Classics in Mathematics Richard S. Ellis
Entropy, Large Deviations, and Statistical Mechanics
Richard S. Ellis
Entr...
28 downloads
338 Views
15MB Size
Report
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!
Report copyright / DMCA form
Classics in Mathematics Richard S. Ellis
Entropy, Large Deviations, and Statistical Mechanics
Richard S. Ellis
Entropy, Large Deviations, and Statistical Meclianics Reprint ofthe 1985 Edition
^ S p rringer i
Richard S. Ellis Department of Mathematics and Statistics University of Massachusetts Amherst, Massachusetts 01003 U.S.A.
Originally published as Vol. 271 in the series Grundlehren der mathematischen Wissenschaften
Mathematics Subject Classification (2000): 82A05,60K35
Library of Congress Control Number: 2005934786
ISSN 14310821 ISBN10 3540290591 Springer Berlin Heidelberg New York ISBN13 9783540290599 Springer Berlin Heidelberg New York This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9,1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable for prosecution under the German Copyright Law. Springer is a part of Springer Science+Business Media springeronline.com © SpringerVerlag Berlin Heidelberg 2006 Printed in Germany The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exemptfiromthe relevant protective laws and regulations and therefore free for general use. Production: LETEX Jelonek, Schmidt & Vockler GbR, Leipzig Cover design: design & production GmbH, Heidelberg Printed on acidfree paper
41/3142/YL  5 4 3 2 1 0
Grundlehren der mathematischen Wissenschaften 271 A Series of Comprehensive Studies in Mathematics
Editors M. Artin S. S. Chern J. M. Frohlich E. Heinz H. Hironaka F. Hirzebruch L. Hormander S. Mac Lane W. Magnus C. C. Moore J. K. Moser M. Nagata W. Schmidt D. S. Scott Ya. G. Sinai J. Tits B. L. van der Waerden M. Waldschmidt S. Watanabe Managing Editors M. Berger B. Eckmann
S. R. S. Varadhan
Grundlehren der mathematischen Wissenschaften A Series of Comprehensive Studies in Mathematics
A Selection 190. 191. 192. 193. 194. 195. 196. 197. 198. 199. 200. 201. 202. 203. 204. 205. 206. 207. 208. 209. 210. 211. 212. 213. 214. 215. 216. 217. 218. 219. 220. 221. 222. 223. 224. 225. 226. 227. 228. 229. 230. 231. 232. 233. 234.
Faith: Algebra: Rings, Modules, and Categories I Faith: Algebra II, Ring Theory Mal'cev: Algebraic Systems Polya/Szego: Problems and Theorems in Analysis I Igusa: Theta Functions Berberian: Baer*Rings Athreya/Ney: Branching Processes Benz: Vorlesungen iiber Geometric der Algebren Gaal: Linear Analysis and Representation Theory Nitsche: Vorlesungen iiber Minimalflachen Dold: Lectures on Algebraic Topology Beck: Continuous Flows in the Plane Schmetterer: Introduction to Mathematical Statistics Schoeneberg: Elliptic Modular Functions Popov: Hyperstability of Control Systems Nikol'skii: Approximation of Functions of Several Variables and Imbedding Theorems Andre: Homologie des Algebres Commutatives Donoghue: Monotone Matrix Functions and Analytic Continuation Lacey: The Isometric Theory of Classical Banach Spaces Ringel: Map Color Theorem Gihman/Skorohod: The Theory of Stochastic Processes I Comfort/Negrepontis: The Theory of Ultrafilters Switzer: Algebraic Topology—Homotopy and Homology Shafarevich: Basic Algebraic Geometry van der Waerden: Group Theory and Quantum Mechanics Schaefer: Banach Lattices and Positive Operators Polya/Szego: Problems and Theorems in Analysis II Stenstrom: Rings of Quotients Gihman/Skorohod: The Theory of Stochastic Process II Duvant/Lions: Inequalities in Mechanics and Physics Kirillov: Elements of the Theory of Representations Mumford: Algebraic Geometry I: Complex Projective Varieties Lang: Introduction to Modular Forms Bergh/Lofstrdm: Interpolation Spaces. An Introduction Gilbarg/Trudinger: Elliptic Partial Differential Equations of Second Order Schutte: Proof Theory Karoubi: KTheory, An Introduction Grauert/Remmert: Theorie der Steinschen Raume Segal/Kunze: Integrals and Operators Hasse: Number Theory Klingenberg: Lectures on Closed Geodesies Lang: Elliptic Curves: Diophantine Analysis Gihman/Skorohod: The Theory of Stochastic Processes III Stroock/Varadhan: Multidimensional Diffusion Processes Aigner: Combinatorial Theory Continued after Index
For Alison, Melissa, and Michael my wife and children who make it all worthwhile
Preface
This book has two main topics: large deviations and equiHbrium statistical mechanics. I hope to convince the reader that these topics have many points of contact and that in being treated together, they enrich each other. Entropy, in its various guises, is their common core. The large deviation theory which is developed in this book focuses upon convergence properties of certain stochastic systems. An elementary example is the weak law of large numbers. For each positive a, P{S„/n > e} converges to zero as n ^ oo, where S„ is the nth partial sum of independent identically distributed random variables with zero mean. Large deviation theory shows that if the random variables are exponentially bounded, then the probabihties converge to zero exponentially fast as n ^ oo. The exponential decay allows one to prove the stronger property of almost sure convergence {S„/n ^ 0 a.s.). This example will be generalized extensively in the book. We will treat a large class of stochastic systems which involve both independent and dependent random variables and which have the following features: probabilities converge to zero exponentially fast as the size of the system increases; the exponential decay leads to strong convergence properties of the system. The most fascinating aspect of the theory is that the exponential decay rates are computable in terms of entropy functions. This identification between entropy and decay rates of large deviation probabihties enhances the theory significantly. Entropy functions have their roots in statistical mechanics. They originated in the work of L. Boltzmann, who in the 1870's studied the relation between entropy and probability in physical systems. Thus statistical mechanics has a strong historical connection with large deviation theory. It also provides a natural context in which the theory can be appHed. Applications of large deviations to models in equilibrium statistical mechanics are presented in Chapters IIIV. These appUcations illustrate convincingly the power of the theory. Equilibrium statistical mechanics is an exciting area of mathematical physics but one which remains inaccessible to many mathematicians. Some texts on the subject provide an introduction to the physics but do not develop the mathematics in much detail or with great rigor. Other texts treat mathematical problems in statistical mechanics with complete rigor but assume an
Vlli
Preface
extensive background in the physics. The uninitiated reader has difficulty understanding how concepts hke ensemble, free energy, or entropy connect up with more familiar concepts in mathematics. My approach in this book is to emphasize strongly the connections between statistical mechanics on the one hand and probability and large deviations on the other. I hope that in so doing, I have succeeded in providing a readable treatment of statistical mechanics which is accessible to a general mathematical audience. My large deviation approach to statistical mechanics was inspired in part by the article ofO. E. Lanford(1973). In recent years, the scope of large deviations has been greatly expanded by M. D. Donsker and S. R. S. Varadhan. This book contains an introduction to their theory. I illustrate the main features in the context of independent identically distributed random vectors taking values in U^. I also present my own large deviation results, which are particularly suited for applications to statistical mechanics. Since readability rather than completeness has been my goal, the large deviation theorems are not stated in the greatest generality. There are two parts to the book. Part I consisting of Chapters IV. Chapter I introduces large deviations by means of elementary examples involving combinatorics and Stirling's formula. Chapter II presents the DonskerVaradhan theory as well as my own large deviation results. The proofs of the theorems in this chapter are detailed and are postponed until Part II. Postponing proofs allows the reader to reach, as soon as possible, interesting apphcations of large deviations to statistical mechanics in Chapters IIIV. Chapter III gives a large deviation analysis of a discrete gas model. Chapters IVV discuss the Ising model of ferromagnetism and related spin systems. The emphasis in these two chapters is upon properties of Gibbs states. While large deviation theory provides a terminology and a set of results that are useful for treating Gibbs states, the book also develops other tools that are needed. These include convexity and moment inequalities. Part II consists of Chapters VIIX. Chapter VI is a summary of the theory of convex functions on IR^. Chapters VIIIX prove the large deviation results stated in Chapter II without proof. The prerequisite for these chapters is a good working knowledge of probability and measure theory. The essential definitions and theorems in probability are listed in Appendix A. The appendix is intended to be a review or an outline for study rather than a detailed exposition. This book can be used as a text. It contains over 100 problems, many of which have hints. Chapters I and II and VIIX are a selfcontained treatment of large deviations and convex functions. Readers primarily interested in spin systems can concentrate upon Chapters IV and V and refer to the statements and proofs of large deviation results as needed. Those portions of Chapters IV and V which do not rely on large deviations are selfcontained. Chapters IV and V can be completely understood without reading Chapter III. This book contains new results and new proofs of known theorems. These include the following: exponential convergence properties of Gibbs states
Preface
IX
[Theorems IV.5.5, IV.6.6, and V.6.1]; a large deviation proof of the Gibbs variational formula [Theorem IV.7.3(a)]; a proof of the central limit theorem for spin systems [Theorem V.7.2(a)]; a level3 large deviation theorem for i.i.d. random variables with a finite state space [Theorem IX.1.1]; a level3 large deviation theorem for Markov chains with a finite state space [Problems IX.6.10IX.6.15]; the solution of the Gibbs variational formula for finiterange interactions on Z via large deviations [Appendix C.6]. Many of the large deviation results and appHcations in the book depend upon my large deviation theorem, Theorem II.6.1. The proof of the level3 theorem in Chapter IX was inspired by statistical mechanics [see Appendix C.6] and information theory. I have had the good fortune of interacting with a number of special people. Todd Baker edited the manuscript with creativity and care. The book benefited greatly from his involvement. Peg Bombardier was my superb typist. She was always cheerful and patient, despite the numerous revisions, and was a pleasure to work with. Alan Sokal read portions of the manuscript and was a big help with the statistical mechanics. I owe a special debt of gratitude to Srinivasa Varadhan. He answered my many questions about large deviations patiently and with insight and showed a strong interest in the book. The encouragement of my family and friends was greatly appreciated. Above all, I thank my wife Alison. Her love is a blessing. I am grateful to Alejandro de Acosta, HansOtto Georgii, Joseph Horowitz, Jonathan Machta, Charles Newman, and R. Tyrrell Rockafellar for reading portions of the manuscript and suggesting improvements. I am also indebted to the many other people, too numerous to mention by name, with whom I have consulted. While writing the book, I received support from the University of Massachusetts, the National Science Foundation, and the Lady Davis Fellowship Trust. Their support is gratefully acknowledged. Richard S. ElUs
Comments on the Use of This Book
At the end of each chapter, there is a Notes section, followed by a Problems section. References to the Notes are indicated by superscripted integers; e.g., entropy"^ refers to Note 4. Near the end of the book, there is a Hst of frequently used symbols. The main large deviation theorems are stated in Chapter II and are proved in Chapters VIIX. Readers interested primarily in large deviations may read Chapters I, II, VIIX while those interested primarily in statistical mechanics may read Chapters IV,
Contents
Preface
vii
Comments on the Use of This Book PART I:
LARGE DEVIATIONS AND STATISTICAL MECHANICS
Chapter I.
