quantile p of the distribution of i=1 |¯ x∗n,i −µFˆn ,i | when Fˆn is ﬁxed at its realized value ∗ ¯ n is the bootstrap sample mean. and x Example 14.2 Let g(µF ) = µF and u(z) = |z|∞ = max1≤i≤p |zi | be the l∞ -norm. The bootstrap simultaneous conﬁdence intervals Dn,g (α) = {µF : |¯ xn,i − µF,i | ≤ c∗n,g (α), i = 1, . . . , p}, have asymptotic simultaneous coverage probability 1 − α, where c∗n,g (α) is a (1 − α)-quantile of the distribution of max1≤i≤p |¯ x∗n,i − µFˆn ,i | when Fˆn is ∗ ¯ n is the bootstrap sample mean. ﬁxed at its realized value and x Example 14.3 This example provides the bootstrap algorithm, easy to implement on a computer, to construct simultaneous conﬁdence intervals on the means µF,i , i = 1, . . . , p. We are given a sample x1 , . . . , xn from F . Bootstrap algorithm • • • •

¯ n = (¯ Calculate x xn,i ). b←1 B ← 2000 (say) Do while b ≤ B. • Draw a bootstrap sample x∗1 , . . . , x∗n from Fˆn . ¯ ∗n = (¯ • Calculate x x∗n,i ). x∗n,i − x ¯n,i | • ub ← max1≤i≤p |¯ • b←b+1 • End. • Order the ub ’s: u(1) ≤ u(2) ≤ · · · ≤ u(B) . • q ← (1 − α)B! ( ·! is the integer part function) • Simultaneous conﬁdence intervals for µF,i with approximate simultaneous coverage probability 1 − α are x ¯n,i − u(q) ≤ µF,i ≤ x ¯n,i + u(q) , i = 1, . . . , p.

14.2 Conﬁdence regions for the variance This time, deﬁne Wn = n1/2 (Sn − ΣF ).

14.2. Conﬁdence regions for the variance

247

The asymptotic distribution of Sn was given in Section 6.3 for all underlying distribution F with ﬁnite fourth moments. The asymptotic distribution is d

Wn → XF , XF ∼ Npp (0, ΩF ), where XF = (xF,ij ), the elements of ΩF given by ijkl kl ij il jk + k11 k11 + k11 k11 cov(xF,ik , xF,jl ) = k1111

with the k’s representing the cumulants of F . The nonparametric bootstrap estimate of the probability distribution under F of Wn is the probability distribution under Fˆn of Wn∗ = n1/2 S∗n − ΣFˆn . Beran and Srivastava (1985) established the consistency of the bootstrap Wn∗ → XF w.p.1. d

A diﬃculty in deriving a conﬁdence region for a function of ΣF is the redundancy of elements due to the symmetry of ΣF . So let uvec(S) = (s11 , s12 , s22 , . . . , s1p , s2p , . . . , spp ) be the vec operator applied only to the upper triangular part of S ∈ Sp . Suppose a conﬁdence region for g(ΣF ) ∈ Rk is desired where g is a function of uvec(ΣF ), which is continuously diﬀerentiable and has ﬁrst derivative g˙ ∈ Rkp(p+1)/2 . Let u : Rk → R be continuous on Rk such that {z ∈ Rk : u(z) = c} has Lebesgue measure 0 for every c ∈ R and u(bz) = bu(z), ∀z ∈ Rk , ∀b > 0. Let ! Wn,g = u n1/2 (g(Sn ) − g(ΣF )) . The delta method (v. Proposition 6.2) immediately yields d

˙ F ) uvec(XF )] Wn,g → u[g(Σ

(14.1)

and ∗ ˙ F ) uvec(XF )] w.p.1, Wn,g → u[g(Σ d

so the bootstrap is consistent [Beran and Srivastava (1985)]. The condition on u implies the limiting distribution in (14.1) is continuous. A bootstrap conﬁdence region for g(ΣF ) having asymptotic coverage probability 1 − α is Dn,g (α) = {g(ΣF ) : u [g(Sn ) − g(ΣF )] ≤ c∗n,g (α)}, where c∗n,g (α) is a (1 − α)-quantile of the distribution of u[g(S∗n ) − g(ΣFˆn )] when Fˆn is ﬁxed at its realized value and S∗n is the bootstrap sample variance.

248

14. Bootstrap conﬁdence regions and tests

Example 14.4 Let g(ΣF ) = ρF,ij =

σF,ij . 1/2 1/2 σF,ii σF,jj

Then, g(Sn ) is the sample correlation coeﬃcient rn,ij and bootstrap conﬁdence regions based on |rn,ij − ρF,ij | have the correct asymptotic coverage probability. Also covered is the Fisher z-transform 1 + ρF,ij 1 = tanh−1 (ρF,ij ). g(ΣF ) = 2 ln 1 − ρF,ij Example 14.5 This example provides the bootstrap algorithm, in an easily programmable form, to construct a conﬁdence interval for the correlation coeﬃcient ρF,ij using the Fisher z-transformation to stabilize the variance. We are given a sample x1 , . . . , xn from F . Bootstrap algorithm • • • • •

Calculate Sn = (sn,ij ). 1/2 1/2 Calculate rn,ij = sn,ij /[sn,ii sn,jj ]. b←1 B ← 2000 (say) Do while b ≤ B. • Draw a bootstrap sample x∗1 , . . . , x∗n from Fˆn . • Calculate S∗n = (s∗n,ij ). 1/2

• • • •

1/2

∗ • Calculate rn,ij = s∗n,ij /[s∗n,ii s∗n,jj ]. −1 ∗ • ub ← tanh (rn,ij ) − tanh−1 (rn,ij ) • b←b+1 End. Order the ub ’s: u(1) ≤ u(2) ≤ · · · ≤ u(B) . q ← (1 − α)B! An approximate (1 − α) conﬁdence interval for ρF,ij is

tanh[tanh−1 (rn,ij ) − u(q) ] ≤ ρF,ij ≤ tanh[tanh−1 (rn,ij ) + u(q) ]. Example 14.6 Let φ1 (ΣF ) > φ2 (ΣF ) > · · · > φp (ΣF ) > 0 be the ordered eigenvalues of ΣF assumed distinct. The vector φ(ΣF ) = (φ1 (ΣF ), . . . , φp (ΣF )) is a continuously diﬀerentiable function of uvec(ΣF ) [Kato (1982), Section 6 of Chapter 2]. The ordered sample eigenvalues are φ(Sn ) = (φ1 (Sn ), . . . , φp (Sn )) . The bootstrap conﬁdence region based on max | ln φi (Sn ) − ln φi (ΣF )|

1≤i≤p

(14.2)

14.3. Tests on the variance

249

has the correct asymptotic coverage probability. Here, u(z) = max1≤i≤p |zi |, z ∈ Rp . The logarithmic transformation stabilizes the variance in the normal model asymptotic for sample eigenvalues (v. Problem 8.9.15). The bootstrap conﬁdence region for φ(ΣF ) corresponding to (14.2) is {φi (ΣF ) : φi (Sn )/An ≤ φi (ΣF ) ≤ φi (Sn )An , i = 1, . . . , p}, where ∗

An = ecn,g (α) and c∗n,g (α) is a (1 − α)-quantile of the distribution of max | ln φi (S∗n ) − ln φi (ΣFˆn )|

1≤i≤p

when Fˆn is ﬁxed at its realized value and S∗n is the bootstrap sample variance. The problem of eﬃciently bootstrapping sample eigenvalues when ΣF may have multiple eigenvalues is still an unresolved problem [Beran and Srivastava (1987), Eaton and Tyler (1991)].

14.3 Tests on the variance Rather than inverting a conﬁdence region, it is sometimes possible to construct bootstrap tests directly from test statistics. This approach [Beran and Srivastava (1985)] to testing structural hypotheses about ΣF is the subject of this section. Assume x1 , . . . , xn are i.i.d. F with ﬁnite fourth moments. Let π : Pp → Pp be a linear projection (π 2 = π), not the identity map. Suppose Tn (Sn ) = n h(Sn ) is a test statistic for the null hypothesis, H0 : ΣF = π(ΣF ). Let F0 be any distribution function satisfying H0 . Example 14.7 Deﬁne the constant linear projection π(ΣF ) = I. Then, the hypothesis H0 : ΣF = I is equivalent to H0 : ΣF = π(ΣF ). Example 14.8 Partition ΣF as σF,11 ΣF = σ F,21 and deﬁne the linear projection π(ΣF ) =

σF,11 0

σ F,21 ΣF,22

0 ΣF,22

.

The hypothesis on the multiple correlation, H0 : R = 0, or H0 : σ 12 = 0 is equivalent to H0 : ΣF = π(ΣF ).

250

14. Bootstrap conﬁdence regions and tests

p Example 14.9 Deﬁne the linear map π(ΣF ) = ( i=1 σF,ii /p) I. Then, the sphericity hypothesis H0 : ΣF = γI, γ > 0 is equivalent to H0 : ΣF = π(ΣF ). The function h deﬁning the test statistic Tn (Sn ) is twice continuously diﬀer˙ F ) = 0. This entiable at uvec(ΣF0 ) ∈ Rp(p+1)/2 , with h(ΣF0 ) = 0 and h(Σ 0 formulation includes the normal model likelihood ratio test in particular. d ¨ ∈ Rp(p+1)/2 denote the second derivative of h and xF = Let h uvec(XF0 ). 0 p(p+1)/2 Then, using the Taylor series, ¨ F )xF . Tn (Sn )|F0 → xF0 h(Σ 0 0 d

We can construct a bootstrap estimate for the null distribution of Tn (Sn ) as follows. Let −1/2

Vn,F = [π(ΣF )]1/2 ΣF

−1/2

Sn ΣF

[π(ΣF )]1/2 .

The bootstrap estimate for the null distribution of Tn (Sn ) is deﬁned to be that of Tn (Vn,Fˆn ). Let dn,h (α, Fˆn ) be a (1 − α)-quantile of Tn (Vn,Fˆn ). Beran and Srivastava (1985) established the consistency of the bootstrap, ¨ F )xF w.p.1. Tn (Vn,Fˆn ) → xF0 h(Σ 0 0 d

Hence, the test which rejects H0 whenever Tn (Sn ) > dn,h (α, Fˆn ) has ¨ F ) = 0. asymptotic size α, provided h(Σ 0 In practice the bootstrap null distribution can be constructed as follows. Let yi = [π(Sn )]1/2 Sn−1/2 xi , i = 1, . . . , n. Let Fˆn,y be the empirical distribution function of the yi ’s. Note that ΣFˆn,y = π(ΣFˆn ), which satisﬁes H0 since π = π 2 . If y1∗ , . . . , yn∗ are d

i.i.d. Fˆn,y and S∗n,y is the sample variance of the yi∗ ’s, then Tn (Vn,Fˆn ) = Tn (S∗n,y ) whose distribution can be approximated by Monte Carlo methods. Example 14.10 We wish to test the hypothesis H0 : ΣF,12 = 0 using the invariant test statistic (v. Section 11.3) Tn

−1 = n tr[Sn,12 S−1 n,22 Sn,21 Sn,11 ]

= n

p1

2 rn,i ,

i=1 2 rn,i

where are the squared sample canonical correlations. The linear projection in this case is deﬁned by Sn,11 0 π(Sn ) = 0 Sn,22

14.3. Tests on the variance

with its square root

1/2

[π(Sn )]

=

1/2

0

0

Sn,11

251

1/2

Sn,22

.

We are given a sample x1 , . . . , xn from F . Bootstrap algorithm • Calculate Sn and partition

Sn =

Sn,11 Sn,21

1/2

1/2

Sn,12 Sn,22

. 1/2

−1/2

• Calculate the square roots Sn , Sn,11 , and Sn,22 and the inverse Sn • • • •

.

−1/2 [π(Sn )]1/2 Sn xi ,

i = 1, . . . , n. Transform yi = b←1 B ← 2000 (say) Do while b ≤ B. • Draw a bootstrap sample y1∗ , . . . , yn∗ from Fˆn,y . • Calculate S∗n and partition ∗ Sn,11 S∗n,12 . S∗n = S∗n,21 S∗n,22 −1

−1

• ub ← n tr[S∗n,12 S∗n,22 S∗n,21 S∗n,11 ] • b←b+1 • End. • Order the ub ’s: u(1) ≤ u(2) ≤ · · · ≤ u(B) . • q ← (1 − α)B! • An approximate size α test rejects H0 : ΣF,12 = 0 whenever Tn > u(q) . It is an easy matter to modify this bootstrap algorithm to bootstrap the test statistic

! −1 −1 −1 −1 Tn = n tr Sn,12 S−1 n,22 Sn,21 Sn,11 I − Sn,12 Sn,22 Sn,21 Sn,11 = n

p1

2 2 rn,i /(1 − rn,i ).

i=1

However, the test based on the largest sample canonical correlation, Tn = 2 , should not be bootstraped unless the user is sure the largest populan rn,1 tion canonical correlation is distinct. In case of multiplicity the population canonical correlations are not a diﬀerentiable function of ΣF [Kato (1982), Section 6 of Chapter 2]. Bootstrap algorithms for estimating the power function of a test statistic can be found in Beran (1986). Nagao and Srivastava (1992) considered high-order asymptotic expansions to the distribution of some test criteria on the variance matrix under local alternatives. For the test of sphericity

252

14. Bootstrap conﬁdence regions and tests

in dimension p = 3, they compared these expansions to the bootstrap approximations for both the normal model likelihood ratio test and Nagao’s test when the distribution is actually multivariate normal or multivariate t.

14.4 Problem 1. John (1971) showed that the test based on J = tr V2 /(tr V)2 ,

n ¯ )(xi − x ¯ ) , is LBI for the hypothesis of sphericwhere V = i=1 (xi − x ity, H0 : ΣF = γI, γ > 0, when the underlying distribution F is multivariate normal. Write down a detailed bootstrap algorithm to evaluate the α critical point of the test J but when F is multivariate student, tp,ν (0, I). Hint: A tp,ν (0, I) distribution can be simulated with Problem 13.6.1.

Appendix A

Assume x ∼ F , y ∼ G, x given by H(t)

|=

Inversion formulas

y on Rn . Then, z = x + y has a d.f., z ∼ H,

= P (x + y ≤ t) = E P (x + y ≤ t|y) = E F (t − y) = F (t − y)dG(y). Rn

Similarly, inverting the roles of x and y, we also have G(t − x)dF (x). H(t) = E G(t − x) = Rn

This leads to the smoothing lemma on convolution. |=

Lemma A.1 (Smoothing lemma) If x is absolutely continuous with y, is absolutely p.d.f. f (t), then z = x + y, where y ∼ G and x continuous with p.d.f. h(t) = E f (t − y). Proof. It follows readily that F (t − y)dG(y) H(t) = Rn = Rn

(−∞,t−y]

f (x)dx dG(y)

254

Appendix A. Inversion formulas

= Rn

(−∞,t]

f (x − y)dx dG(y).

By Tonelli’s theorem, it is posible to interchange the order of integration whereby H(t) = f (x − y)dG(y) dx = E f (x − y)dx. (−∞,t] Rn (−∞,t] 2 We can now establish the inversion formula on Rn . The proof resembles that of Feller (1966, p. 480) for n = 1. Proposition A.1 (Inversion formula) The probability measure Px is given in terms of the characteristic function c(t) = cx (t) by 2 1 Px (a, b] = lim e−it x c(t)e−t t/2N dtdx, n N →∞ (2π) (a,b] Rn

Proof. Take any random t such that t

|=

∀a, b such that Px (∂(a, b]) = 0. x. Then, conditioning yields

E eix t = E E(eix t |x) = E ct (x) = E cx (t). Replace x by x − s for any ﬁxed value of s to ﬁnd Parseval’s relation:

E ct (x − s) = E e−is t cx (t). However, letting t ∼ Nn (0, σ −2 I) with ct (s) = exp(−|s|2 /2σ 2 ), E exp(−|s − x|2 /2σ 2 )

= E e−it s c(t) 2 n/2

σ = e−it s c(t) exp −σ 2 |t|2 /2 dt. 2π Rn

Divide by (2πσ 2 )n/2 to obtain E

1 1 exp(−|s−x|2 /2σ 2 ) = 2 n/2 (2π)n (2πσ )

Rn

e−it s c(t) exp −σ 2 |t|2 /2 dt.

This is of the form E g(s − x) = h(s) in the smoothing lemma where g(s) is the p.d.f. for a Nn (0, σ 2 I). Thus, h(s) is the p.d.f. of x + σz, where z ∼ Nn (0, I), and if we let Pσ denote the probability measure for x + σz, 2 2 1 Pσ (a, b] = e−it s c(t)e−σ |t| /2 dtds, (2π)n (a,b] Rn whereby Slutsky’s theorem with σ = 1/N gives the result.

2

An immediate corollary is the inversion formula for absolutely continuous distribution.

Appendix A. Inversion formulas

255

Corollary A.1 If c(t) is integrable with respect to Lebesgue measure, then 1 f (x) = e−it x c(t)dt. n (2π) Rn Proof. If c(t) is integrable, then the integrand in Proposition A.1 is dominated by an integrable function. By the D.C.T., we can interchange the limit and the integral, which gives the result. 2

Appendix B Multivariate cumulants

B.1 Deﬁnition and properties The moments of a univariate random variable x, µr = E xr , are the coeﬃcients of (it)r /r! in the Taylor series of the characteristic function, cx (t) =

∞

µr (it)r /r!

r=0

whereas the cumulants are the coeﬃcients in the series for Kx (t) ≡ ln[cx (t)], Kx (t) =

∞

kr (it)r /r!,

r=0

provided the expansions are valid. The function Kx (t) is the cumulant generating function. Relations between moments and cumulants are thus obtained by equating the coeﬃcients in the Taylor series of exp(·) in the equation ∞ ∞ r r µr (it) /r! = exp kr (it) /r! . r=0

r=0

We require only the relations between the ﬁrst four moments and cumulants (assuming they exist): µ1 µ2

= k1 , = k2 + k12 ,

B.1. Deﬁnition and properties

µ3 µ4

= k3 + 3k2 k1 + k13 ,

k1

= µ1 , = µ2 − µ21 ,

k2 k3 k4

257

= k4 + 4k3 k1 + 3k22 + 6k2 k12 + k14 ,

= µ3 − 3µ2 µ1 + 2µ31 , = µ4 − 4µ3 µ1 − 3µ22 + 12µ2 µ21 − 6µ41 .

When x is centered, i.e., E x = µ1 = k1 = 0, these simplify to µ2 = k2 , µ3 = k3 , µ4 = k4 + 3k22 ,

k2 = µ2 , k3 = µ3 , k4 = µ4 − 3µ22 . r

r1 p For a random vector x ∈ Rp , product-moments µr1 ,...,r pp = E(x1 · · · xp ) and multivariate cumulants kr1 ,...,rp of order r = i=1 ri are deﬁned similarly,

cx (t)

∞

=

µr1 ,...,rp

(itp )rp (it1 )r1 ··· , r1 ! rp !

kr1 ,...,rp

(itp )rp (it1 )r1 ··· . r1 ! rp !

r1 ,...,rp =0

Kx (t) = ln[cx (t)]

∞

=

r1 ,...,rp =0

Example B.1 For x ∼ Np (µ, Σ), we have Kx (t) = it µ − 12 t Σt, a quadratic function of t, and, thus, all multivariate cumulants of order r > 2 are null. Multivariate cumulants of order 1 are the means, µi , and those of order 2 are the variances, σii , and covariances, σij . Obtaining product-moments in terms of cumulants, and vice versa, is a laborious task which can be greatly simpliﬁed with a “symbolic diﬀerential operator” [Kendall et al. (1987)]. For example, when E x = 0, consider the relation µ4 = k4 + 3k22 , which we write symbolically as µ(r14 ) = k(r14 ) + 3k 2 (r12 ). To obtain a relation between fourth-order product-moments and cumulants of a bivariate distribution, consider the operator r2 ∂(·)/∂r1 . When applied to µ(r14 ), it yields 4µ(r13 r2 ) = 4k(r13 r2 ) + 12k(r12 )k(r1 r2 ), which means, after dividing by 4, µ31 = k31 + 3k20 k11 . Example B.2 The same method can be used to obtain cumulants in terms of product-moments. Considering the relation k4 = µ4 − 3µ22

258

Appendix B. Multivariate cumulants

in symbolic form k(r14 ) = µ(r14 ) − 3µ2 (r12 ), and applying the operator r2 ∂(·)/∂r1 , we get 4k(r13 r2 ) = 4µ(r13 r2 ) − 12µ(r12 )µ(r1 r2 ) or k31 = µ31 − 3µ20 µ11 . Continuing this process, it is possible to obtain relations for trivariate distributions with either operator, r3 ∂(·)/∂r1 or r3 ∂(·)/∂r2 . The operator r3 ∂(·)/∂r1 applied to the last symbolic equation yields 12µ(r12 r2 r3 ) = 12k(r12 r2 r3 ) + 24k(r1 r3 )k(r1 r2 ) + 12k(r12 )k(r2 r3 ), which is equivalent to µ211 = k211 + 2k101 k110 + k200 k011 . The operator r4 ∂(·)/∂r1 ﬁnally gives the relation for fourth-order productmoments and cumulants of a four-dimensional distribution 24µ(r1 r2 r3 r4 ) = 24k(r1 r2 r3 r4 ) + 24k(r3 r4 )k(r1 r2 ) + 24k(r1 r3 )k(r2 r4 ) +24k(r1 r4 )k(r2 r3 ), or µ1111 = k1111 + k0011 k1100 + k1010 k0101 + k1001 k0110 . For fourth-order product-moments of a p-dimensional, p > 4, distribution, we need only specify which four variables enter. For example, µijkl 1111 = E(xi xj xk xl ) satisﬁes ijkl ijkl ijkl ijkl ijkl ijkl ijkl µijkl 1111 = k1111 + k0011 k1100 + k1010 k0101 + k1001 k0110 .

A zero subscript means the superscript variable does not enter, so we can rewrite ijkl kl ij ik jl il jk µijkl 1111 = k1111 + k11 k11 + k11 k11 + k11 k11 .

When a variable is repeated, the indices can be amalgamated. For example, the equation where i = j, iikl kl ii ik il il ik µiikl 1111 = k1111 + k11 k11 + k11 k11 + k11 k11 ,

becomes ikl kl i ik il il ik µikl 211 = k211 + k11 k2 + k11 k11 + k11 k11 ,

and if i = j = k = l, then we recover the initial equation µi4 = k4i + 3(k2i )2 .

B.2. Application to asymptotic distributions

259

Departures from normality is often assessed with the coeﬃcients of skewness, γ1 , and kurtosis, γ2 . For a centered variable x, they are deﬁned as µ3 k3 = 3/2 , γ1 = 3/2 µ2 k2 µ4 k4 γ2 = − 3 = 2. 2 µ2 k2 For a normal variable, γ1 = γ2 = 0. Cumulants of random symmetric matrices can also be deﬁned. For a description of miminal moments and cumulants of symmetric matrices with an application to the Wishart distribution, the reader is referred to Kollo and von Rosen (1995).

B.2 Application to asymptotic distributions Let x1 , . . . , xn i.i.d. x ∈ Rp which has ﬁnite fourth-order moments and Ex = 0 and var x = Σ. The asymptotic distribution of S = n 1 ¯ )(xi − x ¯ ) was derived generally in Section 6.3: i=1 (xi − x (n−1) n1/2 (S − Σ) → Npp (0, var W), d

where W = xx . The only problem is to calculate var W. This can now be done in terms of multivariate cumulants. The block (i, j) of var W is E(xi xj xx ) − E(xi x)E(xj x ) and the element (k, l) of the block (i, j) becomes E(xi xj xk xl ) − E(xi xk )E(xj xl )

ik jl = µijkl 1111 − µ11 µ11 ijkl kl ij il jk = k1111 + k11 k11 + k11 k11 .

The general result is thus ijkl kl ij il jk + k11 k11 + k11 k11 . cov(wik , wjl ) = k1111

B.3 Problems 1. Establish the following: (i) µ11 = k11 and µ21 = k21 . 2 . (ii) µ22 = k22 + k20 k02 + 2k11 (iii) Given µ5 = k5 + 10k3 k2 , calculate µ32 and µ41 . (iv) Obtain µ301 in terms of lower-order cumulants.

260

Appendix B. Multivariate cumulants

2. Demonstrate the kurtosis γ2 of a symmetric contaminated normal density (1 − )(2π)−1/2 exp(− 12 x2 ) + (2πσ)−1/2 exp(− 12 x2 /σ 2 ) is γ2 = 3

[1 + (σ 4 − 1)] − 3. [1 + (σ 2 − 1)]2

3. Evaluate the kurtosis of a Student’s tν distribution as γ2 = 6/(ν − 4), ν > 4.

Appendix C S-plus functions

This appendix describes three S-plus programs which the reader can download from the World Wide Web site www.dms.umontreal.ca/∼bilodeau. Simply download the ﬁle named multivariate and, at the S-plus prompt, type: source(“multivariate”) to compile the functions. 1. U (p; m, n) distribution function. Usage: pu(ζ, p, m, n) Value: The function returns P (U (p; m, n) ≤ ζ). 2. U (p; m, n) quantiles. Usage: qu(α, p, m, n) Value: The function returns the α-quantile , Uα (p; m, n) say, satisfying P (U (p; m, n) ≤ Uα (p; m, n)) = α. It returns as well a Cf actor, frequently used by people relying on the asymptotic result −[n − 12 (p − m + 1)] ln U (p; m, n) → χ2pm , d

to make the approximate χ2pm quantile an exact quantile of −[n − 1 2 (p − m + 1)] ln U (p; m, n). More precisely, Cf actor · χ21−α,pm = −[n − 12 (p − m + 1)] ln Uα (p; m, n). Note that lower quantiles of U (p; m, n) correspond to upper quantiles of χ2pm .

262

Appendix C. S-plus functions

3. Beta Q-Q plot for multivariate normality. Usage: qqbeta(x) The input x is the n × p sample matrix. Value: The function returns the Q-Q plot of the points

d2(i) , [(n − 1)2 /n] betaγi 12 p; 12 (n − p − 1) , i = 1, . . . , n, as described in Section 11.4.1. The graphic device must be activated before using this function. 4. Robust S estimate. Usage: s.estimate(x, r, nr, N samp) The input x denotes the n × p sample matrix. The input r in the interval (0, .5) is the asymptotic breakdown point, nr and N samp are positive integer parameters of the numerical algorithm [Ruppert (1992)]. Values of nr = 3 and N samp = 80p are appropriate for most purposes. The user is urged to experiment with other values of nr and N samp to certify that the s.estimate function returned the global minimum. Value: The function returns the S estimate of location and scatter, µn and Vn , the Mahalanobis distances, distance.mahalanobis, for outlier detection and the objective function, determinant, which the S estimate seeks to minimize. Points with a Mahalanobis distance greater than (χ2.95,p )1/2 should be checked for outliers [Rousseeuw and van Zomeren (1990)]. The implementation uses the biweight ρ(·) function of Section 13.4.2 and determines c0 such that E ρ(|z|)/(c20 /6) = r, where z ∼ Np (0, I), to achieve the desired breakdown point. 5. Asymptotic variance of S estimate. Usage: asymp(p, r) The input p is the number of variables, whereas r is the breakdown point. Value: The function returns the asymptotic variance constants, at the normal distribution, in Proposition 13.11: λ = α/β 2 , σ1 , and σ2 . The constants λ−1 and σ1−1 , in particular, serve as measures of relative eﬃciencies of the location and scatter estimates, respectively, at the normal distribution.

References

[1] Ali, M.M., and R. Ponnapalli (1990). An optimal property of the GaussMarkoﬀ estimator. Journal of Multivariate Analysis 32, 171-176. [2] Anderson, G.A. (1965). An asymptotic expansion for the distribution of the latent roots of the estimated covariance matrix. Annals of Mathematical Statistics 36, 1153-1173. [3] Anderson, T.W. (1963). Asymptotic theory for principal component analysis. Annals of Mathematical Statistics 34, 122-148. [4] Anderson, T.W. (1965). An asymtotic expansion for the distribution of the latent roots of the estimated covariance matrix. Annals of Mathematical Statistics 36, 1153-1173. [5] Anderson, T.W. (1984). An Introduction to Multivariate Statistical Analysis. 2nd ed. John Wiley & Sons, New York. [6] Andrews, D.F., R. Gnanadesikan, and J.L. Warner (1971). Transformations of multivariate data. Biometrics 27, 825-840. [7] Andrews, D.F., R. Gnanadesikan, and J.L. Warner (1973). Methods for assessing multivariate normality. Multivariate Analysis, ed. P.K. Krishnaiah. Academic Press, New York, 95-116. [8] Ash, R. (1972). Real Analysis and Probability. Academic Press, New York. [9] Baringhaus, L., and N. Henze(1991). Limit distributions for measures of multivariate skewness and kurtosis based on projections. Journal of Multivariate Analysis 38, 51-69. [10] Bartlett, M.S. (1937). Properties of suﬃciency and statistical tests. Proceedings of the Royal Society. London. Series A. 160, 268-282. [11] Bartlett, M.S. (1938). Further aspects of the theory of multiple regression. Proceedings of the Cambridge Philosophical Society 34, 33-40.

264

References

[12] Bellman, R. (1960). Introduction to Matrix Analysis. McGraw-Hill, New York. [13] Bentler, P.M. (1983). Some contributions to eﬃcient statistics in structural models: Speciﬁcation and estimation of moment structures. Psychometrika 48, 493-517. [14] Beran, R. (1984). Bootstrap methods in statistics. Jahresbericht der Deutschen Mathematiker-Vereinigung 86, 14-30. [15] Beran, R. (1986). Simulated power functions. Annals of Statistics 14, 151173. [16] Beran, R. (1987). Prepivoting to reduce level error in conﬁdence sets. Biometrika 74, 457-468. [17] Beran, R. (1988). Prepivoting test statistics: a bootstrap view of asymptotic reﬁnements. Journal of the American Statistical Association 83, 687-697. [18] Beran, R., and M.S. Srivastava (1985). Bootstrap tests and conﬁdence regions for functions of a covariance matrix. Annals of Statistics 13, 95-115. [19] Beran, R., and M.S. Srivastava (1987). Correction: Bootstrap tests and conﬁdence regions for functions of a covariance matrix. Annals of Statistics 15, 470-471. [20] Berk, R., and J.T. Hwang (1989). Optimality of the least squares estimator. Journal of Multivariate Analysis 30, 245-254. [21] Berkane, M., K. Oden, and P.M. Bentler (1997). Geodesic estimation in elliptical distributions. Journal of Multivariate Analysis 63, 35-46. [22] Bhat, B.R. (1981). Modern Probability Theory. John Wiley & Sons, New York. [23] Bickel, P.J., and D.A. Freedman (1981). Some asymptotic theory for the bootstrap. Annals of Statistics 9, 1196-1217. [24] Billingsley, P. (1968). Convergence of Probability Measures. John Wiley & Sons, New York. [25] Billingsley, P. (1995). Probability and Measure. 3rd ed. John Wiley & Sons, New York. [26] Bilodeau, M. (1988). On the simultaneous estimation of scale parameters. The Canadian Journal of Statistics 16, 169-174. [27] Bilodeau, M. (1990). On the choice of the loss function in covariance estimation. Statistics & Decisions 8, 131-139. [28] Bilodeau, M. (1995). Minimax estimators of the mean vector in normal mixed linear models. Journal of Multivariate Analysis 52, 73-82. [29] Bilodeau, M.(1996). Some remarks on U (p; m, n) distributions. Statistics and Probability Letters 31, 41-43. [30] Bilodeau M. (1997a). Estimating a multivariate treatment eﬀect under a biased allocation rule. Communications in Statistics, Theory and Methods 26, 1119-1124. [31] Bilodeau, M. (1997b). Robust test for a given variance. Technical Report, Universit´e de Montr´eal.

References

265

[32] Bilodeau, M. (1998). Multivariate ﬂattening for better predictions. Technical Report, Universit´e de Montr´eal. [33] Bilodeau, M., and M.S. Srivastava (1989a). Estimation of the MSE matrix of the Stein estimator. The Canadian Journal of Statistics 16, 153-159. [34] Bilodeau, M., and M.S. Srivastava (1989b). Stein estimation under elliptical distributions. Journal of Multivariate Analysis 28, 247-259. [35] Bilodeau, M., and M.S. Srivastava (1992). Estimation of the eigenvalues of Σ1 Σ−1 2 . Journal of Multivariate Analysis 41, 1-13. [36] Bilodeau, M., and T. Kariya (1989). Minimax estimators in the normal MANOVA model. Journal of Multivariate Analysis 28, 260-270. [37] Bilodeau, M., and T. Kariya (1994). LBI tests of independence in bivariate exponential distributions. Annals of the Institute of Statistical Mathematics 46, 127-136. [38] Blom, G. (1958). Statistical Estimates and Transformed Beta-variables. John Wiley & Sons, New York. [39] Boente, G. (1987). Asymptotic theory for robust principal components. Journal of Multivariate Analysis 21, 67-78. [40] Boulerice, B., and G.R. Ducharme (1997). Smooth tests of goodness-of-ﬁt for directional and axial data. Journal of Multivariate Analysis 60, 154-175. [41] Box, G.E.P. (1949). A general distribution theory for a class of likelihood criteria. Biometrika 36, 317-346. [42] Box, G.E.P., and D.R. Cox (1964). An analysis of transformations. Journal of the Royal Statistical Society B 26, 211-252. [43] Breiman, L., and J.H. Friedman (1997). Predicting multivariate responses in multiple linear regression. Journal of the Royal Statistical Society B 59, 3-54. [44] Brown, P.J. (1980). Aspects of multivariate regression (with discussion). Bayesian Statistics. eds. J.M. Bernardo, M.H. DeGroot, D.V. Lindley, and A.F.M. Smith. Valencia University Press, Valencia. [45] Browne, M.W., and A. Shapiro (1987). Adjustments for kurtosis in factor analysis with elliptically distributed errors. Journal of the Royal Statistical Society B 49, 346-352. [46] Carri`ere, J.F. (1994). Dependent decrement theory. Transactions XLVI, Society of Actuaries, 45-65. [47] Casella, G., and R.L. Berger (1990). Statistical Inference, Duxbury Press, Belmont, California. [48] Chattopadhyay, A.K., and K.C.S. Pillai (1973). Asymptotic expansions for the distributions of characteristic roots when the parameter matrix has several multiple roots. Multivariate analysis III. Academic Press, New York, 117-127. [49] Chikuse, Y. (1976). Asymptotic distributions of the latent roots of the covariance matrix with multiple population roots. Journal of Multivariate Analysis 6, 237-249.

266

References

[50] Cl´eroux, R., and G.R. Ducharme (1989). Vector correlation for elliptical distributions. Communications in Statistics, Theory and Methods 18, 14411454. [51] Coelho, C.A. (1998). The generalized integer gamma distribution–A basis for distributions in multivariate statistics. Journal of Multivariate Analysis 64, 86-102. [52] Cook, R.D., M.E. Johnson (1981). A family of distributions for modelling nonelliptically symmetric multivariate data. Journal of the Royal Statistical Society B 43, 210-218. [53] Copas, J.B. (1975). On the unimodality of the likelihood for the Cauchy distribution. Biometrika 62, 701-704. [54] Courant, R. (1936). Diﬀerential and Integral Calculus II. John Wiley & Sons, New York. [55] Cox, D.R., and N.J.H. Small (1978). Testing multivariate normality. Biometrika 65, 263-272. [56] Cuadras, C.M. (1992). Probability distributions with given multivariate marginals and given dependence structure. Journal of Multivariate Analysis 42, 51-66. [57] Datta, S., N. Mukhopadhyay (1997). On sequential ﬁxed-size conﬁdence regions for the mean vector. Journal of Multivariate Analysis 60, 233-251. [58] Davies, P.L. (1987). Asymptotic behaviour of S-estimates of multivariate location parameters and dispersion matrices. Annals of Statistics 15, 12691292. [59] Davis, A.W. (1971). Percentile approximations for a class of likelihood ratio criteria. Biometrika 58, 349-356. [60] Davison, A.C., and D.V. Hinkley (1997). Bootstrap Methods and their Application. Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press, Cambridge. [61] Donoho, D.L. (1982). Breakdown Properties of Multivariate Location Estimators. Qualifying paper, Harvard University. [62] Ducharme, G.R., P. Milasevic (1987). Spatial median and directional data. Biometrika 74, 212-215. [63] D¨ umbgen, L. (1998). Perturbation inequalities and conﬁdence sets for functions of a scatter matrix. Journal of Multivariate Analysis 65, 19-35. [64] Dykstra, R.L. (1970). Establishing the positive deﬁniteness of the sample covariance matrix. Annals of Mathematical Statistics 41, 2153-2154. [65] Eaton, M.L. (1983). Multivariate Statistics, a Vector Space Approach. John Wiley & Sons, New York. [66] Eaton, M.L. (1988). Concentration inequalities for Gauss-Markov estimators. Journal of Multivariate Analysis 25, 119-138. [67] Eaton, M.L., and M.D. Perlman (1973). The non-singularity of generalized sample covariance matrices. Annals of Statistics 1, 710-717. [68] Eaton, M.L., and D.E. Tyler (1991). On Wielandt’s inequality and its application to the asymptotic distribution of the eigenvalues of a random symmetric matrix. Annals of Statistics 19, 260-271.

References

267

[69] Eaton, M.L., and D.E. Tyler (1994). The asymptotic distribution of singular values with applications to canonical correlations and correspondence analysis. Journal of Multivariate Analysis 50, 238-264. [70] Efron, B. (1969). Student’s t-test under symmetry conditions. Journal of the American Statistical Association 64, 1278-1302. [71] Efron B. (1982). The Jacknife, the Bootstrap and Other Resampling Plans. SIAM, Philadelphia. bibitemefrm Efron B., and C. Morris (1976). Multivariate empirical Bayes and estimation of covariance matrix. Annals of Statistics 4, 22-32. indexaiMorris, C. [72] Efron, B., and R.J. Tibshirani (1993). An Introduction to the Bootstrap. Chapman & Hall, New York. indexaiTibshirani, R.J. [73] Erd´elyi, A., W. Magnus, F. Oberhettinger, and F.G. Tricomi (1953). Higher Transcendental Functions. McGraw-Hill, New York. [74] Escouﬁer, Y. (1973). Le traitement des variables vectorielles. Biometrics 29, 751-760. [75] Fan, Y. (1997). Goodness-of-ﬁt tests for a multivariate distribution by the empirical characteristic function. Journal of Multivariate Analysis 62, 3663. [76] Fang, K.T., S. Kotz, and K.W. Ng (1991). Symmetric Multivariate and Related Distributions. Chapman & Hall, London. [77] Fang, K.T., L.-X. Zhu, and P.M. Bentler (1993). A necessary test of goodness of ﬁt for sphericity. Journal of Multivariate Analysis 45, 34-55. [78] Feller, W. (1966). An Introduction to Probability Theory and Its Applications (Vol. II). John Wiley & Sons, New York. [79] Fisher, R.A. (1953). Dispersion on a sphere. Proceedings of the Royal Society. London. Series A. 217, 295-305. [80] Flury, B. (1997). A First Course in Multivariate Statistics. Springer-Verlag, New York. [81] Frank, M.J. (1979). On the simultaneous associativity of F (x, y) and x + y − F (x, y). Aequationes Math. 19, 194-226. [82] Fraser, D.A.S. (1976). Probability and Statistics: Theory and Applications. DAI Press, Toronto. [83] Fraser, D.A.S., I. Guttman, and M.S. Srivastava (1991). Conditional inference for treatment and error in multivariate analysis. Biometrika 78, 565-572. [84] Fujikoshi, Y. (1970). Asymptotic expansions of the distributions of test statistics in multivariate analysis. Journal of Science of the Hiroshima University. Series A, Mathematics 34, 73-144. [85] Fujikoshi, Y. (1977). An asymptotic expansion for the distributions of the latent roots of the Wishart matrix with multiple population roots. Annals of the Institute of Statistical Mathematics 29, 379-387. [86] Fujikoshi, Y. (1978). Asymptotic expansions for the distributions of some functions of the latent roots of matrices in three situations. Journal of Multivariate Analysis 8, 63-72.

268

References

[87] Fujikoshi, Y. (1988). Comparison of powers of a class of tests for multivariate linear hypothesis and independence. Journal of Multivariate Analysis 26, 48-58. [88] Fujikoshi, Y. (1997). An asymptotic expansion for the distribution of Hotelling’s T 2 -statistic under nonnormality. Journal of Multivariate Analysis 61, 187-193. [89] Fujikoshi, Y., and Y. Watamori (1992). Tests for the mean direction of the Langevin distribution with large concentration parameter. Journal of Multivariate Analysis 42, 210-225. [90] Fujisawa, H. (1997). Improvement on chi-squared approximation by monotone transformation. Journal of Multivariate Analysis 60, 84-89. [91] Fujisawa, H. (1997). Likelihood ratio criterion for mean structure in the growth curve model with random eﬀects. Journal of Multivariate Analysis 60, 90-98. [92] Genest, C., and R.J. MacKay (1986). Copules archim´ediennes et familles de lois bidimensionnelles dont les marges sont donn´ees. Canadian Journal of Statistics 14, 145-159. [93] Genest, C. (1987). Frank’s family of bivariate distributions. Biometrika 74, 549-555. [94] Ghosh, B. K. (1970). Sequential Tests of Statistical Hypotheses. AddisonWesley, Reading, Massachusetts. [95] Giri, N.C. (1996). Multivariate Statistical Analysis. Marcel Dekker, New York. [96] Gnanadesikan, R. (1977). Methods for Statistical Data Analysis of Multivariate Observations. John Wiley & Sons, New York. [97] Gnanadesikan, R., and J.R. Kettenring (1972). Robust estimates, residuals, and outlier detection with multiresponse data. Biometrics 28, 81-124. [98] Gr¨ ubel, R., and D.M. Rocke (1990). On the cumulants of aﬃne equivariant estimators in elliptical families. Journal of Multivariate Analysis 35, 203222. [99] Gunderson, B.K., and R.J. Muirhead (1997). On estimating the dimensionality in canonical correlation analysis. Journal of Multivariate Analysis 62, 121-136 [100] Gupta, A.K., and D. Song (1997). Characterization of p-generalized normality. Journal of Multivariate Analysis 60, 61-71. [101] Gupta, A.K., and D. St. P. Richards (1990). The Dirichlet distributions and polynomial regression. Journal of Multivariate Analysis 32, 95-102. [102] Gupta, A.K., and T. Varga (1992). Characterization of matrix variate normal distributions. Journal of Multivariate Analysis 41, 80-88. [103] Hall, P. (1992). The Bootstrap and Edgeworth Expansion. Springer-Verlag, New York. [104] Hendriks, H., Z. Landsman, and F. Ruymgaart (1996). Asymptotic behavior of sample mean direction for spheres. Journal of Multivariate Analysis 59, 141-152.

References

269

[105] Henze, N., and T. Wagner (1997). A new approach to the BHEP tests for multivariate normality. Journal of Multivariate Analysis 62, 1-23. [106] Henze, N., and B. Zirkler (1990). A class of invariant consistent tests for multivariate normality. Communications in Statistics, Theory and Methods 19, 3595-3617. [107] Hsu, P.L. (1941). On the limiting distribution of the canonical correlations. Biometrika 32, 38-45. [108] Huber, P.J. (1981). Robust Statistics. John Wiley & Sons, New York. [109] Iwashita, T. (1997). Asymptotic null and nonnull distribution of Hotelling’s T 2 -statistic under the elliptical distribution. Journal of Statistical Planning and Inference 61, 85-104.. [110] Iwashita, T., and M. Siotani (1994). Asymptotic distributions of functions of a sample covariance matrix under the elliptical distribution. The Canadian Journal of Statistics 22, 273-283. [111] James, A.T. (1954). Normal multivariate analysis and the orthogonal group. Annals of Mathematical Statistics 25, 40-75. [112] James, A.T. (1969). Tests of equality of latent roots of the covariance matrix. Multivariate Analysis II. Academic Press, New York, 205-218. [113] John, S. (1971). Some optimal multivariate tests. Biometrika 58, 123-127. [114] John, S. (1972). The distribution of a statistic used for testing sphericity of normal distributions. Biometrika 59, 169-174. [115] Johnson, N.L. (1949). Systems of frequency curves generated by methods of translation. Biometrika 36, 149-176. [116] Johnson, N.L., S. Kotz, and A.W. Kemp (1992). Univariate Discrete Distributions. 2nd ed. John Wiley & Sons, New York. [117] Johnson, R.A., and D.W. Wichern (1992). Applied Multivariate Statistical Analysis. 3rd ed. Prentice-Hall, Englewood Cliﬀs, New Jersey. [118] Jolliﬀe, I.T. (1986). Principal Component Analysis. Springer-Verlag, New York. [119] Jordan, S.M., and K. Krishnamoorthy (1995). Conﬁdence regions for the common mean vector of several multivariate normal populations. The Canadian Journal of Statistics 23, 283-297. [120] Kano, Y. (1994). Consistency property of elliptical probability density function. Journal of Multivariate Analysis 51, 139-147. [121] Kano, Y. (1995). An asymptotic expansion of the distribution of Hotelling’s T 2 -statistic under general condition. American Journal of Mathematical and Management Sciences 15, 317-341. [122] Kariya, T. (1985). Testing in the Multivariate General Linear Model. Kinokunia, Tokyo. [123] Kariya, T., and B.K. Sinha (1989). Robustness of Statistical Tests. Academic Press, San Diego. [124] Kariya, T., and M.L. Eaton (1977). Robust tests for spherical symmetry. Annals of Statistics 5, 206-215.

270

References

[125] Kariya, T., R.S. Tsay, N. Terui, and Hong Li (1999). Tests for multinormality with applications to time series. Communications in Statistics, Theory and Methods 28, 519-536. [126] Kariya, T., Y. Fujikoshi, and P.R. Krishnaiah (1987). On tests for selection of variables and independence under multivariate regression models. Journal of Multivariate Analysis 21, 207-237. [127] Kato, T. (1982). A Short Introduction to Perturbation Theory for Linear Operators. Springer-Verlag, New York. [128] Kelker, D. (1970). Distribution theory of spherical distributions and a location-scale parameter generalization. Sankhy¯ a A 32, 419-430. [129] Kendall, M., A. Stuart, and J.K. Ord (1987). Kendall’s Advanced Theory of Statistics. 5th ed. Vol. 1. Oxford University Press, New York. [130] Kent, J.T., and D.E. Tyler (1991). Redescending M-estimates of multivariate location and scatter. Annals of Statistics 19, 2102-2019. [131] Khatri, C.G., and M.S. Srivastava (1974). Asymptotic expansions of the non-null distributions of the likelihood ratio criteria for covariance matrices II. Proc. Carleton University, Ottawa. Metron 36, 55-71. [132] Khatri, C. G., and M.S. Srivastava (1978). Asymptotic expansions for distributions of characteristic roots of covariance matrices. South African Statistical Journal 12, 161-186. [133] Ko, D., and T. Chang (1993). Robust M-estimators on spheres. Journal of Multivariate Analysis 45, 104-136. [134] Koehler, K.J., and J.T. Symanowski (1995). Constructing multivariate distributions with speciﬁc marginal distributions. Journal of Multivariate Analysis 55, 261-282. [135] Kollo, T., and H. Neudecker (1993). Asymptotics of eigenvalues and unitlength eigenvectors of sample variance and correlation matrices. Journal of Multivariate Analysis 47, 283-300. [136] Kollo, T., and D. von Rosen (1995). Minimal moments and cumulants of symmetric matrices: an application to the Wishart distribution. Journal of Multivariate Analysis 55, 149-164. [137] Koltchinskii, V.I., and L. Li (1998). Testing for spherical symmetry of a multivariate distribution. Journal of Multivariate Analysis 65, 228-244. [138] Konishi, S. (1979). Asymptotic expansions for the distributions of statistics based on the sample correlation matrix in principal component analysis. Hiroshima Mathematical Journal 9, 647-700. [139] Konishi, S., and C.G. Khatri (1990). Inferences on interclass and intraclass correlations in multivariate familial data. Annals of the Institute of Statistical Mathematics 42, 561-580. [140] Konishi, S., and C.R. Rao (1991). Inferences on multivariate measures of interclass and intraclass correlations in familial data. Journal of the Royal Statistical Society B 53, 649-659. [141] Konishi, S., and C.R. Rao (1992). Principal component analysis for multivariate familial data. Biometrika 79, 631-641.

References

271

[142] Kotz, S., and I. Ostrovskii (1994). Characteristic functions of a class of elliptical distributions. Journal of Multivariate Analysis 49, 164-178. [143] Kres, H. (1983). Statistical Tables for Multivariate Analysis, a Handbook with References to Applications. Springer-Verlag, New York. [144] Kshirsagar, A.M. (1972). Multivariate Analysis. Marcel Dekker, New York. [145] Kudˆ o, A. (1963). A multivariate analogue of the one-sided test. Biometrika 50, 403-418. [146] Kuwana, Y., and T. Kariya (1991). LBI tests for multivariate normality in exponential power distributions. Journal of Multivariate Analysis 39, 117-134. [147] Lee, Y.-S. (1972). Some results on the distribution of Wilk’s likelihood-ratio criterion. Biometrika 59, 649-664. [148] Lehmann, E.L. (1983). Theory of Point Estimation. John Wiley & Sons, New York. [149] Li, Haijun, M. Scarsini, and M. Shaked (1996). Linkages: A tool for construction of multivariate distributions with given nonoverlapping multivariate marginals. Journal of Multivariate Analysis 56, 20-41. [150] Liu, C. (1995). Missing data imputation using the multivariate t distribution. Journal of Multivariate Analysis 53, 139-158. [151] Liu, C. (1997). ML estimation of the multivariate t distribution and the EM algorithm. Journal of Multivariate Analysis 63, 296-312. [152] Looney, S.W. (1995). How to use test for univariate normality to assess multivariate normality. The American Statistician 49, 64-70. [153] Lopuha¨ a, H.P. (1989). On the relation between S-estimators and Mestimators of multivariate location and covariance. Annals of Statistics 17, 1662-1683. [154] Lopuha¨ a, H.P. (1991). Multivariate τ -estimators for location and scatter. The Canadian Journal of Statistics 19, 307-321. [155] Lopuha¨ a, H.P., and P.J. Rousseeuw (1991). Breakdown points of aﬃne equivariant estimators of multivariate location and covariance matrices. Annals of Statistics 19, 229-248. [156] MacDuﬀy, C.C. (1943). Vectors and Matrices. The Mathematical Association of America, Providence, Rhode Island. [157] Magnus, J.R., and H. Neudecker (1979). The commutation matrix: Some properties and applications. Annals of Statistics 7, 381-394. [158] Malkovich, J.F., and A.A. Aﬁﬁ (1973). On tests for multivariate normality. Journal of the American Statistical Association 68, 176-179. [159] Mardia, K.V. (1970). Measures of multivariate skewness and kurtosis with applications. Biometrika 57, 519-530. [160] Mardia, K.V. (1972). Statistics of Directional Data. Academic Press, London. [161] Mardia, K.V. (1975). Assessment of multinormality and the robustness of Hotelling’s T 2 test. Applied Statistics 24, 163-171.

272

References

[162] Mardia, K.V., J.T. Kent, and J.M. Bibby (1979). Multivariate Analysis. Academic Press, New York. [163] M¨ arkel¨ ainen, T., K. Schmidt, and G.P.H. Styan (1981). On the existence and uniqueness of the maximum likelihood estimate of a vector-valued parameter in ﬁxed sample sizes. Annals of Statistics 9, 758-767. [164] Maronna, R.A. (1976). Robust M-estimators of multivariate location and scatter. Annals of Statistics 4, 51-67. [165] Marshall, A.W., and I. Olkin(1988). Families of multivariate distributions. Journal of the American Statistical Association 83, 834-841. [166] Martin, M.A. (1990). On bootstrap iteration for coverage correction in conﬁdence intervals. Journal of the American Statistical Association 85, 1105-1118. bibitemmat Mathew, T., and K. Nordstr¨ om (1997). Wishart and chi-square distributions associated with matrix quadratic forms. Journal of Multivariate Analysis 61, 129-143. [167] Mauchly, J.W. (1940). Signiﬁcance test for sphericity of a normal n-variate distribution. Annals of Mathematical Statistics 11, 204-209. [168] Muirhead, R.J. (1970). Asymptotic distributions of some multivariate tests. Annals of Mathematical Statistics 41, 1002-1010. [169] Muirhead, R.J. (1982). Aspects of Multivariate Statistical Theory. John Wiley & Sons, New York. [170] Muirhead, R.J., and Y. Chikuse (1975). Asymptotic expansions for the joint and marginal distributions of the latent roots of the covariance matrix. Annals of Statistics 3, 1011-1017. [171] Muirhead, R.J., and C.M. Waternaux (1980). Asymptotic distributions in canonical correlation analysis and other multivariate procedures for nonnormal populations. Biometrika 67, 31-43. [172] Nagao, H. (1973). On some test criteria for covariance matrix. Annals of Statistics 1, 700-709. [173] Nagao, H., and M.S. Srivastava (1992). On the distributions of some test criteria for a covariance matrix under local alternatives and bootstrap approximations. Journal of Multivariate Analysis 43, 331-350. [174] Naito, K. (1998). Approximation of the power of kurtosis test for multinormality. Journal of Multivariate Analysis 65, 166-180. [175] Nelsen, R.B. (1986). Properties of a one-parameter family of bivariate distributions with speciﬁed marginals. Communications in Statistics 15, 3277-3285. [176] Nguyen, T.T. (1997). A note on matrix variate normal distribution. Journal of Multivariate Analysis 60, 148-153. [177] Oakes, D. (1982). A model for association in bivariate survival data. Journal of the Royal Statistical Society B 44, 414-442. [178] Olkin, I., and J.W. Pratt (1958). Unbiased estimation of certain correlation coeﬃcients. Annals of Mathematical Statistics 29, 201-211. [179] Olkin, I., and S.N. Roy (1954). On multivariate distribution theory. Annals of Mathematical Statistics 25, 329-339.

References

273

[180] Perlman, M.D. (1969). One-sided testing problems in multivariate analysis. Annals of Mathematical Statistics 40, 549-567; Correction, Annals of Mathematical Statistics 42 (1971), 1777. [181] Perlman, M.D. (1980). Unbiasedness of the likelihood ratio tests for equality of several covariance matrices and equality of several multivariate normal populations. Annals of Statistics 8, 247-263. [182] Press, W.H. (1992). Numerical Recipies in C: The Art of Scientiﬁc Computing. 2nd ed. Cambridge University Press, New York. [183] Purkayastha, S., and M.S. Srivastava (1995). Asymptotic distributions of some test criteria for the covariance matrix in elliptical distributions under local alternatives. Journal of Multivariate Analysis 55, 165-186. [184] Rao, B.V., and B.K. Sinha (1988). A characterization of Dirichlet distributions. Journal of Multivariate Analysis 25, 25-30. [185] Rao, C.R. (1973). Linear Statistical Inference and Its Applications. 2nd ed. John Wiley & Sons, New York. [186] Redfern, D. (1996). Maple V Release 4. 3rd ed. Springer-Verlag, New York. [187] Reeds, J.A. (1985). Asymptotic number of roots of Cauchy likelihood equations. Annals of Statistics 13, 778-784. [188] Rocke, D.M. (1996). Robustness properties of S-estimators of multivariate location and shape in high dimension. Annals of Statistics 24, 1327-1345. [189] Romeu, J.L., and A. Ozturk (1993). A comparative study of goodness-ofﬁt tests for multivariate normality. Journal of Multivariate Analysis 46, 309-334. [190] Rousseeuw, P.J. (1985). Multivariate estimation with high breakdown point. In Mathematical Statistics and Applications. eds. W. Grossmann, G. Pﬂug, I. Vincze and W. Wertz. Vol. 8. Reidel, Dordrecht, 283-297. [191] Rousseeuw, P.J., and B.C. van Zomeren (1990). Unmasking multivariate outliers and leverage points. Journal of the American Statistical Association 85, 633-639. [192] Rousseeuw, P.J., and V.J. Yohai (1984). Robust regression by means of S-estimators. Robust and Nonlinear Time Series Analysis. Lecture Notes in Statistics Vol. 26. Springer, New York, 256-272. [193] Royston, J.F. (1982). An extension of Shapiro and Wilk’s W test for normality to large samples. Applied Statistics 31, 115-124. [194] Royston, J.F. (1983). Some techniques for assessing multivariate normality based on the Shapiro-Wilk W . Applied Statistics 32, 121-133. [195] Ruppert, D. (1992). Computing S estimators for regression and multivariate location/dispersion. Journal of Computational and Graphical Statistics 1, 253-270. [196] Saw, J.G. (1978). A family of distributions on the m-sphere and some hypothesis tests. Biometrika 65, 69-73. [197] Schoenberg, I.J. (1938). Metric spaces and completely monotone functions. Annals of Mathematics 39, 811-841.

274

References

[198] Sepanski, S.J. (1994). Asymptotics for multivariate t-statistic and Hotelling’s T 2 -statistic under inﬁnite second moments via bootstrapping. Journal of Multivariate Analysis 49, 41-54. [199] Serﬂing, R.J. (1980). Approximation Theorems of Mathematical Statistics. John Wiley & Sons, New York. [200] Shapiro, A., and M.W. Browne (1987). Analysis of covariance structures under elliptical distributions. Journal of the American Statistical Association 82, 1092-1097. [201] Shapiro, S.S., and M.B. Wilk (1965). An analysis of variance test for normality (complete samples). Biometrika 52, 591-611. [202] Shapiro, S.S., and R.S. Francia (1972). An approximate analysis of variance test for normality. Journal of the American Statistical Association 67, 215216. [203] Silvapulle, M.J. (1995). A Hotelling’s T 2 -type statistic for testing against one-sided hypotheses. Journal of Multivariate Analysis 55, 312-319. [204] Singh, K. (1981). On the asymptotic accuracy of Efron’s bootstrap. Annals of Statistics 9, 1187-1195. [205] Siotani, M., T. Hayakawa, and Y. Fujikoshi (1985). Modern Multivariate Statistical Analysis: A Graduate Course and Handbook. American Sciences Press, Columbus, Ohio. [206] Small, N.J.H. (1978). Plotting squared radii. Biometrika 65, 657-658. [207] Spivak, M. (1965). Calculus on Manifolds. Addison-Wesley, New York. [208] Srivastava, M.S. (1967). On ﬁxed-width conﬁdence bounds for regression parameters and the mean vector. Journal of the Royal Statistical Society B 29, 132-140. [209] Srivastava, M.S. (1984). Estimation of interclass correlations in familial data. Biometrika 71, 177-185. [210] Srivastava, M.S., and E.M. Carter (1980). Asymptotic expansions for hypergeometric functions. Multivariate analysis V. North-Holland, AmsterdamNew York, 337-347. [211] Srivastava, M.S., and E.M. Carter (1983). An Introduction to Applied Multivariate Statistics. North-Holland, New York. [212] Srivastava, M.S., and T.K. Hui (1987). On assessing multivariate normality based on Shapiro-Wilk W statistic. Statistics & Probability Letters 5, 15-18. [213] Srivastava, M.S., K.J. Keen, and R.S. Katapa (1988). Estimation of interclass and intraclass correlations in multivariate familial data. Biometrics 44, 141-150. [214] Srivastava, M.S., and C.G. Khatri (1979). An Introduction to Multivariate Statistics. North-Holland, New York. [215] Srivastava, M.S., C.G. Khatri, and E.M. Carter (1978). On monotonicity of the modiﬁed likelihood ratio test for the equality of two covariances. Journal of Multivariate Analysis 8, 262-267. [216] Srivastava, M.S., and D. von Rosen (1998). Outliers in multivariate regression models. Journal of Multivariate Analysis 65, 195-208.

References

275

[217] Srivastava, M.S., and W.K. Yau (1989). Saddlepoint method for obtaining tail probability of Wilks’ likelihood ratio test. Journal of Multivariate Analysis 31, 117-126. [218] Stadje, W. (1993). ML characterization of the multivariate normal distribution. Journal of Multivariate Analysis 46, 131-138. [219] Stahel, W.A. (1981). Robuste Sch¨ atzungen: Inﬁnitesimale Optimalit¨ at und Sch¨ atzungen von Kovarianzmatrizen. Ph. D. thesis, ETH Z¨ urich. [220] Statistical Sciences (1995). S-PLUS Guide to Statistical and Mathematical Analysis, Version 3.3. StatSci, a division of MathSoft, Inc., Seattle, Washington. [221] Stein, C. (1969). Multivariate Analysis I. Technical Report No. 42, Stanford University. [222] Steyn, H.S. (1993). On the problem of more than one kurtosis parameter in multivariate analysis. Journal of Multivariate Analysis 44, 1-22. [223] Stone, M. (1974). Cross-validatory choice and assessment of statistical predictions (with discussion). Journal of the Royal Statistical Society B, 36, 111-147. [224] Strang, G. (1980). Linear Algebra and its Applications. 2nd ed. Academic Press, New York. [225] Sugiura, N. (1973). Derivatives of the characteristic root of a symmetric or a hermitian matrix with two applications in mutivariate analysis. Communications in Statistics 1, 393-417. [226] Sugiura, N. (1976). Asymptotic expansions of the distributions of the latent roots and the latent vector of the Wishart and multivariate F matrices. Journal of Multivariate Analysis 6, 500-525. [227] Sugiura, N., and H. Nagao (1968). Unbiasedness of some test criteria for the equality of one or two covariance matrices. Annals of Mathematical Statistics 39, 1686-1692. [228] Sugiura, N., and H. Nagao (1971). Asymptotic expansion of the distribution of the generalized variance for noncentral Wishart matrix, when Ω = O(n). Annals of the Institute of Statistical Mathematics 23, 469-475. [229] Sutradhar, B.C. (1993). Score test for the covariance matrix of the elliptical t-distribution. Journal of Multivariate Analysis 46, 1-12. [230] Szablowski, P.J. (1998). Uniform distributions on spheres in ﬁnite dimensional Lα and their generalizations. Journal of Multivariate Analysis 64, 103-117. [231] Tang, D. (1994). Uniformly more powerful tests in a one-sided multivariate problem. Journal of the American Statistical Association 89, 1006-1011. [232] Tang, D. (1996). Erratum:“Uniformly more powerful tests in a one-sided multivariate problem” [Journal of the American Statistical Association 89 (1994), 1006-1011]. Journal of the American Statistical Association 91, 1757. [233] Tyler, D.E. (1982). Radial estimates and the test for sphericity. Biometrika 69, 429-436.

276

References

[234] Tyler, D.E. (1983a). Robustness and eﬃciency properties of scatter matrices. Biometrika 70, 411-420. [235] Tyler, D.E. (1983b). The asymptotic distribution of principal components roots under local alternatives to multiple roots. Annals of Statistics 11, 1232-1242. [236] Tyler, D.E. (1986). Breakdown properties of the M-estimators of multivariate scatter. Technical report, Department of Statistics, Rutgers University. [237] Uhlig, H. (1994). On singular Wishart and singular multivariate beta distributions. Annals of Statistics 22, 395-405. [238] van der Merwe, A., and J.V. Zidek (1980). Multivariate regression analysis and canonical variates. Canadian Journal of Statistics, 8, 27-39. [239] von Mises, R. (1918). Uber die “Ganzahligkeit” der Atomegewicht und verwante Fragen. Physikalische Zeitschrift 19, 490-500. [240] Wakaki, H., S. Eguchi, and Y. Fujikoshi (1990). A class of tests for a general covariance structure. Journal of Multivariate Analysis 32, 313-325. [241] Wang, Y., and M.P. McDermott (1998a). A conditional test for a nonnegative mean vector based on a Hotelling’s T 2 -type statistic. Journal of Multivariate Analysis 66, 64-70. [242] Wang, Y., and M.P. McDermott (1998b). Conditional likelihood ratio test for a nonnegative normal mean vector. Journal of the American Statistical Association 93, 380-386. [243] Waternaux, C.M. (1976). Asymptotic distributions of the sample roots for a non-normal population. Biometrika 63, 639-664. [244] Watson, G.S. (1983). Statistics on Spheres. The University of Arkansas Lecture Notes in Mathematical Sciences. John Wiley & Sons, New York. [245] Wielandt, H. (1967). Topics in the Analytic Theory of Matrices (Lecture notes prepared by R.R. Meyer.) University of Wisconsin Press, Madison. [246] Wilks, S.S. (1963). Multivariate statistical outliers. Sankhy¯ a: Series A 25, 407-426. [247] Wolfram, S. (1996). The Mathematica Book. 3rd ed. Wolfram Media, Inc. and Cambridge University Press, New York. [248] Wong, C.S., and D. Liu (1994). Moments for left elliptically contoured random matrices. Journal of Multivariate Analysis 49, 1-23. [249] Yamato, H. (1990). Uniformly minimum variance unbiased estimation for symmetric normal distributions. Journal of Multivariate Analysis 34, 227237.

Author Index

Aﬁﬁ, A.A., 170, 271 Ali, M.M., 66, 263 Anderson, G.A., 132, 263 Anderson, T.W., 132, 183, 263 Andrews, D.F., 94, 170, 263 Ash, R., 63, 263

Boulerice, B., 72, 265 Box, G.E.P., 94, 184, 195, 198, 201, 204, 265 Breiman, L., 154, 156–158, 265 Brown, P.J., 156, 265 Browne, M.W., 228, 265, 274

Baringhaus, L., 171, 263 Bartlett, M.S., 123, 199, 263 Bellman, R., 125, 264 Bentler, P.M., 49, 234, 237, 264, 267 Beran, R., 137, 200, 244, 245, 247, 249–251, 264 Berger, R.L., 86, 147, 265 Berk, R., 66, 264 Berkane, M., 237, 264 Bhat, B.R., 34, 264 Bibby, J.M., 272 Bickel, P.J., 244, 264 Billingsley, P., 20, 78, 264 Bilodeau, M., 156, 158, 181, 235, 264, 265 Blom, G., 186, 265 Boente, G., 224, 265

Carri`ere, J.F., 27, 265 Carter, E.M., 123, 137, 274 Casella, G., 86, 147, 265 Chang, T., 72, 270 Chattopadhyay, A.K., 137, 265 Chikuse, Y., 132, 137, 265, 272 Cl´eroux, R., 192, 266 Coelho, C.A., 184, 266 Cook, R.D., 26, 34, 266 Copas, J.B., 214, 266 Courant, R., 33, 266 Cox, D.R., 94, 170, 171, 265, 266 Cuadras, C.M., 26, 266 Datta, S., 104, 266 Davies, P.L., 222, 224, 266 Davis, A.W., 200, 205, 266 Davison, A.C., 243, 266

278

Author Index

Donoho, D.L., 83, 266 Ducharme, G.R., 72, 192, 265, 266 D¨ umbgen, L., 109, 266 Dykstra, R.L., 88, 266

Hui, T.K., 169, 170, 274 Hwang, J.T., 66, 264

Eaton, M.L., 51, 66, 88, 134, 137, 190, 249, 266, 267, 269 Efron, B., 65, 158, 243, 267 Eguchi, S., 98, 276 Erd´elyi, A., 115, 116, 119, 196, 202, 267 Escouﬁer, Y., 191, 267

James, A.T., 30, 33, 94, 137, 269 John, S., 120, 121, 252, 269 Johnson, M.E., 26, 34, 266 Johnson, N.L., 111, 170, 269 Johnson, R.A., 269 Jolliﬀe, I.T., 161, 269 Jordan, S.M., 138, 269

Fan, Y., 171, 267 Fang, K.T., 49, 208, 267 Feller, W., 254, 267 Fisher, R.A., 72, 267 Flury, B., viii, 267 Francia, R.S., 170, 274 Frank, M.J., 26, 267 Fraser, D.A.S., 86, 96, 104, 147, 245, 267 Freedman, D.A., 244, 264 Friedman, J.H., 154, 156–158, 265 Fujikoshi, Y., 72, 98, 103, 132, 154, 205, 267, 268, 270, 274, 276 Fujisawa, H., 103, 268

Kano, Y., 103, 209, 269 Kariya, T., 51, 154, 156, 158, 171, 209, 227, 265, 269–271 Katapa, R.S., 84, 274 Kato, T., 125, 248, 251, 270 Keen, K.J., 84, 274 Kelker, D., 207, 270 Kemp, A.W., 111, 269 Kendall, M., 257, 270 Kent, J.T., 214, 218, 270, 272 Kettenring, J.R., 170, 185, 194, 268 Khatri, C.G., 12, 13, 30, 31, 83, 119, 123, 137, 233, 236, 270, 274 Ko, D., 72, 270 Koehler, K.J., 26, 270 Kollo, T., 132, 259, 270 Koltchinskii, V.I., 49, 270 Konishi, S., 83, 84, 169, 270 Kotz, S., 111, 208, 267, 269, 271 Kres, H., 98, 271 Krishnaiah, P.R., 154, 270 Krishnamoorthy, K., 138, 269 Kshirsagar, A.M., 271 Kudo, A., 103, 271 Kuwana, Y., 209, 271

Genest, C., 26, 268 Ghosh, B.K., 101, 268 Giri, N.C., 268 Gnanadesikan, R., 94, 170, 185, 186, 194, 263, 268 Gr¨ ubel, R., 213, 268 Gunderson, B.K., 190, 268 Gupta, A.K., 41, 50, 74, 268 Guttman, I., 104, 267 Hall, P., 243, 268 Hayakawa, T., 274 Hendriks, H., 72, 268 Henze, N., 171, 269 Hinkley, D.V., 243, 266 Hsu, P.L., 190, 269 Huber, P.J., 222, 269

Iwashita, T., 103, 231, 269

Landsman, Z., 72, 268 Lee, Y.-S., 202, 271 Lehmann, E.L., 221, 271 Li, Haijun, 26, 271

Author Index

Li, Hong, 171, 270 Li, L., 49, 270 Liu, C., 221, 271 Liu, D., 74, 276 Looney, S.W., 170, 271 Lopuha¨ a, H.P., 222, 224–226, 271 MacDuﬀy, C.C., 30, 271 MacKay, R.J., 26, 268 Magnus, J.R., 76, 271 Magnus, W., 115, 116, 119, 196, 202, 267 Malkovich, J.F., 170, 271 Mardia, K.V., 72, 171, 271, 272 M¨ arkel¨ ainen, T., 214, 272 Maronna, R.A., 222, 224, 272 Marshall, A.W., 26, 272 Mathew, T., 92, 272 Mauchly, J.W., 118, 272 McDermott, M.P., 104, 276 Milasevic, T., 72, 266 Morris, C., 158 Muirhead, R.J., 94, 132, 190, 205, 228, 233, 268, 272 Mukhopadhyay, N., 104, 266 Nagao, H., 123, 128, 140, 234, 251, 272, 275 Naito, K., 171, 272 Nelsen, R.B., 27, 272 Neudecker, H., 76, 132, 270, 271 Ng, K.W., 208, 267 Nguyen, T.T., 74, 272 Nordstr¨ om, K., 92, 272 Oakes, D., 27, 272 Oberhettinger, F., 115, 116, 119, 196, 202, 267 Oden, K., 237, 264 Olkin, I., 26, 94, 115, 272 Ord, J.K., 257, 270 Ostrovskii, I., 208, 271 Ozturk, A., 171, 273

279

Perlman, M.D., 88, 103, 123, 142, 266, 273 Pillai, K.C.S., 137, 265 Ponnapalli, R., 66, 263 Pratt, J.W., 115, 272 Press, W.H., 184, 273 Purkayastha, S., 234, 273 Rao, B.V., 41, 273 Rao, C.R., 84, 270, 273 Redfern, D., 197, 273 Reeds, J.A., 214, 273 Richards, D. St. P., 41, 268 Rocke, D.M., 213, 268, 273 Romeu, J.L., 171, 273 Rousseeuw, P.J., 224, 262, 271, 273 Roy, S.N., 94, 272 Royston, J.F., 169, 170, 273 Ruppert, D., 226, 262, 273 Ruymgaart, F., 72, 268 Saw, J.G., 71, 273 Scarsini, M., 26, 271 Schmidt, K., 214, 272 Schoenberg, I.J., 53, 273 Sepanski, S.J., 103, 274 Serﬂing, R.J., 113, 120, 183, 274 Shaked, M., 26, 271 Shapiro, A., 228, 265, 274 Shapiro, S.S., 169, 170, 274 Silvapulle, M.J., 104, 274 Singh, K., 244, 274 Sinha, B.K., 41, 227, 269, 273 Siotani, M., 231, 269, 274 Small, N.J.H., 170, 171, 185, 266, 274 Song, D., 50, 268 Spivak, M., 28, 29, 33, 274 Srivastava, M.S., 12, 13, 30, 31, 84, 104, 119, 123, 137, 154, 169, 170, 172, 184, 233, 234, 236, 247, 249–251, 264, 265, 267, 270, 272–275 Stadje, W., 86, 275 Stahel, W.A., 83, 275

280

Author Index

Statistical Sciences, 226, 275 Stein, C., 88, 275 Steyn, H.S, 208, 275 Stone, M., 156, 275 Strang, G., 1, 275 Stuart, A., 257, 270 Styan, G.P.H., 214, 272 Sugiura, N., 123, 127, 128, 132, 133, 140, 275 Sutradhar, B.C., 236, 275 Symanowski, J.T., 26, 270 Szablowski, P.J., 50, 275 Tang, D., 103, 275 Terui, N., 171, 270 Tibshirani, R.J., 243 Tricomi, F.G., 115, 116, 119, 196, 202, 267 Tsay, R.S., 171, 270 Tyler, D.E., 132, 134, 137, 190, 210, 214, 215, 218, 224, 226, 228, 231, 249, 266, 267, 270, 275, 276 Uhlig, H., 94, 276 van der Merwe, A., 156, 158, 276 van Zomeren, B.C., 262, 273 Varga, T., 74, 268 von Mises, R., 72, 276 von Rosen, D., 154, 259, 270, 274 Wagner, T., 171, 269 Wakaki, H., 98, 276 Wang, Y., 104, 276 Warner, J.L., 94, 170, 263 Watamori, Y., 72, 268 Waternaux, C.M., 132, 190, 228, 272, 276 Watson, G.S., 72, 276 Wichern, D.W., 269 Wielandt, H., 134, 276 Wilk, M.B., 169, 170, 274 Wilks, S.S., 186, 276 Wolfram, S., 197, 276

Wong, C.S., 74, 276 Yamato, H., 48, 276 Yau, W.K., 184, 275 Yohai, V.J., 224, 273 Zhu, L.-X., 49, 267 Zidek, J.V., 156, 158, 276 Zirkler, B., 171, 269

Subject Index

a.e., 23 absolutely continuous, 23 adjoint, 5 adjusted LRT, 228 aﬃne equivariant, 209 Akaike’s criterion, 190 almost everywhere, 23 ancillary statistic, 118 angular gaussian distribution, 70 asymptotic distribution bootstrap, 243 canonical correlations, 189 correlation coeﬃcient, 81, 82, 230 eigenvalues of R, 168, 242 eigenvalues of S, 130, 242 eigenvalues of S−1 1 S2 , 133 elliptical MLE, 221 Hotelling-T 2 , 101 M estimate, 223 multiple correlation, 112, 230 normal MLE, 213 partial correlation, 117, 230 S estimate, 225 sample mean, 77, 78

sample variance, 80 with multiple eigenvalues, 136 Bartlett correction factor, 199 Bartlett decomposition, 11, 31 basis orthonormal, 3 Basu, 118 Bernoulli numbers, 201 polynomials, 196 trial, 17 beta function, 39 multivariate, 38 univariate, 39 blue multiple regression, 65 multivariate regression, 146 bootstrap correlation coeﬃcient, 248 eigenvalues, 137, 248 means with l1 -norm, 245 means with l∞ -norm, 246 Box-Cox transformation, 94

282

Subject Index

breakdown point, 224 C.E.T., 15, 16 Cr inequality, 33 canonical correlation, 175, 189 Fc distribution, 42 variables, 175 Caratheodory extension theorem, 15, 16 Cauchy-Schwarz inequality, 2 central limit theorem, 78 chain rule for derivatives, 29 change of variables, 29 characteristic function, 21 χ2m , 37 gamma(p, θ), 37 general normal, 45 inversion formula, 21, 24, 254 multivariate normal, 56 uniqueness, 21 Wishart, 90 chi-square, 37 commutation matrix, 75, 81 conditional distributions Dirichlet, 40 elliptical, 208 normal, 62 conditional mean formula, 20 conditional transformations, 31 conditional variance formula, 20 conditions D, 215 D1, 218 E, 228 H, 228 M, 215 M1-M4, 222 S1-S3, 224, 225 conﬁdence ellipsoid, 104, 105, 108 contour, 58 convergence theorems dominated, 18 monotone, 18 convolution, 253

copula, 26 Morgenstern, 34 correlation canonical, 175, 189 coeﬃcient, 67, 81, 82, 230, 248 interclass, 172 intraclass, 48, 64 matrix, 97, 230 multiple, 109 partial, 116 covariance, 20 Cram´er-Wold theorem, 21 cumulant, 80, 211, 256 d.f., 15 delta method, 79 density, 23 multivariate normal, 58 derivative, 28 chain rule, 29 determinant, 4 diagonalization, 6 diﬀerentiation with respect to matrix, 12 vector, 12 dimensionality, 190 Dirichlet, 38, 49 conditional, 40 marginal, 40 distribution absolutely continuous, 23 angular gaussian, 70 Bernoulli, 17 beta, 39 chi-square, 37 noncentral, 45 contaminated normal, 207 copula, 26 Dirichlet, 38, 49 conditional, 40 marginal, 40 discrete, 16 double exponential, 44 elliptical, 207 exchangeable, 47

Subject Index

exponential, 37 F , 42 noncentral, 45 Fc , 42 noncentral, 45, 52 Fisher-von Mises, 72 function, 15 gamma, 37 general normal, 45 inverted Wishart, 97 joint, 25 Kotz-type, 208 Langevin, 72 Laplace, 44 marginal, 25 multivariate Cauchy, 208 multivariate normal, 55 multivariate normal matrix, 74 density, 81 multivariate t, 207, 239 negative binomial, 110 nonsingular normal, 58 permutation invariant, 47 power exponential, 209 singular normal, 62 spherical, 48, 207 standard gamma, 36 standard normal, 44 symmetric, 43 t, 64 U (p; m, n), 150 unif(B n ), 49 unif(S n−1 ), 49 uniform, 24 unif(T n ), 39 Wishart, 87, 97 dominated convergence theorem, 18 double exponential, 44 Efron-Morris, 158 eigenvalue, 6 eigenvector, 6 elliptical distribution, 207 conditional, 208

283

consistency, 209 marginal, 208 empirical characteristic function, 171 empirical distribution, 244 equals-in-distribution, 16 equidistributed, 16 equivariant estimates, 209 estimate blue, 65, 146 M, 222 S, 224, 262 etr, 81 euclidian norm, 2 exchangeable, 47 expected value, 18 of a matrix, 19 of a vector, 19 of an indexed array, 19 exponential distribution scaled, 37 standard, 37 F distribution, 42 Fc distribution, 42 density, 42 familial data, 83, 172 FICYREG, 158 Fisher z-transform, 82, 114, 117, 248 Fisher’s information, 219 Fisher-von Mises distribution, 72 ﬂattening, 157 Gn , 29 gamma function, 36 generalized, 94 scaled, 37 standard, 36 Gauss-Markov multiple regression, 65 multivariate regression, 146 general linear group, 29 general linear hypothesis, 144

284

Subject Index

general normal, 45 generalized gamma function, 94 generalized variance, 93, 96 goodness-of-ﬁt, 72, 171 Gram-Schmidt, 3 group general linear, 29 orthogonal, 8, 48 permutation, 47 rotation, 48 triangular, 10 H¨ older’s inequality, 19 hermitian matrix, 6 transpose, 6 Hotelling-T 2 , 98 one-sided, 103 two-sample, 138 hypergeometric function, 115

|=

i.i.d., 28 iﬀ, 2 image space, 4 imputation, 221 indep ∼ , 38 independence mutual, 27 pairwise , 27 pairwise vs mutual, 34 test, 177, 192, 203 inequality between matrices, 8 Cr , 33 Cauchy-Schwarz, 2 H¨ older, 19 inner product matrix, 5 of complex vectors, 6 of matrices, 145 of vectors, 2 interclass correlation, 172 interpoint distance, 194 intraclass correlation, 48, 64

invariant tests, 102, 120, 122, 138, 140, 151, 178 inverse, 4 partitioned matrix, 11 inversion formulas, 21, 24, 254 inverted Wishart, 97 jacobian, 29, 75 joint distribution, 25 Kendall’s τ , 34 kernel, 4 Kronecker δ, 3 product, 74 Kummer’s formula, 115 kurtosis, 171, 212, 259 parameter, 212 L+ n , 10 Lp , 18 Langevin distribution, 72 Lawley-Hotelling trace test, 154 LBI test for sphericity, 121 least-squares estimate, 66 Lebesgue measure, 23 Leibniz notation, 18 length of a vector, 2 likelihood ratio test asymptotic, 118 linear estimation, 65 linear hypothesis, 144 log-likelihood, 213 LRT, 99 M estimate, 222 asymptotic, 223 Mahalanobis distance, 58, 170, 184, 206, 262 Mallow’s criterion, 190 MANOVA one-way, 159 marginal distribution, 25 matrices, 2 adjoint, 5

Subject Index

commutation, 75, 81 determinant, 4 diagonalization, 6, 7 eigenvalue, 6 eigenvector, 6 hermitian, 6 hermitian transpose, 6 idempotent, 9 image space, 4 inverse, 4 kernel, 4 Kronecker product, 74 nonsingular, 4 nullity, 4 orthogonal, 8, 48 positive deﬁnite, 8 positive semideﬁnite, 8 product, 2 rank, 4 singular value, 8 skew-symmetric, 35 square, 2 square root, 8 symmetric, 2 trace, 2 transpose, 2 triangular, 7 triangular decomposition, 11 unitary, 6 matrix diﬀerentiation, 12 Maxwell-Hershell theorem, 51, 227 mean, 19 minimum volume ellipsoid, 224 missing data, 221 mixture distribution, 21, 46, 53, 56, 111, 207, 209, 238 MLE (Σ, µ), 86, 96 multivariate regression, 147 modulus of a vector, 2 monotone convergence theorem, 18 multiple correlation, 109 asymptotic, 112, 230 invariance, 140 moments, 140

asymptotic, 115 MVUE, 115 multiple regression, 65 multivariate copula, 26 ﬂattening, 157 prediction, 156 regression, 144 multivariate distribution beta, 38 Cauchy, 208 contaminated normal, 207 cumulant, 80, 211, 256 Kotz-type, 208 normal, 55 contour, 58 density, 58 normal matrix, 74 conditional, 82 density, 81 power exponential, 209 t, 207, 239 with given marginals, 26 mutual independence, 27 MVUE R2 , 115 (Σ, µ), 86 Nt process, 37 negative binomial, 110 noncentral chi-square, 45 F , 45 Fc , 45 density, 52 nonsingular matrix, 4 normal, 58 normal general, 45 multivariate, 55 nonsingular, 58 singular, 62 standard, 44 nullity, 4

285

286

Subject Index

one-way classiﬁcation, 158 orthogonal complement, 3 group, 8, 48 matrix, 8, 48 projection, 9, 66 vectors, 2 orthogonal invariance, 48 outlier, 262

|=

Pn , 8 p.d.f., 23 p.f., 16 pairwise independence , 27 partial correlation, 116 asymptotic, 117, 230 permutation group, 47 invariance, 47 perturbation method, 125 Pillai trace test, 154 Poisson process Nt , 37 polar coordinates, 32, 50, 54 positive deﬁnite, 8 semideﬁnite, 8 power transformations, 94 prediction, 156 prediction risk, 156, 157 principal components deﬁnition, 162 sample, 165, 169 probability density function, 23 function, 16 product-moment, 19 projection mutually orthogonal, 10 orthogonal, 9, 66 proportionality test, 139 PSn , 8 Q-Q plot of squared radii, 186 quadratic forms, 66, 67

Radon-Nikodym theorem, 23 rank, 4 Rayleigh’s quotient, 13 rectangles, 15 reﬂection symmetry, 43 regression multiple, 65 multivariate, 144 relative eﬃciency, 236, 262 robust estimates, 222 M type, 222 S type, 224 robustness Hotelling-T 2 , 101, 226 tests on scale matrix, 227 rotation group, 48 rotationally invariant matrix, 210 vector, 49 Roy largest eigenvalue, 154 S estimate, 224, 262 asymptotic, 225 S n−1 , 21, 33 Sn , 47 Sp , 134 sample matrix, 75 sample mean asymptotic, 77, 78 sample variance, 77 asymptotic, 80, 213 scaled distribution exponential, 37 gamma, 37 scaled residuals, 171, 184 score function, 219 Shapiro-Wilk test, 169 simultaneous conﬁdence intervals asymptotic, 109, 139 Bonferroni, 107 eigenvalues by bootstrap, 248 for φ(Σ), 109 linear hypotheses, 104 means by bootstrap, 246 nonlinear hypotheses, 107

Subject Index

robust, 227 Roy-Bose, 106, 139 Scheﬀ´e, 106, 139 singular normal, 62 value, 8 skew-symmetric matrix, 35 skewness, 171, 259 Slutsky theorem, 78 span, 4 SPE prediction risk, 157 Spearman’s ρ, 34 spectral decomposition, 8 SPER prediction risk, 156 spherical distribution, 48, 207 characteristic function, 52 density, 52 density of radius, 223 density of squared radius, 54 square root matrix S1/2 , 8 standard distribution exponential, 37 gamma, 36 normal, 44 statistically independent, 27 Sugiura’s lemma, 127 SVD, 8 symmetric distribution, 43 matrix, 2 t distribution, 64 T 2 of Hotelling, 98 T n , 39 Tn , 38 test equality of means, 159 equality of means and variances, 141, 205 equality of variances, 121, 204 for a given mean, 99 for a given mean vector and variance, 236 for a given variance, 139, 205, 233, 240

Hotelling two-sample, 138 Hotelling-T 2 , 98 independence, 177, 192, 203 Lawley-Hotelling, 154 linear hypothesis, 148, 201 multiple correlation, 110, 241 multivariate normality, 169 Pillai, 154 proportionality, 139 Roy, 154 sphericity, 117, 200 symmetry, 138 total variance, 162 triangular decomposition, 11 group, 10 matrix, 7 U (p; m, n), 150, 261 asymptotic, 184, 201 characterizations, 182 duality, 182 moments, 190 U+ n , 10 UMPI test for multiple correlation, 112 Hotelling-T 2 , 103 unif(B n ), 49 unif(S n−1 ), 49 uniform distribution, 24 unif(T n ), 39 union-intersection test, 160 unit sphere, 21, 33 unitary matrix, 6 uvec operator, 247 variance, 19 generalized, 93, 96 of a matrix, 74 sample, 77 total, 162 vec operator, 73 vector diﬀerentiation, 12 vectors column, 1

287

288

Subject Index

inner product, 2 length, 2 modulus, 2 orthogonal, 2 orthonormal, 3 outer product, 3 row, 1 volume, 23 w.p.1, 52 waiting time process Tn , 38 Wielandt’s inequality, 134

Wishart, 87 characteristic function, 90 density, 93, 97 linear transformation, 88 marginals, 90, 92 moments and cumulants, 259 noninteger degree of freedom, 94 nonsingular, 87 sums, 91

To Rebecca and Deena. D. Brenner

Preface

Our object in writing this book is to present the main results of the modern theory of multivariate statistics to an audience of advanced students who would appreciate a concise and mathematically rigorous treatment of that material. It is intended for use as a textbook by students taking a ﬁrst graduate course in the subject, as well as for the general reference of interested research workers who will ﬁnd, in a readable form, developments from recently published work on certain broad topics not otherwise easily accessible, as, for instance, robust inference (using adjusted likelihood ratio tests) and the use of the bootstrap in a multivariate setting. The references contains over 150 entries post-1982. The main development of the text is supplemented by over 135 problems, most of which are original with the authors. A minimum background expected of the reader would include at least two courses in mathematical statistics, and certainly some exposure to the calculus of several variables together with the descriptive geometry of linear algebra. Our book is, nevertheless, in most respects entirely self-contained, although a deﬁnite need for genuine ﬂuency in general mathematics should not be underestimated. The pace is brisk and demanding, requiring an intense level of active participation in every discussion. The emphasis is on rigorous proof and derivation. The interested reader would proﬁt greatly, of course, from previous exposure to a wide variety of statistically motivating material as well, and a solid background in statistics at the undergraduate level would obviously contribute enormously to a general sense of familiarity and provide some extra degree of comfort in dealing with the kinds of challenges and diﬃculties to be faced in the relatively advanced work

viii

Preface

of the sort with which our book deals. In this connection, a speciﬁc introduction oﬀering comprehensive overviews of the fundamental multivariate structures and techniques would be well advised. The textbook A First Course in Multivariate Statistics by Flury (1997), published by SpringerVerlag, provides such background insight and general description without getting much involved in the “nasty” details of analysis and construction. This would constitute an excellent supplementary source. Our book is in most ways thoroughly orthodox, but in several ways novel and unique. In Chapter 1 we oﬀer a brief account of the prerequisite linear algebra as it will be applied in the subsequent development. Some of the treatment is peculiar to the usages of multivariate statistics and to this extent may seem unfamiliar. Chapter 2 presents in review, the requisite concepts, structures, and devices from probability theory that will be used in the sequel. The approach taken in the following chapters rests heavily on the assumption that this basic material is well understood, particularly that which deals with equality-in-distribution and the Cram´er-Wold theorem, to be used with unprecedented vigor in the derivation of the main distributional results in Chapters 4 through 8. In this way, our approach to multivariate theory is much more structural and directly algebraic than is perhaps traditional, tied in this fashion much more immediately to the way in which the various distributions arise either in nature or may be generated in simulation. We hope that readers will ﬁnd the approach refreshing, and perhaps even a bit liberating, particularly those saturated in a lifetime of matrix derivatives and jacobians. As a textbook, the ﬁrst eight chapters should provide a more than adequate amount of material for coverage in one semester (13 weeks). These eight chapters, proceeding from a thorough discussion of the normal distribution and multivariate sampling in general, deal in random matrices, Wishart’s distribution, and Hotelling’s T 2 , to culminate in the standard theory of estimation and the testing of means and variances. The remaining six chapters treat of more specialized topics than it might perhaps be wise to attempt in a simple introduction, but would easily be accessible to those already versed in the basics. With such an audience in mind, we have included detailed chapters on multivariate regression, principal components, and canonical correlations, each of which should be of interest to anyone pursuing further study. The last three chapters, dealing, in turn, with asymptotic expansion, robustness, and the bootstrap, discuss concepts that are of current interest for active research and take the reader (gently) into territory not altogether perfectly charted. This should serve to draw one (gracefully) into the literature. The authors would like to express their most heartfelt thanks to everyone who has helped with feedback, criticism, comment, and discussion in the preparation of this manuscript. The ﬁrst author would like especially to convey his deepest respect and gratitude to his teachers, Muni Srivastava

Preface

ix

of the University of Toronto and Takeaki Kariya of Hitotsubashi University, who gave their unstinting support and encouragement during and after his graduate studies. The second author is very grateful for many discussions with Philip McDunnough of the University of Toronto. We are indebted to Nariaki Sugiura for his kind help concerning the application of Sugiura’s Lemma and to Rudy Beran for insightful comments, which helped to improve the presentation. Eric Marchand pointed out some errors in the literature about the asymptotic moments in Section 8.4.1. We would like to thank the graduate students at McGill University and Universit´e de Montr´eal, Gulhan Alpargu, Diego Clonda, Isabelle Marchand, Philippe St-Jean, Gueye N’deye Rokhaya, Thomas Tolnai and Hassan Younes, who helped improve the presentation by their careful reading and problem solving. Special thanks go to Pierre Duchesne who, as part of his Master Memoir, wrote and tested the S-Plus function for the calculation of the robust S estimate in Appendix C.