Introduction to Large Deviations
I.L Overview 1.2. Large Deviations for I.I.D. Random Variables with a Finite State Space 1.3. Levels1 and 2 for Coin Tossing 1.4. Levels1 and 2 for I.I.D. Random Variables with a Finite State Space 1.5. Level3: Empirical Pair Measure 1.6. Level3: Empirical Process 1.7. Notes 1.8. Problems Chapter II.
Large Deviation Property and Asymptotics of Integrals
II. 1. 11.2. 11.3. 11.4. 11.5. 11.6.
Introduction Levels1, 2, and 3 Large Deviations for I.I.D. Random Vectors The Definition of Large Deviation Property Statement of Large Deviation Properties for Levels1, 2, and 3 Contraction Principles Large Deviation Property for Random Vectors and Exponential Convergence 11.7. Varadhan's Theorem on the Asymptotics of Integrals 11.8. Notes 11.9. Problems
Chapter III. 111.1. 111.2. 111.3.
x
Large Deviations and the Discrete Ideal Gas
Introduction Physics Prelude: Thermodynamics The Discrete Ideal Gas and the Microcanonical Ensemble
3 3 9 11 13 17 21 26 27 30 30 30 33 37 42 46 50 55 56 59 59 60 64
Xll
Contents
III.4.
Thermodynamic Limit, Exponential Convergence, and Equilibrium Values IIL5. The MaxwellBoltzmann Distribution and Temperature 111.6. The Canonical Ensemble and Its Equivalence with the Microcanonical Ensemble 111.7. A Derivation of a Thermodynamic Equation 111.8. The Gibbs Variational Formula and Principle 111.9. Notes III. 10. Problems Chapter IV. IV.l. IV.2. IV.3. IV.4. IV.5. IV.6. IV.7. IV.8. IV.9.
Introduction An Overview of Ferromagnetic Models FiniteVolume Gibbs States on Z Spontaneous Magnetization for the CurieWeiss Model Spontaneous Magnetization for General Ferromagnets on Z InfiniteVolume Gibbs States and Phase Transitions The Gibbs Variational Formula and Principle Notes Problems
Chapter V. V.l. V.2. V.3. V.4. V.5. V.6. V.7. V.8. V.9. V. 10. V.U. V.12. V.l 3.
Ferromagnetic Models on Z
Magnetic Models on Z^ and on the Circle
Introduction FiniteVolume Gibbs States on Z^, D > 1 Moment Inequalities Properties of the Magnetization and the Gibbs Free Energy Spontaneous Magnetization on iP, D >2, Via the Peierls Argument InfiniteVolume Gibbs States and Phase Transitions InfiniteVolume Gibbs States and the Central Limit Theorem Critical Phenomena and the Breakdown of the Central Limit Theorem Three Faces of the CurieWeiss Model The Circle Model and Random Waves A Postscript on Magnetic Models Notes Problems
66 74 75 78 79 84 86 88 88 88 94 98 103 109 125 131 134 138 138 140 142 146 153 157 162 170 179 190 198 199 203
PART 11: CONVEXITY AND PROOFS OF LARGE DEVIATION THEOREMS Chapter VI. Convex Functions and the LegendreFenchel Transform 211 VI. 1. Introduction VI.2. Basic Definitions VI.3. Properties of Convex Functions
211 211 213
Contents VI.4. VI.5. VI.6. VL7.
Xlll A OneDimensional Example of the LegendreFenchel Transform The LegendreFenchel Transform for Convex Functions on W^ Notes Problems
Chapter VII. VII.L VII.2. VII.3. VIL4. VII.5. VII.6. VII.7. VII.8.
Large Deviations for R a n d o m Vectors
229
Statement of Results Properties of I^ Proof of the Large Deviation Bounds for J = 1 Proof of the Large Deviation Bounds for d > 1 Level1 Large Deviations for I.I.D. Random Vectors Exponential Convergence and Proof of Theorem II.6.3 Notes Problems
229 231 232 235 238 242 243 245
Chapter VIII. VIII. L VIII.2. VIII.3. VIII.4. VIII.5. VIII.6.
Level2 Large Deviations for I.I.D. R a n d o m Vectors
Introduction The Level2 Large Deviation Theorem The Contraction Principle Relating Levels1 and 2 (^ = 1) The Contraction Principle Relating Levels1 and 2(d>2) Notes Problems
Chapter IX. IX. 1. IX.2. IX.3. IX.4. 1X5. IX.6.
218 220 225 226
Level3 Large Deviations for I.I.D. R a n d o m Vectors
Statement of Results Properties of the Level3 Entropy Function Contraction Principles Proof of the Level3 Large Deviation Bounds Notes Problems
250 250 250 253 258 264 265
269 269 270 276 279 287 288
APPENDICES Appendix A: A.l. A.2. A.3. A.4. A.5. A.6. A.7.
Probability
Introduction Measurability Product Spaces ProbabiUty Measures and Expectation Convergence of Random Vectors Conditional Expectation, Conditional Probabihty, and Regular Conditional Distribution The Kolmogorov Existence Theorem
295 295 295 296 297 299 299 301
XIV
A.8. A.9. A. 10. A. 11.
Contents
Weak Convergence of Probability Measures on a Metric Space The Space J^.dR'^f) and the Ergodic Theorem ^Dependent Markov Chains Probability Measures on the Space {1,1}^
Appendix B: Proofs of Two Theorems in Section II.7 B. 1. Proof of Theorem IL7.1 B.2. Proof of Theorem 11.7.2 Appendix C: C.l. C.2. C.3. C.4. C.5. C.6.
Equivalent Notions of InfiniteVolume Measures for Spin Systems
Introduction TwoBody Interactions and InfiniteVolume Gibbs States ManyBody Interactions and InfiniteVolume Gibbs States DLR States The Gibbs Variational Formula and Principle Solution of the Gibbs Variational Formula for FiniteRange Interactions on Z
Appendix D:
Existence of the Specific Gibbs Free Energy
D.l. Existence Along Hypercubes D.2. An Extension
303 306 311 312 319 319 321
323 323 323 325 326 327 330 332 332 334
List of Frequently Used Symbols
335
References
338
Author Index
353
Subject Index
359
Parti
Large Deviations and Statistical Mechanics
Chapter I
Introduction to Large Deviations
I.l.
Overview
One of the common themes of probabihty theory and statistical mechanics is the discovery of regularity in the midst of chaos. The laws of probabihty theory, which include laws of large numbers and central limit theorems, summarize the behavior of a stochastic system in terms of a few parameters (e.g., mean and variance). In statistical mechanics, one derives macroscopic properties of a substance from a probabihty distribution that describes the complicated interactions among the individual constitueht particles. A central concept Hnking the two fields is entropy. ^ The term was introduced into thermodynamics by Clausius in 1865 after many years of intensive work by him and others on the second law of thermodynamics. An early important step in its development and enrichment was the discovery by Boltzmann of a statistical interpretaton of entropy. Boltzmann's discovery, which was published in 1877, has three parts. We have augmented part (c) to include the possibility of phase transitions. (a)
Entropy is a measure of randomness or disorder in a statistical mechanical system. (b) If S is the entropy for a system in a given state and Wis the "thermodynamical probabihty" of that state,* then S = klog W, where ^ is a positive physical constant. (c) The equilibrium states, which are the states of the system observed in nature, are those states with the largest thermodynamical probability and thus the largest entropy. By (a), they are the "most random" states of the system consistent with any constraints which the system must satisfy (e.g., conservation of energy). The existence of more than one equilibrium state corresponds to a phase transition. All the notions of entropy discussed in this book are variations on the Boltzmann theme. In analyzing stochastic or statistical mechanical systems, *The thermodynamical probability is defined to be the number of microstates compatible with the given state (see Wehrl (1978, page 223)).