M. Bilodeau D. Brenner

Contents

Preface List of Tables List of Figures

vii xv xvii

1 Linear algebra 1.1 Introduction . . . . . . . . . . . . . . . 1.2 Vectors and matrices . . . . . . . . . . 1.3 Image space and kernel . . . . . . . . . 1.4 Nonsingular matrices and determinants 1.5 Eigenvalues and eigenvectors . . . . . . 1.6 Orthogonal projections . . . . . . . . . 1.7 Matrix decompositions . . . . . . . . . 1.8 Problems . . . . . . . . . . . . . . . . . 2 Random vectors 2.1 Introduction . . . . . . . . . . . . . 2.2 Distribution functions . . . . . . . . 2.3 Equals-in-distribution . . . . . . . . 2.4 Discrete distributions . . . . . . . . 2.5 Expected values . . . . . . . . . . . 2.6 Mean and variance . . . . . . . . . 2.7 Characteristic functions . . . . . . . 2.8 Absolutely continuous distributions 2.9 Uniform distributions . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . .

. . . . . . . . .

. . . . . . . .

. . . . . . . . .

. . . . . . . .

. . . . . . . . .

. . . . . . . .

. . . . . . . . .

. . . . . . . .

. . . . . . . . .

. . . . . . . .

. . . . . . . . .

. . . . . . . .

. . . . . . . . .

. . . . . . . .

. . . . . . . . .

. . . . . . . .

. . . . . . . . .

. . . . . . . .

1 1 1 3 4 5 9 10 11

. . . . . . . . .

14 14 14 16 16 17 18 21 22 24

xii

Contents

2.10 2.11 2.12 2.13 2.14

Joints and marginals Independence . . . . Change of variables . Jacobians . . . . . . . Problems . . . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

25 27 28 30 33

3 Gamma, Dirichlet, and F distributions 3.1 Introduction . . . . . . . . . . . . . . . 3.2 Gamma distributions . . . . . . . . . . 3.3 Dirichlet distributions . . . . . . . . . . 3.4 F distributions . . . . . . . . . . . . . 3.5 Problems . . . . . . . . . . . . . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

36 36 36 38 42 42

4 Invariance 4.1 Introduction . . . . . . . . . . . . . . . . . . 4.2 Reﬂection symmetry . . . . . . . . . . . . . 4.3 Univariate normal and related distributions 4.4 Permutation invariance . . . . . . . . . . . . 4.5 Orthogonal invariance . . . . . . . . . . . . . 4.6 Problems . . . . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

43 43 43 44 47 48 52

5 Multivariate normal 5.1 Introduction . . . . . . . . . . . . . . . 5.2 Deﬁnition and elementary properties . 5.3 Nonsingular normal . . . . . . . . . . . 5.4 Singular normal . . . . . . . . . . . . . 5.5 Conditional normal . . . . . . . . . . . 5.6 Elementary applications . . . . . . . . 5.6.1 Sampling the univariate normal 5.6.2 Linear estimation . . . . . . . . 5.6.3 Simple correlation . . . . . . . . 5.7 Problems . . . . . . . . . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

55 55 55 58 62 62 64 64 65 67 69

6 Multivariate sampling 6.1 Introduction . . . . . . . . . . . . . . . . . 6.2 Random matrices and multivariate sample 6.3 Asymptotic distributions . . . . . . . . . . 6.4 Problems . . . . . . . . . . . . . . . . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

73 73 73 78 81

7 Wishart distributions 7.1 Introduction . . . . . . . . . . . . . ¯ and S . . . . 7.2 Joint distribution of x 7.3 Properties of Wishart distributions 7.4 Box-Cox transformations . . . . . . 7.5 Problems . . . . . . . . . . . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

85 85 85 87 94 96

. . . . .

. . . . .

. . . . . . . . . .

. . . . .

. . . . .

Contents

xiii

8 Tests on mean and variance 8.1 Introduction . . . . . . . . . . . . . . . . . . 8.2 Hotelling-T 2 . . . . . . . . . . . . . . . . . . 8.3 Simultaneous conﬁdence intervals on means 8.3.1 Linear hypotheses . . . . . . . . . . . 8.3.2 Nonlinear hypotheses . . . . . . . . . 8.4 Multiple correlation . . . . . . . . . . . . . . 8.4.1 Asymptotic moments . . . . . . . . . 8.5 Partial correlation . . . . . . . . . . . . . . . 8.6 Test of sphericity . . . . . . . . . . . . . . . 8.7 Test of equality of variances . . . . . . . . . 8.8 Asymptotic distributions of eigenvalues . . . 8.8.1 The one-sample problem . . . . . . . 8.8.2 The two-sample problem . . . . . . . 8.8.3 The case of multiple eigenvalues . . . 8.9 Problems . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

98 98 98 104 104 107 109 114 116 117 121 124 124 132 133 137

9 Multivariate regression 9.1 Introduction . . . . . . . . . . . . . . . 9.2 Estimation . . . . . . . . . . . . . . . . 9.3 The general linear hypothesis . . . . . 9.3.1 Canonical form . . . . . . . . . 9.3.2 LRT for the canonical problem 9.3.3 Invariant tests . . . . . . . . . . 9.4 Random design matrix X . . . . . . . . 9.5 Predictions . . . . . . . . . . . . . . . . 9.6 One-way classiﬁcation . . . . . . . . . . 9.7 Problems . . . . . . . . . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

144 144 145 148 148 150 151 154 156 158 159

10 Principal components 10.1 Introduction . . . . . . . . . . . . . . 10.2 Deﬁnition and basic properties . . . . 10.3 Best approximating subspace . . . . . 10.4 Sample principal components from S 10.5 Sample principal components from R 10.6 A test for multivariate normality . . 10.7 Problems . . . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

161 161 162 163 164 166 169 172

11 Canonical correlations 11.1 Introduction . . . . . . . . . . . . 11.2 Deﬁnition and basic properties . . 11.3 Tests of independence . . . . . . . 11.4 Properties of U distributions . . . 11.4.1 Q-Q plot of squared radii .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

174 174 175 177 181 184

. . . . .

. . . . .

xiv

Contents

11.5 11.6

Asymptotic distributions . . . . . . . . . . . . . . . . . . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . .

12 Asymptotic expansions 12.1 Introduction . . . . 12.2 General expansions 12.3 Examples . . . . . . 12.4 Problem . . . . . .

189 190

. . . .

. . . .

. . . .

. . . .

. . . .

195 195 195 200 205

13 Robustness 13.1 Introduction . . . . . . . . . . . . . . . . . . . . . 13.2 Elliptical distributions . . . . . . . . . . . . . . . 13.3 Maximum likelihood estimates . . . . . . . . . . . 13.3.1 Normal MLE . . . . . . . . . . . . . . . . 13.3.2 Elliptical MLE . . . . . . . . . . . . . . . 13.4 Robust estimates . . . . . . . . . . . . . . . . . . 13.4.1 M estimate . . . . . . . . . . . . . . . . . . 13.4.2 S estimate . . . . . . . . . . . . . . . . . . 13.4.3 Robust Hotelling-T 2 . . . . . . . . . . . . 13.5 Robust tests on scale matrices . . . . . . . . . . . 13.5.1 Adjusted likelihood ratio tests . . . . . . . 13.5.2 Weighted Nagao’s test for a given variance 13.5.3 Relative eﬃciency of adjusted LRT . . . . 13.6 Problems . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

206 206 207 213 213 213 222 222 224 226 227 228 233 236 238

14 Bootstrap conﬁdence regions and tests 14.1 Conﬁdence regions and tests for the mean 14.2 Conﬁdence regions for the variance . . . . 14.3 Tests on the variance . . . . . . . . . . . . 14.4 Problem . . . . . . . . . . . . . . . . . . .

. . . .

. . . .

. . . .

. . . .

243 243 246 249 252

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

A Inversion formulas

253

B Multivariate cumulants B.1 Deﬁnition and properties . . . . . . . . . . . . . . . . . . B.2 Application to asymptotic distributions . . . . . . . . . . B.3 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . .

256 256 259 259

C S-plus functions

261

References Author Index Subject Index

263 277 281

List of Tables

12.1 Polynomials δs and Bernoulli numbers Bs for asymptotic expansions. . . . . . . . . . . . . . . . . . . . . . . . . . . 12.2 Asymptotic expansions for U (2; 12, n) distributions. . . .

201 203

13.1 Asymptotic eﬃciency of S estimate of scatter at the normal distribution. . . . . . . . . . . . . . . . . . . . . . . . . . . 225 13.2 Asymptotic signiﬁcance level of unadjusted LRT for α = 5%. 238

This page intentionally left blank

List of Figures

2.1 3.1 5.1 5.2

5.3 8.1 8.2

Bivariate Frank density with standard normal marginals and a correlation of 0.7. . . . . . . . . . . . . . . . . . . . . . .

27

Bivariate Dirichlet density for values of the parameters p1 = p2 = 1 and p3 = 2. . . . . . . . . . . . . . . . . . . . . . .

41

Bivariate normal density for values of the parameters µ1 = µ2 = 0, σ1 = σ2 = 1, and ρ = 0.7. . . . . . . . . . . . . . . Contours of the bivariate normal density for values of the parameters µ1 = µ2 = 0, σ1 = σ2 = 1, and ρ = 0.7. Values of c = 1, 2, 3 were taken. . . . . . . . . . . . . . . . . . . A contour of a trivariate normal density. . . . . . . . . . . Power function of Hotelling-T 2 when p = 3 and n = 40 at a level of signiﬁcance α = 0.05. . . . . . . . . . . . . . . . . Power function of the likelihood ratio test for H0 : R = 0 when p = 3, and n = 20 at a level of signiﬁcance α = 0.05.

11.1 Q-Q plot for a sample of size n = 50 from a trivariate normal, N3 (0, I), distribution. . . . . . . . . . . . . . . . . . . . . 11.2 Q-Q plot for a sample of size n = 50 from a trivariate t on 1 degree of freedom, t3,1 (0, I) ≡ Cauchy3 (0, I), distribution.

59

60 61 101 113 187 188

This page intentionally left blank

1 Linear algebra

1.1 Introduction Multivariate analysis deals with issues related to the observations of many, usually correlated, variables on units of a selected random sample. These units can be of any nature such as persons, cars, cities, etc. The observations are gathered as vectors; for each selected unit corresponds a vector of observed variables. An understanding of vectors, matrices, and, more generally, linear algebra is thus fundamental to the study of multivariate analysis. Chapter 1 represents our selection of several important results on linear algebra. They will facilitate a great many of the concepts in multivariate analysis. A useful reference for linear algebra is Strang (1980).

1.2 Vectors and matrices To express the dependence of the x ∈ Rn on its coordinates, we may write any of x1 .. x = (xi , i = 1, . . . , n) = (xi ) = . . xn In this manner, x is envisaged as a “column” vector. The transpose of x is the “row” vector x ∈ Rn

x = (xi ) = (x1 , . . . , xn ) .

2

1. Linear algebra

An m × n matrix A ∈ Rm n may also be denoted in various a11 .. A = (aij , i = 1, . . . , m, j = 1, . . . , n) = (aij ) = . am1

ways:

· · · a1n .. .. . . . · · · amn

The transpose of A is the n × m matrix A ∈ Rnm : a11 · · · am1 .. . .. . A = (aij ) = (aji ) = .. . . a1n · · · amn A square matrix S ∈ Rnn satisfying S = S is termed symmetric. The product of the m × n matrix A by the n × p matrix B is the m × p matrix C = AB for which n cij = aik bkj . k=1

n

is tr A = i=1 aii and one veriﬁes that for A ∈ Rm The trace of A ∈ n n and B ∈ Rm , tr AB = tr BA. In particular, row vectors and column vectors are themselves matrices, so that for x, y ∈ Rn , we have the scalar result Rnn

x y =

n

xi yi = y x.

i=1

This provides the standard inner product, x, y = x y, in Rn with the associated “euclidian norm” (length or modulus) n 1/2 x2i . |x| = x, x1/2 = i=1

The Cauchy-Schwarz inequality is now proved. Proposition 1.1 |x, y| ≤ |x| |y|, ∀x, y ∈ Rn , with equality if and only if (iﬀ ) x = λy for some λ ∈ R. Proof. If x = λy, for some λ ∈ R, the equality clearly holds. If not, 0 < |x − λy|2 = |x|2 − 2λx, y + λ2 |y|2 , ∀λ ∈ R; thus, the discriminant of 2 the quadratic polynomial must satisfy 4x, y2 − 4|x|2 |y|2 < 0. The cosine of the angle θ between the vectors x = 0 and y = 0 is just cos(θ) =

x, y . |x| |y|

Orthogonality is another associated concept. Two vectors x and y in Rn will be said to be orthogonal iﬀ x, y = 0. In contrast, the outer (or tensor) product of x and y is an n × n matrix xy = (xi yj )

1.3. Image space and kernel

3

and this product is not commutative. The concept of orthonormal basis plays a major role in linear algebra. A set {vi } of vectors in Rn is orthonormal if 0, i = j vi vj = δij = 1, i = j. The symbol δij is referred to as the Kronecker delta. The Gram-Schmidt orthogonalization method gives a construction of an orthonormal basis from an arbitrary basis. Proposition 1.2 Let {v1 , . . . , vn } be a basis of Rn . Deﬁne u1 ui

= v1 /|v1 |, = wi /|wi |,

i−1 where wi = vi − j=1 (vi uj )uj , i = 2, . . . , n. Then, {u1 , . . . , un } is an orthonormal basis.

1.3 Image space and kernel Now, a matrix may equally well be recognized as a function either of its column vectors or its row vectors: g1 .. A = (a1 , . . . , an ) = . gm

for aj ∈ Rm , j = 1, . . . , n or gi ∈ Rn , i = 1, . . . , m. If we then write B = (b1 , . . . , bp ) with bj ∈ Rn , j = 1, . . . , p, we ﬁnd that AB = (Ab1 , . . . , Abp ) = (gi bj ) . In particular, for x ∈ Rn , we have expressly that x1 n .. = xi ai Ax = (a1 , . . . , an ) . i=1 xn or g1 x g1 .. .. Ax = . x = . . gm

(1.1)

(1.2)

gm x

The orthogonal complement of a subspace V ⊂ Rn is, by deﬁnition, the subspace V ⊥ = {y ∈ Rn : y ⊥ x, ∀x ∈ V}.

4

1. Linear algebra

Expression (1.1) identiﬁes the image space of A, Im A = {Ax : x ∈ Rn }, with the linear span of its column vectors and the expression (1.2) reveals the kernel, ker A = {x ∈ Rn : Ax = 0}, to be the orthogonal complement of the row space, equivalently ker A = (Im A )⊥ . The dimension of the subspace Im A is called the rank of A and satisﬁes rank A = rank A , whereas the dimension of ker A is called the nullity of A. They are related through the following simple relation: Proposition 1.3 For any A ∈ Rm n , n = nullity A + rank A. Proof. Let {v1 , . . . , vν } be a basis of ker A and extend it to a basis {v1 , . . . , vν , vν+1 , . . . , vn } of Rn . One can easily check {Avν+1 , . . . , Avn } is a basis of Im A. Thus, n = nullity A + rank A. 2

1.4 Nonsingular matrices and determinants We recall some basic facts about nonsingular (one-to-one) linear transformations and determinants. By writing A ∈ Rnn in terms of its column vectors A = (a1 , . . . , an ) with aj ∈ Rn , j = 1, . . . , n, it is clear that A is one-to-one ⇐⇒ a1 , . . . , an is a basis ⇐⇒ ker A = {0} and also from the simple relation n = nullity A + rank A, A is one-to-one ⇐⇒ A is one-to-one and onto. These are all equivalent ways of saying A has an inverse or that A is nonsingular. Denote by σ(1), . . . , σ(n) a permutation of 1, . . . , n and by n(σ) its parity. Let Sn be the group of all the n! permutations. The determinant is, by deﬁnition, the unique function det : Rnn → R, denoted |A| = det(A), that is, (i) multilinear: linear in each of a1 , . . . , an separately (ii) alternating: aσ(1) , . . . , aσ(n) = (−1)n(σ) |(a1 , . . . , an )| (iii) normed: |I| = 1. This produces the formula |A| =

(−1)n(σ) a1σ(1) · · · anσ(n)

σ∈Sn

by which one veriﬁes |AB| = |A| |B| and |A | = |A| .

1.5. Eigenvalues and eigenvectors

5

Determinants are usually calculated with a Laplace development along any given row or column. To this end, let A = (aij ) ∈ Rnn . Now, deﬁne the minor |m(i, j)| of aij as the determinant of the (n−1)×(n−1) “submatrix” obtained by deleting the ith row and the jth column of A and the cofactor |m(i, j)|. Then, the Laplace development of |A| of aij as c(i, j) = (−1)i+j n along the ith row is |A| = j=1 aij ·c(i, j) and a similar development along n the jth column is |A| = i=1 aij · c(i, j). By deﬁning adj(A) = (c(j, i)), the transpose of the matrix of cofactors, to be the adjoint of A, it can be shown A−1 = |A|−1 adj(A). But then Proposition 1.4 A is one-to-one ⇐⇒ |A| = 0. Proof. A is one-to-one means it has an inverse B, |A| |B| = 1 so n |A| = 0. But, conversely, if |A| = 0, suppose Ax = j=1 xj aj = 0, then substituting Ax for the ith column of A n a1 , . . . , xj aj , . . . , an = xi |A| = 0, i = 1, . . . , n j=1 so that x = 0, whereby A is one-to-one.

2

In general, for aj ∈ Rn , j = 1, . . . , k, write A = (a1 , . . . , ak ) and form the “inner product” matrix A A = (ai aj ) ∈ Rkk . We ﬁnd Proposition 1.5 For A ∈ Rnk , 1. ker A = ker A A 2. rank A = rank A A 3. a1 , . . . , ak are linearly independent in Rn ⇐⇒ |A A| = 0. Proof. If x ∈ ker A, then Ax = 0 =⇒ A Ax = 0, and, conversely, if x ∈ ker A A, then A Ax = 0 =⇒ x A Ax = 0 = |Ax|2 =⇒ Ax = 0. The second part follows from the relation k = nullity A + rank A and the 2 third part is immediate as ker A = {0} iﬀ ker A A = {0}.

1.5 Eigenvalues and eigenvectors We now brieﬂy state some concepts related to eigenvalues and eigenvectors. Consider, ﬁrst, the complex vector space Cn . The conjuguate of v = x+iy ∈ C, x, y ∈ R, is v = x − iy. The concepts deﬁned earlier are anologous in this case. The Hermitian transpose of a column vector v = (vi ) ∈ Cn is the row vector vH = (vi ) . The inner product on Cn can then be written v1 , v2 =

6

1. Linear algebra

v1H v2 for any v1 , v2 ∈ Cn . The Hermitian transpose of A = (aij ) ∈ Cm n is AH = (aji ) ∈ Cnm and satisﬁes for B ∈ Cnp , (AB)H = BH AH . The matrix A ∈ Cnn is termed Hermitian iﬀ A = AH . We now deﬁne what is meant by an eigenvalue. A scalar λ ∈ C is an eigenvalue of A ∈ Cnn if there exists a vector v = 0 in Cn such that Av = λv. Equivalently, λ ∈ C is an eigenvalue of A iﬀ |A − λI| = 0, which is a polynomial equation of degree n. Hence, there are n complex eigenvalues, some of which may be real, with possibly some repetitions (multiplicity). The vector v is then termed the eigenvector of A corresponding to the eigenvalue λ. Note that if v is an eigenvector, so is αv, ∀α = 0 in C, and, in particular, v/|v| is a normalized eigenvector. Now, before deﬁning what is meant by A is “diagonalizable” we deﬁne a matrix U ∈ Cnn to be unitary iﬀ UH U = I = UUH . This means that the columns (or rows) of U comprise an orthonormal basis of Cn . We note immediately that if {u1 , . . . , un } is an orthonormal basis of eigenvectors corresponding to eigenvalues {λ1 , . . . , λn }, then A can be diagonalized by the unitary matrix U = (u1 , . . . , un ); i.e., we can write UH AU = UH (Au1 , . . . , Aun ) = UH (λ1 u1 , . . . , λn un ) = diag(λ), where λ = (λ1 , . . . , λn ) . Another simple related property: If there exists a unitary matrix U = (u1 , . . . , un ) such that UH AU = diag(λ), then ui is an eigenvector corresponding to λi . To verify this, note that Aui = U diag(λ)UH ui = U diag(λ)ei = Uλi ei = λi ui . Two fundamental propositions concerning Hermitian matrices are the following. Proposition 1.6 If A ∈ Cnn is Hermitian, then all its eigenvalues are real. Proof. vH Av = (vH Av)H = vH AH v = vH Av, which means that vH Av is real for any v ∈ Cn . Now, if Av = λv for some v = 0 in Cn , then vH Av = λvH v = λ|v|2 . But since vH Av and |v|2 are real, so is λ. 2 Proposition 1.7 If A ∈ Cnn is Hermitian and v1 and v2 are eigenvectors corresponding to eigenvalues λ1 and λ2 , respectively, where λ1 = λ2 , then v1 ⊥ v2 . Proof. Since A is Hermitian, A = AH and λi , i = 1, 2, are real. Then, Av1 = λ1 v1 Av2 = λ2 v2

=⇒ =⇒

v1H AH = v1H A = λ1 v1H =⇒ v1H Av2 = λ1 v1H v2 , v1H Av2 = λ2 v1H v2 .

Subtracting the last two expressions, (λ1 −λ2 )v1H v2 = 0 and, thus, v1H v2 = 0. 2

1.5. Eigenvalues and eigenvectors

7

Proposition 1.7 immediately shows that if all the eigenvalues of A, Hermitian, are distinct, then there exists an orthonormal basis of eigenvectors whereby A is diagonalizable. Toward proving this is true even when the eigenvalues may be of a multiple nature, we need the following proposition. However, before stating it, deﬁne T = (tij ) ∈ Rnn to be a lower triangular matrix iﬀ tij = 0, i < j. Similarly, T ∈ Rnn is termed upper triangular iﬀ tij = 0, i > j. Proposition 1.8 Let A ∈ Cnn be any matrix. There exists a unitary matrix U ∈ Cnn such that UH AU is upper triangular. Proof. The proof is by induction on n. The result is obvious for n = 1. Next, assume the proposition holds for n and prove it is true for n + 1. Let λ1 be an eigenvalue of A and u1 , |u1 | = 1, be an eigenvector. Let U1 = (u1 , Γ) for some Γ such that U1 is unitary (such a Γ exists from the Gram-Schmidt method). Then, λ1 uH H 1 AΓ UH AU = U (λ u , AΓ) = , 1 1 1 1 1 0 B where B = ΓH AΓ ∈ Cnn . From the induction hypothesis, there exists V unitary such that VH BV = T is triangular. Deﬁne 1 0 U2 = 0 V and it is clear that U2 is also unitary. Finally, λ1 uH H H 1 AΓ (U1 U2 ) A(U1 U2 ) = U2 U2 0 B 1 0 λ1 uH 1 1 AΓ = 0 VH 0 0 B λ1 uH 1 AΓV = , 0 T

0 V

which is of the desired form. The proof is complete because U ≡ U1 U2 is unitary. 2 As a corollary we obtain that Hermitian matrices are always diagonalizable. Corollary 1.1 Let A ∈ Cnn be Hermitian. There exists a unitary matrix U such that UH AU = diag(λ). Proof. Proposition 1.8 showed there exists U, unitary, such that UH AU is triangular. However, if A is Hermitian, so is UH AU. The only matrices that are both Hermitian and triangular are the diagonal matrices. 2 In the sequel, we will always use Corollary 1.1 for S ∈ Rnn symmetric. However, ﬁrst note that when S is symmetric all its eigenvalues are real, whereby the eigenvectors can also be chosen to be real, they are the solutions of (S − λI)x = 0. When U ∈ Rnn is unitary, it is called an orthogonal

8

1. Linear algebra

matrix instead. A matrix H ∈ Rnn is said to be orthogonal iﬀ the columns (or rows) of H form an orthonormal basis of Rn , i.e., H H = I = HH . The group of orthogonal matrices in Rnn will be denoted by On = {H ∈ Rnn : HH = I}. We have proven the “spectral decomposition:” Proposition 1.9 If S ∈ Rnn is symmetric, then there exists H ∈ On such that H SH = diag(λ). The columns of H form an orthonormal basis of eigenvectors and λ is the vector of corresponding eigenvalues. Now, a symmetric matrix S ∈ Rnn is said to be positive semideﬁnite, denoted S ≥ 0 or S ∈ PSn , iﬀ v Sv ≥ 0, ∀v ∈ Rn , and it is positive Finally, the deﬁnite, denoted S > 0 or S ∈ Pn , iﬀ v Sv > 0, ∀v = 0. positive semideﬁnite and positive deﬁnite matrices can be characterized in terms of eigenvalues. Proposition 1.10 Let S ∈ Rnn symmetric with eigenvalues λ1 , . . . , λn . 1. S ≥ 0 iﬀ λi ≥ 0, i = 1, . . . , n. 2. S > 0 iﬀ λi > 0, i = 1, . . . , n. Note that if S is positive semideﬁnite, then from Proposition 1.9, we can write S = HDH = (HD1/2 )(HD1/2 ) = (HD1/2 H )2 , 1/2

where D = diag(λi ) and D1/2 = diag(λi ), so that for A = HD1/2 , S = AA , or for B = HD1/2 H , S = B2 . The positive semideﬁnite matrix B is often denoted S1/2 and is the square root of S. If S is positive deﬁnite, 2

we can also deﬁne S−1/2 = HD−1/2 H , which satisﬁes S−1/2 = S−1 . Finally, inequalities between matrices must be understood in terms of positive deﬁniteness; i.e., for matrices A and B, A ≥ B (respectively A > B) means A − B ≥ 0 (respectively A − B > 0). A related decomposition which will prove useful for canonical correlations is the singular value decomposition (SVD). Proposition 1.11 Let A ∈ Rm n of rank A = r. There exists G ∈ Om , H ∈ On such that Dρ 0 A=G H 0 0 where Dρ = diag(ρ1 , . . . , ρr ), ρi > 0, i = 1, . . . , r. Proof. Since A A ≥ 0, there exists H = (h1 , . . . , hn ) ∈ On such that A A = H diag(λ1 , . . . , λr , 0) H ,

1.6. Orthogonal projections

9

where λi > 0, i = 1, . . . , r. For j > r, |Ahj |2 = hj A Ahj = 0 which means Ahj = 0. For j ≤ r, deﬁne ρj = λj and gj = Ahj /ρj . Then, gi gj = hi A Ahj /ρi ρj = δij ; i.e., g1 , . . . , gr are orthonormal. By completing to an orthonormal basis of Rm , we can ﬁnd G = (g1 , . . . , gr , gr+1 , . . . , gm ) ∈ Om . Now, gi Ahj =

or in matrix notation,

G AH =

0, ρj δij ,

Dρ 0

j>r j ≤ r, 0 0

. 2

In the SVD ρ2j , j = 1, . . . , r, are the nonzero eigenvalues of A A and the columns of H are the eigenvectors.

1.6 Orthogonal projections Now recall some basic facts about orthogonal projections. By deﬁnition, an orthogonal projection, P, is simply a linear transformation for which x − Px ⊥ Py, ∀x, y ∈ Rn , but then, equivalently, (x − Px) (Py) = 0, ∀x, y ∈ Rn

⇐⇒ x Py = x P Py, ∀x, y ∈ Rn ⇐⇒ P P = P ⇐⇒ P = P = P2 .

A matrix P such that P = P = P2 is also called an idempotent matrix. Not surprisingly, an orthogonal projection is completely determined by its image. Proposition 1.12 If P1 and P2 are two orthogonal projections, then Im P1 = Im P2 ⇐⇒ P1 = P2 . Proof. It holds since x − P1 x ⊥ P2 y, ∀x, y ∈ Rn =⇒ P2 = P1 P2 , and, similarly, P1 = P2 P1 , whence P1 = P1 = P2 .

2

If X = (x1 , . . . , xk ) is any basis for Im P, we have explicitly P = X(X X)−1 X .

(1.3)

To see this, simply write Px = Xb, and orthogonality, X (x − Xb) = 0, determines the (unique) coeﬃcients b = (X X)−1 X x. In particular, for

10

1. Linear algebra

any orthonormal basis H, P = HH , where H H = Ik . Thus, incidentally, tr P = k and the dimension of the image space is expressed in the trace. However, by this representation we see that for any two orthogonal projections, P1 = HH and P2 = GG , P1 P2 = 0 ⇐⇒ H G = 0 ⇐⇒ G H = 0 ⇐⇒ P2 P1 = 0. Deﬁnition 1.1 P1 and P2 are said to be mutually orthogonal projections iﬀ P1 and P2 are orthogonal projections such that P1 P2 = 0. We write P1 ⊥ P2 when this is the case. Although orthogonal projection and orthogonal transformation are far from synonymous, there is, nevertheless, ﬁnally a very close connection between the two concepts. If we partition any orthogonal transformation H = (H1 , . . . , Hk ), then the brute algebraic fact HH = I = H1 H1 + · · · + Hk Hk represents a precisely corresponding partition of the identity into mutually orthogonal projections. As a last comment on othogonal projection, if P is the orthogonal projection on the subspace V ⊂ Rn , then Q = I−P, which satisﬁes Q = Q = Q2 is also an othogonal projection. In fact, since PQ = 0, then Im Q and Im P are orthogonal subspaces and, thus, Q is the orthogonal projection on V ⊥ .

1.7 Matrix decompositions Denote the groups of triangular matrices with positive diagonal elements as L+ n U+ n

= {T ∈ Rnn : T is lower triangular, tii > 0, i = 1, . . . , n}, = {T ∈ Rnn : T is upper triangular, tii > 0, i = 1, . . . , n}.

An important implication of Proposition 1.2 for matrices is the following matrix decomposition. Proposition 1.13 If A ∈ Rnn is nonsingular, then A = TH for some H ∈ On and T ∈ L+ n . Moreover, this decomposition is unique. Proof. The existence follows from the Gram-Schmidt method applied to the basis formed by the rows of A. The rows of H form the orthonormal basis obtained at the end of that procedure and the elements of T = (tij ) are the coeﬃcients needed to go from one basis to the other. By the GramSchmidt construction itself, it is clear that T ∈ L+ n . For unicity, suppose −1 TH = T1 H1 , where T1 ∈ L+ n and H1 ∈ On . Then, T1 T = H1 H is a + matrix in Ln ∩ On . But, In is the only such matrix (why?). Hence, T = T1 and H = H1 . 2

1.8. Problems

11

A slight generalization of Proposition 1.13 when A ∈ Rpn is of rank A = p is proposed in Problem 1.8.7. Another similar triangular decomposition, known in statistics as the Bartlett decomposition, for positive deﬁnite matrices can now be easily obtained. Proposition 1.14 If S ∈ Pn , then S = TT for a unique T ∈ L+ n. Proof. Since S > 0, then S = HDH , where H ∈ On and D = diag(λi ) 1/2 with λi > 0. Let D1/2 = diag(λi ) and A = HD1/2 . Then, we can write S = AA , where A is nonsingular. From Proposition 1.13, there exists T ∈ L+ n and G ∈ On such that A = TG. But, then, S = TGG T = TT . For −1 + −1 unicity, suppose TT = T1 T1 , where T1 ∈ Ln . Then, T1 TT T1 = I, + 2 which implies that T−1 1 T ∈ Ln ∩ On = {I}. Hence, T = T1 . Other notions of linear algebra such as Kronecker product and “vec” operator will be recalled when needed in the sequel.

1.8 Problems 1. Consider the partitioned matrix S = (sij ) =

S11 S21

S12 . S22

(i) If S11 is nonsingular, prove that |S| = |S11 | · |S22 − S21 S−1 11 S12 |. (ii) For S > 0, prove Hadamard’s inequality, |S| ≤ i sii . (iii) Let S and S11 be nonsingular. Prove that −1 −1 −1 −1 S11 + S−1 −S−1 11 S12 S22.1 S21 S11 11 S12 S22.1 , S−1 = −1 −S−1 S−1 22.1 S21 S11 22.1 where S22.1 = S22 − S21 S−1 11 S12 . (iv) Let S and S22 be nonsingular. Prove that −1 S−1 −S−1 11.2 11.2 S12 S22 , S−1 = −1 −1 −1 −1 −S−1 S−1 22 S21 S11.2 22 + S22 S21 S11.2 S12 S22 where S11.2 = S11 − S12 S−1 22 S21 . Hint: Deﬁne A=

I

−S21 S−1 11

0 I

and B =

I −S−1 11 S12 0 I

and consider the product ASB. 2. Establish with the partitioning (x , x ) , 1 2 S11 S12 S = S21 S22 x =

12

1. Linear algebra

that −1 −1 −1 x S−1 x = (x1 − S12 S−1 22 x2 ) S11.2 (x1 − S12 S22 x2 ) + x2 S22 x2 .

3. For any A ∈ Rpq , B ∈ Rqp , prove the following: (i) |Ip + AB| = |Iq + BA|. Hint: Ip + AB A 0 Iq Ip A 0 Iq + BA

Ip A Ip 0 = , −B Iq B Iq Ip A Ip 0 . = B Iq −B Iq

(ii) The nonzero eigenvalues of AB and BA are the same. 4. Prove Proposition 1.2. 5. Prove Proposition 1.10. 6. Show that if P deﬁnes an orthogonal projection, then the eigenvalues of P are either 0 or 1. 7. Demonstrate the slight generalizations of Proposition 1.13: (i) If A ∈ Rnp is of rank A = p, then A = HT for some T ∈ U+ p and H satisfying H H = Ip . Further, T and H are unique. Hint: For unicity, note that if A = HT = H1 T1 with T1 ∈ U+ p and H1 H1 = Ip , then Im A = Im H = Im H1 and H1 H1 is the orthogonal projection on Im H1 . (ii) If A ∈ Rnp is of rank A = n, then A = TH, where T ∈ L+ n and HH = In . Further, T and H are unique. 8. Assuming A and A + uv are nonsingular, prove (A + uv )−1 = A−1 −

A−1 uv A−1 . (1 + v A−1 u)

9. Vector diﬀerentiation. Let f (x) be a real valued function of x ∈ Rn . Deﬁne ∂f (x)/∂x = (∂f (x)/∂xi ) . Verify (i) ∂a x/∂x = a, (ii) ∂x Ax/∂x = 2Ax, if A is symmetric. 10. Matrix diﬀerentiation [Srivastava and Khatri (1979), p. 37]. Let g(S) be a real-valued function of the symmetric matrix S ∈ Rnn .

1 Deﬁne ∂f (S)/∂S = 2 (1 + δij )∂f (S)/∂sij . Verify (i) ∂tr(S−1 A)/∂S = −S−1 AS−1 , if A is symmetric, (ii) ∂ ln |S|/∂S = S−1 .

1.8. Problems

13

Hint for (ii): S−1 = |S|−1 adj(S). 11. Rayleigh’s quotient. Assume S ≥ 0 in Rnn with eigenvalues λ1 ≥ · · · ≥ λn and corresponding eigenvectors x1 , . . . , xn . Prove: (i) x Sx ≤ λ1 , ∀x = 0. x x (ii) For any ﬁxed j = 2, . . . , n, λn ≤

x Sx ≤ λj , ∀x = 0 x x such that x, x1 = · · · = x, xj−1 = 0. 12. Demonstrate that if A is symmetric and B > 0, then h Ah = λ1 (AB−1 ), |h|=1 h Bh sup

where λ1 (AB−1 ) denotes the largest eigenvalue of AB−1 . 13. Let Am > 0 in Rnn (m = 1, 2, . . .) be a sequence. For any A ∈ Rnn , 2 deﬁne ||A|| = i,j a2ij and let λ1,m ≥ · · · ≥ λn,m be the ordered eigenvalues of Am . Prove that if λ1,m → 1 and λn,m → 1, then limm→∞ ||Am − I|| = 0. 14. In Rp , prove that if |x1 | = |x2 |, then there exists H ∈ Op such that Hx1 = x2 . Hint: When x1 = 0, consider H ∈ Op with ﬁrst row x1 /|x1 |. 15. Show that for any V ∈ Rnn and any m = 1, 2, . . ., (i) if (I − tV) is nonsingular then [Srivastava and Khatri (1979), p. 33] (I − tV)−1 =

m

ti Vi + tm+1 Vm+1 (I − tV)−1 .

i=0

(ii) If V > 0 with eigenvalues λ1 ≥ · · · ≥ λp and |t| < 1/λ1 , then (I − tV)

−1

=

∞ i=0

ti V i .

2 Random vectors

2.1 Introduction A random vector is simply a vector whose components are random variables. The variables are the characteristics of interest that will be observed on each of the selected units in the sample. Questions related to probabilities of a variable to take on some values or probabilities of two or more variables to take on simultaneously values in a set are common in multivariate analysis. Chapter 2 gives a collection of important probability concepts on random vectors such as distribution functions, expected values, characteristic functions, discrete and absolutely continuous distributions, independence, etc.

2.2 Distribution functions First, some basic notations concerning “rectangles” useful to describe the ¯ = R ∪ {±∞} = distribution function of a random vector are given. Let R n ¯ [−∞, ∞]. It is convenient to deﬁne a partial order on R by x ≤ y iﬀ xi ≤ yi , ∀i = 1, . . . , n, and x < y iﬀ xi < yi , ∀i = 1, . . . , n.

2.2. Distribution functions

15

This allows us to express “n-dimensional” rectangles in Rn succinctly: ¯ n. I = (a, b] = {x ∈ Rn : a < x ≤ b} for any a, b ∈ R The interior and closure of I are respectively I ◦ = (a, b) = {x ∈ Rn : a < x < b} and I¯ = [a, b] = {x ∈ Rn : a ≤ x ≤ b} and the boundary of I is the “(n − 1)-dimensional” relative complement ∂I = I¯ − I ◦ . ¯ n ) be denoted by the cartesian Finally let the 2n “corners” of I (a subset of R product a × b = ×ni=1 {ai , bi }. Deﬁnition 2.1 For x distributed on Rn , the distribution function (d.f.) ¯ n. ¯ n → [0, 1], where F (t) = P (x ≤ t), ∀t ∈ R of x is the function F : R This is denoted x ∼ F or x ∼ Fx . A d.f. is automatically right-continuous; thus, if it is known on any dense ¯ n, subset D ⊂ Rn , it is determined everywhere. This is because for any t ∈ R a sequence dn may be chosen in D descending to t: dn ↓ t. From the d.f. may be computed the probability of any rectangle P (a < x ≤ b) = (−1)Na (t) F (t), ∀a < b, t∈a×b

n

where Na (t) = i=1 δ(ai , ti ) counts the number of ti ’s that are ai ’s. The borel subsets of Rn comprise the smallest σ-algebra containing the rectangles B n = σ ((a, b] : a, b ∈ Rn ) . unions of rectangles contains all the The class G n of all countable disjoint ∞ open subsets of Rn , and if we let G = i=1 (ai , bi ] denote a generic element in this class, it follows that P (x ∈ G) =

∞

P (ai < x ≤ bi ).

i=1

By the Caratheodory extension theorem (C.E.T.), the probability of a general borel set A ∈ Bn is then uniquely determined by the formula Px (A) ≡ P (x ∈ A) = inf P (x ∈ G). A⊂G

16

2. Random vectors

2.3 Equals-in-distribution Deﬁnition 2.2 x and y are equidistributed (identically distributed), d

denoted x = y, iﬀ Px (A) = Py (A), ∀A ∈ Bn . On the basis of the previous section, it should be clear that for any dense D ⊂ Rn : d

Proposition 2.1 (C.E.T) x = y ⇐⇒ Fx (t) = Fy (t), ∀t ∈ D. d

Although at ﬁrst glance, = looks like nothing more than a convenient shorthand symbol, there is an immediate consequence of the deﬁnition, deceptively simple to state and prove, that has powerful application in the sequel. Let g : Rn → Ω where Ω is a completely arbitrary space. d

d

Proposition 2.2 (Invariance) x = y =⇒ g(x) = g(y). Proof.

P (g(x) ∈ B) = P x ∈ g −1 (B) = P y ∈ g −1 (B) = P (g(y) ∈ B) . 2

Example 2.1 d

x=y

=⇒

d

xi = yi , i = 1, . . . , n d

=⇒ xi xj = yi yj , i, j = 1, . . . , n n n d =⇒ xri i = yiri , for any ri , i = 1, . . . , n i=1

i=1

=⇒ etc.

2.4 Discrete distributions Deﬁnition 2.3 The probability function (p.f.) of x is the function ¯ n → [0, 1] where p(t) = P (x = t), ∀t ∈ R ¯ n. p:R The p.f. may be evaluated directly from the d.f.: p(t) = lim P (sm < x ≤ t), sm ↑t

where sm ↑ t means s1 < s2 < · · · and sm → t as m → ∞. The subset D = p−1 (0)c where the p.f. is nonzero may contain at most a countable number of points. D is known as the discrete part of x, and x is said to be discrete if it is “concentrated” on D: Deﬁnition 2.4 x is discrete iﬀ P (x ∈ D) = 1.

2.5. Expected values

One may verify that x is discrete ⇐⇒ P (x ∈ A) =

17

p(t), ∀A ∈ Bn .

t∈A∩D

Thus, the distribution of x is entirely determined by its p.f. if and only if it is discrete, and in this case, we may simply write x ∼ p or x ∼ px .

2.5 Expected values For any event A, we may consider the indicator function 1, x ∈ A IA (x) = 0, x ∈ A. It is clear that IA (x) is itself a discrete random variable, referred to as a Bernoulli trial, for which P (IA (x) = 1) = Px (A) and P (IA (x) = 0) = 1 − Px (A). This is denoted IA (x) ∼ Bernoulli (Px (A)) and we deﬁne E IA (x) = Px (A). For any k mutually disjoint and exhaustive events A1 , . . . , Ak and k real numbers a1 , . . . , ak , we may form the simple function s(x) = a1 IA1 (x) + · · · + ak IAk (x). Obviously, s(x) is also discrete with P (s(x) = ai ) = Px (Ai ), i = 1, . . . , k. By requiring that E be linear, we (are forced to) deﬁne E s(x) = a1 Px (A1 ) + · · · + ak Px (Ak ). The most general function for which we need ever compute an expected value may be directly expressed as a limit of a sequence of simple functions. Such a function g(x) is said to be measurable and we may explicitly write g(x) = lim sN (x), N →∞

where convergence holds pointwise, i.e., for every ﬁxed x. If g(x) is nonnegative, it can be proven that we may always choose the sequence of simple functions to be themselves non-negative and nondecreasing as a sequence whereupon we deﬁne E g(x) = lim E sN (x) = sup E sN (x). N →∞

N

Then, in general, we write g(x) as the diﬀerence of its positive and negative parts g(x) = g + (x) − g − (x),

18

2. Random vectors

deﬁned by +

g (x) −

g (x)

= =

g(x), 0, −g(x), 0,

g(x) ≥ 0 g(x) < 0, g(x) ≤ 0 g(x) < 0,

and ﬁnish by deﬁning E g + (x) − E g − (x), if E g + (x) < ∞ or E g − (x) < ∞ E g(x) = “undeﬁned,” otherwise. We may sometimes use the Leibniz notation E g(x) = g(t)dPx (t) = g(t)dF (t). One should verify the fundamental inequality |E g(x)| ≤ E |g(x)|. Let ↑ denote convergence of a monotonically nondecreasing sequence. Something is said to happen for almost all x if it fails to happen on a set A such that Px (A) = 0. The two main theorems concerning “continuity” of E are the following: Proposition 2.3 (Monotone convergence theorem (M.C.T.)) Suppose 0 ≤ g1 (x) ≤ g2 (x) ≤ · · ·. If gN (x) ↑ g(x), for almost all x, then E gN (x) ↑ E g(x). Proposition 2.4 (Dominated convergence theorem (D.C.T.)) If gN (x) → g(x), for almost all x, and |gN (x)| ≤ h(x) with E h(x) < ∞, then E |gN (x) − g(x)| → 0 and, thus, also E gN (x) → E g(x). It should be clear by the process whereby expectation is deﬁned (in stages) that we have d

Proposition 2.5 x = y ⇐⇒ E g(x) = E g(y), ∀g measurable.

2.6 Mean and variance n n i=1 ti xi for each (ﬁxed) t ∈ R ,

n 1/2 2 and the “euclidean norm” (length) |x| = . By any of three i=1 xi equivalent ways, for p > 0 one may say that the pth moment of x is ﬁnite: Consider the “linear functional” t x =

E |t x|p < ∞, ∀t ∈ Rn

⇐⇒ E |xi |p < ∞, i = 1, . . . , n ⇐⇒ E |x|p < ∞. n To show this, one must realize that |xi | ≤ |x| ≤ i=1 |xi | and Lp = {x ∈ Rn : E |x|p < ∞} is a linear space (v. Problem 2.14.3). From the simple inequality ar ≤ 1 + ap , ∀a ≥ 0 and 0 < r ≤ p, if we let a = |x| and take expectations, we get E |x|r ≤ 1 + E |x|p . Hence, if for

2.6. Mean and variance

19

p > 0, the pth moment of x is ﬁnite, then also the rth moment is ﬁnite, for any 0 < r ≤ p. A product-moment of order p for x = (x1 , . . . , xn ) is deﬁned by E

n

xpi i , pi ≥ 0, i = 1, . . . , n,

i=1

n

pi = p.

i=1

A useful inequality to determine that a product-moment is ﬁnite is H¨ older’s inequality: Proposition 2.6 (H¨ older’s inequality) For any univariate random variables x and y, 1 1 + = 1. r s n From this inequality, if the pth moment of x ∈ R is ﬁnite, then all productmoments of order p are also ﬁnite. This can be veriﬁed for n = 2, as H¨older’s inequality gives 1/r

E |xy| ≤ (E |x|r )

p1 /p

E |xp11 xp22 | ≤ (E |x1 |p )

1/s

· (E |y|s )

p2 /p

· (E |x2 |p )

, r > 1,

, pi ≥ 0, i = 1, 2, p1 + p2 = p.

The conclusion for general n follows by induction. If the ﬁrst moment of x is ﬁnite we deﬁne the mean of x by def µ = E x = (E xi ) = (µi ). If the second moment of x is ﬁnite, we deﬁne the variance of x by def Σ = var x = (cov(xi , xj )) = (σij ) . In general, we deﬁne the expected value of any multiply indexed array of univariate random variables, ξ = (xijk··· ), componentwise by E ξ = (E xijk··· ). Vectors and matrices are thus only special cases and it is obvious that Σ = E (x − µ)(x − µ) = E xx − µµ . It is also obvious that for any A ∈ Rm n, E Ax = Aµ and var Ax = AΣA . In particular, E t x = t µ and var t x = t Σt ≥ 0, ∀t ∈ Rn . Now, the reader should verify that more generally cov(s x, t x) = s Σt and that considered as a function of s and t, the left-hand side deﬁnes a (pseudo) inner product. Thus, Σ is automatically positive semideﬁnite, Σ ≥ 0. But by this, we may immediately write Σ = HDH with H orthogonal and D = diag(λ), where the columns of H comprise an orthonormal basis of “eigenvectors” and the components of λ ≥ 0 list the corresponding

20

2. Random vectors

“eigenvalues.” Accordingly, we may always “normalize” any x with Σ > 0 by letting z = D−1/2 H (x − µ), which represents a three-stage transformation of x in which we ﬁrst relocate −1/2 independently along by µ, then rotate by H , and, ﬁnally, rescale by λi each axis. We ﬁnd, of course, that E z = 0 and var z = I. The linear transformation z = Σ−1/2 (x − µ) also satisﬁes E z = 0 and var z = I. When the vector x ∈ Rn is partitioned as x = (y , z ) , where y ∈ Rr , z ∈ Rs , and n = r + s, it is useful to deﬁne the covariance between two vectors. The covariance matrix between y and z is, by deﬁnition, cov(y, z) = (cov(yi , zj )) ∈ Rrs . Then, we may write

var(x) =

var(y) cov(z, y)

cov(y, z) var(z)

.

Sometimes, expected value of y is easier to calculate by conditioning on another random vector z. In this regard, the conditional mean theorem and conditional variance theorem are stated. A general proof of the conditional mean theorem can be found in Billingsley (1995, Section 34). Proposition 2.7 (Conditional mean formula) E[E(y|z)] = E y. An immediate consequence is the conditional variance formula. Proposition 2.8 (Conditional variance formula) var y = E[var(y|z)] + var[E(y|z)]. Example 2.2 Deﬁne a group variable I such that P (I = 1) = 1 − , P (I = 2) = . Conditionally on I, assume x|I = 1 ∼ N (µ1 , σ12 ), x|I = 2 ∼ N (µ2 , σ22 ).

Then fx (x)

=

1 1 exp − 2 (x − µ1 )2 σ1 2σ1 1 1 +(2π)−1/2 exp − 2 (x − µ2 )2 σ2 2σ2 (1 − )(2π)−1/2

2.7. Characteristic functions

21

is a mixture or -contaminated normal density. It follows from the construction of x that E x = E[E(x|I)] = (1 − )µ1 + µ2 ≡ µ, var x = E[var(x|I)] + var[E(x|I)] = (1 − )σ12 + σ22 + (1 − )(µ1 − µ)2 + (µ2 − µ)2 .

2.7 Characteristic functions We require only the most basic facts about characteristic functions. Deﬁnition 2.5 The characteristic function of x is the function c : Rn → C deﬁned by

c(t) = cx (t) = E eit x . Note: 1. c(0) = 1, |c(t)| ≤ 1 and c(−t) = c(t). 2. c(t) is uniformly continuous:

|c(t) − c(s)| = E ei(t−s) x − 1 eis x ≤ E ei(t−s) x − 1 .

Since ei(t−s) x − 1 ≤ 2, continuity follows by the D.C.T. Uniformity holds since ei(t−s) x − 1 depends only on t − s. The main result is perhaps the “inversion formula” proven in Appendix A: 1 N →∞ (2π)n

Px (a, b] = lim

2

e−it x c(t)e−t t/2N dtdx,

Rn

(a,b]

∀a, b such that Px (∂(a, b]) = 0. Thus, the C.E.T. may be applied immediately to produce the technically equivalent: d

Proposition 2.9 (Uniqueness) x = y ⇐⇒ cx (t) = cy (t), ∀t ∈ Rn . Now if we consider the linear functionals of x: t x with t ∈ Rn , it is clear that ct x (s) = cx (st), ∀s ∈ R, t ∈ Rn , so that the characteristic function of x determines all those of t x, t ∈ Rn and vice versa. Let S n−1 = {s ∈ Rn : |s| = 1} be the “unit sphere” in Rn , and we have d

d

Proposition 2.10 (Cram´ er-Wold) x = y ⇐⇒ t x = t y, ∀t ∈ S n−1 .

22

2. Random vectors

Proof. Since ct x (s) = cx (st), ∀s ∈ R, t ∈ Rn , it is clear that x = y ⇐⇒ t x = t y, ∀t ∈ Rn . d

Since t x = |t|

t |t|

d

x, ∀t = 0, it is also clear that

t x = t y, ∀t ∈ Rn ⇐⇒ t x = t y, ∀t ∈ S n−1 . d

d

2 By this result, it is clear that one may reduce a good many issues concerning random vectors to the univariate level. In the speciﬁc matter of computation, the reader should know that in the special case of a univariate random variable X: If E e±δX < ∞ for any δ > 0, the Laplace transform of X is determined in the strip |Re(z)| ≤ δ as the (absolutely convergent) power series LX (z) =

∞

E X n z n /n!,

n=0

and since such a power series is completely determined by its coeﬃcients, we ﬁnd that one may legitimately obtain the characteristic function cX (t) = LX (it), ∀t ∈ R, by merely observing the coeﬃcients in an expansion of the moment-generating function since they are necessarily the same as those of the Laplace transform: mX (t) = LX (t), ∀|t| ≤ δ. Example 2.3 Suppose fz (s) = (2π)−1/2 e−s moment-generating function (m.g.f.), ﬁnding 2

mz (t) = et

/2

2

/2

. One easily computes the

,

which has the obvious expansion for every t, whereupon 2

cz (t) = e−t

/2

.

(2.1)

2.8 Absolutely continuous distributions Lebesgue measure, λ, is the extension to all borel sets of our natural sense of volume measure in Rn . Thus, we deﬁne λ(a, b]

=

n i=1

(bi − ai ), ∀a < b in Rn ,

2.8. Absolutely continuous distributions

λ(G)

=

∞

λ(ai , bi ], ∀G =

i=1

∞

23

(ai , bi ] in G n ,

i=1

and λ(A) = inf λ(G), ∀A in Bn . A⊂G

As before, the C.E.T. guarantees that λ is a measure on B n . We will often denote Lebesgue measure explicitly as volume: λ(A) = vol(A). Incidentally, something is said to happen “almost everywhere” (a.e.) if the set where it fails to happen has zero volume. Now, the general conception of a random vector continuously distributed in space is that the probabilities of events will depend continuously on the volume of the events. Thus, Deﬁnition 2.6 x is absolutely continuous, denoted x λ, iﬀ ∀ > 0, ∃δ > 0 such that vol(A) < δ =⇒ P (x ∈ A) < . But, in that case, Proposition 2.11 x λ ⇐⇒ vol(A) = 0 =⇒ P (x ∈ A) = 0. Proof. Assume x λ. If vol(A) = 0 but P (x ∈ A) = 0 we may take = P (x ∈ A)/2 to ﬁnd the contradiction that P (x ∈ A) < . Conversely, suppose vol(A) = 0 =⇒ P (x ∈ A) = 0 but that x λ. Then, ∃0 > 0 such that ∀n, ∃An with vol(An ) < 1/2n but P (x ∈ An ) ≥ 0 . Letting A = limAn = ∩∞ n=1 ∪k≥n Ak , since ∪k≥n Ak is a monotone sequence we ﬁnd 2 the contradiction that vol(A) = 0 but P (x ∈ A) ≥ 0 . Thus, a distribution which depends continuously on volume satisﬁes the relatively simple criterion x λ ⇐⇒ vol(A) = 0 =⇒ P (x ∈ A) = 0. However, it is on this particular criterion, by the theorem of RadonNikodym, that absolute continuity is characterized ﬁnally in terms of densities: Proposition 2.12 (Radon-Nikodym) x is absolutely continuous ⇐⇒ there is a (a.e.-unique) probability density function (p.d.f.) f : Rn → [0, ∞) such that f (t)dt, ∀A ∈ Bn . P (x ∈ A) = A

But since the p.d.f. then determines such a distribution completely, we may simply write x ∼ f or x ∼ fx . It is, of course, by the extension process that deﬁnes expectation (in stages) that automatically E g(x) = g(t)f (t)dt, ∀g measurable,

24

2. Random vectors

such that E g(x) is deﬁned. Now in particular, the distribution function may itself be expressed as t F (t) = P (x ≤ t) = f (s)ds, ∀t ∈ Rn . −∞ In practice, we will often be able to invoke the fundamental theorems of calculus to obtain an explicit representation of the p.d.f. by simply diﬀerentiating the d.f.: 1. By the ﬁrst fundamental theorem of calculus, f (t) = ∂ n F (t)/∂t1 · · · ∂tn at every t where f (t) is continuous. 2. Also, by the second fundamental theorem, if f (t) = ∂ n F (t)/∂t1 · · · ∂tn exists and is continuous (a.e.) on some rectangle I, then P (x ∈ A) = f (t)dt, ∀A ⊂ I. A

Finally, in relation to the inversion formula, when the characteristic func tion c(t) is absolutely integrable, i.e., Rn |c(t)|dt < ∞, the corresponding distribution function is absolutely continuous with p.d.f. (v. Appendix A): 1 f (s) = e−it s c(t)dt. (2.2) n (2π) Rn

2.9 Uniform distributions The most fundamental absolutely continuous distribution would, of course, be conveyed by volume measure itself. Consider any event C for which 0 < vol(C) < ∞. Deﬁnition 2.7 x is uniformly distributed on C, denoted x ∼ unif(C), iﬀ P (x ∈ A) = vol(AC)/vol(C), ∀A ∈ Bn . If vol(∂C) = 0, as is often the case, we may just as well include as exclude ¯ then x, y, and z are it, so if x ∼ unif(C), y ∼ unif(C ◦ ) and z ∼ unif(C), equidistributed: d

d

x = y = z. Now, for x ∼ unif(C) we may immediately reexpress each probability as an integral: P (x ∈ A) = k · IC (t)dt, ∀A ∈ Bn , A

2.10. Joints and marginals

where

IC (t) =

25

1, t ∈ C 0, t ∈ C,

is the indicator function for C and k = vol(C)−1 . We thus have an explicit determination of “the” density for x: f (t) = k IC (t), ∀t ∈ Rn . Example 2.4 For x ∼ unif([0, 1]) on Rn , the p.d.f. may be expressed as a simple product f (t) = I[0,1] (t) =

n

I[0,1] (ti ), ∀t ∈ Rn ,

i=1

from which F (t) =

n

ti I[0,1] (ti ) + I(1,∞) (ti ) , ∀t ∈ Rn .

i=1

2.10 Joints and marginals Consider xi ∼ Fi on Rni , i = 1, . . . , k, with x1 k .. n x= ∼ F on R where n = ni . . i=1 xk x is called the joint of x1 , . . . , xk which are, in turn, called marginals of x. Since it is clear that

t1 . P (x ≤ t) = P (x1 ≤ t1 , . . . , xk ≤ tk ), ∀t = .. , tk

we will, by a slight abuse of our notation, write

t1 . F (t) = F (t1 , . . . , tk ), ∀t = .. tk

to reﬂect this “partitioning.” In this way, the distribution function is said to express the joint distribution of x1 , . . . , xk , and the marginals may be recovered on the simple substitution of ∞ in all but the ith place: Fi (s) = F (∞, . . . , s, . . . , ∞), ∀s ∈ Rni .

26

2. Random vectors

In the special case where x is absolutely continuous with p.d.f. f (t) = f (t1 , . . . , tk ), it follows that each xi is also absolutely continuous with p.d.f. fi (s) that is obtained by “integrating out” the other variables: ∞ ∞ fi (s) = dtj , ∀s ∈ Rni . ··· f (t1 , . . . , s, . . . , tk ) −∞ −∞ 1≤j≤k j=i

This is by direct application of Fubini’s theorem whereby we may interchange the order of integration in a product integral to verify that fi (s)ds, ∀A ∈ Bni P (xi ∈ A) = A

and, of course, in particular,

Fi (s) =

s

fi (u)du, ∀s ∈ Rni .

−∞ Koehler and Symanowski (1995) presented a method for constructing multivariate distributions with any speciﬁc set of univariate marginals. It provides a rich class of distributions for modeling multivariate data as well as a basis for easily simulating correlated observations. The inclusion of diﬀerent association parameters for diﬀerents subsets of variables allows for many diﬀerent patterns of associations. Their work follows those of Genest and MacKay (1986) and Marshall and Olkin (1988), among others. A tool called linkage [Li et al. (1996)] can be used for the construction of multivariate distributions with given multivariate marginals; Cuadras (1992) found related results.

Example 2.5 The bivariate parametric family of d.f.’s on [0, 1]2 of Cook and Johnson (1981) is deﬁned by −1/α 1 1 , α > 0. (2.3) F (t1 , t2 ; α) = α + α − 1 t1 t2 The case α = 0 can be deﬁned by continuity. It has marginals F1 (t1 )

= F (t1 , 1; α) = t1 ,

F2 (t2 )

= F (1, t2 ; α) = t2 ,

which are identically distributed as unif([0, 1]). Multivariate distributions on [0, 1] with uniform marginals are often referred to as copulas. The slight modiﬁcation −1/α 1 1 + − 1 , α > 0, F (t1 , t2 ; α) = F1 (t1 )α F2 (t2 )α is a bivariate distribution with arbitrary marginals F1 and F2 . The bivariate parametric family of d.f.’s on [0, 1]2 of Frank (1979) [v. also Genest (1987)] (αt1 − 1)(αt2 − 1) F (t1 , t2 ; α) = logα 1 + , α > 0, (2.4) (α − 1)

2.11. Independence

27

0.2 2 0.1 1 0 0

-2 -1

-1

0 1

-2 2

Figure 2.1. Bivariate Frank density with standard normal marginals and a correlation of 0.7.

(the case α = 1 can be deﬁned by continuity), where logα (·) denotes logarithm in base α, is also a copula. Such distributions have found applications in modeling survival data [Oakes (1982), Carri`ere (1994)]. Figure 2.1 is a graph of a bivariate Frank density with standard normal marginals. The association parameter α = 0.00296 using Nelsen (1986) corresponds to a correlation of 0.7.

2.11 Independence Deﬁnition 2.8 x1 , . . . , xk are mutually statistically independent iﬀ P (x1 ∈ A1 , . . . , xk ∈ Ak ) =

k

P (xi ∈ Ai ), ∀Ai ∈ Bni , i = 1, . . . , k. |=

i=1

x2 . By the extension Denote pairwise independence (k = 2) simply x1 process that deﬁnes expectation, we ultimately ﬁnd: Proposition 2.13 x1 , . . . , xk are independent ⇐⇒ E

k i=1

∀g1 , . . . , gk such that E |gi (xi )| < ∞.

gi (xi ) =

k i=1

E gi (xi ),

28

2. Random vectors

and also (chieﬂy by the C.E.T.) Proposition 2.14 x1 , . . . , xk are independent ⇐⇒ F (t) =

k

Fi (ti ), ∀t ∈ Rn .

i=1

In the special case where each xi is absolutely continuous with p.d.f. fi (ti ), we may conclude that x is as well, and we have: Proposition 2.15 x1 , . . . , xk are independent ⇐⇒ f (t) =

k

fi (ti ), ∀t ∈ Rn .

i=1

Finally, independence may also be characterized: Proposition 2.16 x1 , . . . , xk are independent ⇐⇒ cx (t) =

k

cxi (ti ), ∀t ∈ Rn .

i=1

Example 2.6 For x ∼ unif([0, 1]) on R , it is clear that x1 , . . . , xn are independently and identically distributed (i.i.d.) as unif([0, 1]). n

|=

Example 2.7 Let x = (x1 , x2 ) have Frank’s d.f. (2.4). Proposition 2.14 x2 iﬀ α = 1. yields, after elementary calculus, x1

2.12 Change of variables We recall some basic calculus [Spivak (1965), p. 16]. Let A ⊂ Rn be open. Deﬁnition 2.9 The derivative of φ : A → Rm , at x ∈ A, is the unique linear transformation φ (x) ∈ Rm n such that φ(x + h) − φ(x) = φ (x)h + o(h) or, equivalently, lim

h→0

φ(x + h) − φ(x) − φ (x)h |h|

= 0.

When φ (x) exists, all partial derivatives ∂φi (x)/∂xj exist. This determines the derivative componentwise as φ (x) = (∂φi (x)/∂xj ) . A condition for φ (x) to exist is that all partial derivatives ∂φi (x)/∂xj exist in an open neighborhood of x and are continuous at x. There are, of course, various notations for derivatives, all acceptable: φ (x) = Dφ(x) = ∂φ(x)/∂x.

2.12. Change of variables

29

The derivative satisﬁes the “chain rule” (φ ◦ ψ) (x) = φ (ψ(x)) ψ (x). In the very special case m = n, the jacobian of φ : Rn → Rn is, by deﬁnition, the absolute value of the determinant of φ (x) and is denoted by |φ (x)|+ . Another common notation for the jacobian of the transformation y = φ(x) is J(y → x) = |φ (x)|+ . From the chain rule, it is made clear that if z = φ(y) and y = ψ(x), then J(z → x) = J(z → y) · J(y → x), J(y → x) = [J(x → y)]−1 . At any rate, we have an important and general result, easy to state, but the proof of which is by no means trivial [Spivak (1965), p. 67]. Proposition 2.17 Let φ : A → Rn be one-to-one and continuously diﬀerentiable on A. If f : φ(A) → R is integrable, then f (x)dx = f (φ(y)) |φ (y)|+ dy. φ(A) A It is this result that is applied directly to obtain the standard “change of variables” formula for absolutely continuous random vectors. Proposition 2.18 If x ∼ f on Rn and C = {x : f (x) > 0} is open, for any φ : C → Rn one-to-one and bi-diﬀerentiable with inverse ψ : φ(C) → C, let y = φ(x). Then, y ∼ g with g(y) = f (ψ(y)) |ψ (y)|+ . Proof. P (y ∈ B)

= P (φ(x) ∈ B) = P (x ∈ ψ(B)) = f (x)dx = f (ψ(y)) |ψ (y)|+ dy. ψ (B) B 2

By an abuse of notation, y (and x) have two diﬀerent meanings in Proposition 2.18: y is a random vector in y ∼ g, whereas it is any given point of Rn in the density g(y). Now, if the function φ in question is simply a linear transformation, φ(x) = Ax, it is already its own derivative everywhere on Rn , φ (x) = A, and the formula for change of variables greatly simpliﬁes. Suppose that A : Rn → Rn is a nonsingular transformation. The group of all such nonsingular transformations is known as the general linear group and denoted by Gn = {A ∈ Rnn : A is nonsingular} = {A ∈ Rnn : |A| = 0}.

30

2. Random vectors

Two examples are as follows: Example 2.8 If x ∼ f and y = Ax, we ﬁnd y ∼ g, where g(y) = f (A−1 y) |A|−1 + . Example 2.9 x ∼ unif(C), C ⊂ Rn =⇒ Ax + b ∼ unif(AC + b) where AC + b = {Ax + b : x ∈ C}.

2.13 Jacobians The derivation of jacobians is the diﬃcult part in making transformations. It can be a daunting task. This section is directed to the derivation of more complicated jacobians. It can be skipped on a ﬁrst reading and consulted when needed in the sequel. Although jacobians are useful for densities, our approach is to derive distributions without appealing, whenever possible, to densities. Derivations of densities appear mainly in the form of problems. Proposition 2.19 The jacobian of the transformation V = AWA , W ∈ Rnn symmetric and A ∈ Rnn constant, is J(V → W) = |A|n+1 + . Proof. The transformation is linear and, thus, the jacobian is necessarily a polynomial in the elements of A, p(A) say. If W = BUB , then from the chain rule, we have J(V → U) = J(V → W) · J(W → U), i.e., p(AB) = p(A)p(B). The only polynomials in the elements of a matrix satisfying this multiplicative rule are the integer powers of the determinant [MacDuﬀy (1943, p. 50)]. Hence, p(A) = |A|k , for some integer k. We can ﬁnd k by choosing A = aI. Since V = a2 W and there are 12 n(n+1) distinct elements, then J(V → W) = an(n+1) = |aI|n+1 . We found k = n + 1. 2 James (1954) also used MacDuﬀy’s characterization of the determinant for skew-symmetric matrices. At this point, we make some comments concerning the diﬀerential of a function of several variables. Our development here closely resembles that of Srivastava and Khatri (1979, p. 26). For n a real-valued nfunction y = f (x), x ∈ R , the diﬀerential is deﬁned as dy = df = i=1 ∂f (x)/∂xi · dxi . For a vector-valued function y = f (x), x and y in Rn , the diﬀerential is deﬁned componentwise, i.e., n df1 i=1 ∂f1 (x)/∂xi · dxi . .. dy = df = .. = = ∂f (x)/∂x · dx, . n dfn i=1 ∂fn (x)/∂xi · dxi where ∂f (x)/∂x = (∂fi (x)/∂xj ) ∈ Rnn is the usual derivative of f (x). Hence, dy is a linear function of dx with jacobian J(dy → dx) = |∂f (x)/∂x|+ = J(y → x).

2.13. Jacobians

31

Note that x and y could be replaced by any “vectorized” array or matrix. For example, for F(X) = (fij (X)) ∈ Rm n , we can deﬁne the diﬀerential componentwise, i.e., dF = (dfij ). The reader can then check (v. Problem 2.14.15) F = GH =⇒ dF = G · dH + dG · H. As an example, consider the inverse transformation. Proposition 2.20 The jacobian of the transformation V = W−1 , W ∈ −(n+1) . Rnn nonsingular and symmetric, is J(V → W) = |W|+ Proof. Since VW = I, then V · dW + dV · W = 0, which implies dV = −W−1 · dW · W−1 . Hence, from Proposition 2.19, J(V → W) = J(dV → −(n+1) dW) = |W|+ . 2 The jacobian of “conditional transformations” [Srivastava and Khatri (1979), p. 29], used to prove Propositions 2.22 and 2.23, may provide simpliﬁcations in some cases. Proposition 2.21 Let xi and yi in Rpi , i = 1, . . . , r, be related through the system of “conditional transformations” y1 = f1 (x1 ), y2 = f2 (y1 , x2 ), .. . yr = fr (y1 , . . . , yr−1 , xr ), where each fi is diﬀerentiable. Then, J(y1 , . . . , yr → x1 , . . . , xr ) =

r

J(yi → xi ).

i=1

Proof. The jacobian has the triangular form 0 ··· 0 ∂y1 /∂x1 0 ∗ ∂y2 /∂x2 · · · , J = . .. .. .. . . . . . ∗ ∗ · · · ∂yr /∂xr + r and, thus, we get J = i=1 |∂yi /∂xi |+ immediately.

2

As an example of jacobian via conditional transformations, consider the Bartlett decomposition of W > 0 in Rnn as W = TT for a unique T ∈ L+ n (v. Proposition 1.14). Due to symmetry, W has eﬀectively n(n + 1)/2 elements and, thus, the decomposition gives a transformation f : Rn(n+1)/2 → Rn(n+1)/2 deﬁned by f (W) = T.

32

2. Random vectors

Proposition 2.22 The jacobian of the transformation f (W) = T is J(W → T) = 2n

n

tn−i+1 . ii

i=1

Proof. Partition W and T in conformity so that w11 w21 t11 0 t11 = w21 W22 t21 T22 0

t21 T22

.

Observe the system of conditional transformations w11

= t211 ,

w21

= w11 t21 ,

W22

1/2

= w21 w21 /w11 + T22 T22

from which 1/2

J(W → T) = (2t11 )(w11 )n−1 J(W22 → T22 ) = 2tn11 J(W22 → T22 ). 2

The conclusion follows by induction.

As another example, consider the transformation to polar coordinates on Rn , x → (r, θ1 , . . . , θn−1 ) given by x1 x2 x3 .. . xn−1 xn

= r sin(θ1 ) sin(θ2 ) · · · sin(θn−2 ) sin(θn−1 ), = r sin(θ1 ) sin(θ2 ) · · · sin(θn−2 ) cos(θn−1 ), = r sin(θ1 ) sin(θ2 ) · · · cos(θn−1 ),

= r sin(θ1 ) cos(θ2 ), = r cos(θ1 ),

where r > 0 is the “radius” and 0 < θi ≤ π, i = 1, . . . , n − 2, 0 < θn−1 ≤ 2π are the “angles”. The jacobian J(x → r, θ) is facilitated with the system of conditional transformations y1 y2 y3 .. . yn−1 yn

= x21 + · · · + x2n−2 + x2n−1 + x2n = x21 + · · · + x2n−2 + x2n−1 = x21 + · · · + x2n−2

= r2 , = y1 sin2 (θ1 ), = y2 sin2 (θ2 ),

= x21 + x22 = x21

= yn−2 sin2 (θn−2 ), = yn−1 sin2 (θn−1 ).

Proposition 2.23 The jacobian of the transformation to polar coordinates in Rn is J(x → r, θ) = rn−1 sinn−2 (θ1 ) sinn−3 (θ2 ) · · · sin(θn−2 ).

2.14. Problems

33

Proof. We give the main idea and the reader is asked in Problem 2.14.11 to complete the details. We have J(y → r, θ) = J(y → x) · J(x → r, θ). The jacobian J(y → x) is trivial and J(y → r, θ) is evaluated using Proposition 2.21 on conditional transformations. 2 Let S n−1 = {s ∈ Rn : |s| = 1} be the “unit sphere” in Rn . The superscript n − 1 refers to the dimension of this surface. At times, we would like to bypass the angles and consider directly the transformation f : Rn \{0} → (0, ∞) × S n−1 , x → (r, u) deﬁned by r = |x| and u = x/|x| ∈ S n−1 . Since [Courant (1936), p. 302] R g(|x|)dx = g(r)rn−1 drdu, 0

|x|≤R

S n−1

where du is the “area element” of S n−1 , then rn−1 is the jacobian. Proposition 2.24 The jacobian of the transformation x → (r, u) is J(x → r, u) = rn−1 . The jacobians of other transformations on k-surfaces (manifolds) in Rn are useful for sampling distributions of eigenvalues, for example, but their full understanding requires a knowledge of diﬀerential forms and integration on manifolds [Spivak (1965), James (1954)]. This will not be pursued here.