4
I. Introduction to Large Deviations
one must extrapolate from a microscopic level, on which the system is defined, to a macroscopic level, on which the laws describing the behavior of the system are formulated. Boltzmann shows that entropy is the bridge between these two levels. We will illustrate these ideas and outline the main themes of this book in terms of a basic stochastic model. This discussion is intended for the reader who has a knowledge of probability theory consistent with Appendix A. Other readers may still perceive the global picture without turning to Appendix A at this time. That task may be postponed until the end of the first chapter. Each of our systems is modeled microscopically by a collection of random variables {X^; a e j ^ } which are defined on a probability space (Q, J^, P) and which take values in a space F. Q is a nonempty set, J*^ is a crfield of subsets of Q, and P is a probability measure on (Q, 0. /(z) is called the entropy function for the coin tossing model. We now interpret it in view of our earUer remarks on entropy. For any zeU and large n, QniA^^e} is approximately exp( —«J^(z,e)). Since F(z, s) ^ I(z) as e > 0, we may heuristically write (1.2)
e.Mz,a}^exp(«/(z))
for n large and e small. If z =/= m^, then /(z) is positive and exp( —^2/(z)) ^ 0 as n^ CO. This is consistent with the exponential decay of Qn{A^^^} for 0 < 8 < \z — mp\. The heuristic formula (1.2) shows that to a small value of /(z) there corresponds a large probability Qni^z^e) c>r, in other words, a high multiplicity of microstates. In this sense, /(z) is a measure of the multiplicity of microstates compatible with the macrostate z. For another interpretation, given points z^ and Z2 in [0,1], it is reasonable to call z^ more random than Z2 if/(^i) < /(^2)j that is, if there are more microstates compatible with z^ than with Z2 Thus /(z) also measures the randomness of the macrostate z. The equiUbrium state m^ = ^ is that macrostate which is compatible with the most microstates. In fact, I(m^) = 0 = min{/(z):zeU}
and
/(z) > 0
for z ^ m^.
Thus the equilibrium state, being the unique minimum point of/, is the most random macrostate. Points z outside of [0,1] are forbidden values for SJn: 'l^Qn{A,,e} = 0, then set \ogQ„{A.,e} =
1.1. Overview
7
if ^ n [0,1] is empty, then Q^{A} = 0 and /(z) = oo for each zeA. For z ^ nip and 0 < e < z — m^, Qn{^z,e} is called a /ar^^ deviation probability since the event {COGQ: »S„(CO)/WG^^ J corresponds to a fluctuation or deviation of SJn of order z — m^l away from the Hmiting mean. It is a very rare event since Qn{A^^^} » 0 exponentially. An equivalent statement of the WLLN is that the distributions {Q„; « = 1,2, . . . } converge weakly to the unit point measure at m^ (written Qn =^^m )• ^^ this book, we will study much more general but analogous situations. A sequence of probability measures {Q„',n = 1,2, . . . } on a complete separable metric space ^ will converge weakly to the unit point measure at some XQE^. {Q„} will have a large deviation property in the sense that Qn{K} will decay exponentially for all closed sets A^ in ^ which do not contain XQ. The decay rate will be given by —mt^^j^I(x), where I(x) is some nonnegative function on ^ with a unique minimum point at XQ (I(Xf)) = 0). I(x) is called the entropy function of the measures {Qn}. In a series of important papers beginning in 1975, Donsker and Yaradhan have identified three levels of large deviations which fit into the general framework just described. These levels will be treated in detail in this book. Let {1} ;7 G Z } be a sequence of i.i.d. random vectors taking values in IR"^. Let p be the distribution of X^ and Pp the corresponding infinite product measure on Q = {U'f. Level1. Define g^^^ to be the distribudon of 5„(a;)/« = Yj=i ^j(^)/^ on IR^ and assume that JiRrfxp(Jx) is finite. Then by the WLLN the sequence iQjji); n = 1,2,...} converges weakly to 3^ , where m^ is the mean ^^dxp(dx) of p. If, furthermore, the moment generating function ^^dQxp 0 Urn P
{CDEQ:
\S„(co)/n  m \ > e} = 0,
(1.7) lim P { C O G Q : n^oo
max
L„ ^(co) — pJ > e} = 0.
i= l,...,r
'
The vector (pi,p2? • • • ?Pr) is the hmiting mean of the random vector (L„ j(co), . . . ,L„ X^)). The probabilities in (1.7) represent large deviations since they involve fluctuations of order s of the respective /2sums away from the limiting means, and s is fixed. Below we show by elementary combinatorial arguments that each of these probabilities decays exponentially.
1.3.
Levels1 and 2 for Coin Tossing
Coin tossing is defined by the state space T = (0,1} and the measure p = 1^0 + i^i The value 0 represents a tail and 1 a head. The proof for this simple case will set a pattern of proof for the more general large deviation results which follow. We have p^ = p2= ^, m^ = i , L„ ^(co) = 1 — 5„(co)/«, and L„ 2( s}.
Let Q[^^ be the P^distribution of SJn on U and define the closed set A = {zeU: \z — m^l > e},
where 0 < e < ^.
The set yl n [0,1] is nonempty and Ql^^^{A} = Pp{\SJn — m^l > e} is positive for all sufficiently large n. Since A does not contain m^, Ql^^^{A} ^ 0. According to the next theorem, Q[^^{A} decays exponentially, and the decay rate is given in terms of the entropy function (1 9)
I^'\z) = ^'^""^^^'^ + ^^ ~ ^)^^g(2(l  z)) 00
for Z G [ 0 , 1], for z^ [0,1],
where 01ogO = 0. Ip^\z) is convex, is symmetric about z = w^ = ^, and attains its minimum value of 0 at the unique point z = mp. /j^Ms depicted in Figure LI.
12
I. Introduction to Large Deviations
Theorem 1.3.1. limUogQi'^A} n^oofl
= ]imhogP^{\SJn n^ooH
 m j > s} =
^
'^
mml^'Xz). zeA
Since the set A is closed and does not contain m^^min^^^ I^p^\z) > I^J^Xm^ = 0. Hence Q\I~\A] converges to zero exponentially fast as n^ oo. Proof. Let Q„ be the finite configuration space consisting of all sequences CO = (a>i,CO2, ...,co„), with each COJEF = {0,1}; thus Q„ = F". If n^P^ is the finite product measure on Q„ with identical onedimensional marginals p, then Q^„^^{A} = n„P^{coeQ„: S„{co)/neA}, For fixed n and coeQ„, S„(co) may take any value ke{0,\, ... ,n}. Sn(co)/n is in A if and only if k is in the set A^ = {^6{0,1, ... ,n}: \k/n — i  > e}. For keA^, define C(n,k) = n\/(k\(n — k)\). There are C(n,k) points co in Q„ for which S^(oji) = k, and n„Pp{a>} = 2~" for each coeQ^. Hence (1.10) ei,^'{^}= I
7r„P,{coeQ„:5„(co)/« = A:/«}= ^ C(«,A:)^.
keA„
keAfj
^
Since there are no more than n + 1 terms in the sum, maxC(^,/:)^ < Q'^^^A) < (« + l)maxC(«,/:)4^, and since log is an increasing function (1.11)
max i l o g ( c ( „ , ; t ) l ) ] < i l o g e a , { ^ } < M ! L ± 1 ) + keA„
max log(C(«,A:)^ keA„
Thus the asymptotic behavior of Ql,^^{A} is governed by the asymptotic behavior of the largest summand in (1.10). Entropy arises by the following lemma. Lemma 1.3.2. Uniformly in ke [0,1, ..., n], UogC(n,k) n
=  ^ l o g ^  fl  ^ V o g f 1   V o(^^] n n \ n j \ n j \ n
j
as «  00.
Proof. Since C(n, 0) = C(n, n) = 1 and C{n, 1) = C(«, n — I) = n, the lemma holds for all ^2 > 1 and k = 0, \, n — I, n. A weak form of Stirling's approximation states that for all n>2, log(n\) = nlogn — n + P„, where \P„\ = 0(logn) [Problem L8.1]. Hence for 2 < /c < «  2, logC(n,k) n
= log«  '^logk  "^^login n n
 k) + kpn Pkn
Pnu)
1.4. Levels1 and 2 for I.I.D. Random Variables with a Finite State Space
13
Write . k, I n — k, I iogn = — l o g logn n n n and combine these terms with the other log terms to give
ilogC(«,.)=^log^(l4)log(l^) + l(^„^,^„.,). For 2 < / : < « — 2, the last term can be bounded by 0{n~^ log/t) uniformly in k. This completes the proof. n The lemma shows that
1,0. (C(,a)l) = ,«4  ^.0,^  (,  ^),og(l  ^) .
oi^^)
The first three terms are exactly —I^^\k/n), where Ip^Hs defined in (1.9). Thus
Since n~^ Iogn and n~^ \og(n + 1) both tend to zero, we have by (1.11) (1.13)
limloge,'>U} = I i m m a x f  / S ^ M  I l =  l i m min/i^M  I
For each «the set {z G [0, l ] : z = Z:/^ for some A: G ^ „ } is a subset of ^4 n [0,1]. Since Ip^\z) = oo for z$[0,1], we conclude using Problem 1.8.2 that
limlogel^U^}= n*aon
min zeAn[0,l]
h'\z)=minh'\z). ^
zsA
n ^
1.4. Levels1 and 2 for I.I.D. Random Variables w^ith a Finite State Space In general, the state space F equals {x^, ^2, .. ., x j , where x^ < X2< • • • < x^ are real numbers. The exponential decay rates of the two probabilities in (1.7) are expressed in terms of a function called the relative entropy. Let ^ ( F ) denote the set of all subsets of F and Ji{T) the set of probabihty measures on ^ ( F ) . Each veJiiT) has the form v^^^^Vj^ where v^> 0 and Yli=\ ^i = 1  ^ ( r ) may be identified with the compact convex subset of W consisting of all vectors v = (v^, . . . , v^) which satisfy v^ > 0 and Y!i=i "^t— ^• The relative entropy^ of v with respect to the measure p = Y!i=i Pi^x (Pt > ^) is defined by I(^^\v) = ^ V, log^ i=l
We have the following properties.