2.14 Problems 1. Show that |E g(x)| ≤ E |g(x)| for any g : Rn → R such that E |g(x)| < ∞. 2. Prove the Cr inequality: For x and y distributed on Rk , E |x + y|r ≤ Cr [E |x|r + E |y|r ] , r > 0, where

Cr =

1, 0 0, a ≥ 0, b ≥ 0. 3. For each p > 0 let Lp denote the collection of all random vectors on Rk for which the pth moment exists: E |x|p < ∞. Prove the following basic facts: (i) Lp is a vector space. (ii) For any 0 < r ≤ p, Lp ⊆ Lr .

34

2. Random vectors

(iii) E |x|p < ∞ ⇐⇒ E |xi |p < ∞, i = 1, . . . , k ⇐⇒ E |t x|p < ∞, ∀t ∈ Rk . (iv) E |a x|p = 0, for some a ∈ Rk =⇒ P (x ∈ a⊥ ) = 1. (v) For any x ∈ L1 , |E x| ≤ E |x|. Indicate also the precise circumstances under which equality occurs. 4. Prove that if the pth moment (p > 0) of x ∈ Rn is ﬁnite, then all product-moments of x of order p are ﬁnite. 5. For x distributed on Rn , consider the p.d.f. fx (x) = c|x|2 · I[0,1] (x). (i) Determine c. (ii) Determine E xand var x. n i (iii) Determine E i=1 xi . Hint: E g(x) = cE |u|2 g(u), where u ∼ unif([0, 1]). p m n n 6. Let A ∈ Rm n , B ∈ Rq , and C ∈ Rq be constant and X ∈ Rp , x ∈ R , q and y, z ∈ R be random. Check the following:

(i) E(AXB + C) = A(E X)B + C (ii) cov(Ax, By) = A cov(x, y)B (iii) cov(x, y + z) = cov(x, y) + cov(x, z). 7. Prove the conditional variance formula. 8. Pairwise versus mutual independence [Bhat (1981)]. Let x and y be i.i.d. random variables taking the values +1 and −1 with probability 1/2 each. Deﬁne z = xy.

|=

|=

|=

i) Establish x, y, z are pairwise independent, but not mutually independent. ii) Does x z and y z imply xy z? |=

9. Let x = (x1 , x2 ) have the d.f. of Cook and Johnson (1981) as in x2 iﬀ α = 0. expression (2.3). Demontrate x1 10. Given a bivariate copula d.f. C(t1 , t2 ), two measures of association are Spearman’s ρ and Kendall’s τ , t1 t2 dC(t1 , t2 ) − 3, ρ = 12 [0,1]2 C(t1 , t2 )dC(t1 , t2 ) − 1, τ = 4 [0,1]2

|ρ| ≤ 1 and |τ | ≤ 1. Now, let |α| < 1/3 in the bivariate Morgenstern copula C(t1 , t2 ) = t1 t2 [1 + 3α(1 − t1 )(1 − t2 )]. Verify this copula is parameterized by Spearman’s measure, or α = 12 t1 t2 dC(t1 , t2 ) − 3. [0,1]2

2.14. Problems

35

11. Complete the proof of Proposition 2.23. 12. Demonstrate the jacobian of the transformation T1 = AT 2 ∈ 2nfor T + i L+ n and A = (aij ) constant also in Ln is J(T1 → T2 ) = i=1 aii . 13. Demonstrate the jacobian of the transformation U1 = AU2 for U2 ∈ + U+ n and A = (aij ) constant also in Un is J(U1 → U2 ) =

n

n−i+1 aii .

i=1

14. Demonstrate the jacobian of the transformation V = AWA , where W ∈ Rnn is skew-symmetric, i.e., W = −W , is J(V → W) = |A|n−1 + . 15. Suppose F(X) = (fij (X)) ∈ Rm n and deﬁne the diﬀerential componentwise, i.e., dF = (dfij ). Demonstrate that F = GH =⇒ dF = G · dH + dG · H. 16. Let

V=

V11 V21

V12 V22

>0

and deﬁne the transformation f : (V11 , V12 , V22 ) → (V11.2 , V12 , V22 ), −1 where V11.2 = V11 − V12 V22 V21 .

(i) Prove f deﬁnes a one-to-one mapping. (ii) Obtain J(V11 , V12 , V22 → V11.2 , V12 , V22 ) = 1.

3 Gamma, Dirichlet, and F distributions

3.1 Introduction This chapter introduces some basic probability distributions useful in statistics. The gamma distribution, in particular, is the building block of many other distributions such as chi-square, F , and Dirichlet. The Dirichlet distribution, as deﬁned in Section 3.3, has the important physical interpretation of proportion of time waited in a Poisson process. However, it has other applications such as the distribution of spacing variables (v. Problem 3.5.3) and the distribution theory (v. Section 4.5) related to spherical distributions, which play an important role in robustness.

3.2 Gamma distributions Deﬁnition 3.1 Standard gamma: z ∼ gamma(p) or z ∼ G(p) on p > 0 “degrees of freedom” iﬀ Γ(p)−1 z p−1 e−z , z > 0 fz (z) = 0, z ≤ 0. The integrating constant, as it depends on p > 0, is known as the gamma function and is, in fact, deﬁned by ∞ Γ(p) = tp−1 e−t dt, p > 0. 0

3.2. Gamma distributions

One may verify some basic properties: Γ(p + 1) = pΓ(p), Γ(2) = Γ(1) = 1, Γ( 12 ) =

√

37

π

and, in particular, Γ(n) = (n − 1)!. Obviously, E z r = Γ(p + r)/Γ(p), ∀r > −p, so that E z = var z = p. A more general gamma distribution is obtained by simply rescaling the standard gamma. Deﬁnition 3.2 Scaled gamma: x ∼ gamma(p, θ) or x ∼ G(p, θ) on p > 0 “degrees of freedom” and “scale” θ > 0 iﬀ x = θz, z ∼ G(p). Obviously,

fx (x) =

Γ(p)−1 θ−p xp−1 e−x/θ , 0,

x>0 x ≤ 0,

and, thus, the characteristic function is cx (t) = (1 − iθt)−p with the “convolution” of gamma distributions as a corollary. Corollary 3.1 If xi , i = 1, . . . , n, are independent G(pi , θ), then n n xi ∼ G pi , θ . i=1

i=1

In the special case where p = 1 we have the exponential distributions. Deﬁnition 3.3 Standard exponential : Scaled exponential :

z ∼ exp(1) iﬀ z ∼ G(1). x ∼ exp(θ) iﬀ x = θz, z ∼ exp(1).

The chi-square distribution is another special case. d

Deﬁnition 3.4 Chi-square: y ∼ χ2m or y = χ2m iﬀ y = 2z, z ∼ G( 12 m). Equivalently, y ∼ χ2m iﬀ y ∼ G( 12 m, 2), and the chi-square is a special case of the scaled gamma above. Thus, the gamma distribution occurs in common statistical practice as the chi-square (2×gamma≡chi-square). The characteristic function of y ∼ χ2m is immediate: cy (t) = (1 − i2t)−m/2 . One should, however, also recall how it describes “waiting time” in a Poisson process. Recall that the Poisson process Nt arises ﬁrst on purely physical considerations as a description of the number of “successes” in what is eﬀectively an inﬁnite number of independent bernoulli trials over the ﬁxed time period t where the average number is known to be proportional to t. On these

38

3. Gamma, Dirichlet, and F distributions

assumptions, Xn ∼ binomial(n, pn ) and E Xn → λt, whereby d

Xn → Nt , where Nt ∼ Poisson(λt). One then has the (conjugate) waiting time process Tn to describe the amount of time to wait until at least n “successes.” Since Tn > t ⇐⇒ Nt < n, we ﬁnd P (Tn > t) = P (Nt < n) =

n−1

e−λt (λt)i /i!

i=0

and diﬀerentiating produces the p.d.f. fn (t) = λe−λt (λt)n−1 /(n − 1)!, t > 0, whereby we discover Tn ∼ G(n, λ−1 ) or, equivalently, zn = λTn ∼ G(n). The exponential itself is just as well predicated on a diﬀerent intuition in that one may show that it is the unique distribution that has “no memory” in the explicit sense that x ∼ exp(θ) ⇐⇒ P (x > s + t|x > t) = P (x > s) > 0, ∀s, t > 0.

3.3 Dirichlet distributions If the gamma is intuitively a waiting time, the Dirichlet, otherwise known as the multivariate beta, is simply the proportion of time waited. Deﬁnition 3.5 Dirichlet: x ∼ Dn (p; pn+1 ) or x ∼ betan (p; pn+1 ), p = (p1 , . . . , pn ) , pi > 0, i = 1, . . . , n + 1 iﬀ d

x= with zi

1 z T

n+1 indep ∼ G(pi ), i = 1, . . . , n + 1, z = (z1 , . . . , zn ) , and T = i=1 zi .

The notation

indep ∼ means “independently distributed as.”

Proposition 3.1 The joint p.d.f. of x and T can be described as n+1 n+1 1 t i=1 pi −1 e−t , t > 0 i.e., T ∼ G , fT (t) = pi n+1 Γ( i=1 pi ) i=1

3.3. Dirichlet distributions

39

|=

pn+1 −1 n+1 n n Γ( i=1 pi ) pi −1 fx (x) = n+1 xi xi , x ∈ T n, 1− i=1 Γ(pi ) i=1 i=1 n where T n = {x ∈ Rn : xi > 0, T. i=1 xi < 1}. Moreover, x Proof. Using independence, the joint p.d.f. of the zi ’s is n+1 n+1 p −1 1 · zi i · exp − zi , zi > 0, ∀i. fz,zn+1 (z, zn+1 ) = n+1 i=1 Γ(pi ) i=1 i=1 We simply transform from (z1 , . . . , zn+1 ) to (x1 , . . . , xn , t), where zi = txi , i = 1, . . . , n and

zn+1 = t 1 −

n

xi

.

i=1

The jacobian is given by ∂z1 , . . . , zn+1 ∂x1 , . . . , xn , t +

∂tx1 , . . . , txn , t(1 − ni=1 xi ) = ∂x1 , . . . , xn , t + 0 x1 t .. .. . . = t x 0 nn −t · · · −t 1 − i=1 xi 0 x1 t .. .. . . = tn . = t xn 0 0 ··· 0 1

Thus, the joint p.d.f. of (x, T ) is n n n+1 1 xipi −1 (1 − xi )pn+1 −1 t i=1 pi −1 e−t , x ∈ T n , t > 0, n+1 i=1 Γ(pi ) i=1 i=1

and the conclusions are reached.

2

Note that the Dirichlet where all the parameters are 1 is simply the uniform distribution on the triangular region T n , Dn (1; 1) ≡ unif(T n ). Also, d

the Dirichlet distribution generalizes the beta distribution, D1 (p1 ; p2 ) = beta(p1 ; p2 ), with p.d.f. fx (x) = B(p1 , p2 )−1 xp1 −1 (1 − x)p2 −1 , 0 < x < 1, where B(p1 , p2 ) = Γ(p1 )Γ(p2 )/Γ(p1 + p2 ), is the beta function. The converse of Proposition 3.1 is almost obvious (by inverse change of variables); it need only be stated.

40

3. Gamma, Dirichlet, and F distributions

n i=1

xi )

|=

Proposition 3.2 If zi = T xi , i = 1, . . . , n and zn+1 = T (1 − n+1 T , then with T ∼ G( i=1 pi ), x ∼ Dn (p; pn+1 ), and x zi

indep ∼ G(pi ), i = 1, . . . , n + 1.

Four useful corollaries are also stated and the reader is asked to prove n+1 them. For x ∼ Dn (p; pn+1 ), let p = i=1 pi denote the “grand total,” noting that, by deﬁnition, d 1 x= z T n+1 indep with zi ∼ G(pi ), i = 1, . . . , n + 1, z = (z1 , . . . , zn ) , and T = i=1 zi . We ﬁnd the following: Corollary 3.2 (Marginal Dirichlet) If x1 = (xi1 , . . . , xik ) denotes any subset of the coordinates, then x1 ∼ Dk (p1 ; q) with p1 = (pi1 , . . . , pik ) and k p = q + j=1 pij . Corollary 3.3 If x = (x1 , . . . , xm ) is “partitioned” in any manner whatever so that we may write xi ∼ Dki (pi ; qi ), i = 1, . . . , m, deﬁne y by letting yi = xi 1, i.e., the total of the components of xi , with corresponding ri = pi 1. We ﬁnd y ∼ Dm (r; pn+1 ) with r = (r1 , . . . , rm ) . n Corollary 3.4 If S = x 1 = i=1 xi and again x1 = (xi1 , . . . , xik ) , d

k < n, is any subset, let w1 = S1 x1 . We ﬁnd w1 ∼ Dk (p1 ; r) with k p1 = (pi1 , . . . , pik ) as before but this time, p − pn+1 = r + j=1 pij . Corollary 3.5 (Conditional Dirichlet) If x = (x1 , x2 ) ∼ Dn ((p1 , p2 ) ; pn+1 ) , where x1 , p1 ∈ Rr and x2 , p2 ∈ Rs , n = r + s, then x1 | x2 ∼ Dr (p1 ; pn+1 ). 1 − x2 1

|=

We easily compute the moments of a Dirichlet distribution. By the converse representation in Proposition 3.2, if T ∼ G(p), x ∼ Dn (p; pn+1 ), and T , we have x d

T x = z with zi

indep ∼ G(pi ), i = 1, . . . , n.

This gives Tr

n

d

xri i =

i=1

n i=1

ziri with r =

n

ri ,

i=1

so that E Tr E

n i=1

xri i =

n i=1

E ziri for ri > −pi , i = 1, . . . , n.

3.3. Dirichlet distributions

41

6 4

1

2

0.8 0.6

0 0 0.2

0.4 0.4 0.2

0.6 0.8 10

Figure 3.1. Bivariate Dirichlet density for values of the parameters p1 = p2 = 1 and p3 = 2.

We ﬁnd E

n

n xri i =

i=1

E ziri . E Tr

i=1

In particular, E xi E x2i and E xi xj

= pi /p, ∀i, = (pi + 1)pi /(p + 1)p, ∀i, = pi pj /(p + 1)p, ∀i = j,

and letting θ = p1 p gives E x = θ and var x =

1 diag(θ) − θθ . p+1

Figure 3.1 exhibits a bivariate Dirichlet density. Various characterizations of Dirichlet distributions can be found in the literature [Rao and Sinha (1988), Gupta and Richards (1990)].

42

3. Gamma, Dirichlet, and F distributions

3.4 F distributions The ratio of two independent gammas is described by the F distribution, intuitively a relative waiting time. d y1 /s1 y2 /s2 ,

Deﬁnition 3.6 F distribution: F ∼ F (s1 , s2 ) iﬀ F = i = 1, 2.

yi

indep 2 ∼ χsi ,

One may easily obtain the moments of an F distribution and, in particular, its mean and variance. Sometimes, the distributions are more easily expressed in terms of the canonical Fc distribution: d

Deﬁnition 3.7 Canonical Fc distribution: F ∼ Fc (s1 , s2 ) iﬀ F = y1 /y2 , indep yi ∼ χ2si , i = 1, 2. One should also verify the simple relation F ∼ Fc (s1 , s2 ) ⇐⇒ (1 + F )−1 ∼ beta( 12 s2 ; 12 s1 ). The noncentral chi-square and F distributions useful to describe the nonnull distribution of some tests are deﬁned in Section 4.3.

3.5 Problems 1. If y ∼ χ2m , then E y h = 2h Γ

1

2m

+ h /Γ 12 m , h > − 12 m.

2. Prove Corollary 3.1. 3. Assume x ∼ unif([0, 1]) in Rn . (i) Deﬁne y by y1 = x(1) = min ({x1 , . . . , xn }) and

yi = x(i) = min {x1 , . . . , xn } − {x(1) , . . . , x(i−1) } , i = 2, . . . , n. Determine the distribution of y. (ii) Deﬁne z by z1 = y1 and zi = yi − yi−1 , i = 2, . . . , n, and determine the distribution of z. (iii) Determine E x, var x, E y, and var y as well as E z and var z. 4. Prove Corollaries 3.2, 3.3, 3.4, and 3.5. 5. Show the simple equivalence F ∼ Fc (s1 , s2 ) ⇐⇒ (1 + F )−1 ∼ beta( 12 s2 ; 12 s1 ). 6. Obtain the density of F ∼ Fc (s1 , s2 ):

Γ 12 (s1 + s2 ) F s1 /2−1 , F > 0. f (F ) = 1 1 Γ 2 s1 Γ 2 s2 (1 + F )(s1 +s2 )/2

4 Invariance

4.1 Introduction Invariance is a distributional property of a random vector acted upon by a group of transformations. The simplest group of transformations {+1, −1} leads to symmetric distributions by deﬁning a random variable to be symd metric iﬀ x = −x. Groups of transformations acting on random vectors commonly encountered are the permutations and orthogonal transformations. The permutation invariance gives the “exchangeable” random vectors and the invariance by orthogonal transformations deﬁnes the spherical distributions. Of great importance is the orthogonal group, since it speciﬁes the physical basis for normality in the Maxwell-Hershell theorem. Spherical distributions will play a central role later in Chapter 13 to build the elliptical models useful in the study of robustness.

4.2 Reﬂection symmetry d

Deﬁnition 4.1 x is (reﬂection) symmetric iﬀ x = −x. d

One immediately notes that if x = −x and E |x| < ∞, then E x = 0 (why?). The distribution of a symmetric random variable x is completely determined by the distribution of its modulus |x|, as the next proposition shows.

4. Invariance d

d

Proposition 4.1 x = −x ⇐⇒ x = s|x| with s

|=

44

|x|, s ∼ unif{±1}.

Proof. (=⇒) : Let F be the d.f. of x. Then, P (s|x| ≤ t)

1 2

{P (|x| ≤ t) + P (|x| ≥ −t)} P (−t ≤ x ≤ t) + 1, t ≥ 0 1 = 2· P (|x| ≥ −t), t 2, s1 (s2 − 2) s22 (s1 + 2δ)2 + (s1 + 4δ)(s2 − 2) , s2 > 4. 2 2 s1 (s2 − 2)2 (s2 − 4)

3. Obtain the density of F ∼ Fc (s1 , s2 ; δ) using Problem 3.5.6:

∞ k Γ 12 (s1 + s2 + 2k) F (s1 +2k)/2−1 −δ δ

1 1 f (F ) = e , k! Γ 2 (s1 + 2k) Γ 2 s2 (1 + F )(s1 +s2 +2k)/2 k=0 F > 0. 4. Assume x = (x1 , x2 ) has a spherical distribution. Show that x1 also has a spherical distribution. 5. Let x ∈ Rn have a spherical distribution with a ﬁnite rth moment. Demonstrate that all product-moments of x, E(xs11 · · · xsnn ), of order n s = i=1 si ≤ r are null provided one of the si is odd. 6. Let x = (x1 , . . . , xn ) have a spherical distribution. Prove the following: (i) cx (t) = cx1 (|t|) is a function of |t|. (ii) If x is absolutely continuous, x ∼ f , then f (x) = f (|x|e1 ) depends on x only through |x|.

4.6. Problems

53

7. Assume x ∈ Rn is rotationally invariant. Prove the mixture characterization ∞ cu1 (|t|r)dF (r), cx (t) = 0

where u = (u1 , . . . , un ) ∼ unif(S n−1 ) and F is the distribution function of |x| on [0, ∞). This means any rotationally invariant distribution is a mixture of uniform distributions on spheres of varying radius r ≥ 0 [Schoenberg (1938)]. 8. Let x ∼ unif(B n ), where B n = {s : s s ≤ 1} is the “unit ball” in Rn . (i) Deﬁne y by yi = x2i , i = 1, . . . , n, and determine the distribution of y, the marginal distribution of each yi , i = 1, . . . , n, and, ﬁnally, the distribution of R2 = |x|2 = x x. (ii) Obtain vol(B n ) using (i) and indicate the special cases n = 1, 2, 3. (iii) Determine E x and var x as well as E R2 and var R2 . Hint: Realize that y is “concentrated” on T n = {y : yi ≥ 0,

n

yi ≤ 1}.

i=1

9. Assume x ∈ Rn is permutationally invariant ∀n and E |x|2 < ∞. Let S = g(x), where g : Rn → R is any (permutation) symmetric function, i.e., g(Jt) = g(t), ∀J ∈ Sn , ∀t ∈ Rn . (i) Prove ρ ≥ 0. w.p.1 = E (f (Jx) | S) , ∀J ∈ Sn . w.p.1 ¯) ≤ 0. (iii) cov(x1 , x2 | x (ii) E (f (x) | S)

10. Assume x ∈ Rn has a “rotationally invariant” distribution such that d P (x = 0) = 0, i.e., Hx = x, ∀H ∈ On . Let R = |x| and z = x/R. |=

(i) Prove that z has the same distribution as if x had been unif(B n ). z. (ii) Prove that R (iii) Determine E z and var z. (iv) Determine E x and var x in terms of E R2 . Partition x = (x1 , x2 ) , x1 ∈ Rk and x2 ∈ Rn−k and let Ri = |xi |, i = 1, 2. (v) Determine the distribution of R12 /R2 and R12 /R22 . 11. Assume u ∼ unif(S n−1 ) and u = (u1 , u2 ) , u1 ∈ Rk . (i) Prove that the density of u1 is f (u1 ) =

Γ( 12 n) (1 − u1 u1 )(n−k)/2−1 , 0 < u1 u1 < 1. π k/2 Γ[ 12 (n − k)]

54

4. Invariance

Hint: Show (u21 , . . . , u2k ) ∼ Dk ( 12 1; 12 (n − k)) and consider the one-to-many transformation u2i → ±ui . (ii) Prove |u1 |2 ∼ beta( 12 k; 12 (n − k)). 12. Assume u = (u1 , u2 , u3 ) ∼ unif(S 2 ). Show that u1 ∼ unif(−1, 1). Does this hold in other dimensions?

|=

13. Let x = (x1 , . . . , xn ) have a spherical density fx (x) = g(|x|2 ) for some function g : [0, ∞) → [0, ∞). Let x = ru, where r ≥ 0 denotes “radius” and u ∈ S n−1 represents “direction.” Prove the following using J(x → r, u) = rn−1 : u. (i) r (ii) r2 has density fr2 (s) = 12 ωn sn/2−1 g(s), s > 0, where ωn is the “area” of the unit sphere S n−1 . (iii) With the special case x1 , . . . , xn i.i.d. N (0, 1), ﬁnd the “area” ωn . (iv) What is the density of u? 14. Let x ∈ Rn have a spherical density fx (x) = g(|x|2 ) and x → r, θ1 , . . . , θn−1 be the transformation to polar coordinates as in Proposition 2.23. Prove θn−1 ∼ unif(0, 2π). What can be said about the other angles? 15. Prove the following concerning spherical distributions: (i) If g(|x|2 ) is a density on Rn for some g : [0, ∞) → [0, ∞), then ∞ n−1 r g(r2 )dr = Γ 12 n /(2π n/2 ). 0 (ii) If the kth moment of x is ﬁnite, i.e., E |x|k < ∞, then ∞ rn+k−1 g(r2 )dr < ∞. 0

(iii) If the second moment of x is ﬁnite, then var x = αI, where α = E x21 . From Problem 4.6.6, cx (t) = φ(t t) for some function φ. Prove α = −2φ (0).

5 Multivariate normal

5.1 Introduction This chapter is entirely devoted to the multivariate normal distribution. In Section 5.2, the basic properties are demonstrated. Then, Sections 5.3 and 5.4 make the distinction between the nonsingular and the singular cases. In the nonsingular case, the density is derived while we explain the geometry of the singular case. Section 5.5 contains the conditional distribution in all its generality. Finally, the last section reaps the ﬁrst beneﬁts by considering some applications in univariate sampling, regression, and elementary correlation.

5.2 Deﬁnition and elementary properties Let Σ = (σij ) ∈ Rnn be symmetric, positive semideﬁnite, and µ ∈ Rn . Deﬁnition 5.1 Multivariate normal: x ∼ Nn (µ, Σ) iﬀ t x ∼ N (t µ, t Σt), ∀t ∈ Rn . Note that x has product-moments of any order by the fact that this is true of t x, ∀t ∈ Rn . Proposition 5.1 x ∼ Nn (µ, Σ) =⇒ E x = µ and var x = Σ.

56

5. Multivariate normal

Proof. Setting t = ei = (0, . . . , 1, . . . , 0) , we ﬁnd the individual component xi = ei x: xi ∼ N (µi , σii ), where µi = E xi , σii = var xi . Similarly, setting t to be a vector with 1’s in the ith and jth components, i = j, and 0’s elsewhere, we ﬁnd xi + xj ∼ N (µi + µj , σii + σjj + 2σij ), i = j. However, since on the other hand for i = j, var (xi +xj ) = var xi +var xj + 2 2cov(xi , xj ), then cov(xi , xj ) = σij . Proposition 5.2 Let A : Rn → Rm , linear, and x ∼ Nn (µ, Σ). Then, Ax ∼ Nm (Aµ, AΣA ). Proof. Let y = Ax and merely note that s y = (A s) x ∼ N (s Aµ, s AΣA s) , ∀s ∈ Rm . 2 By specializing A to be the projection onto any particular subset of coordinates, we deduce immediately that all the marginal distributions are normal. As a simple corollary on rotational invariance, we have z ∼ Nn (0, σ 2 I) =⇒ Hz = z, ∀H ∈ On . d

The characteristic function for x ∼ Nn (µ, Σ) derives from the univariate level (4.1):

cx (t) = ct x (1) = exp − 12 t Σt + it µ . Example 5.1 Although all marginals of x have a univariate normal distribution, the vector x itself may not have a multivariate normal distribution. Consider a random vector x whose distribution is a mixture of two multivariate normal distributions, cx (t) = αcx1 (t) + (1 − α)cx2 (t), 0 < α < 1, where x1 x2

∼ Nn (0, (1 − ρ1 )I + ρ1 11 ) , ∼ Nn (0, (1 − ρ2 )I + ρ2 11 ) .

Then, cxi (ti ) = αcz (ti ) + (1 − α)cz (ti ) = cz (ti ), where z ∼ N (0, 1), d

which shows xi = z, i = 1, . . . , n, but x does not have a multivariate normal distribution. Other counterexamples can be given using copulas [v. Example 2.5].

5.2. Deﬁnition and elementary properties

57

As a special case, we have the characteristic function for z ∼ Nn (0, σ 2 I): n n

exp(− 12 t2i σ 2 ) = czi (ti ) cz (t) = exp − 12 σ 2 t t = i=1

i=1

by which An implication is that if z = (z1 , z2 ) ∼ Nn (0, I), then z1 density for z becomes

|=

z ∼ Nn (0, σ 2 I) ⇐⇒ z1 , . . . , zn i.i.d. N (0, σ 2 ). z2 , and the

n

(2π)−1/2 exp − 12 zi2 . fz (z) = (2π)−n/2 exp − 12 z z = i=1

It is also clear from the characteristic function that the family of multivariate normal is closed under translation: x ∼ Nn (µ, Σ) =⇒ x + b ∼ Nn (µ + b, Σ), ∀b ∈ Rn . Now, suppose that x ∼ Nn (µ, Σ) and write Σ = HDH with H orthogonal and D = diag(λ). We ﬁnd, of course, that y = H (x − µ) ∼ Nn (0, D), and if we then let A = HD1/2 , we deduce the representation: d

Proposition 5.3 x ∼ Nn (µ, Σ) ⇐⇒ x = Az + µ for any A such that AA = Σ, z ∼ Nn (0, I). Finally, partition x = (x1 , x2 ) , where x1 ∈ Rn1 and x2 ∈ Rn2 , n = n1 +n2 , with corresponding µ1 Σ11 Σ12 µ= and Σ = . µ2 Σ21 Σ22 Concerning independence, we have the following necessary and suﬃcient condition.

x1

|=

Proposition 5.4 Let x = (x1 , x2 ) ∼ Nn (µ, Σ). Then, x2 ⇐⇒ Σ12 = 0.

x1

|=

Proof. (=⇒): x2

=⇒ E g1 (x1 )g2 (x2 ) = E g1 (x1 )E g2 (x2 ), ∀g1 , g2 =⇒ Σ12 = 0.

(⇐=): Assume Σ12 = 0. Write Σii = Aii Aii , i = 1, 2. Then, Σ = AA , where A11 0 A= . 0 A22 Using the representation A11 z1 + µ1 x1 d = Az + µ = , x2 A22 z2 + µ2

where z ∼ Nn (0, I), it is clear that since z1

z2 , then x1

|=

5. Multivariate normal

|=

58

x2 .

2

Another simple proof based on characteristic functions is proposed in Problem 5.7.3.

5.3 Nonsingular normal n When x ∼ Nn (µ, Σ) and |Σ| = |AA | = |A|2 = |D| = i=1 λi > 0, deﬁne z = A−1 (x−µ) whereby we have an explicit density for x by simple change of variables:

fx (x) = fz A−1 (x − µ) · J(z → x) = (2π)−n/2 |Σ|−1/2 exp − 12 (x − µ) Σ−1 (x − µ) and, of course, from Proposition 4.4, then also (x − µ) Σ−1 (x − µ) = z z ∼ χ2n . The quantity [(x − µ) Σ−1 (x − µ)]1/2 is often called the Mahalanobis distance of x to µ. Example 5.2 The bivariate density function is just a special case. For µ1 σ12 ρσ1 σ2 µ= , Σ= , µ2 ρσ1 σ2 σ22 we ﬁnd −1

Σ

1 = (1 − ρ2 )

σ1−2 −ρ/σ1 σ2

−ρ/σ1 σ2 σ2−2

.

Thus, the bivariate density takes the form fx (x1 , x2 )

=

2 x1 − µ1 1 1 1 exp − 2π σ1 σ2 (1 − ρ2 )1/2 2(1 − ρ2 ) σ1 2 x1 − µ1 x2 − µ2 x2 − µ2 −2ρ . + σ1 σ2 σ2

A plot of this density is given in Figure 5.1. The contours, which consists of the set of points of equal probability density, of a multivariate normal are the points x of equal Mahalanobis distance to µ, (x − µ) Σ−1 (x − µ) = c2 , for any constant c > 0. Letting y = H x, ν = H µ, where H diagonalizes Σ, H ΣH = D, then the contours are the ellipsoids p i=1

(yi − νi )2 /di = c2

5.3. Nonsingular normal

0.2 0.15 0.1 0.05 0

59

2 1 0

-2 -1

-1

0 1

-2 2

Figure 5.1. Bivariate normal density for values of the parameters µ1 = µ2 = 0, σ1 = σ2 = 1, and ρ = 0.7. 1/2

centered at ν with principal axes of half length cdi eigenvectors in H = (h1 , . . . , hp ).

supported by the

Example 5.3 The contours of the bivariate normal density are in parametric form, and in the y coordinates, 1/2 d1 sin θ y1 ν1 , 0 ≤ θ ≤ 2π. = +c 1/2 y2 ν2 d2 cos θ Thus, the contours in the original x coordinates are just 1/2 1/2 h11 d1 sin θ + h12 d2 cos θ x1 µ1 , 0 ≤ θ ≤ 2π. = +c 1/2 1/2 x2 µ2 h21 d1 sin θ + h22 d2 cos θ A contour plot is given in Figure 5.2. Example 5.4 Using the transformation to polar coordinates on p. 32, the contours of the trivariate normal density are in parametric form, and in the y coordinates, 1/2 d1 sin θ1 sin θ2 y1 ν1 1/2 y2 = ν2 + c d2 sin θ1 cos θ2 , 0 ≤ θ1 ≤ π, 0 ≤ θ2 ≤ 2π. 1/2 y3 ν3 d cos θ1 3

60

5. Multivariate normal

x2 3 2 1 -3

-2

-1

1

2

3

x1

-1 -2 -3 Figure 5.2. Contours of the bivariate normal density for values of the parameters µ1 = µ2 = 0, σ1 = σ2 = 1, and ρ = 0.7. Values of c = 1, 2, 3 were taken.

Thus, the contours in the original x1 µ1 h11 x2 = µ2 + c h21 x3 µ3 h31

x coordinates are just 1/2 d1 sin θ1 sin θ2 h12 h13 h22 h23 d21/2 sin θ1 cos θ2 , 1/2 h32 h33 d3 cos θ1

0 ≤ θ1 ≤ π, 0 ≤ θ2 ≤ 2π. The contour plot corresponding to c = 1 is given in Figure 5.3 when µ = 0 and 13 −4 2 Σ = −4 13 −2 . 2 −2 10 The corresponding eigenvalues of d1 = 18, d2 = d3 = 9 give the typical ellipsoidal contours. Still assuming |Σ| > 0, we apply the Gram-Schmidt process to the basis formed by the row vectors of A = HD1/2 , obtaining (uniquely) A = TG (v. Proposition 1.13) with T ∈ L+ n , G ∈ On , where L+ n = {T ∈ Gn : T is lower triangular, tii > 0, i = 1, . . . , n}. Then, Σ = AA = TT for a unique T ∈ L+ n (v. Proposition 1.14). We have the “triangular” representation: Proposition 5.5 (Triangular representation) d

x ∼ Nn (µ, Σ) ⇐⇒ x = Tz + µ with T ∈ L+ n such that Σ = TT and z ∼ Nn (0, I).

5.3. Nonsingular normal

2 0 -2

2

0

-2

-2 0 2 Figure 5.3. A contour of a trivariate normal density.

61

62

5. Multivariate normal

5.4 Singular normal Now, for x ∼ N (µ, σ 2 ), we know that σ 2 = 0 ⇐⇒ x = µ w.p.1. This holds, since if σ 2 = 0, P (x = µ) = limn→∞ P (|x − µ| < 1/n), but P (|x − µ| ≥ 1/n) ≤ σ 2 n2 = 0, ∀n. Thus, the normal family includes the “trivial” (constant) random variables as special cases. By Cram´er-Wold Proposition 2.10, this also holds for random vectors x ∈ Rn with E x = µ and var x = Σ: Σ = 0 ⇐⇒ x = µ w.p.1. However, if x ∼ Nn (µ, Σ) with |Σ| = 0, we may write D1 0 H1 Σ = HDH = (H1 , H2 ) = H1 D1 H1 , H2 0 0 where D1 = diag(λ1 , . . . , λr ) comprises the nonzero eigenvalues, H1 = (h1 , . . . , hr ) gives a basis for the column space of Σ, Im Σ, and H2 = (hr+1 , . . . , hn ) gives a basis for the kernel, ker Σ. One should note that Im H2 Im H1

=

(Im H1 )⊥ = ker Σ,

=

(Im H2 )⊥ = Im Σ.

H2 (x − µ)

|=

Then it is clear that H2 ΣH2 = 0 and, thus, we ﬁnd that H2 (x − µ) = 0 w.p.1 or, equivalently, x − µ ∈ (Im H2 )⊥ w.p.1., whereas H1 (x − µ) ∼ Nr (0, D1 ) has a nonsingular normal distribution. Of course, this is yet equivalent to saying that x ∈ µ + Im Σ w.p.1 and one can then almost visualize x in this r-dimensional aﬃne subspace of Rn , r < n. A curious fact in this case is that vol(µ + Im Σ) = 0 but Px (µ + Im Σ) = 1, therefore x cannot be absolutely continuous (v. Proposition 2.11). It is worth recalling at this point that any constant random vector is automatically statistically independent of any other random vector, and so we might notice, in particular, the rather odd looking fact that x.

5.5 Conditional normal By a suitable permutation, one may rearrange an arbitrary multivariate normal x so that any subset x1 of its coordinates are brought to the fore, and the overall distribution is expressed by µ1 Σ11 Σ12 x1 ∼ Nn , , x1 ∈ Rn1 , x2 ∈ Rn2 , n = n1 + n2 . x2 µ2 Σ21 Σ22 We derive the conditional distribution of x1 given x2 . First suppose Σ22 is nonsingular and note that for any B, I 0 I −B Σ11 Σ12 −B I Σ21 Σ22 0 I

5.5. Conditional normal

=

Σ11 − Σ12 B − BΣ21 + BΣ22 B Σ21 − Σ22 B

Σ12 − BΣ22 Σ22

so by deliberately setting B = Σ12 Σ−1 22 , we ﬁnd µ1 − Bµ2 Σ11.2 x1 − Bx2 ∼ Nn , x2 µ2 0

0 Σ22

63

,

,

where Σ11.2 = Σ11 − Σ12 Σ−1 22 Σ21 . d

However, independence means x1 − Bx2 | x2 = x1 − Bx2 , so that we have x1 − Bx2 | x2 ∼ Nn1 (µ1 − Bµ2 , Σ11.2 ). Since we may legitimately treat x2 as though constant (the full justiﬁcation of this depending on the fact that we have a “regular” conditional distribution to which the Fubini theorem applies [Ash (1972)]) we may conclude that x1 | x2 ∼ Nn1 (µ1 + Σ12 Σ−1 22 (x2 − µ2 ), Σ11.2 ). In the singular case, if |Σ22 | = 0, we may always write D 0 H1 = H1 DH1 , Σ22 = (H1 , H2 ) 0 0 H2 where D ∈ Rkk is nonsingular, and we may then take what is called a “pseudo-inverse” for Σ22 : −1 D 0 H1 = H1 D−1 H1 . = (H , H ) Σ− 1 2 22 H2 0 0 w.p.1 We then have, of course, H2 (x2 − µ2 ) = 0 and also µ1 Σ11 x1 Σ12 H1 ∼ N , n1 +k H1 x2 H1 µ2 H1 Σ21 D to which the results in the nonsingular case apply, immediately showing µ1 − Bµ2 Σ11.2 0 x1 − Bx2 ∼ N , , n1 +k H1 x2 H1 µ2 0 D |=

|=

− where B = Σ12 Σ− 22 and Σ11.2 = Σ11 − Σ12 Σ22 Σ21 . But from this, H1 x2 , and thus, overall, x1 − Bx2 x2 . We arrive at the x1 − Bx2 completely general conclusion:

Proposition 5.6 x1 | x2 ∼ Nn1 (µ1.2 , Σ11.2 ), where µ1.2 Σ11.2

= µ1 + Σ12 Σ− 22 (x2 − µ2 ), = Σ11 − Σ12 Σ− 22 Σ21 .

64

5. Multivariate normal

5.6 Elementary applications 5.6.1

Sampling the univariate normal

Observe that x1 , . . . , xn i.i.d. N (µ, σ 2 ) ⇐⇒ x ∼ Nn (µ1, σ 2 I). √ Letting H = ( 1/ n, Γ ) ∈ On for some Γ and √ w = H x ∼ Nn ( nµe1 , σ 2 I), obviously w1 , . . . , wn are independent, and, of course, √ x w1 = n¯ and

d

with its trivial algebraic corollary √ d n(¯ x − µ)/sx = (n − 1)z/χn−1 , z ∼ N (0, 1), and z

s2x

|=

¯ x ¯ ∼ N (µ, σ 2 /n), (n − 1)s2x = σ 2 χ2n−1 , and x

|=

w22 + · · · + wn2 = |w|2 − w12 = |x|2 − n¯ x2 = |x − x ¯1|2 = (n − 1)s2x , n ¯)2 /(n − 1) is the sample variance. Thus, we have where s2x = i=1 (xi − x the basic statistical result

χ2n−1 .

|=

We make the following deﬁnition (W.S. Gosset, “Student,” 1908): d d √ Deﬁnition 5.2 t-Distribution: t = tp iﬀ t = pz/χp , where z ∼ N (0, 1) and z χ2p . Thus, by deﬁnition, d

√

d

n(¯ x−µ)/sx = tn−1 is a pivotal quantity for µ. Clearly,

d

t = tp ⇐⇒ t = −t and t2 ∼ F (1, p). This provides a quick way of obtaining the integral moments of tp . The Student’s t-distribution sometimes plays a role in the dependent case. The intraclass correlation model is one such example.

Example 5.5 Assume x ∼ Nn µ1, σ 2 [(1− ρ)I + ρ11 ] , where −1/(n − n n 1) ≤ ρ ≤ 1. Let x ¯ = i=1 xi /n, s2x = i=1 (xi − x ¯)2 /(n − 1), and t = √ n(¯ x − µ)/sx . We determine a constant c such that ct ∼ tn−1 . With the orthogonal transformation above, we still have w w1 w22 + · · · + wn2

= H x, √ = n¯ x, =

(n − 1)s2x .

√ Since H 1 = ( n, 0 ) , the distribution of w is √ nµ w ∼ Nn , σ 2 [(1 − ρ)I + ρ diag(n, 0, . . . , 0)] . 0

5.6. Elementary applications

65

|=

|=

Hence, w1 (w2 , . . . , wn ) , which implies x ¯ s2x . The distribution of x ¯ 2 and sx are given by √ √ n¯ x ∼ N ( nµ, σ 2 [(1 − ρ) + ρn]), (n − 1)s2x ∼ σ 2 (1 − ρ)χ2n−1 .

Finally, we can conclude that ct ∼ tn−1 by deﬁning 1/2 1−ρ c= . (1 − ρ) + ρn In fact, the Student’s t-distribution has nothing to do with normal distributions. It is more related to the concept of spherical symmetry, as in the next example [Efron (1969)].

d

the representation x = Ru, where u ∼ unif(S n−1 ) and R

|=

invariant” distribution Example 5.6 Assume x ∈ Rn has a “rotationally √ n¯ x /s ∼ tn−1 , where, as usual, and P (x = 0) = 0. We establish that x n n ¯)2 /(n − 1). Using Proposition 4.10, x ¯ = i=1 xi /n and s2x = i=1 (xi − x u, is valid.

d

Hence, (¯ x, sx ) = (R¯ u, Rsu ) and the distribution of √ x √ u ¯ d √ R¯ u ¯ n = n = n sx Rsu su √ does not depend on R. Thus, n¯ x/sx ∼ tn−1 since this is the case when x ∼ Nn (0, I).

5.6.2

Linear estimation

Consider now the problem of linear estimation in the so-called multiple regression model. Let V ⊂ Rn be any k-dimensional vector subspace and y = µ + e, E e = 0, var e = σ 2 I, and µ ∈ V. ˆ ˆ where µ ˆ= Let θ = Tµ, where T ∈ Rm n , and consider the estimate θ = Tµ, Py is the orthogonal projection of y on V (v. Section 1.6). We prove that among all possible unbiased linear estimates of θ, the regression estimate ˆ has the minimum variance. In this sense, θ ˆ is the “best” linear unbiased θ estimate (blue). ˆ = blue(θ). Proposition 5.7 (Gauss-Markov) θ ˜ = By is unbiased for θ ⇐⇒ BP = TP. But then, Proof. θ ˆ = σ 2 TPT = σ 2 BPB ≤ σ 2 BB = var θ ˜ var θ with equality iﬀ ˜ = θ, ˆ BQB = 0 ⇐⇒ BQ = 0 ⇐⇒ B = BP ⇐⇒ θ where Q = I − P.

2

66

5. Multivariate normal

ˆ = Py = blue(µ) with var µ ˆ = σ 2 P. For example, µ Now, expressing µ = Xβ with respect to any basis X = (x1 , . . . , xk ) for V and recalling the representation (1.3) for P, the coeﬃcients are uniquely ˆ = B0 µ ˆ = determined as β = B0 µ, where B0 = (X X)−1 X . But then, β 2 −1 ˆ ˆ B0 y = blue(β), where var β = σ (X X) . Obviously, β i = blue(β i ), i = 1, . . . , k. Another optimality property of the “Gauss-Markov” estimate was recently discovered [Berk and Hwang (1989), Eaton (1988), Ali and Ponnapalli (1990]: The probability of the Gauss-Markov estimate of θ falling inside any ﬁxed ellipsoid centered at θ is greater than or equal to the probability that any linear unbiased estimate of θ falls inside the same elˆ = Py lipsoid. It is interesting to remark that the Gauss-Markov estimate µ is also the least-squares estimate. This follows from a general property of orthogonal projections: Proposition 5.8 ˆ 2, min |y − µ|2 = |y − µ| µ∈V ˆ = Py is the orthogonal projection of y on V. where µ Proof. For all µ ∈ V, |y − µ|2

ˆ + (µ ˆ − µ)|2 = |(y − µ) ˆ 2 + |µ ˆ − µ|2 + 2(y − µ) ˆ (µ ˆ − µ) = |y − µ| 2 2 ˆ − µ| ˆ + |µ = |y − µ|

ˆ ∈ V ⊥ and µ ˆ − µ ∈ V. Hence, since y − µ ˆ 2 , ∀µ ∈ V, |y − µ|2 ≥ |y − µ| ˆ with equality if µ = µ.

2

Finally, since Q = I − P gives the orthogonal projection on V ⊥ , we ﬁnd ˆ = Qy = Qe =⇒ |y − µ| ˆ 2 = e Qe, y−µ so that ˆ 2 = E e Qe = E tr Qee = tr QE ee = (n − k)σ 2 . E |y − µ| Thus, we determine the unbiased estimate sˆ2 of σ 2 by ˆ 2. (n − k)ˆ s2 = |y − µ| ˆ y − µ) ˆ = cov(Py, Qy) = σ 2 PQ = 0. Before It is also clear that cov(µ, ˆ and sˆ2 , stating the joint distribution under normality of our estimates µ we prove the following lemma on quadratic forms. Lemma 5.1 Let z ∼ Nn (µ, I) and Q ∈ Rnn be an orthogonal projection of rank Q = m. Then, z Qz ∼ χ2m (δ), where δ = µ Qµ/2.

5.6. Elementary applications

67

Proof. Let H = (h1 , . . . , hm ) be an orthonormal basis for Im Q and write Q = HH , where H H = Im . Then, z Qz = (H z) (H z) = |e|2 , where 2 e = H z ∼ Nm (H µ, I). Hence, |e|2 ∼ χ2m (δ) with δ = |H µ|2 /2. If, in addition, we assume normality, y = Xβ + e, e ∼ Nn (0, σ 2 I),

we have the general result

ˆ ∼ Nk β, σ 2 (X X)−1 , (n − k)ˆ ˆ β s2 ∼ σ 2 χ2n−k , and β

ˆ y − µ,

|=

ˆ = B0 y = B0 µ, ˆ (n − k)ˆ ˆ 2 = e Qe, and µ ˆ β s2 = |y − µ|

|=

then since

s

with corollary ˆ 2 |X(β − β)| ∼ F (k, n − k). 2 kˆ s We close this section with a slight generalization of Lemma 5.1. Corollary 5.1 Assume x ∼ Nn (µ, Σ), Σ > 0, and A is symmetric such that AΣA = A and rank ΣA = m. Then, x Ax ∼ χ2m (δ), where δ = µ Aµ/2. Proof. Letting z = Σ−1/2 x and B = Σ1/2 AΣ1/2 , then x Ax = z Bz, where z ∼ Nn (Σ−1/2 µ, I), and the conclusion follows from Lemma 5.1 since B is an orthogonal projection of rank m. 2

5.6.3

Simple correlation

Let (xi , yi ) i.i.d. (x, y), i = 1, . . . , n, be any “bivariate” sample. The correlation coeﬃcient ρ

= =

cor(x, y) cov(x, y) var(x) var(y)

is usually estimated by the sample correlation coeﬃcient n ¯)(yi − y¯) i=1 (xi − x r = n 1/2 n 1/2 2 [ i=1 (xi − x ¯) ] [ i=1 (yi − y¯)2 ] (x − x ¯1) (y − y¯1) . = |x − x ¯1| |y − y¯1| Note that r is just the cosine of the angle between the residual vectors x−x ¯1 and y − y¯1. The main (nonparametric) reason for using r as an w.p.1 estimate of ρ is its (strong) consistency: r → ρ as n → ∞.

68

5. Multivariate normal

Now, suppose that ρ = 0 so that x r=

|=

If, in addition, we assume normality, a “pivotal statistic” may be derived. First, notice that since r is invariant with respect to relocation and rescaling in both x and y, we may suppose at the outset that 0 1 ρ x ∼ N2 , . 0 ρ 1 y y. Then,

(Qx) (Qy) (x − x ¯1) (y − y¯1) , = |x − x ¯1| |y − y¯1| |Qx||Qy|

where Q = In − n−1 11 is an orthogonal projection of rank Q = n − 1. We can write (v. Section 1.6) Q = HH with H H = In−1 . Then, r=

z w (H x) (H y) , = |H x||H y| |z||w|

where z = H x and w = H y are independent Nn−1 (0, I). Finally, letting |=

d

u = z/|z| and v = w/|w|, we have u v, and from Corollary 4.3, u = v ∼ unif(S n−2 ). Therefore, using Proposition 4.8, r = u v = u1 = d

z1 . |z|

Thus, √

z1 r d = 1/2 , 2 2 1−r z22 + · · · + zn−1

where the zi ’s are i.i.d. N (0, 1) and √ r d n − 2√ = tn−2 . 1 − r2 We have proved: Proposition 5.9 If (xi , yi ) , i = 1, . . . , n, are i.i.d. as a bivariate normal with ρ = 0, then √ r d n − 2√ = tn−2 . 1 − r2 However, if ρ = 0 and 2 0 σ x , ∼ N2 0 ρστ y

ρστ τ2

,

then we may apply this result to the linear transformation 0 1 − ρ2 0 x/σ − ρy/τ ∼ N2 , 0 0 1 y/τ

5.7. Problems

using

69

n

r˜ =

((xi − x ¯)/σ − ρ(yi − y¯)/τ ) (yi − y¯) . !1/2 1/2 2 n 2 ¯)/σ − ρ(yi − y¯)/τ ) [ i=1 (yi − y¯) ] i=1 ((xi − x i=1

n

We ﬁnd √

r˜ z1 d = 2 , 2 2 (z2 + · · · + zn−1 )1/2 1 − r˜

where the zi ’s are i.i.d. N (0, 1) and, by direct computation, √

r − ρc sy /τ r˜ . =√ , where c = 2 2 s 1 − r˜ 1−r x /σ

Thus, we obtain the result √

(r − ρc) d n − 2√ = tn−2 . 1 − r2

This is actually a pivotal for β = ρσ/τ . Later, the reader will be able to prove that √ d n(r − ρ) → (1 − ρ2 )z, z ∼ N (0, 1) (v. Problem 6.4.8), which can be used to obtain an approximate conﬁdence interval for ρ. The exact distribution of r is treated in Section 8.4 in the more general context of multiple correlation coeﬃcient.

5.7 Problems 1. Plot the contours of the N2 (µ, Σ) distribution when 1 µ = , 2 2 1 Σ = . 1 4 |=

y and consider 2. Let x ∼ Nn (µ1, σ 2 I), y ∼ Nn (ν1, τ 2 I), and x n ¯)(yi − y¯) i=1 (xi − x . r = n 1/2 n 1/2 2 [ i=1 (xi − x ¯) ] [ i=1 (yi − y¯)2 ] (i) Determine the distribution of r. (ii) Determine E r and var r. 3. Prove Proposition 5.4 with characteristic functions. 4. Obtain the integral moments of the tp distribution.

70

5. Multivariate normal

5. Let x be such that E x = µ and var x = Σ. Show that min E |x − c|2 = tr Σ c

and that the minimum is attained at c = µ. 6. Assume

x1 x2

∼ Nn

µ1 µ2

Σ11 , Σ21

Σ12 Σ22

.

Demonstrate that min E |x1 − (Cx2 + d)|2 = tr Σ11.2 C,d

− is attained at C = Σ12 Σ− 22 and d = µ1 − Σ12 Σ22 µ2 . n 2 7. Assume that z ∼ Nn (0, I) and let z¯ = i=1 zi /n, (n − 1)s = n 2 ¯) . i=1 (zi − z

(i) Prove z¯, s, (z1 − z¯)/s are mutually independent. (ii) Determine the distribution of (z1 − z¯)/s. √ Hint: Let H = (1/ n, (e1 − n−1 1)/ (n − 1)/n, Γ) ∈ On , for some matrix Γ, w = H z, and note that (w2 , . . . , wn ) is rotationally invariant. 8. Assume y = Xβ + e, e ∼ Nn (0, σ 2 I), where, as usual, the columns of X ∈ Rnk are linearly independent and let C ∈ Rrk be of rank r. Show that ˆ − d) [C(X X)−1 C ]−1 (Cβ ˆ − d) (Cβ ∼ F (r, n − k; δ), rˆ s2 where δ=

(Cβ − d) [C(X X)−1 C ]−1 (Cβ − d) . 2σ 2

9. Let x ∼ Nn (µ1, σ 2 I). (i) Assume y is ﬁxed, y ∈ span{1}. Find the distribution of r=

(x − x ¯1) (y − y¯1) . |x − x ¯1| |y − y¯1|

(ii) This time assume y has any distribution satisfying

and x

|=

P (y ∈ span{1}) = 1 y, and determine the distribution of r.

10. Angular gaussian distribution. The angular gaussian distribution is obtained by the projection of x ∼ Nn (0, Λ) onto the unit sphere S n−1 ; i.e., the angular gaussian density is that of u = x/|x|.

5.7. Problems

71

(i) Prove that the angular gaussian density is f (u) =

Γ( 12 n) −1/2 −1 −n/2 |Λ| (u Λ u) , u ∈ S n−1 . 2π n/2

(ii) What is the special case Λ = I? (iii) Prove that the angular gaussian distribution can also be obtained by projecting (onto S n−1 ) x with density fx (x) = |Λ|−1/2 g(x Λ−1 x). The word gaussian is misleading here; symmetry is the key. 11. Rotationally symmetric distributions on spheres [Saw (1978)]. This class of distributions will be those for which the density is constant on those points u ∈ S n−1 satisfying u θ = δ, ∀δ ∈ [−1, 1] and some ﬁxed θ ∈ S n−1 . (i) For some ﬁxed λ ≥ 0, consider the function g(λ, ·) : [−1, 1] → [0, ∞). Prove 1 (1 − t2 )(n−3)/2 ωn−1 g(λ, u θ)du = g(λ, t) 1 1 dt, B( 2 , 2 (n − 1)) −1 S n−1 where B(·, ·) denotes the beta function and ωn = 2π n/2 /Γ( 12 n) is the “area” of S n−1 . Hint: ωn−1 g(λ, u θ)du = E g(λ, u θ) = E g(λ, u1 ), S n−1

where u = (u1 , . . . , un ) ∼ unif(S n−1 ) and use Problem 4.6.11. (ii) Deduce that f (u) = ωn−1 g(λ, u θ) is a density on S n−1 if 1 (1 − t2 )(n−3)/2 g(λ, t) 1 1 dt = 1. B( 2 , 2 (n − 1)) −1 Denote this distribution u ∼ Gn (λ, θ). (iii) What are the “contours” of a Gn (λ, θ) distribution? (iv) If g(λ, t) is an increasing function of t, prove Gn (λ, θ) is unimodal. What is the mode? (v) Prove: u ∼ Gn (λ, θ) =⇒ Hu ∼ Gn (λ, Hθ), ∀H ∈ On . (vi) Obtain the ﬁrst two moments of u ∼ Gn (λ, θ), E u = ρ1 θ,

= {(1 − ρ2 )I + (nρ2 − 1)θθ }/(n − 1),

E uu where

ρi =

1

−1

ti g(λ, t)

(1 − t2 )(n−3)/2 dt < ∞, i = 1, 2. B( 12 , 12 (n − 1))

5. Multivariate normal d

Hint: Use the representation u = tθ +(1−t2 )1/2 ζ, where t = u θ and ζ is distributed uniformly on the sphere orthogonal to θ, ζ [Watson (1983), p. 44]. t (vii) Prove

fx (x) = g(λ, θ x/|x|)(2π)−n/2 exp − 12 x x |=

72

is a density on Rn by transforming to polar coordinates x → (r, u), r ≥ 0, u ∈ S n−1 . (viii) Demonstrate that the distribution Gn (λ, θ) can be obtained by projecting the distribution for x ∼ fx onto S n−1 ; i.e., if x ∼ fx , then u = x/|x| ∼ Gn (λ, θ). Remark: The very special case g(λ, t) = exp(λt) yields the Langevin distribution also known, for n = 2 and 3, as the Fisher-von Mises distribution on the circle and sphere [Fisher (1953), von Mises (1918)]. Tests for the mean direction, θ, of the Langevin distribution are discussed by Fujikoshi and Watamori (1992). Robust estimators of (λ, θ) for the Langevin distribution include the circular median [Mardia (1972)], the normalized spatial median [Ducharme and Milasevic (1987)], and the M-estimator on spheres [Ko and Chang (1993)]. Goodness-of-ﬁt for directional data using smooth tests was considered by Boulerice and Ducharme (1997). Asymptotic behavior of sample mean direction on spheres, without symmetry condition on the p.d.f., was recently derived by Hendriks at al. (1996).

6 Multivariate sampling

6.1 Introduction The basic tools for manipulating random samples from a multivariate distribution are developed in this chapter. We introduce random matrices in Section 6.2 and show the usefulness of the “vec operator” and Kronecker product in this regard. Also, the matrix variate normal distribution is deﬁned and its basic properties are explained. Section 6.3 deals with theorems in the “asymptotic world” as the sample size goes to inﬁnity. These are the central limit theorem, a general Slutsky theorem, and the so-called delta method.

6.2 Random matrices and multivariate sample For A = (aij ) = (a1 , . . . , aq ) ∈ Rpq , we may always regard A as a vector in Rpq where we deﬁne a1 .. . vec(A) = . aq This operation is obviously linear Rpq → Rpq and we may regard A and vec(A) as synonymous.

74

6. Multivariate sampling

For

x1 X = (xij ) = ...

xp

random on Rpq , we may denote the mean of X by E X = M = (µij ). However, the variance of X is a quadruply indexed array consisting of all covariances of the individual entries var X = (σijkl ) = (cov(xij , xkl )) . Since there is no inherent order to this array, we ﬁnd it convenient to impose one by equating var X = var vec(X ) = Ω = (Ωij ) = (cov(xi , xj )) . The element in position (k, l) of the block Ωij is cov(xik , xjl ). One must be very careful to remember that Ω is pq × pq. For instance, if we write X ∼ Nqp (M, Ω), we really mean that vec(X ) ∼ Npq (vec(M ), Ω). In fact, this will be the deﬁnition. Moments of a multivariate normal matrix, Nqp (M, Ω), were given by Wong and Liu (1994). Characterization of a multivariate normal matrix distribution via conditioning is discussed by Gupta and Varga (1992) and Nguyen (1997). The Kronecker product will be very handy for manipulating random matrices. The Kronecker product of A ∈ Rpq and B ∈ Rrs is a block-matrix with the block in position (i, j) being aij B, A ⊗ B = (aij B) ∈ Rpr qs . One can verify the basic properties. Lemma 6.1 The Kronecker product satisﬁes the following: (i) (aA) ⊗ (bB) = ab(A ⊗ B), a, b ∈ R (ii) (A + B) ⊗ C = (A ⊗ C) + (B ⊗ C) (iii) (A ⊗ B) ⊗ C = A ⊗ (B ⊗ C) (iv) (A ⊗ B) = A ⊗ B , (v) (AB) ⊗ (CD) = (A ⊗ C)(B ⊗ D) (vi) (A ⊗ B)−1 = A−1 ⊗ B−1 , whenever A and B are nonsingular. (vii) If v = 0 and u = 0 are eigenvectors of A and B, respectively, Av = λv, and Bu = γu, then v⊗u is an eigenvector of A⊗B corresponding to the eigenvalue λγ. (viii) tr (A ⊗ B) = (tr A)(tr B) (ix) |A ⊗ B| = |A|q |B|p , A ∈ Rpp , B ∈ Rqq

6.2. Random matrices and multivariate sample

75

(x) If A > 0 and B > 0, then A ⊗ B > 0. The following lemma will also be useful for handling random matrices. Its proof is left as an exercise. Lemma 6.2 A ∈ Rrp , X ∈ Rpq , and B ∈ Rqs =⇒ vec(AXB) = (B ⊗ A)vec(X). As a corollary useful for densities (v. Problem 6.4.4) we also have: Corollary 6.1 Let A ∈ Rpp , X ∈ Rpq , and B ∈ Rqq . If Y = AXB, then J(Y → X) = |A|q+ |B|p+ . Proof. Since vec(Y) = vec(AXB) = (B ⊗ A)vec(X), then J(Y → X) = J(vec(Y) → vec(X)) = |B ⊗ A|+ = |A|q+ |B|p+ . 2 Example 6.1 Consider a sample x1 , . . . , xn i.i.d. x, where x ∼ Np (µ, Σ) and forms the “sample matrix” x1 .. X = . . xn

Then, we see that X ∼ Npn (1µ , In ⊗ Σ). Example 6.2 As another example, suppose that z ∼ Np (0, I) and form the “outer product” matrix W = zz = (z1 z, . . . , zp z). Then, obviously, E W = var z = I, but the variance of W depends on the fourth-order moments of z. Since E zi = E zi3 = 0, E zi2 = 1, and E zi4 = 3, it follows easily that E zi z E zi zj zz

= ei , = δij I + ei ej + ej ei ,

from which cov(zi z, zj z) = δij I + ej ei . At this point it becomes useful to deﬁne the “commutation matrix” Kp , a block-matrix whose block in position (i, j) is ej ei ∈ Rpp , 2

Kp = (ej ei ) ∈ Rpp2 .

76

6. Multivariate sampling

For example, for p = 2, we have 1 0 K2 = ··· 0 0

0 0 ··· 1 0

.. . .. . ··· .. . .. .

0

0

0 ··· . 0

1 ··· 0 0

1

This enables one to write succinctly [Magnus and Neudecker (1979)] var W = (I + Kp ). To generalize slightly, suppose that x ∼ Np (0, Σ) and let W = xx . Since d

W = xx = Azz A , where z ∼ Np (0, I) and Σ = AA , the variance of W becomes var W

= =

var Azz A = var (A ⊗ A)vec(zz ) (A ⊗ A)(I + Kp )(A ⊗ A ).

However, since Kp commutes with A ⊗ A (why?) (v. Problem 6.4.2), then, ﬁnally, var W = (I + Kp )(Σ ⊗ Σ). We can also write this expression componentwise as cov(wik , wjl ) = σij σkl + σkj σil , where Σ = (σij ). Suppose that 0 1 x , ∼ N2 0 ρ y

ρ 1

(6.1)

.

We may use the above result to determine the variance of (x2 , y 2 , xy). This is needed later in obtaining the asymptotic distribution of the sample correlation coeﬃcient. For x ∼ unif(B n ), W = xx , the above method may be adapted to help determine E W and var W. The distribution for linear transformations of multivariate normal matrices is straightforward with Lemma 6.2. Proposition 6.1 If A ∈ Rrp , X ∼ Nqp (M, Ω), and B ∈ Rqs , then AXB ∼ Nsr (AMB, (A ⊗ B )Ω(A ⊗ B)) . Proof. Since vec(X ) ∼ Npq (vec(M ), Ω), then vec ((AXB) )

= (A ⊗ B )vec(X ) ∼ Nrs ((A ⊗ B )vec(M ), (A ⊗ B )Ω(A ⊗ B)) .

The proof is complete as (A ⊗ B )vec(M ) = vec ((AMB) ).

2

6.2. Random matrices and multivariate sample

77

Example 6.3 Assuming X ∼ Nqp (M, A ⊗ B), A ≥ 0 is in Rpp and B ≥ 0 is in Rqq . We evaluate E XX . Let X = (x1 , . . . , xq ) and observe, with the choice A = Ip and B = ei , that xi ∼ Np (mi , bii A), where M = (m1 , . . . , mq ). Then, E XX =

q

E xi xi =

i=1

q

(bii A + mi mi )

i=1

leads to the expression E XX = (tr B)A + MM . We now turn to considerations of convergence. For the general sample x1 , . . . , xn i.i.d. x, where E x = µ and var x = Σ, the strong law of large numbers (S.L.L.N.) provides the sample mean as a natural estimate ¯ = nj=1 xj /n for µ: x w.p.1 ¯ → µ. x Of course, Wi = xi xi , i = 1, . . . , n, i.i.d. xx , where E xx = Σ + µµ are n and the S.L.L.N. applies to W = j=1 xj xj /n so that w.p.1 W → Σ + µµ . ˆ =W−x ¯x ¯ , we ﬁnd Then, obviously, if we let Σ ˆ w.p.1 Σ → Σ. ¯x ¯ = Σ/n + µµ , so that However, E x E

n ˆ Σ=Σ n−1

and it has become customary to use this “unbiased” estimate. The reader should have no particular diﬃculty in showing that as explicit functions of the sample matrix, these (unbiased and consistent) estimates may be expressed by ¯= x

n ˆ 1 1 X 1 and S ≡ X QX, Σ= n n−1 (n − 1)

where Q = I − n−1 11 . The estimate S is the sample variance, which is often written as 1 ¯ )(xi − x ¯ ) . (xi − x (n − 1) i=1 n

S=

As an expression of “pythagorus,” we ﬁnd X = QX + PX, where P = I − Q = n−1 11

78

6. Multivariate sampling

and, thus, ¯. X X = X QX + X PX = (n − 1)S + n¯ xx

6.3 Asymptotic distributions The central limit theorem (C.L.T.) states that for any sample x1 , . . . , xn i.i.d. x, where E x = µ and var x = Σ, √ d n(¯ x − µ) → z, where z ∼ Np (0, Σ). d

Now, recall the very general fact that if xn → x on Rp and g : Rp → Rq d is any continuous (with Px probability 1)1 function, then g(xn ) → g(x) on Rq . Note that since matrices in Rpq are really only vectors in Rpq , this result is considerably more general than it might appear at ﬁrst. Thus, if Σ is nonsingular (the singular case goes through as well; v. Problem 6.4.10), x − µ) → χ2p . n(¯ x − µ) Σ−1 (¯ d

There is another very basic fact that derives from the Cram´er-Wold theorem and the (univariate) Slutsky theorem. d

d

Lemma 6.3 (Multivariate Slutsky) If Xn → X on Rpq and Yn → C on Rrs where C is any constant matrix, then d

(Xn , Yn ) → (X, C) on Rpq × Rrs . Proof. From Cram´er-Wold Proposition 2.10, for any linear combination d tij xn,ij → tij xij , i,j

i,j

skl yn,kl

d

→

k,l

skl ckl ,

k,l

and from the univariate Slutsky theorem, d tij xn,ij + skl yn,kl → tij xij + skl ckl . i,j

i,j

k,l

k,l

Using Cram´er-Wold again, the conclusion is reached.

2

A more general statement on metric spaces can be found in Billingsley (1968, p. 27). It follows, of course, that for any continuous function, d

g(Xn , Yn ) → g(X, C). 1 Let

Cg = {t ∈ Rp : g is continuous at t}. Then, g is continuous with Px probability 1 means that Px (Cg ) = P (x ∈ Cg ) = 1.

6.3. Asymptotic distributions

79

As a simple example n(¯ x − µ) S−1 (¯ x − µ) → χ2p . d

One more general proposition: √ d Proposition 6.2 (Delta method) If n(xn − c) → z on Rp and g : p q R → R is diﬀerentiable at c, then √ d n (g(xn ) − g(c)) → Dg(c) z. Proof. This is simply because by the very deﬁnition of the derivative at c, the function h(t)/|t − c|, t = c k(t) = 0, t = c, where h(t) = (g(t) − g(c)) − Dg(c) (t − c), is continuous at c, and we may, therefore, write √ √ √ n (g(xn ) − g(c)) = Dg(c) n(xn − c) + k(xn )| n(xn − c)|. d

Using Slutsky’s theorem, we may conclude that since k(xn ) → 0 and √ d | n(xn − c)| → |z|, √ d n (g(xn ) − g(c)) → Dg(c) z. 2 This, of course, applies directly to the C.L.T. to give √ d n (g(¯ x) − g(µ)) → Nq (0, Dg(µ)ΣDg(µ) ) . However, consider a more elaborate application: Let x1 , . . . , xn be i.i.d. x as before with E x = 0 and var x = Σ. Then, let Wi = xi xi , i = 1, . . . , n, and W = xx so that Wn W W1 , . . . , are i.i.d. x x1 xn with

E

and

var

W x

=

=

Σ 0

var W cov(vec(W), x) cov(x, vec(W)) Σ

By the C.L.T., √

W x

n

W−Σ ¯ x

d

→ Npp+1 (0, Ω)

≡ Ω.

80

6. Multivariate sampling

and the reader may then use Lemma 6.3 to ﬁnd that √ √ √ √ d ˆ − Σ) = n(W − Σ) − √1 ( n¯ n(Σ x)( n¯ x) → Npp (0, var W) n and, of course, √

d

n(S − Σ) → Npp (0, var W).

Note that since the function S is unchanged if x is replaced by x − µ, this result is automatically valid for the more general case where E x = µ. The expression for var W was given in Example 6.2 for the normal case, and the elliptical case is treated in the sequel in Example 13.6. Unfortunately, var W is seldom of a particular tractable form. It depends on the fourth-order multivariate cumulants of x. The relation between product-moments and multivariate cumulants is rather technical and is relegated to Appendix B. There, it is proven generally for W = (wij ) = xx that ijkl ik jl kl ij il jk cov(wik , wjl ) = µijkl 1111 − µ11 µ11 = k1111 + k11 k11 + k11 k11 ,

where the µ’s are the product-moments and the k’s are the cumulants of x. Example 6.4 For a sample of size n from a bivariate distribution with ﬁnite fourth-order moments, we ﬁnd the asymptotic distribution 2 2 σ1 s1 √ d n s12 − σ12 → N3 (0, Ω), s22 σ22 where

µ14 − (µ12 )2 Ω= · ·

12 1 µ12 31 − µ11 µ2 12 2 µ22 − (µ12 11 ) ·

1 2 µ12 22 − µ2 µ2 12 2 µ12 . 13 − µ11 µ2 2 µ4 − (µ22 )2

The product-moments are µ14 µ12

µ12 31 µ12 11 µ22 µ12 22

= E (x1 − µ1 )4 , = E (x1 − µ1 )2 = σ12 , = E (x1 − µ1 )3 (x2 − µ2 ), = E (x1 − µ1 )(x2 − µ2 ) = σ12 , = E (x2 − µ2 )2 = σ22 ,

= E (x1 − µ1 )2 (x2 − µ2 )2 , etc.

¯ and S will not be asymptotically independent unless all thirdIn general, x order product-moments of x in cov (vec(W), x) are null. But this is exactly the case when z = Σ−1/2 x has a spherical distribution since cov (vec(W), x)

=

cov (vec(xx ), x)

6.4. Problems

=

81

(Σ1/2 ⊗ Σ1/2 ) cov (vec(zz ), z) Σ1/2

and all third-order product-moments of z are null (v. Problem 4.6.5). However, if the underlying random vector x is already normal, then things reduce considerably. For p = 2, the correlation coeﬃcient, r, is a very simple function of S and, thus, it should be straightforward for the reader to obtain the asymptotic distribution of r (v. Problems 6.4.8-6.4.9). In fact, since this function is unchanged if the individual coordinates are normalized, we may assume at the outset that 0 1 ρ x , . ∼ N2 0 ρ 1 y

6.4 Problems 1. Prove Lemma 6.2: If A ∈ Rrp , X ∈ Rpq , and B ∈ Rqs , then vec(AXB) = (B ⊗ A)vec(X). 2. Let A, B ∈ Rpp and Kp be the “commutation matrix.” Show the following: p p (i) Kp = i=1 j=1 ei ej ⊗ ej ei , (ii) Kp vec(A) = vec(A ), (iii) Kp (A ⊗ B) = (B ⊗ A)Kp , (iv) tr A B = [vec(A)] vec(B), (v) If A is symmetric, tr A2 = [vec(A)] 12 (I + Kp )vec(A). |=

3. Show that if Z ∼ Nqp (0, I) and P and Q in Rpp are orthogonal QZ. projections such that PQ = 0, then PZ Hint: Obtain PZ var . QZ 4. Obtain the p.d.f. of X ∼ Nqp (M, A ⊗ B), where A > 0 is in Rpp and B > 0 is in Rqq : pq q p f (X) = (2π)− 2 |A|− 2 |B|− 2 etr − 12 A−1 (X − M)B−1 (X − M) , where etr(·) ≡ exp[tr(·)]. Hint: Let X = A1/2 ZB1/2 + M, where Z ∼ Nqp (0, Ip ⊗ Iq ), and use Corollary 6.1. 5. Assume E E = 0 and var E = In ⊗ Σ, Σ ≥ 0 is in Rpp . Show that (i) var E = Σ ⊗ In , (ii) E E AE = (tr A)Σ.

82

6. Multivariate sampling

6. Assume Σ22 is nonsingular and Σ11 n (X1 , X2 ) ∼ Np1 +p2 1(µ1 , µ2 ), In ⊗ Σ21

Σ12 Σ22

,

where Xi ∈ Rnpi , µi ∈ Rpi , and Σij ∈ Rppij , i, j = 1, 2. Prove X1 | X2 ∼ Npn1 (1µ1 + (X2 − 1µ2 )B , In ⊗ Σ11.2 ) , with B = Σ12 Σ−1 22 . 7. For W = xx , in each case determine E W and var W: 1 ρ , (i) x ∼ N2 (0, Σ) and Σ = ρ 1 1 (ii) x ∼ unif(S ), (iii) x ∼ unif(B 2 ). 8. Assume (xi , yi ) , i = 1, . . . , n, are i.i.d. 0 1 x ∼ N2 , 0 ρ y

ρ 1

and let r be the sample correlation coeﬃcient. Prove the asymptotic

√ d result n(r − ρ) → N 0, (1 − ρ2 )2 . 9. Fisher’s z-transform is z

=

ζ

=

1+r , 1−r 1+ρ tanh−1 (ρ) = 12 log . 1−ρ

tanh−1 (r) =

1 2

log

(i) Show that it is a “variance stabilizing transformation” for the √ d correlation coeﬃcient: n − 3(z − ζ) → N (0, 1). (ii) Use the fact that z is a monotone function of r to obtain an approximate (1 − α)100% conﬁdence interval for ρ, zα/2 zα/2 , tanh z + , tanh z − (n − 3)1/2 (n − 3)1/2 where P (N (0, 1) > zα/2 ) = α/2. 10. Let x1 , . . . , xn i.i.d. x, where E x = µ and var x = Σ. Prove that n(¯ x − µ) Σ− (¯ x − µ) → χ2r , d

where r = rank Σ. 11. Demonstrate the following representation of Mahalanobis distance: n h xi − n1 j=1 h xj = di , sup 2 1/2 |h|=1 n n 1 1 k=1 h xk − n j=1 h xj (n−1)

6.4. Problems

83

¯ ) S−1 (xi − x ¯ )]1/2 is the Mahalanobis distance from where di = [(xi − x ¯. xi to x Remark: This was used by Stahel (1981) and Donoho (1982) to suggest the robust estimate of location as a weighted average n w(ui )xi ˆ = i=1 µ , n i=1 w(ui ) where w(·) is a positive and strictly decreasing function and |h xi − medj (h xj )| . |h|=1 medk |h xk − medj (h xj )|

ui = sup

The notation “med” refers to the ordinary median. 12. Multivariate familial data [Konishi and Khatri (1990)]. Suppose a random sample of n families on x = (x1 , . . . , xp ) ∈ Rp with E x = µ and var x = Σ. Let x1i .. Zi = . , i = 1, . . . , n, xki ,i

denote the measurements on the ith family with ki ≥ 1 siblings, where xji = (x1ji , . . . , xpji ) , j = 1, . . . , ki , is the score of the jth child on p characteristics. It is assumed that Z1 , . . . , Zn are mutually independent and E Zi var Zi

= 1ki µ , =

(Iki ⊗ Σ) + (1ki 1ki − Iki ) ⊗ Σs .

The matrix Σs reﬂects the dependence among siblings. For the estimation of Σ, let ¯1 x ki ¯ = .. ¯ i )(xji − x ¯ i ) , = (xji − x , V X i . ¯ n x

j=1

¯ i = (x1i + · · · + xki ,i )/ki . Further, let B ∈ Rnn , B ≥ 0, such where x that B1n = 0. (i) Prove that

ˆ = (tr B)−1 Σ

¯ + ¯ BX X

n

ωi Vi

,

i=1

where the weights ω1 , . . . , ωn are non-negative constants, satisﬁes n ˆ = Σ+(tr B)−1 ωi (ki − 1) − tr[B(In − D−1 EΣ n )] (Σ−Σs ), i=1

84

6. Multivariate sampling

where Dn = diag(k1 , . . . , kn ). ˆ is unbiased for Σ. (ii) Find a condition on the weights so that Σ (iii) The corresponding estimate of Σs is given by n −1 ¯ ˆ ¯ νi V i , Σs = (tr B) X BX + i=1

where ν1 , . . . , νn are constants. Prove that for weights satisfying the condition n νi (ki − 1) + tr(BD−1 n ) = 0, i=1

ˆ s is unbiased for Σs . Σ A multivariate familial model for interclass correlation, with a “mother” for each family, was considered earlier by Srivastava et al. (1988). Principal component analysis for the model described here was developed by Konishi and Rao (1992). A general description of principal components is given in Chapter 10.

7 Wishart distributions

7.1 Introduction As before,

x1 X = ...

xn

¯ and S, denotes the sample matrix from which x n¯ x = X 1, ¯, (n − 1)S = X X − n¯ xx provide consistent unbiased estimates for µ and Σ, respectively. In Section 7.2, the maximum likelihood estimates of µ and Σ are derived assuming x1 , . . . , xn i.i.d. x with x ∼ Np (µ, Σ), Σ > 0. The fundamental ¯ and S is proved in Proposiresult about the joint distribution of x tion 7.1. The basic properties of Wishart distributions are studied in Section 7.3. Section 7.4 presents the Box-Cox transformation to enhance the multivariate normality of the data.

¯ and S 7.2 Joint distribution of x ¯ and S are “optimal” in some respects. Denote With underlying normality, x V = (n − 1)S. Using the notation exp[tr(·)] = etr(·), the p.d.f. for X can

86

7. Wishart distributions

be written in various ways: f (X)

=

− np 2

(2π)

−n 2

|Σ|

exp − 12

n

(xi − µ) Σ−1 (xi − µ)

i=1

= =

−1 n n ¯ |Σ|− 2 e− 2 µ Σ µ exp − 12 tr Σ−1 X X + nµ Σ−1 x ' & np n (7.1) x − µ)(¯ x − µ) ] Σ−1 . (2π)− 2 |Σ|− 2 etr − 12 [V + n(¯ (2π)−

np 2

By general properties of exponential families [Fraser (1976), pp. 339, 342, 406, or Casella and Berger (1990), pp. 254-255, 263], it is plain that ¯ ) (or any one-to-one function such as (S, x ¯ )) is minimal suﬃcient (X X, x and complete for (Σ, µ), so that by the Rao-Blackwell/Lehmann-Scheﬀ´e ¯ and S have minimum theorems, among all unbiased estimates of µ and Σ, x ¯ ) is the MVUE (Minimum Variance Unbiased variance. We say that (S, x Estimate) of (Σ, µ). ˆ and Furthermore, to obtain the maximum likelihood estimates (MLE) µ ˆ when n − 1 ≥ p, we minimize Σ 1 x − µ) Σ−1 (¯ x − µ) (7.2) VΣ−1 + (¯ n ˆ = x ¯ , so we need only and (since the last term is ≥ 0) it is clear that µ minimize 1 ln |nV−1 Σ| + tr VΣ−1 , n where the constant, ln |nV−1 |, was added. The condition n − 1 ≥ p ensures that V is nonsingular w.p.1. This is proved later in Corollary 7.2. But then, letting T = nV−1 Σ, we need only determine the T that minimizes ln |Σ| + tr

ln |T| + tr T−1 . However, this is accomplished when all the eigenvalues of T are 1 so that T = I and we conclude altogether ˆ = 1 V. ˆ =x ¯ and Σ µ n Remark: It is a well-known result, which can be traced back to Gauss, that the only location family, f (x − θ), of p.d.f. on R for which x ¯ is a MLE of θ originates from the normal density. This MLE characterization of normal density also holds on Rp [Stadje (1993)]. ¯ and S. It is obvious that Let us consider the exact distribution of x 1 ¯ ∼ Np µ, Σ . x n d

We begin by representing x = Az + µ, for any AA = Σ, which, for the d sample matrix, means that X = ZA + 1µ . Thus, (¯ x, Sx ) = (A¯ z + µ, ASz A ). d

7.3. Properties of Wishart distributions

87

|=

|=

|=

However, in Z = (zij ), all the components are i.i.d. N (0, 1) so that even the columns are mutually independent. Thus, with orthogonal projections P = QZ (v. Problem 6.4.3), n−1 11 and Q = I − n−1 11 , it is clear that PZ ¯ Sz , hence and since n¯ z = Z P1 and (n − 1)Sz = Z QQZ, we see that z ¯ Sx . x If next we express Q = HH , where H gives an orthonormal basis for ⊥ 1 (of dimension (n − 1)), it is made plain that (n − 1)Sz = Z HH Z = U U, where Z H = U = (u1 , . . . , un−1 ), ui i.i.d. Np (0, I). Accordingly, we make the following deﬁnition. Deﬁnition 7.1 Wishart distribution: m d zi zi , zi i.i.d. Np (0, I). W ∼ Wp (m) iﬀ W = i=1

V

∼ Wp (m, Σ) iﬀ V = AWA , Σ = AA , W ∼ Wp (m). d

Thus, we have the fundamental statistical result:

¯ ¯ ∼ Np (µ, Σ/n), (n − 1)S ∼ Wp (n − 1, Σ), and x x

|=

Proposition 7.1 For xi i.i.d. Np (µ, Σ), i = 1, . . . , n, S.

One may, of course, go to some trouble to obtain an explicit density for the Wishart. However, one needs primarily to understand some of its basic properties and the density will not really reveal very much.

7.3 Properties of Wishart distributions The distribution of the trace of W ∼ Wp (m) follows almost immediately from the deﬁnition. Proposition 7.2 W ∼ Wp (m) =⇒ tr W ∼ χ2mp . Proof. By deﬁnition of Wp (m), d

tr W = tr

m i=1

zi zi =

m

zi zi ,

i=1

where zi are i.i.d. Np (0, I). From Proposition 4.4, zi zi ∼ χ2p . Corollary 3.1 then gives tr W ∼ χ2mp . 2 Now, a useful lemma to determine when V ∼ Wp (m, Σ) is nonsingular w.p.1. is the following: Lemma 7.1 Z = (zij ) ∈ Rnn with zij i.i.d. N (0, 1) =⇒ P (|Z| = 0) = 0.