Pi
where 0 log 0 = 0.
14
I. Introduction to Large Deviations
Proposition 1.4.1. (a) Ip^\v) is a convex function ofveJiiY). (b) Ip^\v) measures the discrepancy between v and p in the sense that Ip^Kv) > 0 with equality if and only if v = p. Thus Ip^\v) attains its infimum over Ji{T) at the unique measure v = p. Proof (a) I^^\v) equals YJi^'^^^^tlPi^Ph where h{x) is the convex function xlogx, X > 0. Let pi and v be probability measures on ^ ( F ) . Then for 0 e } i = l , . . .,r
where 0 < £
s}
where 0 < £ < min{mp — x^,x^ — nip}.
The set A^ n [ x i , x j is nonempty and Q[^^{Ai} = Pp{\SJn — m^] > s} is positive for all sufficiently large n. Since A^ does not contain m^, 2i^H^i} "^ 0. According to Theorem 1.4.3, gj,^^{^i} decays exponentially and the decay rate is given in terms of a function I^^^ calculated from Ip^\v) by a variational formula min{/^^^(v): V 6 ^ ( r ) , Y xv; = z]
for
ze[xi,xS\,
00
forz^[xi,xJ.
/^^^ is called the levelX entropy function. It is welldefined, and it measures the discrepancy between z and nip in the sense that Ip^\z) > 0 with equality if and only if z = m^. Thus, the point m^ is the unique minimum point of Ip^\z). In addition Ip^\z) is a continuous convex function of Z G [ X I , X J . These properties are proved in Sections VII.5 and VIII.3. In Section II.4 we give another formula for /^^^ in terms of a LegendreFenchel transform [see (2.14)]. For coin tossing (F = {0,1}, p = ^d^ + \d^), formula (1.17) for /^^^ reduces to (1.9). Indeed the only measure v G ^ ( F ) which satisfies the constraint ^2^^ x,v, = ZG[0,1] is V = (1  z)3o h zd,. Hence by (1.17) I^'\z) = / f >((1  z)So + zd,) = (1  z)log(2(l  z)) + zlog(2z) for ZG [ 0 , 1 ] .
Formula (1.17), which relates the level1 and level2 entropy functions, is called a contraction principle. It will be seen to follow directly from (1.6), which expresses SJn as the mean of the empirical measure L„. Here is the large deviation theorem for levels1 and 2. Theorem 1.4.3. (1.18)
\im\ogQi'^{A,}
= l i m  l o g P , {  5 >  m, > £} =  m i n /^^^(z),
*The point m^ = ^[=1 Xipi is in the open interval (xi,x^) since p^ > 0, Y}=i A = 1
16
I. Introduction to Large Deviations
Mmlog Ql^^{A^} = limlogPp{ max \L„ .p,\>E}=
 m i n Ij,^\v).
(1.19) Since the set AI is closed and does not contain mp.mm^^^ J^p^\z) > I^p^\m^ = 0. Similarly the minimum oflf"* over A2 is positive. Hence both 2!,^^{^i} and Q^n^i^i] converge to zero exponentially fast. It is instructive to interpret Theorem 1.4.3 with reference to the discussion in Section I.l. Think of {X^'JeZ} as giving the successive outcomes of a gambling game in an infinite number of plays of the game separated by a constant time interval. For level1 the macrostates are all real numbers Z 6 [ x i , x j . These correspond to a macroscopic description of the game in terms of the expected value of the outcome of a single play. The microscopic /?sums are {*S„(CO)/A2}. The P^probability that SJn is close to z behaves for large n like exp( —^/^X^)) [see (1.12)]. /^^^ is the entropy function and the mean m^ = YJi=i ^tPt ^^ ^^^ equilibrium state. For level2 the macrostates are all probability measures veJiiT). Each v is a candidate for the distribution of the r outcomes x^, . .., x/m. each play of the game. The microscopic ^sums are {L^ico, •)}• The P^probabiUty that L„ is close to v behaves for large n like exp( —«/^^^(v)) [see (1.21)]. /^^^ is the entropy function and the measure p is the equiUbrium state. Proof of Theorem L4.3. First consider level2. For fixed n and co and ie {1, . . . , r}, let /TJ be the number of times Xj appears in the sequence X^ (co), . . . , X^{oj). Then L„ i(co) = kjn, and L„(co, •) is in A2 if and only ifk = (k^,..., k^ is in the set A2 n = \^ = (ki, .. . ,k^):kie{0,1, I
.. . ,n}, Y ki = n, max —^ — pA > e}. i=i i=i,...,r\n I J
For fixed k6^2,/i? define
There are C(n,k) points co = (co^,oj2,   • ,cOn) ^^ the finite configuration space Q„ = F" for which L„i(a)) = kJn for each /. Let n^Pp be the finite product measure on Q„ with identical onedimensional marginals p. We have Q?^{^2}=
I
7r„P,{a;GQ„:L,,,(co) = /:,//2 for e a c h / } =
^
(1.20) The next lemma is proved like Lemma L3.2 [Problem L8.1(b)]. Lemma L4.4. Uniformly ink = (k^, ... ,kr),
logC(/7,k)=  y ^ l o g ^ + of^^S^^ n {^i n n \ n J
asn^00.
C(n,k)pK
1,5. Level3: Empirical Pair Measure
17
The lemma implies that for each k, Uog(C(n,k)p')) n
= i^flogp, i=in\
 log^) + nj
o f ^ . \ n J
Define the measure v^/^ = ^=1 (ki/n)S^.eJ^(r). The sum in the last display is exactly —/p^^(Vk/„), where Ip^\\ir^ is the relative entropy of v^/„ with respect to p. Since L„i((o) = kjn for each / if and only if L„(co, •) = v,,/„, we see that (1.21)
^iogeK/„} =  / ; ' * ( v + ^ ( ^ ) 
In the sum (1.20) for Qi^^{A2} there are no more than (n + l y terms. As in the proof for coin tossing, we conclude that logei^>{^.} = max {/f>(v,„)} + o ( l ^ ) + o f ^ ^ i i ^ ) . For each n the set { v e ^ ( r ) : v = v^/„ for some ke^^j „} is a subset of A2. Problem 1.8.2 yields (1.19): limlloge 0, X L = I ^tj = U and X?=i ^y = YJ=I '^ki for each /. ^s(F^) may be identified with the compact convex subset of IR**^ consisting of all vectors T = {T^^. ; ZJ* = 1, . . . , r} which satisfy TIJ > 0, X L  = I ^ij = 1, and Xj=i ^0 = Zfc=i ^fci for each /. Let gj,^^ be the Ppdistribution of M„ on ^s(F^) and define the closed set ^3 = { T G ^ , ( F 2 ) :
(1.26)
max
li^^.  p^pj] > e},
iJl,,r where 0 < s
e} is positive for all sufficiently large n. Let 7i2Pp be the product measure in ^ , ( r ^ ) with onedimensional marginals p {7t2Pp{xi,Xj} = PtPj). Since A^ does not contain n2Pp, the ergodic hmit (1.25) impHes that QlifH^a} ^ 0. A level3 large deviation problem is to determine the decay rate of this probabihty. In the next section, we will consider other level3 problems which involve the empirical frequencies of strings {xi^,Xi^, . . . ,Xj^} of arbitrary length A: = 3,4, . . . . The probabilities Q[^\ {A^} decay exponentially and the decay rate is given in terms of a function I^^l which is a natural extension of the level2 function I^^\ The latter determines the decay rate of level2 probabiHties. For xe ^ , ( r ^ ) , set (V,), = Y.U ^ij and define (1.27) ,;.(,) ^ ^ , ^ ^ l , g _  _ ^ where the sum runs over all / and7 for which (vjj > 0. Ip^^ix) is welldefined (OlogO = 0) and equals the relative entropy of T with respect to {(vj^py}. We have the following properties [Problems IX.6.1IX.6.3]. I^^li^) is a convex function of T. I^^li^) measures the discrepancy between TeJ^^ir^) and 712Pp in the sense that Ip^li^) > 0 with equahty if and only if T = 712Pp. Thus the measure 712Pp is the unique minimum point of I^^l on ^ ^ ( r ^ ) . Here is a level3 large deviation theorem for the empirical pair measure. We sketch the proof in the case where F consists of two points (r = 2). Theorem 1.5.1. \imlogQi^}{A^}
= limlogPp{ max
\M^jj  p^pjl > s}
yv.ns)
= min/^f](x). Since the set A^ is closed and does not contain 7i2Pp, min^^^ ^p^ii'^) > ^p^ii^i^p) = ^' Hence 21,^2 {^3} converges to zero exponentially fast. Proof for r = 2. For fixed n and co, define Nij = nM^ijia)). M„(co, •) is in A^ if and only if N = {A/^^; ij' = 1,2} is in the set A,,„ = \N= 2
{N,JJ=
1,2}: N,e{0, I...,
n}, t
^u = ^^
2
Z ^ j = Z ^ki
for each /,
max {N^n  p^Pjl > s
Now fix NeA^^„. Let ki = Yj=i A^/ and define y(n,N) to be the number of points CO in the finite configuration space Q„ = F" for which M„jj(a)) = Nij/n for each / a n d / Then 251^3} =
Z
7r„Pp{coGQ„: M„,f^/a;) = N^j/n
for each / and;}
20
I. Introduction to Large Deviations
where p^ = p\^p^2' The asymptotics of y{n,N) are given by the next lemma. Define D„ = {N\ y(n, N)>0}. then NED^. Uniformly in NeD^,
Lemma 1.5.2. IfN^2>^^ (1.29) ^
Uogy{n,N)=Y^\og^n ^'^ ^ i'tin
Y%og^ + o(^^^ ^j ^ ^ n ^ n \ ^ J
°n
asn^oo.
Define the measure T^/„ = YJj=i (^»^{xf,x,.}e^,(r^). Since M„jj(co) = Nij/n for each / and7 if and only if M„(a), •) = T^/„, the lemma implies that for NED„
^logei?U%/J = llog(y(n,N)p^)
= /^fJC^/J +
o(^^
As in the proof of Theorem 1.4.3, we have Um^loge{^3} =  l i m
min
/,) = ^"f Vm«)(),
where T^ is the identity mapping and T^ = T(T^~^) for /c = 2 , . . . , «  1. For
1.6. Level3: Empirical Process
23
each Borel subset B of Q, R„(a),B) is the relative frequency with which X(n,(jo), TX{n,co), . . . , r""^X(w, co) is in B. Since Xin.oS) is periodic of period n, R(n, •) is for each co a strictly stationary probabiHty measure. It is called the empirical process corresponding to A^i(co), . . . , X„(a>). Let /r be a positive integer less than n. In order to interpret formula (1.33), define r^">(co) to be the ^tuple (Z^(co), A>+i(co), . . . ,Z^+;,_i(a;)) if JSG {1,2,...,« — / : + 1} and to be the cyclic ^tuple (Xp{co),..., Xn{co), X^ (co),..., Xp+j^_i_^(cQ)) if j5e{« —/c + 2,. . . , « } . If Z is the cylinder set [coeQ: ^ 1 = Xi,.(^2 = ^i^' • • • '^fc =  ^ i j . * then
1 " i^„(a;,X) =  X ^y, •) is the onedimensional marginal of i^„(co, •)• If 2^ is the twodimensional cylinder set {coeQ:co^ = x^, 0)2 = •X:J}, then (1.35)
i^„(co,E) = M„(co, {x„x,.}) = M„jj(co).
Thus M„(co, ') is the twodimensional marginal of i^„(co, •). If Z is any cylinder set, then the ergodic theorem implies that limi^„(co,Z) = P,{Z}
P,a.s.
[Theorems A.9.2(c) and Corollary A.9.8]. Let Z^, . . . ,Z^ be cylinder sets such that 0 < Pp{^k} < I, k = I, ..., N, and define the closed set^ (1.36)
^3 = {PeJ^M'^rnaxJP{^,}
 />,{ZJ > e},
Let Qi^^ be the P^distribution of i^„ on ^^(Q). For all sufficiently small e > 0 the set B2 is nonempty, and for all sufficiently large n Ql,'^{B,} = P,{a;eQ:^jnax^i^„(co,Z,)  P,{Z,} > e} > 0. Since B^ does not contain P^, the ergodic theorem impUes that Ql^^{B^} ^ 0. In fact, the probabilities decay exponentially, and the decay rate is given in terms of a function I^^\ which we now define. Let a be a positive integer and n^ the projection of F^ onto F"^ defined by *We use CO since co labels the empirical process. ^The topology on A/s(Q) is the topology of weak convergence [Sections A.8A.9].
24
I. Introduction to Large Deviations
n^co = (coi, . . . ,o)J. If P is a measure in Jl^i^, measure n^P on ^(F'') by requiring n,P{F) = P{n;'F}
=
P{(D:(CO,,
then define a probability ...
,COJGF}
for subsets F of F^. The measure n^P is called the adimensional marginal of P. We consider the quantity
which is the relative entropy of n^P with respect to n^Pp. In Chapter IX, we prove that the limit (1.37)
/(i') = limi/^i' The latter is the Shannon entropy of V, H(v). (b) Let P be a Markov chain with transition matrix {y^^} and invariant measure v [Example 1.6.1 (b)]. Then Ip^\P) = logr — h(P), where h(P) = ZL=i^t7tjlogyoFor any measure p in Jf(r) with each Pi> 0 and for any PeJ^^(Q), one can express Ip^KP) as an expectation involving relative entropy [see (2.20)]. Here is the level3 large deviation theorem. It will be proved in Chapter IX.
1.6. Level3: Empirical Process
25
Theorem 1.6.3. For all sufficiently small £ > 0 \im\ogQ^^\B,] n^cofl
(1.38)
= lim^logP^lcoeQ: max i?„(co,2^)  P , { E J  > ^} n^con
'^
k=l,...,N
^
=min/^3'(P). P G B^
Since the set B^ is closed and does not contain P^, miiipg^ ^p^\P) > 0. Hence Q^n^^z] converges to zero exponentially fast.
I^^KP^
=
The theorem can be interpreted as follows. For level3, the macrostates are all strictly stationary probability measures PeJ^^QL). Each P is a candidate for describing the probabilistic structure of a gambling game on Q [see page 16]. The microscopic «sums are the empirical processes [Rnioj, •)}• The P^probability that R^ioj, •) is close to PeMJS^ behaves for large n like exp( —«/^^^(P)). /^^^ is the entropy function and the measure Pp is the equiUbrium state. We end this chapter by discussing the relationship between Theorem 1.6.3 and the previous large deviation theorems. Theorem 1.5.1 treated the empirical pair measure M„(<JO, ').lf P is SL measure in ^s(Q), then the twodimensional marginal 712^^ is a probability measure on F^, and since P is strictly stationary, 712^ has equal onedimensional marginals. Hence 712^ belongs to ^ , ( F ^ ) . In the set B^ in (1.36), let { I J run through all twodimensional cylinder sets {cdeQ:cd^ = Xi,cd2 = Xj}, i,j = I, . . . , r. Then n^B^ = {xeJiJ s}.
ij=l,,r
The latter is the set A^ in (1.26). Since n2Rn(o), •) = M„(co, •) for each co, it follows that P,{coGQ:M„(a>, ')eA^} = Pp{cDeQ:R„{cD, ')eB,}, By Theorem L5.1 the decay rate of the first probability is — min^g^^/^f^(T). This must equal the decay rate o{Pp{co e Q: Rn(oj, •) ^ ^3}? which by Theorem L6.3 equals — min {IJ^^\P):PGB^}. The latter can be rewritten as minmin{/^^^(P):PG^,(Q),7C2P = T}. T 6 B^
Thus, one expects that for x e ^ s ( r ^ ) (1.39) /(v) = min VSA2
I^^KP)
P^^3
= min mm{I^^\P):PEJ^,(Q\n^P
= v}.
^^^2
In Chapter IX, we prove for veJ^iT) the following contraction principle which is consistent with the last display: (1.40)
I^'Kv) = t nlog^ = min{I^'\P): i=l
PeJiM.n.P
= v}.
Pi
Finally, for level1, recall the contraction principle (1.17) relating /^^^ and /^^^:forzG[xi,xJ /^^>(z) = min/;^)(v): veJi{T\
^ x,v, = z\.
Comparing this with (1.40), we conclude that for zG [x^, x j , (1.41)
/(z) = min/;^>(i>): PeJiM.
f x,{7i,P\ = z  .
This completes our discussion of large deviations for i.i.d. random variables with a finite outcome space. An interesting feature of this chapter was the use of combinatorics to calculate explicit formulas for the entropy functions /^^^, /^^\ and /jf]. In the next chapter, the three levels of large deviations will be generalized to random vectors taking values in R"^. The theory will be applied to statistical mechanics in later parts of the book.
1.7. Notes 1 (page 3). Wehrl (1978) discusses the historical and physical backgrounds of entropy together with modern developments. He Hsts many references. 2 (page 13). Relative entropy I^^Kv), introduced by KuUback and Leibler in 1951, is also known as the KuUbackLeibler information number. It plays an important role in statistics, especially in large sample theories of estimation and testing [Kullback (1959), Bahadur (1967, 1971)]. / f ^(v) measures the statistical distance between v and p. The smaller this distance, the harder
1.7. Notes
27
it is to discriminate between v and p. Applications of large deviations to statistics are discussed by Bahadur (1971), Chernoff (1972), and individual articles in " G r a n d e s deviations et applications statistiques," Asterisque 68, Societe Mathematique de France, Paris, 1979.
3 (page 14). Shannon entropy H(v) was first defined by Shannon (1948) and independently by Wiener (1948). The form of H(v) can be derived from a set of axioms which a reasonable measure of randomness should satisfy; see, e.g., Khinchin (1957, pages 913). 4 (page 24). In information theory, the mean entropy h(P), PeJiX^)^ measures the amount of information per symbol in a message which is generated according to P [Khinchin (1957), MeEliece (1977)]. Large deviation bounds, known as Chernoff bounds [see Problem VII.8.9], are widely used [Wozencraft and Jacobs (1965), Gallager (1968)]. In ergodic theory, h{P) is the KolmogorovSinai invariant of the dynamical system (Q, =^, P, T) [Martin and England (1981)].
1.8. Problems 1.8.1. (a) Given «e{2,3, . . . , } , prove that nXogn — n\ \ < \og{n!) < {n + 1) \og{n + 1) — (« + 1) + 1 by considering the area under the graph oflogx, X > 1. Deduce the weak form of Stirling's approximation:log(«!) = nXogn — n { 0{\ogn), n>2. (b) Prove Lemma 1.4.4. 1.8.2. This problem shows how to complete the proofs of Theorems 1.3.1, 1.4.3. and 1.5.1. Let ^ be a compact subset of W^ a n d / a realvalued function on A which is continuous relative to A. Let {A„;« = 1,2,...} be closed subsets of A such that for any aeA there exists a sequence a„eAn with a„^ a SLS n^co. Prove that lim„_^ mm^^^^f(x) = min^^^f(x). 1.8.3. Let p = {pij; /,7 = 1, . . . , r} be a set of positive numbers such that I}j=iPij = 1 Define v^ = Yj=iPij ^^d fij = Y}=iPij' Let p = { A ; / = 1, . . . , r} be a sequence of positive numbers such that ^  = 1 A = 1. Prove that
i,j=l
i=l
PiPj
Pi
7=1
Pj
with equality if and only ifpij = Vifij for each / a n d / If each Pi = 1/r, then we conclude that the randomness in/7 (as measured by Shannon entropy) is no greater than the sum of the randomness in v and the randomness in p, and that equality holds if and only if p is product measure.