88

7. Wishart distributions

Proof. The proof proceeds by induction. The result is true for n = 1, as z11 has an absolutely continuous distribution. Next, partition z11 z12 Z= z21 Z22 n−1 and assume the result holds for Z22 ∈ Rn−1 . Then,

P (|Z| = 0)

= P (|Z| = 0, |Z22 | = 0) + P (|Z| = 0, |Z22 | = 0) = P (z11 = z12 Z−1 22 z21 , |Z22 | = 0) = E P (z11 = z12 Z−1 22 z21 , |Z22 | = 0 | z12 , z21 , Z22 ) =

0. 2

A slight generalization is contained in Corollary 7.1 Z = (zij ) ∈ Rnn with zij i.i.d. N (0, 1) =⇒ P (|Z| = t) = 0, ∀t. Proof. P (|Z| = t) = E P z11 = z12 Z−1 22 z21 +

t , |Z22 | = 0 | z12 , z21 , Z22 |Z22 |

= 0. 2 It should be observed that Lemma 7.1 and Corollary 7.1 remain valid if Z has any absolutely continuous distribution. We can now prove [Stein (1969), Dykstra (1970)]: Proposition 7.3 W ∼ Wp (m), m ≥ p =⇒ W is nonsingular w.p.1. d

Proof. The representation W = Z Z, where Z = (z1 , . . . , zm ) and zi ’s are i.i.d. Np (0, I), gives rank W = rank Z Z = rank Z ≥ rank (z1 , . . . , zp ) d

whence rank W

w.p.1 = p

w.p.1 = p.

2

Its corollary gives a condition on the sample size and the population variance for the sample variance matrix S to be nonsingular w.p.1. Corollary 7.2 V ∼ Wp (m, Σ), m ≥ p, |Σ| = 0 =⇒ |V| = 0 w.p.1. Eaton and Perlman (1973) established that the sample variance matrix S is nonsingular w.p.1 for independent observations, which are not necessarily normal or identically distributed. Concerning linear transformations of Wishart matrices, we have

7.3. Properties of Wishart distributions

89

Proposition 7.4 V ∼ Wp (m, Σ), B ∈ Rqp =⇒ BVB ∼ Wq (m, BΣB ). d

Proof. Let W ∼ Wp (m). Since V = AWA , for any AA = Σ, then BVB = (BA)W(BA) ∼ Wq (m, BAA B ). d

2 Example 7.1 Suppose W ∼ Wp (m). What is E WAW for a ﬁxed A ≥ 0? d

Since HWH = W, for all H ∈ Op , we see that WAW

= HWH AHWH , ∀H ∈ Op d

= HWDWH , d

where H was chosen to diagonalize A, H AH = D = diag(λi ). Thus, d m E WAW = H(E WDW)H . But using W = i=1 zi zi , where zi ∼ Np (0, I) are independent, we ﬁnd E zi zi Dzj zj E WDW = i,j

=

E zi zi Dzi zi +

E zi zi Dzj zj

i =j

i

= mE xx Dxx + m(m − 1)E xx Dyy , where x and y are i.i.d. Np (0, I). However, λi E x2i xx E xx Dxx = i

=

λi (I + 2ei ei )

i

= and E xx Dyy

=

(tr A)I + 2D

λi E xi yi xy

i

=

λi ei ei

i

= D. Hence, E WDW = m(tr A)I + m(m + 1)D and, ﬁnally, we obtain E WAW = m(tr A)I + m(m + 1)A. The characteristic function of Wishart distributions also follows from basic principles.

90

7. Wishart distributions

Example 7.2 The characteristic function of V ∼ Wp (m, Σ), evaluated d

at S symmetric, is deﬁned by cV (S) = E exp(i tr SV). Write V = m A j=1 zj zj A , for any AA = Σ, and diagonalize A SA = HDH to obtain m cV (S) = E exp i tr SA zj zj A j=1

= E exp i tr HDH

= E exp i tr D

m

zj zj

= E exp i tr D

m

zj zj since H zj = zj

j=1 p m

(H zj )(H zj )

j=1

= E exp i

j=1

m

d

2 zjk dk , where D = diag(d1 , . . . , dp )

j=1 k=1

=

=

m

p

j=1 k=1 p m

cχ21 (dk ) (1 − 2idk )−1/2

j=1 k=1

= |I − 2iD|−m/2 = |I − 2iSΣ|−m/2 . Hence, the characteristic function is given by cV (S) = |I − 2iΣS|−m/2 . Now, consider some results concerning the marginals of a Wishart. For this reason, partition V ∈ Rpp as V11 V12 , V= V21 V22 where V11 ∈ Rrr and V22 ∈ Rss , r + s = p. The matrix Σ is partitioned similarly. Proposition 7.5 V ∼ Wp (m, Σ) =⇒ V11 ∼ Wr (m, Σ11 ). Proof. Choose B = ( Ir

0 ) ∈ Rrp in Proposition 7.4.

Concerning independence, we have:

2

Proposition 7.6 V ∼ Wp (m, Σ) and Σ12 = 0 =⇒ V11

|=

7.3. Properties of Wishart distributions

91

V22 .

Proof. By the very deﬁnition, X X X X Y d V = U U = ( X Y ) = , Y Y X Y Y

|=

where U = (u1 , . . . , um ) and the ui ’s are i.i.d Np (0, Σ). Then, it suﬃces to recall (v. Problem 6.4.6) that for a multivariate normal, Σ12 = 0 implies Y. 2 X The previous two propositions are not surprising if we consider their statistical interpretation. First, the distribution of V is associated with the sample variance based on all p components, whereas that of V11 corresponds to a sample variance but considering only the ﬁrst r components. Second, if Σ12 = 0, then the distributions of V11 and V22 are associated with sample variances based on two independent subvectors of dimension r and s, r + s = p. The next proposition, the proof of which is left as an exercise, relates to sums of independently distributed Wishart matrices. It has to do with the way one would pool information from independent samples to estimate the population variance (v. Problem 8.9.1). indep ∼ Wp (mi , Σ), i = 1, . . . , k, then k k Vi ∼ Wp mi , Σ .

Proposition 7.7 If Vi

i=1

i=1

Lemma 7.2 Let H = (h1 , . . . , hr ), where the hi ’s are orthonormal in Rn and Z ∼ Npn (0, In ⊗ Ip ). Then, 1. H Z ∼ Npr (0, Ir ⊗ Ip ), 2. Z HH Z ∼ Wp (r). Proof. Using Proposition 6.1, H Z ∼ Npr (0, (H H) ⊗ Ip ). This proves part 1 because H H = Ir . Since Z hi , i = 1, . . . , r, arei.i.d. Np (0, I), part r 2 follows from the Wishart deﬁnition: Z HH Z = i=1 (Z hi )(Z hi ) ∼ 2 Wp (r). Proposition 7.8 Let

x1 X = ... ∼ Npn (0, In ⊗ Σ).

xn

If V ⊂ Rn is a linear subspace, dim V = r, and P is the orthogonal projection on V, then X PX ∼ Wp (r, Σ).

92

7. Wishart distributions

Proof. Choose an orthonormal basis H = (h1 , . . . , hr ) for V and observe d

that P = HH and r = rank P = dim V. Write Σ = AA and X = ZA , where Z ∼ Npn (0, In ⊗ Ip ). Therefore, using Lemma 7.2, X PX

= X HH X d = AZ HH ZA = AWA , where W ∼ Wp (r). d

Hence, X PX ∼ Wp (r, Σ).

2

General results on Wishart and chi-square distributions associated with matrix quadratic forms are available in Mathew and Nordstr¨ om (1997). A fundamental result on marginals useful in the sequel is now stated, but ﬁrst −1 recall the notation V11.2 = V11 − V12 V22 V21 , where V was partitioned as on page 90. Proposition 7.9 If V ∼ Wp (m, Σ), m ≥ p, Σ > 0, then V11.2

and V11.2

|=

V21 | V22 V22

∼ Wr (m − s, Σ11.2 ), ∼ Nrs (V22 Σ−1 22 Σ21 , V22 ⊗ Σ11.2 ), ∼ Ws (m, Σ22 ),

(V21 , V22 ).

Proof. As before, write d

V=UU=

X Y

(X Y) =

X X X Y Y X Y Y

,

where U ∼ Npm (0, Im ⊗ Σ). Thus, X | Y ∼ Nrm (YΣ−1 22 Σ21 , Im ⊗ Σ11.2 ) (v. Problem 6.4.6). Let P = Y(Y Y)−1 Y be the orthogonal projection on the column space of Y and Q = I − P, rank Q = m − s. It is clear, since Y = PY, that V11.2 V21 V22

= X [I − Y(Y Y)−1 Y ]X = (QX) (QX), = Y X = (PY) (PX), = Y Y = (PY) (PY).

|=

|=

Since Y X | Y ∼ Nrs ((Y Y)Σ−1 22 Σ21 , (Y Y)⊗Σ11.2 ) depends only on Y Y, −1 s then V21 | V22 ∼ Nr (V22 Σ22 Σ21 , V22 ⊗ Σ11.2 ). From Proposition 7.8 and QY = 0, V11.2 | Y ∼ Wr (m − s, Σ11.2 ), which does not depend on Y; hence, V11.2 Y and V11.2 ∼ Wr (m − s, Σ11.2 ), unconditionally. It is clear V22 ∼ Ws (m, Σ22 ). Only independence remains to be shown. QX. To see this, note PQ = 0 and However, conditionally on Y, PX P P var X|Y = (P , Q ) ⊗ Σ11.2 Q Q P ⊗ Σ11.2 0 = . 0 Q ⊗ Σ11.2

Hence, given Y, V11.2

|=

7.3. Properties of Wishart distributions

93

V21 . Finally, using Proposition 2.13,

E [f (V11.2 ) · g(V21 , V22 )]

= E [E f (V11.2 ) g(V21 , V22 ) | Y] = E {E [f (V11.2 ) | Y] E [g(V21 , V22 ) | Y]} = E {Ef (V11.2 ) E [g(V21 , V22 ) | Y]} = Ef (V11.2 ) · Eg(V21 , V22 ), 2

which proves independence.

Proposition 7.9 with r = 1 and s = p−1 can be used to prove inductively several results concerning Wishart distributions. Here are two corollaries: the distribution of the generalized variance, |V|, and the Wishart density. Corollary 7.3 If V ∼ Wp (m, Σ), m ≥ p, Σ > 0, then |V| ∼ |Σ|

p

χ2m−p+i ;

i=1

i.e., |V|/|Σ| is distributed as a product of p mutually independent chi-square variables.

|=

Proof. The result obviously holds for p = 1. Assume it holds for p − 1. Let r = 1 and s = p − 1 in Proposition 7.9. Then, |V| = v11.2 |V22 |, where V22 . From the v11.2 ∼ σ11.2 χ2m−p+1 , V22 ∼ Wp−1 (m, Σ22 ), and v11.2 induction hypothesis, |V22 | ∼ |Σ22 |

p−1

χ2m−p+1+i = |Σ22 | d

i=1

p

χ2m−p+i

i=2

2

and the conclusion follows. Corollary 7.4 If W ∼ Wp (m), m ≥ p, then the p.d.f. of W is

1 |W|(m−p−1)/2 etr − 12 W , W > 0, 2m p where Γp (u) = π p(p−1)/4 i=1 Γ u − 12 (i − 1) , u > 12 (p − 1). fW (W) =

1

2mp/2 Γp

(7.3)

Proof. The result holds for p = 1, as the density reduces to a chi-square density. Let r = 1 and s = p − 1 in Proposition 7.9, then w21

w11.2 | W22 W22

∼ χ2m−p+1 , ∼ Np−1 (0, W22 ), ∼ Wp−1 (m).

Thus, the joint p.d.f. of (w11.2 , w21 , W22 ) is 1 (m−p+1)/2−1 exp(− 12 w11.2 ) w 2(m−p+1)/2 Γ[ 12 (m − p + 1)] 11.2 −1 W22 w21 ) · (2π)−(p−1)/2 |W22 |−1/2 exp(− 12 w21

94

7. Wishart distributions

·

1 |W22 |(m−p)/2 etr(− 12 W22 ). m(p−1)/2 2 Γp−1 ( 12 m)

Make the change of variables (w11.2 , w21 , W22 ) → (w11 , w21 , W22 ) with jacobian J(w11.2 , w21 , W22 → w11 , w21 , W22 ) = J(w11.2 → w11 ) = 1 −1 W22 w21 while using the relations |W| = w11.2 |W22 | and w11.2 = w11 −w21 to get the result. 2 The reader should check that Γp (u) is a generalized gamma function in the sense that Γ1 (u) = Γ(u), u > 0. The density of V ∼ Wp (m, Σ), Σ > 0, m ≥ p, follows directly from the transformation V = AWA , for any AA = Σ, and the jacobian in Proposition 2.19 (v. Problem 7.5.7). James (1954) and Olkin and Roy (1954) proposed a constructive proof by jacobians of transformations on k-surfaces (manifolds). It requires a knowledge of diﬀerential forms and integration on k-surfaces which goes beyond the scope of this book. The theory of singular Wishart distributions (m < p) is available in Uhlig (1994). The function (7.3) is a density function even when the number of degrees of freedom m ∈ R, possibly noninteger, satisﬁes m > p − 1 [Muirhead (1982), p. 62].

7.4 Box-Cox transformations A method [Andrews et al. (1971)] that is an extension of the technique of Box and Cox (1964) is described for obtaining data-based transformations of multivariate observations to enhance the normality of their distribution. Speciﬁcally, power transformations of the original variables are estimated to eﬀect both marginal and joint normality. The likelihood method, used by Box and Cox (1964) for the univariate problem, is the one adopted here for the multivariate case. The simple family of power transformations deﬁned by λj (λj ) xj = (xj − 1)/λj , λj = 0, λj = 0, ln xj , j = 1, . . . , p, will be considered. Each variable xj must be non-negative, otherwise, with a known lower bound, we may add a constant suﬃciently large, aj , and consider xj + aj as the original variable. Let x1 .. X = . = (xij ) ∈ Rnp , xn

7.4. Box-Cox transformations

95

X(λ)

(λ) x1 (λj ) n . = .. = (xij ) ∈ Rp , (λ) xn

be the sample matrices of the original and transformed data, respectively, where λ = (λ1 , . . . , λp ) is the unknown vector of power transformation parameters. If λ is the vector of parameters yielding joint normality, Np (µ, Σ), the density of X(λ) is from (7.1): f (X(λ) ) = (2π)− 2 |Σ|− 2 ! ) ( ·etr − 12 V(λ) + n(¯ x(λ) − µ)(¯ x(λ) − µ) Σ−1 , np

n

where 1 (λ) x , n i=1 i n

¯ (λ) x

=

V(λ)

=

n

(λ)

(xi

(λ)

¯ (λ) )(xi −x

¯ (λ) ) . −x

i=1

The jacobian of the transformation, J(X(λ) → X), is J=

p n

λ −1

xijj

.

j=1 i=1

Hence, the density of the genuine data X is f (X) = f (X(λ) ) · J = (2π)−

np 2

! ) ( n |Σ|− 2 etr − 12 V(λ) + n(¯ x(λ) − µ)(¯ x(λ) − µ) Σ−1 · J.

The log-likelihood of (Σ, µ, λ) is, up to an additive constant, n 1 ln |Σ| − tr(V(λ) Σ−1 ) 2 2 p n n (λ) − µ) Σ−1 (¯ x(λ) − µ) + (λj − 1) ln xij . x − (¯ 2 j=1 i=1

l(Σ, µ, λ) = −

(7.4)

For a speciﬁed λ, the maximum likelihood estimate of (Σ, µ), exactly as for (7.2), is given by ¯ (λ) . V(λ) /n, x If these estimates are substituted in (7.4), the maximized log-likelihood function is, up to an additive constant, lmax (λ) = −

p n n ln |V(λ) | + (λj − 1) ln xij , 2 j=1 i=1

(7.5)

96

7. Wishart distributions

a function of p parameters which can be computed and studied. The maxˆ may be obtained by numerically maximizing imum likelihood estimate λ (7.5). Also, conﬁdence regions for λ may be obtained. One such (1−α)100% conﬁdence region for λ based on asymptotic considerations [Fraser (1976), p. 357] is ˆ − lmax (λ) ≤ 1 χ2 {λ : lmax (λ) 2 1−α,p }, where χ21−α,p is the (1 − α)-quantile of a χ2p distribution. The likelihood criterion used here speciﬁes joint normality rather than marginal normality as the goal of the transformation.

7.5 Problems ˆ were derived by ˆ and Σ 1. The maximum likelihood estimates µ minimizing 1 ln |Σ| + tr VΣ−1 + (¯ x − µ) Σ−1 (¯ x − µ). n Derive the MLE, this time by calculus, using the vector and matrix diﬀerentiation rules of Problems 1.8.9-1.8.10. 2. Prove Proposition 7.7. 3. Show that if V ∼ Wp (m, Σ), then: (i) var V = m[Σ ⊗ Σ + (σ j σ i )] = m(I + Kp )(Σ ⊗ Σ) where Kp is the “commutation matrix.” (ii) Prove Proposition 7.7 again, but using characteristic functions this time. 4. Let W ∼ Wp (m), m ≥ p. Prove: |=

(i) 1/(t W−1 t) ∼ χ2m−p+1 , for any t, |t| = 1. W and px (0) = 0, then x is independent of (ii) If x (x x) ∼ χ2m−p+1 . (x W−1 x) d

Hint: HWH = W, ∀H ∈ Op . 5. Assume W ∼ Wp (m), m ≥ p, and A ≥ 0. Prove: (i) E W = mI, (ii) E W−1 = I/(m − p − 1). 6. Moments of generalized variance. (i) Let W ∼ Wp (m), m ≥ p. Prove E |W|h = 2ph

Γp ( 12 m + h) , h > 12 (p − m − 1). Γp ( 12 m)

7.5. Problems

97

Hint: E |W|h has an integrand of the form of a Wp (m + 2h) density. Use the normalizing constant cp,m = [2mp/2 Γp ( 12 m)]−1 . (ii) If V ∼ Wp (m, Σ), m ≥ p, Σ > 0, then E |V|h = |Σ|h 2ph

Γp ( 12 m + h) , h > 12 (p − m − 1). Γp ( 12 m)

7. Wishart density. Obtain the p.d.f. of V ∼ Wp (m, Σ), m ≥ p, Σ > 0: fV (V) =

2mp/2 Γ

1

1 |V|(m−p−1)/2 etr − 12 Σ−1 V , m/2 p 2 m |Σ|

V > 0. 8. Let W ∼ Wp (m) and consider the correlation matrix R = (rij ), where wij rij = 1/2 1/2 . wii wjj Demonstrate that the density of R is f (R) =

[Γ( 12 m)]p |R|(m−p−1)/2 . Γp ( 12 m)

Hint: Use the transformation W → w11 , . . . , wpp , R. 9. Inverted Wishart distribution. Derive the p.d.f. of U = V−1 , where V ∼ Wp (m, Σ), m ≥ p, Σ > 0:

1 |U|−(m+p+1)/2 etr − 12 Σ−1 U−1 , fU (U) = mp/2 1 m/2 2 Γp 2 m |Σ| U > 0. 10. Assume W ∼ Wp (m) and deﬁne W = TT for a unique T ∈ L+ p. (i) Prove tij ∼ N (0, 1), 1 ≤ i < j ≤ p, and t2ii ∼ χ2m−i+1 , 1 ≤ i ≤ p, are all mutually independent. (ii) Using (i) prove tr W ∼ χ2pm . (iii) Using (i) again,prove that if V ∼ Wp (m, Σ), m ≥ p, Σ > 0, p then |V| ∼ |Σ| i=1 χ2m−p+i .

8 Tests on mean and variance

8.1 Introduction ¯ ) on a good footing in Chapter 7, we Having laid the distribution of (S, x now present inference problems such as the Hotelling-T 2 test on the mean vector, the simultaneous conﬁdence intervals on means, the inference about multiple and partial correlation coeﬃcients, the test of sphericity, and the test of equality of variances. In some cases, the tests are optimal in some sense. This is the case of the Hotelling-T 2 test and the test of multiple correlation, which are shown to be uniformly most powerful invariant (UMPI). The asymptotic distribution of eigenvalues, both in the one-sample and two-sample cases, is treated in Section 8.8. Tables of critical points with references to applications for most multivariate tests are available in Kres (1983). The approach adopted here rests mainly on likelihood ratio tests, although other general and valid testing procedures based on minimization of divergence measures exist in the literature [Wakaki et al. (1990)].

8.2 Hotelling-T 2 Now, assume that x1 , . . . , xn are i.i.d. x with x ∼ Np (µ, Σ) and Σ > 0. The properties of Wishart distributions in Chapter 7 provide an easy way to obtain the distribution of the Hotelling-T 2 statistic x − µ0 ) S−1 (¯ x − µ0 ), T 2 = n(¯

(8.1)

8.2. Hotelling-T 2

99

¯ and S are the unbiased estimate for µ and Σ. This is needed where x to test the hypothesis H0 : µ = µ0 against all alternatives or to build a conﬁdence ellipsoid for µ. In fact, the following proposition shows that the Hotelling-T 2 statistic is a monotone function of the likelihood ratio test (LRT) statistic. As usual, let V=

n

¯ )(xi − x ¯ ) (xi − x

i=1

be the matrix of sums of squares and cross-products. Proposition 8.1 The likelihood ratio statistic for H0 : µ = µ0 against H1 : µ = µ0 is −n/2 1 2 T . Λ= 1+ (n − 1) ˆ = 1 V. However, ˆ =x ¯ and Σ Proof. The unrestricted MLE of (Σ, µ) is µ n the MLE of Σ under H1 is obtained from (7.2) by minimizing 1 VΣ−1 + (¯ x − µ0 ) Σ−1 (¯ x − µ0 ) n 1 x − µ0 ) ]Σ−1 . = ln |Σ| + tr [V + n(¯ x − µ0 )(¯ n Using the same technique as on page 86 we ﬁnd ln |Σ| + tr

ˆ ˆ = Σ =

1 x − µ0 ) ] [V + n(¯ x − µ0 )(¯ n n 1 (xi − µ0 )(xi − µ0 ) . n i=1

Thus, with (7.1), the LRT becomes Λ = = =

ˆ ˆ µ0 ) L(Σ, ˆ µ) ˆ L(Σ, ˆ ˆ −n/2 exp(− 1 np) |Σ| 2 ˆ −n/2 exp(− 12 np) |Σ|

x − µ0 )(¯ x − µ0 ) ]|−n/2 | n1 [V + n(¯ 1 | n V|−n/2

= |I + nV−1 (¯ x − µ0 )(¯ x − µ0 ) |−n/2 −n/2 1 T2 = 1+ , (n − 1) where the last equality made use of Problem 1.8.3.

2

The distribution of T 2 is a direct consequence of the following proposition.

8. Tests on mean and variance

Proposition 8.2 If z ∼ Np (δ, I), W ∼ Wp (m), m ≥ p, and z

|=

100

W, then

z W−1 z ∼ Fc (p, m − p + 1; δ δ/2). Proof. Using an orthogonal transformation H = (z/|z|, Γ) ∈ Op , we get immediately z W−1 z

= (Hz) (HWH )−1 (Hz) = |z|2 e1 V−1 e1 , |=

where V = HWH . Since the conditional distribution V | z ∼ Wp (m) does z. Letting not depend on z, then V ∼ Wp (m), unconditionally, and V 11 v v21 , V−1 = 21 v V22 z W−1 z = |z|2 v 11 = z z/v11.2 , where the last equality made use of Problem 1.8.1. The conclusion follows since z z ∼ χ2p (δ δ/2) and, by Proposition 7.9, v11.2 ∼ χ2m−p+1 . 2 As a corollary, we obtain the distribution of Hotelling-T 2 . Corollary 8.1 The non-null distribution of T 2 for n ≥ p + 1 is T 2 /(n − 1) ∼ Fc (p, n − p; δ), with δ = n(µ − µ0 ) Σ−1 (µ − µ0 )/2. d

Proof. In terms of the sample matrix, as on page 86, X = ZA +1µ , where d Z ∼ Npn (0, In ⊗ Ip ) and Σ = AA , and, thus, (¯ x, Sx ) = (A¯ z + µ, ASz A ). Therefore, T2 (n − 1)

x − µ0 ) = n(¯ x − µ0 ) [(n − 1)S]−1 (¯

= n (A¯ z + µ − µ0 ) [(n − 1)ASz A ]−1 (A¯ z + µ − µ0 ) ¯ + A−1 (µ − µ0 ) . ¯ + A−1 (µ − µ0 ) [(n − 1)Sz ]−1 z = n z d

The proof follows from Proposition 8.2, as ¯ + A−1 (µ − µ0 ) ∼ Np n1/2 A−1 (µ − µ0 ), I n1/2 z (n − 1)Sz

∼ Wp (n − 1) 2

are independent.

Example 8.1 The power function of an α signiﬁcance level Hotelling-T 2 test may now be evaluated as a function of δ = n(µ − µ0 ) Σ−1 (µ − µ0 )/2 in the following manner: β = P (Fc (p, n − p; δ) ≥ tα ) ,

8.2. Hotelling-T 2

101

beta 1 0.8 0.6 0.4 0.2 2

4

6

8

10

delta

Figure 8.1. Power function of Hotelling-T 2 when p = 3 and n = 40 at a level of signiﬁcance α = 0.05.

where tα = [p/(n − p)]Fα (p, n − p) is the critical point. Proposition 4.7 and Problem 3.5.6 yields β

∞

δk P (Fc (p + 2k, n − p) ≥ tα ) k! k=0 ∞ −1 F (p+2k)/2−1 δk ∞ 1 = e−δ B 2 (p + 2k), 12 (n − p) dF. k! tα (1 + F )(n+2k)/2 k=0

=

e−δ

Numerical evaluation in Mathematica for p = 3, n = 40 and α = 0.05 produced the plot in Figure 8.1. The robustness of Hotelling-T 2 is easily established. Without normality, assuming x1 , . . . , xn are i.i.d. x, E x = µ0 and var x = Σ, the asymptotic d distributions in Section 6.3 gave T 2 → χ2p . This asymptotic distribution is the same regardless of the underlying distribution of x. A diﬀerent situation arises in presence of “contamination.” Assume the simple situation x1 , . . . , xn−1 are i.i.d. Np (µ0 , Σ), but there is one (or more) contaminated observation xn ∼ Np (µ0 +γ, Σ). We assume Σ known for the x − µ0 ) ∼ Np (γ/n1/2 , Σ) sake of simplicity. It is easily checked that n1/2 (¯ and, thus, from Corollary 5.1, T 2 = n(¯ x − µ0 ) Σ−1 (¯ x − µ0 ) ∼ χ2p (γ Σ−1 γ/2n). Since P (χ2p (δ) ≥ c) is monotone increasing in δ [Ghosh (1970), p. 302], all other parameters being ﬁxed, it follows that T 2 will reject H0 : µ = µ0 with probability converging to 1 as |γ| → ∞ (for ﬁxed n) even though all observations, but one, have mean µ0 . A procedure which is insensitive to

102

8. Tests on mean and variance

contamination of the data consists of building an Hotelling-T 2 test T 2 = n(µn − µ0 )Σ−1 n (µn − µ0 ) from a robust estimate (Σn , µn ) such as an M-estimate or S-estimate. These truly robust tests are studied in Chapter 13. We end this section with a discussion of invariant tests on the mean vector. Consider the canonical problem of testing H0 : µ = 0 against H1 : µ = 0. The group of transformations Gp acts on the observations as xi → Axi , where A ∈ Gp . This transformation induces the following transformations on the minimal suﬃcient statistic (¯ x, S) and parameters, ¯ → A¯ x x, S → ASA , and µ → Aµ, Σ → AΣA . Note that the hypotheses are preserved because µ = 0 iﬀ Aµ = 0, for any A ∈ Gp . We deﬁne a test function f (¯ x, S) to be invariant iﬀ it yields the same value on the original as on the transformed data, i.e., f (y, W) = f (Ay, AWA ), ∀A ∈ Gp , ∀(y, W) ∈ Rp × Pp . This has important implications. First, the choice A = S−1/2 yields ¯ , I). f (¯ x, S) = f (S−1/2 x Now, there exists an orthogonal transformation H ∈ Op (v. Problem 1.8.14) ¯ = (¯ ¯ )1/2 e1 . Choosing now A = H, we ﬁnd such that HS−1/2 x x S−1 x f (¯ x, S)

¯ , HH ) = f (HS−1/2 x ¯ )1/2 e1 , I), = f ((¯ x S−1 x

which shows that any invariant test function depends on the data only ¯. through T 2 = n¯ x S−1 x Second, selecting A = Σ−1/2 gives ¯ , Σ−1/2 SΣ−1/2 ), f (¯ x, S) = f (Σ−1/2 x where ¯ ∼ Np (Σ−1/2 µ, n−1 I) Σ−1/2 x (n − 1)Σ−1/2 SΣ−1/2

∼ Wp (n − 1).

Using the same argument, there exists an orthogonal transformation H ∈ Op such that HΣ−1/2 µ = (µ Σ−1 µ)1/2 e1 . Choosing this time A = H, we ﬁnd ¯ , HΣ−1/2 SΣ−1/2 H ), f (¯ x, S) = f (HΣ−1/2 x where ¯ ∼ Np ((µ Σ−1 µ)1/2 e1 , n−1 I) HΣ−1/2 x (n − 1)HΣ−1/2 SΣ−1/2 H

∼ Wp (n − 1),

and, thus, the power function of any invariant test depends on (µ, Σ) only through the parameter function δ = nµ Σ−1 µ/2. These results are summarized.

8.2. Hotelling-T 2

103

Proposition 8.3 For testing H0 : µ = 0 against H1 : µ = 0, any invariant test with respect to the group Gp depends on the minimal suﬃcient ¯ . Moreover, the power function x S−1 x statistic (¯ x, S) only through T 2 = n¯ of any invariant test depends on (µ, Σ) only through the parameter function δ = nµ Σ−1 µ/2. The LRT is obviously invariant. In the class of invariant tests, it is possible to show that the Hotelling-T 2 is uniformly most powerful. We say that T 2 is the UMPI test. Proposition 8.4 For testing H0 : µ = 0 against H1 : µ = 0, the ¯ , is UMPI. x S−1 x Hotelling-T 2 test, T 2 = n¯ Proof. It has already been established that T 2 is invariant and that all invariant tests, depending on (¯ x, S), are a function of T 2 . The problem thus reduces to ﬁnding the UMP test for H0 : δ = 0 based on one observation from x ≡ T 2 /(n − 1) ∼ Fc (p, n − p; δ), where δ = nµ Σ−1 µ/2. The density of x was given in Problem 4.6.3:

∞ Γ 12 (n + 2k) x(p+2k)/2−1 δk

1 1 e−δ , x > 0. f (x; δ) = k! Γ 2 (p + 2k) Γ 2 (n − p) (1 + x)(n+2k)/2 k=0 From the Neyman-Pearson lemma, the most powerful test for H0 : δ = 0 rejects H0 for large values of the ratio

1 k ∞ k x f (x; δ) −δ δ Γ 2 (n + 2k) e ≥ c2 . = c1 f (x; 0) k! Γ 12 (p + 2k) (1 + x) k=0 Since this ratio is monotone increasing in x, this is equivalent to rejecting H0 for large values of x, x ≥ c3 . This rejection region does not depend on δ and, thus, the test is uniformly most powerful. 2 An asymptotic expansion of the distribution function of T 2 was obtained whose ﬁrst term is χ2p under the elliptical distribution [Iwashita (1997)] and for general non-normality [Fujikoshi (1997), Kano (1995)]. Improvement to the chi-square approximation by monotone transformation of T 2 is also possible [Fujisawa (1997)]. For a modiﬁcation to T 2 with the same chisquare asymptotic distribution but in the case of inﬁnite second moment, refer to Sepanski (1994). Kudˆ o (1963) was the ﬁrst to propose a multivariate analogue, when Σ is known, to the one-sided t-test. The multivariate problem is to test the null hypothesis, H0 : µ = 0, against the one-sided alternative hypothesis, H1 : µ ≥ 0, where µ ≥ 0 is interpreted componentwise. It can be stated even more generally in terms of cone. The LRT for the one-sided problem, with unknown Σ, was obtained by Perlman (1969). Tang (1994, 1996) discussed unbiasedness and invariance of tests in the one-sided multivariate problem. Silvapulle (1995) derived the null distribution of a Hotelling-T 2

104

8. Tests on mean and variance

type statistic. There is a conditional test by Fraser, Guttman and Srivastava (1991); v. also Wang and McDermott (1998a, 1998b).

8.3 Simultaneous conﬁdence intervals on means Let x1 , . . . , xn be i.i.d. Np (µ, Σ). For any preassigned level β = 1−α, deﬁne the quantile Fα (p1 , p2 ) by the equation P (F (p1 , p2 ) ≥ Fα (p1 , p2 )) = α. By applying Hotelling’s result, we have exactly (n − 1)p Fα (p, n − p) = β, x − µ) ≤ P n(¯ x − µ) S−1 (¯ (n − p) whereby we see that in β × 100% of such experiments, the “true” µ lies in the random ellipsoid ¯ ) S−1 (µ − x ¯ ) ≤ cα }, {µ ∈ Rp : n(µ − x where we have simply let cα =

(n − 1)p Fα (p, n − p). (n − p)

We are β × 100% “conﬁdent” that our particular observed ellipsoid, ¯ ) S−1 (µ − x ¯ ) ≤ cα }, CR(µ; β) = {µ ∈ Rp : n(µ − x contains µ since P (µ ∈ CR(µ; β)) = β. Sequential ﬁxed-size conﬁdence regions for the mean vector were investigated by Srivastava (1967) and Datta and Mukhopadhyay (1997).

8.3.1

Linear hypotheses

In many experiments, one simply wishes to compare the various components of µ to each other. For instance, one may ask: “Is µ1 equal to µ3 ?” “Is the diﬀerence between µ2 and the average of µ1 and µ3 equal to 3.1?” Answering the ﬁrst question amounts to testing the hypothesis H0 : µ1 − µ3 = 0, while the second question is equivalent to testing the hypothesis H0 : 2µ2 − (µ1 + µ3 ) = 6.2. Questions like these are said to be linear in µ, and in the general case, there would be a certain speciﬁed vector a ∈ Rp and constant c ∈ R for which we would wish to test H0 : a µ = c.

8.3. Simultaneous conﬁdence intervals on means

105

Given x1 , . . . , xn i.i.d. Np (µ, Σ), if we let yi = a xi , i = 1, . . . , n, then clearly y1 , . . . , yn are i.i.d. N1 (a µ, a Σa). Obviously, one may apply the univariate results directly to the y data so that √ d H0 is correct iﬀ n(¯ y − c)/sy = tn−1 and, of course,

sy sy CI(a µ; β) = y¯ − √ tα/2,n−1 , y¯ + √ tα/2,n−1 . n n

¯ and s2y = a Sa. One should notice that, conveniently, y¯ = a x Realistically, one would have more than one such question to consider, and so that there would be r speciﬁed vectors ai ∈ Rp , i = 1, . . . , r, with corresponding constants ci ∈ R, i = 1, . . . , r, for which we would wish simultaneously to test H0 : a1 µ = c1 , . . . , ar µ = cr . This is clearly equivalent to testing H0 : A µ = c, where A = (a1 , . . . , ar ) and c = (c1 , . . . , cr ) . Letting yi = A xi , i = 1, . . . , n, then, clearly, y1 , . . . , yn are i.i.d. Nr (A µ, A ΣA). Under the assumption that the vectors a1 , . . . , ar are linearly independent, one may apply the multivariate results above directly to the y data so that y − c) S−1 y − c) = H0 is correct iﬀ n(¯ y (¯ d

(n − 1)r F (r, n − r) n−r

and a conﬁdence ellipsoid for ν = A µ is ¯ ) S−1 ¯ ) ≤ kα }, CR(ν; β) = {ν ∈ Rr : n(ν − y y (ν − y where kα =

(n − 1)r Fα (r, n − r). (n − r)

¯ = A x ¯ and Sy = A SA. Notice that y For obvious pragmatic reasons, one might in practice wish to have individual conﬁdence intervals for each component νi , i = 1, . . . , r. Thus, we would like to specify r intervals for these quantities in which we are simultaneously conﬁdent. The following lemma is needed. Lemma 8.1 Assume S ∈ Pp and A = (a1 , . . . , ar ) ∈ Rpr is of rank r. Then (ai x)2 ≤ x A(A SA)−1 A x ≤ x S−1 x, ∀x ∈ Rp . ai Sai

106

8. Tests on mean and variance

Proof. Using Rayleigh’s quotient (v. Problem 1.8.12),

x A(A SA)−1 A x = λ1 A(A SA)−1 A S = 1 −1 xS x x =0 sup

and the right inequality follows. Let y = A x = (y1 , y2 ) and B = A SA ∈ Pr partitioned as b11 b21 B= b21 B22 with inverse (v. Problem 1.8.1) −1 −1 b11 + b−2 −1 11 b21 B22.1 b21 B = −1 −1 −b11 B22.1 b21

−1 −b−1 11 b21 B22.1 −1 B22.1

.

Then, y B−1 y =

y12 y2 (a x)2 −1/2 −1/2 + ||y1 b11 B22.1 b21 − B22.1 y2 ||2 ≥ 1 = 1 , b11 b11 a1 Sa1 2

which is the left inequality. Since by simple algebra in Lemma 8.1, we have the inequalities (¯ yi − νi )2 ≤ (¯ y − ν) S−1 y − ν) ≤ (¯ x − µ) S−1 (¯ x − µ), y (¯ ai Sai we ﬁnd that

P

| y¯i − νi | 1/2 n ≤ kα , i = 1, . . . , r (ai Sai )1/2

≥ P n(¯ y − ν) S−1 y − ν) ≤ kα = β. y (¯ 1/2

Therefore, we are at least β × 100% conﬁdent in simultaneously presenting the r observed “Roy-Bose” intervals 1/2

1/2

kα kα (a Sai )1/2 ≤ νi ≤ y¯i + 1/2 (ai Sai )1/2 , i = 1, . . . , r. (8.2) n1/2 i n One should note that kα ≤ cα (why?), so we do somewhat better using kα . The constant cα , however, allows all possible linear combinations since y¯i −

¯ − a µ)2 (a x = (¯ x − µ) S−1 (¯ x − µ). a Sa a =0 sup

Therefore, we are at least β × 100% conﬁdent in simultaneously presenting all of the observed “Scheﬀ´e” intervals 1/2

1/2

cα cα ¯ + 1/2 (a Sa)1/2 , ∀a ∈ Rp . (a Sa)1/2 ≤ a µ ≤ a x (8.3) n1/2 n Although the “Scheﬀ´e” intervals are wider, they can be useful in making a great number of unplanned comparisons between means. ¯− a x

8.3. Simultaneous conﬁdence intervals on means

107

Actually, if the number, r, of questions that one asks is very small, we can sometimes even improve on kα . Let Ti = n1/2

(¯ y i − νi ) , i = 1, . . . , r. (ai Sai )1/2

d

Note Ti = tn−1 , i = 1, . . . , r, but they are not independent. Then, if we deﬁne the event Ai = {|Ti | ≤ t0 }, P (|Ti | ≤ t0 , i = 1, . . . , r)

= P (∩ri=1 Ai ) = 1 − P (∪ri=1 Aci ) r P (Aci ) ≥ 1−

(8.4)

i=1

=

1 − r P (|tn−1 | > t0 ) .

The inequality (8.4) is the Bonferroni inequality. If we deliberately equate the ﬁnal term to β = 1 − α and then solve for t0 , we ﬁnd that α P (tn−1 > t0 ) = , 2r or, equivalently, t0 = tα/2r,n−1 . Therefore, one can see that if we let bα = t2α/2r,n−1 , we will still be at least β × 100% conﬁdent if, instead of the “Roy-Bose” intervals (8.2), we present the r “Bonferroni” intervals 1/2

1/2

bα bα (ai Sai )1/2 ≤ νi ≤ y¯i + 1/2 (ai Sai )1/2 , i = 1, . . . , r. 1/2 n n Note that the relative length of “Roy-Bose” to “Bonferroni” is obviously bα /kα , and in a particular application, one would use the method with the shorter intervals. For non-normal data x1 , . . . , xn i.i.d. x with E x = µ and var x = Σ, large sample “Roy-Bose” and “Scheﬀ´e” simultaneous conﬁdence intervals can be constructed similarly (v. Problems 8.9.5 and 8.9.6) by appealing ﬁrst to the central limit theorem. y¯i −

8.3.2

Nonlinear hypotheses

In certain experiments, one might wish to compare the various components of µ to each other in ways that are plainly nonlinear. For instance, one may ask: “Is µ1 equal to µ23 ?” “Is the diﬀerence between µ2 and the product of µ1 and µ3 equal to 5.7?” The ﬁrst question corresponds to the hypothesis H0 : µ1 − µ23 = 0

108

8. Tests on mean and variance

and the second to H0 : µ2 − µ1 µ3 = 5.7. In each case, there is a certain function g : Rp → R and constant c ∈ R for which we are entertaining the hypothesis H0 : g(µ) = c. If we actually had r such hypotheses to consider in the same experiment, then there would, of course, be r speciﬁed real-valued functions gi , i = 1, . . . , r, with constants ci ∈ R, i = 1, . . . , r, for which we would wish simultaneously to test H0 : g1 (µ) = c1 , . . . , gr (µ) = cr . By letting g : R → Rr deﬁned by g(x) = (g1 (x), . . . , gr (x)) be continuously diﬀerentiable at µ, and c ∈ Rr , this is simply equivalent to p

H0 : g(µ) = c. The results in this section are asymptotic, so we assume possibly nonnormal data x1 , . . . , xn i.i.d. x with E x = µ and var x = Σ. One may not apply the speciﬁc results in Section 8.3.1 directly to this situation, but one may yet apply the same essential logic as formulated in that section √ d by appealing to the central limit theorem, n(¯ x − µ) → Np (0, Σ), and the √ d x) − g(µ)) → Nr (0, Σg ), where delta method in Proposition 6.2, n (g(¯ x), ν = g(µ), and Σg = [Dg(µ)]Σ[Dg(µ)] . This time, if we let y = g(¯ x)]S[Dg(¯ x)] then, as on page 79, Sg = [Dg(¯ 2 n(y − ν) S−1 g (y − ν) → χr . d

Thus, with dα = χ2α,r , the conﬁdence ellipsoid for ν, CR(ν; β) = {ν ∈ Rr : n(ν − y) S−1 g (ν − y) ≤ dα }, has an asymptotic coverage probability of β, i.e., P (ν ∈ CR(ν; β)) → β as n → ∞. To have individual conﬁdence intervals on each component νi , i = 1, . . . , r, we have the inequality in Lemma 8.1: (yi − νi )2 ≤ (y − ν) S−1 g (y − ν). Sg,ii From this purely algebraic fact, it follows that 1/2 | yi − νi | 1/2 ≤ dα , i = 1, . . . , r ≥ β. lim P n 1/2 n→∞ Sg,ii

8.4. Multiple correlation

109

Thus, asymptotically, we are at least β × 100% conﬁdent in simultaneously presenting the r observed intervals 1/2

yi −

1/2

dα dα 1/2 1/2 Sg,ii ≤ νi ≤ yi + 1/2 Sg,ii , i = 1, . . . , r. 1/2 n n

If r is quite small, one might try to improve on dα using a Bonferroni approach. The construction of simultaneous conﬁdence intervals on functions φ(Σ) is treated quite generally in D¨ umbgen (1998). Asymptotic considerations for the Wishart model show that the resulting conﬁdence bounds are substantially smaller than those obtained by inverting likelihood ratio tests.

8.4 Multiple correlation

|=

The multiple correlation coeﬃcient R is the maximum correlation possible between a variable x1 and a linear combination, t x2 , of a vector x2 . Not surprisingly, with underlying normality, the likelihood ratio test of H0 : ˆ x2 will be a function of the sample multiple correlation coeﬃcient R. x1 Assume x1 ∈ R and x2 ∈ Rp−1 have a joint normal distribution, σ11 σ 21 x1 ∼ Np 0, , x2 σ 21 Σ22 where Σ22 = A2 > 0. We have set the mean to 0 without any loss of generality. Since the simple correlation coeﬃcient is invariant to rescaling of each variable, we can assume at the outset that var t x2 = t Σ22 t = 1 and solve max cor(x1 , t x2 ).

t Σ22 t=1

For any t such that t Σ22 t = 1, cor2 (x1 , t x2 ) = (σ 21 t)2 /σ11 = A−1 σ 21 , At2 /σ11 ≤ σ 21 Σ−1 22 σ 21 /σ11 .

|=

The last inequality follows from the Cauchy-Schwarz inequality given in Proposition 1.1. It is an equality iﬀ At ∝ A−1 σ 21 , or, equivalently, 2 t ∝ Σ−1 22 σ 21 . The maximum correlation possible is R ≥ 0, where R = −1 σ 21 Σ22 σ 21 /σ11 , and is called the multiple correlation coeﬃcient between x1 and x2 . It should be noted immediately that the maximum correlation is achieved by t x2 = E (x1 | x2 ) = σ 21 Σ−1 22 x2 , i.e., by the conditional mean of x1 given x2 . In order to test H0 : R = 0 (equivalently, H0 : σ 21 = 0 or x2 ), the sample variance, based on a random sample of size n, H0 : x1 v11 v21 (n − 1)S ≡ V = , v21 V22

110

8. Tests on mean and variance

is partitioned and is distributed as V ∼ Wp (n − 1, Σ). In the obvious ˆ ≥ 0 is called ˆ 2 = v V−1 v21 /v11 and R manner, the sample version is R 21 22 the sample multiple correlation coeﬃcient. Proposition 8.5 The likelihood ratio test Λ rejects H0 for small values of ˆ 2 )n/2 . Λ = (1 − R Proof. Based on the likelihood (7.1) from x1 , . . . , xn i.i.d. Np (µ, Σ), Σ > ˆ =x ¯ . Without constraints, the MLE of Σ is 0, the MLE of µ is always µ ˆ = 1 V, but when σ 21 = 0, the constrained MLE becomes Σ n 1 v11 0 ˆ ˆ . Σ= 0 V22 n Thus, ˆ ˆ µ) ˆ L(Σ, Λ= ˆ ˆ L(Σ, µ)

−1

ˆ ˆ ˆ ) ˆ −n/2 etr(− 1 VΣ |Σ| 2 = ˆ −n/2 etr(− 1 VΣ ˆ −1 ) |Σ| 2 −n/2 exp(− 12 np) v11 |V22 | = |V| exp(− 12 np) n/2 v11.2 ˆ 2 )n/2 , = = (1 − R v11

where the last equality made use of |V| = v11.2 |V22 |.

2

ˆ 2 in which negative binomial Of greater interest is the distribution of R probabilities intervene. The reader should recall at this point that a negative binomial variable represents the number of failures, k, before the rth success in a sequence of independent bernoulli trials. Deﬁnition 8.1 Negative binomial: x ∼ nb(r, p), r > 0 and 0 ≤ p ≤ 1, iﬀ the probability function of x is given by r+k−1 r pk = P (x = k) = p (1 − p)k , k = 0, 1, . . . . k In Deﬁnition 8.1, r need not be an integer. In that case, the combination factor is calculated via the gamma function: r+k−1 Γ(r + k) (r)k = , = k k!Γ(r) k! where (r)0 = 1 and (r)k = r(r + 1) · · · (r + k − 1) for k = 1, 2, . . .. Recall that Fc (s1 , s2 ) denotes the canonical Fc distribution (v. Deﬁnition 3.7). Proposition 8.6 ˆ2 R P ≤t = ˆ2 1−R

∞ k=0

pk · P (Fc (p − 1 + 2k, n − p) ≤ t) ,

8.4. Multiple correlation

ˆ2 ≤ t P R

=

∞

111

pk · P beta 12 (p − 1 + 2k); 12 (n − p) ≤ t ,

k=0

|=

where pk are the negative binomial probabilities

1 2 (n − 1) k (1 − R2 )(n−1)/2 R2k , k = 0, 1, . . . . pk = k! ˆ 2 ), then R ˆ 2 /(1 − R ˆ 2 ) = v V−1 v21 /v11.2 . Proof. Since v11.2 = v11 (1 − R 21 22 2 (v21 , V22 ) and v11.2 ∼ σ11.2 χn−p . With the help of Proposition 7.9, v11.2 We also have v21 | V22 ∼ Np−1 (V22 Σ−1 22 σ 21 , σ11.2 V22 ) from which −1/2

−1/2

σ11.2 V22

−1/2

1/2

v21 | V22 ∼ Np−1 (σ11.2 V22 Σ−1 22 σ 21 , I)

−1 V22 v21 | V22 ∼ σ11.2 χ2p−1 (δ), where and, therefore, v21 −1 δ = σ 21 Σ−1 22 V22 Σ22 σ 21 /(2σ11.2 ).

Hence, conditional on V22 , ˆ 2 ) ∼ Fc (p − 1, n − p; δ). ˆ 2 /(1 − R R Using Proposition 4.7, ∞ ˆ2 R δk ≤ t | V22 = e−δ P (Fc (p − 1 + 2k, n − p) ≤ t) . P ˆ2 k! 1−R k=0

To obtain the unconditional distribution, take expectations on both sides with respect to the distribution of V22 . First, we need the distribution of δ. Since V22 ∼ Wp−1 (n − 1, Σ22 ), then R2 R2 d 1 2 1 δ∼ χ = G 2 (n − 1), . (1 − R2 ) 2 n−1 (1 − R2 ) The expectation computation is immediate (v. Problem 8.9.10) if we use a result well known in bayesian inference [Johnson et al. (1992), p. 204] that if K given δ is Poisson(δ) and δ ∼ G(p, θ), then the marginal of K is negative binomial, K ∼ nb(p, (1 + θ)−1 ). Hence, δk , k! completing the proof of the ﬁrst result. The second result follows with the obvious monotone transformation. 2 pk = P (K = k) = E P (K = k | δ) = E e−δ

ˆ 2 ) is distributed as a negative binomial mixture of canonical ˆ 2 /(1−R Thus, R ˆ 2 is a negative binomial mixture of beta Fc distributions, whereas that of R ˆ distributions. The moments of R (v. Problem 8.9.9) follow directly from the later characterization. The null distribution is just a special case.

112

8. Tests on mean and variance

ˆ 2 /(1 − R ˆ 2 ) ∼ Fc (p − 1, n − p). Proposition 8.7 Assuming R = 0, R The exact distribution of the simple correlation coeﬃcient, introduced earlier in Section 5.6.3, when ρ = 0 is just another special case when p = 2. The invariance of the multiple correlation coeﬃcient is discussed in Problem 8.9.13. Proposition 8.8 For testing H0 : R = 0 against H1 : R > 0, the test ˆ is UMPI. which rejects for large values of R ˆ is clearly invariant and it was established in ProbProof. The statistic R ˆ lem 8.9.13 that all invariant tests, depending on (¯ x, V), are a function of R. The problem thus reduces to ﬁnding the UMP test based on one observation ˆ The density of x ≡ R ˆ 2 follows from Proposition 8.6, from R. f (x; R2 ) =

∞ k=0

pk

B( 12 (p

1 1 1 x 2 (p−1+2k)−1 (1 − x) 2 (n−p)−1 , 1 − 1 + 2k), 2 (n − p))

0 < x < 1. From the Neyman-Pearson lemma, the most powerful test rejects H0 for large values of the ratio

∞ Γ 12 (n − 1 + 2k) k f (x; R2 ) x ≥ c2 . = c1 pk 1 f (x; 0) Γ 2 (p − 1 + 2k) k=0 Since this ratio is monotone increasing in x this is equivalent to rejecting H0 for large values of x, x ≥ c3 . This rejection region does not depend on R and, thus, the test is uniformly most powerful. 2 Example 8.2 The power function of the likelihood ratio test for H0 : R = 0 may be evaluated with Proposition 8.6: 1 ∞

−1 pk β = B 12 (p − 1 + 2k), 12 (n − p) tα k=0 (p−1+2k)/2−1

·x (1 − x)(n−p)/2−1 dx,

1 where tα = betaα 2 (p − 1); 12 (n − p) is the critical point. A numerical evaluation in Mathematica of β for p = 3 and n = 20 at the signiﬁcance level α = 0.05 gave the plot in Figure 8.2. For large samples, the asymptotic distribution provides a simpler disˆ 2 is asymptotically tribution. By the delta method in Proposition 6.2, R normal since it is a function of the sample variance S, which is itself asymptotically normal. However, rather than calculating the derivatives, it is somewhat easier to use op (n−1/2 ) asymptotic expansions. This technique is illustrated in the following proof. ˆ2, Proposition 8.9 The null and alternative asymptotic distributions of R when sampling from a multivariate normal distribution, are given by

8.4. Multiple correlation

113

beta 1 0.8 0.6 0.4 0.2

0.2

0.4

0.6

0.8

R

Figure 8.2. Power function of the likelihood ratio test for H0 : R = 0 when p = 3, and n = 20 at a level of signiﬁcance α = 0.05.

d ˆ 2 − R2 ) → (i) n1/2 (R N 0, 4R2 (1 − R2 )2 , d

ˆ 2 → χ2 . (ii) If R = 0, then nR p−1 Proof. By invariance arguments (v. Problem 8.9.13), assume without loss of generality, 1 Re1 , Σ= Re1 Ip−1 where e1 = (1, 0, . . . , 0) ∈ Rp−1 . Since n1/2 (S − Σ) → Z = d

z11 z21

z21 Z22

,

where Z ∼ Npp (0, (I + Kp )(Σ ⊗ Σ)), then we can write the op (n−1/2 ) expansions 1 + n−1/2 z11 + op (n−1/2 ),

s11

=

s21

= Re1 + n−1/2 z21 + op (n−1/2 ),

S22

= I + n−1/2 Z22 + op (n−1/2 ), p

where op (n−1/2 ) is such that n1/2 · op (n−1/2 ) → 0 [Serﬂing (1980), p. 9]. Straightforward algebra, with the aid of Problem 1.8.15, then gives s21 S−1 22 s21 s11 = [1 + n−1/2 z11 + op (n−1/2 )]−1 · [Re1 + n−1/2 z21 + op (n−1/2 )]

114

8. Tests on mean and variance

·[I + n−1/2 Z22 + op (n−1/2 )]−1 · [Re1 + n−1/2 z21 + op (n−1/2 )] =

[1 − n−1/2 z11 + op (n−1/2 )] · [Re1 + n−1/2 z21 + op (n−1/2 )] ·[I − n−1/2 Z22 + op (n−1/2 )] · [Re1 + n−1/2 z21 + op (n−1/2 )]

= R2 + 2n−1/2 Rz21 − n−1/2 R2 z22 − n−1/2 R2 z11 + op (n−1/2 ). d ˆ 2 −R2 ) → Thus, n1/2 (R 2Rz21 −R2 z22 −R2 z11 , but since (v. equation (6.1)) (z21 , z11 , z22 ) ∼ N3 (0, Ω), where 1 + R2 2R 2R 2 2R2 , Ω = 2R 2 2R 2R 2

the linear combination with a = (2R, −R2 , −R2 ) yields 2Rz21 − R2 z22 − R2 z11 ∼ N (0, a Ωa), whereby a direct evaluation provides a Ωa = 4R2 (1−R2 )2 . This proves (i). d

To prove (ii), note that when R = 0, n1/2 s21 → z21 , where z21 ∼ Np−1 (0, I). p

p

d −1/2 −1/2 S22 s21 →

However, since S22 → I and s11 → 1, then n1/2 s11 d d ˆ2 → nR |z21 |2 = χ2p−1 .

z21 and 2

ˆ and of Fisher’s As a corollary, we get the asymptotic distribution of R z-transform. ˆ and of its Fisher’s z Corollary 8.2 The asymptotic distributions of R transform, when sampling from a multivariate normal distribution, are given by

d ˆ − R) → N 0, (1 − R2 )2 , (i) n1/2 (R d ˆ − tanh−1 (R) → (ii) n1/2 tanh−1 (R) N (0, 1). Proof. It follows directly from the delta method applied to the square root 2 transformation and to the tanh−1 transformation. More general results on the asymptotic distributions of correlation coefﬁcients obtained from any asymptotically normal equivariant estimate of variance, not necessarily S, will be given in Chapter 13 for a sample from an elliptical distribution.

8.4.1

Asymptotic moments

ˆ 2 in Proposition 8.6 can be used The mixture beta characterization of R ˆ 2 in terms of those of beta to obtain immediately the exact moments of R distributions (v. Problem 8.9.9). Simple approximations for large n are, however, possible as is now shown.

8.4. Multiple correlation

115

Using Proposition 8.6, we have ∞

2 ˆ pk · P beta 12 (n − p); 12 (p − 1 + 2k) ≤ t . P 1−R ≤t = k=0

From the moments of x ∼ beta(a, b) given by E xh =

(a)h Γ(a + b)Γ(a + h) = , h = 1, 2, . . . , Γ(a)Γ(a + b + h) (a + b)h

we can write ˆ2 h

E (1 − R ) =

∞

1

2 (n

k=0

1 − 1) k 2 (n − p) h 2 (n−1)/2 2k . (1 − R ) R 1 k! 2 (n − 1 + 2k) h

The hypergeometric function 2 F1 (a, b; c; z)

≡

∞ (a)k (b)k z k (c)k k!

k=0

after some simple algebra allows to write

1 2 (n − p) h 2 h ˆ E (1 − R ) = 1 2 (n − 1) h ·(1 − R2 )(n−1)/2 2 F1 ( 12 (n − 1), 12 (n − 1); 12 (n − 1) + h; R2 ). Then, upon using Kummer’s formula [Erd´elyi et al. (1953), p. 105] 2 F1 (a, b; c; z)

we ﬁnally ﬁnd ˆ2 h

E (1 − R ) =

1

(n

21 2 (n

= (1 − z)(c−a−b) 2 F1 (c − a, c − b; c; z), − p) h (1 − R2 )h 2 F1 (h, h; 21 (n − 1) + h; R2 ). − 1) h

For h = 1, we then obtain ˆ2) E (1 − R

= =

(n − p) (1 − R2 ) 2 F1 (1, 1; 12 (n + 1); R2 ) (n − 1) 2R2 (n − p) 2 −2 (1 − R ) 1 + + O(n ) (n − 1) (n + 1)

and ˆ 2 = R2 + (p − 1) (1 − R2 ) − 2 (n − p) R2 (1 − R2 ) + O(n−2 ). ER (n − 1) (n2 − 1) ˆ 2 is biased, as it overestimates R2 . The MVUE This expression shows that R of R2 [Olkin and Pratt (1958)] is ˆ 2 ) 2 F1 (1, 1; 1 (n − p + 2); 1 − R ˆ 2 ). ˆ 2 ) = 1 − (n − 3) (1 − R U (R 2 (n − p)

116

8. Tests on mean and variance

ˆ is close to The MVUE has the drawback of taking negative values when R 0. In fact, using the relation [Erd´elyi et al. (1953), p. 61] 2 F1 (a, b; c; 1)

=

Γ(c) Γ(c − a − b) , Γ(c − a) Γ(c − b)

it is easily established that U (0) = −(p−1)/(n−p−2) and, of course, U (1) = 1. A similar expansion for h = 2 can be done to obtain an asymptotic ˆ2. expansion for var R

8.5 Partial correlation Assume two subsets of variables x1 ∈ Rp1 and x2 ∈ Rp2 have a joint normal distribution, µ1 x1 Σ11 Σ12 ∼ Np , . x2 µ2 Σ21 Σ22 The partial correlation coeﬃcient between variables xi and xj , in the subset x1 , is just the ordinary simple correlation ρ between xi and xj but with the variables in the subset x2 held ﬁxed. This will be denoted by ρij|x2 . It can be expressed in terms of Σ if one recalls the conditional normal of Section 5.5: x1 | x2 ∼ Np1 (µ1 + Σ12 Σ−1 22 (x2 − µ2 ), Σ11.2 ). Writing Σ11.2 = (σij|x2 ), where σij|x2 denotes the (i, j) element of Σ11.2 , then σij|x2 . ρij|x2 = 1/2 1/2 σii|x2 σjj|x2 Using Proposition 7.9, we already know that since (n − 1)S = V ∼ Wp (n − 1, Σ), then V11.2 ∼ Wp1 (n − 1 − p2 , Σ11.2 ), where V was partitioned in conformity as V11 V12 V= . V21 V22 ˆ of Σ, it is clear that the MLE of Since V is proportional to the MLE Σ ρij|x2 is just rij|x2 =

vij|x2 1/2 1/2 vii|x2 vjj|x2

,

where V11.2 = (vij|x2 ) and vij|x2 denotes the (i, j) element of V11.2 . Considering the distribution of V11.2 , the distribution of rij|x2 is the same as

8.6. Test of sphericity

117

for a simple correlation coeﬃcient but with n − p2 in place of n. We have proved: Proposition 8.10 2 rij|x2 P ≤t = 2 1 − rij|x 2 2 P rij|x ≤t 2

=

∞ k=0 ∞

pk · P (Fc (1 + 2k, n − p2 − 2) ≤ t) ,

pk · P beta 12 (1 + 2k); 12 (n − p2 − 2) ≤ t ,

k=0

where pk are the negative binomial probabilities

1 2 (n − p2 − 1) k (1 − ρ2ij|x2 )(n−p2 −1)/2 ρ2k pk = ij|x2 , k = 0, 1, . . . . k! For large samples, as for the simple correlation coeﬃcient, it follows from Problem 6.4.8 that

d n1/2 rij|x2 − ρij|x2 → N 0, (1 − ρ2ij|x2 )2 . A Fisher’s z-transform as for the simple correlation coeﬃcient in Problem 6.4.9 is deﬁnitely possible for a partial correlation coeﬃcient.

8.6 Test of sphericity Assume x ∼ Np (µ, Σ), Σ > 0, and consider testing the hypothesis that the p variables in x = (x1 , . . . , xp ) are independent and have the same variance: H0 : Σ = γI, γ > 0. Based on a random sample x1 , . . . , xn , regardless of the hypothesis H0 , as ˆ =x ¯ . Now, without constraint, long as Σ > 0, the MLE of µ is always µ ˆ = 1 V, where, as usual, the MLE of Σ is Σ n V=

n

¯ )(xi − x ¯ ) . (xi − x

i=1

However, under H0 , the MLE is obtained by solving

max |γI|−n/2 etr − 12 γ −1 V . γ>0

Taking logarithms, the function to maximize is − 12 np ln γ − 12 γ −1 tr V,

118

8. Tests on mean and variance

and the solution is easily calculated, γˆ = tr V/np. Therefore, the likelihood ratio, ﬁrst derived by Mauchly (1940), becomes

ˆ |ˆ γ I|−n/2 etr − 12 γˆ −1 V L(ˆ γ I, µ) = Λ = ˆ µ) ˆ −n/2 etr − 1 Σ ˆ −1 V ˆ L(Σ, |Σ| 2 n/2 exp(− 12 np) | n1 V| = . 1 |( np tr V)I| exp(− 12 np) Thus, ˜ ≡ Λ2/n Λ

|V| = 1 = ( p tr V)p

1/p p i=1 li p 1 i=1 li p

p ,

where l1 ≥ · · · ≥ lp are the ordered eigenvalues of V. The LRT compares the geometric and arithmetic means of those eigenvalues; they coincide when V has the structure as in H0 . Proposition 8.11 The LRT for testing H0 : Σ = γI, γ > 0 against ˜ = |V|/( 1 tr V)p . H1 : Σ > 0 rejects H0 for small values of Λ p At this point, we remind the reader about the general expression for the asymptotic degrees of freedom for likelihood ratio tests. In general, for testing H0 : θ ∈ Θ0 against H1 : θ ∈ Θc0 under regularity conditions, then, d

under H0 , −2 ln Λ → χ2f as the sample size n → ∞. The degrees of freedom f is the diﬀerence between the number of free parameters in Θ = Θ0 ∪ Θc0 and the number of free parameters in Θ0 . From the general theory of LRT, it is clear that the asymptotic null distribution is −2 ln Λ → χ2f , f = 12 p(p + 1) − 1. d

Lemma 8.2 When Σ = γI, γ > 0, tr V

|=

Better approximations can be obtained by calculating the moments of Λ (or ˜ as in Section 12.3. The moments are easily calculated with the following Λ) lemma. |V|/(tr V)p .

Proof. When Σ = γI, clearly the distribution of |V|/(tr V)p does not depend on γ. From the likelihood for (µ, γ), which forms an exponential family, the minimal suﬃcient and complete statistic is (¯ x, tr V). The 2 conclusion follows using Basu’s1 theorem.

|=

1 Basu’s theorem: If T is complete and suﬃcient for the family P = {P : θ ∈ Θ}, θ A, for any ancillary statistic A. By deﬁnition, a statistic A is ancillary iﬀ its then T distribution does not depend on θ.

8.6. Test of sphericity

But then, since

˜ Λ

then

1 tr V p

˜h · E EΛ

119

p = |V|,

1 tr V p

ph = E |V|h ,

from which ˜h = EΛ

E |V|h . E ( p1 tr V)ph

Proposition 8.12 When Σ = γI, γ > 0, then

1 Γp 12 (n − 1) + h h ph Γ 2 (n − 1)p ˜

E Λ =p . Γ 12 (n − 1)p + ph Γp 12 (n − 1) Proof. We have V ∼ γWp (n − 1). The proof follows directly from the above remark in conjunction with Corollary 7.3, Proposition 7.2, and the deﬁnition of Γp (·) on page 93. Moments of chi-square distributions can be obtained from Section 3.2 (v. Problem 3.5.1). 2 ˜ can be characterized as a product of Finally, the exact distribution of Λ independent beta variables [Srivastava and Khatri (1979), p. 209]. ˜ is Proposition 8.13 The exact null distribution of Λ d ˜= Λ

p−1

beta[ 12 (n − 1 − i), i( 12 + p1 )];

(8.5)

i=1

˜ is distributed as the product of p − 1 mutually independent beta i.e., Λ variables. Proof. We make use of the multiplicative formula of Gauss [Erd´elyi et al. (1953), p. 4] 1

Γ(mz) = (2π)−(m−1)/2 mmz− 2

m−1

Γ(z +

r m ),

m = 2, 3, . . . ,

r=0

with m = p and z = 12 (n − 1), 12 (n − 1) + h. We can then rewrite the moments as p−1 1 p r 1 1 r=0 Γ[ 2 (n − 1) + p ] i=1 Γ[ 2 (n − 1) + h − 2 (i − 1)] h ˜ p = EΛ 1 1 p−1 1 r i=1 Γ[ 2 (n − 1) − 2 (i − 1)] r=0 Γ[ 2 (n − 1) + h + p ] p−1 1 1 1 i i=1 Γ[ 2 (n − 1) + h − 2 i] Γ[ 2 (n − 1) + p ] . = p−1 1 1 1 i i=1 Γ[ 2 (n − 1) − 2 i] Γ[ 2 (n − 1) + h + p ]

120

8. Tests on mean and variance

It is then straightforward to check that all moments of order h > 0 on the d left and right sides of = in (8.5) are the same. Since the domain is the bounded interval [0, 1], there is a unique distribution with these moments [Serﬂing (1980), p. 46]. 2 The group Op × Rp × (R\{0}) transforms the data as xi → aHxi + b, for any H ∈ Op , b ∈ Rp , and a = 0. It preserves normality and induces transformations on the minimal suﬃcient statistic (¯ x, V) and parameters (µ, Σ) ¯ → aH¯ as x x + b, V → a2 HVH , µ → aHµ + b, and Σ → a2 HΣH . Thus, the transformation also preserves the sphericity. A test function f (¯ x, V) is said to be invariant with respect to this group of transformations when it takes the same value on the original data as on the transformed data, i.e., f (y, W) = f (aHy + b, a2 HWH ), ∀(H, b, a) ∈ Op × Rp × (R\{0}), ∀(y, W) ∈ Rp × Pp . This invariance property yields formidable simpliﬁcations. First, if we diagonalize V = HDH , where D = diag(l1 , . . . , lp ), then −1/2 ¯ , we ﬁnd choosing a = lp and b = −aH x f (¯ x, V)

¯ + b, a2 H VH) = f (aH x ˜ = f (0, D),

˜ = diag(l1 /lp , . . . , lp−1 /lp , 1) depends on the sample only through where D the ratios l1 /lp , . . . , lp−1 /lp . So, any invariant test can be written as a function of li /lp , i = 1, . . . , p − 1. Second, if we diagonalize Σ = GDλ G , where Dλ lists the eigenvalues −1/2 and b = −aG µ, λ1 ≥ · · · ≥ λp on its diagonal, then choosing a = λp we ﬁnd ˜ ), ¯ + b ∼ Np (0, n−1 D aG x λ 2 ˜ a G VG ∼ Wp (n − 1, Dλ ), ˜ where D λ = diag(λ1 /λp , . . . , λp−1 /λp , 1). Thus, the non-null distribution of any invariant test depends on (µ, Σ) only through the ratios λ1 /λp , . . . , λp−1 /λp . These invariance results are summarized in a proposition. Proposition 8.14 With respect to the above group of transformations, any invariant test depends on the minimal suﬃcient statistic (¯ x, V) only through the ratios l1 /lp , . . . , lp−1 /lp of eigenvalues of V. The power function of any invariant test depends on (µ, Σ) only through the ratios λ1 /λp , . . . , λp−1 /λp of eigenvalues of Σ. The LRT is obviously invariant. There is no uniformly most powerful invariant (UMPI) test for the sphericity hypothesis, but John (1971) showed the test based on J = tr V2 /(tr V)2 is locally most powerful (best)

8.7. Test of equality of variances

121

invariant (LBI). The null distribution of the LBI test is given in John (1972).

8.7 Test of equality of variances The data consist of a independent a samples xi1 , . . . , xini i.i.d. Np (µi , Σi ), Σi > 0, i = 1, . . . , a. Let n = i=1 ni be the total number of observations. The hypothesis in question here is the equality of variances H0 : Σ1 = · · · = Σa which is being tested against all alternatives. Since the samples are independent, the likelihood function can be built immediately from (7.1), L(Σ1 , . . . , Σa , µ1 , . . . , µa ) a ' & ni , |Σi |− 2 etr − 12 [Vi + ni (¯ xi − µi )(¯ xi − µi ) ] Σ−1 ∝ i i=1

where, as usual, ¯i x

=

Vi

=

ni 1 xij , ni j=1 ni

¯ i )(xij − x ¯ i ) , i = 1, . . . , a. (xij − x

j=1

Without the restriction speciﬁed in H0 , the parameters are unrelated and, ˆ i = 1 Vi . Under ¯ i and Σ ˆi = x thus, the unrestricted MLE is just the usual µ ni H0 , however, we have Σ1 = · · · = Σa = Σ, for some unknown Σ, and, thus, ˆ ˆˆ = x ˆ = 1 V, where V = a Vi pools all the variances together. ¯ i and Σ µ i i=1 n Thus, the LRT becomes Λ = = =

¯1, . . . , x ¯a) L( n1 V, . . . , n1 V, x ¯1, . . . , x ¯a) L( n11 V1 , . . . , n1a Va , x a | 1 V|−ni /2 exp(− 12 np) ai=1 1n −ni /2 exp(− 1 np) i=1 | ni Vi | 2 a ni /2 pn/2 n i=1 |Vi | . a pni /2 |V|n/2 n i i=1

Proposition 8.15 The LRT for testing H0 : Σ1 = · · · = Σa rejects the hypothesis for small values of a |Vi |ni /2 npn/2 Λ = i=1 n/2 . a pn /2 |V| n i i=1

i

122

8. Tests on mean and variance

The group Gp × (Rp )a transforms the observations as xij → Axij + bi , for any A ∈ Gp and bi ∈ Rp , i = 1, . . . , a. This obviously preserves the ¯ i → A¯ xi + bi , Vi → AVi A , normality and transforms the statistics as x and V → AVA , and induces the parameter transformation Σi → AΣi A . The hypothesis H0 is thus also preserved by this group of transformations. Therefore, the LRT statistic evaluated at the transformed data Axij + bi is a ni /2 npn/2 i=1 |AVi A | Λ = pni /2 a |AVA |n/2 i=1 ni a ni /2 npn/2 i=1 |Vi | = , pni /2 a |V|n/2 i=1 ni which is identical to the LRT statistic evaluated at the original data xij . We say that the LRT is invariant with respect to this group of transformations. ¯ a , V1 , . . . , Va ) is termed invariant iﬀ In general, a test function f (¯ x1 , . . . , x f (y1 , . . . , ya , W1 , . . . , Wa ) = f (Ay1 + b1 , . . . , Aya + ba , AW1 A , . . . , AWa A ), ∀(A, b1 , . . . , ba ) ∈ Gp ×(Rp )a , ∀(y1 , . . . , ya , W1 , . . . , Wa ) ∈ (Rp )a ×(Pp )a . This has important consequences. First, by deliberately choosing A = Σ−1/2 , where Σ1 = · · · = Σa = xi , it is clear that AVi A ∼ Wp (ni − 1) do Σ under H0 , and bi = −A¯ not involve any unknown parameters. Thus, the null distribution of any invariant test function ¯ a , V1 , . . . , V a ) f (¯ x1 , . . . , x = f (A¯ x1 + b1 , . . . , A¯ xa + ba , AV1 A , . . . , AVa A ) = f (0, . . . , 0, AV1 A , . . . , AVa A ) such as Λ is parameter free. Note that we need only consider test ¯ a , V1 , . . . , Va ) is suﬃcient for functions of this form since (¯ x1 , . . . , x (µ1 , . . . , µa , Σ1 , . . . , Σa ). −1/2 −1/2 Second, in the special case a = 2, diagonalize V1 V2 V1 = HDH , where H ∈ Op and D = diag(l1 , . . . , lp ) contains the eigenvalues of V1−1 V2 . −1/2 xi , we ﬁnd that This time by deliberately choosing A = H V1 , bi = −A¯ for any invariant test ¯ 2 , V1 , V2 ) f (¯ x1 , x

= f (0, 0, AV1 A , AV2 A ) = f (0, 0, I, D)

is a function of l1 , . . . , lp only. Thus, any invariant test function depends on −1/2 −1/2 = the data only through l1 , . . . , lp . Similarly, diagonalizing Σ1 Σ2 Σ1 GDλ G , where G ∈ Gp and Dλ contains the eigenvalues λ1 , . . . , λp of −1/2 , AV1 A ∼ Wp (n1 − 1) and Σ−1 1 Σ2 , we have after choosing A = G Σ1

8.7. Test of equality of variances

123

AV2 A ∼ Wp (n2 − 1, Dλ ). Thus, the non-null distribution of any invariant test function depends on (µ1 , µ2 , Σ1 , Σ2 ) only through the eigenvalues of Σ−1 1 Σ2 . Proposition 8.16 With respect to the group of transformations Gp × (Rp )2 , any invariant test for testing H0 : Σ1 = Σ2 depends on ¯ 2 , V1 , V2 ) only through the eigenvalues l1 , . . . , lp of V1−1 V2 . The (¯ x1 , x power function of any invariant test depends on (µ1 , µ2 , Σ1 , Σ2 ) only through the eigenvalues λ1 , . . . , λp of Σ−1 1 Σ2 . For example, the LRT when a = 2 can be written Λ=

npn/2 pn1 /2 pn2 /2 n2

n1

p

n /2

li 2 . (1 + li )n/2 i=1

An alternative invariant test function [Nagao (1973)] is 2 a n ni tr Vi V−1 − I . N = 12 ni i=1 Continuing now with the moments of the null distribution of the LRT, we comment ﬁrst on a result of unbiasedness. Although the LRT is a biased test, Perlman (1980) proved that the slight modiﬁcation a |Vi |mi /2 mpm/2 ∗ , Λ = i=1 m/2 a pmi /2 |V| i=1 mi where the sample sizes ni are replaced by the corresponding degrees of a freedom mi = ni − 1 and m = i=1 mi = n − a, yields an unbiased test. We will, thus, concentrate on the latter. It was Bartlett (1937) who ﬁrst proposed the use of the modiﬁed LRT, Λ∗ . For a = 2, unbiasedness of Λ∗ was established earlier by Sugiura and Nagao (1968), whereas Srivastava, Khatri, and Carter (1978) proved a monotonicity property stronger than unbiasedness. The null moments of Λ∗ is a simple consequence of invariance coupled with the normalizing constant cp,m = [2mp/2 Γp ( 12 m)]−1 of a Wp (m) p.d.f. Proposition 8.17 Under H0 , the moments of the modiﬁed LRT Λ∗ are given by a Γp ( 12 m) Γp [ 12 mi (1 + h)] . pmi h/2 Γ [ 1 m(1 + h)] Γp ( 12 mi ) p 2 i=1 i=1 mi

E Λ∗ h = a

mpmh/2

Proof. Under H0 , by invariance, we can assume Σi = I and Vi ∼ Wp (mi ) are independently distributed. Thus, from the Wp (mi ) densities, we have a ∗h cp,mi ··· |V|−mh/2 EΛ = i=1

V1 >0

Va >0

124

8. Tests on mean and variance

·

a

|Vi |[mi (1+h)−p−1]/2 etr(− 12 Vi )dV1 · · · dVa .

i=1

The integrand is seen to contain the p.d.f. of Vi ∼ Wp (mi (1 + h)) independently distributed. However, when this is the case V ∼ Wp (m(1 + h)). Thus, we ﬁnd a cp,mi ∗h E |V|−mh/2 , E Λ = a i=1 i=1 cp,mi (1+h) where V ∼ Wp (m(1 + h)). Using the moments of the generalized variance in Problem 7.5.6 and simplifying, the conclusion is reached. 2 An accurate approximation to the null distribution of the modiﬁed LRT Λ∗ by asymptotic expansion of high order is discussed in Example 12.4.

8.8 Asymptotic distributions of eigenvalues Based on a random sample x1 , . . . , xn from Np (µ, Σ), n several tests on the ¯ )(xi − x ¯ ) ∼ variance Σ are a function of the eigenvalues of V = i=1 (xi − x Wp (n−1, Σ). It was seen that an invariant test for sphericity, H0 : Σ = σ 2 I, depends only on (l1 /lp , . . . , lp−1 /lp ) where l1 ≥ · · · ≥ lp are the eigenvalues of V. Also, in the two independent samples problem, x11 , . . . , x1n1 x21 , . . . , x2n2

i.i.d. Np (µ1 , Σ1 ), i.i.d. Np (µ2 , Σ2 ),

an invariant test for the equality of variances, H0 : Σ1 = Σ2 , depends only on the eigenvalues of V1−1 V2 , where Vi =

ni

¯ i )(xij − x ¯ i ) ∼ Wp (ni − 1, Σi ), i = 1, 2. (xij − x

j=1

The distribution of eigenvalues of various random matrices thus plays an important role in testing hypotheses.

8.8.1

The one-sample problem

We investigate the asymptotic distribution of the eigenvalues l1 , . . . , lp of V ∼ Wp (m, Σ). We already know there exists H ∈ Op such that H ΣH = Λ, where Λ = diag(λ1 , . . . , λp ), and since V and H VH have the same eigenvalues, we can assume at the outset that Σ = Λ is diagonal. An eﬀective method for such problems is to write S=

V = Λ + m−1/2 V(1) , m

8.8. Asymptotic distributions of eigenvalues

125

where V(1) = m1/2 (S − Λ) is Op (1), and expand the eigenvalues of S, li /m, around λi in powers of m−1/2 . This is called the perturbation method [Bellman (1960), Kato (1982)]. We now clearly outline the steps to obtain an approximation, with remainder of the order O(m−1 ), to the distribution function of a nearly arbitrary function f (l/m) of l = (l1 , . . . , lp ) . Step 1: Perturbation method More generally, consider a diagonal matrix Λ = diag(λ1 , . . . , λp ) and assume that the perturbation of Λ can be expressed as a power series in as follows: R = Λ + V(1) + 2 V(2) + O(3 ), where V(j) , j = 1, 2, are symmetric and is a small real number. We shall discuss the case when λα is distinct from the other p − 1 eigenvalues. Let lα be the αth eigenvalue of R and cα = (c1α , . . . , cpα ) the corresponding normalized eigenvector with cαα > 0. The quantities lα and cα can be assumed of the form [Bellman (1960), p. 61] 2 (2) 3 = λα + λ(1) α + λα + O( ), p p (1) (2) = eα + aiα ei + 2 aiα ei + O(3 ),

lα cα

i=1

(8.6) (8.7)

i=1

where ei = (0, . . . , 1, . . . , 0) is the ith canonical basis vector. We determine (1) (2) (1) (2) the unknown coeﬃcients λα , λα , aiα , and aiα by substituting (8.6) and (8.7) into the equation Rcα = lα cα and equating the coeﬃcients of the powers of . This gives [Λ + V(1) + 2 V(2) + O(3 )][eα +

p

(1)

aiα ei + 2

p

i=1

2 (2) 3 = [λα + λ(1) α + λα + O( )][eα +

p

(2)

aiα ei + O(3 )]

i=1

(1)

aiα ei + 2

p

i=1

(2)

aiα ei + O(3 )],

i=1

and equating the coeﬃcients, we obtain the equations λ α eα p

(1)

(1) aiα λi ei + vα

= λ α eα , p (1) = aiα λα ei + λ(1) α eα ,

i=1

and p i=1

(2)

aiα λi ei +

(8.8) (8.9)

i=1

p i=1

=

(1) (1)

aiα vi

p i=1

(2)

+ vα

(2) aiα λα ei

+

p i=1

(1)

(2) aiα λ(1) α ei + λ α eα ,

(8.10)

126

8. Tests on mean and variance (j)

(j)

where V(j) = (v1 , . . . , vp ), j = 1, 2. The αth component of (8.9) yields (1) (1) (1) a(1) αα λα + vαα = aαα λα + λα (1)

(1)

from which λα = vαα . The component i = α of the same equation yields (1)

(1)

(1)

aiα λi + viα = aiα λα , (1)

(1)

from which aiα = −viα λiα , i = α, where λiα = 1/(λi − λα ). (1)

(1)

Note that aαα can be chosen arbitrarily and we set aαα = 0 here. The (2) (2) unknown quantities λα and aiα can be determined similarly using (8.10). The expansions (8.6)-(8.7) from the perturbation analysis thus take the ﬁnal form 2 (1) (1) (2) lα = λα + vαα + 2 vαα + λαβ vαβ + O(3 ), (8.11) β =α

ciα

(1) (1) (1) (1) (1) = −λiα viα + 2 λiα viα vαα + λαβ viβ vβα β =α

3

cαα

+O( ), i = α, 2 (1) = 1 + 2 − 12 λ2αβ vαβ + O(3 ). β =α

Returning to our one-sample problem, assuming λα is distinct, the eigenvalue lα /m of S can be expanded by setting V(2) = 0 as (1) 2 (1) lα /m = λα + m−1/2 vαα + m−1 λαβ vαβ + Op (m−3/2 ). (8.12) β =α

Step 2: Taylor series of f (l/m) Assuming f (·) is continuously diﬀerentiable in a neighborhood of λ = (λ1 , . . . , λp ) , we can write the Taylor series around λ, f (l/m) = f (λ) + Df (λ)(l/m − λ) + 12 (l/m − λ) D2 f (λ)(l/m − λ) + Op (m−3/2 ). Upon using (8.12), this becomes f (l/m)

−1/2

= f (λ) + m

p

(1) fi vii

+m

i=1

+ 12 m−1

p p i=1 j=1

−1

p i=1

(1) (1)

fi

(1) 2

λiβ viβ

β =i

fij vii vjj + Op (m−3/2 ),

(8.13)

8.8. Asymptotic distributions of eigenvalues

127

where Df (λ) D2 f (λ)

= =

(f1 , . . . , fp ) = (∂f (λ)/∂λi ) ,

(fij ) = ∂ 2 f (λ)/∂λi ∂λj .