28
I. Introduction to Large Deviations
1.8.4. (a) Prove Jensen's inequality: if vvi, ^2, . . . , w^ and >^i, J^2? • • • > J^r ^re nonnegative numbers such that Yj=i ^j— 1' ^^^^ ( Z ^jyjjlog(Z
y^jyA < Z
^jyj^^^yy
IHint: Let h(x) = xlogx. If a = Yj=i ^jyj > 0, then h(yj) > h(a) + h\a) • (yja).] (b) [Renyi (1970b, page 556)]. Let v = {Vjij = 1, . . . ,r} be a sequence of nonnegative numbers such that Yj=i ^j~ ^ ^^^ y ~ {^/fc} be an r x r doubly stochastic matrix (y^^ > 0,Yj=i Iju = 1 for each k, Y!k=i Jju = 1 for eachy) and define fi^ = Yj=i ^j)!/fc ^^^ Pi — V^ for / = 1, . . . , r. Prove that Z f^k^ogfi, < t ^Aogvj k=l
and
fJ'Kfi) < / f >(v).
j=l
See Voight (1981) for generalizations. 1.8.5 [J. K. Percus (private communication)], (a) Derive formula (1.31). iHint: Expand _Bjj^_R
(1  c)(l  B,,B,Jil
22
 c)) = (1  5 i i + 1  5^2) X (B,2B2xm

c)"',
n>0
where 1  c = 1  (B^^ + B^^  ^ n ^ ^ j ) = (1  B,,){\ (b) For r>2, let fc; = X5=i ^r Show that
B^jll
forallA;j> 1,
where [di^ — Nij/kj']^^ is the aacofactor of the r x r matrix {3^ — Nu/kj}. \_Hint: U B = {Bij} is an r x r positive matrix with 0 < Zl/=i ^u ^ 1' ^^^^ Tr(/  B)^ = Xa=i U  B\JdQt(I  B\ where [/  B\^ is the aacofactor of I — B. Use the formula (which holds in the sense of distribution theory) 1 det(/^)
exp[/«x,j;> 
<x,Byy)^dxdy'—.']
(2ny
L8.6. Let B = {Bij} be a positive 2 x 2 matrix and X(B) the largest eigenvalue of B in absolute value. By Lemma IX.4.1, 2.(B) is positive and logX(B) = lim„^oo n~^ logTr ^". As in the proof of Lemma L5.2, (1.42)
T r 5 " = l.y(n,N)
H Bl'jiJ. ij=l
1.8. Problems
29
where the sum runs over all A^ = {Niy, ij = 1 , 2 } such that A^^G {0, 1, . . . , n), Yj,j=i ^ij = ^' ^^^ Tj=i ^ij ~ Zfc=i ^ki for each /. Using Lemma 1.5.2, prove that (1.43)
logA(^) = m a x  x T , ^ l o g ^ ? ^ : T 6 ^ , , J ,
where (vj^ = Yjj=i '^tj ^^^ ^s, i is the set of T = {T^^; iJ = 1 , 2 } satisfying T^^ > 0, Yj,j=i '^ij = 1»^^^ Z?=i '^y ^ Zfe=i '^fei fo^ ^^^^ ^' ^^^ s ^ ^ i^ (143) runs over all / and j for which (vj, > 0. Theorem IX.4.4 and Problem IX.6.4 are generalizations. 1.8.7. Verify the calculations in Examples 1.6.2(a), (b).
Chapter II
Large Deviation Property and Asymptotics of Integrals
II. 1. Introduction The previous chapter treated three levels of large deviations for independent, identically distributed (i.i.d.) random variables with a finite state space. The main results show the exponential decay of large deviation probabiHties. A level1 example is Pp{\SJn — mp > e}, where 5„ is the «th partial sum of the random variables and nip is their common mean. Levels2 and 3 treat analogous probabiHties for the empirical measures {L„} and the empirical processes {i^„}, respectively. One purpose of the present chapter is to expand the scope of large deviations and to prepare the groundwork for appUcations to statistical mechanics later in the book.^ This chapter extends the ideas of Chapter I by considering random vectors taking values in IR'^, where rf > 1 is a fixed integer. R^ is a natural state space for stochastic models, including models in statistical mechanics. The theory, which is somewhat technical and detailed, is presented in a manner that closely parallels the development in Chapter I, where elementary proofs based upon combinatorics were possible because of the finite state space. The reader who followed that development should find the new theorems familiar looking and plausible. In order to reach illuminating applications of the theory to statistical mechanical models, the proofs of the theorems will be postponed until Chapters VIIX. The last section of this chapter shows how one can apply large deviations to evaluate the asymptotics of certain configuration space integrals that depend on a parameter. The main theorem in this section is due to Varadhan (1966). In the next few chapters we will apply Varadhan's theorem to study integrals associated with the statistical mechanical models (partition functions).
II.2. Levels1, 2, and 3 Large Deviations for I.I.D. Random Vectors The presentation of the three levels of large deviations for vectorvalued random variables involves some topology. We next outUne the information we need, saving a more thorough treatment for Appendix A.
II.2. Levels1, 2, and 3 Large Deviations for I.LD. Random Vectors
31
The basic underlying model in this chapter is a sequence {Xfje Z} of i.i.d. random vectors defined on a probability space (Q, J^, P) and taking values in IR"^, where (i is a fixed positive integer. Since all of our results depend only on the distribution of the random vectors, we may suppose (Q, ^ , P) to be the following standard model. Q is the space of all sequences a> = {cOj 'JeZ} with each cOj e W^. That is, Q is the infinite product space Yljez ^"^ C)r simply (U^)^. We endow Q. with the product topology. An open base for this topology consists of all sets Yijez ^j^ where all but a finite number of the factors equal U^ and the remainder, say Uj^, ..., U^^, are open sets in U^. Convergence in this topology is equivalent to coordinatewise convergence in U^. ^ is the Borel crfield of Q. The Borel (Tfield, denoted by J'(Q), is defined as the afield generated by the open subsets of Q. J*(Q) coincides with the afield generated by the product cyUnder sets (2.1)
2:^,fc= {coGQ:co^+iGFi,...,a;^+fcG/;},
where m and k are integers with k > 1 and F^, .. . ,F^ are Borel subsets of [R^. P is an infinite product measure on J*(Q) with identical onedimensional marginals, which we denote by p. Thus, for example,
P{^m,.]=\\p{Fj]. 7= 1
We write Pp for P. Xjj'eZ, is theyth coordinate function on Q defined by Xj((jo) = (Oj. The set {Xj 'JeZ} is a sequence of i.i.d. random vectors taking values in R^. This sequence is called the coordinate representation process. Each A} has distribution p: Pp{ojea\X^{(D)eB]
= p{B}
for Borel subsets B of U^,
Q can serve as a configuration space for physical systems. In statistical mechanical models, Q may be the configuration space for the velocities of a system of particles [Chapter III] or for the spins at sites of a ferromagnet [Chapters IV and V]. A microscopic model for such a system is completed by choice of a probability measure P on (Q, J^). The measure P, called an ensemble, expresses the microscopic interactions among the particles or spins. Product measure corresponds to the absence of interactions. It models a system of independent particles (ideal gas) or independent spins. When interactions are present, an ensemble which is not product measure must be chosen. We now define the three levels for a sequence of i.i.d. random vectors {1}'JeZ) on a probabiUty space {Q, J^, P). Let p be the distribution of X^. Level1 large deviations treat the microscopic «sums Sn(oji)/n = Yj=i Xj(coi)/n, « = 1, 2, . . . . The law of large numbers states that if J^^ \\x\\ p{dx) is finite,* then * IIjc = (Zf^i xfy^^, the Euclidean norm of x.
32
(2.2)
II. Large Deviation Property and Asymptotics of Integrals
S„((D)/n > m =
xp(dx)
Pa.s.
Denote by gj,^^ the distribution of SJn on U^. The hmit (2.2) impUes that if A is any Borel set whose closure does not contain m^, then Ql^^{A} ^0 as n^ CO. Hence the distributions {2J,^^;« = 1,2, . . . ) converge weakly to the unit point measure 5^ on IR'^ [Theorem A.8.2]. Level1 large deviations study the exponential decay of the probabiHties Ql^^{A}. Level2 large deviations treat the empirical measures L„(co, •) = ^~^ Zj=i ^xi(o)(')^ « = 1, 2, . . . , which are level2 microscopic «sums. For each Borel subset B of W^, L„(CD,B) is the relative frequency with which X^(CD), . . . , X„((D) is in B. Therefore, for each co, L„(co, •) is an element of ^(U^), the set of probabiHty measures on ^(U^). The mean of the empirical measure is l^dxL„((o,dx) = S„(co)/n. The set ^(U^) with the topology of weak convergence is a complete separable metric space [Example A.8.4(a)]. It is easily checked that if ^ is a Borel subset otJ^iW^), then the set L~^A = {(JOGQ:L„((D, ')eA} belongs to the crfield J^. The distribution of L„ is the measure Q\^^ defined by Qi^^{A} = PpiL'^A} for A a Borel subset ofJi{U^). For level2, the Hmit (2.2) is replaced by (2.3)
L,{(Dr)^P
Pa.s.