Step 3: Expansion of the characteristic function The characteristic function of m1/2 [f (l/m) − f (λ)] thus becomes E exp{itm1/2 [f (l/m) − f (λ)]} p p it (1) (1) 2 fi vii fi λiβ viβ = E exp it exp √ m i=1 i=1 β =i p p (1) (1) +1 fij vii vjj + Op (m−1 ) 2

i=1 j=1

p it (1) (1) 2 = E exp it fi vii fi λiβ viβ 1+ √ m i=1 i=1 β =i p p (1) (1) + 12 fij vii vjj + Op (m−1 ) .

p

(8.14)

i=1 j=1

We need then to evaluate the following expectations in (8.14): p (1) fi vii , E exp it E exp it E exp it

i=1 p i=1 p

(1) fi vii

(1) 2

· viβ , β = i,

(8.15) (8.16)

(1) fi vii

(1) (1)

· vii vjj .

(8.17)

i=1

Step 4: Sugiura’s lemma [Sugiura (1973)] Let V ∼ Wp (m, Σ) and S = V/m. Lemma 8.3 Let g(S) be an analytic function at S = Σ and put T = m1/2 (S − Σ). Deﬁne a matrix of diﬀerential operators by

∂ = 12 (1 + δij )∂/∂ij applied to the function g(Γ) of a symmetric matrix Γ = (γij ). Then, for any symmetric matrix A and suﬃciently large m, 2 d2j−1 (it)2j−1 E g(S)etr(itAT) = etr[−t2 (AΣ)2 ] · 1 + m−1/2 j=1

128

8. Tests on mean and variance

+m−1

3

g2j (it)2j + O(m−3/2 ) g(Γ)|Γ=Σ ,

j=1

where each coeﬃcient is given by d1

=

d3

=

g0

=

g2

=

g4

=

g6

=

2 tr (ΣAΣ∂), 4 tr (ΣA)3 , 3 tr (Σ∂)2 , 1 4 tr (ΣA)2 Σ∂ + d21 , 2 2 tr (ΣA)4 + d1 d3 , 1 2 d . 2 3

Before presenting the proof, we comment on Taylor series and diﬀerential operators. An analytic function g(x), at x0 , of a real variable x can be written as a Taylor series g(x)

= g(x0 ) +

∞ g (j) (x0 ) j=1

(x−x0 )∂

= e =

j!

(x − x0 )j

· g(x)|x=x0

[1 + (x − x0 )∂ + 12 (x − x0 )2 ∂ 2 + · · ·]g(x)|x=x0 ,

where ∂ j g(x)|x=x0 = ∂ j g(x0 )/∂xj is the jth derivative of g evaluated at x0 . In the same way, for a function g(S) analytic at Σ, of a symmetric matrix S, we have g(S) = {etr(S − Σ)∂} g(Γ)|Γ=Σ . p

Proof. Note that S → Σ. Taylor series expansion of g(S) at Σ gives g(S) = {etr(S − Σ)∂} g(Γ)|Γ=Σ . After multiplying by etr(itAT) and taking expectations with respect to V ∼ Wp (m, Σ), we get E etr(itAT)g(S)

= |I − 2m−1/2 itAΣ − 2m−1 Σ∂|−m/2 ( ) · etr(−m1/2 itAΣ − Σ∂) g(Γ)|Γ=Σ .

The above determinant can be arranged according to powers of m as in Sugiura and Nagao (1971). 2 Evaluation of (8.15)

8.8. Asymptotic distributions of eigenvalues

129

Let A = diag(f1 , . . . , fp ), g(Γ) ≡ 1, and Σ = Λ, and note that tr AV(1)

=

p

(1)

fi vii ,

i=1

tr(AΛ)2

=

p

fi2 λ2i ≡ τ 2 /2 (say).

i=1

Since d1 g(Γ) = 0, the lemma yields p (1) E exp it fi vii = exp(− 12 t2 τ 2 )[1 + m−1/2 d3 (it)3 + O(m−1 )], i=1

p where d3 = (4/3) i=1 fi3 λ3i . Evaluation of (8.16) 2 , i = β. Note that g(Λ) = 0 and Let A = diag(f1 , . . . , fp ) and g(Γ) = mγiβ the diﬀerential operator d1 , d1 = 2

p

fk λ2k ∂kk ,

k=1

is a linear combination of ∂kk ≡ ∂/∂kk and, thus, d1 g(Γ) = 0. For similar reasons, we also have g2 g(Γ) = 0 and d3 g(Γ)|Γ=Λ = g4 g(Γ)|Γ=Λ = g6 g(Γ)|Γ=Λ = 0, which implies p (1) (1) 2 fi vii · viβ = exp(− 12 t2 τ 2 )[m−1 g0 + O(m−3/2 )]g(Γ)|Γ=Λ . E exp it i=1

However, g0 is the diﬀerential operator g0 = tr(Λ∂)2 =

p p

2 λk λl ∂kl

k=1 l=1

and, thus, g0 g(Γ)|Γ=Λ =

p p

2 2 λk λl ∂kl (mγiβ )|Γ=Λ = mλi λβ .

k=1 l=1

Hence, we get E exp it

p

(1) fi vii

(1) 2

· viβ

= exp(− 12 t2 τ 2 )[λi λβ + O(m−1/2 )].

i=1

Evaluation of (8.17) Similarly, letting A = diag(f1 , . . . , fp ) and g(Γ) = m(γii − λi )(γjj − λj ), we ﬁnd p (1) (1) (1) E exp it · vii vjj = exp(− 12 t2 τ 2 )[2λ2i δij + O(m−1/2 )]. fi vii i=1

130

8. Tests on mean and variance

We now return to the expansion of the characteristic function in (8.14). Hence, altogether, the expansion of the characteristic function becomes E exp{itm1/2 [f (l/m) − f (λ)]/τ } = exp(− 12 t2 τ 2 )[1 + m−1/2

2 a2j−1 j=1

τ 2j−1

(it)2j−1 + O(m−1 )],

where a1 a3

=

p

=

fi λiβ λi λβ +

i=1 β =i p

4 3

p

fii λ2i ,

(8.18)

i=1

fi3 λ3i .

(8.19)

i=1

Step 5: Inversion of the characteristic function Using the inversion formula (2.2), an expansion for the density function of s = m1/2 [f (l/m) − f (λ)]/τ is f (s)

=

1 2π

+m

∞

−∞ −1/2

1

2

e−its e− 2 t dt ∞ 2 1 2 a2j−1 1 e−its e− 2 t (it)2j−1 dt + O(m−1 ) 2j−1 2π τ −∞ j=1

= φ(s) − m−1/2

2 a2j−1 j=1

τ 2j−1

φ(2j−1) (s) + O(m−1 ),

and similarly for the distribution function of s, F (s) = Φ(s) − m−1/2

2 a2j−1 j=1

τ 2j−1

Φ(2j−1) (s) + O(m−1 ),

where φ and Φ are respectively the density function and distribution function of the standard normal distribution. We have proved: Proposition 8.18 Let f (·) be continuously diﬀerentiable in a neighbor2 hood p of 2λ. 2 If the population eigenvalues λα are all distinct and τ = 2 i=1 fi λi = 0, then the distribution function of s = m1/2 [f (l/m) − f (λ)]/τ can be expanded for large m as Φ(s) − m−1/2

2 a2j−1 j=1

τ 2j−1

Φ(2j−1) (s) + O(m−1 ).

8.8. Asymptotic distributions of eigenvalues

131

Corollary 8.3 Let f (·) be continuously diﬀerentiable in a neighborhood of p λ. If the population eigenvalues λα are all distinct and τ 2 = 2 i=1 fi2 λ2i = 0, then the limiting distribution is given by s = m1/2 [f (l/m) − f (λ)]/τ → N (0, 1). d

For an individual eigenvalue the expansion follows immediately. Corollary 8.4 Let lα be the αth largest eigenvalue of V ∼ Wp (m, Λ). If λα is distinct from all other p − 1 eigenvalues, the distribution function of √ s = m1/2 (lα /m − λα )/( 2λα ) can be expanded for large m as Φ(s) − m−1/2

2 a2j−1 j=1

where a1

=

τ 2j−1

Φ(2j−1) (s) + O(m−1 ),

λα λβ /(λα − λβ ),

β =α

4 3 λ . 3 α The sample eigenvalues are asymptotically independent, as the following corollary shows. a3

=

Corollary 8.5 Let l = (l1 , . . . , lp ) be the eigenvalues of V ∼ Wp (m, Λ). If the population eigenvalues λα are all distinct, then the joint limiting distribution is given by (m/2)1/2 Λ−1 (l/m − λ) → Np (0, I). d

Proof. From (8.12), we can write (1)

(1) m1/2 (l/m − λ) = (v11 , . . . , vpp ) + Op (m−1/2 ) ≡ v(1) + Op (m−1/2 ).

The asymptotic distribution of V(1) was derived in Section 6.3, and for the d marginal v(1) , we ﬁnd, using (6.1), v(1) → Np (0, 2Λ2 ). 2 The asymptotic expansion in Corollary 8.4 gives the ﬁrst two terms of a more accurate approximation, with remainder O(m−3/2 ), Φ(s) − m−1/2

2 a2j−1 j=1

τ

Φ(2j−1) (s) + m−1 2j−1

3 b2j (2j) Φ (s) + O(m−3/2 ), 2j τ j=1

where a1 and a3 are given in Corollary 8.4 and b2 = 2λ2α λβ /(λα − λβ ) − 2λ3α λβ /(λα − λβ )2 β =α

β =α

132

8. Tests on mean and variance

3 + λ2α λ2β /(λα − λβ )2 + λ2α λγ λβ /(λα − λγ )(λα − λβ ), 2 γ δ1 , then 0 ≤ αj − βj ≤ ρ2 (C)/(βj − δ1 ), j = 1, . . . , q, 0 ≤ δr−i − αp−i ≤ ρ2 (C)/(βq − δr−i ), i = 0, . . . , r − 1. A proof of Wielandt’s inequality can be found in Eaton and Tyler (1991), who used it to ﬁnd the asymptotic distribution of the eigenvalues of symmetric random matrices in the case of multiple eigenvalues. The matrix A can be viewed as a perturbation of a block-diagonal matrix, namely A = A0 + E, where B 0 0 C A0 = and E = . 0 D C 0 By Wielandt’s inequality, the eigenvalues of A0 are perturbed quadratically in E when A0 is perturbed linearly in E. Generally, eigenvalues are only perturbed linearly when the matrix is perturbed linearly. The quadratic perturbation of eigenvalues in Wielandt’s inequality is due to the special structure of E relative to A0 . Let Sp be the set of p × p real symmetric matrices. Consider a sequence of random matrices Sn ∈ Sp , and assume that Wn = n1/2 (Sn − Σ) → W, d

for some Σ ∈ Sp , and hence W ∈ Sp . Given Σ ∈ Sp , let φ(Σ) = (φ1 (Σ), . . . , φp (Σ)) be the vector of ordered eigenvalues φ1 (Σ) ≥ φ2 (Σ) ≥ · · · ≥ φp (Σ). The asymptotic distribution of n1/2 (φ(Sn ) − φ(Σ)) is studied.

(8.21)

8.8. Asymptotic distributions of eigenvalues

135

We consider in the ﬁrst place the case Σ = diag(d1 Ip1 , . . . , dk Ipk ), p = p1 + · · · + pk , where d1 > d2 > · · · > dk represent the distinct eigenvalues of Σ with the multiplicity of di being pi , i = 1, . . . , k. To reﬂect the block structure of Σ, consider the partitioned matrix S S ··· S n,11

n,12

Sn,21 Sn = .. . Sn,k1

Sn,22 .. .

n,1k

· · · Sn,2k .. .. , . . · · · Sn,kk

Sn,k2

where Sn,ij ∈ Rppij . Lemma 8.4 For k = 2,

φ(Sn ) −

φ(Sn,11 ) φ(Sn,22 )

is Op (n−1 ).

Proof. Let An = {Sn | φp1 (Sn,11 ) > φ1 (Sn,22 )}. Since φ is continuous and, p p p from (8.21), Sn,11 → d1 Ip1 and Sn,22 → d2 Ip2 , it follows that φp1 (Sn,11 ) → p d1 and φ1 (Sn,22 ) → d2 . Thus, P (An ) → 1, so attention can be restricted to An , n = 1, 2, . . .. For Sn ∈ An , Wielandt’s inequality implies for 1 ≤ i ≤ p1 , 0 ≤ φi (Sn ) − φi (Sn,11 ) ≤ ρ2 (Sn,12 )/ (φi (Sn,11 ) − φ1 (Sn,22 )) . By (8.21), Sn,12 is Op (n−1/2 ), and since ρ is continuous, it follows that ρ2 (Sn,12 ) is Op (n−1 ). Since p

φi (Sn,11 ) − φ1 (Sn,11 ) → d1 − d2 > 0, then φi (Sn ) − φi (Sn,11 ) is Op (n−1 ), i = 1, . . . , p1 . The proof of φp−j (Sn ) − 2 φp2 −j (Sn,22 ) is Op (n−1 ), j = 0, . . . , p2 − 1, is analogous. By applying Lemma 8.4, k − 1 times, the following asymptotic equivalence result is obtained. The vector 1pi ∈ Rpi is the vector of ones. Proposition 8.21 n1/2 (φ(Sn ) − φ(Σ)) = Zn + Rn , where

φ(Sn,11 ) − d1 1p1 .. Zn = n1/2 . φ(Sn,kk ) − dk 1pk

and the remainder term Rn is Op (n−1/2 ). p

Since Rn → 0, using Slutsky’s theorem the asymptotic distribution of n1/2 (φ(Sn ) − φ(Σ))

136

8. Tests on mean and variance

is that of the leading term Zn . Considering the partitioned W = (Wij ), Wij ∈ Rppij , i, j = 1, . . . , k, we have immediately from (8.21) that Sn,11 − d1 Ip1 W11 d .. .. . n1/2 → . . Sn,kk − dk Ipk Now, because the function

Wkk

φ(W11 ) .. G(W) = .

φ(Wkk ) is continuous and since φ n1/2 (Sn,11 − d1 Ip1 ) = n1/2 (φ(Sn,11 ) − d1 1p1 ) , it follows that d

Zn = G(Wn ) → G(W). We have proved: d

Proposition 8.22 If n1/2 (Sn − Σ) → W and Σ is diagonal, then n1/2 (φ(Sn ) − φ(Σ)) → G(W). d

In the general case where Σ is not diagonal, there exists H ∈ Op such that Σ = H diag (d1 Ip1 , . . . , dk Ipk ) H ≡ HDH . d

From (8.21), n1/2 (HSn H − D) → HWH . Since φ(HSn H ) = φ(Sn ) and φ(Σ) = φ(D), we obtain the general result n1/2 (φ(Sn ) − φ(Σ)) → G(HWH ). d

This general result is summarized. d

Proposition 8.23 If n1/2 (Sn − Σ) → W, then n1/2 (φ(Sn ) − φ(Σ)) → G(HWH ), d

where H diagonalizes Σ, Σ = H diag (d1 Ip1 , . . . , dk Ipk ) H . An important special case is when W in (8.21) is a multivariate normal matrix and all eigenvalues of Σ are distinct. In that case, HWH is also a multivariate normal matrix. Also, since all eigenvalues of Σ have multiplicity 1, then G(HWH ) is just a p-dimensional marginal of HWH and, hence, has a p-dimensional normal distribution.

8.9. Problems

137

Example 8.3 It was seen in Chapter 6 that when sampling from a Np (µ, Σ) distribution, the asymptotic distribution of the sample variance d

S is n1/2 (S − Σ) → W, where W ∼ Npp (0, (I + K)(Σ ⊗ Σ)). We derive the asymptotic distribution of the eigenvalues of S when all eigenvalues of Σ are distinct. Using Proposition 8.23, we have d n1/2 (φ(S) − φ(Σ)) → G(HWH ), where H diagonalizes Σ, HΣH ≡ D (say). But from Proposition 6.1, HWH ∼ Npp (0, (I + K)(D ⊗ D)). From (6.1), (HWH )ii ∼ N (0, 2d2i ), i = 1, . . . , p, and are independently distributed. Since all eigenvalues of Σ are distinct, then n1/2 (φ(S) − φ(Σ)) → Np (0, 2D2 ), d

which is the result proven previously in Corollary 8.5. Example 8.4 We now derive the asymptotic distribution of the r smallest eigenvalues of S when the smallest eigenvalue of Σ has multiplicity r, φ(Σ) = (φ1 (Σ), . . . , φp−r (Σ), λ, . . . , λ) . As in Example 8.3, HWH ∼ Npp (0, (I + Kp )(D ⊗ D)). Hence, the lower right r × r block of HWH is distributed as Nrr (0, λ2 (I + Kr )). Using Proposition 8.23, we have ﬁnally n1/2 λ−1 (φp−r+1 (S) − λ, . . . , φp (S) − λ) → w, d

where w is distributed as the eigenvalues of a Nrr (0, (I + Kr )) distribution. An application of Wielandt’s inequality to bootstrapping eigenvalues can be found in Eaton and Tyler (1991), who extend the work of Beran and Srivastava (1985, 1987). Earlier papers on the case of multiple eigenvalues include those of James (1969), Chattopadhyay and Pillai (1973), Chikuse (1976), Khatri and Srivastava (1978), and Srivastava and Carter (1980).

8.9 Problems 1. Two-sample T 2 . Let x1 , . . . , xn i.i.d. Np (µ, Σ) and y1 , . . . , ym i.i.d. Np (τ , Σ), Σ > 0, be two independent samples. Deﬁne the sample variances Sx

=

1 ¯ )(xi − x ¯ ) , (xi − x (n − 1) i=1

Sy

=

1 ¯ )(yi − y ¯ ) , (yi − y (m − 1) i=1

n

m

138

8. Tests on mean and variance

and Spool =

1 [(n − 1)Sx + (m − 1)Sy ]. (n + m − 2)

Determine (i) the distribution of Spool , (ii) the distribution of −1 1 1 ¯ ) S−1 ¯) (¯ x−y x−y T2 = + pool (¯ n m used for testing H0 : µ = τ against H1 : µ = τ . 2. Invariance of two-sample T 2 . This is a continuation of Problem 8.9.1. ¯ , Spool ) is minimal suﬃcient for (µ, τ , Σ). (i) Prove (¯ x, y (ii) Consider (A, b) in the group of transformations Gp × Rp acting ¯ → A¯ ¯ → A¯ as x x + b, y y + b, and Spool → ASpool A . Prove that this group of transformations leaves the testing problem ¯ , Spool ) invariant and that any invariant test depends on (¯ x, y only through T 2 . (iii) Prove that any invariant test has a power function depending on (µ, τ , Σ) only through (µ − τ ) Σ−1 (µ − τ ). 3. Common mean vector. For independent samples xi1 , . . . , xini , i.i.d. Np (µ, Σi ), i = 1, . . . , a, ¯ i ) be the from distributions with a common mean vector µ, let (Si , x MVUE from each sample. Consider estimating the common mean µ with respect to the weighted least-squares criterion min µ

a

ci ni (¯ xi − µ) S−1 xi − µ) i (¯

i=1

a for some constants ci > 0, i=1 ci = 1. Establish the estimate of µ is given by −1 a a −1 −1 ˜= ¯i . ci ni Si ci ni Si x µ i=1

i=1

Remark: Jordan and Krishnamoorthy (1995) built an exact β×100% ˜ conﬁdence region centered at µ. 4. Test of symmetry. Assume x1 , . . . , xn i.i.d. Np (µ, Σ), Σ > 0. Choose C ∈ Rp−1 of p rank C = p − 1 such that C1 = 0. Prove the following:

8.9. Problems

139

(i) ker C = span{1} and, therefore, H0 : Cµ = 0 ⇐⇒ H0 : µ1 = · · · = µp . (ii) Any p − 1 columns of C are linearly independent, which implies that C = A(Ip−1 , −1), for some nonsingular A ∈ Rp−1 p−1 . Conclude that, thereafter, the value of x) (CSC )−1 (C¯ x), T 2 = n(C¯ n n ¯ )(xi − x ¯ ) , ¯ = i=1 xi /n and (n − 1)S = i=1 (xi − x where x does not depend on the choice of C. (iii) The null (under H0 ) distribution of T 2 is T 2 /(n − 1) ∼ Fc (p − 1, n − p + 1). 5. Assume x1 , . . . , xn i.i.d. x ∈ Rp (possibly non-normal) with E x = µ and var x = Σ. Establish that, asymptotically, we are at least (1 − α) × 100% conﬁdent in simultaneously presenting all of the observed “Scheﬀ´e” intervals: 1/2 1/2 χ2α,p χ2α,p 1/2 ¯− ¯+ (a Sa) ≤aµ≤ax (a Sa)1/2 , ax n n ∀a ∈ Rp . 6. Assume x1 , . . . , xn i.i.d. x ∈ Rp (possibly non-normal) with E x = µ and var x = Σ. Let a1 , . . . , ar be linearly independent in Rp . Establish that, asymptotically, we are at least (1 − α) × 100% conﬁdent in simultaneously presenting the observed “Roy-Bose” intervals: 1/2 1/2 χ2α,r χ2α,r 1/2 ¯− ¯+ ai x (ai Sai ) ≤ a i µ ≤ ai x (ai Sai )1/2 , n n i = 1, . . . , r. 7. Test of proportionality. Given x1 , . . . , xn a random sample from Np (µ, Σ), Σ > 0, obtain the likelihood ratio test Λ for H0 : Σ = γΣ0 , γ > 0 where Σ0 > 0 is a known matrix: Λ with V =

n

i=1 (xi

2/n

=

|Σ−1 0 V|/

versus

H1 : Σ > 0,

1 tr Σ−1 0 V p

¯ )(xi − x ¯ ) as usual. −x

8. Test for a given variance. Let x1 , . . . , xn i.i.d. Np (µ, Σ), Σ > 0.

p ,

140

8. Tests on mean and variance

(i) Prove that the LRT for H0 : Σ = I versus H1 : Σ = I is given by Λ = (e/n)pn/2 |V|n/2 etr(− 12 V),

n ¯ )(xi − x ¯ ) . where V = i=1 (xi − x Remark: This test is biased, but Sugiura and Nagao (1968) have shown the slight modiﬁcation Λ∗ = (e/m)pm/2 |V|m/2 etr(− 12 V), where m = n − 1, gives an unbiased test. (ii) Using the Wishart density in Problem 7.5.7, prove that mph/2 Γp [ 12 m(1 + h)] 2e |Σ|mh/2 E Λ∗ h = . m |I + hΣ|m(1+h)/2 Γp ( 12 m) Hint: The integrand has the form of a Wishart density. Simply ﬁnd the normalizing constant. (iii) Under H0 , mph/2 Γp [ 12 m(1 + h)] 2e ∗h EΛ = (1 + h)−mp(1+h)/2 . m Γp ( 12 m) Remark: An accurate approximation to the null distribution of Λ∗ using those moments is given in Example 12.5. ˆ2: 9. Use Proposition 8.6 to obtain the moments of R

∞ Γ 12 (p − 1 + 2k) + h Γ 12 (n − 1 + 2k) 2h ˆ

, pk ER = 1 1 Γ (p − 1 + 2k) Γ (n − 1 + 2k) + h 2 2 k=0 where pk are the negative binomial probabilities given in Proposition 8.6. 10. Demonstrate that if K given δ is Poisson(δ) and δ ∼ G(p, θ), then the marginal of K is the negative binomial K ∼ nb(p, (1 + θ)−1 ). 11. Write Nagao’s test for the equality of two variances, H0 : Σ1 = Σ2 , as a function of the eigenvalues l1 , . . . , lp of V1−1 V2 , where as usual ni ¯ i )(xij − x ¯ i ) , i = 1, 2. Vi = j=1 (xij − x 12. Write the LBI test for sphericity as a function of l1 /lp , . . . , lp−1 /lp , n where l1 ≥ · · · ≥ lp are the ordered eigenvalues of V = i=1 (xi − ¯ ) . ¯ )(xi − x x 13. Invariance of multiple correlation. Let xi = (xi1 , xi2 ) , i = 1, . . . , n, be i.i.d. Np (µ, Σ), Σ > 0. Consider the group of transformations axi1 + b xi → , Axi2 + b for any a = 0, b ∈ R, A ∈ Gp−1 , and b ∈ Rp−1 .

8.9. Problems

141

(i) Show that this transformation induces the following transformations on the suﬃcient statistics and parameters: a¯ x1 + b ¯ → x , A¯ x2 + b 2 a v11 av21 A , V → aAv21 AV22 A aµ1 + b µ → , Aµ2 + b 2 aσ 21 A a σ11 . Σ → aAσ 21 AΣ22 A −1/2

−1/2

x1 , A = V22 , and b = −A¯ x2 to (ii) Choose a = v11 , b = −a¯ prove that any invariant test f (¯ x, V) depends on the data only −1/2 1/2 through u = V22 v21 /v11 . (iii) Prove that there exists an orthogonal transformation H ∈ Op−1 ˆ 1 (v. Problem 1.8.14). such that Hu = Re (iv) Choosing further a = 1, b = 0, A = H, and b = 0, prove that ˆ any invariant test is necessarily a function of R. (v) Prove that the non-null distribution of any invariant test depends on the parameters only through R. 14. Test of equality of means and variances. The data consists of a independent samplesxi1 , . . . , xini i.i.d. a Np (µi , Σi ), Σi > 0, i = 1, . . . , a. Let n = i=1 ni be the total number of observations. The hypothesis is the equality of the distributions H0 : µ1 = · · · = µa ; Σ1 = · · · = Σa which is being tested against all alternatives. Let ¯i x

=

ni 1 xij , ni j=1

1 ¯i, ni x n i=1 a

¯ = x

be the ith sample mean and overall mean, respectively, and Vi

=

ni

¯ i )(xij − x ¯ i ) , i = 1, . . . , a, (xij − x

j=1

B

=

a

¯ )(¯ ¯ ) ni (¯ xi − x xi − x

i=1

be the usual “within” and “between” sums of squares, respectively.

142

8. Tests on mean and variance

(i) Prove that the LRT is a |Vi |ni /2 nnp/2 Λ = ai=1 . | i=1 Vi + B|n/2 a nni p/2 i=1

i

Remark: Perlman (1980) proved this LRT yields an unbiased test. (ii) Use the group of transformations xij → Axij + a, for any A ∈ Gp , a ∈ Rp , to argue that the null distribution of Λ (or any other invariant test) can be obtained by setting µi = 0 and Σi = I without loss of generality. (iii) Establish that V1 , . . . , Va , B are mutually independent whenever Σ1 = · · · = Σa holds. Moreover, verify that, under H0 , Vi ∼ Wp (ni − 1) and B ∼ Wp (a − 1). Hint: Let xi1 X1 . Xi = ... and X = .. Xa xini be the sample matrices. For appropriately chosen orthogonal projections Qi , i = 1, . . . , a, and Q, so that Vi = Xi Qi Xi and B = X QX, write Q1 0 · · · 0 Q1 X1 0 Q2 · · · 0 .. .. .. .. .. . = X . . . . Qa Xa 0 0 · · · Qa QX Q ≡ CX. Then use Proposition 6.1 and verify that CC is block-diagonal. (iv) Obtain the null moments of Λ: Γp [ 1 ni (1 + h) − 1 ] Γp [ 12 (n − 1)] 2 2 . 1 ni ph/2 Γ [ 1 n(1 + h) − 1 ] Γ [ (n − 1)] p p i n 2 2 i=1 2 i=1 i

E Λh = a

a

nnph/2

Hint: Recall the normalizing constant cp,m of a Wp (m) p.d.f. and use a similar argument as in the proof of Proposition 8.17 to establish E Λh = a

nnph/2

a

ni ph/2 i=1 i=1 ni

cp,ni −1 cp,ni (1+h)−1

E |W|−nh/2 ,

where W ∼ Wp (n(1 + h) − 1). Remark: The null moments are invoked in Problem 12.4.1 to develop an accurate approximation to the null distribution of Λ.

8.9. Problems

143

15. Assume the population eigenvalue λα is distinct from all other p − 1 eigenvalues. Establish that the logarithmic transformation is a variance stabilizing transformation for the sample eigenvalue lα of V ∼ Wp (m, Λ).

9 Multivariate regression

9.1 Introduction Multivariate regression with p responses as opposed to p multiple regressions is getting increasingly more attention, especially in the context of prediction. In this chapter, we generalize the multiple regression model of Section 5.6.2 to the multivariate case. The estimation method of Section 9.2 relies also on orthogonal projections. The model considered is Y = XB + E, where Y ∈ Rnp , B ∈ Rkp , and X ∈ Rnk of rank X = k is ﬁxed. The error term E is such that E E = 0 and var E = In ⊗ Σ with Σ > 0 in Rpp . The observation vectors consisting of the rows of Y are thus uncorrelated. The Gauss-Markov estimate is derived ﬁrst. Then, assuming normality, the maximum likelihood estimates of B and Σ are obtained together with the fundamental result about their joint distribution. Section 9.3 derives the likelihood ratio test for the general linear hypothesis H0 : CB = 0 against all alternatives where C ∈ Rrk of rank C = r in the above model. In the last sections, we discuss the practical and more commonly encountered situation of k random (observed) predictors and the problem of prediction of p responses from the same set of k predictors. Finally, an application to the MANOVA one-way classiﬁcation model is treated as a special case.

9.2. Estimation

145

The multivariate regression model can be seen as a set of p correlated multiple regression models of Section 5.6.2. With the partition Y

=

(y1 , . . . , yp ),

B

=

(β 1 , . . . , β p ),

E =

(9.1)

(e1 , . . . , ep ),

we can rewrite Y = XB + E as yi = Xβ i + ei , i = 1, . . . , p, where ei ∼ Nn (0, σii I). However, the p multiple regression models are correlated since cov(ei , ej ) = σij I. Testing a relationship between the various β i ’s will require one to treat the p models as one multivariate regression model.

9.2 Estimation First, observe that Rnp is a linear space on which we deﬁne the usual inner product Y, Z = tr(Y Z) =

p n

yij zij , for any Y, Z ∈ Rnp .

i=1 j=1

The mean of Y is in a subspace V = {XA : A ∈ Rkp }. To deﬁne the orthogonal projection of Y on V, a basis for V is needed. Partition X into columns X = (x1 , . . . , xk ). An element XA ∈ V, where

a1 A = ...

ak

is partitioned into rows, is of the form XA

=

k

xi ai =

i=1

=

p k i=1 j=1

k

p xi aij ej

i=1

aij (xi ej );

j=1

146

9. Multivariate regression

hence, {xi ej ∈ Rnp : i = 1, . . . , k; j = 1, . . . , p} spans V. Moreover, linear independence holds since p p k k aij (xi ej ) = 0 =⇒ xi aij ej = 0 i=1 j=1

i=1

=⇒

p

j=1

aij ej = 0, ∀i

j=1

=⇒ aij = 0, ∀i, j. ˆ of Y on V, note that Next, to calculate the orthogonal projection, say XB, ⊥ ˆ ˆ since Y − XB ∈ V , then xi ej , Y − XB = 0, ∀i, j. But, this means that ! ˆ = x (Y − XB)e ˆ j = 0, ∀i, j. tr ej xi (Y − XB) i ˆ = 0, which gives Therefore, X (Y − XB) ˆ = (X X)−1 X Y. B ˆ = (βˆ1 , . . . , βˆp ), where With the partition in (9.1) we ﬁnd B βˆi = (X X)−1 X yi . For the purpose of estimation of the regression coeﬃcients, the p multiple regression models can be treated separately. ˆ = X(X X)−1 X Y ≡ The orthogonal projection of Y on V becomes XB −1 PY, where P = X(X X) X . This provides the “orthogonal direct sum” Y = PY + QY with Q = I − P as usual. The Gauss-Markov property is the subject of Proposition 9.1. Consider ˆ = D(PY)F. the parameter C = D(XB)F and its natural estimate C ˆ is the “best” (blue) in the sense Among all linear unbiased estimates, C that it has the minimum variance: ˆ = blue(C). Proposition 9.1 (Gauss-Markov) C ˜ = GYH be any linear unbiased estimate. Then, Proof. Let C ˜ = C, ∀B ∈ Rk EC p

⇐⇒ GXBH = DXBF, ∀B ∈ Rkp ⇐⇒ GPYH = DPYF, ∀Y ∈ Rnp ⇐⇒ H Y PG = F Y PD , ∀Y ∈ Rnp ⇐⇒ [(GP) ⊗ H ]vec(Y ) = [(DP) ⊗ F ]vec(Y ) ⇐⇒ (GP) ⊗ H = (DP) ⊗ F .

Now, we have ˆ = var C

[(DP) ⊗ F ](In ⊗ Σ)[(DP) ⊗ F]

9.2. Estimation

= =

147

[(GP) ⊗ H ](In ⊗ Σ)[(GP) ⊗ H] ˜ (GPG ) ⊗ (H ΣH) ≤ (GG ) ⊗ (H ΣH) = var C,

with equality iﬀ GP = G

⇐⇒ [(GP) ⊗ H ]vec(Y ) = (G ⊗ H )vec(Y ), ∀Y ∈ Rnp ⇐⇒ [(DP) ⊗ F ]vec(Y ) = (G ⊗ H )vec(Y ), ∀Y ∈ Rnp ˆ = C. ˜ ⇐⇒ C 2

ˆ = blue(B); also, In Proposition 9.1, if F = I and D = (X X)−1 X , then B −1 ˆ if F = ej and D = ei (X X) X , then bij = blue(bij ), and so on. ˆ is obviously unbiased for B; furthermore, noting that The estimate B QY = QE and using Problem 6.4.5, E Y QY = E E QE = (tr Q)Σ = (n − k)Σ. Thus, S ≡ Y QY/(n − k) is an unbiased estimate of Σ. Under normality, these estimates are optimal in the sense that they have minimum variance among all unbiased estimates. The likelihood for (B, Σ) from Y is L(B, Σ) ∝ |Σ|−n/2 etr − 12 Σ−1 (Y − XB) (Y − XB) ∝ |Σ|−n/2 etr − 12 Σ−1 B X XB & ' · exp − 12 tr Σ−1 · (Y Y) − 2Σ−1 B · (X Y) . From general properties of exponential families [Fraser (1976), pp. 339, 342, 406 or Casella and Berger (1990), pp. 254-255, 263], the statistic (Y Y, X Y) is minimal suﬃcient and complete for (B, Σ). Of course, any ˆ S) is also minimal suﬃcient and complete. one-to-one function such as (B, Thus, from Rao-Blackwell/Lehmann-Scheﬀ´e theorems, among all unbiased ˆ and S have minimum variance. estimates, B Using the decomposition Y = PY + QY, where PQ = 0, the loglikelihood can be written as ' 1 & n l(B, Σ) = cte− ln |Σ|− tr Σ−1 [Y QY + (PY − XB) (PY − XB)] . 2 2 ˆ and Σ ˆ when Thus, to obtain the maximum likelihood estimates (MLE) B n−k ≥ p (for V ≡ (n−k)S = Y QY to be nonsingular w.p.1), we minimize 1 1 VΣ−1 + tr (PY − XB)Σ−1 (PY − XB) , n n ˆ = (X X)−1 X Y, so we and since the last term is ≥ 0, it is clear that B need only minimize ln |Σ| + tr

1 ln |Σ| + tr VΣ−1 . n

148

9. Multivariate regression

However, we already solved a similar problem in Chapter 7 when we derived the maximum likelihood estimates of the mean and variance of a multivariate normal distribution. Using the same result, we ﬁnd that ˆ = 1V Σ n is the maximum likelihood estimate of Σ. Proposition 9.2 gives the joint ˆ and S. distribution of B

ˆ Moreover, B

|=

ˆ Proposition 9.2 With underlying normality, the joint distribution of B and S is

ˆ ∼ N k B, (X X)−1 ⊗ Σ , B p (n − k)S ∼ Wp (n − k, Σ). S.

ˆ QY, which implies B

|=

and, thus, PY

|=

ˆ = (X X)−1 X Y Proof. Since Y ∼ Npn (XB, In ⊗ Σ), the distribution of B follows from Proposition 6.1. Next, since (n − k)S = Y QY = E QE, the distribution of (n − k)S is a direct consequence of Proposition 7.8. Since PQ = 0, we obtain immediately that P P var Y = ⊗ Ip (In ⊗ Σ) ((P, Q) ⊗ Ip ) Q Q P 0 P⊗Σ 0 = ⊗Σ= 0 Q 0 Q⊗Σ S.

2

9.3 The general linear hypothesis Consider now the problem of testing the general linear hypothesis H0 : CB = 0 against all alternatives where C ∈ Rrk of rank C = r in the multivariate regression model Y ∼ Npn (XB, In ⊗ Σ) with X ∈ Rnk of rank X = k. The likelihood ratio test will be more easily expressed using a “canonical” form for this problem.

9.3.1

Canonical form

The canonical form is obtained by transforming the original response Y, in two steps, so that in the new model X becomes X0 and C reduces to

9.3. The general linear hypothesis

C0 , where

X0 =

Ik 0

149

and

C0 = (Ir , 0).

Step 1 : This step is to reduce X to X0 . The Gram-Schmidt method applied to the columns of X (v. Problem 1.8.7) gives X = H1 U, where U ∈ U+ k and H1 H1 = Ik . There exists Γ1 ∈ Rnn−k such that (H1 , Γ1 ) ∈ On . Let ˜ = H1 Y, Y Γ1 ˜ In ⊗ Σ) with B ˜ the ˜ ∼ N n (X0 B, ˜ = UB. But since CB = CU−1 B, then Y p ˜B ˜ = 0, where C ˜ = CU−1 . hypothesis H0 : CB = 0 becomes H0 : C ˜ to C0 . Once again, the GramStep 2 : The second step is to reduce C ˜ ˜ = LH2 , where H2 H = Schmidt method applied to the rows of C yields C 2 k−r + Ir , L ∈ Lr . There exists Γ2 ∈ Rk such that H2 ∈ Ok . Γ2 Let

˜ ˜ = Y

H2 Γ2 0

0 ˜ Y, In−k

˜ ˜ ˜ ∼ N n (E Y, ˜ In ⊗ Σ), where then Y p

H2 ˜ ˜ ˜ = Γ2 B EY 0

or

˜ ˜ ˜1 ˜1 Y B ˜ ˜ ˜2 E Y ˜2 , ≡ B ˜ ˜ 0 Y3

˜ ˜ ˜˜ = ˜ was partitioned in conformity with B ˜ ∈ Rr and B ˜ 1 = H2 B where Y 2 p ˜ ∈ Rk−r . Now, to transform the hypothesis, note that Γ2 B p ˜ ˜ ˜ = LB ˜B ˜ = LH2 B ˜ 1 = 0. ˜ 1 = 0 ⇐⇒ B C ˜ ˜˜ are all ˜ 1 = 0. Because the rows of Y Hence, the hypothesis becomes H0 : B independent, an equivalent problem in its canonical form, with the obvious change of notation, is to test H0 : M1 = 0 against H1 : M1 = 0

150

9. Multivariate regression

based on Z1 Z2 Z3

∼ Nst (M1 , It ⊗ Σ), ∼ Nsu (M2 , Iu ⊗ Σ), ∼ Nsv (0, Iv ⊗ Σ), v ≥ s,

where Z1 , Z2 , and Z3 are independent.

9.3.2

LRT for the canonical problem

We can now obtain a simple expression for the likelihood ratio test. In the canonical model, the likelihood function for (M1 , M2 , Σ) is L(M1 , M2 , Σ) ∝ |Σ|−n/2 etr − 12 Σ−1 (Z1 − M1 ) (Z1 − M1 ) ·etr − 12 Σ−1 (Z2 − M2 ) (Z2 − M2 ) ·etr − 12 Σ−1 Z3 Z3 , where n = t + u + v. Note that in this form, the minimal and suﬃcient statistic is (Z1 , Z2 , Z3 Z3 ). For maximum likelihood estimates when v ≥ s (for Z3 Z3 to be nonsingular w.p.1), we minimize 2 − l(M1 , M2 , Σ) n

1 tr Σ−1 (Z1 − M1 ) (Z1 − M1 ) n 1 1 + tr Σ−1 (Z2 − M2 ) (Z2 − M2 ) + tr Σ−1 Z3 Z3 . n n

= cte + ln |Σ| +

Since each term is ≥ 0, it follows that the maximum likelihood estimates are ˆ 2 = Z2 , and Σ ˆ = Z Z3 /n. ˆ 1 = Z1 , M M 3 Also, when M1 = 0, the maximum likelihood estimates become ˆ ˆ ˆ ˆ 1 = 0, M ˆ 2 = Z2 , and Σ ˆ = (Z Z1 + Z Z3 )/n. M 1 3 Therefore, the LRT is the test which rejects H0 for small values of Λ=

ˆ ˆ ˆ ˆ −n/2 ˆ 2 , Σ) ˆ ˆ 1, M ˆ L(M |Z3 Z3 |n/2 |Σ| = . = ˆ 1, M ˆ 2 , Σ) ˆ ˆ −n/2 |Z1 Z1 + Z3 Z3 |n/2 L(M |Σ| d

|=

Deﬁnition 9.1 U-distribution: U ∼ U (p; m, n) iﬀ U = |W1 |/|W1 + W2 |, W2 , m + n > p. where W1 ∼ Wp (n), W2 ∼ Wp (m), and W1 Properties of U -distributions are deferred to Section 11.4. Going back to the original model, the likelihood ratio test can be exˆ Σ, ˆ and X. Composing the two transformations pressed in terms of B,

9.3. The general linear hypothesis

151

˜ ˜ we obtain ˜ and Y ˜ → Y, Y → Y ˜ ˜1 Y H2 H1 Y ˜ ˜ 2 = Γ2 H1 Y , Y ˜ Γ1 Y ˜3 Y and after long but straightforward calculations, the LRT is expressed as

Λ2/n

˜ ˜ ˜ ˜ Y |Y 3 3| ˜ ˜ ˜ ˜ ˜ ˜ Y ˜ ˜ |Y 1 1 + Y3 Y3 |

=

ˆ |nΣ|

=

ˆ+ |nΣ

ˆ C [C(X X)−1 C ]−1 CB| ˆ B

.

The null distribution of the LRT statistic follows directly from the canonical form of the model and the deﬁnition of U -distributions. Proposition 9.3 The null distribution of the LRT statistic Λ for testing H0 : CB = 0 against H1 : CB = 0, where C ∈ Rrk of rank C = r, in the model Y ∼ Npn (XB, In ⊗ Σ), with X ∈ Rnk of rank X = k, is Λ2/n ∼ U (p; r, n − k). When n is large, a simple approximation can be used for the null distribution of Λ. From the LRT general theory, we can immediately write −2 ln Λ → χ2pr , n → ∞. d

9.3.3

Invariant tests

The problem in its canonical form is to test H0 : M1 = 0 against H1 : M1 = 0

(9.2)

based on Z1 Z2 Z3

∼ Nst (M1 , It ⊗ Σ), ∼ Nsu (M2 , Iu ⊗ Σ), ∼ Nsv (0, Iv ⊗ Σ), v ≥ s,

where Z1 , Z2 , and Z3 are independent. Since v ≥ s, Z3 Z3 is nonsingular w.p.1 and let m ≡ min(s, t) = rank Z1 w.p.1. The group Gs × Rus × Ot × Ou × Ov transforms the variables as Z1 → H1 Z1 A, Z2 → H2 Z2 A + B, and Z3 → H3 Z3 A for any (A, B, H1 , H2 , H3 ) ∈ Gs × Rus × Ot × Ou × Ov . This induces the parameter transformations M1 → H1 M1 A, M2 → H2 M2 A + B, and Σ → A ΣA. Thus, we will say that a test function is invariant iﬀ f (Z1 , Z2 , Z3 ) = f (H1 Z1 A, H2 Z2 A + B, H3 Z3 A),

152

9. Multivariate regression

∀(A, B, H1 , H2 , H3 ) ∈ Gs × Rus × Ot × Ou × Ov , ∀(Z1 , Z2 , Z3 ) ∈ Rts × Rus × Rvs . The choice B = −H2 Z2 A shows that any invariant test does not depend on Z2 , f (Z1 , Z2 , Z3 ) = f (H1 Z1 A, 0, H3 Z3 A). Since rank Z3 = s w.p.1., then using Problem 1.8.7, there exists U ∈ U+ s and H ∈ Rvs satisfying H H = Is such that Z3 = HU. The choice A = U−1 G, where G ∈ Gs is arbitrary for now, yields f (Z1 , Z2 , Z3 ) = f (H1 Z1 U−1 G, 0, H3 HG). From the singular value decomposition (Proposition 1.11) there exists G ∈ Os and H1 ∈ Ot such that D 0 H1 (Z1 U−1 )G = , 0 0 where D is diagonal and contains the square root of the nonzero eigenvalues of (Z1 U−1 )(Z1 U−1 ) = Z1 (Z3 Z3 )−1 Z1 . We thus have D 0 , 0, H3 HG . f (Z1 , Z2 , Z3 ) = f 0 0 Finally, since (HG) (HG) = Is , the s columns of HG are orthonormal in Rv , and by completing to an orthonormal basis of Rv , there exists Γ such that (HG) H3 ≡ ∈ Ov Γ and

H3 HG =

Thus, altogether, we ﬁnd f (Z1 , Z2 , Z3 ) = f

D 0

Is 0

0 0

.

, 0,

Is 0

,

which shows that any invariant test depends on (Z1 , Z2 , Z3 ) only through the nonzero eigenvalues of Z1 (Z3 Z3 )−1 Z1 . Invariance permits a reduction of the parameter space also. Since the transformed parameters are M1 → H1 M1 A, M2 → H2 M2 A+B, and Σ → A ΣA, the choice B = −H2 M2 A shows that the non-null distribution of any invariant test is independent of M2 . Similarly, using the singular value decomposition, there exists G ∈ Os and H1 ∈ Ot such that D 0 H1 (M1 Σ−1/2 )G = , 0 0

9.3. The general linear hypothesis

153

where D is diagonal and contains the square root of the nonzero eigenvalues of (M1 Σ−1/2 )(M1 Σ−1/2 ) = M1 Σ−1 M1 . Thus, the choice A = Σ−1/2 G shows that the non-null distribution of any invariant test depends on (M1 , M2 , Σ) only through the nonzero eigenvalues of M1 Σ−1 M1 . We have proved: Proposition 9.4 For the hypothesis testing situation (9.2) and group of transformations described above, any invariant test depends on (Z1 , Z2 , Z3 ) only through the nonzero eigenvalues of Z1 (Z3 Z3 )−1 Z1 . Moreover, the nonnull distribution of any invariant test depends on (M1 , M2 , Σ) only through the nonzero eigenvalues of M1 Σ−1 M1 . The non-null distribution of those eigenvalues is complicated except in the case m ≡ min(s, t) = 1, where there is only one such eigenvalue. From Proposition 9.4, assume without loss of generality that D 0 t Z1 ∼ Ns , It ⊗ Is , 0 0 ∼ Nsv (0, Iv ⊗ Is ),

Z3

where D contains the square root of the nonzero eigenvalues of M1 Σ−1 M1 . (i) The case t = 1. Here, we have Z1 ∼ Ns ((M1 Σ−1 M1 )1/2 e1 , Is ) and Z3 Z3 ∼ Ws (v). Hence, from Proposition 8.2 on Hotelling’s test, the conclusion is Z1 (Z3 Z3 )−1 Z1 ∼ Fc (s, v − s + 1; M1 Σ−1 M1 /2). (ii) The case s = 1. Here, the distributions are Z1 ∼ Nt ((M1 M1 /σ 2 )1/2 e1 , It ), where Σ = σ 2 was set, and Z3 Z3 ∼ χ2v . Thus, by deﬁnition, Z1 Z1 ∼ Fc (t, v; M1 M1 /2σ 2 ). Z3 Z3 Another equivalent expression for the LRT is Λ2/n

=

|Z3 Z3 | + Z3 Z3 |

|Z1 Z1

= |Is + Z1 Z1 (Z3 Z3 )−1 |−1 m = (1 + li )−1 , i=1

154

9. Multivariate regression

where l1 ≥ · · · ≥ lm are the ordered nonzero eigenvalues of Z1 (Z3 Z3 )−1 Z1 . Thus, Λ takes on small values when those eigenvalues are large. Other possible tests could be used such as the following: m Lawley-Hotelling: T02 = i=1 li = tr Z1 Z1 (Z3 Z3 )−1 Pillai: Roy:

V =

m

li i=1 1+li

= tr Z1 Z1 (Z1 Z1 + Z3 Z3 )−1

l1 = largest eigenvalue of Z1 (Z3 Z3 )−1 Z1 .

None of these tests has a power function which dominates the others over the whole parameter space or even locally [Fujikoshi (1988)]. However, it is easy to see that the asymptotic (as v → ∞, s, t, and u are ﬁxed) null distribution of the three tests −2 ln Λ, vT02 , and vV is χ2st . From the LRT d

general theory, we already know −2 ln Λ → χ2st . Under H0 , we can assume p Σ = Is without loss of generality. From the law of large numbers, Z3 Z3 /v → Is ; hence, we have from Lemma 6.3 v tr Z1 Z1 (Z3 Z3 )−1 → tr Z1 Z1 = χ2st . d

d

The same argument applies to vV . The asymptotic null distribution of Roy’s test is quite diﬀerent and is given as Problem 9.7.5 together with an interpretation as a union-intersection test. In the very special case m ≡ min(s, t) = 1, these three tests are equivalent to the LRT, which is uniformly most powerful invariant (UMPI). The proof of this UMPI property is the same as for the Hotelling-T 2 test (v. Proposition 8.4) since the non-null distribution in both cases (i) and (ii) above is a noncentral canonical Fc distribution. Kariya et al. (1987) considered hypotheses related to selection and independence under multivariate regression models. Breiman and Friedman (1997) presented several methods of predicting responses in a multivariate regression model. The likelihood ratio test for detecting a single outlier (a shift in the mean) in a multivariate regression model was obtained by Srivastava and von Rosen (1998).

9.4 Random design matrix X When the prediction variables X, just as the dependent variables Y, are observed, then it is appropriate to consider X as a random matrix. The model most commonly encountered assumes Y = XB + E, E ∼ Npn (0, In ⊗ Σ),

9.4. Random design matrix X

155

|=

where the errors are independently distributed of the prediction variables, i.e., E X. When X has an absolutely continuous distribution, the argument in the proof of Proposition 7.5 shows that X X is nonsingular w.p.1. The conditional model Y|X ∼ Npn (XB, In ⊗ Σ) is thus identical to the case of a ﬁxed X. Using Proposition 9.2, we ﬁnd the following properties of the same estimates: ˆ is unbiased, (i) B ˆ = E E(B|X) ˆ EB = E B = B. (ii) (n − k)S ∼ Wp (n − k, Σ). Indeed, (n − k)S|X ∼ Wp (n − k, Σ)

ˆ (iii) B

|=

and this conditional distribution does not depend on X. S. With Proposition 2.13, ˆ · h(S) E g(B)

ˆ = E E[g(B)h(S)|X] ˆ = E{E[g(B)|X]E[h(S)|X]} ˆ = E{E[g(B)|X]E[h(S)]} ˆ · E h(S). = E g(B)

Moreover, for testing the general linear hypothesis H0 : CB = 0, the conditional null distribution of Λ2/n does not depend on X and, thus, Λ2/n ∼ U (p; r, n − k), unconditionally. The non-null distribution of Λ2/n , however, will depend on the distribution of X, as is the case for p = 1 as exempliﬁed by Problem 5.7.8, where the noncentrality parameter of the distribution of the F -test depends on X. ˆ may be evaluated as Example 9.1 The variance of B ˆ var B

ˆ ) = var vec(B ˆ )|X] ˆ )|X] + var E[vec(B = E var[vec(B = E[(X X)−1 ⊗ Σ] + var vec(B )

= E (X X)−1 ⊗ Σ.

The last expectation may be evaluated directly in some cases. For example, if X ∼ Nkn (0, In ⊗ Ω), Ω > 0, then with Problem 7.5.5 and since X X ∼ Wk (n, Ω), E (X X)−1 = (n − k − 1)−1 Ω−1 .

156

9. Multivariate regression

9.5 Predictions

Y = XB + E, E ∼ Npn (0, In ⊗ Σ), X

|=

The problem of predicting several, possibly correlated, responses from the same set of predictors is becoming increasingly important. Applications by Breiman and Freidman (1997) include prediction of changes in the valuations of stocks in 60 industry groups by using over 100 econometric variables as predictors. Or, in chemometrics, the prediction of 6 output characteristics of the polymers produced as predicted by 22 predictor variables. Another example by Brown (1980, pp. 247-292) lists electoral results for all 71 Scottish constituencies in the Bristish general elections of February and October 1974. Data consist of total votes for each of the four parties (Conservative, Labour, Liberal, and Nationalist) in each election, together with a categorical variable listing the location of the constituency by six regions, and the size of the electorate in each constituency. The objective is to use the February and October results from part of the constituencies to predict the remaining October results from the corresponding February data. Research papers related to predictions include Stone (1974), van der Merwe and Zidek (1980), Bilodeau and Kariya (1989), and Breiman and Friedman (1997). Assume the “centered” model E.

For the sake of simplicity, we take X centered, X ∼ Nkn (0, In ⊗ Ω). For given values of the prediction variables, x = (x1 , . . . , xk ), it is desired to obtain a prediction of the dependent variables, y = (y1 , . . . , yp ). Using the ˆ an obvious prediction method is Gauss-Markov (GM) estimate B, ˆ = (x βˆ1 , . . . , x βˆp ) ˆ = x B y

|=

|=

so that prediction of the ith variable is done considering only the ith multiple regression model. Assuming the “future” observation follows the same e, model, i.e., y = x B + e , where x ∼ Nk (0, Ω), e ∼ Np (0, Σ), and x (X, E), one can evaluate the risk and is independent of the past, (x, e) of the GM prediction as y − y) E (ˆ y − y) Σ−1 (ˆ

ˆ − B) − e ]Σ−1 [(B ˆ − B) x − e] = E [x (B ! ˆ − B)Σ−1 (B ˆ − B) Ω + p = E tr (B ! ˆ − B)Σ−1 (B ˆ − B) Ω + p. = tr E (B

The SPER (sum of Squares of Prediction Error when the independent variable is Random) risk is obtained on subtracting p from the above: ˆ = E tr (B ˆ − B)Σ−1 (B ˆ − B) Ω. RSPER (B)

9.5. Predictions

157

ˆ − B)Σ−1/2 ∼ N k (0, (X X)−1 ⊗ I), Example 6.3 gives Letting U = (B p E(UU |X) = p(X X)−1 . Finally, with Example 9.1, we get ˆ = pk/(n − k − 1). RSPER (B) A closely related risk function, more tractable mathematically, is SPE deﬁned by ˆ = E tr (B ˆ − B)Σ−1 (B ˆ − B) X X. RSPE (B) ˜ = BA ˆ Smaller risk may be achieved with an estimate B for a certain ˜ = A y ˆ is seen A ∈ Rpp . The corresponding prediction for each variable in y to be a linear combination (multivariate ﬂattening) of the p prediction equations. Example 9.2 Multivariate ﬂattening. Assuming (B, Σ, Ω) is known, then an optimal multivariate ﬂattening [Breiman and Friedman (1997)] would be solution of ˆ − y) Σ−1 (A y ˆ − y). min E (A y A

Now, since ˆ − y) Σ−1 (A y ˆ − y) = (Σ−1/2 A y ˆ − Σ−1/2 y) (Σ−1/2 A y ˆ − Σ−1/2 y), (A y letting C = Σ−1/2 A and z = Σ−1/2 y, the optimization problem becomes equivalent to y |2 . min E |z − Cˆ C

With Problem 5.7.6, the solution is readily obtained: ˆ ) · [var y ˆ ]−1 . C = cov(z, y Each factor is evaluated as ˆ) cov(z, y

ˆ) = Σ−1/2 cov(y, y −1/2 = Σ E yˆ y ˆ = Σ−1/2 E (B x + e)x B −1/2 ˆ E B xx B = Σ = Σ−1/2 B ΩB

and, similarly, ˆ var y

ˆy ˆ = E y = E [B + (X X)−1 X E] xx [B + (X X)−1 X E] = B ΩB + E E X(X X)−1 xx (X X)−1 X E

= B ΩB + E x (X X)−1 x Σ (v. P roblem 6.4.5) = B ΩB + [k/(n − k − 1)]Σ (v. P roblem 7.5.4).

158

9. Multivariate regression

Hence, altogether, the optimal A is given by A = [(B ΩB) + rΣ]−1 (B ΩB), r = k/(n − k − 1). ˆ and modiﬁcations thereof are given in the above papers. Sample-based A In particular, for small r, A ≈ I − r(B ΩB)−1 Σ (v. Problem 1.8.15) and van der Merwe and Zidek (1980) established that the sample-based ˆ −1 S ˆ = I − r(n − k)(B ˆ X XB) A which they called FICYREG (FIltered Canonical Y REGression) leads to smaller SPE risk than GM for r = (k − p − 1)/(n − k + p + 1) provided n > k > p + 1. Bilodeau and Kariya (1989) proposed the modiﬁed Efron-Morris (1976) ˆ −1 S − b(n − k)S/tr(B ˆ X XB) ˆ ˆ = I − r(n − k)(B ˆ X XB) A and showed that it leads still to a smaller SPE risk than FICYREG for r = (k − p − 1)/(n − k + p + 1) and b = (p − 1)/(n − k + p + 1). Note that the choice b = 0 reduces to FICYREG. Breiman and Friedˆ built from cross-validation (CV) and generalized man (1997) considered A cross-validation (GCV). Their large-scale simulations point strongly toward the superiority of CV and GCV over other commonly used prediction techniques. The GCV in particular seems very promising since its evaluation is nearly as simple as GM. The CV, in contrast, is computationally intensive. The unbiased estimate of the SPE risk for the GCV predictions was recently obtained by Bilodeau (1998).

9.6 One-way classiﬁcation In this section, the one-factor univariate analysis of variance is generalized to test the equality of several means of multivariate normal populations. Let yi1 , . . . , yini i.i.d. Np (µi , Σ), i = 1, . . . , a, be a independent samples from multivariate normal distributions with common variance Σ > 0. In matrix notation, let y11 .. . y1n µ1 1 . .. Y = .. , X = diag(1n1 , . . . , 1na ), B = . . y µa a1 . .. yan a

9.7. Problems

159

Then, the a samples can be written as the multivariate regression model

where n =

a i=1

Y = XB + E, E ∼ Npn (0, In ⊗ Σ), ni . The hypothesis of equality of means H0 : µ1 = · · · = µa

can be translated into a general linear hypothesis. Deﬁne C = (Ia−1 , −1a−1 ); then the hypothesis becomes H0 : CB = 0, where C ∈ Raa−1 of rank C = a−1. Using the canonical formulation of this problem, the reader can verify that the LRT is |SSw | , Λ2/n = |SSw + SSb | where SSw SSb

ni a

=

¯ i )(yij − y ¯ i ) , (yij − y

i=1 j=1 a

¯ )(¯ ¯ ) ni (¯ yi − y yi − y

=

i=1

are the usual “within” and “between” sums of squares with ¯i = y

ni

¯= yij /ni and y

j=1

a

¯ i /n. ni y

i=1

The other analysis-of-variance models such as the two-way classiﬁcation model can be generalized similarly to test the eﬀect of each factor or the presence of interactions between factors. We will not pursue this any further here.

9.7 Problems 1. Show that the likelihood ratio test statistic Λ for testing the general linear hypothesis can be written Λ2/n =

ˆ |nΣ| ˆ ˆ +B ˆ C [C(X X)−1 C ]−1 CB| |nΣ

.

ˆ = PY is the orthogonal projection of Y on V = 2. The estimate XB ˆ is also the solution {XA : A ∈ Rkp }. Use this fact to prove that XB of the least-squares problem min tr Ω(Y − V) (Y − V),

V∈V

160

9. Multivariate regression

for any ﬁxed Ω ∈ Rpp , Ω > 0. |=

3. Prove that if Z1 ∼ Nst (M1 , It ⊗Σ), Z3 ∼ Nsv (0, Iv ⊗Σ), and Z1 Z3 , where Σ > 0 and v ≥ s + 2, then t 1 Is + M M1 Σ−1 . E Z1 Z1 (Z3 Z3 )−1 = v−s−1 v−s−1 1 4. Using the canonical model, prove that the LRT Λ for the hypothesis of the equality of several multivariate means is Λ2/n = |SSw |/|SSw + SSb |, as described in Section 9.6. What is the null distribution of this test? 5. The general linear hypothesis in its canonical form is to test H0 : M1 = 0 against H1 : M1 = 0 based on Z1

∼ Nst (M1 , It ⊗ Σ),

Z2

∼ Nsu (M2 , Iu ⊗ Σ),

Z3

∼ Nsv (0, Iv ⊗ Σ), v ≥ s,

where Z1 , Z2 , and Z3 are independent. (i) Prove the asymptotic result as v → ∞ concerning Roy’s test: d

If M1 = 0 then vl1 → α1 , where l1 is the largest eigenvalue of Z1 (Z3 Z3 )−1 Z1 and α1 is the largest eigenvalue of a random matrix W ∼ Ws (t). (ii) Union-intersection test. (a) For a given h ∈ Rt , |h| = 1, deﬁne Hh,0 : M1 h = 0 and Hh,1 : M1 h = 0. Prove H0 H1

= ∩h {Hh,0 : |h| = 1}, = ∪h {Hh,1 : |h| = 1}.

(b) For a given h, |h| = 1, prove that the LRT for Hh,0 against Hh,1 accepts Hh,0 for small values of Rh = h Z1 (Z3 Z3 )−1 Z1 h. Demonstrate the null distribution Rh ∼ Fc (s, v −s+1) does not depend on h. (c) The union-intersection test accepts H0 iﬀ sup|h|=1 Rh ≤ c for some constant c. Demonstrate the union-intersection test statistic sup|h|=1 Rh = l1 is, in fact, Roy’s test. Remark: For a given h, the test based on Rh is UMPI for testing Hh,0 against Hh,1 (the non-null distribution of Rh is a noncentral canonical Fc distribution just like Hotelling’s-T 2 ; v. Proposition 8.4), but Roy’s test is not generally UMPI for testing H0 against H1 .

10 Principal components

10.1 Introduction In this chapter we assume that x ∈ Rp with E x = µ and var x = Σ = (σij ). When the dimension p is large, the principal components method seeks to replace x by y ∈ Rk , where k < p (and hopefully much smaller), without losing too much “information.” This is sometimes particularly useful for a graphical description of the data since it is much easier to view vectors of low dimension. Section 10.2 deﬁnes principal components and gives their interpretation as normalized linear combinations with maximum variance. In Section 10.3, we explain an optimal property of principal components as best approximating subspace of dimension k in terms of squared prediction error. Section 10.4 introduces the sample principal components; they give the coordinates of the projected data which is closest, in terms of euclidian distance, to the original data. Section 10.5 treats the sample principal components calculated from the correlation matrix. Finally, Section 10.6 presents a simple test for multivariate normality which generalizes the univariate Shapiro and Wilk’s statistic. A book entirely devoted to principal component analysis is that of Jolliﬀe (1986).

162

10. Principal components

10.2 Deﬁnition and basic properties The total variance of x is deﬁned as p p E |x − µ|2 = var xi = σii = tr Σ. i=1

i=1

Recall that Σ ≥ 0 can be written as Σ = HDH , where H = D =

(h1 , . . . , hp ) ∈ Op , diag(λ1 , . . . , λp ),

and λ1 ≥ · · · ≥ λp are the ordered eigenvalues of Σ. Since we are only interested in var x, we will assume throughout this chapter that µ = 0. If we let h1 x y = H x = ... , hp x

p p var y = D. Then i=1 var yi = i=1 λi = tr Σ, so x and y have the same “total variance.” Moreover, the variables yi ’s are uncorrelated, cov(hi x, hj x) = hi Σhj = λj hi hj = λj δij . Deﬁnition 10.1 The variables yi = hi x, i = 1, . . . , p, are, by deﬁnition, the principal components of x. p p Since HH = I, then x = ( i=1 hi hi )x = i=1 yi hi and the principal components can be viewed as the coordinates of x withrespect to the k orthonormal basis {h1 , . . . , hp } of Rp . When the ratio i=1 λi /tr Σ is close to 1, then (y1 , . . . , yk ) can eﬀectively replace x without losing much in terms of “total variance.” The principal components can also be got sequentially as follows. First a normalized linear combination t x, |t| = 1, is sought such that var t x = t Σt is maximum. Since for all t, |t| = 1, p p p 2 2 λi (t hi ) ≤ λ1 (t hi ) = λ1 t hi hi t = λ1 |t|2 = λ1 ; t Σt = i=1

i=1

i=1

hence, maxt t=1 t Σt = λ1 , which is attained for t = h1 . So, the ﬁrst principal component y1 = h1 x is the normalized linear combination with maximum variance. Now, given yi = hi x, i = 1, . . . , k, another linear combination s x, |s| = 1, is sought which maximizes the variance s Σs and is uncorrelated with y1 , . . . , yk . Note that cov(s x, yi ) = λi s hi , i = 1, . . . , k. As above, for all s ⊥ h1 , . . . , hk , |s| = 1, we have s Σs =

p i=k+1

λi (s hi )2 ≤ λk+1

p i=k+1

(s hi )2 = λk+1

p i=1

(s hi )2 = λk+1 .

10.3. Best approximating subspace

163

Hence, max

s s=1 s⊥h1 ,...,hk

s Σs = λk+1

is attained for s = hk+1 , which means that yk+1 = hk+1 x is the normalized linear combination with maximum variance among all those uncorrelated with y1 , . . . , yk .

10.3 Best approximating subspace The orthogonal projection of x on the subspace spanned by the ﬁrst k eigenvectors, Pk x, is k k hi hi x = yi hi . Pk x = i=1

i=1

Proposition 10.1 shows that Pk x gives the best approximation to x by a subspace of dimension at most k in terms of squared prediction error. Before stating the result, we present a lemma. Denote by Pk⊥ the set of all orthogonal projections P ∈ Rpp of rank P = k. Lemma 10.1 Let Σ ≥ 0 in Rpp with eigenvalues λ1 ≥ · · · ≥ λp . Then, k

max tr ΣP =

P∈Pk⊥

min tr Σ(I − P)

P∈Pk⊥

are attained at P =

k i=1

λi ,

i=1 p

=

λi

i=k+1

hi hi , where

Σ = HDH , H = (h1 , . . . , hp ) ∈ Op , D = diag(λ1 , . . . , λp ). Proof. Take any P ∈ Pk⊥ . Let A = (a1 , . . . , ak ) whose columns form an orthonormal basis for Im P, then P = AA . Now, tr ΣP = tr HDH AA = tr D(H A)(H A) , and note that G = (g1 , . . . , gk ) ≡ H A has orthonormal columns too, i.e., G G = Ik . Therefore, tr ΣP =

tr D

k

gi gi =

i=1

≤

k i=1

max

g g=1 g⊥g1 ,...,gi−1

k

gi Dgi

i=1

g Dg =

k i=1

λi

164

10. Principal components

(when i = 1 the orthogonality condition is void) with equality if gi = ei , which means A = HG = (h1 , . . . , hk ). This shows the ﬁrst part related to the maximum. The second part is immediate. 2 Proposition 10.1 Assume x ∼ Np (0, Σ), Σ > 0, and let B ∈ Rkp of rank B = k, C ∈ Rpk . Then, min E |x − CBx|2 = B,C

p

λi

i=k+1

is attained when CB = Pk . Proof. Fix B. We have x 0 Σ ∼ Np+k , Bx 0 BΣ

ΣB BΣB

and x | Bx ∼ Np (0, Σ − ΣB (BΣB )−1 BΣ). Using Problem 5.7.6, min E |x − CBx|2 = tr Σ − ΣB (BΣB )−1 BΣ C = tr Σ I − AB (BΣB )−1 BA = tr Σ(I − P), where A = HD1/2 H and P = AB (BΣB )−1 BA, and the extremum is reached at C = (ΣB )(BΣB )−1 . Now, P is an orthogonal projection of rank k. From Lemma 10.1, tr Σ(I − P) is minimized when AB (BΣB )−1 BA =

k

hi hi

i=1

or B (BΣB )−1 B = Finally, CB = Σ[B (BΣB )−1 B] =

k

k

λ−1 i hi hi .

i=1

i=1

hi hi = Pk .

2

Obviously, if µ = 0 in Proposition 10.1, the best approximation of rank k is Pk (x − µ) + µ, which represents the orthogonal projection of x on the aﬃne subspace span{h1 , . . . , hk } + µ.

10.4 Sample principal components from S The variance Σ is usually unknown. Sample principal components can be obtained from the estimate S = V/m, m = n − 1, where, as usual, V=

n i=1

¯ )(xi − x ¯ ) . (xi − x

10.4. Sample principal components from S

165

Since S ≥ 0, write ˆ , ˆ diag(l1 /m, . . . , lp /m) H S=H where ˆ p ) ∈ Op ˆ 1, . . . , h ˆ = (h H and l1 ≥ · · · ≥ lp are the ordered eigenvalues of V. The sample principal ˆ x, i = 1, . . . , p. components of x are deﬁned as h i p Let V ⊂ R be a k-dimensional subspace and denote by V + a = {x + a : x ∈ V} the corresponding aﬃne subspace. What is the aﬃne subspace V +a of dimension k such that the orthogonal projection of the data on V + a is “closest” to the original data? First, we must specify what is meant by “closest.” As a measure of distance, take the usual euclidian distance d(V, a) =

n

ˆ i |2 , |xi − x

i=1

ˆ i − a) + a is the orthogonal projection of xi on V + a. ˆ i = P(x where x Proposition 10.2 Among all k-dimensional subspaces V and vectors a ∈ ˆ 1, . . . , h ˆ k }. ¯ and V = span{h Rp , the distance d(V, a) is minimized for a = x n ˆ the orthogonal projection on ¯ )(xi − x ¯ ) , P Proof. Deﬁne V = i=1 (xi − x ˆ ˆ V, and Q = I − P. Then, d(V, a)

=

n

ˆ i − a) − a|2 |xi − P(x

i=1

=

n

ˆ i − Qa| ˆ 2 |Qx

i=1

=

n

ˆ i − Q¯ ˆ x) + (Q¯ ˆ x − Qa)| ˆ 2 |(Qx

i=1

=

n i=1

=

ˆ i − Q¯ ˆ x|2 + |Qx

n

ˆ x − Qa| ˆ 2 |Q¯

i=1

ˆ + n(¯ ˆ x − a). tr QV x − a) Q(¯

¯ . Also, The two terms in the last expression are non-negative; hence, a = x ˆ ih ˆ . ˆ = k h 2 from Lemma 10.1 and since V ∝ S, P i i=1 p k by the The ratio f (λ) = i=1 λi / i=1 λi of total variance explained p k ﬁrst k principal components is estimated by f (l/m) = i=1 li / i=1 li . A large sample (1 − α) × 100% conﬁdence interval on this ratio f (λ), when all population eigenvalues λα are distinct, can be constructed with Proposition 8.18.

166

10. Principal components

We end this section with a word of caution: Principal components are not invariant with respect to individual rescaling of the p variables in x; that is, if w = Φx, where Φ = diag(φ1 , . . . , φp ), then ΦΣΦ does not have the same eigen-structure as Σ. This means, for example, that the interesting projections of the data found with Proposition 10.2 may look entirely diﬀerent after rescaling. Also, if the ﬁrst variable x1 has a variance much larger than the variances of all other variables, x2 , . . . , xp , then the ﬁrst principal component y1 will be approximately equivalent to x1 . Principal components are thus most meaningful when all variables are measured in the same units and have variances of the same magnitude. For this reason, principal components are often calculated from the sample correlation matrix R rather than the sample variance S.

10.5 Sample principal components from R If we let S0 Σ0

= =

diag(s11 , . . . , spp ), diag(σ11 , . . . , σpp ),

then the population and sample correlation matrices are given by −1/2

R = S0 ρ =

−1/2

SS0

,

−1/2 −1/2 Σ0 ΣΣ0 .

Then, as in the previous section, we can decompose ρ = G diag(γ1 , . . . , γp ) G , ˆ diag(f1 , . . . , fp ) G ˆ , R = G and deﬁne the sample principal components from the standardized vari−1/2 ¯ ), and R as g ˆi z, i = 1, . . . , p, where G = (g1 , . . . , gp ) ables, z = S0 (x − x ˆ and similarly for G. The ratio of total variance (of the standardized vari√ ¯i )/ sii ) explained by the ﬁrst k principal components ables zi = (xi − x becomes f (γ) =

k

γi /p.

i=1

The construction of a conﬁdence interval on this ratio f (γ) thus necessitates the asymptotic distribution of the eigenvalues fi of the sample correlation matrix R. This is now derived using the perturbation method of Section 8.8. Using the Taylor series x−1/2 = a−1/2 − 12 a−3/2 (x − a) + · · · ,

10.5. Sample principal components from R

167

we have directly −1/2

S0

−1/2

= [I − 12 Σ0

−1/2

(S0 − Σ0 )Σ0

−1/2

+ · · ·]Σ0

.

Deﬁne −1/2

V

=

(vij ) = n1/2 (Σ0

V0

=

diag(v11 , . . . , vpp ),

−1/2

SΣ0

− ρ),

and note that V is Op (1). Then, we can write R = =

−1/2

[I − 12 n−1/2 V0 + · · ·]Σ0

−1/2

[Σ + (S − Σ)]Σ0

[I − 12 n−1/2 V0 + · · ·]

[I − 12 n−1/2 V0 + · · ·](ρ + n−1/2 V)[I − 12 n−1/2 V0 + · · ·]

= ρ + n−1/2 (V − 12 ρV0 − 12 V0 ρ) + Op (n−1 ), from which G RG = Γ + n−1/2 V(1) + Op (n−1 ), where Γ = V(1)

diag(γ1 , . . . , γp ), (1)

(vij ) = G (V − 12 ρV0 − 12 V0 ρ)G.

=

(10.1)

Equation (8.11) in the perturbation method then leads, assuming γα to be a distinct eigenvalue, to the expansion (1) fα = γα + n−1/2 vαα + Op (n−1 ),

or, in vector form, assuming all eigenvalues γα to be distinct, to the expansion n1/2 (f − γ)

=

(1)

(1) (v11 , . . . , vpp ) + Op (n−1/2 )

≡ v(1) + Op (n−1/2 ). Now, since V is asymptotically normal with mean 0, so is V(1) and its marginal v(1) . We need only calculate the asymptotic variance of v(1) . From (10.1) and the relation ρG = GΓ, we have (1) vαα

= gα Vgα − γα gα V 0 gα p p p 2 = gjα gkα vjk − γα gjα vjj ; j=1 k=1

j=1

hence, (1) (1) cov(vαα , vββ )

=

p p p p j=1 k=1 i=1 l=1 p p

+γα γβ

j=1 i=1

gjα gkα giβ glβ cov(vjk , vil ) 2 2 gjα giβ cov(vjj , vii )

168

10. Principal components

−γα

p p p

2 gjα giβ glβ cov(vjj , vil )

j=1 i=1 l=1

−γβ

p p p

2 giβ gjα gkα cov(vii , vjk ).

i=1 j=1 k=1 d

Since V → Npp (0, (I + Kp )(ρ ⊗ ρ)), we ﬁnd upon using (6.1) that (1)

(1) lim cov(vαα , vββ )

n→∞

=

p p p p j=1 k=1 i=1 l=1 p p

+γα γβ

gjα gkα giβ glβ (ρkl ρji + ρjl ρki ) 2 2 gjα giβ 2ρ2ji

j=1 i=1 p p p

−γα

2 gjα giβ glβ 2ρjl ρji

j=1 i=1 l=1

−γβ

p p p

2 giβ gjα gkα 2ρik ρij .

i=1 j=1 k=1

Finally, with the simple relations p p

gkα glβ ρkl

= γα δαβ ,

k=1 l=1 p

glβ ρjl

= γβ gjβ ,

l=1

we obtain the simpliﬁcation (1)

(1) , vββ ) = 2γα γβ δαβ − (γα + γβ ) lim cov(vαα

n→∞

p

2 2 gjα gjβ

j=1

+

p p

2 2 gjα giβ ρ

j=1 i=1

We summarize the result. Proposition 10.3 Let f = (f1 , . . . , fp ) be the eigenvalues of the sample correlation matrix R. If the eigenvalues γα of the population correlation matrix ρ are all distinct, then the joint limiting distribution is n1/2 (f − γ) → Np (0, Ω), d

10.6. A test for multivariate normality

where Ω = (ωαβ ) is given by ωαβ = 2γα γβ δαβ − (γα + γβ )

p

2 2 gjα gjβ +

j=1

p p

169

2 2 2 gjα giβ ρji .

j=1 i=1

k The limiting distribution of a function such as f (f ) = i=1 fi /p for the ratio of total variance explained by the ﬁrst k principal components is easily derived by the delta method [v. Problem 10.7.5]. Problem 13.6.19 provides the asymptotic distribution of n1/2 (f − γ) when sampling from an elliptical distribution. Konishi (1979) obtained, with Sugiura’s lemma, a more accurate approximation with remainder O(n−1 ), similar to that of Proposition 8.18, for the distribution function of s = (n − 1)1/2 (f (f ) − f (γ)) , where f (·) is a continuously diﬀerentiable function in a neighborhood of γ.

10.6 A test for multivariate normality Shapiro and Wilk’s (1965) W statistic has been found to be the best omnibus test for detecting departures from univariate normality. Royston (1983) extends the application of W to testing multivariate normality, but the procedure involves a certain approximation which needs to be justiﬁed. The procedure of Srivastava and Hui (1987) does not require such an approximation and has a simple asymptotic null distribution and the calculations are straightforward. Srivastava and Hui (1987) proposed two test statistics for testing multivariate normality. These are based on principal components and may be considered as a generalization of the Shapiro-Wilk statistic. As in Section 10.4, write ˆ , m = n − 1, ˆ diag(l1 /m, . . . , lp /m) H S=H where ˆ p ) ∈ Op . ˆ 1, . . . , h ˆ = (h H The sample principal components of xj , j = 1, . . . , n, are deﬁned as ˆ xj , i = 1, . . . , p, j = 1, . . . , n. Thus, under the null hypothesis of yij = h i multivariate normality, we can treat yi1 , . . . , yin , i = 1, . . . , p, as p approximately independent samples. For sample i, the univariate Shapiro-Wilk statistic is deﬁned as 2 n m aj yi(j) , i = 1, . . . , p, W (i) = nli j=1

170

10. Principal components

where aj ’s are the constants tabulated in Shapiro and Wilk (1965) and yi(1) ≤ yi(2) ≤ · · · ≤ yi(n) are the ordered values of yi1 , . . . , yin . For n > 50, the values of aj are given by Shapiro and Francia (1972) and up to 2000 by Royston (1982). From Shapiro and Wilk (1968), we note that for each i, W (i) can be transformed to an approximate standard normal variable G(W (i)) by using Johnson’s (1949) SB system, W (i) − , G(W (i)) = γ + δ ln 1 − W (i) where γ, δ, and can be found in Table 1 of Shapiro and Wilk (1968) up to n = 50. For n > 50, values of γ, δ, and can be obtained with the help of the results in Shapiro and Francia (1972) and Royston (1982). Let M1 = −2

p

ln [Φ (G(W (i)))] ,

i=1

where Φ(·) is the distribution of a standard normal variable. Note that if U ∼ unif(0, 1), then −2 ln U ∼ χ22 . Srivastava and Hui (1987) proposed M1 as their ﬁrst test statistic for testing multivariate normality, where M1 is approximately distributed as χ22p under the hypothesis of multivariate normality. Large values of M1 will indicate non-normality. Next, they observed that small values of W (i) indicate a departure from normality for variate i. Thus, they considered the minimum of all components and proposed M2 = min W (i) 1≤i≤p

as the second test statistic. The null distribution of M2 is approximately given by p

P (M2 ≤ t) = 1 − [1 − Φ (G(t))] .

(10.2)

For p =2, 4, and 6 and n =10, 25, and 50, a simulation study [Srivastava and Hui (1987)] found that the null distribution of both M1 and M2 are well approximated by χ22p and (10.2), respectively. Examples of the use of M1 and M2 on data sets are provided by Looney (1995) with the necessary SAS procedures or FORTRAN subroutines. Most tests for multivariate normality are functions of the squared radii ¯ ), (or squared Mahalanobis distances of xi to x ¯ ) S−1 (xi − x ¯ ), i = 1, . . . , n. d2i = |zi |2 = (xi − x Some graphical procedures [Andrews et al. (1973), Cox and Small (1978), Gnanadesikan and Kettenring (1972)] are based on d2i . One such Q-Q plot is described in Section 11.4.1. Malkovich and Aﬁﬁ (1973) considered

10.6. A test for multivariate normality

171

the supremum of the standardized skewness and kurtosis over all linear combinations t x, {E (t x − t µ)3 }2 , (t Σt)3 t∈S E (t x − t µ)4 = max − 3 . 2 p−1 (t Σt) t∈S

β1M

=

β2M

max p−1

The tests are based on the sample versions M β1,n

=

M β2,n

=

max b1,n (t),

t∈S p−1

max |b2,n (t) − 3|,

t∈S p−1

respectively, where b1,n (t)

=

b2,n (t)

=

{ n1 1 n

n

n

¯ 3 2 ) } i=1 (t xi − t x , 3 (t St)

¯ 4 ) i=1 (t xi − t x . 2 (t St)

Mardia’s kurtosis test [Mardia (1970)] is a function of d2i and his skewness test is a function of the scaled residuals ¯ ), i = 1, . . . , n. zi = S−1/2 (xi − x Mardia’s measures of multivariate skewness and kurtosis are n 1 ¯ ) S−1 (xj − x ¯ )}3 , B1,n = {(xi − x n2 i,j=1 1 ¯ ) S−1 (xi − x ¯ )}2 , {(xi − x n i=1 n

B2,n

=

respectively. The tests of multivariate normality based on multivariate M and B1,n , are inconsistent against each ﬁxed non-normal skewness, β1,n elliptical distribution [Baringhaus and Henze (1991)]. However, the tests M and B2,n , are consistent. An approxibased on multivariate kurtosis, β2,n M against elliptically symmetric mation formula of the power of the test β2,n distributions was derived by Naito (1998). Cox and Small (1978) proposed tests based on linearity of regression rather than directly on normality. An omnibus test based on empirical characteristic function of the scaled residuals was also proposed [Henze and Zirkler (1990), v. also Henze and Wagner (1997)]. Goodness-of-ﬁt tests for a general multivariate distribution by the empirical characteristic function was treated by Fan (1997). A characterization of multivariate normality by hermitian polynomials was recently proposed by Kariya et al. (1997) to build an omnibus test. A comparative study of goodness-of-ﬁt tests for multivariate normality was carried out by Romeu and Ozturk (1993).

172

10. Principal components

10.7 Problems 1. In morphometric studies, it is often the case that all variables are positively correlated. Prove that if Σ has all positive covariances, σij > 0 for i = j, then all the coeﬃcients in h1 of the ﬁrst principal component may be taken non-negative. 2. For Σ ≥ 0 in Rpp with spectral decomposition Σ = HDH as in k Section 10.2, prove that Θ = i=1 λi hi hi is the matrix of rank k such that p p (σij − θij )2 ||Σ − Θ||2 = i=1 j=1

is minimum. Hint: ||Σ − Θ||2 = tr(D − E)(D − E) , where E = H ΘH. 3. Assume x ∈ Rp has density fx (x) = |Λ|−1/2 g[(x − µ1) Λ−1 (x − µ1)], where Λ = σ 2 [(1 − ρ)I + ρ11 ]. Prove there exists H ∈ Op such that y = Hx has density p y2 (y1 − p1/2 µ)2 −1/2 −(p−1)/2 fy (y) = λ1 λ2 g + i=2 i . λ1 λ2 4. Parent-child interclass correlation [Srivastava (1984)]. Assume x ∈ Rp+1 has density fx (x) = |Σ|−1/2 exp[(x − µ) Σ−1 (x − µ)], where

µ = Σ =

µm µs 1

,

2 σms 1 σm 2 σms 1 σs [(1 − ρss )I + ρss 11 ]

.

Here, “m” stands for mother and “s” means siblings. Let A = (1/p, Γ ) ∈ Rp for some Γ satisfying Γ1 = 0 and ΓΓ = Ip−1 . 2 (i) Interpret the parameters (µm , µs , σm , σms , ρss ). (ii) Prove that if 1 0 ˜ , A= 0 A

˜ A ˜ = diag(Ω, γ 2 Ip−1 ), where then AΣ s 2 σm σms Ω = , σms η 2

10.7. Problems

γs2 η2

173

= σs2 (1 − ρss ), =

[1 + (p − 1)ρss ]σs2 /p. |=

˜ is such that (y1 , y2 ) ∼ N2 ((µm , µs ) , Ω), (iii) Deduce that y = Ax 2 (y3 , . . . , yp+1 ). yi ∼ N (0, γs ) (i = 3, . . . , p + 1), and (y1 , y2 ) (iv) What are the implications for maximum likelihood estimation of the unknown parameters in i)? Remark: The yi ’s are not the principal components but are closely related to the concept. 5. Let f = (f1 , . . . , fp ) be the eigenvalues of the sample correlation matrix ρ matrix R. If the eigenvalues γα of the population correlation k are all distinct, then ﬁnd the limiting distribution of i=1 fi /p for the ratio of total variance explained by the ﬁrst k principal components.

11 Canonical correlations

11.1 Introduction The objective of canonical correlation analysis is to get a simple description of the structure of correlation between subsets of variables. Assume that two subsets of variables x1 and x2 have a joint normal distribution,

x1 x2

∼ Np

µ1 µ2

Σ11 , Σ21

Σ12 Σ22

.

The analysis searches for a pair of linear combinations t1 x1 and t2 x2 with maximum correlation. This is the ﬁrst canonical correlation. Having found such a pair, the analysis is pursued one step further by searching for a second pair of linear combinations with maximum correlation among all those uncorrelated with the ﬁrst pair. The correlation found is the second canonical correlation. The argument is repeated until all possible correlations are exhausted. This analysis is explained in detail in Section 11.2. In Section 11.3, tests of independence between x1 and x2 are derived. Not surprisingly, the tests proposed will be functions of the sample canonical correlations. Section 11.4 uses advantageously the context of testing independence to derive simple proofs of the properties of U (p; m, n) distributions introduced earlier in Section 9.3.2. As a by-product we also obtain a method of constructing Q-Q plots of squared radii for a visual inspection of multivariate normality. Asymptotic distributions of sample canonical correlations is the subject of Section 11.5.

11.2. Deﬁnition and basic properties

175

11.2 Deﬁnition and basic properties Assume Σjj > 0, Σij ∈ Rppij , i, j = 1, 2. Without any loss of generality, suppose p1 ≤ p2 . Write Σjj = A2j , where Aj > 0, j = 1, 2. Now using the SVD (v. Proposition 1.11), we have −1 A−1 1 Σ12 A2 = G(Dρ , 0)H ,

where Dρ = diag(ρ1 , . . . , ρp1 ), ρ1 ≥ · · · ≥ ρp1 ≥ 0, =

(g1 , . . . , gp1 ) ∈ Op1 ,

H =

(h1 , . . . , hp2 ) ∈ Op2 .

G

If we deﬁne u = G A−1 1 x1 = (u1 , . . . , up1 ) , −1 v = H A2 x2 = (v1 , . . . , vp2 ) ,

then

I (D , 0) p ρ 1 u . var = Dρ v Ip2 0

Obviously, var ui = var vj = 1 and cor(ui , vj ) = ρi δij , i = 1, . . . , p1 , j = 1, . . . , p2 . Deﬁnition 11.1 The variables u1 , . . . , up1 and v1 , . . . , vp2 are deﬁned to be the canonical variables. The numbers ρi ’s, 1 ≥ ρ1 ≥ ρ2 ≥ · · · ≥ ρp1 ≥ 0, are the canonical correlations. Note that the number of nonzero canonical correlations is rank Σ12 ≡ c. In a similar manner as the principal components were interpreted, the canonical variables can also be derived sequentially. First, we seek linear combinations t1 x1 and t2 x2 such that cor(t1 x1 , t2 x2 ) is maximal. But, in general, since cor(x, y) is invariant with respect to linear transformations, x → ax + b, y → cy + d, a, c > 0, we may assume at the outset that var tj xj = tj Σjj tj = 1, j = 1, 2. Introducing the ellipsoids Ej = {tj : tj Σjj tj = 1}, j = 1, 2, the problem is thus max t1 Σ12 t2 .

t1 ∈E1 t2 ∈E2

For tj ∈ Ej , |Aj tj | = 1, j = 1, 2, the Cauchy-Schwarz inequality gives (t1 Σ12 t2 )2

−1 2 = A1 t1 , A−1 1 Σ12 A2 h −1 2 ≤ |A−1 1 Σ12 A2 h| ,

176

11. Canonical correlations

−1 2 where h = A2 t2 has norm 1. Letting B = A−1 1 Σ12 A2 , then |Bh| = h B Bh, where Dρ2 0 −1 −1 −1 Σ A ) (A Σ A ) = H B B = (A−1 H . 12 2 12 2 1 1 0 0

Thus, from the method used for principal components, we ﬁnd h B Bh ≤ ρ21 with equality when h = h1 . This gives t2 x2 = h1 A−1 2 x2 = v1 . Finally, the Cauchy-Schwarz inequality is, in fact, an equality iﬀ A1 t1 ∝ Bh1 or, equivalently, t1

−1 −1 ∝ A−1 1 A1 Σ12 A2 h1 = A−1 1 G(Dρ , 0)H h1

= A−1 1 G(Dρ , 0)e1 = ρ1 A−1 1 g1 , which, in turn, gives t1 x1 = g1 A−1 1 x1 = u1 . We have proved that (u1 , v1 ) is the pair of linear combinations with maximum correlation ρ1 . Second, having found pairs of linear combinations −1 (ui , vi ) = (gi A−1 1 x1 , hi A2 x2 ), i = 1, . . . , k, k < rank Σ12 ≡ c,

another pair (t1 x1 , t2 x2 ) is sought with maximum correlation among all those uncorrelated with the preceding pairs; i.e., the restriction cov(tj xj , ui ) = cov(tj xj , vi ) = 0, i = 1, . . . , k; j = 1, 2, is imposed. This last restriction is characterized in terms of orthogonality: cov(t1 x1 , ui ) = t1 Σ11 A−1 1 gi = t1 A1 gi = 0 ⇐⇒ t1 ⊥ A1 gi .

Similarly, cov(t2 x2 , vi ) = 0 iﬀ t2 ⊥ A2 hi . We note that when t1 ⊥ A1 gi , the other condition, cov(t1 x1 , vi ) = 0, is automatically satisﬁed: cov(t1 x1 , vi )

= t1 Σ12 A−1 2 hi −1 = t1 A1 (A−1 1 Σ12 A2 )hi = t1 A1 G(Dρ , 0)H hi = ρi t1 A1 gi = 0.

Similarly, when t2 ⊥ A2 hi then cov(t2 x2 , ui ) = 0 is automatically satisﬁed. So the problem becomes max t1 Σ12 t2 ,

⊥ t1 ∈E1 ⊥ t2 ∈E2

where E1⊥ E2⊥

= {t1 ∈ E1 : t1 ⊥ A1 g1 , . . . , A1 gk }, = {t2 ∈ E2 : t2 ⊥ A2 h1 , . . . , A2 hk }.

11.3. Tests of independence

177

The Cauchy-Schwarz inequality gives for tj ∈ Ej⊥ , (t1 Σ12 t2 )2 = A1 t1 , Bh2 ≤ h B Bh, −1 and h = A2 t2 . But, using the where, as before, B = A−1 1 Σ12 A2 orthogonality restrictions, Dρ2 0 h B Bh = t2 A2 H H A2 t2 0 0 c = (t2 A2 hi )2 ρ2i ≤ ρ2k+1 , i=k+1

with equality when h = A2 t2 = hk+1 , which yields t2 x2 = hk+1 A−1 2 x2 = vk+1 . As before, the Cauchy-Schwarz inequality becomes an equality iﬀ −1 A1 t1 ∝ Bhk+1 , which implies t1 = A−1 1 gk+1 and t1 x1 = gk+1 A1 x1 = uk+1 . The solution is the pair of canonical variables (uk+1 , vk+1 ). Repeating the second stage for k = 1, . . . , c − 1, all the pairs of canonical variables (ui , vi ), i = 1, . . . , c, can be generated. Each pair of canonical variables is identiﬁed with the pair of linear combinations of x1 and x2 with maximum correlation among all those uncorrelated with the preceding pairs. Finally, the canonical correlations can be characterized as solutions of a determinant equation. In fact, the nonzero squared canonical correlations ρ2i , i = 1, . . . , c, are the nonzero eigenvalues of −1 −1 −1 −1 −1 −1 B B = (A−1 1 Σ12 A2 ) (A1 Σ12 A2 ) = A2 Σ21 Σ11 Σ12 A2 .

Hence, the nonzero ρ2i are the nonzero solutions λ of the equation −1 |Σ12 Σ−1 22 Σ21 Σ11 − λI| = 0.

11.3 Tests of independence Based on a random sample of size n from a Np1 +p2 (µ, Σ), where Σ11 Σ12 Σ= Σ21 Σ22 with Σij ∈ Rppij , we construct a test of independence reﬂected by the hypothesis, H0 : Σ12 = 0 ⇐⇒ H0 : ρ1 = · · · = ρp1 = 0, against all alternatives. The unbiased estimator S of Σ is partitioned in conformity as V11 V12 (n − 1)S ≡ V = V21 V22

178

11. Canonical correlations

ˆ = V/n is and we know already that V ∼ Wp (n − 1, Σ). The MLE Σ proportional to S. Without any restriction, the MLE of Σij , i, j = 1, 2, is ˆ ij = Vij /n. However, under H0 , the restricted MLE’s are given by Σ ˆ ˆ ˆ ˆ 11 = Σ ˆ 11 , Σ ˆ 22 = Σ ˆ 22 , Σ ˆ 12 = 0. Σ The LRT takes the form Λ = =

ˆ ˆ ˆ ˆ 22 , Σ ˆ 12 ) ˆ 11 , Σ L(¯ x, Σ ˆ 11 , Σ ˆ 22 , Σ ˆ 12 ) L(¯ x, Σ ˆ 22 |−n/2 ˆ 11 |−n/2 |Σ |Σ . ˆ −n/2 |Σ|

ˆ ∝ V and using the relation |V| = |V11 ||V22.1 |, Thus, since Σ Λ2/n

=

|V11.2 | |V22.1 | = |V22 | |V11 |

−1 −1 = |I − V12 V22 V21 V11 | p1 = (1 − ri2 ) i=1

is a function of the sample canonical correlations ri ’s, where ri2 is a solution λ of the equation −1 −1 V21 V11 − λI| = 0. |V12 V22

They satisfy w.p.1, 1 > r12 > · · · > rp21 > 0. Consider now the invariant tests. The group Gp1 × Gp2 × Rp1 × Rp2 transforms the observations as B1 0 xi1 b1 xi1 + → xi2 0 B2 xi2 b2 B1 xi1 + b1 = , i = 1, . . . , n, B2 xi2 + b2 for any (B1 , B2 , b1 , b2 ) ∈ Gp1 ×Gp2 ×Rp1 ×Rp2 . This induces the following transformations on the minimal suﬃcient statistic (¯ x, V): ¯1 ¯ 1 + b1 x B1 x → , ¯2 ¯ 2 + b2 B2 x x B1 0 V11 V12 B1 0 V11 V12 → V21 V22 V21 V22 0 B2 0 B2 B1 V11 B1 B1 V12 B2 = . B2 V21 B1 B2 V22 B2 A test function f (¯ x, V) is invariant iﬀ B1 W11 B1 B1 y1 + b1 , f (y, W) = f B2 y2 + b2 B2 W21 B1

B1 W12 B2 B2 W22 B2

,

11.3. Tests of independence

179

∀(B1 , B2 , b1 , b2 ) ∈ Gp1 × Gp2 × Rp1 × Rp2 , ∀(y, W) ∈ Rp × Pp . The choice ¯ i , i = 1, 2, immediately yields bi = −Bi x B1 V11 B1 B1 V12 B2 f (¯ x, V) = f 0, . B2 V21 B1 B2 V22 B2 Using the same arguments as in the deﬁnition of canonical correlations, let Vii = A2i , where Ai > 0, i = 1, 2, and consider the SVD −1 A−1 1 V12 A2 = G(Dr , 0)H ,

where Dr = diag(r1 , . . . , rp1 ), 1 > r1 > · · · > rp1 > 0, and we still assume and p1 ≤ p2 without loss of generality. Then, the choice B1 = G A−1 1 B2 = H A−1 2 ﬁnally gives Ip1 (Dr , 0) ; f (¯ x, V) = f 0, Dr Ip2 0 i.e., any invariant test is a function of the sample canonical correlations ri ’s. A similar argument shows that the power function of any invariant test depends only on the population canonical correlations ρi ’s. Proposition 11.1 With respect to the block-diagonal group of transformations above, any invariant test depends on the minimal suﬃcient statistic (¯ x, V) only through the sample canonical correlations ri ’s. The power function of any invariant test depends on (µ, Σ) only through the population canonical correlations ρi ’s. We now derive the null distribution of the LRT test. Proposition 11.2 Under the hypothesis of independence, H0 : Σ12 = 0 and n − 1 > min(p1 , p2 ), Λ2/n ∼ U (p2 ; p1 , n − 1 − p1 ). Proof. By invariance, assume without loss of generality that Σ = I and let m = n − 1. Write V11 V12 d X1 = (X1 , X2 ), V21 V22 X2 where (X1 , X2 ) ∼ Npm (0, Im ⊗ Ip ). The conditional distribution of X2 given X1 is X2 | X1 ∼ Npm2 (0, Im ⊗ Ip2 ). w.p.1 = p1 . Therefore, we have the SVD D G X 1 H = , 0

Now, for X1 ∈ Rm p1 , rank X1

180

11. Canonical correlations

where G ∈ Om , H ∈ Op1 , and D ∈ Rpp11 is diagonal and nonsingular. Thus, ˜1 DH X G X 1 = ≡ , 0 0 ˜ 1 ∈ Rp1 is nonsingular. Since G (a function of X1 ) is orthogonal, where X p1 G X2 | X1 ∼ Npm2 (0, Im ⊗ Ip2 ),

1 . Then, Y where Y ∈ Rpp12 and Z ∈ Rm−p p2

V22 V11 V12

|=

which does not depend on X1 and so G X2 ∼ Npm2 (0, Im ⊗ Ip2 ) unconditionally. Now, partition Y G X 2 = , Z Z and

= X2 X2 = (G X2 ) (G X2 ) = Y Y + Z Z, ˜X ˜ = X1 X1 = (G X1 ) (G X1 ) = X 1 1, ˜ = X X2 = (G X1 ) (G X2 ) = X Y. 1

1

Finally, Λ2/n

=

|V22.1 | |V22 |

=

˜X ˜ 1 (X ˜ −1 X ˜ Y| |Y Y + Z Z − Y X |Z Z| 1 1) 1 = , |Y Y + Z Z| |Y Y + Z Z|

where Z Z ∼ Wp2 (m − p1 ), and Y Y ∼ Wp2 (p1 ). By deﬁnition, Λ2/n ∼ U (p2 ; p1 , m − p1 ). 2 −1 −1 Let R = V12 V22 V21 V11 . As for multivariate regression, other invariant tests can be constructed such as

tr R =

p1

ri2 ,

i=1

tr R(I − R)−1

=

r12

=

p1 i=1

ri2 , (1 − ri2 )

max{r12 , . . . , rp21 }.

Again, none of these tests has a power function which uniformly dominates the others. It is shown in Example 14.10 how to perform a bootstrap test using the test statistics tr R or tr R(I − R)−1 .

11.4. Properties of U distributions

181

11.4 Properties of U distributions We end this chapter with some properties and characterizations useful for the tabulation and moments of U distributions. These simpliﬁed proofs are from Bilodeau (1996). Assume x1 ∈ Rp1 is ﬁxed and x2 ∼ Np2 (0, Σ22 ), p = p1 + p2 , Σ22 > 0. Based on a random sample of size n, say X ∈ Rnp , the matrix of sums of squares and cross-products V is partitioned in conformity as X1 V11 V12 V=XX= (X1 , X2 ) = . X2 V21 V22 When n > min(p1 , p2 ) and rank X1 = p1 , consider ˜= Λ

|V22.1 | |V11.2 | |V| = = . |V11 ||V22 | |V22 | |V11 |

Proposition 11.3 If n > min(p1 , p2 ) and rank X1 = p1 , then ˜ ∼ U (p2 ; p1 , n − p1 ). Λ Proof. Assume without loss of generality Σ22 = I and thus X2 ∼ Npn2 (0, I). Now, X1 ∈ Rnp1 has rank X1 = p1 . Its singular value decomposition is D G X 1 H = , 0 where G ∈ On , H ∈ Op1 , and D ∈ Rpp11 is diagonal and nonsingular. Thus, ˜1 DH X G X 1 = ≡ 0 0

1 . Then, Y where Y ∈ Rpp12 and Z ∈ Rn−p p2 ˜ ˜ ˜ V11 = X1 X1 , and V12 = X1 Y. Finally,

|=

˜ 1 ∈ Rp1 is nonsingular. Since G (a function of X1 ) is orthogonal, where X p1 G X2 ∼ Npn2 (0, I). Partition Y G X2 = , Z Z and V22 = Y Y + Z Z,

˜ ˜ ˜ −1 ˜ |Z Z| ˜ = |V22.1 | = |Y Y + Z Z − Y X1 (X1 X1 ) X1 Y| = Λ , |V22 | |Y Y + Z Z| |Y Y + Z Z|

˜ ∼ U (p2 ; p1 , n− where Z Z ∼ Wp2 (n−p1 ), Y Y ∼ Wp2 (p1 ). By deﬁnition, Λ 2 p1 ).

|=

Proposition 11.3 remains valid if X1 has any absolutely continuous distribution (and thus has rank X1 = p1 w.p.1 (v. Lemma 7.1 and the remark ˜ does X2 . It suﬃces to notice the distribution of Λ on page 88)) and X1 not depend on X1 .

182

11. Canonical correlations

˜ = |V11.2 |/|V11 |, if X1 is normal and X2 is ﬁxed, Vice versa, writing Λ ˜ ∼ U (p1 ; p2 , n − p2 ). The rank X2 = p2 , it is clear the same proof yields Λ d

duality property asserts that, in fact, U (p1 ; p2 , n − p2 ) = U (p2 ; p1 , n − p1 ). As a by-product, we show the “duality” property: d

|=

Corollary 11.1 U (p; m, n) = U (m; p, m + n − p) when m + n > p. X2 . Since X1 is Proof. Assume X1 and X2 are both normal and X1 ˜ ∼ U (p2 ; p1 , n− ˜ ∼ U (p1 ; p2 , n−p2 ), and since X2 is also normal, Λ normal, Λ d ˜ being unique, U (p1 ; p2 , n−p2 ) = p1 ). The distribution of Λ U (p2 ; p1 , n−p1 ). 2 Substitute (p, m, m + n) for (p1 , p2 , n). In order to obtain a characterization of U distributions as a product of independent beta variables, we prove the following lemma. Lemma 11.1 If n ≥ p, d

U (p; 1, n) = beta

1

2 (n

− p + 1); 12 p .

Proof. When m = 1, recalling the identity |I + AB| = |I + BA| (v. Problem 1.8.3), d

where z

|=

U (p; 1, n) =

|W| = |I + W−1 zz |−1 = (1 + z W−1 z)−1 , |W + zz |

W, z ∼ Np (0, I), and W ∼ Wp (n). Using Proposition 8.2, z W−1 z ∼ Fc (p, n − p + 1).

Finally, using Problem 3.5.5, (1 + z W−1 z)−1 ∼ beta

1

2 (n

− p + 1); 12 p . 2

Proposition 11.4 A variable distributed as U (p; m, n), n ≥ p, has the two characterizations m

d beta 12 (n − p + i); 12 p U (p; m, n) = i=1

and d

U (p; m, n) =

p

beta

1

2 (n

− p + i); 12 m ;

i=1

i.e., a U (p; m, n) variable has the same distribution as a product of independent beta variables. Proof. The second representation follows from the ﬁrst and the duality property of U distributions. We need only show the ﬁrst representation. Its proof proceeds by induction on m. From Lemma 11.1, the result is true for

11.4. Properties of U distributions

183

d

U (p; m, n) = =

|W| , W ∼ Wp (n), Z ∼ Npm (0, Im ⊗ Ip ), Z |W + Z Z| |W| |(W + z1 z1 )| ≡ U1 · U2 , · |W + z1 z1 | |(W + z1 z1 ) + Z2 Z2 |

where

Z=

z1 Z2

|=

m = 1. Assume the result is true for m − 1 and show it holds for m. By deﬁnition, W

,

|=

|=

Z2 . Consider z1 ∼ Np (0, I), Z2 ∼ Npm−1 (0, Im−1 ⊗ Ip ), and z1 now the distribution of U2 . Let W1 = W + z1 z1 and W2 = Z2 Z2 . Then, W1 ∼ Wp (n + 1), W2 ∼ Wp (m − 1), and W1 W2 . Therethe induction hypothesis gives U2 ∼ fore, U2 ∼ U (p; m − 1, n + 1) and m−1 1 1 i=1 beta 2 (n + 1 − p + i); 2 p . Translating i → i + 1, U2 ∼

m

beta

1

2 (n

− p + i); 12 p .

i=2

|=

|=

|=

|=

|=

|=

|=

The factor missing for i = 1 is U1 . The proof is complete if we prove U2 . First, note that if U1 W1 , then U1 , W1 , and Z2 are mutuU1 U2 . So, we prove U1 W1 . But, if ally independent and, therefore, U1 d V ∼ Wp (n, Σ), Σ > 0, x ∼ Np (0, Σ), and V x, then (V, x) = (Y Y, x), n x. In the model for (Y, x), Y Y + xx where Y ∼ Np (0, In ⊗ Σ), Y is complete and suﬃcient for Σ. Therefore, V + xx is complete and sufﬁcient for Σ. Using Basu’s theorem in the footnote on page 118, V + xx is independent of any ancillary statistic such as |V|/|V + xx |. This proves W1 . 2 U1 This representation is useful for ﬁnding the distribution function or quantiles of a U (p; m, n) distribution since ln U (p; m, n) can be represented as a convolution of simple distributions. Of course, it is advantageous to use the representation with min(p, m) number of factors. This number of factors can be reduced further by 12 by grouping adjacent factors by pairs [Anderson (1984), p. 304]. The following lemma allows the pairing. Lemma 11.2 For n > 1, [beta(n − 1; m)]2 = beta( 12 (n − 1); 12 m) · beta( 12 n; 12 m). d

Proof. It is straightforward to check that all moments of order h > 0 on d the left and right sides of = are the same (v. Problem 11.6.3). Since the domain is the bounded interval [0, 1], there is a unique distribution with these moments [Serﬂing (1980), p. 46]. 2 The following representation has a reduced number of factors as p is even or odd.

184

11. Canonical correlations

Corollary 11.2 For n ≥ p, a U (p; m, n) variable can be represented as r

beta( 12 (n − p + 1); 12 m) ·

i=1 r

[beta(n + 1 − 2i; m)]2 ,

if p = 2r

[beta(n + 1 − 2i; m)]2 ,

if p = 2r + 1.

i=1

Proof. The proof for p = 2r is as follows. From Proposition 11.4, we have d

U (p; m, n) =

r

beta( 12 (n − p + 2i − 1); 12 m) beta( 12 (n − p + 2i); 12 m),

i=1

and from Lemma 11.2, d

U (p; m, n) =

r

[beta(n − p + 2i − 1; m)]2 .

i=1

The conclusion follows after reversing the index i → r − i + 1. The proof for p odd is identical except for the ﬁrst isolated factor. 2 The asymptotic distribution as n → ∞ of U (p; m, n) should be clear from the asymptotic distribution of the likelihood ratio statistic in Proposition 11.2, −n ln U (p; m, n) → χ2pm . d

(11.1)

The slight modiﬁcation −[n − 12 (p − m + 1)] ln U (p; m, n) → χ2pm d

is often used as an improved approximation since it has a remainder of order O(n−2 ), whereas the remainder in (11.1) is O(n−1 ). The general asymptotic expansion of order O(n−α ) [Box (1949)] is treated in Section 12.3. As an alternative to asymptotic expansion an S-plus program in Appendix C uses the fast Fourier transform [Press (1992)] to compute the density of U (p; m, n) by convolution and thus calculates exact probabilities (up to a discretization of the beta variables) and quantiles. Srivastava and Yau (1989) presented the saddlepoint method for obtaining tail probabilities. An exact closed form solution without series representation was also recently derived [Coelho (1998)].

11.4.1

Q-Q plot of squared radii

The scaled residuals of n observations, xi , may be deﬁned as ¯ ), i = 1, . . . , n. zi = S−1/2 (xi − x ¯ ) are Then, the squared radii (or squared Mahalanobis distances of xi to x ¯ ) S−1 (xi − x ¯ ), i = 1, . . . , n. d2i = |zi |2 = (xi − x

11.4. Properties of U distributions

185

Note that if xi ∼ Np (µ, Σ), then di is an ancillary statistic; i.e., the distribution of di , say F (·), does not depend on (µ, Σ). One aspect of multivariate normality can thus be tested with a Q-Q plot of the ordered d2i against the quantiles of the distribution F (·) [Small (1978)]. Gnanadesikan and Kettenring (1972) derived the following result. Lemma 11.3 If x1 , . . . , xn are i.i.d. Np (µ, Σ), then

n d2i ∼ beta 12 p; 12 (n − p − 1) . 2 (n − 1) Proof. 1−

n d2 (n − 1)2 i

n ¯ )(xi − x ¯ ) |/|V| (xi − x (n − 1) = |W1 |/|W1 + W2 |,

= |V −

n n ¯ )(xi − x ¯ ) , W1 = V − W2 , and W2 = (n−1) where V = i=1 (xi − x (xi − ¯ ) . Assume without loss of generality that µ = 0 and Σ = I. Thus, ¯ )(xi − x x with Z ∼ Npn (0, In ⊗ Ip ), (W1 , W2 ) = (Z (Q − H)Z, Z HZ) , d

where H = Q

n (ei − n−1 1)(ei − n−1 1) , (n − 1)

= I − n−1 11 .

The following can be veriﬁed easily: (i) H is idempotent of rank 1, (ii) Q is idempotent of rank n − 1, (iii) Q(ei − n−1 1) = (ei − n−1 1) and, thus, QH = H, (iv) Q − H is idempotent of rank n − 2, (Q − H)H = 0. |=

Thus, W1 W2 , W1 ∼ Wp (n − 2), and W2 ∼ Wp (1) (v. Proposition 7.8 and Problem 6.4.3), which implies

d |W1 |/|W1 + W2 | ∼ U (p; 1, n − 2) = beta 12 (n − p − 1); 12 p . 2 Consider the ordered

d2i , d2(1) ≤ d2(2) ≤ · · · ≤ d2(n) .

Assuming d2i , i = 1, . . . , n, are i.i.d. according to the distribution in Lemma 11.3, one could evaluate the expected order statistics, E d2(i) . Then, the Q-Q plot consists of a graph of the points d2(i) , E d2(i) , i = 1, . . . , n.

186

11. Canonical correlations

To simplify matters, we can assign to d2(i) a cumulative probability of i/n and approximate E d2(i) by the quantile γi = i/n of the distribution [(n − 1)2 /n] beta

1

1 2 p; 2 (n

− p − 1) .

Blom (1958) has shown how to select α and β so that the expected order statistic E d2(i) may be well approximated by the quantile γi = (i − α)/(n − α − β + 1).

(11.2)

For beta, the distribution at hand, the indicated choice is α

=

β

=

(p − 2) , 2p (n − p − 2) . 2(n − p − 1)

(11.3)

Thus, the recommended Q-Q plot is the graph of the points

d2(i) , [(n − 1)2 /n] betaγi 12 p; 12 (n − p − 1) , i = 1, . . . , n, where betaα (a; b) denotes the quantile α of a beta(a; b) distribution and γi is given by (11.2) and (11.3). The Splus function qqbeta in Appendix C produces the Q-Q plot. One should not forget, however, that the d2i are correlated, but from Wilks (1963), cor(d2i , d2j ) = −

1 , i = j, (n − 1)

and the correlation, of the order O(n−1 ), is negligible for moderate to large sample sizes. A Q-Q plot approaching a 45◦ straight line is consistent with multivariate normality. Figure 11.1 gives the Q-Q plot for 50 observations generated from a N3 (0, I) distribution and Figure 11.2 is the Q-Q plot for 50 observations generated from a trivariate Cauchy distribution. These are easily generated with Example 13.2. The deviations from the straight line are clearly more systematic in Figure 11.2 associated with a distribution with heavier “tails” than the multivariate normal. For large n, the beta distribution can be approximated by a χ2p distribution. Gnanadesikan (1977, p. 172) remarked that in the bivariate case n = 25 may provide a suﬃciently large sample for this chi-squared approximation to be adequate. However, n = 100 does not seem large enough for p = 4, for there is a marked deviation from linearity when the ordered d2i are plotted against expected order statistics of chi-squared, and this eﬀect becomes more marked as p increases. We therefore recommend the use of the beta distribution.

11.4. Properties of U distributions

187

10

•

8

•

• 6

•

4 0

2

Beta quantiles

•

• •• ••• •• •• • • •• 0

•• •• • • ••• •• •

2

• •• • •• •••

4

•

• • •

6

8

10

12

14

Squared distances Figure 11.1. Q-Q plot for a sample of size n = 50 from a trivariate normal, N3 (0, I), distribution.

11. Canonical correlations

•

10

188

8

•

• •

6 4 0

2

Beta quantiles

•

• •• • •• •• • •• •• ••••• •• ••• •• •••• • 0

••

• •

5

•

•

10

15

20

25

Squared distances Figure 11.2. Q-Q plot for a sample of size n = 50 from a trivariate t on 1 degree of freedom, t3,1 (0, I) ≡ Cauchy3 (0, I), distribution.

11.5. Asymptotic distributions

189

11.5 Asymptotic distributions Assuming xi , i = 1, . . . , n, are i.i.d. Np (µ, Σ), we derive the asymptotic distribution of rα2 , α = 1, . . . , p1 , when ρα is distinct from all other canonical correlations. A squared sample canonical correlation, rα2 , is a value of l for which there is a nonzero solution c to the equation −1 (S−1 11 S12 S22 S21 − l I)c = 0.

(11.4)

d

Using the result n1/2 (S − Σ) → W of Section 6.3, where W ∼ Npp (0, (I + K)(Σ ⊗ Σ)), we write S11 S22

= I + n−1/2 W11 ,

S12

=

= I + n−1/2 W22 , (Dρ , 0) + n−1/2 W12 .

−1/2 W22 + Op (n−1 ) and similarly for Using Problem 1.8.15, S−1 22 = I − n −1 −1/2 , S11 . Keeping terms up to order n −1 S−1 11 S12 S22 S21

= D2ρ + n−1/2 −D2ρ W11 + (Dρ , 0)W21 Dρ Dρ −(Dρ , 0)W22 + W12 + Op (n−1 ). 0 0

We now apply the perturbation method as in Section 8.8.1 and obtain from (8.11) the expansion αα αα αα − ρ2α w22 + 2ρα w21 ] + Op (n−1 ), rα2 = ρ2α + n−1/2 [−ρ2α w11 αα is the element (α, α) of the matrix Wij . From (6.1), we have where wij αα αα αα (w11 , w22 , w21 ) → N3 (0, Ω), d

where

2 Ω = 2ρ2α 2ρα

2ρ2α 2 2ρα

2ρα 2ρα . 1 + ρ2α

Finally, deﬁning the linear combination vector a = (−ρ2α , −ρ2α , 2ρα ) , we d

obtain n1/2 (rα2 − ρ2α ) → N (0, a Ωa), whereby a direct calculation shows a Ωa = 4ρ2α (1 − ρ2α )2 . We have shown: Proposition 11.5 The asymptotic distribution of the squared sample canonical correlation rα2 , α = 1, . . . , p1 , assuming ρα is distinct from all d

other canonical correlations is n1/2 (rα2 − ρ2α ) → N (0, 4ρ2α (1 − ρ2α )2 ).

190

11. Canonical correlations

Various extensions of Proposition 11.5 to the joint distribution of sample canonical correlations can be envisaged. The simplest extension is to the joint distribution of r12 , . . . , rp21 when all canonical correlations are distinct, ρ1 > · · · > ρp1 . Corollary 11.3 The asymptotic joint distribution of r12 , . . . , rp21 when all population canonical correlations are distinct, ρ1 > · · · > ρp1 , is n1/2 (r12 − ρ21 , . . . , rp21 − ρ2p1 )

d → Np1 0, 4 diag ρ21 (1 − ρ21 )2 , . . . , ρ2p1 (1 − ρ2p1 )2 . Proof. It suﬃces to consider the asymptotic covariance of two squared sample canonical correlations, rα2 and rβ2 , when the population canonical correlations, ρα and ρβ , are of multiplicity 1. But, it is immediate from the proof of Proposition 11.5 that ββ ββ ββ αα αα αα − ρ2α w22 + 2ρα w21 , −ρ2β w11 − ρ2β w22 + 2ρβ w21 )=0 cov(−ρ2α w11 ββ αα since from (6.1), all the covariances satisfy, cov(wij , wkl ) = 0, i, j, k, l = 1, 2. 2

Hsu (1941) derived the asymptotic joint density when 1 > ρ1 > · · · > ρc > ρc+1 = · · · = ρp1 = 0. Muirhead and Waternaux (1980) obtained the asymptotic joint distribution when all population canonical correlations are distinct, as in Proposition 11.3 but for any underlying distribution with ﬁnite fourth moments. Eaton and Tyler (1994), assuming an underlying elliptical distribution or, in fact, any other distribution with ﬁnite fourth moments, derived the asymptotic joint distribution in full generality, ρ1 ≥ · · · ≥ ρc > ρc+1 = · · · = ρp1 = 0, using an extension of Wielandt’s inequality to singular values. In canonical correlation analysis, the number of nonzero population correlations is called the dimensionality. Asymptotic distributions of the dimensionality estimated by Mallow’s criterion and Akaike’s criterion were derived [Gunderson and Muirhead (1997)] for non-normal multivariate populations with ﬁnite fourth moments.

11.6 Problems 1. Obtain the hth moment, h > 0, of U ∼ U (p; m, n), n ≥ p, m Γ 12 (n − p + i) + h Γ 12 (n + i) h = EU · 1 1 Γ (n − p + i) Γ (n + i) + h 2 2 i=1

11.6. Problems

=

p Γ i=1

191

1 Γ 12 (m + n − p + i) 2(n − p + i) +h . · 1 Γ 12 (n − p + i) Γ 2 (m + n − p + i) + h

2. Establish the following exact results concerning U distributions: n[1 − U (1; m, n)] m U (1; m, n) (n − p + 1)[1 − U (p; 1, n)] p U (p; 1, n) (n − 1)[1 − U (2; m, n)1/2 ] m U (2; m, n)1/2 (n − p + 1)[1 − U (p; 2, n)1/2 ] p U (p; 2, n)1/2

∼ F (m, n), ∼ F (p, n − p + 1), ∼ F (2m, 2(n − 1)), ∼ F (2p, 2(n − p + 1)).

3. Prove that [beta(n − 1, m)]2 and beta( 12 (n − 1), 12 m) · beta( 12 n, 12 m), a product of two independent betas, have the same moments of order h > 0. 4. For x = (x1 , x2 ) = (x1 , . . . , xp1 ; xp1 +1 , . . . , xp1 +p2 ) establish that simple correlation and multiple correlation coeﬃcients are bounded above as (i) |ρxi ,xj | ≤ ρ1 , i = 1, . . . , p1 , j = p1 + 1, . . . , p1 + p2 , (ii) Rxi ,x2 ≤ ρ1 , i = 1, . . . , p1 , where ρ1 is the largest canonical correlation. 5. Let Σ12 = ρ1p1 1p2 , Σii = ρ1pi 1pi + (1 − ρ)Ipi , i = 1, 2, corresponding to the equicorrelated case. Determine the canonical variables corresponding to the nonzero canonical correlation. Hint: Σ11 1p1 = [1 + (p1 − 1)ρ]1p1 . 6. Let x1 , . . . , xn be i.i.d. Np (µ, Σ), where Σ11 Σ12 Σ= Σ21 Σ22 with Σ12 ∈ Rpp12 , p = p1 + p2 . For testing H0 : Σ12 = 0 against H1 : Σ12 = 0, consider the test statistic [Escouﬁer (1973)] E=

tr(S12 S21 ) , [tr(S211 )]1/2 [tr(S222 )]1/2

where S is the sample variance partitioned as Σ. Prove: (i) E is invariant under the group of transformations xi1 H1 0 xi1 b1 xi = → + xi2 xi2 b2 0 H2

192

11. Canonical correlations

for any (H1 , H2 , b1 , b2 ) ∈ Op1 × Op2 × Rp1 × Rp2 . (ii) If H0 holds, then the distribution of E is the same as when diag(λi ) 0 Σ= , 0 diag(γj ) where λi and γj are, respectively, the eigenvalues of Σ11 and Σ22 . d (iii) Under H0 , n1/2 S12 → Z = (zij ), where zij are independently distributed as N (0, λi γj ), i = 1, . . . , p1 , j = 1, . . . , p2 . (iv) Conclude the null distribution −1/2 p −1/2 p p1 p2 1 2 d 2 λ2i γj2 λi γj zij . nE→ i=1

j=1

i=1 j=1

Remark: Unlike for canonical correlations, the asymptotic null distribution depends on unknown parameters because of the lack of invariance of E (the group Op1 × Op2 × Rp1 × Rp2 is only a subgroup of Gp1 ×Gp2 ×Rp1 ×Rp2 ). The asymptotic distribution of E for sampling from an elliptical distribution was derived by Cl´eroux and Ducharme (1989). 7. Test of mutual independence of several subvectors. This problem given in the form of a project derives the exact null distribution of the likelihood ratio test for mutual independence. Consider a random sample of size n ≥ p + 1 from Np (µ, Σ), where Σ11 Σ12 · · · Σ1r Σ21 Σ22 · · · Σ2r Σ= .. .. .. ... . . . r

Σr1

···

Σr2

Σrr

with Σij ∈ Rppij , p = j=1 pj . We wish to test H0 : Σij = 0, 1 ≤ i < j ≤ r, versus all alternatives. (i) Prove the likelihood ratio test Λ for H0 can be written |V| , i=1 |Vii |

Λ2/n = r

n ¯ )(xi − x ¯ ) ∼ Wp (n − 1, Σ). where as usual V = i=1 (xi − x ˜ = Λ2/n , (ii) Obtain the exact null moments of Λ 1 Γpi ( 1 m) 2 ˜ h = Γp ( 2 m + h) EΛ , 1 Γp ( 2 m) i=1 Γpi ( 12 m + h) r

where m = n − 1.

11.6. Problems

193

Hint: ˜h = EΛ

cp,m

r

cp,m+2h

i=1

where cp,m = [2pm/2 Γp ( 12 m)]−1 is of a Wp (m) density and where Vii Wpi (m + 2h). (iii) Deﬁne the upper left corner of V to V11 V12 · · · V21 V22 · · · ˜ ii = . V .. .. .. . . Vi1

Vi2

···

E |Vii |−h , the normalizing constant are mutually independent be V1i V2i ∈ Rpp¯¯ii , .. . Vii

˜ ˜ where p¯i = p1 +· · ·+pi , and note that rVrr = V and V11 = V11 . Derive the equivalent form Λ2/n = i=2 Ui , where Ui =

˜ ii | |V , i = 2, . . . , r. ˜ i−1,i−1 | |Vii ||V

(iv) Use Proposition 11.2 to obtain immediately under H0 Ui ∼ U (pi ; p¯i−1 , n − 1 − p¯i−1 ) , i = 2, . . . , r. (v) When Σ=

˜ r−1,r−1 Σ 0

0 Σrr

,

|=

|=

|=

˜ r−1,r−1 , Vrr ) is suﬃcient and complete and that prove that (V Ur is ancillary. Conclude that, under H0 , Ur Vr−1,r−1 , Vrr ). U2 , . . . , Ur−1 ) under H0 . (vi) Using (iii), prove that Ur U2 , . . . , Ui−1 ), i = (vii) Repeat this argument to prove Ui 3, . . . , r, whence, altogether, U2 , . . . , Ur are mutually independent under H0 . (viii) Use Proposition 11.4 to obtain the exact null distribution d ˜= Λ

pi r

beta

1

2 (n

− p¯i−1 − j); 12 p¯i−1 .

i=2 j=1

Note that a further representation with a reduced number of factors as pi is odd or even is immediate from Corollary 11.2. n/2 (ix) Prove Ui is the likelihood ratio test for Hi : Σli = 0, l = 1, . . . , i − 1 when it is known that all the hypotheses Hi+1 , · · ·, Hr are true. Note that H0 = ∩ri=2 Hi .

194

11. Canonical correlations

(x) Use (viii) to obtain the equivalent expression for the exact null ˜ moments of Λ: pi r 1 1 Γ[ (n − p ¯ − j) + h] Γ[ (n − j)] i−1 2 2 ˜h = . EΛ Γ[ 12 (n − p¯i−1 − j)] Γ[ 12 (n − j) + h] i=2 j=1 8. Generalized squared interpoint distance [Gnanadesikan and Kettenring (1972)] Let d2ij = (xi −xj ) S−1 (xi −xj ) be the generalized squared interpoint distance (or squared Mahalanobis distance) between xi and xj , i = j. Prove that if x1 , . . . , xn are i.i.d. Np (µ, Σ), then

1 d2ij ∼ beta 12 p; 12 (n − p − 1) . 2(n − 1)

12 Asymptotic expansions

12.1 Introduction The exact distribution of likelihood ratio tests in multivariate analysis is often too complicated to be of any practical use. An asymptotic expansion due to Box (1949) is rather simple and easy to program on a computer to obtain the distribution function to any degree of accuracy. This approximation is applied on several of the testing situations previously encountered. In at least one situation where the exact distribution is known, an evaluation of the approximation is carried out for small to moderate sample sizes.

12.2 General expansions The method can be used whenever the likelihood ratio criterion Λ (or a suitable power W ) has moment of order h of the form b yj h a y j j=1 k=1 Γ[xk (1 + h) + ζk ] , (12.1) E W h = K a b xk x k=1 k j=1 Γ[yj (1 + h) + ηj ] where b j=1

yj =

a k=1

xk ,

196

12. Asymptotic expansions

and K is just a constant (not depending on h) so that E W 0 = 1. Equation (12.1) is usually obtained for real h; it is, however, generally valid on the domain where the functions are analytic. This means if we let M = −2 log W , then we can write the characteristic function of ρM , for a constant 0 < ρ ≤ 1 to be determined later, as cρM (t)

= E W −2itρ b

y y j aj=1 jxk k=1 xk

= K

−2itρ a k=1

Γ[xk (1 − 2itρ) + ζk ]

j=1

Γ[yj (1 − 2itρ) + ηj ]

b

.

Taking logarithms and deﬁning βk = (1 − ρ)xk ,

j = (1 − ρ)yj ,

(12.2)

the cumulant generating function is KρM (t) = log cρM (t) = g(t) − g(0),

(12.3)

where g(t)

=

2itρ[

a

k=1

+

a

xk log xk −

b

yj log yj ]

j=1

log Γ[ρxk (1 − 2it) + βk + ζk ]

k=1

−

b

log Γ[ρyj (1 − 2it) + j + ηj ],

j=1

with g(0) = − log K =

a

log Γ[ρxk + βk + ζk ] −

b

log Γ[ρyj + j + ηj ].

j=1

k=1

We use the asymptotic expansion in z as |z| → ∞ [Erd´elyi et al. (1953), p. 48] for bounded h, √ log Γ(z + h) = log 2π + (z + h − 12 ) log z − z (12.4) −

l

(−1)α

α=1

Bα+1 (h) −α z + O(z −(l+1) ), | arg z| < π. α(α + 1)

The terms Br (h) are the Bernoulli polynomials deﬁned to be the coeﬃcients in the Taylor series ∞

zr zehz = B (h) , |z| < 2π. r ez − 1 r=0 r! The reader can verify the ﬁrst few Bernoulli polynomials B0 (h)

=

1,

12.2. General expansions

B1 (h) B2 (h) B3 (h)

197

= h − 12 ,

= h2 − h + 16 ,

= h3 − 32 h2 + 12 h,

B4 (h)

= h4 − 2h3 + h2 −

B5 (h)

= h5 −

B6 (h)

= h6 −

1 30 , 5 4 5 3 1 2 h + 3 h − 6 h, 3h5 + 52 h4 − 12 h2

+

1 42 .

Bernoulli polynomials can be generated at will with modern symbolic computations software such as the function bernoulli(r,h); in Maple [Redfern (1996)] or BernoulliB[h,r] in Mathematica [Wolfram (1996)]. Let (z, h) = (ρxk (1 − 2it), βk + ζk ), (ρyj (1 − 2it), j + ηj ), (ρxk , βk + ζk ), and (ρyj , j + ηj ) in turn in (12.4). We assume that xk and yj are terms behaving as O(n), where n is the sample size. This will have to be checked in each application. When ρ = 1, then βk = j = 0, and h is bounded in all cases. Later, ρ will be allowed to depend on the sample size n and we will need to check that βk and j are bounded. Then substitute the four expansions for log Γ(z + h) in g(t) and g(0) of (12.3) to obtain, after long but straightforward simpliﬁcations, KρM (t) = − 12 f log(1 − 2it) +

l

ωα [(1 − 2it)−α − 1] + O(n−(l+1) ), (12.5)

α=1

where

f = −2

a

ζk −

b

ηj − 12 (a − b) ,

(12.6)

j=1

k=1

a b (−1)α+1 Bα+1 (βk + ζk ) Bα+1 (j + ηj ) − . ωα = α(α + 1) (ρxk )α (ρyj )α j=1

(12.7)

k=1

Note that ωα = O(n−α ) if xk and yj are O(n), and βk and j are O(1). The next step consists of deriving the characteristic function cρM by exponentiation of KρM and then using the inversion formula (2.2) to derive the p.d.f.: cρM (t)

= eKρM (t) =

(1 − 2it)−f /2 ·[1 + O(n

=

exp[ωα (1 − 2it)−α ]

α=1 −(l+1)

−f /2

(1 − 2it)

l

l

exp(−ωα )

α=1

)]

l ∞ α=1 k=0

l ∞ ωαk ωk −αk (1 − 2it) (−1)k α k! k! α=1 k=0

198

12. Asymptotic expansions

·[1 + O(n−(l+1) )]. The expansion of order O(n−(l+1) ) is then obtained by keeping and collecting terms of order up to O(n−l ). Let us illustrate the procedure for an expansion of order O(n−2 ) and leave the higher-order expansion to the symbolic calculators Mathematica [Wolfram (1996)] or Maple [Redfern (1996)]. For the expansion of order O(n−2 ), set l = 1 and note that cρM (t)

=

(1 − 2it)−f /2 [1 + ω1 (1 − 2it)−1 ](1 − ω1 ) + O(n−2 )

= (1 − 2it)−f /2 {1 + ω1 [(1 − 2it)−1 − 1]} + O(n−2 ) = cf (t) + ω1 [cf +2 (t) − cf (t)] + O(n−2 ), (12.8) where cf (t) = (1 − 2it)−f /2 denotes the characteristic function of χ2f on f degrees of freedom. Then, by the inversion formula (2.2), ∞ 1 fρM (s) = cρM (t)e−its dt 2π −∞ = gf (s) + ω1 [gf +2 (s) − gf (s)] + O(n−2 ),

(12.9)

where gf (t) is the p.d.f. of χ2f . Finally, by integration on (−∞, x], the d.f. takes the form P (ρM ≤ x) = Gf (x) + ω1 [Gf +2 (x) − Gf (x)] + O(n−2 ), where Gf (t) is the d.f. of χ2f . A full justiﬁcation of the last two integrations would require one to show that the remainders in (12.8) and (12.9) are O(n−2 ) uniformly in t and s, respectively; v. Box (1949) for details. The whole purpose of introducing ρ in the expansion is to reduce the number of terms. In the above example, one can choose ρ to annihilate the term of order O(n−1 ), i.e., to make ω1 = 0. Recalling (12.2) and B2 (h) = h2 − h + 16 , we have a b 1 −1 −1 ω1 = xk B2 (βk + ζk ) − yj B2 (j + ηj ) 2ρ j=1 k=1 a b 1 2 1 −(1 − ρ)f + = x−1 yj−1 (ηj2 − ηj + 16 ) . k (ζk − ζk + 6 ) − 2ρ j=1 k=1

Thus, ω1 vanishes by choosing a b 2 1 x−1 yj−1 (ηj2 − ηj + 16 ) . ρ = 1 − f −1 k (ζk − ζk + 6 ) − k=1

(12.10)

j=1

Even though ρ now depends on xk and yj , assumed to be of order O(n), the asymptotic expansion is still valid since for this choice of ρ, βk = (1 − ρ)xk and j = (1 − ρ)yj are terms of order O(1) and, thus, ωα is still O(n−α ).

12.2. General expansions

199

Proposition 12.1 If W has moments (12.1), where xk and yj are terms O(n), then with the choice of ρ in (12.10), P (ρM ≤ x) = Gf (x) + O(n−2 ).

(12.11)

The asymptotic expansions of order O(n−2 ) were proposed by Bartlett (1938). The factor ρ which annihilates the term ω1 of order O(n−1 ) is referred to as the Bartlett correction factor. We now give the general result for an expansion of order O(n−6 ) (l = 5) calculated with the help of Maple [Redfern (1996)]: P (ρM ≤ x) = Gf (x) + ω1 [Gf +2 (x) − Gf (x)] + ω2 [Gf +4 (x) − Gf (x)] + 12 ω12 [Gf +4 (x) − 2Gf +2 (x) + Gf (x)] + ω3 [Gf +6 (x) − Gf (x)] +ω2 ω1 [Gf +6 (x) − Gf +4 (x) − Gf +2 (x) + Gf (x)] + 16 ω13 [Gf +6 (x) − 3Gf +4 (x) + 3Gf +2 (x) − Gf (x)] +ω4 [Gf +8 (x) − Gf (x)] + ω3 ω1 [Gf +8 (x) − Gf +6 (x) − Gf +2 (x) + Gf (x)] + 12 ω2 ω12 [Gf +8 (x) − 2Gf +6 (x) + 2Gf +2 (x) − Gf (x)]

1 4 + 24 ω1 [Gf +8 (x) − 4Gf +6 (x) + 6Gf +4 (x) − 4Gf +2 (x)] + Gf (x)]

+ 12 ω22 [Gf +8 (x) − 2Gf +4 (x) + Gf (x)] + ω5 [Gf +10 (x) − Gf (x)] +ω4 ω1 [Gf +10 (x) − Gf +8 (x) − Gf +2 (x) + Gf (x)]

+ω3 ω2 [Gf +10 (x) − Gf +6 (x) − Gf +4 (x) + Gf (x)] + 12 ω3 ω12 [Gf +10 (x) − 2Gf +8 (x) + Gf +6 (x) −Gf +4 (x) + 2Gf +2 (x) − Gf (x)] 1 + 6 ω2 ω13 [Gf +10 (x) − 3Gf +8 (x) + 2Gf +6 (x)

+2Gf +4 (x) − 3Gf +2 (x) + Gf (x)] 1 + 120 ω15 [Gf +10 (x) − 5Gf +8 (x) + 10Gf +6 (x) −10Gf +4 (x) + 5Gf +2 (x) − Gf (x)] 1 2 + 2 ω2 ω1 [Gf +10 (x) − Gf +8 (x) − 2Gf +6 (x) +2Gf +4 (x) + Gf +2 (x) − Gf (x)]

+O(n−6 ). When ρ is chosen as in (12.10) so that ω1 = 0, then things reduce considerably. Proposition 12.2 If W has moments (12.1) where xk and yj are terms O(n), then with the choice of ρ in (12.10),

P (ρM ≤ x) = Gf (x) + ω2 [Gf +4 (x) − Gf (x)] + ω3 [Gf +6 (x) − Gf (x)] +ω4 [Gf +8 (x) − Gf (x)] + 12 ω22 [Gf +8 (x) − 2Gf +4 (x) + Gf (x)] +ω5 [Gf +10 (x) − Gf (x)] + ω3 ω2 [Gf +10 (x) − Gf +6 (x) − Gf +4 (x) + Gf (x)] +O(n−6 ).

200

12. Asymptotic expansions

Automatic correction of coverage probability of conﬁdence intervals [Martin (1990) or of error in rejection probability of tests [Beran (1987, 1988)] is now made possible by the resampling or “bootstrap” technology (v. Chapter 14). Bootstrap of a Bartlett corrected likelihood ratio test reduces level error in Proposition 12.1 from O(n−2 ) to O(n−3 ) automatically without further analytical expansions.

12.3 Examples We now present some examples. Example 12.1 Test of sphericity. The likelihood ratio test (LRT) of sphericity was derived in Proposition 8.11 and its moments are given in Proposition 8.12. Now, it is simply a matter of rewriting things in the form (12.1) to obtain the asymptotic expansion. Hence, for W = Λm/n , m = n − 1, E Wh

Γp [ 12 m + 12 mh] Γ[ 12 mp + 12 pmh] p 1 1 1 k=1 Γ[ 2 m + 2 mh − 2 (k − 1)] pmh/2 = Kp Γ[ 12 mp + 12 pmh] p Γ[ 1 m(1 + h) − 12 (k − 1)] = Kppmh/2 k=1 2 1 , Γ[ 2 mp(1 + h)] = Kppmh/2

so that we have the form (12.1) with a = p,

xk = 12 m,

ζk = − 12 (k − 1),

b = 1,

y1 = 12 mp,

η1 = 0.

p

Observe that k=1 xk = y1 is satisﬁed and xk and y1 are terms behaving as O(n). The asymptotic expansion with remainder O(n−6 ) as in Proposition 12.2 is now a simple matter of calculating with (12.6) and (12.10), f

=

1 2 (p

ρ

=

1−

+ 2)(p − 1), 2p2 + p + 2 , 6pm

βk and 1 in (12.2), and, ﬁnally, ωα , α = 2, 3, 4, 5, in (12.7). Of course, one could go to great lengths to obtain the most simpliﬁed algebraic expressions in terms of p and n. For example, Davis (1971) using properties of Bernoulli polynomials showed ωα

=

2(−1)α α(α + 1)(α + 2)ρα

12.3. Examples

s

Bs

0

1

1

− 12

2

1 6

3

0

4

1 − 30

5

0

6

1 42

7

0

201

δs − 12 p 1 4 p(p

+ 1)

1 − 16 p(2p2 + 3p − 1) 1 16 p(p

− 1)(p + 1)(p + 2)

1 − 192 p(6p4 + 15p3 − 10p2 − 30p + 3) 1 128 p(p

− 1)(p + 1)(p + 2)(2p2 + 2p − 7)

1 − 768 p(6p6 + 21p5 − 21p4 − 105p3 + 21p2 + 147p − 5) 1 768 p(p

− 1)(p + 1)(p + 2)(3p4 + 6p3 − 23p2 − 26p + 62)

Table 12.1. Polynomials δs and Bernoulli numbers Bs for asymptotic expansions.

·

α+1 s=1

α+2 Bs (1 − ρ)α+1−s δs + 12 (s + 1) s−1 ( 12 m)1−s , s+1 p

where Bs ≡ Bs (0) are the Bernoulli numbers and the δs are certain polynomials in p deﬁned by Box (1949) (v. Table 12.1). Example 12.2 Asymptotics for U (p; m, n) distributions. The LRT for the general linear hypothesis in multivariate regression was described in Proposition 9.3 as Λ2/n ∼ U (p; r, n − k). Another example is the LRT for independence between two subvectors in Proposition 11.2 where Λ2/n ∼ U (p2 ; p1 , n − 1 − p1 ). Thus, we derive the asymptotic expansion for W ∼ [U (p; m, n − c)]n/2 , where n − c ≥ p, which includes both cases. Now, the moments of U distributions were given in Problem 11.6.1. Hence, E Wh

= E [U (p; m, n − c)]nh/2 p Γ[ 1 (n − c − p + k) + 12 nh] = K p k=1 1 2 1 j=1 Γ[ 2 (m + n − c − p + j) + 2 nh] p Γ[ 1 n(1 + h) + 12 (−c − p + k)] = K p k=1 1 2 , 1 j=1 Γ[ 2 n(1 + h) + 2 (m − c − p + j)]

which is of the form (12.1) with

202

12. Asymptotic expansions

a = p,

xk = 12 n,

ζk = 12 (−c − p + k),

ηj = 21 (m − c − p + j). p p Note, again, that xk and yj are O(n) and k=1 xk = j=1 yj is satisﬁed. Using (12.6) and (12.10), we have b = p,

f ρ

yj = 12 n,

= pm, = 1 − n−1 [c − 12 (m − p − 1)].

An interesting peculiarity in this case which derives from a symmetry property of Bernoulli polynomials [Erd´elyi et al. (1953), p. 37], namely Bα (1 − h) = (−1)α Bα (h), is that ω2α−1 = 0, α = 1, 2, . . . , which means the series involves only terms of even powers of n−1 [Lee (1972)]. To see this, ﬁrst note βk = (1 − ρ)xk βk + ζk

= =

1 1 2 [c − 2 (m − p − 1)], − 12 [ 12 (m + p − 1) − k],

and, similarly, k + ηk = 12 [ 12 (m − p + 1) + k]. Therefore, we ﬁnd (note that k → p − k + 1 reverses the order of terms in the following sums) p

B2α (βk + ζk )

=

k=1

= = =

p k=1 p k=1 p k=1 p

B2α − 12 [ 12 (m + p − 1) − k]

B2α − 12 [ 12 (m + p − 1) − (p − k + 1)]

B2α 1 + 12 [ 12 (m + p − 1) − (p − k + 1)] B2α (k + ηk ),

k=1

and from (12.7), ω2α−1 = 0, α = 1, 2, . . .. Thus, the expansion in Proposition 12.2 further reduces to P (ρM ≤ x) = Gf (x) + ω2 [Gf +4 (x) − Gf (x)] + ω4 [Gf +8 (x) − Gf (x)] + 12 ω22 [Gf +8 (x) − 2Gf +4 (x) + Gf (x)] + O(n−6 ),

(12.12)

where from the same symmetry property of Bernoulli polynomials, one can easily establish that ω2α =

p

22α B2α+1 12 [ 12 (m − p + 1) + k] , α = 1, 2, . . . . 2α α(2α + 1)(ρn) k=1

12.3. Examples

O(n−2 ) O(n−4 ) O(n−6 ) exact

n=2

n=5

n = 10

n = 15

n = 20

n = 30

.9500 .8107 .7168 .4714

.9500 .8848 .8642 .8315

.9500 .9220 .9182 .9139

.9500 .9345 .9334 .9322

.9500 .9402 .9397 .9393

.9500 .9451 .9449 .9448

203

Table 12.2. Asymptotic expansions for U (2; 12, n) distributions.

A small-scale numerical evaluation of (12.12) would help to determine how large n should be for the asymptotics of U (p; m, n) distributions to be accurate. Fix p = 2 and m = 12, and vary n =2, 5, 10, 15, 20, and 30. The asymptotic distribution of −n log U (2; 12, n) is χ224 . So, we choose the critical point x = χ2.95,24 = 36.41502. The evaluation of P − n + 12 (12 − 2 − 1) log U (2; 12, n) ≤ 36.41502 using (12.12) led to Table 12.2. The exact values were obtained for p = 2 with the transformation in Problem 11.6.2: P (−ρn log U (2; 12, n) ≤ x) = P U (2; 12, n) ≥ e−x/ρn (n − 1)(1 − y 1/2 ) = P F (24, 2(n − 1)) ≤ , 12y 1/2 with y = e−x/ρn . The approximations of order O(n−6 ) can thus be used in practice for n as small as 10 in this case. They are nearly exact to four decimal places for n = 30. Example 12.3 Test of mutual independence between subvectors. This is a continuation of Problem 11.6.7, where in item (ii), we found the ˜ = Λ2/n , moments of Λ r 1 Γpj ( 12 m) ˜ h = Γp ( 2 m + h) E Λ , Γp ( 12 m) j=1 Γpj ( 12 m + h)

with m = n − 1. This can be written in the form of (12.1) for W = Λ as p Γ[ 1 n(1 + h) − 12 k] h nh/2 ˜ pj 2 1 = K r k=1 E W =E Λ 1 , j=1 l=1 Γ[ 2 n(1 + h) − 2 l] with the identiﬁcation a = p, b=

r j=1

pj = p,

xk = 12 n,

ζk = − 12 k, k = 1, . . . , a,

yjl = 12 n,

ηjl = − 12 l, j = 1, . . . , r, l = 1, . . . , pj .

204

12. Asymptotic expansions

The constants f and ρ can be veriﬁed with (12.6) and (12.10): pj p r − 12 k − − 12 l = 12 Σ2 , f = −2 j=1 l=1

k=1

1−

ρ =

2Σ3 + 9Σ2 , 6nΣ2

r where Σs ≡ ps − j=1 psj . For simpliﬁed algebraic expressions of ω2 through ω6 , the reader is referred to Box (1949). Example 12.4 Test of equality of variances. The null moments of the modiﬁed likelihood ratio test Λ∗ for the hypothesis H0 : Σ1 = · · · = Σa were obtained in Proposition 8.17. Thus, for W = Λ∗ , we can write a Γp ( 12 m) Γp [ 12 mi (1 + h)] pmi h/2 Γ [ 1 m(1 + h)] Γp ( 21 mi ) p 2 i=1 i=1 mi h a p a 1 m/2 1 1 i=1 ( 2 m) i=1 l=1 Γ[ 2 mi (1 + h) − 2 (l − 1)] = K a p 1 p 1 1 mi /2 i=1 l=1 ( 2 mi ) j=1 Γ[ 2 m(1 + h) − 2 (j − 1)]

E Wh

a

=

mpmh/2

which is of the form (12.1) with the identiﬁcation a = pa,

xkl = 12 mk ,

ζkl = − 12 (l − 1), k = 1, . . . , a, l = 1, . . . , p,

b = p,

yj = 12 m,

ηj = − 12 (j − 1), j = 1, . . . , p.

The degrees of freedom f and ρ in (12.6) and (12.10) are p p a ζkl − ηj − 1 (pa − p) f = −2 2

k=1 l=1

= ρ

=

j=1

1 2 p(p

+ 1)(a − 1), p p a 2 1 x−1 yj−1 (ηj2 − ηj + 16 ) 1 − f −1 kl (ζkl − ζkl + 6 ) − k=1 l=1

=

1−

(2p2 + 3p − 1) 6(p + 1)(a − 1)

a

k=1

1 1 − mk m

j=1

.

Values of ωα can be calculated from (12.7) in simpliﬁed algebraic form but this is unnecessary since they can be easily programmed for the computer to evaluate the expansion in Proposition 12.2. Note ﬁnally that since we require (1 − ρ)xkl and (1 − ρ)yj to remain bounded, the expansion is asymptotic as m → ∞ while mk /m → αk for some proportions 0 < αk < 1 such a that k=1 αk = 1.

12.4. Problem

205

The basic idea in asymptotic expansions was to represent the cumulant generating function in the form (12.5). This is often possible even though the moments may not be of the form (12.1). Example 12.5 An example is provided by the modiﬁed likelihood ratio test for a given variance in Problem 8.9.8, where mph/2 Γp [ 12 m(1 + h)] 2e E Λ∗ h = (1 + h)−mp(1+h)/2 , m = n − 1. m Γp ( 12 m) For W = Λ∗ , Davis (1971) showed that Proposition 12.2 holds with f

=

1 2 p(p

ρ

=

1−

ωα

=

α+1 α + 2 2(−1)α (1 − ρ)α+1−s δs ( 12 m)1−s , (ω1 = 0). α(α + 1)(α + 2)ρα s=1 s + 1

+ 1),

2p2 + 3p − 1 , 6m(p + 1)

The δs are the same as those of Table 12.1. For asymptotic expansions of the null distribution of Lawley-Hotelling and Pillai trace tests, the reader is referred to Muirhead (1970) and Fujikoshi (1970).

12.4 Problem 1. This problem develops the asymptotic expansion of the LRT for the equality of means and variances H0 : µ1 = · · · = µa ; Σ1 = · · · = Σa between a multivariate normal populations. The LRT Λ, together with its moments, are given in Problem 8.9.14. Using the same notation and W = Λ, establish the following:

(i) The moments of W have the equivalent form h p a p 1 n/2 Γ[ 1 nl (1 + h) − 12 k] j=1 ( 2 n) k=1 h p l=1 1 2 E W = K p a 1 n /2 1 l k=1 l=1 ( 2 nl ) j=1 Γ[ 2 n(1 + h) − 2 j] (ii) Perform the usual identiﬁcation to conclude the validity of Proposition 12.2 with f

=

ρ =

1 2 (a

− 1)p(p + 3),

(2p2 + 9p + 11) 1− 6(a − 1)(p + 3)

a 1 1 . − n n i=1 i

13 Robustness

13.1 Introduction Many inference methods were presented in previous chapters for multivariate normal populations. A question of theoretical and utmost practical importance is the eﬀect of non-normality on the inference. For example, what happens if the likelihood ratio test of sphericity, derived assuming normality, is performed, but, in fact, the population follows a multivariate student distribution on 10 degrees of freedom? Is the signiﬁcance level of α = 5%, say, still close to 5%? The theory of robustness gives answers as to how sensitive multivariate normal inferences are to departures from normality. Most importantly, it proposes some remedies, i.e., more robust procedures. In Section 13.2, we present some non-normal models often used in robustness, the so-called elliptical distributions. The rest of the chapter is devoted to robust estimation and adjusted likelihood ratio tests. A robust analysis of data is useful in several ways. It can validate or rebuﬀ data analysis done on classical assumptions of multivariate normality. It also comes into play in the identiﬁcation of outliers, which is a challenging task for data sets with more than two variables. Robust estimates of location vector and scale matrix serve this role admirably. They can be used to evaluate robust Mahalanobis distances from an observation vector xi to the location vector. Points with large Mahalanobis distances can then be singled out and scrutinized.

13.2. Elliptical distributions

207

13.2 Elliptical distributions Suppose that x ∈ Rp has a density fx (x) = |Λ|−1/2 g[(x − µ) Λ−1 (x − µ)], where g : [0, ∞) → [0, ∞) is a ﬁxed function independent of µ and Λ = (Λij ) and depends on x only through (x − µ) Λ−1 (x − µ). Denote this elliptical distribution by x ∼ Ep (µ, Λ). The main reference for elliptical distributions is Kelker (1970). The aﬃne linear transformation y = Bx + b with B ∈ Gp and b ∈ Rp has density fy (y) = |BΛB |−1/2 g[(y − Bµ − b) (BΛB )−1 (y − Bµ − b)]. Thus, y ∼ Ep (Bµ + b, BΛB ); i.e., the transformation x → Bx + b induces the parameter transformation µ → Bµ + b and Λ → BΛB . In particular, z = Λ−1/2 (x − µ) ∼ Ep (0, I) has a spherical or rotationally invariant distribution. Elliptical distributions are a location scale generalization of spherical distributions. Thus, for example, if z ∼ Ep (0, I) with characteristic function necessarily of the form cz (t) = φ(t t) (v. Problem 4.6.6), then x = Λ1/2 z + µ ∼ Ep (µ, Λ) has characteristic function cx (t) = exp(it µ)φ(t Λt). Moreover, if z has a ﬁnite second moment, E z = 0 and var z = αI, for some constant α, implies E x = µ and var x ≡ Σ = αΛ. An important implication is that all elliptical distributions with ﬁnite second moments have the same correlation matrix. The constant α = −2φ (0) (v. Problem 4.6.15) is easily found by diﬀerentiation of cz (t). Examples of spherical distributions commonly used in robustness are members of the normal mixture family with density ∞ (2πw)−p/2 exp(− 12 w−1 x x)dF (w), fx (x) = 0

where F (·) is the “mixing” distribution function on [0, ∞). These can be d

|=

simulated easily using the representation x = w1/2 z, where w ∼ F , z ∼ z (v. Problem 13.6.1). Np (0, I), w Example 13.1 Obviously, P (w distribution.

= σ 2 ) = 1 yields the Np (0, σ 2 I)

Example 13.2 The two-point distribution, P (w = 1) = 1 − , P (w = σ 2 ) = for some “contamination” proportion 0 < < 1, yields the symmetric contaminated normal distribution. Example 13.3 The multivariate t on ν degrees of freedom denoted tp,ν is obtained with νw−1 ∼ χ2ν . The reader is asked to show in Problem 13.6.1

208

13. Robustness

that x has density fx (x) = cp,ν (1 + x x/ν)−(ν+p)/2 , x ∈ Rp , where cp,ν = (νπ)−p/2 Γ 12 (ν + p) /Γ 12 ν . The general multivariate tp,ν (µ, Λ) is obtained by relocating and rescaling, y = Λ1/2 x + µ, and has density −(ν+p)/2 fy (y) = cp,ν |Λ|−1/2 1 + (y − µ) Λ−1 (y − µ)/ν , y ∈ Rp . The multivariate t on 1 degree of freedom is also known as the multivariate Cauchy distribution. The Kotz-type distributions form another important class of elliptical distributions [Fang et al. (1991), p. 76]. Their characteristic function was obtained recently by Kotz and Ostrovskii (1994). Elliptical distributions that can be expanded as a power series are deﬁned in Steyn (1993) and used to deﬁne other nonelliptical distributions with heterogeneous kurtosis. The following result gives the marginal and conditional distributions for an Ep (µ, Λ) distribution. Let x = (x1 , x2 ) with xi ∈ Rpi , i = 1, 2, p = p1 + p2 , and partition µ and Λ in conformity as (µ , µ ) , 1 2 Λ11 Λ12 . Λ = Λ21 Λ22 µ =

Proposition 13.1 The marginal and conditional distributions of an Ep (µ, Λ) distribution are elliptical: (i) x2 ∼ Ep2 (µ2 , Λ22 ), (ii) x1 |x2 ∼ Ep1 (µ1.2 , Λ11.2 ), where µ1.2 Λ11.2

= µ1 + Λ12 Λ−1 22 (x2 − µ2 ), = Λ11 − Λ12 Λ−1 22 Λ21 .

The conditional variance is of the form var(x1 |x2 ) = w(x2 )Λ11.2 , for some function w(x2 ) ∈ R which depends on x2 only through the quadratic form (x2 − µ2 ) Λ−1 22 (x2 − µ2 ). Proof. Letting t = (0 , t2 ) in cx (t) = exp(it µ)φ(t Λt), we ﬁnd cx2 (t2 ) = exp(it2 µ2 )φ(t2 Λ22 t2 ) and, thus, x2 ∼ Ep2 (µ2 , Λ22 ). For the conditional distribution, let z = x1 − [µ1 + Λ12 Λ−1 22 (x2 − µ2 )] with jacobian J(x → z, x2 ) = 1. Upon using Problem 1.8.2, the conditional density z|x2 is −1 |Λ|−1/2 g[z Λ−1 11.2 z + (x2 − µ2 ) Λ22 (x2 − µ2 )] , |Λ22 |−1/2 fx2 (x2 )

13.2. Elliptical distributions

209

where fx2 (x2 ) depends only on (x2 − µ2 ) Λ−1 22 (x2 − µ2 ). Thus, we have z|x2 ∼ Ep1 (0, Λ11.2 ) and, in turn, x1 |x2 ∼ Ep1 (µ1.2 , Λ11.2 ). 2 Example 13.4 The univariate power exponential distribution has p.d.f. 2α −1/2 1 x − µ exp − 2 1/2 , α > 0. (13.1) fx (x) = c1,α Λ Λ A multivariate extension seems to be ( α ) . fx (x) = cp,α |Λ|−1/2 exp − 12 (x − µ) Λ−1 (x − µ)

(13.2)

This elliptical distribution has an advantage of generating distributions with heavier and lighter tails than the multivariate normal by taking α < 1 or α > 1, whereas many other elliptical distributions including the multivariate t cannot generate lighter-tail distributions. Kuwana and Kariya (1991) used this property to derive a locally best invariant test of multivariate normality (α = 1). Taking α = 0.5 simply, in (13.2), E (x1 − µ1 )2 = 4(p + 1)Λ11 , which depends on p (v. Problem 13.6.5); the corresponding moment in (13.1) is E (x − µ)2 = 8Λ11 , with Λ = Λ11 . So, the marginal distribution of x1 in (13.2) is not that of x in (13.1). The inconsistency takes place for many other elliptical distributions. Kano (1994) characterized the consistency property of elliptical distributions: An elliptical family is consistent iﬀ it is a normal mixture family. In particular, the multivariate normal and multivariate t families are consistent. In Proposition 13.1 the marginal is elliptical but possibly of a diﬀerent functional form since the characteristic function φ may be related to p. For the estimation of (µ, Λ), it seems natural to ask that location and scatter estimates transform in exactly the same manner as the parameters; i.e., that they be “aﬃne equivariant” as described in the following deﬁnition. Formally, let x1 .. X = . ∈ Rnp xn

be the sample matrix. ˆ ˆ Deﬁnition 13.1 The location and scatter estimates µ(X) and Λ(X) are aﬃne equivariant iﬀ for all B ∈ Gp and b ∈ Rp , ˆ + 1b ) µ(XB ˆ + 1b ) Λ(XB

ˆ = Bµ(X) + b, ˆ = BΛ(X)B .

210

13. Robustness

When the underlying distribution belongs to an elliptical family, the distribution of aﬃne equivariant estimates has a special structure. In particular, the general form of the mean and variance estimates can be characterized for ﬁnite samples. To establish this general form, we need to extend the notion of rotational invariance of random vectors in Section 4.4 to symmetric random matrices. Deﬁnition 13.2 A random symmetric matrix W is rotationally invariant d iﬀ W = HWH , ∀H ∈ Op . The following lemma [Tyler (1982)] characterizes the general form of the mean and variance of any rotationally invariant random matrix. Proposition 13.2 Let W ∈ Rpp symmetric be rotationally invariant with ﬁnite second moments. Then, there exist constants η, σ1 ≥ 0, and σ2 ≥ −2σ1 /p such that E W var W

= ηI, = σ1 (I + Kp ) + σ2 vec(I)[vec(I)] .

Proof. For the mean, let E W ≡ A. By rotational invariance, A = HAH , ∀H ∈ Op . Hence, x Ax = x HAH x = y Ay, ∀x, y ∈ Rp , |x| = |y| = 1. Choosing x = hi and y = hj , the ith and jth eigenvectors of A corresponding to eigenvalues λi and λj , respectively, we get λi = λj ≡ η (say). This means A = ηI. For the variance, let Ω ≡ var W = Ωijkl ei ej ⊗ ek el , where cov(wki , wlj ) = Ωijkl . Note that {ei ej ⊗ ek el , i, j, k, l = 1, . . . , p} 2

d

d

forms a basis for Rpp2 . Since W = HWH , ∀H ∈ Op , then vec(W) = (H ⊗ H)vec(W) and, thus, Ω = (H ⊗ H)Ω(H ⊗ H ), or Ωijkl ei ej ⊗ ek el , Ωijkl hi hj ⊗ hk hl = where H = (h1 , . . . , hp ). By choosing for some m, hm = −em and hr = er , r = m, we obtain Ωijkl = 0 unless i = j = k = l, i = j and k = l, i = k and j = l, or i = l and j = k. By choosing H to give a permutation of the rows, we obtain Ωiiii = σ0 , ∀i = 1, . . . , p, Ωiikk = σ1 for i = k, Ωijij = σ2 for i = j, and Ωijji = σ3 for i = j. Thus, ei ei ⊗ ei ei + σ1 ei ei ⊗ ek ek Ω = σ0 i

+σ2

i =j

i =k

ei ej ⊗ ei ej + σ3

i =j

ei ej ⊗ ej ei

13.2. Elliptical distributions

211

= σ1 I + σ2 vec(I)[vec(I)] + σ3 Kp +(σ0 − σ1 − σ2 − σ3 ) ei ei ⊗ e i ei . i

Since ∀H ∈ Op , (H ⊗ H)I(H ⊗ H ) = I, (H ⊗ H)vec(I)[vec(I)] (H ⊗ H ) = vec(I)[vec(I)] , (H ⊗ H)Kp (H ⊗ H ) = Kp , and (H ⊗ H)

ei ei

⊗

ei ei

(H ⊗ H ) =

i

ei ei

⊗

ei ei

,

i

for some H ∈ Op , it follows that σ0 − σ1 − σ2 − σ3 = 0. Also, since W is symmetric, cov(wij , wji ) = var wij , which implies σ1 = σ3 . Therefore, Ω = σ1 (I + Kp ) + σ2 vec(I)[vec(I)] . The conditions on σ1 and σ2 follow since Ω is positive semideﬁnite.

2

The variance of W = (wij ) can be written componentwise with the Kronecker delta cov(wki , wlj ) = σ1 (δij δkl + δkj δil ) + σ2 δki δlj . The form of var W states that the oﬀ-diagonal elements of W are uncorrelated with each other and uncorrelated with the diagonal elements. Each oﬀ-diagonal element has variance σ1 . The diagonal elements all have variance 2σ1 + σ2 with the covariance between any two diagonal elements being σ2 . Example 13.5 A simple example is W ∼ Wp (m) which is rotationally invariant with var W = m(I + Kp ). Example 13.6 Assume x ∼ Ep (0, Λ) and let W = xx . Then var W = (Λ1/2 ⊗ Λ1/2 )var(zz )(Λ1/2 ⊗ Λ1/2 ), where z ∼ Ep (0, I). Using Proposition 13.2 var(zz ) is evaluated with σ1 σ2

= =

var(z1 z2 ) = E(z12 z22 ) = µ22 ,

cov(z12 , z22 ) = E(z12 z22 ) − E(z12 )E(z22 ) = µ22 − µ22 .

In terms of cumulants we have σ1 = k22 + k22 and σ2 = k22 . These cumulants are easily found with the Taylor series ln φ(t21 + t22 )

(it1 )2 (it2 )2 (it1 )4 (it2 )4 + k2 + k4 + k4 2! 2! 4! 4! (it1 )2 (it2 )2 + o(|t|4 ). +k22 2! 2!

= k2

212

13. Robustness

The reader can verify by diﬀerentiation (v. Problem 13.6.3) k2 k4 k22

= −2φ (0) = α, = =

12(φ (0) − φ (0)2 ), 4(φ (0) − φ (0)2 ).