This Umit states that the sequence {L„(co, ')\n= \,1, ...] converges weakly to p Pa.s. It is a consequence of the ergodic theorem [Corollary A.9.8]. The limit (2.3) imphes that if ^ is any Borel subset of ^((R^) whose closure does not contain p, then Q^^\A] ^ 0 as n^ co. Hence the distributions {gj,^^;w = 1,2, . . . } converge weakly to the unit point measure S^ on J^(U'^). Level2 large deviations study the exponential decay of the probabilities e[,^H^}Level3 large deviations treat the empirical processes
(2.4)
i?„(«,) = "E Vm»)(),
n=l,2, ...,
n /c = 0 which are level3 microscopic «sums. T is the shift mapping on (W^)^ and X(n,co) is the periodic point in (IR'^)^ obtained by repeating (X^(co), X2((jo), . . . , X„((o)) periodically [cf.. Section 1.6]. For each Borel subset B of (IR^)^, R„(co, B) is the relative frequency with which X{n, co), TX{n, co), . . . , r^X{n, CO) is in B. Therefore, for each co, i?„(co, •) is an element of ^ . ( ( ( R ^ ) , the set of strictly stationary probability measures on 0^{{U^Y), All the onedimensional marginals of Rn{(o, •) equal L„(a;, •)• The set Ji^i^^Y) with the topology of weak convergence is a complete separable metric space [Theorem A.9.2(a)]. It is easily checked that if ^ is a Borel subset of Ji,{UY, then the set R^^A = {(oeQ:R^(co, ')EA} belongs to the afield J^. The distribution of i^„ is the measure Qi^^ defined by Qi^^{A} = Pp{R~^A} for A a Borel subset of J^,((UY), For level3, the hmits (2.2) and (2.3) are replaced by (2.5)
R,(co,')^P^
Pa.s.
II.3. The Definition of Large Deviation Property
33
Table II. 1. Definitions of Levels1, 2, and 3 Random A.s. quantity limit
Metric space 3C
Distribution
Weak limit
Level
e„ on ^(X)
of{e„}
1
SJn
rUp
W^
Q^:^{B} = P ^ e ^
Qi"=>S„,
2
Ln
P
Ji{W)
Qi'>{B}=^P{L„eB}
Qi'^^5,
3
Rn
Pp
j^sm'f)
Q'„'^{B} = P{R„eB}
ei^'^P,
This limit states that the sequence {R„(o), ');n = 1,2, ...} converges w^eakly to P^ Pa.s. It is a consequence of the ergodic theorem [Corollary A.9.8]. The limit (2.5) implies that if A is any Borel subset ofJ^,((^Y) whose closure does not contain F^, then Q\j^^{A} ^ 0 as w ^ oo. Hence the distributions {Q[^^;« = 1,2, . . . } converge v^eakly to the unit point measure 3p on J^^aW^)^). Level3 large deviations study the exponential decay of the probabiHties Q[^\A}. Thus for each level there is a random quantity SJn, L„, or R^. Each quantity takes values in a complete separable metric space ^ = W^, Ji{U^), or Ji^{{R'^Y), and has an a.s. limit which is a point XQ in SC {XQ = m^, p, or Pp). The distributions of these random quantities are measures gj,^\ Qj,^\ or Q\^^^ on the Borel subsets of the respective space ^ , and they converge weakly to the corresponding unit point measure S^^ on ^. Large deviations study the asymptotic behaviors of the distributions Ql,^\ 2i^\ and Q[^\ This information is summarized in Table II. 1. In the next section, we give an abstract formulation of a class of large deviation problems, which includes the three levels of large deviations for i.i.d. random vectors. We return to the three levels in Section II.4.
II.3. The Definition of Large Deviation Property A central theme of Chapter I was to express the exponential decay of large deviation probabilities in terms of entropy functions. For each of the three levels of large deviations we had Hmits of the form (2.6)
lim logQ^^{A,} = min/^^>(x), n^CO fl
/ = 1, 2, 3.
^^^i
The functions /^^^, fp^\ and /^^^ are the entropy functions corresponding to the three levels and the sets A^ are closed sets not containing the unique minimum point of the respective function I^p\ In each case, the minimum of /^'^ over A^ is positive. Hence the limit (2.6) impHes that the large deviation probabihties 2i'H^i} decay exponentially. The righthand side is the decay rate. For example, in Chapter I, A^ is the set {zeU\ \z — m^\ > s} and I^^^ is defined by the contraction principle (1.17). The point m^ is the unique min
34
II. Large Deviation Property and Asymptotics of Integrals
imum point of/^^ The limit (2.6) for / = 1 can be used to show the exponential decay of probabilities of other sets. Let K be any closed set in U which does not contain the point m^. Since A^ is a subset oi A^ for sufficiently small £, the limit (2.6) implies that there exists a number N = N(K) > 0 such that (2.7)
Q['^{A:} < ^"^
for all sufficiently large n.
In applications, these bounds for closed sets are often as important as limits of the form (2.6). Our first task is a formulation of such bounds for distributions on W^, The bounds will be expressed directly in terms of entropy functions on W^ which for now we assume, and later will prove, to exist. This formulation will suggest how to prove hmits like (2.6) and will motivate the general concept of large deviation property. Large deviation property will be defined later in the section. Let {2„;« = 1,2, . . . } be a sequence of Borel probability measures on IR'^. Suppose that there exists a nonnegative realvalued function I(x), XG [R^ such that inf^^^dl(x) = 0 and (2.8)
limsuploge„{A:} <  inf/(x)
for each nonempty closed set K in U^. Assume also that I(x) attains its infimum of 0 at some point in IR"^. Such a point is called a minimum point of /; a minimum point need not be unique. What additional properties of I{x) guarantee that a bound of the form (2.7) holds whenever A^ does not contain a minimum point of/? It suffices if the following hold. (a) I{x) is lower semicontinuous onW^\ x^^x implies Uminf^^^o ^(^n) ^ I{x). (b) I{x) has compact level sets: the set {xeR'^: I{x) < b] is compact for each real number b. Indeed, i f / ( ^ ) denotes the infimum of/(x) over K, then (2.8) implies that (2.9)
Qn{K]< exp(nI(K)/2)
for all sufficiently large n.
If we prove that I(K) is positive, then we will have a bound of the form (2.7). The set K^ = Kn {xeU':I{x) < I(K) + 1} is nonempty and since I(x) has compact level sets, K^ is compact. According to a standard result, a lower semicontinuous function attains its infimum over a nonempty compact set [Problem II.9.2]. Since I(K^) = I(K), there exists a point ZEK^ ^ K such that I(z) = I(K). 1{I(K) = 0, then z must equal a minimum point of /. This contradicts the assumption that K contains no minimum point of/. Thus I(K) is positive, and we are done. In many cases, functions I(x) arise which take the value oo. Thus it may happen that in the argument just given, the infimum I(K) equals oo. But then (2.8) implies that for any number A^ > 0, Qn{K} < ^""^ for all sufficiently large n. Along with the upper bound (2.8), it is reasonable to consider the complementary lower bound for each nonempty open set G in (R"^:
II.3. The Definition of Large Deviation Property
(2.10)
35
liminfloga{G} > inf/(x). "*oo
n
xeG
Let us deduce a simple criterion on a nonempty Borel subset A of W^ which implies that if both bounds (2.8) and (2.10) are vahd, then (2.11)
\im\ogQ,{A}=
inf/(x).
n*ao n
xeA
Denote the closure of ^ by c l ^ and the interior of ^ by i n t ^ . Since cXA^ ^ ^ i n t ^ , it follows that — inf I{x) > limsuplogQ„{cl^} > limsuplogQ„{^} xec\A
(2.12)
n^ao
It
n^cc
fl
^ ^ > liminflog2„M} > l i m i n f  l o g g j i n t ^ } > «>oo
yi
n*(X)
yi
inf I{x). xeint^
If the two extreme terms are equal, then the limit (2.11) follows. We return to the level1 problem studied in Chapter I. Let g„ = Q\^^ be the distribution of n~^ YJ)=I ^•' where X^, X2, . . . are i.i.d. random vectors taking values in a finite set Y = {xi,X2, . . . , x j with x^ < X2 < • • • < x^. According to Theorems IL4.1 and IL5.1 below, the bounds (2.8) and (2.10) are vahd for Qi^^; I equals the entropy function /^^^ defined in (1.17). Let A^ be the closed set {z G [R: \z — mp\ > e},whereO < s < min{mp — x^^x^ — nip}. Since the infimum of/^^^(z) over ^^ equals the infimum of/^^^(z) over int^^, the bounds (2.8) and (2.10) yield limlogei^H^i}=inf/;^^(z). In addition, since /^^^ is a continuous function on the compact set [ x i , x j [Theorem 11.4.1(c)], /^^^ is lower semicontinuous on U and has compact level sets. This discussion leads us to the definition of large deviation property. We want the definition to be general enough to cover the three levels of large deviations summarized in Table II. 1. Let ^ be a complete separable metric space and {Q„;« = 1,2,... } a sequence of probabiUty measures on the Borel subsets of ^ . Suppose that as « > 00, {Q„} converges weakly to the unit point measure 3^^ at some point XQ in ^ . Thus for most Borel sets A, Qn{A] ^ 0 as w > 00. In the examples we consider, Qn{A] converges to zero exponentially fast as « > 00 with an exponential rate depending on the set A. We allow for the possibiHty of other scaUng constants besides n. The situation is abstracted in the following definition.^ Definition IL3.1. Let ^ be a complete separable metric space, ^ ( ^ ) the Borel ofield of ^ , and {Q„; /i = 1,2, . . . } a sequence of probability measures on ^(^). {2„} is said to have a large deviation property if there exist a sequence of positive numbers {(2„; n= 1,2,...} which tend to 00 and a function I{x) which maps ^ into [0, 00] such that the following hypotheses hold.