The kurtosis of z1 is k4 (φ (0) − φ (0)2 ) = 3 ≡ 3k, k22 φ (0)2 where k represents a kurtosis parameter. Thus, k4 = 3kα2 and k22 = kα2 . Finally, we obtain σ1 = (1 + k)α2 and σ2 = kα2 from which var(zz ) = α2 (1 + k)(I + Kp ) + α2 k vec(I)[vec(I)] and var W = (1 + k)(I + Kp )(Σ ⊗ Σ) + k vec(Σ)[vec(Σ)] , where Σ = αΛ is the variance of x. ˆ ˆ Corollary 13.1 If µ(X) and Λ(X) are aﬃne equivariant with ﬁnite second moment and x1 , . . . , xn are i.i.d. Ep (µ, Λ), then there exist constants η, β ≥ 0, σ1 ≥ 0 and σ2 ≥ −2σ1 /p such that ˆ E µ(X) ˆ var µ(X) ˆ E Λ(X)

= µ,

ˆ var Λ(X)

= σ1 (I + Kp )(Λ ⊗ Λ) + σ2 vec(Λ)[vec(Λ)] .

= βΛ, = ηΛ,

d

Proof. First, X = ZΛ1/2 + 1µ , where z1 .. Z= . zn

d

1/2 ˆ ˆ ˆ and zi ’s are i.i.d. Ep (0, I). Hence, µ(X) = µ(ZΛ +1µ ) = Λ1/2 µ(Z)+µ. ˆ Obviously, µ(Z) is a rotationally invariant random vector. Using the result ˆ ˆ of Section 4.5, E µ(Z) = 0 and var µ(Z) = βI, for some β ≥ 0. Therefore, d 1/2 ˆ ˆ ˆ ˆ = Λ1/2 Λ(Z)Λ , where E µ(X) = µ and var µ(X) = βΛ. Similarly, Λ(X) ˆ Λ(Z) is a rotationally invariant matrix whose mean and variance have the ˆ general form in Proposition 13.2. Hence, E Λ(X) = ηΛ, for some η, and ! ˆ ˆ var Λ(X) = var (Λ1/2 ⊗ Λ1/2 )vec(Λ(Z)) ! ˆ = (Λ1/2 ⊗ Λ1/2 ) var Λ(Z) (Λ1/2 ⊗ Λ1/2 )

= σ1 (I + Kp )(Λ ⊗ Λ) + σ2 vec(Λ)[vec(Λ)]

13.3. Maximum likelihood estimates

for some σ1 ≥ 0, σ2 ≥ −2σ1 /p.

213

2

Complicated expressions using tensor methods for third-order and fourth-order cumulants of aﬃne equivariant estimates in elliptical families were obtained by Gr¨ ubel and Rocke (1990). ˆ Another way of writing var Λ(X) is to give the covariances between any ˆ ˆ ij ): two elements of Λ(X) = (Λ ˆ ki , Λ ˆ lj ) = σ1 (Λij Λkl + Λkj Λil ) + σ2 Λki Λlj . cov(Λ One should note that a reasonable estimate of Λ assumed positive deﬁnite ˆ should satisfy Λ(X) > 0 w.p.1, and in that case, η > 0.

13.3 Maximum likelihood estimates Assume x1 , . . . , xn i.i.d. x ∼ Ep (µ, Λ) with var x = αΛ = Σ. The simplest but ineﬃcient method to estimate (µ, Λ) would be to use the MLE under a ˆ = 1 n (xi − x ¯ and Σ ¯ )(xi − x ¯ ) . A more eﬃcient Np (µ, Σ) distribution, x i=1 n procedure would be the MLE under the “true” Ep (µ, Λ) model. These two possibilities are now investigated.

13.3.1

Normal MLE

When x has ﬁnite fourth-order moments, the general discussion of Section 6.3 showed that ˆ −Σ var W 0 d Σ p+1 n1/2 → N 0, , p 0 Σ ¯ − µ x where W = xx . From the calculation of var W in Example 13.6, it follows that

where n

|=

d ˆ − Σ, x ¯ − µ) → (N, n), n1/2 (Σ

N,

n ∼ Np (0, Σ) N ∼ Npp (0, (1 + k)(I + Kp )(Σ ⊗ Σ) + k vec(Σ)[vec(Σ)] ).

13.3.2 Elliptical MLE For x ∼ Ep (µ, Λ) deﬁned with a known function g(·) the log-likelihood for (µ, Λ) is simply ln (µ, Λ) = cte +

n i=1

ln g[(xi − µ) Λ−1 (xi − µ)] − 12 n ln |Λ|.

(13.3)

214

13. Robustness

Diﬀerention with respect to µ and Λ (v. Problems 1.8.9 and 1.8.10) leads to the equations n

ˆ −1 (xi − µ) ˆ u(si )Λ

= 0

ˆ −1 (xi − µ)(x ˆ −1 ˆ −1 − 1 nΛ ˆ i − µ) ˆ Λ u(si )Λ 2

= 0,

i=1 1 2

n i=1

ˆ −1 (xi − µ). ˆ Λ ˆ Thus, the MLE where u(s) = −2g (s)/g(s) and si = (xi − µ) ˆ estimating equations ˆ Λ)) satisﬁes the implicit (because si depends on (µ, ˆ = µ ˆ Λ =

ave [u(si )xi ] /ave [u(si )] ,

(13.4)

ˆ i − µ) ˆ ]. ave [u(si )(xi − µ)(x

(13.5)

The notation “ave” means arithmetic average over i = 1, . . . , n. Example 13.7 The multivariate Student’s tp,ν has g(s) ∝ (1+s/ν)−(ν+p)/2 and u(s) = (ν + p)/(ν + s). Note that u(s) ≥ 0 and is strictly decreasing. It acts as a weight function, giving more weight to data points with small squared Mahalanobis distances. The existence and unicity of a solution to the estimating equations is a diﬃcult problem. For the location-only problem, it is known in the univariate case [Reeds (1985)] that the estimating equation is susceptible to multiple solutions. Uniqueness of the solution in the univariate locationscale Cauchy (ν = 1) problem was established by Copas (1975) and for ν > 1 by M¨ arkel¨ ainen et al. (1981). The approach presented here is that of Kent and Tyler (1991), which works equally well in the multivariate case. The location-scale problem is very tricky, but the scale-only problem is quite simple. We will thus concentrate on the latter problem. Scale-only problem For the scale-only problem, we assume without any loss of generality that µ = 0. The log-likelihood reduces to l(A) = cte +

n

ln g(xi A−1 xi ) − 12 n ln |A|

i=1

and the estimating equation simpliﬁes to ˆ = ave [u(si )xi x ] , A i

(13.6)

ˆ −1 xi . Let ψ(s) = su(s) and assume where u(·) is as before and si = xi A that lim ψ(s) = a0 > 0.

s→∞

13.3. Maximum likelihood estimates

215

This condition is satisﬁed for the tp,ν distribution, as lims→∞ ψ(s) = ν + p. The following condition on the data is to ensure the existence of a solution to (13.6). It speciﬁes that the data points should not be too concentrated in low-dimensional linear subspaces of Rp . Let Pn (·) denote the empirical distribution of x1 , . . . , xn , i.e., for any borel set B ⊂ Rp 1 I(xi ∈ B). n i=1 n

Pn (B) =

Condition D. For all linear subspaces V ⊂ Rp with dim V ≤ p − 1, Pn (V) < 1 − [p − dim V]/a0 .

The existence of a solution under condition D is proved in Kent and Tyler (1991). Proving existence is the most diﬃcult part, but uniqueness of the solution and convergence of a numerical algorithm is much simpler. ˆ > 0 such that Proposition 13.3 Under condition D, there exists A ˆ ≤ l(A), ∀A > 0. l(A) Note that when sampling from an absolutely continuous distribution condition D is satisﬁed w.p.1 for a0 > p and sample sizes n ≥ p since for any subspace V, k = dim V ≤ p − 1, Pn (V)

w.p.1 k k (p − k) (p − k) ≤ . ≤ =1− 1. Since su(s) is strictly increasing and u(s) is nonincreasing, it follows that for x = 0, u(x A−1 x) ≤ u(λ−1 1 x x) ≤

λ−1 1 x x u(λ−1 1 x x) < λ1 u(x x), λ−1 1 x x

where the ﬁrst inequality used Rayleigh’s quotient. This implies A = ave u(xi A−1 xi )xi xi < λ1 ave [u(xi xi )xi xi ] = λ1 I.

216

13. Robustness

This gives the contradiction λ1 < λ1 , and so λ1 ≤ 1. A similar argument shows λp ≥ 1. Thus, A = I. 2 Under conditions D and M, the unique solution can be found by regarding the estimating equation as a ﬁxed-point equation. Given a starting value A0 > 0, deﬁne the iterative numerical algorithm Am+1 = ave u(xi A−1 m xi )xi xi , m = 0, 1, . . . . Proposition 13.5 Under conditions D and M, for any A0 > 0, Am converges as m → ∞ to the unique solution (the MLE) of (13.6). Proof. Conditions D and M ensure existence and uniqueness of a solution ˆ Since xi → Bxi , B ∈ Gp , induces the new solution A ˆ → BAB ˆ , A. ˆ one can assume without loss of generality that A = I is the solution. Let λ1,m ≥ · · · ≥ λp,m be the eigenvalues of Am , m = 1, 2, . . .. Step 1: The following results are established: (i) λ1,m ≤ 1 =⇒ λ1,m+1 ≤ 1, (ii) λ1,m > 1 =⇒ λ1,m+1 < λ1,m , (iii) λp,m ≥ 1 =⇒ λp,m+1 ≥ 1, and (iv) λp,m < 1 =⇒ λp,m+1 > λp,m . Note that (iii) and (iv) imply Am+1 > 0 whenever Am > 0. To prove (i), if λ1,m ≤ 1, then −1 x A−1 m x ≥ λ1,m x x ≥ x x, ˆ and since u(s) is nonincreasing, u(x A−1 m x) ≤ u(x x). Given that A = I is the solution, this implies Am+1 ≤ ave [u(xi xi )xi xi ] = I. Thus, λ1,m+1 ≤ 1. −1 The proof of (iii) is similar. To prove (ii), since x A−1 m x ≥ λ1,m x x, u(s) is nonincreasing and su(s) is strictly increasing, it follows that if λ1,m > 1, then −1 u(x A−1 m x) ≤ u(λ1,m x x) ≤ λ1,m u(x x),

with the second inequality strict for x = 0. This implies Am+1 < λ1,m ave [u(xi xi )xi xi ] = λ1,m I. Thus, λ1,m+1 < λ1,m . The proof of (iv) is similar. Step 2: We shall now show that (v) lim sup λ1,m ≤ 1, (vi) lim inf λp,m ≥ 1, from which it follows that λ1,m → 1 and λp,m → 1, so that Am → I (v. Problem 1.8.13). Given A > 0, let λ1 (A) denote the largest eigenvalue and deﬁne φ(A) = ave u(xi A−1 xi )xi xi .

13.3. Maximum likelihood estimates

217

Step 1 implies that if λ1 (A) > 1 and B = φ(A), then λ1 (B) < λ1 (A). In view of step 1, statement (v) requires proof only in the case in which λ1,m = λ1 (Am ) > 1, ∀m. Note that λ1,m is a decreasing sequence in this case. Let λ∗ = lim λ1,m ≥ 1 and suppose, if possible, that λ∗ > 1. From step 1, the eigenvalues of the sequence Am are bounded away from 0 and ∞. Thus, we can ﬁnd a convergent subsequence Amj → B0 say, where B0 > 0. Further, Amj +1 = φ(Amj ) → φ(B0 ) = B1 , say. Since λ1,m is decreasing, λ1 (B0 ) = lim λ1,mj = λ∗ and λ1 (B1 ) = lim λ1,mj +1 = λ∗ . However, step 1 implies that λ1 (B1 ) < λ1 (B0 ), giving a contradiction. Hence, (v) follows. Item (vi) is proved similarly. 2 Location-scale problem Results for location scale are derived by embedding the p-dimensional location-scale problem into a (p + 1)-dimensional scale-only problem. For given Λ ∈ Pp , µ ∈ Rp , and γ > 0, let Λ + γ −1 µµ γ −1 µ A= (13.7) ∈ Rp+1 p+1 γ −1 γ −1 µ and observe that any A ∈ Pp+1 can be written in this form. On using the inverse of a partitioned matrix (v. Problem 1.8.1), one ﬁnds Λ−1 −Λ−1 µ A−1 = . −µ Λ−1 γ + µ Λ−1 µ Now deﬁne the artiﬁcial vectors yi = (xi , 1) ∈ Rp+1 and note that yi A−1 yi = (xi − µ) Λ−1 (xi − µ) + γ.

(13.8)

Let A(1) be deﬁned as in (13.7) but with γ = 1. Upon using (13.8) and |A(1) | = |Λ|, the objective function (13.3) can be expressed as ln (µ, Λ) = l(A(1) ) = cte +

n

1 ln g(yi A−1 (1) yi − 1) − 2 n ln |A(1) |.

(13.9)

i=1

Thus, the problem of maximizing (13.3) over µ ∈ Rp and Λ ∈ Pp is equivalent to maximizing l(A(1) ) over A(1) ∈ Pp+1 with the restriction that the (p + 1, p + 1) element of A(1) be 1. Moreover, the estimating equations (13.4) and (13.5) can be rewritten in a single estimating equation as −1 ˆ ˆ ˆ ˆ µ γˆ −1 µ ˆ = Λ + γˆ µ (13.10) = ave[u(si )yi yi ], A ˆ γˆ −1 γˆ −1 µ ˆ −1 (xi − µ), ˆ Λ ˆ as in the original where γˆ −1 = ave[u(si )] with si = (xi − µ) location-scale formulation. Using (13.8), the single estimating equation can be reexpressed as ! ˆ = ave u∗ (y A ˆ −1 yi ; γˆ )yi y , (13.11) A i i

218

13. Robustness

where u∗ (s; γ) = u(s − γ), for s ≥ γ. This looks very similar to the estimating equation of a scale-only problem, the diﬀerence being that the function u∗ (·; γˆ ) depends on the data through γˆ . The next condition for existence of a solution is just the previous condition D on yi ’s recast in terms of xi ’s. Condition D1. For all translated linear subspaces (hyperplanes) H ⊂ Rp with dim H ≤ p − 1, Pn (H) < 1 − (p − dim H)/a0 .

This time if a0 > p + 1, n ≥ p + 1, then condition D1 is satisﬁed w.p.1 when sampling from an absolutely continuous distribution. Proposition 13.6 If conditions D1 and M hold, then there exists a soluˆ > 0 to (13.11). This solution is unique if (s + γˆ )u(s) is ˆ ∈ Rp and Λ tion µ strictly increasing in s ≥ 0 for γˆ −1 = ave[u(si )] deﬁned above. A diﬃculty in applying Proposition 13.6 is the strictly increasing condition ˆ the ˆ Λ) which depends on the unknown γˆ . However, given a solution (µ, condition guarantees that no other solutions exist. For the tp,ν distribution, ν ≥ 1, we prove that γˆ is independent of the data (ˆ γ = 1) and the condition is thus automatically satisﬁed for ν > 1 since (s + γˆ )u(s) = (ν + p)(s + 1)/(s + ν) is strictly increasing. Lemma 13.1 For the tp,ν distribution ν ≥ 1, γˆ = 1. Proof. If γu ≥ γ and (s + γu )u(s) is strictly increasing and condition M ˆ −1 and holds, then (s + γ)u(s) is also strictly increasing. Multiplying by Λ taking the trace of (13.5), we get ave[si u(si )] = p. Thus, ∀b > 0, p = ave[(si + b)u(si )] − bˆ γ −1 , which implies γl ≤ γˆ ≤ γu , where γu−1

=

γl−1

=

sup inf [(s + b)u(s) − p]/b, b>0 s>0

inf sup[(s + b)u(s) − p]/b.

b>0 s>0

Letting b = ν, we obtain 1 ≤ γl ≤ γˆ ≤ γu ≤ 1.

2

The Cauchy case, ν = 1, has (s + 1)u(s) = p + 1, which is not strictly increasing. It requires a special treatment, but the MLE is also unique under condition D1 [Kent and Tyler (1991)]. For the tp,ν case, since γˆ is independent of the data, this means that when condition D1 is satisﬁed, the ﬁxed-point algorithm still converges to the MLE. So, for any starting

13.3. Maximum likelihood estimates

219

values µ0 and Λ0 > 0, the iterative equations & ' ave u (xi − µm ) Λ−1 m (xi − µm ) xi & ' , µm+1 = ave u (xi − µm ) Λ−1 m (xi − µm ) ' & Λm+1 = ave u (xi − µm ) Λ−1 m (xi − µm ) (xi − µm )(xi − µm ) converge to the MLE. Asymptotics for the MLE The general theory of maximum likelihood coupled with the fact that the MLE is aﬃne equivariant tells us that for some constants β, σ1 , and σ2 , d ˆ − Λ, µ ˆ − µ) → (N, n), n1/2 (Λ

where n ∼ Np (0, βΛ) N ∼ Npp (0, σ1 (I + Kp )(Λ ⊗ Λ) + σ2 vec(Λ)[vec(Λ)] ) . |=

Using Fisher’s information, these constants can now be evaluated and we can also show that N n; i.e., they are asymptotically independent. The score function is the derivative of l(µ, Λ) = cte + ln g[(x − µ) Λ−1 (x − µ)] −

1 2

ln |Λ|

|=

with respect to (µ, Λ) and its variance is called Fisher’s information and is denoted by I(µ, Λ). It is also well known that the asymptotic variance is the inverse of Fisher’s information. Let us show that I(µ, Λ) is block-diagonal and thus N n. We have ∂l/∂µ = u(s)Λ−1 (x − µ) ∂l/∂Λ = − 12 Λ−1 + 12 u(s)Λ−1 (x − µ)(x − µ) Λ−1 , where s = (x − µ) Λ−1 (x − µ). The constants β, σ1 , and σ2 being independent of (µ, Λ), it suﬃces to evaluate the variance of the score while d

assuming (µ, Λ) = (0, I) and x = z ∼ Ep (0, I). The expectation E{∂l/∂µi · ∂l/∂Λjk } involves only ﬁrst-order and third-order product moments of z, which is spherical. Since these moments are all null, it follows that I(0, I) is blockdiagonal with blocks I1 and I2 , say. We then calculate β from I1−1 . Now, & ' I1 = E (∂l/∂µ) (∂l/∂µ) = E u2 (s)zz = E[su2 (s)]E[uu ], d

|=

where we have let z = s1/2 u, where s = |z|2 , u ∼ unif(S p−1 ), and u. Then, I1 = p−1 E[su2 (s)]I if we note that E uu = p−1 I (v. Probs lem 13.6.4). Thus, we have shown that β = p/E[su2 (s)]. We now evaluate

220

13. Robustness

σ1 and σ2 from I2 . We would like to identify var N with I2−1 , but we must ﬁrst eliminate the redundant elements of the symmetric N for var N to become nonsingular. For this reason, deﬁne Aj

(0, Ij ) : p × j, j = 1, . . . , p, diag(Ap , . . . , A1 ) : p2 × 12 p(p + 1),

= =

Mp

and verify that for any symmetric A ∈ Rpp , Mp vec(A) is the 12 p(p + 1)dimensional vector formed by stacking the columns of A after deleting the upper triangular part of A. Now, var(Mp vec(N)) = Mp var(N)Mp . It is easy to check that Mp Mp = I, Mp Kp Mp = Dp , Mp vec(Ip ) = ap , where αj

=

(1, 0, . . . , 0) ∈ Rj , j = 1, . . . , p,

ap

= =

(αp , . . . , α1 ) : 12 p(p + 1) × 1, diag(ap ).

Dp Then, we can identify

I2−1 = σ1 (I + Dp ) + σ 2 ap ap . Using the inverse of a perturbed matrix (v. Problem 1.8.8), we have with the relations (I + Dp )−1 ap = 12 ap and ap ap = p, I2

= σ1−1 (I + Dp )−1 − σ2 [4σ12 (1 + 12 pσ2 σ1−1 )]−1 ap ap = i1 (I + Dp )−1 + i2 ap ap ,

where i1 = σ1−1 ,

i2 = −σ2 [4σ12 (1 + 12 pσ2 σ1−1 )]−1 .

(13.12)

As an example for p = 2, we thus have the identiﬁcation I2

= i1 (I + Dp )−1 + i2 ap ap 1 0 i2 2 i1 + i2 = 0 i1 0 1 0 2 i1 + i2 i2 2 ∂l ∂Λ11

∂l ∂Λ11

∂l ∂l ∂l = E ∂Λ11 ∂Λ21 ∂Λ21 ∂l ∂Λ11

∂l ∂Λ22

Thus, in general for i = j, i1 i2

∂l ∂Λ21

∂l ∂Λ21 2

∂l ∂Λ22

∂l ∂Λ11 ∂l ∂Λ21

' & = E (∂l/∂Λij )2 , = E {(∂l/∂Λii )(∂l/∂Λjj )} .

∂l ∂Λ22

∂l ∂Λ22 ∂l ∂Λ22 2

.

13.3. Maximum likelihood estimates

221

These are evaluated with d

∂l/∂Λij

= u(s)zi zj = ψ(s)ui uj ,

∂l/∂Λii

= − 12 + 12 u(s)zi2 = − 12 + 12 ψ(s)u2i . d

Using Problem 13.6.4, E u2i = p−1 and E u2i u2j = [p(p + 2)]−1 , i = j, the ﬁnal result is thus i1 i2

= [p(p + 2)]−1 E ψ 2 (s), = − 14 + [p(p + 2)]−1 E ψ 2 (s)

if we note that E ψ(s) = p. The constants σ1 and σ2 are obtained by solving equation (13.12). The density of s was given in Problem 4.5.13. We have proved that under regularity conditions for the MLE [Lehmann (1983), pp. 429-430] Proposition 13.7

where N

|=

d ˆ − Λ, µ ˆ − µ) → (N, n), n1/2 (Λ

n and

n ∼ Np (0, βΛ) , N ∼ Npp (0, σ1 (I + Kp )(Λ ⊗ Λ) + σ2 vec(Λ)[vec(Λ)] ) , with β σ1 σ2

= p/E[su2 (s)], = p(p + 2)/E[ψ 2 (s)], = −2σ1 (1 − σ1 )/[2 + p(1 − σ1 )]

and s has density π p/2 1 p−1 g(s), s > 0. s2 Γ( 12 p) The parameter σ1 of the asymptotic variance will play a major role as an index of relative eﬃciency for robust tests. Example 13.8 For the tp,ν distribution, the reader can check σ1 = 1 + 2/(p + ν). The maximum likelihood estimation of the multivariate tp,ν distribution with possibly missing data and unknown degrees of freedom was treated by Liu (1997). Missing data imputation using the multivariate tp,ν distribution was also the subject of Liu (1995).

222

13. Robustness

13.4 Robust estimates An alternative approach to MLE consists of robust location and scatter estimates such as the M estimate [Maronna (1976), Huber (1981)] or the S estimate [Davies (1987), Lopuha¨a (1989)]. The theoretical proofs for existence, unicity, consistency, and asymptotic normality of these estimates go beyond the scope of this book. Of importance to us, however, is to show how √ easily these aﬃne equivariant and n-asymptotically normal estimates can serve as the building block to robust tests on location and scatter. They are succintly introduced now and invoked later to construct robust tests.

13.4.1

M estimate

Let x1 , . . . , xn i.i.d. x ∼ Ep (µ, Λ) and z ∼ Ep (0, I). The idea behind M estimate is to modify the MLE estimating equations to gain robustness. The M estimate of location and scatter are deﬁned as solution to the equations µn

=

Vn

=

ave [u1 (ti )xi ] /ave [u1 (ti )] , ave u2 (t2i )(xi − µn )(xi − µn ) ,

(13.13) (13.14)

1/2 . where ti = (xi − µn ) Vn−1 (xi − µn ) The M estimates are obviously aﬃne equivariant. Interestingly, they include, as a particular case, the MLE estimate with the functions u1 (t) = −2g (t2 )/g(t2 ) and u2 (t2 ) = u1 (t). Deﬁne ψi (s) = sui (s), i = 1, 2. The following conditions on the functions are needed and will always be assumed: M1. u1 and u2 are non-negative, nonincreasing, and continuous on [0, ∞). M2. ψ1 and ψ2 are bounded. Let Ki = sups≥0 ψi (s). M3. ψ2 is nondecreasing and is strictly increasing in the interval where ψ2 < K2 . M4. There exists s0 such that ψ2 (s20 ) > p and that u1 (s) > 0 for s ≤ s0 (and, hence, K2 > p). Example 13.9 The tp,ν MLE has ψ1 (s) = (ν + p)s/(ν + s2 ) and ψ2 (s) = (ν + p)s/(ν + s). It is easy to verify M1 through M4. Example 13.10 Huber’s ψ function is deﬁned as ψ(s, k) = max[−k, min(s, k)]. Let k > 0 be a constant and take ψ1 (s) = ψ(s, k) and ψ2 (s) = ψ(s, k2 ). A further condition on the data is needed for existence of the M estimate.

13.4. Robust estimates

223

Condition D2. There exists a > 0 such that for every hyperplane H, dim H ≤ p − 1, p Pn (H) ≤ 1 − − a. K2 When sampling from an absolutely continuous distribution, condition D2 is satisﬁed w.p.1 for n suﬃciently large. Proposition 13.8 If condition D2 is satisﬁed, there exists a solution (µn , Vn ) to (13.13) and (13.14). Moreover, µn belongs to the convex hull of {x1 , · · · , xn }. Proposition 13.9 Assume condition D2 and g is decreasing. Let (µn , Vn ) be a solution to (13.13) and (13.14), then (µn , Vn ) → (µ, V) almost surely, where V = σ −1 Λ with σ being the solution to E ψ2 (σt2 ) = p and t = |z|. The reason for the presence of σ is that Vn is consistent for a certain multiple of Λ, σ −1 Λ say, deﬁned by the implicit equation V = E u2 [(x − µ) V−1 (x − µ)](x − µ)(x − µ) . Multiplying by V−1 and taking trace yields E ψ2 (σ|z|2 ) = p. This expectation can be evaluated as a simple integral if one recalls the density of t = |z| (v. Problem 4.6.13): f (t) =

2π p/2 p−1 2 t g(t ), t ≥ 0. Γ( 12 p)

Proposition 13.10 Assume sψi (s) are bounded (i = 1, 2) and g is decreasing such that E ψ1 (σ 1/2 t) > 0. Then, n1/2 (Vn − V, µn − µ) → (N, n), where n

|=

d

N and

n ∼ Np 0, (α/β 2 )V , N ∼ Npp (0, σ1 (I + Kp )(V ⊗ V) + σ2 vec(V)[vec(V)] ) ,

with σ being the solution to E ψ2 (σt2 ) = p, where α β σ1 σ2

= p−1 E ψ12 (σ 1/2 t), !

= E 1 − p−1 u1 (σ 1/2 t) + p−1 ψ1 (σ 1/2 t) , = a1 (p + 2)2 (2a2 + p)−2 , & ' (a1 − 1) − 2a1 (a2 − 1)[p + (p + 4)a2 ](2a2 + p)−2 , = a−2 2

and a1 a2

= [p(p + 2)]−1 E ψ22 (σt2 ), = p−1 E [σt2 ψ2 (σt2 )].

224

13. Robustness

Those results are due to Maronna (1976), but Tyler (1982) found the asymptotic variance parameters σ1 and σ2 in Proposition 13.10. Asymptotic theory for robust principal components was developed by Tyler (1983b) and Boente (1987).

13.4.2

S estimate

Recently, Davies (1987) and Lopuha¨ a (1989) investigated properties of the S estimate for multivariate location and scatter. As before, consider a random sample x1 , . . . , xn i.i.d. x ∼ Ep (µ, Λ) and z ∼ Ep (0, I). Again, let t = |z|. In the context of regression, Rousseeuw and Yohai (1984) obtained an asymptotically normal and robust estimate from a function ρ assumed to satisfy the following: S1: ρ is symmetric, has a continuous derivative ψ, and ρ(0) = 0. S2: There exists a ﬁnite constant c0 > 0 such that ρ is strictly increasing on [0, c0 ] and constant on [c0 , ∞). Let a0 = sup ρ. A typical ρ function is Tukey’s biweight 2 t /2 − t4 /(2c20 ) + t6 /(6c40 ) if |t| ≤ c0 ρ(t) = c20 /6 if |t| ≥ c0 . The S estimate (µn , Vn ) is deﬁned as the solution of the optimization 1/2 problem where ti = (xi − µn ) Vn−1 (xi − µn ) : 1 ρ(ti ) = b0 n i=1 n

min |Vn | subject to

over all µn ∈ Rp and Vn > 0. The constant b0 , 0 < b0 < a0 , chosen so that 0 < b0 /a0 ≡ r ≤ (n − p)/2n, leads to a ﬁnite-sample breakdown point [Lopuha¨ a and Rousseeuw (1991)] of ∗n = nr/n. The choice r = (n−p)/2n results in the maximal breakdown point (n − p + 1)/2!/n (asymptotically 50%). Roughly speaking, the breakdown point is the minimum percentage of contaminated data necessary to bring the estimate beyond any given bound. The sample mean requires only one point and thus has a breakdown point 1/n, or asymptotically 0%. To obtain simultaneously a breakdown point of ∗n = nr/n and a consistent estimate of scale, i.e., Vn → Λ w.p.1, for a given Ep (µ, Λ) distribution the constant c0 is chosen so that E ρ(t)/a0 = r and then b0 is set to E ρ(t). A geometrical interpretation of S estimate can be given with the ellipsoidal contours of an Ep (µ, Λ). First, the volume of a p-dimensional ellipsoid z Λ−1 z ≤ 1 is |Λ|1/2 2π p/2 /[pΓ( 12 p)]; thus, minimizing |Λ| corresponds to ﬁnding a minimum volume ellipsoid [Rousseeuw (1985)]. Second, if we could allow discontinuous ρ, then ρ(t) = 1 − I[−c0 ,c0 ] (t) would count the points outside the ellipsoid. So, for r = 25%, the optimization would

13.4. Robust estimates

225

ﬁnd the minimum volume ellipsoid containing 75% of the data. An S estimate is thus a smoothed √ version of a minimum volume ellipsoid. The smoothing is done to get n-asymptotically normal estimates. Assuming further S3: ρ has a second derivative ψ , both ψ (t) and u(t) = ψ(t)/t are bounded and continuous, the asymptotic normality of the S estimate holds. Proposition 13.11 Assume S1 through S3. Let V = Λ and assume E ψ (t) > 0, E ψ (t)t2 + (p + 1)ψ(t)t > 0. Let α β

= p−1 E ψ 2 (t), = E 1 − p−1 u(t) + p−1 ψ (t) , p(p + 2)E[ψ 2 (t)t2 ] , E 2 [ψ (t)t2 + (p + 1)ψ(t)t]

σ1

=

σ2

= −2p−1 σ1 + 4

2

E [ρ(t) − b0 ] , E 2 [ψ(t)t]

then n1/2 (Vn − V, µn − µ) → (N, n), where n

|=

d

N and

n ∼ Np 0, (α/β 2 )V N ∼ Npp (0, σ1 (I + Kp )(V ⊗ V) + σ2 vec(V)[vec(V)] ) .

r = .5 r = .3 r = .1

p=1

p=2

p = 10

26.9% 40.5% 49.1%

37.7% 77.0% 98.9%

91.5% 98.0% 99.9%

Table 13.1. Asymptotic eﬃciency of S estimate of scatter at the normal distribution.

According to Lopuha¨ a (1989) the asymptotic eﬃciency for the estimation of the scatter as measured by the index σ1 (or 2σ1 + σ2 for p = 1) are as in Table 13.1 at the normal distribution. The asymptotic eﬃciency of the location estimate are even higher. For the S estimate, a high breakdown point corresponds to a low eﬃciency and vice versa. Let us mention that S estimates are able to achieve the asymptotic variance of M estimates. However, S estimates can have a

226

13. Robustness

high breakdown point in any dimension, whereas the asymptotic breakdown point of an M estimate is at most 1/(p + 1) [Tyler (1986)]. Lopuha¨ a (1991) deﬁnes τ estimates which can have the same high breakdown point as S estimates√but can attain simultaneously high eﬃciency. The τ estimates are also n-asymptotically normal. An S-plus [Statistical Sciences, (1995)] function, s.estimate, to evaluate S estimate is described in Appendix C. The implementation follows the recommendations of Ruppert (1992) to increase the speed of numerical convergence of this numerically intensive problem. The S-plus function asymp evaluates the asymptotic variance constants λ = α/β 2 , σ1 , and σ2 , at the normal distribution.

Robust Hotelling-T 2

13.4.3

Assume x1 , . . . , xn are i.i.d. x ∼ Ep (µ, Λ). Consider a test of hypothesis on the mean, H0 : µ = µ0 against H1 : µ = µ0 , using a robust version of the classical Hotelling-T 2 . Assume (Vn , µn ) is a robust aﬃne equivariant and asymptotically normal estimate (M or S estimate for example), n1/2 (Vn − V, µn − µ) → (N, n), where n

|=

d

N and

n ∼ Np 0, (α/β 2 )V , N ∼ Npp (0, σ1 (I + Kp )(V ⊗ V) + σ2 vec(V)[vec(V)] ) .

Proposition 13.12 Under the sequence of contiguous alternatives H1,n : µ = µ0 + n−1/2 γ, TR2 = n(µn − µ0 ) Vn−1 (µn − µ0 ), where (Vn , µn ) is asymptotically normal as above, satisﬁes 2 β −1 d α γV γ . TR2 → 2 · χ2p β 2α d

In particular, TR2 →

α 2 β 2 χp

under H0 .

Proof. Let X and Y be the sample matrices under H0 and H1,n , respectively. Then, we can write Y = X + n−1/2 1γ . d

Aﬃne equivariance of the estimate immediately gives µn (Y) = µn (X) + n−1/2 γ, d d

Vn (Y) = Vn (X) → V w.p.1. Since

d d n1/2 (µn (Y) − µ0 ) = n1/2 (µn (X) − µ0 ) + γ → Np γ, (α/β 2 )V ,

13.5. Robust tests on scale matrices

227

it follows from Corollary 5.1 on quadratic forms (with A = (β 2 /α)V−1 ) that 2 β −1 d α γV γ . TR2 = n(µn (Y) − µ0 ) Vn−1 (Y)(µn (Y) − µ0 ) → 2 χ2p β 2α 2 Another type of robustness found in the literature assumes an elliptical distribution on the whole data matrix X ∈ Rnp with mean 1µ and variance of the form I ⊗ Σ. Under weak assumptions on the p.d.f., the classical Hotelling-T 2 (8.1) remains UMPI and the null distribution of T 2 is the same as if xi had been i.i.d. Np (µ, Σ), i.e., T 2 ∼ Fc (p, n − p) [v. Corollary 8.1]. The main diﬀerence in the two approaches resides in that the observations under an elliptical distribution on X cannot be independent, although they are uncorrelated, unless the elliptical distribution is normal. Independence and spherical symmetry do not go together, except in the normal case, by virtue of the Maxwell-Hershell theorem [v. Proposition 4.11]. One may consult the book by Kariya and Sinha (1989) on this type of robustness for many statistical tests. Having found the asymptotic null distribution of Hotelling-T 2 , it is now a simple matter to extend the results of Section 8.3 to construct robust simultaneous conﬁdence intervals on means. For example, asymptotically, we are at least (1 − γ) × 100% conﬁdent in simultaneously presenting all of the observed “Scheﬀ´e” intervals: 1/2 1/2 α χ2γ,p α χ2γ,p 1/2 (a Vn a) ≤ a µ ≤ a µn + (a Vn a)1/2 , a µn − β2 n β2 n ∀a ∈ Rp . Realistically, the parametric family Ep (µ, Λ) is unknown. Thus, α and β will have to be replaced by consistent estimates.

13.5 Robust tests on scale matrices Assume x1 , . . . , xn are i.i.d. x ∼ Ep (µ, Λ). Consider a test of hypothesis on Λ which is of the general form h(Λ) = 0, where h(Λ) ∈ Rq is a continuously diﬀerentiable function. We will assume µ = 0. Under a Np (0, Λ) distribution, recall that a likelihoodratio test on Λ is based n uniquely on the likelihood statistic Sn = n1 i=1 xi xi . We know that nSn ∼ Wp (n, Λ). Thus, Sn has density (up to a multiplicative constant) |Λ|−n/2 etr(− 12 nΛ−1 Sn ). So, we deﬁne f (A, Λ) = |Λ|−n/2 etr(− 12 Λ−1 A), fh (A) = sup f (A, Λ), h(Λ)=0

Lh (A)

=

fh (A) . f (A, A)

228

13. Robustness

Note that Lh (Sn ) is the likelihood ratio test for H0 : h(Λ) = 0 when x ∼ Np (0, Λ). The idea to build a robust test when x ∼ Ep (0, Λ) is ˆ n ) where Λ ˆ n could be Sn [Muirhead and to use the test statistic Lh (Λ Waternaux, (1980)] or, preferably, a more robust estimate [Tyler (1983a)]. Other approaches which will not be considered here include those based on minimum discrepancy test statistics [Browne and Shapiro (1987), Shapiro and Browne (1987)].

13.5.1

Adjusted likelihood ratio tests

A general method of making a simple correction to the likelihood ratio test is possible for hypotheses satisfying the following condition H on the function h. Condition H. h(Γ) = h(γΓ), ∀γ > 0, ∀Γ > 0. Examples of hypothesis satisfying condition H are the test of sphericity and the test of covariance. Example 13.11 The test of sphericity H0 : Λ = γI for some unknown γ can be written as H0 : h(Λ) = 0 with hij (Λ) = Λij /Λpp , 1 ≤ i < j ≤ p, and hii (Λ) = Λii /Λpp − 1, i = 1, . . . , p − 1. Here, we have q = 12 (p − 1)p + p − 1. Example 13.12 The test of covariance between two subvectors H0 : Λ12 = 0, where Λ12 ∈ Rpp12 can be written as H0 : h(Λ) = 0 by choosing h(Λ) = −1/2

vec(Λ11

−1/2

Λ12 Λ22

). Obviously, q = p1 p2 .

Condition H is not dependent on the location or the spread of the elliptical contours, but concerns only the direction and relative lengths of the ˆ n is also necessary. axes of the contours. A condition E on the estimate Λ However, as we encountered in M and S estimation, we usually have an estimate Vn of a multiple V of Λ. Note that hypothesis H0 : h(Λ) = 0 is equivalent to H0 : h(V) = 0 under condition H. d

Condition E. Vn is aﬃne equivariant and n1/2 (Vn − V) → Z, where Z ∼ Npp (0, σ1 (I + Kp )(V ⊗ V) + σ2 vec(V)[vec(V)] ).

Normal and elliptical MLE, the M estimate, and the S estimate satisfy condition E under regularity conditions. An estimate of h(V) is h(Vn ), whose asymptotic distribution follows from the delta method (v. Proposition 6.2). A diﬃculty is the redundance of variables due to the symmetry of V. For this reason, the following derivative will be very useful. Deﬁne da/db = (dai /dbj ), where i varies over rows and j runs over columns. The derivative of h(V) with respect to V is deﬁned as h (V) = 12 [d h(V)/d vec(V)](I + Jp ) ∈ Rqp2 ,

13.5. Robust tests on scale matrices

229

p where Jp = i=1 ei ei ⊗ ei ei and ei ∈ Rp is a vector of zero but a 1 in position i. An example when q = 1 and p = 2 is enlightening: .. . 0 0 2 0 .. 0 . 0 0 1 1 2 (dh/ds11 , dh/ds21 , dh/ds12 , dh/ds22 ) · · · · · · · · · · · · · · · .. 0 . 1 0 0 .. . 0 2 0 0

= dh/ds11 , 12 dh/ds21 , 12 dh/ds12 , dh/ds22 is the usual gradiant of h taking into account the symmetry. Before stating the result, we need a lemma on gradiants. Lemma 13.2 Let f : Rp → R be continuously diﬀerentiable. Then, df (x)/dx, x = 0 for all x in a neighborhood of x0 iﬀ f (x) = f (αx) for all x and αx in a neighborhood of x0 . Proof. It suﬃces to notice that the contours of f are rays coming out of the origin and that the gradiant is a perpendicular vector to the contour.2 d

Proposition 13.13 Under conditions H and E, n1/2 [h(Vn ) − h(V)] → Zh , where Zh ∼ Nq (0, 2σ1 [h (V)](V ⊗ V)[h (V)] ) . Proof. As in Proposition 6.2, we can write n1/2 [h(Vn ) − h(V)] = h (V) n1/2 vec(Vn − V) + op (1). Therefore, n1/2 [h(Vn ) − h(V)] → h (V) vec(Z). d

From Lemma 13.2 and condition H, we have h (V) vec(V) = 0. Thus, the asymptotic variance is var h (V) vec(Z)

= σ1 [h (V)](I + Kp )(V ⊗ V)[h (V)] , = σ1 [h (V)](V ⊗ V)(I + Kp )[h (V)] .

Applying the identity Kp vec(A) = vec(A ) to the columns of [h (V)] gives (I + Kp )[h (V)] = 2[h (V)] and the conclusion follows. 2 An important consequence of condition H is the asymptotic variance which becomes independent of σ2 . This means that for Vn satisfying condition E, all the asymptotic distributions of h(Vn ), under condition H, such as a simple correlation coeﬃcient, a multiple correlation coeﬃcient, a ratio of eigenvalues, etc., are the same as those for the sample variance S, when

230

13. Robustness

sampling from a multivariate normal distribution, except for the factor σ1 in the asymptotic variance. For completeness, the results for correlations are now given. Denote by rij the simple correlation deﬁned from the scale estimate Vn = (vn,ij ), i.e., vn,ij rij = 1/2 1/2 , vn,ii vn,jj and let ρij =

Λij 1/2 1/2 Λii Λjj

be the correlation for the Ep (µ, Λ) distribution. Proposition 13.14 Assume condition E holds on Vn . Then, 1/2

n1/2 (rij − ρij ) → σ1 d

· N (0, (1 − ρ2ij )2 ).

From the delta method it is also clear that an arbitrary number of correlation coeﬃcients is jointly asymptotically normal. Thus, it suﬃces to consider the case of two correlation coeﬃcients rij and rkl . Proposition 13.15 Assume condition E holds on Vn . Then, (1 − ρ2ij )2 ω rij − ρij d 1/2 n1/2 → σ1 · N2 0, , rkl − ρkl ω (1 − ρ2kl )2 where the asymptotic covariance ω is given by ω = ρij ρkl + ρkj ρil − ρlj (ρij ρkj + ρil ρkl ) − ρki (ρij ρil + ρkj ρkl ) + 12 ρki ρlj (ρ2ij + ρ2il + ρ2kj + ρ2kl ). Proof. Assume V = (ρij ) without loss of generality. Write down the asymptotic distribution vn,ij ρij vn,ii 1 1 d 1/2 vn,jj n − → N6 (0, Ω) vn,kl ρkl vn,kk 1 vn,ll 1 for a certain Ω and apply the delta method.

2

ˆ ≡ R(V ˆ n ) and parSimilarly, for the multiple correlation coeﬃcient R tial correlation coeﬃcient rij|x2 ≡ rij|x2 (Vn ), obtained from Vn satisfying condition E, we can write the asymptotic distributions:

d 1/2 ˆ 2 − R2 ) → n1/2 (R σ1 · N 0, 4R2 (1 − R2 )2 ,

d 1/2 n1/2 rij|x2 − ρij|x2 → σ1 · N 0, (1 − ρ2ij|x2 )2 .

13.5. Robust tests on scale matrices

231

Higher-order asymptotic distributions for functions of the sample variance S can also be derived with the use of zonal polynomials [Iwashita and Siotani (1994)]. For the same reason, adjustment to the likelihood ratio test will take a rather simple form. The asymptotic distribution of the modiﬁed likelihood ratio test Lh (Vn ), where Vn may be a robust estimate, is obtained with the equivalent form of Wald’s test for the same hypothesis. Let un ∼ vn mean p un − vn → 0. The following result on Wald’s formulation holds regardless of condition H. Proposition 13.16 Let An > 0, n = 1, 2, . . ., be such that n1/2 (An − d

A) → (·) for a ﬁxed A > 0 satisfying h(A) = 0. If rank h (Γ) = q, ∀Γ in a neighborhood of A, then −2 ln Lh (An ) ∼ n[h(An )] [Ch (An )]−1 h(An ), where Ch (Γ) = 2[h (Γ)](Γ ⊗ Γ)[h (Γ)] . Proof. This is a generalization of Wald’s formulation for the asymptotic behavior of the likelihood ratio statistic. Refer to Tyler (1983a) for details. 2 Corollary 13.2 Assume conditions H and E. Then: d

(i) under H0 , −2 ln Lh (Vn ) → σ1 χ2q , (ii) under the sequence of contiguous alternatives Λn = Λ+n−1/2 B, where h(Λ) = 0 and B is a ﬁxed symmetric matrix, −2 ln Lh (Vn ) → σ1 χ2q (δh (Λ, B)/2σ1 ) , d

where δh (Λ, B) = [vec(B)] [h (Λ)] [Ch (Λ)]−1 h (Λ)vec(B). d

Proof. From conditions H and E, n1/2 (Vn − V) → Z and n1/2 [h(Vn ) − d

h(V)] → Zh , where Zh ∼ Nq (0, σ1 Ch (V)). Under H0 : h(V) = 0, p

d

n1/2 h(Vn ) → Zh , and since h (·) is continuous, Ch (Vn ) → Ch (V). Hence, we have [n1/2 h(Vn )] [Ch (Vn )]−1 [n1/2 h(Vn )] → Zh [Ch (V)]−1 Zh = σ1 χ2q . d

d

For contiguous alternatives, under condition H, the noncentrality parameter is invariant with respect to scalar multiplication δh (Λ, B) = 2 δh (αΛ, αB), ∀α > 0. As a particular case for Sn which has σ1 = 1 + k, we have, under H0 , d ˆ → −2 ln Lh (Sn )/(1 + k) χ2q

232

13. Robustness

for some consistent estimate kˆ of the kurtosis parameter. So, in the class of Ep (0, Λ) with ﬁnite fourth-order moments, this adjusted LRT is robust in the sense that the asymtotic distribution is the same as if x ∼ Np (0, Λ). Note that a consistent estimate kˆ can be obtained by the method of moment with the identity 1 + k = pE(s2 )/[(p + 2)E 2 (s)],

(13.15)

2

where s = |z| and z ∼ Ep (0, I) has fourth-order moments (v. Problem 13.6.12). More generally, the test statistic −2 ln Lh (Vn )/ˆ σ1 will be referred to as an adjusted LRT. The test of sphericity can serve as an example to illustrate Proposition 13.16 and Corollary 13.2. Wald’s formulation is generally obtained by a Taylor series of −2 ln Lh (Vn ) around V, satisfying H0 : h(V) = 0. For the test of sphericity, we have −2 ln Lh (Vn ) = −n ln |Vn | + pn ln(p−1 tr Vn ). = γI + n−1/2 Zn , Under H0 : V = γI and condition E, we can write Vn ∞ where Zn is bounded in probability. Since ln(1 + x) = i=1 (−1)i+1 xi /i, −1 < x < 1, it follows that for a ﬁxed symmetric A, ln |I + tA| =

∞

(−1)i+1 tr(Ai )ti /i

i=1

for all t suﬃciently small. Hence, we get the expansion −2 ln Lh (Vn )

= d

→

1 −2 [tr(Z2n ) 2γ

− p−1 (tr Zn )2 ] + Op (n−1/2 )

1 −2 [tr(Z2 ) 2γ

− p−1 (tr Z)2 ],

where Z ∼ γNpp (0, σ1 (I + Kp ) + σ2 vec(I)[vec(I)] ). From the relations (v. Problem 6.4.2) tr Z2 = [vec(Z)] 12 (I + Kp ) vec(Z) and tr Z = [vec(I)] vec(Z), it follows that 1 −2 [tr(Z2 ) 2γ

− p−1 (tr Z)2 ] = [vec(Z)] A vec(Z),

where A = 12 γ −2 { 12 (I + Kp ) − p−1 vec(I)[vec(I)] } is a quadratic form. This is Wald’s equivalent formulation for this test. The asymptotic result −2 ln Lh (Vn )/σ1 → χ2q , q = 12 (p − 1)(p + 2) d

follows from Corollary 5.1 on quadratic forms. When condition H is not satisﬁed, simple adjustments to the LRT is generally not possible, as the following corollary shows. Corollary 13.3 Under H0 : h(V) = 0 and condition E, −2 ln Lh (Vn ) → σ1 χ2q−1 + [σ1 + σ2 δh (V, V)]χ21 , d

|=

χ21 . The term δh (V, V) = 0 iﬀ for some neighborhood of V, with χ2q−1 h(Γ) = h(γΓ) for all Γ and γΓ in this neighborhood.

13.5. Robust tests on scale matrices

233

Proof. Take a closer look at the distribution of Zh in Proposition 13.13 d when condition H is not satisﬁed. Under H0 , we still have n1/2 h(Vn ) → h (V) vec(Z) but with an added term in the variance: var h (V) vec(Z)

= σ1 [h (V)](V ⊗ V)(I + Kp )[h (V)] +σ2 h (V) vec(V)[vec(V)] [h (V)] ≡ Dh (V).

Using Proposition 13.16, the equivalent Wald’s formulation is −2 ln Lh (Vn ) ∼ [n1/2 h(Vn )] [Ch (Vn )]−1 [n1/2 h(Vn )], and, thus, −2 ln Lh (Vn ) → Zh [Ch (V)]−1 Zh , d

where Zh ∼ Nq (0, Dh (V)). The result follows since [Ch (V)]−1 Dh (V) has eigenvalues σ1 of multiplicity (q − 1) and σ1 + σ2 δh (V, V). The second statement follows since h (V) vec(V) = 0 iﬀ the stated condition holds. 2

13.5.2

Weighted Nagao’s test for a given variance

In this section, we consider an example where the condition H on the hypothesis is not satisﬁed, but a simple test, robust to large kurtosis, can still be built. For testing the hypothesis, H0 : Σ = I, against H1 : Σ = I, the modiﬁed likelihood ratio test based on n i.i.d. vectors from a Np (µ, Σ) distribution is (v. Problem 8.9.8) Λ∗ = epm/2 |Sn |m/2 etr(− 12 mSn ), m = n − 1, n ¯ )(xi − x ¯ ) /m. It is invariant to orthogonal where Sn = i=1 (xi − x transformations, unbiased, and −2 ln Λ∗ is asymptotically distributed as a noncentral chi-square [Khatri and Srivastava (1974)], χ2f (δ), with f = p 1 2 i=1 di /4, under the sequence of local alternatives 2 p(p + 1) and δ = Σn = I + n−1/2 D, D = diag(d1 , . . . , dp ).

(13.16)

However, Muirhead (1982, p. 365) showed that if the sample came from an elliptical distribution, Ep (µ, Σ), with kurtosis 3k, then the asymptotic null distribution is kp d ∗ ˆ χ2 + χ2f −1 , −2 ln Λ /(1 + k) → 1 + 2(1 + k) 1 where χ21 and χ2f −1 are independently distributed and kˆ is a consistent estimate of k. A generalization to robust estimates of scale is proposed in Problem 13.6.16. Therefore, even the adjusted test statistic −2 ln Λ∗ /(1 + ˆ is not robust to non-normality of the data, especially for large values k) of k or long-tailed distribution. Moreover, the procedure of estimating k

234

13. Robustness

in the asymptotic distribution and calculating the critical points of the convolution of ˆ kp χ21 and χ2f −1 , 1+ ˆ 2(1 + k) as if kˆ were a constant, is obviously not a valid procedure. A new test statistic W is proposed, which is also invariant to orthogonal transformations and has an asymptotic null distribution χ2f for all underlying elliptical distributions with ﬁnite fourth moments. The asymptotic non-null distribution under the sequence of local alternatives (13.16) is noncentral chi-square. It is asymptotically fully eﬃcient at the normal distribution as compared to the modiﬁed likelihood ratio test. Let Sn = I + n−1/2 Un , Sn = (sij ), Un = (uij ). Then, when H0 is true, the asymptotic distribution of √ √ un = (u11 / 2, . . . , upp / 2, u12 , . . . , u1p , u23 , . . . , u2p , . . . , up−1,p ) ∈ Rf when the observations xi are drawn from an elliptical distribution with kurtosis 3k is Nf (0, Γ), where Ω 0 Γ= 0 (1 + k)If −p with Ω = (1+k)Ip + 12 k11 , 1 = (1, . . . , 1) ∈ Rp . Then, from Corollary 5.1, we have under H0 , u11 d . u2ij /(1 + k) → χ2f . un Γ−1 un = 12 (u11 , . . . , upp )Ω−1 .. + i<j upp The test statistic proposed [Bentler (1983)] is s11 − 1 n .. ˆ ˆ −1 W = (s11 − 1, . . . , spp − 1)Ω s2ij /(1 + k), + n . 2 i<j spp − 1 ˆ p + 1 k11 ˆ and kˆ is a consistent estimate of k. ˆ = (1 + k)I where Ω 2 Note that when kˆ ≡ 0, then W reduces to Nagao’s (1973) test statistic (n/2) tr(Sn − I)2 . Asymptotic expansions of Nagao’s test for elliptical distributions were derived by Purkayastha and Srivastava (1995). The test statistic W can be seen as a weighted form of Nagao’s statistic with the diagonal and oﬀ-diagonal elements of the sample variance matrix, Sn , being assigned diﬀerent weights.

13.5. Robust tests on scale matrices

235

The identity (13.15) leads by the method of moments to the consistent and orthogonally invariant estimate n ¯ |4 |xi − x (13.17) kˆ = pn ni=1 2 − 1. ¯ |2 ) ( i=1 |xi − x ˆ , it has a ˆ perturbed by a rank 1 matrix, namely 1 k11 ˆ is (1 + k)I Since Ω 2 known inverse which leads to the equivalent expression W =

kˆ n ˆ −n tr(Sn − I)2 /(1 + k) (tr Sn − p)2 , ˆ ˆ + kp] ˆ 2 2 (1 + k)[2(1 + k)

showing that W is invariant to orthogonal transformations, xi → Hxi for any orthogonal matrix H. Thus, without loss of generality, we can take for W the sequence of local alternatives (13.16) with a diagonal matrix D. The following result was given in Bilodeau (1997b). Proposition 13.17 Under the sequence of local alternatives Σn = I + n−1/2 D, D = diag(d1 , . . . , dp ), the asymptotic distribution of W is noncentral chi-square, W → χ2f (d Ω−1 d/4), d

where f

=

d = Ω =

1 2 p(p

+ 1), (d1 , . . . , dp ) , (1 + k)I + 12 k11 .

1/2

Proof. Let xi = Σn zi , where zi ∼ Ep (0, I). Also, let x1 z1 .. .. X = . and Z = . xn

zn

be the sample matrices, and Sn (X) and Sn (Z) be the sample variance matrices obtained from X and Z, respectively. Then, we have 1/2 1/2 Un (X) ≡ n1/2 [Sn (X) − I] = Σ1/2 (Σn − I), n Un (Z)Σn + n d

where Un (Z) → Npp (0, (1 + k)(I + K) + k vec(I)vec(I) ), Σn → I, and n1/2 (Σn − I) = D. Hence, the asymptotic result Un (X) → Npp (D, (1 + k)(I + K) + k vec(I)vec(I) ) d

ˆ is obtained. Since W is a continuous function of Un (X) and k, ˆ W = g(Un (X), k),

236

13. Robustness

the conclusion follows from Lemma 6.3 and classical results on quadratic forms if kˆ in (13.17) is consistent under the same sequence of local alterp 1/2 ¯ = Σn z ¯ and since z ¯ → 0, Σn → I, we natives. This is now shown. From x p ¯ → 0. Thus, the asymptotic equivalences have x 1 1 1 1 ¯ |2 ∼ ¯ |4 ∼ |xi − x |xi |2 , |xi − x |xi |4 , n i=1 n i=1 n i=1 n i=1 n

n

n

n

p

where u ∼ v means u − v → 0, hold. But now, since 1 1 1 |zi |j ≤ |xi |j ≤ (1 + n−1/2 d(p) )j/2 |zi |j , n i=1 n i=1 n i=1 n

(1 + n−1/2 d(1) )j/2

n

n

where d(1) = min{di } and d(p) = max{di }, we also have the equivalences 1 1 1 1 |xi |2 ∼ |zi |2 , |xi |4 ∼ |zi |4 . n i=1 n i=1 n i=1 n i=1 n

n

n

n

p Thus, 1 + kˆ → pE|zi |4 /E 2 |zi |2 = 1 + k, which completes the proof.

2

When k = 0, the test statistic W is asymptotically distributed, under the sequence of local alternatives (13.16), as χ2f (d d/4). Therefore, W is asymptotically fully eﬃcient at the normal distribution as compared to the modiﬁed likelihood ratio test, −2 ln Λ∗ . Sutradhar (1993) discusses the score test of the multivariate t. For testing the hypothesis H0 : µ = 0 and Σ = I against H1 : µ = 0 ¯ under the sequence of local or Σ = I, consider the test statistic W + n¯ x x alternatives µn = n−1/2 τ , Σn = I + n−1/2 D, where D = diag(d1 , . . . , dp ). Then, it can be established along the same lines ¯ → χ2f (δ), W + n¯ x x d

where f = 12 p(p + 3), δ = d Ω−1 d/4 + τ τ /2, and d and Ω are as in ¯ is thus robust in the class of elliptical Proposition 13.17. The test W +n¯ x x distributions with ﬁnite fourth moments. Its full eﬃciency at the normal distribution as compared to the likelihood ratio test follows immediately by comparing the asymptotic non-null distributions of the two tests [Khatri and Srivastava (1974)].

13.5.3

Relative eﬃciency of adjusted LRT

Under condition H, the adjusted LRT based on Sn has noncentrality parameter δh (Λ, B)/2(1 + k)

13.5. Robust tests on scale matrices

237

and for k moderately large, it is expected to have low power. A measure of eﬃciency can be derived by comparing the adjusted LRT with the exact LRT derived under a particular Ep (0, Λ) with known density deﬁned by the function g(·). The likelihood for Λ built from the sample matrix x1 .. X= . xn

is Lg (Λ) = |Λ|−n/2

n

g(xi Λ−1 xi ).

i=1

Then, ˆn Λ

=

arg min Lg (Λ),

ˆ ˆn Λ

=

arg min Lg (Λ)

Λ>0

h(Λ)=0

are respectively the restricted and unrestricted elliptical MLE of Λ. Then, the optimal procedure is the LRT derived for a given g(·): Lh,g (X) =

ˆ ˆ n) Lg (Λ . ˆ n) Lg (Λ

Then, Wald’s classical formulation for this “elliptical” LRT is ˆ n )] [Ch (Λ ˆ n )]−1 h(Λ ˆ n )/σ1,g −2 ln Lh,g (X) ∼ n[h(Λ under H0 or under the sequence of alternatives Λn = Λ + n−1/2 B. The parameter σ1,g is the value of σ1 in the asymptotic variance of the MLE given in Proposition 13.7, i.e., σ1,g = p(p + 2)/E[ψ 2 (s)]. Corollary 13.4 Assume condition H holds. Then: d

(i) under H0 , −2 ln Lh,g (X) → χ2q , (ii) under the sequence of contiguous alternatives Λn = Λ+n−1/2 B, where h(Λ) = 0 and B is a ﬁxed symmetric matrix, −2 ln Lh,g (X) → χ2q (δh (Λ, B)/2σ1,g ) , d

where, as before, δh (Λ, B) = [vec(B)] [h (Λ)] [Ch (Λ)]−1 h (Λ) vec(B). When g(·) is known, another test which is ﬁrst-order eﬃcient and asymptotically distributed as chi-square is the minimum geodesic distance test [Berkane et al. (1997)]. The proof of Corollary 13.4 is identical to that of Corollary 13.2. The ˆ to the asymptotic eﬃciency of the adjusted LRT −2 ln Lh (Sn )/(1 + k)

238

13. Robustness

q=1 q=2 q=3

ν =5

ν=6

ν=7

ν=8

ν = 30

.26 .37 .46

.17 .22 .27

.13 .17 .20

.11 .14 .16

.06 .06 .06

Table 13.2. Asymptotic signiﬁcance level of unadjusted LRT for α = 5%.

elliptical LRT can thus be measured by the ratio of the noncentrality parameters, i.e., σ1,g /(1 + k). For the tp,ν density, it was evaluated that σ1,g = 1 + 2/(p + ν), whereas 1 + k = (ν − 2)/(ν − 4) for a relative eﬃciency of (ν − 4)(ν + 2)/[(ν − 2)(ν + p)]. For p = 2 and ν = 5, this gives an eﬃciency of 33%. This is due to the poor robustness property of Sn . This adjusted LRT cannot really be thought of as a robust test because of its low eﬃciency. To obtain a truly robust adjusted LRT, one has to replace Sn by an eﬃcient robust estimate, i.e., one with a σ1 close to σ1,g . We conclude this analysis by guarding the practitioner against assuming indiscriminantly the normality of the data and using the “optimal” test for normality. If the data came from an elliptical distribution with a kurtosis parameter k and the hypothesis (satisfying condition H) was H0 : h(Λ) = 0, where h(Λ) ∈ Rq , then what was supposed to be an α = 5% signiﬁcance level test would be, in fact, for large samples, a test of signiﬁcance level: P (−2 ln Lh (Sn ) ≥ χ2.95,q )

P (−2 ln Lh (Sn )/(1 + k) ≥ χ2.95,q /(1 + k))

→ P χ2q ≥ χ2.95,q /(1 + k) . =

For a tp,ν distribution with 1 + k = (ν − 2)/(ν − 4), the signiﬁcance level may be far from 5%, as evidenced by Table 13.2. The situation worsens as ν decreases, which means the tails become heavier or q increases, which is related to the complexity of the hypothesis. For q = 3 and ν = 5, tossing a coin is nearly as reliable!

13.6 Problems 1. Demonstrate the following on normal mixture representation: d

|=

z, then (i) If x = w1/2 z, where w ∼ F , z ∼ Np (0, I), and w ∞ (2πw)−p/2 exp(− 12 w−1 x x)dF (w), fx (x) = 0

where F (·) is a distribution function on [0, ∞). (ii) If νw−1 ∼ χ2ν , then x = w1/2 z ∼ tp,ν has density

where cp,ν

fx (x) = cp,ν (1 + x x/ν)−(ν+p)/2 , x ∈ Rp ,

= (νπ)−p/2 Γ 12 (ν + p) /Γ 12 ν .

13.6. Problems

239

2. Assume x ∼ tp,ν (µ, Λ) where x is partitioned as x = (x1 , x2 ) , xi ∈ Rpi , i = 1, 2, p = p1 + p2 . Demonstrate the following: (i) E x = µ, var x = [ν/(ν − 2)]Λ, ν > 2. (ii) p−1 (x − µ) Λ−1 (x − µ) ∼ F (p, ν). (iii) The marginal distribution is x2 ∼ tp2 ,ν (µ2 , Λ22 ), where (µ , µ ) , 1 2 Λ11 Λ12 Λ = Λ21 Λ22 µ =

are partitioned in conformity. (iv) The conditional distribution is

x1 |x2 ∼ tp1 ,ν+p2 µ1 + Λ12 Λ−1 22 (x2 − µ2 ), h(x2 )Λ11.2 , where h(x2 ) = [ν/(ν + p2 )] · [1 + (x2 − µ2 ) Λ−1 22 (x2 − µ2 )/ν]. Determine E(x1 |x2 ), var(x1 |x2 ) and the condition for their existence. 3. Verify by diﬀerentiation of ln φ(t21 + t22 ) the cumulants = −2φ (0),

k2 k4

= =

k22

12(φ (0) − φ (0)2 ), 4(φ (0) − φ (0)2 ),

where φ(t21 + t22 ) is the characteristic function of a bivariate rotationally invariant vector. 4. Obtain E uu = p−1 I and var(uu ) =

1 2 (I + Kp ) − 2 vec(I)[vec(I)] , p(p + 2) p (p + 2)

where u ∼ unif(S p−1 ). 5. For the multivariate power exponential family (13.2), prove the following: (i) The normalizing constant is cp,α =

αΓ(p/2) π p/2 2p/2α Γ(p/2α)

.

(ii) The variance of x is var x =

21/α Γ[(p/2 + 1)/α] Λ. pΓ(p/2α)

(iii) For α = 1/2, verify the assertion in Example 13.4. Hint: Use the representation in polar coordinates in Proposition 4.10 together with Problems 4.6.13 and 13.6.4.

240

13. Robustness

ˆ = (Λ ˆ ij ) is aﬃne equivariant with variance 6. Check that when Λ ˆ = σ1 (I + Kp )(Λ ⊗ Λ) + σ2 vec(Λ)[vec(Λ)] , var Λ then ˆ ki , Λ ˆ lj ) = σ1 (Λij Λkl + Λkj Λil ) + σ2 Λki Λlj . cov(Λ 7. Prove if a0 > p + 1, n ≥ p + 1, then condition D1 is satisﬁed w.p.1 when sampling from an absolutely continuous distribution. 8. Verify conditions M1 through M4 for u1 and u2 corresponding to the MLE under the tp,ν distribution. 9. Verify conditions M1 through M4 for Huber’s ψ function in Example 13.10. 10. Assume z ∼ Ep (0, I) with density g(|z|2 ). Deﬁne u(s) = −2g (s)/g(s) and ψ(s) = su(s). Prove E ψ(|z|2 ) = p. Hint: Integrate by parts. 11. For the tp,ν distribution, verify that the asymptotic variance parameter σ1 of the elliptical MLE in Proposition 13.7 is σ1 = 1 + 2/(p + ν). 12. Prove 1 + k = pE(s2 )/[(p + 2)E 2 (s)], where s = |z|2 and z ∼ Ep (0, I) has fourth-order moments. 13. Deﬁne u(s) = −2g (s)/g(s) and ψ(s) = su(s). Let s = |z|2 , where z ∼ Ep (0, I). Derive an upper bound for σ1,g /(1 + k), the index of relative eﬃciency, by going through the following steps: (i) E ψ 2 (s) ≥ [E ψ(s)]2 , (ii) E(s2 ) E[ψ 2 (s)] ≥ [Esψ(s)]2 = (p + 2)2 E 2 (s), (iii) σ1,g /(1 + k) ≤ min{1, (1 + 2p−1 )(1 + k)−1 }. (iv) Interpret the bound in (iii). 14. Demonstrate that if ρ(t) = t2 and b0 = p in the deﬁnition of the S estimate, then the solution is the normal MLE. 15. Demonstrate that the S estimate (µn , Vn ) is necessarily a solution of the equations ave [u(ti )(xi − µn )] = 0 ave [pu(ti )(xi − µn )(xi − µn ) − v(ti )Vn ] = 0, 1/2 where ti = (xi − µn ) Vn−1 (xi − µn ) and v(t) = tψ(t) − ρ(t) + b0 . M and S estimates are close relatives! 16. Test for a given variance. This is a continuation of Problem 8.9.8. The LRT under the Np (0, Λ) for H0 : Λ = I versus H1 : Λ = I is given by Lh (Sn ) = epn/2 |Sn |n/2 etr(− 12 nSn ),

13.6. Problems

241

n where Sn = n1 i=1 xi xi . Suppose that, in fact, xi ∼ Ep (0, Λ) and ˆ n satisfying condition E instead one decides to use a robust estimate Λ of Sn . Then, demonstrate that under H0 ,

where χ21

|=

d ˆ n) → (σ1 + 12 σ2 p)χ21 + σ1 χ2p(p+1)/2−1 , −2 ln Lh (Λ

χ2p(p+1)/2−1 by following these steps:

ˆ n = I + n−1/2 Zn and use the Taylor series for ln |I + tA| (i) Write Λ around t = 0 to show that under H0 , d 1 ˆ n) → −2 ln Lh (Λ 2 [vec(Z)] vec(Z),

where Z ∼ Npp (0, σ1 (I + Kp ) + σ2 vec(I)[vec(I)] ). (ii) Demonstrate √ √ var(z11 / 2, . . . , zpp / 2, z12 , . . . , z1p , z23 , . . . , z2p , . . . , zp−1,p ) is given by

σ1 Ip + 12 σ2 11 0

0 σ1 Ip(p−1)/2

≡ Ω.

(iii) Verify the eigenvalues of Ω are σ1 of multiplicity 12 p(p + 1) − 1 and σ1 + 12 σ2 p of multiplicity 1. 17. Test of multiple correlation. The LRT under (x1 , x2 ) ∼ Np (0, Λ) for H0 : R2 = 0 versus H1 : R2 = 0 is given by

where Sn = the partition

1 n

n

ˆ 2 (Sn ))n/2 , Lh (Sn ) = (1 − R

i=1

ˆ 2 (Sn ) = s S−1 s21 /s11 in terms of xi xi and R 21 22 Sn =

s11 s21

s21 S22

.

Suppose, in fact, that (x1 , x2 ) ∼ Ep (0, Λ) and one decides to use ˆ n satisfying condition E instead of Sn . Then, a robust estimate Λ demonstrate that under H0 , d ˆ n) → σ1 χ2p−1 , −2 ln Lh (Λ

by following these steps: (i) Argue that one can assume Λ = I. (ii) Using a Taylor series, prove that ˆ n ) ∼ nR ˆ n ). ˆ 2 (Λ −2 ln Lh (Λ d ˆ n) → ˆ 2 (Λ z z, where z ∼ Np−1 (0, σ1 I) to (iii) Finally, prove that nR conclude.

242

13. Robustness

n 18. Let l = (l1 , . . . , lp ) be the eigenvalues of Sn = n1 i=1 xi xi , calculated from a sample of an Ep (0, Λ) distribution, Λ = diag(λ1 , . . . , λp ). If the population eigenvalues λα are all distinct, then prove that the joint limiting distribution is given by n1/2 Λ−1 (l − λ) → Np (0, 2(1 + k)I + k11 ). d

19. Suppose the sample is taken from an elliptical distribution with kurtosis 3k. Let f = (f1 , . . . , fp ) be the eigenvalues of the sample correlation matrix R = (rij ). If the eigenvalues γα of the population correlation matrix ρ = (ρij ) = G diag(γ1 , . . . , γp ) G , where G = (gij ) ∈ Op , are all distinct, then prove that the joint limiting distribution is n1/2 (f − γ) → Np (0, (1 + k)Ω), d

where Ω = (ωαβ ) is given by ωαβ = 2γα γβ δαβ − (γα + γβ )

p j=1

2 2 gjα gjβ +

p p j=1 i=1

2 2 2 gjα giβ ρji .

14 Bootstrap conﬁdence regions and tests

An important part of multivariate analysis deals with conﬁdence regions and tests of hypotheses on the mean vector and variance matrix. The classical theoretical developments for such procedures rest mainly upon the multivariate normality assumption. Without multivariate normality, the asymptotic distribution of many tests becomes more complex and often leads to untabulated limit distributions. The bootstrap conﬁdence regions and tests on the mean vector and variance matrix have the desired asymptotic levels under very mild conditions. We will present the bootstrap technique main ideas without formal proofs. The interested reader should consult the cited references. General references for the bootstrap are Efron (1982), who made the technique widely applicable by using modern computational power, Efron and Tibshirani (1993) and Hall (1992). The book by Davison and Hinkley (1997) has S-plus code which may prove useful.

14.1 Conﬁdence regions and tests for the mean Let x = (x1 , . . . , xp ) ∼ F with mean µF = (µF,i ) and variance ΣF = (σF,ij ). Let x1 , . . . , xn be i.i.d. F and ¯n x

=

1 xi , n i=1

Sn

=

1 ¯ n )(xi − x ¯ n ) , (xi − x (n − 1) i=1

n

n

244

14. Bootstrap conﬁdence regions and tests

be the sample mean and sample variance, respectively. Deﬁne the “pivot” xn − µF )|. wn = n1/2 |Sn−1/2 (¯ Then, by the central limit theorem, d

wn → |z|, z ∼ Np (0, I). The empirical distribution function of the data x1 , . . . , xn is denoted by 1 I(xi ≤ t), Fˆn (t) = n i=1 n

where I(xi ≤ t) is the indicator function. In other words, Fˆn is the discrete distribution function with equal probability 1/n at the points xi , i = 1, . . . , n. Then, for x∗ ∼ Fˆn , we have E x∗

= µFˆn =

1 xi , n i=1

var x∗

= ΣFˆn =

1 ¯ n )(xi − x ¯ n ) . (xi − x n i=1

n

n

The nonparametric bootstrap estimate of the probability distribution of wn under F , Jn (F ), is the bootstrap estimate Jn (Fˆn ), which can be interpreted ¯ ∗n and S∗n be the sample as follows. Let x∗i , i = 1, . . . , n, be i.i.d. Fˆn , and let x ∗ mean and sample variance, respectively, of the xi ’s. Then, Jn (Fˆn ) is the probability law under Fˆn of x∗n − µFˆn )|. wn∗ = n1/2 |S∗n −1/2 (¯ In practice, Jn (Fˆn ) may be approximated to any degree of accuracy with resampling by Monte Carlo methods. The consistency of the bootstrap was established by Beran (1984, example 3) using a triangular array version of the C.L.T., wn∗ → |z| w.p.1, d

which means that wn and wn∗ converge in distribution to the same limit. However, it was Singh (1981), and Bickel and Freedman (1981) who ﬁrst established the consistency of the bootstrap in the univariate situation. Let cn (α, Fˆn ) be a (1 − α)-quantile of the bootstrap distribution Jn (Fˆn ). By the consistency of the bootstrap, if Dn (α) is the bootstrap conﬁdence region for µF , (¯ xn − µF )| ≤ cn (α, Fˆn )}, Dn (α) = {µF : n1/2 |S−1/2 n then lim P (µF ∈ Dn (α)) = 1 − α.

n→∞

14.1. Conﬁdence regions and tests for the mean

245

The bootstrap conﬁdence region Dn (α) handles all norms, | · |, on Rp with equal ease. Most often though, the euclidian norm is intended, and in that case, the ellipsoidal bootstrap conﬁdence region can be written as Dn (α) = {µF : n(¯ xn − µF ) S−1 xn − µF ) ≤ c2n (α, Fˆn )}. n (¯ A (1−α)-acceptance region, A(µ0 ), for testing the hypothesis H0 : µF = µ0 against H1 : µF = µ0 may be obtained by inverting the conﬁdence region, Dn (α), in the usual way [Fraser (1976), p. 580]. Here, the test which rejects H0 : µF = µ0 iﬀ µ0 ∈ Dn (α) is a test with asymptotic type I error probability α. More generally, suppose a conﬁdence region on g(µF ) ∈ Rk , k ≤ p, is wanted where g : Rp → Rk is a continuously diﬀerentiable function and has ﬁrst derivative g˙ ∈ Rkp . Let u : Rk → R be continuous on Rk such that {z ∈ Rk : u(z) = c} has Lebesgue measure 0 for every c ∈ R. Consider the statistic ! wn,g = u n1/2 (g(¯ xn ) − g(µF )) . The central limit theorem coupled with the delta method yields d

˙ F ) zF ], zF ∼ Np (0, ΣF ). wn,g → u[g(µ The condition imposed on u ensures that the limit distribution is continuous. Using arguments as in Beran (1984), it can be established that the bootstrap estimate is consistent, i.e., ! d ∗ ˙ F ) zF ] w.p.1. wn,g = u n1/2 g(¯ x∗n ) − g(µFˆn ) → u[g(µ To construct the bootstrap conﬁdence region, let cn,g (α, Fˆn ) be a (1 − α)quantile of the bootstrap distribution. A bootstrap conﬁdence region for g(µF ) having asymptotic coverage probability 1 − α is ! Dn,g (α) = {g(µF ) : u n1/2 (g(¯ xn ) − g(µF )) ≤ cn,g (α, Fˆn )}. In the examples to be considred, the function u has the additional property, u(bz) = bu(z), ∀z ∈ Rk , ∀b > 0. Then, the factor n1/2 may be omitted and we may write equivalently Dn,g (α) = {g(µF ) : u [g(¯ xn ) − g(µF )] ≤ c∗n,g (α)}, where c∗n,g (α) is a (1 − α)-quantile of the distribution of u[g(¯ x∗n ) − g(µFˆn )] ∗ ¯ n is the bootstrap sample mean. when Fˆn is ﬁxed at its realized value and x Judicious choices of u and g give interesting conﬁdence regions, as the following examples show.

246

14. Bootstrap conﬁdence regions and tests

p Example 14.1 Let g(µF ) = µF and u(z) = |z|1 = i=1 |zi | be the l1 norm. Then, the bootstrap conﬁdence region p ∗ Dn,g (α) = µF : |¯ xn,i − µF,i | ≤ cn,g (α) , i=1

has asymptotic coverage probability 1−α, where c∗n,g (α) is a (1−α)-quantile p of the distribution of i=1 |¯ x∗n,i −µFˆn ,i | when Fˆn is ﬁxed at its realized value ∗ ¯ n is the bootstrap sample mean. and x Example 14.2 Let g(µF ) = µF and u(z) = |z|∞ = max1≤i≤p |zi | be the l∞ -norm. The bootstrap simultaneous conﬁdence intervals Dn,g (α) = {µF : |¯ xn,i − µF,i | ≤ c∗n,g (α), i = 1, . . . , p}, have asymptotic simultaneous coverage probability 1 − α, where c∗n,g (α) is a (1 − α)-quantile of the distribution of max1≤i≤p |¯ x∗n,i − µFˆn ,i | when Fˆn is ∗ ¯ n is the bootstrap sample mean. ﬁxed at its realized value and x Example 14.3 This example provides the bootstrap algorithm, easy to implement on a computer, to construct simultaneous conﬁdence intervals on the means µF,i , i = 1, . . . , p. We are given a sample x1 , . . . , xn from F . Bootstrap algorithm • • • •

¯ n = (¯ Calculate x xn,i ). b←1 B ← 2000 (say) Do while b ≤ B. • Draw a bootstrap sample x∗1 , . . . , x∗n from Fˆn . ¯ ∗n = (¯ • Calculate x x∗n,i ). x∗n,i − x ¯n,i | • ub ← max1≤i≤p |¯ • b←b+1 • End. • Order the ub ’s: u(1) ≤ u(2) ≤ · · · ≤ u(B) . • q ← (1 − α)B! ( ·! is the integer part function) • Simultaneous conﬁdence intervals for µF,i with approximate simultaneous coverage probability 1 − α are x ¯n,i − u(q) ≤ µF,i ≤ x ¯n,i + u(q) , i = 1, . . . , p.

14.2 Conﬁdence regions for the variance This time, deﬁne Wn = n1/2 (Sn − ΣF ).

14.2. Conﬁdence regions for the variance

247

The asymptotic distribution of Sn was given in Section 6.3 for all underlying distribution F with ﬁnite fourth moments. The asymptotic distribution is d

Wn → XF , XF ∼ Npp (0, ΩF ), where XF = (xF,ij ), the elements of ΩF given by ijkl kl ij il jk + k11 k11 + k11 k11 cov(xF,ik , xF,jl ) = k1111

with the k’s representing the cumulants of F . The nonparametric bootstrap estimate of the probability distribution under F of Wn is the probability distribution under Fˆn of Wn∗ = n1/2 S∗n − ΣFˆn . Beran and Srivastava (1985) established the consistency of the bootstrap Wn∗ → XF w.p.1. d

A diﬃculty in deriving a conﬁdence region for a function of ΣF is the redundancy of elements due to the symmetry of ΣF . So let uvec(S) = (s11 , s12 , s22 , . . . , s1p , s2p , . . . , spp ) be the vec operator applied only to the upper triangular part of S ∈ Sp . Suppose a conﬁdence region for g(ΣF ) ∈ Rk is desired where g is a function of uvec(ΣF ), which is continuously diﬀerentiable and has ﬁrst derivative g˙ ∈ Rkp(p+1)/2 . Let u : Rk → R be continuous on Rk such that {z ∈ Rk : u(z) = c} has Lebesgue measure 0 for every c ∈ R and u(bz) = bu(z), ∀z ∈ Rk , ∀b > 0. Let ! Wn,g = u n1/2 (g(Sn ) − g(ΣF )) . The delta method (v. Proposition 6.2) immediately yields d

˙ F ) uvec(XF )] Wn,g → u[g(Σ

(14.1)

and ∗ ˙ F ) uvec(XF )] w.p.1, Wn,g → u[g(Σ d

so the bootstrap is consistent [Beran and Srivastava (1985)]. The condition on u implies the limiting distribution in (14.1) is continuous. A bootstrap conﬁdence region for g(ΣF ) having asymptotic coverage probability 1 − α is Dn,g (α) = {g(ΣF ) : u [g(Sn ) − g(ΣF )] ≤ c∗n,g (α)}, where c∗n,g (α) is a (1 − α)-quantile of the distribution of u[g(S∗n ) − g(ΣFˆn )] when Fˆn is ﬁxed at its realized value and S∗n is the bootstrap sample variance.

248

14. Bootstrap conﬁdence regions and tests

Example 14.4 Let g(ΣF ) = ρF,ij =

σF,ij . 1/2 1/2 σF,ii σF,jj

Then, g(Sn ) is the sample correlation coeﬃcient rn,ij and bootstrap conﬁdence regions based on |rn,ij − ρF,ij | have the correct asymptotic coverage probability. Also covered is the Fisher z-transform 1 + ρF,ij 1 = tanh−1 (ρF,ij ). g(ΣF ) = 2 ln 1 − ρF,ij Example 14.5 This example provides the bootstrap algorithm, in an easily programmable form, to construct a conﬁdence interval for the correlation coeﬃcient ρF,ij using the Fisher z-transformation to stabilize the variance. We are given a sample x1 , . . . , xn from F . Bootstrap algorithm • • • • •

Calculate Sn = (sn,ij ). 1/2 1/2 Calculate rn,ij = sn,ij /[sn,ii sn,jj ]. b←1 B ← 2000 (say) Do while b ≤ B. • Draw a bootstrap sample x∗1 , . . . , x∗n from Fˆn . • Calculate S∗n = (s∗n,ij ). 1/2

• • • •

1/2

∗ • Calculate rn,ij = s∗n,ij /[s∗n,ii s∗n,jj ]. −1 ∗ • ub ← tanh (rn,ij ) − tanh−1 (rn,ij ) • b←b+1 End. Order the ub ’s: u(1) ≤ u(2) ≤ · · · ≤ u(B) . q ← (1 − α)B! An approximate (1 − α) conﬁdence interval for ρF,ij is

tanh[tanh−1 (rn,ij ) − u(q) ] ≤ ρF,ij ≤ tanh[tanh−1 (rn,ij ) + u(q) ]. Example 14.6 Let φ1 (ΣF ) > φ2 (ΣF ) > · · · > φp (ΣF ) > 0 be the ordered eigenvalues of ΣF assumed distinct. The vector φ(ΣF ) = (φ1 (ΣF ), . . . , φp (ΣF )) is a continuously diﬀerentiable function of uvec(ΣF ) [Kato (1982), Section 6 of Chapter 2]. The ordered sample eigenvalues are φ(Sn ) = (φ1 (Sn ), . . . , φp (Sn )) . The bootstrap conﬁdence region based on max | ln φi (Sn ) − ln φi (ΣF )|

1≤i≤p

(14.2)

14.3. Tests on the variance

249

has the correct asymptotic coverage probability. Here, u(z) = max1≤i≤p |zi |, z ∈ Rp . The logarithmic transformation stabilizes the variance in the normal model asymptotic for sample eigenvalues (v. Problem 8.9.15). The bootstrap conﬁdence region for φ(ΣF ) corresponding to (14.2) is {φi (ΣF ) : φi (Sn )/An ≤ φi (ΣF ) ≤ φi (Sn )An , i = 1, . . . , p}, where ∗

An = ecn,g (α) and c∗n,g (α) is a (1 − α)-quantile of the distribution of max | ln φi (S∗n ) − ln φi (ΣFˆn )|

1≤i≤p

when Fˆn is ﬁxed at its realized value and S∗n is the bootstrap sample variance. The problem of eﬃciently bootstrapping sample eigenvalues when ΣF may have multiple eigenvalues is still an unresolved problem [Beran and Srivastava (1987), Eaton and Tyler (1991)].

14.3 Tests on the variance Rather than inverting a conﬁdence region, it is sometimes possible to construct bootstrap tests directly from test statistics. This approach [Beran and Srivastava (1985)] to testing structural hypotheses about ΣF is the subject of this section. Assume x1 , . . . , xn are i.i.d. F with ﬁnite fourth moments. Let π : Pp → Pp be a linear projection (π 2 = π), not the identity map. Suppose Tn (Sn ) = n h(Sn ) is a test statistic for the null hypothesis, H0 : ΣF = π(ΣF ). Let F0 be any distribution function satisfying H0 . Example 14.7 Deﬁne the constant linear projection π(ΣF ) = I. Then, the hypothesis H0 : ΣF = I is equivalent to H0 : ΣF = π(ΣF ). Example 14.8 Partition ΣF as σF,11 ΣF = σ F,21 and deﬁne the linear projection π(ΣF ) =

σF,11 0

σ F,21 ΣF,22

0 ΣF,22

.

The hypothesis on the multiple correlation, H0 : R = 0, or H0 : σ 12 = 0 is equivalent to H0 : ΣF = π(ΣF ).

250

14. Bootstrap conﬁdence regions and tests

p Example 14.9 Deﬁne the linear map π(ΣF ) = ( i=1 σF,ii /p) I. Then, the sphericity hypothesis H0 : ΣF = γI, γ > 0 is equivalent to H0 : ΣF = π(ΣF ). The function h deﬁning the test statistic Tn (Sn ) is twice continuously diﬀer˙ F ) = 0. This entiable at uvec(ΣF0 ) ∈ Rp(p+1)/2 , with h(ΣF0 ) = 0 and h(Σ 0 formulation includes the normal model likelihood ratio test in particular. d ¨ ∈ Rp(p+1)/2 denote the second derivative of h and xF = Let h uvec(XF0 ). 0 p(p+1)/2 Then, using the Taylor series, ¨ F )xF . Tn (Sn )|F0 → xF0 h(Σ 0 0 d

We can construct a bootstrap estimate for the null distribution of Tn (Sn ) as follows. Let −1/2

Vn,F = [π(ΣF )]1/2 ΣF

−1/2

Sn ΣF

[π(ΣF )]1/2 .

The bootstrap estimate for the null distribution of Tn (Sn ) is deﬁned to be that of Tn (Vn,Fˆn ). Let dn,h (α, Fˆn ) be a (1 − α)-quantile of Tn (Vn,Fˆn ). Beran and Srivastava (1985) established the consistency of the bootstrap, ¨ F )xF w.p.1. Tn (Vn,Fˆn ) → xF0 h(Σ 0 0 d

Hence, the test which rejects H0 whenever Tn (Sn ) > dn,h (α, Fˆn ) has ¨ F ) = 0. asymptotic size α, provided h(Σ 0 In practice the bootstrap null distribution can be constructed as follows. Let yi = [π(Sn )]1/2 Sn−1/2 xi , i = 1, . . . , n. Let Fˆn,y be the empirical distribution function of the yi ’s. Note that ΣFˆn,y = π(ΣFˆn ), which satisﬁes H0 since π = π 2 . If y1∗ , . . . , yn∗ are d

i.i.d. Fˆn,y and S∗n,y is the sample variance of the yi∗ ’s, then Tn (Vn,Fˆn ) = Tn (S∗n,y ) whose distribution can be approximated by Monte Carlo methods. Example 14.10 We wish to test the hypothesis H0 : ΣF,12 = 0 using the invariant test statistic (v. Section 11.3) Tn

−1 = n tr[Sn,12 S−1 n,22 Sn,21 Sn,11 ]

= n

p1

2 rn,i ,

i=1 2 rn,i

where are the squared sample canonical correlations. The linear projection in this case is deﬁned by Sn,11 0 π(Sn ) = 0 Sn,22

14.3. Tests on the variance

with its square root

1/2

[π(Sn )]

=

1/2

0

0

Sn,11

251

1/2

Sn,22

.

We are given a sample x1 , . . . , xn from F . Bootstrap algorithm • Calculate Sn and partition

Sn =

Sn,11 Sn,21

1/2

1/2

Sn,12 Sn,22

. 1/2

−1/2

• Calculate the square roots Sn , Sn,11 , and Sn,22 and the inverse Sn • • • •

.

−1/2 [π(Sn )]1/2 Sn xi ,

i = 1, . . . , n. Transform yi = b←1 B ← 2000 (say) Do while b ≤ B. • Draw a bootstrap sample y1∗ , . . . , yn∗ from Fˆn,y . • Calculate S∗n and partition ∗ Sn,11 S∗n,12 . S∗n = S∗n,21 S∗n,22 −1

−1

• ub ← n tr[S∗n,12 S∗n,22 S∗n,21 S∗n,11 ] • b←b+1 • End. • Order the ub ’s: u(1) ≤ u(2) ≤ · · · ≤ u(B) . • q ← (1 − α)B! • An approximate size α test rejects H0 : ΣF,12 = 0 whenever Tn > u(q) . It is an easy matter to modify this bootstrap algorithm to bootstrap the test statistic

! −1 −1 −1 −1 Tn = n tr Sn,12 S−1 n,22 Sn,21 Sn,11 I − Sn,12 Sn,22 Sn,21 Sn,11 = n

p1

2 2 rn,i /(1 − rn,i ).

i=1

However, the test based on the largest sample canonical correlation, Tn = 2 , should not be bootstraped unless the user is sure the largest populan rn,1 tion canonical correlation is distinct. In case of multiplicity the population canonical correlations are not a diﬀerentiable function of ΣF [Kato (1982), Section 6 of Chapter 2]. Bootstrap algorithms for estimating the power function of a test statistic can be found in Beran (1986). Nagao and Srivastava (1992) considered high-order asymptotic expansions to the distribution of some test criteria on the variance matrix under local alternatives. For the test of sphericity

252

14. Bootstrap conﬁdence regions and tests

in dimension p = 3, they compared these expansions to the bootstrap approximations for both the normal model likelihood ratio test and Nagao’s test when the distribution is actually multivariate normal or multivariate t.

14.4 Problem 1. John (1971) showed that the test based on J = tr V2 /(tr V)2 ,

n ¯ )(xi − x ¯ ) , is LBI for the hypothesis of sphericwhere V = i=1 (xi − x ity, H0 : ΣF = γI, γ > 0, when the underlying distribution F is multivariate normal. Write down a detailed bootstrap algorithm to evaluate the α critical point of the test J but when F is multivariate student, tp,ν (0, I). Hint: A tp,ν (0, I) distribution can be simulated with Problem 13.6.1.

Appendix A

Assume x ∼ F , y ∼ G, x given by H(t)

|=

Inversion formulas

y on Rn . Then, z = x + y has a d.f., z ∼ H,

= P (x + y ≤ t) = E P (x + y ≤ t|y) = E F (t − y) = F (t − y)dG(y). Rn

Similarly, inverting the roles of x and y, we also have G(t − x)dF (x). H(t) = E G(t − x) = Rn

This leads to the smoothing lemma on convolution. |=

Lemma A.1 (Smoothing lemma) If x is absolutely continuous with y, is absolutely p.d.f. f (t), then z = x + y, where y ∼ G and x continuous with p.d.f. h(t) = E f (t − y). Proof. It follows readily that F (t − y)dG(y) H(t) = Rn = Rn

(−∞,t−y]

f (x)dx dG(y)

254

Appendix A. Inversion formulas

= Rn

(−∞,t]

f (x − y)dx dG(y).

By Tonelli’s theorem, it is posible to interchange the order of integration whereby H(t) = f (x − y)dG(y) dx = E f (x − y)dx. (−∞,t] Rn (−∞,t] 2 We can now establish the inversion formula on Rn . The proof resembles that of Feller (1966, p. 480) for n = 1. Proposition A.1 (Inversion formula) The probability measure Px is given in terms of the characteristic function c(t) = cx (t) by 2 1 Px (a, b] = lim e−it x c(t)e−t t/2N dtdx, n N →∞ (2π) (a,b] Rn

Proof. Take any random t such that t

|=

∀a, b such that Px (∂(a, b]) = 0. x. Then, conditioning yields

E eix t = E E(eix t |x) = E ct (x) = E cx (t). Replace x by x − s for any ﬁxed value of s to ﬁnd Parseval’s relation:

E ct (x − s) = E e−is t cx (t). However, letting t ∼ Nn (0, σ −2 I) with ct (s) = exp(−|s|2 /2σ 2 ), E exp(−|s − x|2 /2σ 2 )

= E e−it s c(t) 2 n/2

σ = e−it s c(t) exp −σ 2 |t|2 /2 dt. 2π Rn

Divide by (2πσ 2 )n/2 to obtain E

1 1 exp(−|s−x|2 /2σ 2 ) = 2 n/2 (2π)n (2πσ )

Rn

e−it s c(t) exp −σ 2 |t|2 /2 dt.

This is of the form E g(s − x) = h(s) in the smoothing lemma where g(s) is the p.d.f. for a Nn (0, σ 2 I). Thus, h(s) is the p.d.f. of x + σz, where z ∼ Nn (0, I), and if we let Pσ denote the probability measure for x + σz, 2 2 1 Pσ (a, b] = e−it s c(t)e−σ |t| /2 dtds, (2π)n (a,b] Rn whereby Slutsky’s theorem with σ = 1/N gives the result.

2

An immediate corollary is the inversion formula for absolutely continuous distribution.

Appendix A. Inversion formulas

255

Corollary A.1 If c(t) is integrable with respect to Lebesgue measure, then 1 f (x) = e−it x c(t)dt. n (2π) Rn Proof. If c(t) is integrable, then the integrand in Proposition A.1 is dominated by an integrable function. By the D.C.T., we can interchange the limit and the integral, which gives the result. 2

Appendix B Multivariate cumulants

B.1 Deﬁnition and properties The moments of a univariate random variable x, µr = E xr , are the coeﬃcients of (it)r /r! in the Taylor series of the characteristic function, cx (t) =

∞

µr (it)r /r!

r=0

whereas the cumulants are the coeﬃcients in the series for Kx (t) ≡ ln[cx (t)], Kx (t) =

∞

kr (it)r /r!,

r=0

provided the expansions are valid. The function Kx (t) is the cumulant generating function. Relations between moments and cumulants are thus obtained by equating the coeﬃcients in the Taylor series of exp(·) in the equation ∞ ∞ r r µr (it) /r! = exp kr (it) /r! . r=0

r=0

We require only the relations between the ﬁrst four moments and cumulants (assuming they exist): µ1 µ2

= k1 , = k2 + k12 ,

B.1. Deﬁnition and properties

µ3 µ4

= k3 + 3k2 k1 + k13 ,

k1

= µ1 , = µ2 − µ21 ,

k2 k3 k4

257

= k4 + 4k3 k1 + 3k22 + 6k2 k12 + k14 ,

= µ3 − 3µ2 µ1 + 2µ31 , = µ4 − 4µ3 µ1 − 3µ22 + 12µ2 µ21 − 6µ41 .

When x is centered, i.e., E x = µ1 = k1 = 0, these simplify to µ2 = k2 , µ3 = k3 , µ4 = k4 + 3k22 ,

k2 = µ2 , k3 = µ3 , k4 = µ4 − 3µ22 . r

r1 p For a random vector x ∈ Rp , product-moments µr1 ,...,r pp = E(x1 · · · xp ) and multivariate cumulants kr1 ,...,rp of order r = i=1 ri are deﬁned similarly,

cx (t)

∞

=

µr1 ,...,rp

(itp )rp (it1 )r1 ··· , r1 ! rp !

kr1 ,...,rp

(itp )rp (it1 )r1 ··· . r1 ! rp !

r1 ,...,rp =0

Kx (t) = ln[cx (t)]

∞

=

r1 ,...,rp =0

Example B.1 For x ∼ Np (µ, Σ), we have Kx (t) = it µ − 12 t Σt, a quadratic function of t, and, thus, all multivariate cumulants of order r > 2 are null. Multivariate cumulants of order 1 are the means, µi , and those of order 2 are the variances, σii , and covariances, σij . Obtaining product-moments in terms of cumulants, and vice versa, is a laborious task which can be greatly simpliﬁed with a “symbolic diﬀerential operator” [Kendall et al. (1987)]. For example, when E x = 0, consider the relation µ4 = k4 + 3k22 , which we write symbolically as µ(r14 ) = k(r14 ) + 3k 2 (r12 ). To obtain a relation between fourth-order product-moments and cumulants of a bivariate distribution, consider the operator r2 ∂(·)/∂r1 . When applied to µ(r14 ), it yields 4µ(r13 r2 ) = 4k(r13 r2 ) + 12k(r12 )k(r1 r2 ), which means, after dividing by 4, µ31 = k31 + 3k20 k11 . Example B.2 The same method can be used to obtain cumulants in terms of product-moments. Considering the relation k4 = µ4 − 3µ22

258

Appendix B. Multivariate cumulants

in symbolic form k(r14 ) = µ(r14 ) − 3µ2 (r12 ), and applying the operator r2 ∂(·)/∂r1 , we get 4k(r13 r2 ) = 4µ(r13 r2 ) − 12µ(r12 )µ(r1 r2 ) or k31 = µ31 − 3µ20 µ11 . Continuing this process, it is possible to obtain relations for trivariate distributions with either operator, r3 ∂(·)/∂r1 or r3 ∂(·)/∂r2 . The operator r3 ∂(·)/∂r1 applied to the last symbolic equation yields 12µ(r12 r2 r3 ) = 12k(r12 r2 r3 ) + 24k(r1 r3 )k(r1 r2 ) + 12k(r12 )k(r2 r3 ), which is equivalent to µ211 = k211 + 2k101 k110 + k200 k011 . The operator r4 ∂(·)/∂r1 ﬁnally gives the relation for fourth-order productmoments and cumulants of a four-dimensional distribution 24µ(r1 r2 r3 r4 ) = 24k(r1 r2 r3 r4 ) + 24k(r3 r4 )k(r1 r2 ) + 24k(r1 r3 )k(r2 r4 ) +24k(r1 r4 )k(r2 r3 ), or µ1111 = k1111 + k0011 k1100 + k1010 k0101 + k1001 k0110 . For fourth-order product-moments of a p-dimensional, p > 4, distribution, we need only specify which four variables enter. For example, µijkl 1111 = E(xi xj xk xl ) satisﬁes ijkl ijkl ijkl ijkl ijkl ijkl ijkl µijkl 1111 = k1111 + k0011 k1100 + k1010 k0101 + k1001 k0110 .

A zero subscript means the superscript variable does not enter, so we can rewrite ijkl kl ij ik jl il jk µijkl 1111 = k1111 + k11 k11 + k11 k11 + k11 k11 .

When a variable is repeated, the indices can be amalgamated. For example, the equation where i = j, iikl kl ii ik il il ik µiikl 1111 = k1111 + k11 k11 + k11 k11 + k11 k11 ,

becomes ikl kl i ik il il ik µikl 211 = k211 + k11 k2 + k11 k11 + k11 k11 ,

and if i = j = k = l, then we recover the initial equation µi4 = k4i + 3(k2i )2 .

B.2. Application to asymptotic distributions

259

Departures from normality is often assessed with the coeﬃcients of skewness, γ1 , and kurtosis, γ2 . For a centered variable x, they are deﬁned as µ3 k3 = 3/2 , γ1 = 3/2 µ2 k2 µ4 k4 γ2 = − 3 = 2. 2 µ2 k2 For a normal variable, γ1 = γ2 = 0. Cumulants of random symmetric matrices can also be deﬁned. For a description of miminal moments and cumulants of symmetric matrices with an application to the Wishart distribution, the reader is referred to Kollo and von Rosen (1995).

B.2 Application to asymptotic distributions Let x1 , . . . , xn i.i.d. x ∈ Rp which has ﬁnite fourth-order moments and Ex = 0 and var x = Σ. The asymptotic distribution of S = n 1 ¯ )(xi − x ¯ ) was derived generally in Section 6.3: i=1 (xi − x (n−1) n1/2 (S − Σ) → Npp (0, var W), d

where W = xx . The only problem is to calculate var W. This can now be done in terms of multivariate cumulants. The block (i, j) of var W is E(xi xj xx ) − E(xi x)E(xj x ) and the element (k, l) of the block (i, j) becomes E(xi xj xk xl ) − E(xi xk )E(xj xl )

ik jl = µijkl 1111 − µ11 µ11 ijkl kl ij il jk = k1111 + k11 k11 + k11 k11 .

The general result is thus ijkl kl ij il jk + k11 k11 + k11 k11 . cov(wik , wjl ) = k1111

B.3 Problems 1. Establish the following: (i) µ11 = k11 and µ21 = k21 . 2 . (ii) µ22 = k22 + k20 k02 + 2k11 (iii) Given µ5 = k5 + 10k3 k2 , calculate µ32 and µ41 . (iv) Obtain µ301 in terms of lower-order cumulants.

260

Appendix B. Multivariate cumulants

2. Demonstrate the kurtosis γ2 of a symmetric contaminated normal density (1 − )(2π)−1/2 exp(− 12 x2 ) + (2πσ)−1/2 exp(− 12 x2 /σ 2 ) is γ2 = 3

[1 + (σ 4 − 1)] − 3. [1 + (σ 2 − 1)]2

3. Evaluate the kurtosis of a Student’s tν distribution as γ2 = 6/(ν − 4), ν > 4.

Appendix C S-plus functions

This appendix describes three S-plus programs which the reader can download from the World Wide Web site www.dms.umontreal.ca/∼bilodeau. Simply download the ﬁle named multivariate and, at the S-plus prompt, type: source(“multivariate”) to compile the functions. 1. U (p; m, n) distribution function. Usage: pu(ζ, p, m, n) Value: The function returns P (U (p; m, n) ≤ ζ). 2. U (p; m, n) quantiles. Usage: qu(α, p, m, n) Value: The function returns the α-quantile , Uα (p; m, n) say, satisfying P (U (p; m, n) ≤ Uα (p; m, n)) = α. It returns as well a Cf actor, frequently used by people relying on the asymptotic result −[n − 12 (p − m + 1)] ln U (p; m, n) → χ2pm , d

to make the approximate χ2pm quantile an exact quantile of −[n − 1 2 (p − m + 1)] ln U (p; m, n). More precisely, Cf actor · χ21−α,pm = −[n − 12 (p − m + 1)] ln Uα (p; m, n). Note that lower quantiles of U (p; m, n) correspond to upper quantiles of χ2pm .

262

Appendix C. S-plus functions

3. Beta Q-Q plot for multivariate normality. Usage: qqbeta(x) The input x is the n × p sample matrix. Value: The function returns the Q-Q plot of the points

d2(i) , [(n − 1)2 /n] betaγi 12 p; 12 (n − p − 1) , i = 1, . . . , n, as described in Section 11.4.1. The graphic device must be activated before using this function. 4. Robust S estimate. Usage: s.estimate(x, r, nr, N samp) The input x denotes the n × p sample matrix. The input r in the interval (0, .5) is the asymptotic breakdown point, nr and N samp are positive integer parameters of the numerical algorithm [Ruppert (1992)]. Values of nr = 3 and N samp = 80p are appropriate for most purposes. The user is urged to experiment with other values of nr and N samp to certify that the s.estimate function returned the global minimum. Value: The function returns the S estimate of location and scatter, µn and Vn , the Mahalanobis distances, distance.mahalanobis, for outlier detection and the objective function, determinant, which the S estimate seeks to minimize. Points with a Mahalanobis distance greater than (χ2.95,p )1/2 should be checked for outliers [Rousseeuw and van Zomeren (1990)]. The implementation uses the biweight ρ(·) function of Section 13.4.2 and determines c0 such that E ρ(|z|)/(c20 /6) = r, where z ∼ Np (0, I), to achieve the desired breakdown point. 5. Asymptotic variance of S estimate. Usage: asymp(p, r) The input p is the number of variables, whereas r is the breakdown point. Value: The function returns the asymptotic variance constants, at the normal distribution, in Proposition 13.11: λ = α/β 2 , σ1 , and σ2 . The constants λ−1 and σ1−1 , in particular, serve as measures of relative eﬃciencies of the location and scatter estimates, respectively, at the normal distribution.

References

[1] Ali, M.M., and R. Ponnapalli (1990). An optimal property of the GaussMarkoﬀ estimator. Journal of Multivariate Analysis 32, 171-176. [2] Anderson, G.A. (1965). An asymptotic expansion for the distribution of the latent roots of the estimated covariance matrix. Annals of Mathematical Statistics 36, 1153-1173. [3] Anderson, T.W. (1963). Asymptotic theory for principal component analysis. Annals of Mathematical Statistics 34, 122-148. [4] Anderson, T.W. (1965). An asymtotic expansion for the distribution of the latent roots of the estimated covariance matrix. Annals of Mathematical Statistics 36, 1153-1173. [5] Anderson, T.W. (1984). An Introduction to Multivariate Statistical Analysis. 2nd ed. John Wiley & Sons, New York. [6] Andrews, D.F., R. Gnanadesikan, and J.L. Warner (1971). Transformations of multivariate data. Biometrics 27, 825-840. [7] Andrews, D.F., R. Gnanadesikan, and J.L. Warner (1973). Methods for assessing multivariate normality. Multivariate Analysis, ed. P.K. Krishnaiah. Academic Press, New York, 95-116. [8] Ash, R. (1972). Real Analysis and Probability. Academic Press, New York. [9] Baringhaus, L., and N. Henze(1991). Limit distributions for measures of multivariate skewness and kurtosis based on projections. Journal of Multivariate Analysis 38, 51-69. [10] Bartlett, M.S. (1937). Properties of suﬃciency and statistical tests. Proceedings of the Royal Society. London. Series A. 160, 268-282. [11] Bartlett, M.S. (1938). Further aspects of the theory of multiple regression. Proceedings of the Cambridge Philosophical Society 34, 33-40.

264

References

[12] Bellman, R. (1960). Introduction to Matrix Analysis. McGraw-Hill, New York. [13] Bentler, P.M. (1983). Some contributions to eﬃcient statistics in structural models: Speciﬁcation and estimation of moment structures. Psychometrika 48, 493-517. [14] Beran, R. (1984). Bootstrap methods in statistics. Jahresbericht der Deutschen Mathematiker-Vereinigung 86, 14-30. [15] Beran, R. (1986). Simulated power functions. Annals of Statistics 14, 151173. [16] Beran, R. (1987). Prepivoting to reduce level error in conﬁdence sets. Biometrika 74, 457-468. [17] Beran, R. (1988). Prepivoting test statistics: a bootstrap view of asymptotic reﬁnements. Journal of the American Statistical Association 83, 687-697. [18] Beran, R., and M.S. Srivastava (1985). Bootstrap tests and conﬁdence regions for functions of a covariance matrix. Annals of Statistics 13, 95-115. [19] Beran, R., and M.S. Srivastava (1987). Correction: Bootstrap tests and conﬁdence regions for functions of a covariance matrix. Annals of Statistics 15, 470-471. [20] Berk, R., and J.T. Hwang (1989). Optimality of the least squares estimator. Journal of Multivariate Analysis 30, 245-254. [21] Berkane, M., K. Oden, and P.M. Bentler (1997). Geodesic estimation in elliptical distributions. Journal of Multivariate Analysis 63, 35-46. [22] Bhat, B.R. (1981). Modern Probability Theory. John Wiley & Sons, New York. [23] Bickel, P.J., and D.A. Freedman (1981). Some asymptotic theory for the bootstrap. Annals of Statistics 9, 1196-1217. [24] Billingsley, P. (1968). Convergence of Probability Measures. John Wiley & Sons, New York. [25] Billingsley, P. (1995). Probability and Measure. 3rd ed. John Wiley & Sons, New York. [26] Bilodeau, M. (1988). On the simultaneous estimation of scale parameters. The Canadian Journal of Statistics 16, 169-174. [27] Bilodeau, M. (1990). On the choice of the loss function in covariance estimation. Statistics & Decisions 8, 131-139. [28] Bilodeau, M. (1995). Minimax estimators of the mean vector in normal mixed linear models. Journal of Multivariate Analysis 52, 73-82. [29] Bilodeau, M.(1996). Some remarks on U (p; m, n) distributions. Statistics and Probability Letters 31, 41-43. [30] Bilodeau M. (1997a). Estimating a multivariate treatment eﬀect under a biased allocation rule. Communications in Statistics, Theory and Methods 26, 1119-1124. [31] Bilodeau, M. (1997b). Robust test for a given variance. Technical Report, Universit´e de Montr´eal.

References

265

[32] Bilodeau, M. (1998). Multivariate ﬂattening for better predictions. Technical Report, Universit´e de Montr´eal. [33] Bilodeau, M., and M.S. Srivastava (1989a). Estimation of the MSE matrix of the Stein estimator. The Canadian Journal of Statistics 16, 153-159. [34] Bilodeau, M., and M.S. Srivastava (1989b). Stein estimation under elliptical distributions. Journal of Multivariate Analysis 28, 247-259. [35] Bilodeau, M., and M.S. Srivastava (1992). Estimation of the eigenvalues of Σ1 Σ−1 2 . Journal of Multivariate Analysis 41, 1-13. [36] Bilodeau, M., and T. Kariya (1989). Minimax estimators in the normal MANOVA model. Journal of Multivariate Analysis 28, 260-270. [37] Bilodeau, M., and T. Kariya (1994). LBI tests of independence in bivariate exponential distributions. Annals of the Institute of Statistical Mathematics 46, 127-136. [38] Blom, G. (1958). Statistical Estimates and Transformed Beta-variables. John Wiley & Sons, New York. [39] Boente, G. (1987). Asymptotic theory for robust principal components. Journal of Multivariate Analysis 21, 67-78. [40] Boulerice, B., and G.R. Ducharme (1997). Smooth tests of goodness-of-ﬁt for directional and axial data. Journal of Multivariate Analysis 60, 154-175. [41] Box, G.E.P. (1949). A general distribution theory for a class of likelihood criteria. Biometrika 36, 317-346. [42] Box, G.E.P., and D.R. Cox (1964). An analysis of transformations. Journal of the Royal Statistical Society B 26, 211-252. [43] Breiman, L., and J.H. Friedman (1997). Predicting multivariate responses in multiple linear regression. Journal of the Royal Statistical Society B 59, 3-54. [44] Brown, P.J. (1980). Aspects of multivariate regression (with discussion). Bayesian Statistics. eds. J.M. Bernardo, M.H. DeGroot, D.V. Lindley, and A.F.M. Smith. Valencia University Press, Valencia. [45] Browne, M.W., and A. Shapiro (1987). Adjustments for kurtosis in factor analysis with elliptically distributed errors. Journal of the Royal Statistical Society B 49, 346-352. [46] Carri`ere, J.F. (1994). Dependent decrement theory. Transactions XLVI, Society of Actuaries, 45-65. [47] Casella, G., and R.L. Berger (1990). Statistical Inference, Duxbury Press, Belmont, California. [48] Chattopadhyay, A.K., and K.C.S. Pillai (1973). Asymptotic expansions for the distributions of characteristic roots when the parameter matrix has several multiple roots. Multivariate analysis III. Academic Press, New York, 117-127. [49] Chikuse, Y. (1976). Asymptotic distributions of the latent roots of the covariance matrix with multiple population roots. Journal of Multivariate Analysis 6, 237-249.

266

References

[50] Cl´eroux, R., and G.R. Ducharme (1989). Vector correlation for elliptical distributions. Communications in Statistics, Theory and Methods 18, 14411454. [51] Coelho, C.A. (1998). The generalized integer gamma distribution–A basis for distributions in multivariate statistics. Journal of Multivariate Analysis 64, 86-102. [52] Cook, R.D., M.E. Johnson (1981). A family of distributions for modelling nonelliptically symmetric multivariate data. Journal of the Royal Statistical Society B 43, 210-218. [53] Copas, J.B. (1975). On the unimodality of the likelihood for the Cauchy distribution. Biometrika 62, 701-704. [54] Courant, R. (1936). Diﬀerential and Integral Calculus II. John Wiley & Sons, New York. [55] Cox, D.R., and N.J.H. Small (1978). Testing multivariate normality. Biometrika 65, 263-272. [56] Cuadras, C.M. (1992). Probability distributions with given multivariate marginals and given dependence structure. Journal of Multivariate Analysis 42, 51-66. [57] Datta, S., N. Mukhopadhyay (1997). On sequential ﬁxed-size conﬁdence regions for the mean vector. Journal of Multivariate Analysis 60, 233-251. [58] Davies, P.L. (1987). Asymptotic behaviour of S-estimates of multivariate location parameters and dispersion matrices. Annals of Statistics 15, 12691292. [59] Davis, A.W. (1971). Percentile approximations for a class of likelihood ratio criteria. Biometrika 58, 349-356. [60] Davison, A.C., and D.V. Hinkley (1997). Bootstrap Methods and their Application. Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press, Cambridge. [61] Donoho, D.L. (1982). Breakdown Properties of Multivariate Location Estimators. Qualifying paper, Harvard University. [62] Ducharme, G.R., P. Milasevic (1987). Spatial median and directional data. Biometrika 74, 212-215. [63] D¨ umbgen, L. (1998). Perturbation inequalities and conﬁdence sets for functions of a scatter matrix. Journal of Multivariate Analysis 65, 19-35. [64] Dykstra, R.L. (1970). Establishing the positive deﬁniteness of the sample covariance matrix. Annals of Mathematical Statistics 41, 2153-2154. [65] Eaton, M.L. (1983). Multivariate Statistics, a Vector Space Approach. John Wiley & Sons, New York. [66] Eaton, M.L. (1988). Concentration inequalities for Gauss-Markov estimators. Journal of Multivariate Analysis 25, 119-138. [67] Eaton, M.L., and M.D. Perlman (1973). The non-singularity of generalized sample covariance matrices. Annals of Statistics 1, 710-717. [68] Eaton, M.L., and D.E. Tyler (1991). On Wielandt’s inequality and its application to the asymptotic distribution of the eigenvalues of a random symmetric matrix. Annals of Statistics 19, 260-271.

References

267

[69] Eaton, M.L., and D.E. Tyler (1994). The asymptotic distribution of singular values with applications to canonical correlations and correspondence analysis. Journal of Multivariate Analysis 50, 238-264. [70] Efron, B. (1969). Student’s t-test under symmetry conditions. Journal of the American Statistical Association 64, 1278-1302. [71] Efron B. (1982). The Jacknife, the Bootstrap and Other Resampling Plans. SIAM, Philadelphia. bibitemefrm Efron B., and C. Morris (1976). Multivariate empirical Bayes and estimation of covariance matrix. Annals of Statistics 4, 22-32. indexaiMorris, C. [72] Efron, B., and R.J. Tibshirani (1993). An Introduction to the Bootstrap. Chapman & Hall, New York. indexaiTibshirani, R.J. [73] Erd´elyi, A., W. Magnus, F. Oberhettinger, and F.G. Tricomi (1953). Higher Transcendental Functions. McGraw-Hill, New York. [74] Escouﬁer, Y. (1973). Le traitement des variables vectorielles. Biometrics 29, 751-760. [75] Fan, Y. (1997). Goodness-of-ﬁt tests for a multivariate distribution by the empirical characteristic function. Journal of Multivariate Analysis 62, 3663. [76] Fang, K.T., S. Kotz, and K.W. Ng (1991). Symmetric Multivariate and Related Distributions. Chapman & Hall, London. [77] Fang, K.T., L.-X. Zhu, and P.M. Bentler (1993). A necessary test of goodness of ﬁt for sphericity. Journal of Multivariate Analysis 45, 34-55. [78] Feller, W. (1966). An Introduction to Probability Theory and Its Applications (Vol. II). John Wiley & Sons, New York. [79] Fisher, R.A. (1953). Dispersion on a sphere. Proceedings of the Royal Society. London. Series A. 217, 295-305. [80] Flury, B. (1997). A First Course in Multivariate Statistics. Springer-Verlag, New York. [81] Frank, M.J. (1979). On the simultaneous associativity of F (x, y) and x + y − F (x, y). Aequationes Math. 19, 194-226. [82] Fraser, D.A.S. (1976). Probability and Statistics: Theory and Applications. DAI Press, Toronto. [83] Fraser, D.A.S., I. Guttman, and M.S. Srivastava (1991). Conditional inference for treatment and error in multivariate analysis. Biometrika 78, 565-572. [84] Fujikoshi, Y. (1970). Asymptotic expansions of the distributions of test statistics in multivariate analysis. Journal of Science of the Hiroshima University. Series A, Mathematics 34, 73-144. [85] Fujikoshi, Y. (1977). An asymptotic expansion for the distributions of the latent roots of the Wishart matrix with multiple population roots. Annals of the Institute of Statistical Mathematics 29, 379-387. [86] Fujikoshi, Y. (1978). Asymptotic expansions for the distributions of some functions of the latent roots of matrices in three situations. Journal of Multivariate Analysis 8, 63-72.

268

References

[87] Fujikoshi, Y. (1988). Comparison of powers of a class of tests for multivariate linear hypothesis and independence. Journal of Multivariate Analysis 26, 48-58. [88] Fujikoshi, Y. (1997). An asymptotic expansion for the distribution of Hotelling’s T 2 -statistic under nonnormality. Journal of Multivariate Analysis 61, 187-193. [89] Fujikoshi, Y., and Y. Watamori (1992). Tests for the mean direction of the Langevin distribution with large concentration parameter. Journal of Multivariate Analysis 42, 210-225. [90] Fujisawa, H. (1997). Improvement on chi-squared approximation by monotone transformation. Journal of Multivariate Analysis 60, 84-89. [91] Fujisawa, H. (1997). Likelihood ratio criterion for mean structure in the growth curve model with random eﬀects. Journal of Multivariate Analysis 60, 90-98. [92] Genest, C., and R.J. MacKay (1986). Copules archim´ediennes et familles de lois bidimensionnelles dont les marges sont donn´ees. Canadian Journal of Statistics 14, 145-159. [93] Genest, C. (1987). Frank’s family of bivariate distributions. Biometrika 74, 549-555. [94] Ghosh, B. K. (1970). Sequential Tests of Statistical Hypotheses. AddisonWesley, Reading, Massachusetts. [95] Giri, N.C. (1996). Multivariate Statistical Analysis. Marcel Dekker, New York. [96] Gnanadesikan, R. (1977). Methods for Statistical Data Analysis of Multivariate Observations. John Wiley & Sons, New York. [97] Gnanadesikan, R., and J.R. Kettenring (1972). Robust estimates, residuals, and outlier detection with multiresponse data. Biometrics 28, 81-124. [98] Gr¨ ubel, R., and D.M. Rocke (1990). On the cumulants of aﬃne equivariant estimators in elliptical families. Journal of Multivariate Analysis 35, 203222. [99] Gunderson, B.K., and R.J. Muirhead (1997). On estimating the dimensionality in canonical correlation analysis. Journal of Multivariate Analysis 62, 121-136 [100] Gupta, A.K., and D. Song (1997). Characterization of p-generalized normality. Journal of Multivariate Analysis 60, 61-71. [101] Gupta, A.K., and D. St. P. Richards (1990). The Dirichlet distributions and polynomial regression. Journal of Multivariate Analysis 32, 95-102. [102] Gupta, A.K., and T. Varga (1992). Characterization of matrix variate normal distributions. Journal of Multivariate Analysis 41, 80-88. [103] Hall, P. (1992). The Bootstrap and Edgeworth Expansion. Springer-Verlag, New York. [104] Hendriks, H., Z. Landsman, and F. Ruymgaart (1996). Asymptotic behavior of sample mean direction for spheres. Journal of Multivariate Analysis 59, 141-152.

References

269

[105] Henze, N., and T. Wagner (1997). A new approach to the BHEP tests for multivariate normality. Journal of Multivariate Analysis 62, 1-23. [106] Henze, N., and B. Zirkler (1990). A class of invariant consistent tests for multivariate normality. Communications in Statistics, Theory and Methods 19, 3595-3617. [107] Hsu, P.L. (1941). On the limiting distribution of the canonical correlations. Biometrika 32, 38-45. [108] Huber, P.J. (1981). Robust Statistics. John Wiley & Sons, New York. [109] Iwashita, T. (1997). Asymptotic null and nonnull distribution of Hotelling’s T 2 -statistic under the elliptical distribution. Journal of Statistical Planning and Inference 61, 85-104.. [110] Iwashita, T., and M. Siotani (1994). Asymptotic distributions of functions of a sample covariance matrix under the elliptical distribution. The Canadian Journal of Statistics 22, 273-283. [111] James, A.T. (1954). Normal multivariate analysis and the orthogonal group. Annals of Mathematical Statistics 25, 40-75. [112] James, A.T. (1969). Tests of equality of latent roots of the covariance matrix. Multivariate Analysis II. Academic Press, New York, 205-218. [113] John, S. (1971). Some optimal multivariate tests. Biometrika 58, 123-127. [114] John, S. (1972). The distribution of a statistic used for testing sphericity of normal distributions. Biometrika 59, 169-174. [115] Johnson, N.L. (1949). Systems of frequency curves generated by methods of translation. Biometrika 36, 149-176. [116] Johnson, N.L., S. Kotz, and A.W. Kemp (1992). Univariate Discrete Distributions. 2nd ed. John Wiley & Sons, New York. [117] Johnson, R.A., and D.W. Wichern (1992). Applied Multivariate Statistical Analysis. 3rd ed. Prentice-Hall, Englewood Cliﬀs, New Jersey. [118] Jolliﬀe, I.T. (1986). Principal Component Analysis. Springer-Verlag, New York. [119] Jordan, S.M., and K. Krishnamoorthy (1995). Conﬁdence regions for the common mean vector of several multivariate normal populations. The Canadian Journal of Statistics 23, 283-297. [120] Kano, Y. (1994). Consistency property of elliptical probability density function. Journal of Multivariate Analysis 51, 139-147. [121] Kano, Y. (1995). An asymptotic expansion of the distribution of Hotelling’s T 2 -statistic under general condition. American Journal of Mathematical and Management Sciences 15, 317-341. [122] Kariya, T. (1985). Testing in the Multivariate General Linear Model. Kinokunia, Tokyo. [123] Kariya, T., and B.K. Sinha (1989). Robustness of Statistical Tests. Academic Press, San Diego. [124] Kariya, T., and M.L. Eaton (1977). Robust tests for spherical symmetry. Annals of Statistics 5, 206-215.

270

References

[125] Kariya, T., R.S. Tsay, N. Terui, and Hong Li (1999). Tests for multinormality with applications to time series. Communications in Statistics, Theory and Methods 28, 519-536. [126] Kariya, T., Y. Fujikoshi, and P.R. Krishnaiah (1987). On tests for selection of variables and independence under multivariate regression models. Journal of Multivariate Analysis 21, 207-237. [127] Kato, T. (1982). A Short Introduction to Perturbation Theory for Linear Operators. Springer-Verlag, New York. [128] Kelker, D. (1970). Distribution theory of spherical distributions and a location-scale parameter generalization. Sankhy¯ a A 32, 419-430. [129] Kendall, M., A. Stuart, and J.K. Ord (1987). Kendall’s Advanced Theory of Statistics. 5th ed. Vol. 1. Oxford University Press, New York. [130] Kent, J.T., and D.E. Tyler (1991). Redescending M-estimates of multivariate location and scatter. Annals of Statistics 19, 2102-2019. [131] Khatri, C.G., and M.S. Srivastava (1974). Asymptotic expansions of the non-null distributions of the likelihood ratio criteria for covariance matrices II. Proc. Carleton University, Ottawa. Metron 36, 55-71. [132] Khatri, C. G., and M.S. Srivastava (1978). Asymptotic expansions for distributions of characteristic roots of covariance matrices. South African Statistical Journal 12, 161-186. [133] Ko, D., and T. Chang (1993). Robust M-estimators on spheres. Journal of Multivariate Analysis 45, 104-136. [134] Koehler, K.J., and J.T. Symanowski (1995). Constructing multivariate distributions with speciﬁc marginal distributions. Journal of Multivariate Analysis 55, 261-282. [135] Kollo, T., and H. Neudecker (1993). Asymptotics of eigenvalues and unitlength eigenvectors of sample variance and correlation matrices. Journal of Multivariate Analysis 47, 283-300. [136] Kollo, T., and D. von Rosen (1995). Minimal moments and cumulants of symmetric matrices: an application to the Wishart distribution. Journal of Multivariate Analysis 55, 149-164. [137] Koltchinskii, V.I., and L. Li (1998). Testing for spherical symmetry of a multivariate distribution. Journal of Multivariate Analysis 65, 228-244. [138] Konishi, S. (1979). Asymptotic expansions for the distributions of statistics based on the sample correlation matrix in principal component analysis. Hiroshima Mathematical Journal 9, 647-700. [139] Konishi, S., and C.G. Khatri (1990). Inferences on interclass and intraclass correlations in multivariate familial data. Annals of the Institute of Statistical Mathematics 42, 561-580. [140] Konishi, S., and C.R. Rao (1991). Inferences on multivariate measures of interclass and intraclass correlations in familial data. Journal of the Royal Statistical Society B 53, 649-659. [141] Konishi, S., and C.R. Rao (1992). Principal component analysis for multivariate familial data. Biometrika 79, 631-641.

References

271

[142] Kotz, S., and I. Ostrovskii (1994). Characteristic functions of a class of elliptical distributions. Journal of Multivariate Analysis 49, 164-178. [143] Kres, H. (1983). Statistical Tables for Multivariate Analysis, a Handbook with References to Applications. Springer-Verlag, New York. [144] Kshirsagar, A.M. (1972). Multivariate Analysis. Marcel Dekker, New York. [145] Kudˆ o, A. (1963). A multivariate analogue of the one-sided test. Biometrika 50, 403-418. [146] Kuwana, Y., and T. Kariya (1991). LBI tests for multivariate normality in exponential power distributions. Journal of Multivariate Analysis 39, 117-134. [147] Lee, Y.-S. (1972). Some results on the distribution of Wilk’s likelihood-ratio criterion. Biometrika 59, 649-664. [148] Lehmann, E.L. (1983). Theory of Point Estimation. John Wiley & Sons, New York. [149] Li, Haijun, M. Scarsini, and M. Shaked (1996). Linkages: A tool for construction of multivariate distributions with given nonoverlapping multivariate marginals. Journal of Multivariate Analysis 56, 20-41. [150] Liu, C. (1995). Missing data imputation using the multivariate t distribution. Journal of Multivariate Analysis 53, 139-158. [151] Liu, C. (1997). ML estimation of the multivariate t distribution and the EM algorithm. Journal of Multivariate Analysis 63, 296-312. [152] Looney, S.W. (1995). How to use test for univariate normality to assess multivariate normality. The American Statistician 49, 64-70. [153] Lopuha¨ a, H.P. (1989). On the relation between S-estimators and Mestimators of multivariate location and covariance. Annals of Statistics 17, 1662-1683. [154] Lopuha¨ a, H.P. (1991). Multivariate τ -estimators for location and scatter. The Canadian Journal of Statistics 19, 307-321. [155] Lopuha¨ a, H.P., and P.J. Rousseeuw (1991). Breakdown points of aﬃne equivariant estimators of multivariate location and covariance matrices. Annals of Statistics 19, 229-248. [156] MacDuﬀy, C.C. (1943). Vectors and Matrices. The Mathematical Association of America, Providence, Rhode Island. [157] Magnus, J.R., and H. Neudecker (1979). The commutation matrix: Some properties and applications. Annals of Statistics 7, 381-394. [158] Malkovich, J.F., and A.A. Aﬁﬁ (1973). On tests for multivariate normality. Journal of the American Statistical Association 68, 176-179. [159] Mardia, K.V. (1970). Measures of multivariate skewness and kurtosis with applications. Biometrika 57, 519-530. [160] Mardia, K.V. (1972). Statistics of Directional Data. Academic Press, London. [161] Mardia, K.V. (1975). Assessment of multinormality and the robustness of Hotelling’s T 2 test. Applied Statistics 24, 163-171.

272

References

[162] Mardia, K.V., J.T. Kent, and J.M. Bibby (1979). Multivariate Analysis. Academic Press, New York. [163] M¨ arkel¨ ainen, T., K. Schmidt, and G.P.H. Styan (1981). On the existence and uniqueness of the maximum likelihood estimate of a vector-valued parameter in ﬁxed sample sizes. Annals of Statistics 9, 758-767. [164] Maronna, R.A. (1976). Robust M-estimators of multivariate location and scatter. Annals of Statistics 4, 51-67. [165] Marshall, A.W., and I. Olkin(1988). Families of multivariate distributions. Journal of the American Statistical Association 83, 834-841. [166] Martin, M.A. (1990). On bootstrap iteration for coverage correction in conﬁdence intervals. Journal of the American Statistical Association 85, 1105-1118. bibitemmat Mathew, T., and K. Nordstr¨ om (1997). Wishart and chi-square distributions associated with matrix quadratic forms. Journal of Multivariate Analysis 61, 129-143. [167] Mauchly, J.W. (1940). Signiﬁcance test for sphericity of a normal n-variate distribution. Annals of Mathematical Statistics 11, 204-209. [168] Muirhead, R.J. (1970). Asymptotic distributions of some multivariate tests. Annals of Mathematical Statistics 41, 1002-1010. [169] Muirhead, R.J. (1982). Aspects of Multivariate Statistical Theory. John Wiley & Sons, New York. [170] Muirhead, R.J., and Y. Chikuse (1975). Asymptotic expansions for the joint and marginal distributions of the latent roots of the covariance matrix. Annals of Statistics 3, 1011-1017. [171] Muirhead, R.J., and C.M. Waternaux (1980). Asymptotic distributions in canonical correlation analysis and other multivariate procedures for nonnormal populations. Biometrika 67, 31-43. [172] Nagao, H. (1973). On some test criteria for covariance matrix. Annals of Statistics 1, 700-709. [173] Nagao, H., and M.S. Srivastava (1992). On the distributions of some test criteria for a covariance matrix under local alternatives and bootstrap approximations. Journal of Multivariate Analysis 43, 331-350. [174] Naito, K. (1998). Approximation of the power of kurtosis test for multinormality. Journal of Multivariate Analysis 65, 166-180. [175] Nelsen, R.B. (1986). Properties of a one-parameter family of bivariate distributions with speciﬁed marginals. Communications in Statistics 15, 3277-3285. [176] Nguyen, T.T. (1997). A note on matrix variate normal distribution. Journal of Multivariate Analysis 60, 148-153. [177] Oakes, D. (1982). A model for association in bivariate survival data. Journal of the Royal Statistical Society B 44, 414-442. [178] Olkin, I., and J.W. Pratt (1958). Unbiased estimation of certain correlation coeﬃcients. Annals of Mathematical Statistics 29, 201-211. [179] Olkin, I., and S.N. Roy (1954). On multivariate distribution theory. Annals of Mathematical Statistics 25, 329-339.

References

273

[180] Perlman, M.D. (1969). One-sided testing problems in multivariate analysis. Annals of Mathematical Statistics 40, 549-567; Correction, Annals of Mathematical Statistics 42 (1971), 1777. [181] Perlman, M.D. (1980). Unbiasedness of the likelihood ratio tests for equality of several covariance matrices and equality of several multivariate normal populations. Annals of Statistics 8, 247-263. [182] Press, W.H. (1992). Numerical Recipies in C: The Art of Scientiﬁc Computing. 2nd ed. Cambridge University Press, New York. [183] Purkayastha, S., and M.S. Srivastava (1995). Asymptotic distributions of some test criteria for the covariance matrix in elliptical distributions under local alternatives. Journal of Multivariate Analysis 55, 165-186. [184] Rao, B.V., and B.K. Sinha (1988). A characterization of Dirichlet distributions. Journal of Multivariate Analysis 25, 25-30. [185] Rao, C.R. (1973). Linear Statistical Inference and Its Applications. 2nd ed. John Wiley & Sons, New York. [186] Redfern, D. (1996). Maple V Release 4. 3rd ed. Springer-Verlag, New York. [187] Reeds, J.A. (1985). Asymptotic number of roots of Cauchy likelihood equations. Annals of Statistics 13, 778-784. [188] Rocke, D.M. (1996). Robustness properties of S-estimators of multivariate location and shape in high dimension. Annals of Statistics 24, 1327-1345. [189] Romeu, J.L., and A. Ozturk (1993). A comparative study of goodness-ofﬁt tests for multivariate normality. Journal of Multivariate Analysis 46, 309-334. [190] Rousseeuw, P.J. (1985). Multivariate estimation with high breakdown point. In Mathematical Statistics and Applications. eds. W. Grossmann, G. Pﬂug, I. Vincze and W. Wertz. Vol. 8. Reidel, Dordrecht, 283-297. [191] Rousseeuw, P.J., and B.C. van Zomeren (1990). Unmasking multivariate outliers and leverage points. Journal of the American Statistical Association 85, 633-639. [192] Rousseeuw, P.J., and V.J. Yohai (1984). Robust regression by means of S-estimators. Robust and Nonlinear Time Series Analysis. Lecture Notes in Statistics Vol. 26. Springer, New York, 256-272. [193] Royston, J.F. (1982). An extension of Shapiro and Wilk’s W test for normality to large samples. Applied Statistics 31, 115-124. [194] Royston, J.F. (1983). Some techniques for assessing multivariate normality based on the Shapiro-Wilk W . Applied Statistics 32, 121-133. [195] Ruppert, D. (1992). Computing S estimators for regression and multivariate location/dispersion. Journal of Computational and Graphical Statistics 1, 253-270. [196] Saw, J.G. (1978). A family of distributions on the m-sphere and some hypothesis tests. Biometrika 65, 69-73. [197] Schoenberg, I.J. (1938). Metric spaces and completely monotone functions. Annals of Mathematics 39, 811-841.

274

References

[198] Sepanski, S.J. (1994). Asymptotics for multivariate t-statistic and Hotelling’s T 2 -statistic under inﬁnite second moments via bootstrapping. Journal of Multivariate Analysis 49, 41-54. [199] Serﬂing, R.J. (1980). Approximation Theorems of Mathematical Statistics. John Wiley & Sons, New York. [200] Shapiro, A., and M.W. Browne (1987). Analysis of covariance structures under elliptical distributions. Journal of the American Statistical Association 82, 1092-1097. [201] Shapiro, S.S., and M.B. Wilk (1965). An analysis of variance test for normality (complete samples). Biometrika 52, 591-611. [202] Shapiro, S.S., and R.S. Francia (1972). An approximate analysis of variance test for normality. Journal of the American Statistical Association 67, 215216. [203] Silvapulle, M.J. (1995). A Hotelling’s T 2 -type statistic for testing against one-sided hypotheses. Journal of Multivariate Analysis 55, 312-319. [204] Singh, K. (1981). On the asymptotic accuracy of Efron’s bootstrap. Annals of Statistics 9, 1187-1195. [205] Siotani, M., T. Hayakawa, and Y. Fujikoshi (1985). Modern Multivariate Statistical Analysis: A Graduate Course and Handbook. American Sciences Press, Columbus, Ohio. [206] Small, N.J.H. (1978). Plotting squared radii. Biometrika 65, 657-658. [207] Spivak, M. (1965). Calculus on Manifolds. Addison-Wesley, New York. [208] Srivastava, M.S. (1967). On ﬁxed-width conﬁdence bounds for regression parameters and the mean vector. Journal of the Royal Statistical Society B 29, 132-140. [209] Srivastava, M.S. (1984). Estimation of interclass correlations in familial data. Biometrika 71, 177-185. [210] Srivastava, M.S., and E.M. Carter (1980). Asymptotic expansions for hypergeometric functions. Multivariate analysis V. North-Holland, AmsterdamNew York, 337-347. [211] Srivastava, M.S., and E.M. Carter (1983). An Introduction to Applied Multivariate Statistics. North-Holland, New York. [212] Srivastava, M.S., and T.K. Hui (1987). On assessing multivariate normality based on Shapiro-Wilk W statistic. Statistics & Probability Letters 5, 15-18. [213] Srivastava, M.S., K.J. Keen, and R.S. Katapa (1988). Estimation of interclass and intraclass correlations in multivariate familial data. Biometrics 44, 141-150. [214] Srivastava, M.S., and C.G. Khatri (1979). An Introduction to Multivariate Statistics. North-Holland, New York. [215] Srivastava, M.S., C.G. Khatri, and E.M. Carter (1978). On monotonicity of the modiﬁed likelihood ratio test for the equality of two covariances. Journal of Multivariate Analysis 8, 262-267. [216] Srivastava, M.S., and D. von Rosen (1998). Outliers in multivariate regression models. Journal of Multivariate Analysis 65, 195-208.

References

275

[217] Srivastava, M.S., and W.K. Yau (1989). Saddlepoint method for obtaining tail probability of Wilks’ likelihood ratio test. Journal of Multivariate Analysis 31, 117-126. [218] Stadje, W. (1993). ML characterization of the multivariate normal distribution. Journal of Multivariate Analysis 46, 131-138. [219] Stahel, W.A. (1981). Robuste Sch¨ atzungen: Inﬁnitesimale Optimalit¨ at und Sch¨ atzungen von Kovarianzmatrizen. Ph. D. thesis, ETH Z¨ urich. [220] Statistical Sciences (1995). S-PLUS Guide to Statistical and Mathematical Analysis, Version 3.3. StatSci, a division of MathSoft, Inc., Seattle, Washington. [221] Stein, C. (1969). Multivariate Analysis I. Technical Report No. 42, Stanford University. [222] Steyn, H.S. (1993). On the problem of more than one kurtosis parameter in multivariate analysis. Journal of Multivariate Analysis 44, 1-22. [223] Stone, M. (1974). Cross-validatory choice and assessment of statistical predictions (with discussion). Journal of the Royal Statistical Society B, 36, 111-147. [224] Strang, G. (1980). Linear Algebra and its Applications. 2nd ed. Academic Press, New York. [225] Sugiura, N. (1973). Derivatives of the characteristic root of a symmetric or a hermitian matrix with two applications in mutivariate analysis. Communications in Statistics 1, 393-417. [226] Sugiura, N. (1976). Asymptotic expansions of the distributions of the latent roots and the latent vector of the Wishart and multivariate F matrices. Journal of Multivariate Analysis 6, 500-525. [227] Sugiura, N., and H. Nagao (1968). Unbiasedness of some test criteria for the equality of one or two covariance matrices. Annals of Mathematical Statistics 39, 1686-1692. [228] Sugiura, N., and H. Nagao (1971). Asymptotic expansion of the distribution of the generalized variance for noncentral Wishart matrix, when Ω = O(n). Annals of the Institute of Statistical Mathematics 23, 469-475. [229] Sutradhar, B.C. (1993). Score test for the covariance matrix of the elliptical t-distribution. Journal of Multivariate Analysis 46, 1-12. [230] Szablowski, P.J. (1998). Uniform distributions on spheres in ﬁnite dimensional Lα and their generalizations. Journal of Multivariate Analysis 64, 103-117. [231] Tang, D. (1994). Uniformly more powerful tests in a one-sided multivariate problem. Journal of the American Statistical Association 89, 1006-1011. [232] Tang, D. (1996). Erratum:“Uniformly more powerful tests in a one-sided multivariate problem” [Journal of the American Statistical Association 89 (1994), 1006-1011]. Journal of the American Statistical Association 91, 1757. [233] Tyler, D.E. (1982). Radial estimates and the test for sphericity. Biometrika 69, 429-436.

276

References

[234] Tyler, D.E. (1983a). Robustness and eﬃciency properties of scatter matrices. Biometrika 70, 411-420. [235] Tyler, D.E. (1983b). The asymptotic distribution of principal components roots under local alternatives to multiple roots. Annals of Statistics 11, 1232-1242. [236] Tyler, D.E. (1986). Breakdown properties of the M-estimators of multivariate scatter. Technical report, Department of Statistics, Rutgers University. [237] Uhlig, H. (1994). On singular Wishart and singular multivariate beta distributions. Annals of Statistics 22, 395-405. [238] van der Merwe, A., and J.V. Zidek (1980). Multivariate regression analysis and canonical variates. Canadian Journal of Statistics, 8, 27-39. [239] von Mises, R. (1918). Uber die “Ganzahligkeit” der Atomegewicht und verwante Fragen. Physikalische Zeitschrift 19, 490-500. [240] Wakaki, H., S. Eguchi, and Y. Fujikoshi (1990). A class of tests for a general covariance structure. Journal of Multivariate Analysis 32, 313-325. [241] Wang, Y., and M.P. McDermott (1998a). A conditional test for a nonnegative mean vector based on a Hotelling’s T 2 -type statistic. Journal of Multivariate Analysis 66, 64-70. [242] Wang, Y., and M.P. McDermott (1998b). Conditional likelihood ratio test for a nonnegative normal mean vector. Journal of the American Statistical Association 93, 380-386. [243] Waternaux, C.M. (1976). Asymptotic distributions of the sample roots for a non-normal population. Biometrika 63, 639-664. [244] Watson, G.S. (1983). Statistics on Spheres. The University of Arkansas Lecture Notes in Mathematical Sciences. John Wiley & Sons, New York. [245] Wielandt, H. (1967). Topics in the Analytic Theory of Matrices (Lecture notes prepared by R.R. Meyer.) University of Wisconsin Press, Madison. [246] Wilks, S.S. (1963). Multivariate statistical outliers. Sankhy¯ a: Series A 25, 407-426. [247] Wolfram, S. (1996). The Mathematica Book. 3rd ed. Wolfram Media, Inc. and Cambridge University Press, New York. [248] Wong, C.S., and D. Liu (1994). Moments for left elliptically contoured random matrices. Journal of Multivariate Analysis 49, 1-23. [249] Yamato, H. (1990). Uniformly minimum variance unbiased estimation for symmetric normal distributions. Journal of Multivariate Analysis 34, 227237.

Author Index

Aﬁﬁ, A.A., 170, 271 Ali, M.M., 66, 263 Anderson, G.A., 132, 263 Anderson, T.W., 132, 183, 263 Andrews, D.F., 94, 170, 263 Ash, R., 63, 263

Boulerice, B., 72, 265 Box, G.E.P., 94, 184, 195, 198, 201, 204, 265 Breiman, L., 154, 156–158, 265 Brown, P.J., 156, 265 Browne, M.W., 228, 265, 274

Baringhaus, L., 171, 263 Bartlett, M.S., 123, 199, 263 Bellman, R., 125, 264 Bentler, P.M., 49, 234, 237, 264, 267 Beran, R., 137, 200, 244, 245, 247, 249–251, 264 Berger, R.L., 86, 147, 265 Berk, R., 66, 264 Berkane, M., 237, 264 Bhat, B.R., 34, 264 Bibby, J.M., 272 Bickel, P.J., 244, 264 Billingsley, P., 20, 78, 264 Bilodeau, M., 156, 158, 181, 235, 264, 265 Blom, G., 186, 265 Boente, G., 224, 265

Carri`ere, J.F., 27, 265 Carter, E.M., 123, 137, 274 Casella, G., 86, 147, 265 Chang, T., 72, 270 Chattopadhyay, A.K., 137, 265 Chikuse, Y., 132, 137, 265, 272 Cl´eroux, R., 192, 266 Coelho, C.A., 184, 266 Cook, R.D., 26, 34, 266 Copas, J.B., 214, 266 Courant, R., 33, 266 Cox, D.R., 94, 170, 171, 265, 266 Cuadras, C.M., 26, 266 Datta, S., 104, 266 Davies, P.L., 222, 224, 266 Davis, A.W., 200, 205, 266 Davison, A.C., 243, 266

278

Author Index

Donoho, D.L., 83, 266 Ducharme, G.R., 72, 192, 265, 266 D¨ umbgen, L., 109, 266 Dykstra, R.L., 88, 266

Hui, T.K., 169, 170, 274 Hwang, J.T., 66, 264

Eaton, M.L., 51, 66, 88, 134, 137, 190, 249, 266, 267, 269 Efron, B., 65, 158, 243, 267 Eguchi, S., 98, 276 Erd´elyi, A., 115, 116, 119, 196, 202, 267 Escouﬁer, Y., 191, 267

James, A.T., 30, 33, 94, 137, 269 John, S., 120, 121, 252, 269 Johnson, M.E., 26, 34, 266 Johnson, N.L., 111, 170, 269 Johnson, R.A., 269 Jolliﬀe, I.T., 161, 269 Jordan, S.M., 138, 269

Fan, Y., 171, 267 Fang, K.T., 49, 208, 267 Feller, W., 254, 267 Fisher, R.A., 72, 267 Flury, B., viii, 267 Francia, R.S., 170, 274 Frank, M.J., 26, 267 Fraser, D.A.S., 86, 96, 104, 147, 245, 267 Freedman, D.A., 244, 264 Friedman, J.H., 154, 156–158, 265 Fujikoshi, Y., 72, 98, 103, 132, 154, 205, 267, 268, 270, 274, 276 Fujisawa, H., 103, 268

Kano, Y., 103, 209, 269 Kariya, T., 51, 154, 156, 158, 171, 209, 227, 265, 269–271 Katapa, R.S., 84, 274 Kato, T., 125, 248, 251, 270 Keen, K.J., 84, 274 Kelker, D., 207, 270 Kemp, A.W., 111, 269 Kendall, M., 257, 270 Kent, J.T., 214, 218, 270, 272 Kettenring, J.R., 170, 185, 194, 268 Khatri, C.G., 12, 13, 30, 31, 83, 119, 123, 137, 233, 236, 270, 274 Ko, D., 72, 270 Koehler, K.J., 26, 270 Kollo, T., 132, 259, 270 Koltchinskii, V.I., 49, 270 Konishi, S., 83, 84, 169, 270 Kotz, S., 111, 208, 267, 269, 271 Kres, H., 98, 271 Krishnaiah, P.R., 154, 270 Krishnamoorthy, K., 138, 269 Kshirsagar, A.M., 271 Kudo, A., 103, 271 Kuwana, Y., 209, 271

Genest, C., 26, 268 Ghosh, B.K., 101, 268 Giri, N.C., 268 Gnanadesikan, R., 94, 170, 185, 186, 194, 263, 268 Gr¨ ubel, R., 213, 268 Gunderson, B.K., 190, 268 Gupta, A.K., 41, 50, 74, 268 Guttman, I., 104, 267 Hall, P., 243, 268 Hayakawa, T., 274 Hendriks, H., 72, 268 Henze, N., 171, 269 Hinkley, D.V., 243, 266 Hsu, P.L., 190, 269 Huber, P.J., 222, 269

Iwashita, T., 103, 231, 269

Landsman, Z., 72, 268 Lee, Y.-S., 202, 271 Lehmann, E.L., 221, 271 Li, Haijun, 26, 271

Author Index

Li, Hong, 171, 270 Li, L., 49, 270 Liu, C., 221, 271 Liu, D., 74, 276 Looney, S.W., 170, 271 Lopuha¨ a, H.P., 222, 224–226, 271 MacDuﬀy, C.C., 30, 271 MacKay, R.J., 26, 268 Magnus, J.R., 76, 271 Magnus, W., 115, 116, 119, 196, 202, 267 Malkovich, J.F., 170, 271 Mardia, K.V., 72, 171, 271, 272 M¨ arkel¨ ainen, T., 214, 272 Maronna, R.A., 222, 224, 272 Marshall, A.W., 26, 272 Mathew, T., 92, 272 Mauchly, J.W., 118, 272 McDermott, M.P., 104, 276 Milasevic, T., 72, 266 Morris, C., 158 Muirhead, R.J., 94, 132, 190, 205, 228, 233, 268, 272 Mukhopadhyay, N., 104, 266 Nagao, H., 123, 128, 140, 234, 251, 272, 275 Naito, K., 171, 272 Nelsen, R.B., 27, 272 Neudecker, H., 76, 132, 270, 271 Ng, K.W., 208, 267 Nguyen, T.T., 74, 272 Nordstr¨ om, K., 92, 272 Oakes, D., 27, 272 Oberhettinger, F., 115, 116, 119, 196, 202, 267 Oden, K., 237, 264 Olkin, I., 26, 94, 115, 272 Ord, J.K., 257, 270 Ostrovskii, I., 208, 271 Ozturk, A., 171, 273

279

Perlman, M.D., 88, 103, 123, 142, 266, 273 Pillai, K.C.S., 137, 265 Ponnapalli, R., 66, 263 Pratt, J.W., 115, 272 Press, W.H., 184, 273 Purkayastha, S., 234, 273 Rao, B.V., 41, 273 Rao, C.R., 84, 270, 273 Redfern, D., 197, 273 Reeds, J.A., 214, 273 Richards, D. St. P., 41, 268 Rocke, D.M., 213, 268, 273 Romeu, J.L., 171, 273 Rousseeuw, P.J., 224, 262, 271, 273 Roy, S.N., 94, 272 Royston, J.F., 169, 170, 273 Ruppert, D., 226, 262, 273 Ruymgaart, F., 72, 268 Saw, J.G., 71, 273 Scarsini, M., 26, 271 Schmidt, K., 214, 272 Schoenberg, I.J., 53, 273 Sepanski, S.J., 103, 274 Serﬂing, R.J., 113, 120, 183, 274 Shaked, M., 26, 271 Shapiro, A., 228, 265, 274 Shapiro, S.S., 169, 170, 274 Silvapulle, M.J., 104, 274 Singh, K., 244, 274 Sinha, B.K., 41, 227, 269, 273 Siotani, M., 231, 269, 274 Small, N.J.H., 170, 171, 185, 266, 274 Song, D., 50, 268 Spivak, M., 28, 29, 33, 274 Srivastava, M.S., 12, 13, 30, 31, 84, 104, 119, 123, 137, 154, 169, 170, 172, 184, 233, 234, 236, 247, 249–251, 264, 265, 267, 270, 272–275 Stadje, W., 86, 275 Stahel, W.A., 83, 275

280

Author Index

Statistical Sciences, 226, 275 Stein, C., 88, 275 Steyn, H.S, 208, 275 Stone, M., 156, 275 Strang, G., 1, 275 Stuart, A., 257, 270 Styan, G.P.H., 214, 272 Sugiura, N., 123, 127, 128, 132, 133, 140, 275 Sutradhar, B.C., 236, 275 Symanowski, J.T., 26, 270 Szablowski, P.J., 50, 275 Tang, D., 103, 275 Terui, N., 171, 270 Tibshirani, R.J., 243 Tricomi, F.G., 115, 116, 119, 196, 202, 267 Tsay, R.S., 171, 270 Tyler, D.E., 132, 134, 137, 190, 210, 214, 215, 218, 224, 226, 228, 231, 249, 266, 267, 270, 275, 276 Uhlig, H., 94, 276 van der Merwe, A., 156, 158, 276 van Zomeren, B.C., 262, 273 Varga, T., 74, 268 von Mises, R., 72, 276 von Rosen, D., 154, 259, 270, 274 Wagner, T., 171, 269 Wakaki, H., 98, 276 Wang, Y., 104, 276 Warner, J.L., 94, 170, 263 Watamori, Y., 72, 268 Waternaux, C.M., 132, 190, 228, 272, 276 Watson, G.S., 72, 276 Wichern, D.W., 269 Wielandt, H., 134, 276 Wilk, M.B., 169, 170, 274 Wilks, S.S., 186, 276 Wolfram, S., 197, 276

Wong, C.S., 74, 276 Yamato, H., 48, 276 Yau, W.K., 184, 275 Yohai, V.J., 224, 273 Zhu, L.-X., 49, 267 Zidek, J.V., 156, 158, 276 Zirkler, B., 171, 269

Subject Index

a.e., 23 absolutely continuous, 23 adjoint, 5 adjusted LRT, 228 aﬃne equivariant, 209 Akaike’s criterion, 190 almost everywhere, 23 ancillary statistic, 118 angular gaussian distribution, 70 asymptotic distribution bootstrap, 243 canonical correlations, 189 correlation coeﬃcient, 81, 82, 230 eigenvalues of R, 168, 242 eigenvalues of S, 130, 242 eigenvalues of S−1 1 S2 , 133 elliptical MLE, 221 Hotelling-T 2 , 101 M estimate, 223 multiple correlation, 112, 230 normal MLE, 213 partial correlation, 117, 230 S estimate, 225 sample mean, 77, 78

sample variance, 80 with multiple eigenvalues, 136 Bartlett correction factor, 199 Bartlett decomposition, 11, 31 basis orthonormal, 3 Basu, 118 Bernoulli numbers, 201 polynomials, 196 trial, 17 beta function, 39 multivariate, 38 univariate, 39 blue multiple regression, 65 multivariate regression, 146 bootstrap correlation coeﬃcient, 248 eigenvalues, 137, 248 means with l1 -norm, 245 means with l∞ -norm, 246 Box-Cox transformation, 94

282

Subject Index

breakdown point, 224 C.E.T., 15, 16 Cr inequality, 33 canonical correlation, 175, 189 Fc distribution, 42 variables, 175 Caratheodory extension theorem, 15, 16 Cauchy-Schwarz inequality, 2 central limit theorem, 78 chain rule for derivatives, 29 change of variables, 29 characteristic function, 21 χ2m , 37 gamma(p, θ), 37 general normal, 45 inversion formula, 21, 24, 254 multivariate normal, 56 uniqueness, 21 Wishart, 90 chi-square, 37 commutation matrix, 75, 81 conditional distributions Dirichlet, 40 elliptical, 208 normal, 62 conditional mean formula, 20 conditional transformations, 31 conditional variance formula, 20 conditions D, 215 D1, 218 E, 228 H, 228 M, 215 M1-M4, 222 S1-S3, 224, 225 conﬁdence ellipsoid, 104, 105, 108 contour, 58 convergence theorems dominated, 18 monotone, 18 convolution, 253

copula, 26 Morgenstern, 34 correlation canonical, 175, 189 coeﬃcient, 67, 81, 82, 230, 248 interclass, 172 intraclass, 48, 64 matrix, 97, 230 multiple, 109 partial, 116 covariance, 20 Cram´er-Wold theorem, 21 cumulant, 80, 211, 256 d.f., 15 delta method, 79 density, 23 multivariate normal, 58 derivative, 28 chain rule, 29 determinant, 4 diagonalization, 6 diﬀerentiation with respect to matrix, 12 vector, 12 dimensionality, 190 Dirichlet, 38, 49 conditional, 40 marginal, 40 distribution absolutely continuous, 23 angular gaussian, 70 Bernoulli, 17 beta, 39 chi-square, 37 noncentral, 45 contaminated normal, 207 copula, 26 Dirichlet, 38, 49 conditional, 40 marginal, 40 discrete, 16 double exponential, 44 elliptical, 207 exchangeable, 47

Subject Index

exponential, 37 F , 42 noncentral, 45 Fc , 42 noncentral, 45, 52 Fisher-von Mises, 72 function, 15 gamma, 37 general normal, 45 inverted Wishart, 97 joint, 25 Kotz-type, 208 Langevin, 72 Laplace, 44 marginal, 25 multivariate Cauchy, 208 multivariate normal, 55 multivariate normal matrix, 74 density, 81 multivariate t, 207, 239 negative binomial, 110 nonsingular normal, 58 permutation invariant, 47 power exponential, 209 singular normal, 62 spherical, 48, 207 standard gamma, 36 standard normal, 44 symmetric, 43 t, 64 U (p; m, n), 150 unif(B n ), 49 unif(S n−1 ), 49 uniform, 24 unif(T n ), 39 Wishart, 87, 97 dominated convergence theorem, 18 double exponential, 44 Efron-Morris, 158 eigenvalue, 6 eigenvector, 6 elliptical distribution, 207 conditional, 208

283

consistency, 209 marginal, 208 empirical characteristic function, 171 empirical distribution, 244 equals-in-distribution, 16 equidistributed, 16 equivariant estimates, 209 estimate blue, 65, 146 M, 222 S, 224, 262 etr, 81 euclidian norm, 2 exchangeable, 47 expected value, 18 of a matrix, 19 of a vector, 19 of an indexed array, 19 exponential distribution scaled, 37 standard, 37 F distribution, 42 Fc distribution, 42 density, 42 familial data, 83, 172 FICYREG, 158 Fisher z-transform, 82, 114, 117, 248 Fisher’s information, 219 Fisher-von Mises distribution, 72 ﬂattening, 157 Gn , 29 gamma function, 36 generalized, 94 scaled, 37 standard, 36 Gauss-Markov multiple regression, 65 multivariate regression, 146 general linear group, 29 general linear hypothesis, 144

284

Subject Index

general normal, 45 generalized gamma function, 94 generalized variance, 93, 96 goodness-of-ﬁt, 72, 171 Gram-Schmidt, 3 group general linear, 29 orthogonal, 8, 48 permutation, 47 rotation, 48 triangular, 10 H¨ older’s inequality, 19 hermitian matrix, 6 transpose, 6 Hotelling-T 2 , 98 one-sided, 103 two-sample, 138 hypergeometric function, 115

|=

i.i.d., 28 iﬀ, 2 image space, 4 imputation, 221 indep ∼ , 38 independence mutual, 27 pairwise , 27 pairwise vs mutual, 34 test, 177, 192, 203 inequality between matrices, 8 Cr , 33 Cauchy-Schwarz, 2 H¨ older, 19 inner product matrix, 5 of complex vectors, 6 of matrices, 145 of vectors, 2 interclass correlation, 172 interpoint distance, 194 intraclass correlation, 48, 64

invariant tests, 102, 120, 122, 138, 140, 151, 178 inverse, 4 partitioned matrix, 11 inversion formulas, 21, 24, 254 inverted Wishart, 97 jacobian, 29, 75 joint distribution, 25 Kendall’s τ , 34 kernel, 4 Kronecker δ, 3 product, 74 Kummer’s formula, 115 kurtosis, 171, 212, 259 parameter, 212 L+ n , 10 Lp , 18 Langevin distribution, 72 Lawley-Hotelling trace test, 154 LBI test for sphericity, 121 least-squares estimate, 66 Lebesgue measure, 23 Leibniz notation, 18 length of a vector, 2 likelihood ratio test asymptotic, 118 linear estimation, 65 linear hypothesis, 144 log-likelihood, 213 LRT, 99 M estimate, 222 asymptotic, 223 Mahalanobis distance, 58, 170, 184, 206, 262 Mallow’s criterion, 190 MANOVA one-way, 159 marginal distribution, 25 matrices, 2 adjoint, 5

Subject Index

commutation, 75, 81 determinant, 4 diagonalization, 6, 7 eigenvalue, 6 eigenvector, 6 hermitian, 6 hermitian transpose, 6 idempotent, 9 image space, 4 inverse, 4 kernel, 4 Kronecker product, 74 nonsingular, 4 nullity, 4 orthogonal, 8, 48 positive deﬁnite, 8 positive semideﬁnite, 8 product, 2 rank, 4 singular value, 8 skew-symmetric, 35 square, 2 square root, 8 symmetric, 2 trace, 2 transpose, 2 triangular, 7 triangular decomposition, 11 unitary, 6 matrix diﬀerentiation, 12 Maxwell-Hershell theorem, 51, 227 mean, 19 minimum volume ellipsoid, 224 missing data, 221 mixture distribution, 21, 46, 53, 56, 111, 207, 209, 238 MLE (Σ, µ), 86, 96 multivariate regression, 147 modulus of a vector, 2 monotone convergence theorem, 18 multiple correlation, 109 asymptotic, 112, 230 invariance, 140 moments, 140

asymptotic, 115 MVUE, 115 multiple regression, 65 multivariate copula, 26 ﬂattening, 157 prediction, 156 regression, 144 multivariate distribution beta, 38 Cauchy, 208 contaminated normal, 207 cumulant, 80, 211, 256 Kotz-type, 208 normal, 55 contour, 58 density, 58 normal matrix, 74 conditional, 82 density, 81 power exponential, 209 t, 207, 239 with given marginals, 26 mutual independence, 27 MVUE R2 , 115 (Σ, µ), 86 Nt process, 37 negative binomial, 110 noncentral chi-square, 45 F , 45 Fc , 45 density, 52 nonsingular matrix, 4 normal, 58 normal general, 45 multivariate, 55 nonsingular, 58 singular, 62 standard, 44 nullity, 4

285

286

Subject Index

one-way classiﬁcation, 158 orthogonal complement, 3 group, 8, 48 matrix, 8, 48 projection, 9, 66 vectors, 2 orthogonal invariance, 48 outlier, 262

|=

Pn , 8 p.d.f., 23 p.f., 16 pairwise independence , 27 partial correlation, 116 asymptotic, 117, 230 permutation group, 47 invariance, 47 perturbation method, 125 Pillai trace test, 154 Poisson process Nt , 37 polar coordinates, 32, 50, 54 positive deﬁnite, 8 semideﬁnite, 8 power transformations, 94 prediction, 156 prediction risk, 156, 157 principal components deﬁnition, 162 sample, 165, 169 probability density function, 23 function, 16 product-moment, 19 projection mutually orthogonal, 10 orthogonal, 9, 66 proportionality test, 139 PSn , 8 Q-Q plot of squared radii, 186 quadratic forms, 66, 67

Radon-Nikodym theorem, 23 rank, 4 Rayleigh’s quotient, 13 rectangles, 15 reﬂection symmetry, 43 regression multiple, 65 multivariate, 144 relative eﬃciency, 236, 262 robust estimates, 222 M type, 222 S type, 224 robustness Hotelling-T 2 , 101, 226 tests on scale matrix, 227 rotation group, 48 rotationally invariant matrix, 210 vector, 49 Roy largest eigenvalue, 154 S estimate, 224, 262 asymptotic, 225 S n−1 , 21, 33 Sn , 47 Sp , 134 sample matrix, 75 sample mean asymptotic, 77, 78 sample variance, 77 asymptotic, 80, 213 scaled distribution exponential, 37 gamma, 37 scaled residuals, 171, 184 score function, 219 Shapiro-Wilk test, 169 simultaneous conﬁdence intervals asymptotic, 109, 139 Bonferroni, 107 eigenvalues by bootstrap, 248 for φ(Σ), 109 linear hypotheses, 104 means by bootstrap, 246 nonlinear hypotheses, 107

Subject Index

robust, 227 Roy-Bose, 106, 139 Scheﬀ´e, 106, 139 singular normal, 62 value, 8 skew-symmetric matrix, 35 skewness, 171, 259 Slutsky theorem, 78 span, 4 SPE prediction risk, 157 Spearman’s ρ, 34 spectral decomposition, 8 SPER prediction risk, 156 spherical distribution, 48, 207 characteristic function, 52 density, 52 density of radius, 223 density of squared radius, 54 square root matrix S1/2 , 8 standard distribution exponential, 37 gamma, 36 normal, 44 statistically independent, 27 Sugiura’s lemma, 127 SVD, 8 symmetric distribution, 43 matrix, 2 t distribution, 64 T 2 of Hotelling, 98 T n , 39 Tn , 38 test equality of means, 159 equality of means and variances, 141, 205 equality of variances, 121, 204 for a given mean, 99 for a given mean vector and variance, 236 for a given variance, 139, 205, 233, 240

Hotelling two-sample, 138 Hotelling-T 2 , 98 independence, 177, 192, 203 Lawley-Hotelling, 154 linear hypothesis, 148, 201 multiple correlation, 110, 241 multivariate normality, 169 Pillai, 154 proportionality, 139 Roy, 154 sphericity, 117, 200 symmetry, 138 total variance, 162 triangular decomposition, 11 group, 10 matrix, 7 U (p; m, n), 150, 261 asymptotic, 184, 201 characterizations, 182 duality, 182 moments, 190 U+ n , 10 UMPI test for multiple correlation, 112 Hotelling-T 2 , 103 unif(B n ), 49 unif(S n−1 ), 49 uniform distribution, 24 unif(T n ), 39 union-intersection test, 160 unit sphere, 21, 33 unitary matrix, 6 uvec operator, 247 variance, 19 generalized, 93, 96 of a matrix, 74 sample, 77 total, 162 vec operator, 73 vector diﬀerentiation, 12 vectors column, 1

287

288

Subject Index

inner product, 2 length, 2 modulus, 2 orthogonal, 2 orthonormal, 3 outer product, 3 row, 1 volume, 23 w.p.1, 52 waiting time process Tn , 38 Wielandt’s inequality, 134

Wishart, 87 characteristic function, 90 density, 93, 97 linear transformation, 88 marginals, 90, 92 moments and cumulants, 259 noninteger degree of freedom, 94 nonsingular, 87 sums, 91

Our partners will collect data and use cookies for ad personalization and measurement. Learn how we and our ad partner Google, collect and use data. Agree & close