36
II. Large Deviation Property and Asymptotics of Integrals
(a) I(x) is lower semicontinuous on ^. (b) I(x) has compact level sets. (c) limsupn^^ a~^ log Q„{K} < —inf^g^/(x) for each closed set A^in ^ . * (d) liminf„^oo a~^ logQn{G} > —mf^^Ql(x) for each open set G in ^ . I(x) is called an entropy function of {Q„}. We note several consequences of the definition. The infimum ofI(x) over 0C equals 0. This follows from the upper and lower large deviation bounds (c)(d) with K= G = 3C. I(x) attains its infimum over any nonempty closed set (the infimum may be oo). This follows from hypotheses (a) and (b) by the same argument given earlier in the case ^ = IR'^ [page 34]. According to the following theorem, if a large deviation property holds, then the entropy function is unique.^ Theorem II.3.2. Let ^ be a complete separable metric space and {Q^;n = 1,2,...} a sequence of probability measures on ^{^). If for a fixed sequence [a^.n = 1,2,...} {Qf^} has a large deviation property with entropy function I and with entropy function / , then I(x) = J(x)for all xe^. Proof Fix a point xe^ and define Kj and Gj to be the closed ball and the open ball, respectively, with radius l/j and center x (7 = 1,2, . . . ) . Since
liKj^,)
< KGj) < liminfloge„{G,} < limsuploge„{^,} < "00
J(Kjl
a„
where if ^ is a subset of ^ , then 1(A) and J(A) denote the infimum of / and of/over A, respectively. Repeating the argument with / a n d /interchanged, we see that J(Kj) < I(Kj^,) < J(Kj^,) < I(Kj^,) < . . . . Hence to prove that I(x) = J(x), it suffices to show that I(Kj) ^ I(x) and J(Kj) ^ J(x) as 7 ^ 00. We prove the first. The second is proved the same way. Since {I(K^ ;7 = 1,2, . . . } is an increasing sequence and x is a point in each set K^, X\v[ij^^I(K^ exists as an extended realvalued number and lim^^oo A^j) ^ ^W /attains its infimum over the closed set Kj. Hence there exists a point x^ in K^ such that I(x^ = I(Kj). The sequence {xj;j = 1,2, ...} converges to x and by lower semicontinuity *If ATor G equals the empty set (/>, then (c)(d) hold trivially if we set log QM) = log 0 = — 00 and inf^g^/(x) = 00.
II.4. Statement of Large Deviation Properties for Levels1, 2, and 3
37
I(x) < lim I(x,) = lim I{K:) < I(x). Thus I(Kj) ^ I(x), and the proof is complete.
n
We remarked above that if I(x) is an entropy function, then the infimum of I(x) over the whole space ^ equals 0. Since ^ is closed, I(x) attains its infimum of 0 at some point of ^ . Such a point is called a minimum point of/. Recall the case ^ = IR"^ which was considered at the beginning of this section. We proved using hypotheses (a)(c) of large deviation property that Qn{K} decays exponentially for any closed set ^ not containing a minimum point of /. The same proof works for any complete separable metric space ^ , and we have the following result. Theorem II.3.3. Let ^ be a complete separable metric space, {Q^;n = 1,2,... } a sequence ofprobability measures on ^{SC), andl(x) a function which maps ^ into [0, oo] such that inf_^g^/(x) = 0. Suppose that hypotheses (a)(c) of large deviation property are valid for some positive sequence a„^ oo. Then for any closed set Knot containing a minimum point of I, there exists a number N = N(K) > 0 such that Q„{K} < e'""^^ for all sufficiently large n. We end this section with a condition on a Borel set A which guarantees that lim„_^oo ^/7^ log Qn{A} = — inf^^^ /(x). The theorem is proved as in (2.12). Theorem 11.3.4. Let ^ be a complete separable metric space and {Q^;n = 1,2, . . . } a sequence of probability measures on ^{^), Suppose that {2„} has a large deviation property with constants {a„;n = 1,2, ...} and entropy function L Call a Borel subset A of 3C an Icontinuity set if inf I{x) = inf I(x). xe c\ A
xe int A
If A is an Icontinuity set, then lim^^o^ a~^ log Qn(^) =
II.4.
—inf^^^I(x).
Statement of Large Deviation Properties for Levels1, 2, and 3
The microscopic «sums and distributions determining the three levels were summarized in Table II. 1. This section states the large deviation property for each level, then summarizes the results in Table II.2. Level1. We denote by < •, • > the Euclidean inner product on W^ and by  •  the EucHdean norm. Theorem II.4.1.* Let X^, X2, . .. be a sequence of i.i.d. random vectors taking values in U^ and let peJ^(U^) be the distribution ofX^. For teU^, define * Parts (a) and (b) of Theorem IL4.1 are proved in Section VIL5; part (c) is proved in Theorem VIIL3.3(c).
38
II. Large Deviation Property and Asymptotics of Integrals
(2.13)
c^(t) = log^{exp}  log
expp(Jx).
Assume that Cp{t) is finite for all t and define (2.14)
/;i>(z) = s u p { < ^ , z >  c / 0 }
for
zeW.
Let S„ denote the nth partial sum ^"=1 A}, « = 1, 2, . . . . Then the following conclusions hold. (a) {Q^n^], the distributions of{SJn} on U^, have a large deviation property with an = n and entropy function I^^K (b) ll^\z) is a convex function of z. Ip^\z) measures the discrepancy between z and the mean"^ m^ = ^^dxp(dx) in the sense that Ip^\z) > 0 with equality if and only ifz = mp. (c) If p is supported on a finite set Y = {xi,X2, • . . ?^r} — ^ with x^ < X2 < '' • < x^, then Ip^^ is finite and continuous on the interval \_Xi,x^ and Il^\z)= o o / o r z ^ [ x i , x J . The hypothesis that Cp(t) be finite for all / is satisfied if, for example, X^ is bounded. In Chapter I, we defined the level1 entropy function Ip^^hy the contraction principle (1.17). A new aspect of Theorem II.4.1 is formula (2.14) for I^^\ Although the proof of the theorem is postponed, we will motivate this formula by assuming that {Q[^^} has a large deviation property with some entropy function /^^^ and then deducing (2.14) as a consequence. For the contraction principle, see Theorem II.5.1. If {Qn^^} has a large deviation property with entropy function /^^\ then it is tempting to think of Q[^\dz) as having the asymptotic form exp( — nfp^\z))dz. Since X^, X2, . . . are i.i.d., log£{exp} = log fl E{cxp{t,Xjy} n
n
j=i
= logE{cxp{t,X,}}
= Cp(t).
Using the heuristic formula Ql^^\dz) ^ Qxp( — nll,^\z))dz, we may write Cp(t) = log^{exp(«(z)}.
The function Cp{t) is a convex function on U^ [Example VII. 1.2]. Formula *J(Rdxp((ix) < 00 since Cp{t) < 00 for all /.
II.4. Statement of Large Deviation Properties for Levels1, 2, and 3
39
(2.15) expresses this convex function as the LegendreFenchel transform of Ip^\z). When the theory of these transforms is developed, we will see that formula (2.15) can be inverted to yield formula (2.14) for I^^\ In Section II.7, we will justify the manipulations leading to (2.15). The function Cp(t) is wellknown in probabihty theory as the logarithm of the moment generating function of p (also called the cumulant generating function of p). By analogy with statistical mechanics, we call Cp(t) the free energy function of p. The reason for this choice of terminology will be clear in Section IV.5. Here are two examples of entropy functions calculated by (2.14). Example II.4.2. (a) For coin tossing, p = i^o + 2^1 ^^^ ^p(0 = log[2(1 + e^y]. Fix z real. Cp{f) is convex and tz — Cp{f) is concave. Hence tz — Cp{t) attains its supremum at t if and only if 0 = {tz
 Cp{f)) = z  dp{f).
The derivative Cp{i) = e^/(l + e^) has range (0,1) as t runs through (— 00, 00). For ZG(0, 1) Cp(t) equals z at the unique value t = t(z) = log(z/(l — z)). By (2.14) fp'\z) = t(z) • z  Cp(t(z)) = zlog(2z) + (1  z)log(2(l  z)). For z = 0 or 1, the supremum in (2.14) is not attained but /^^(0) = /^^^(l) = log2; for z ^ [ 0 , 1 ] , /^'^(z) = 00 [Theorem VIII.3.3(b)]. Hence we obtain the formula for /^^^ in (1.9). (b) Given m real and cr^ > 0, consider the Gaussian probabihty measure on U p(dz) = N(m, G^){dz) = ,^^ 2>.i/2^^P
{iKG^y
(z — mY iG^
dz\
m — \^ xp(dx) and G^ = JK (^ — mYp{dx). A standard calculation shows that c^{t) = ^G^t^ + mt. For any z real the supremum in (2.14) is attained at r  (z  m)/G^ and fp'\z) = (z  mY/lG^ Theorem II.4.1 has an interesting implication concerning the convergence of {SJn}. For any £ > 0, let K be the closed set {ze [R'^: z  m^ > s}. The distributions {Q[^^} of {SJn} have a large deviation property and the entropy function /^^(z) attains its infimum of 0 at the unique point z = nip. By Theorem II.3.3, there exists a number TV = N(8) > 0 such that (2.16)
Qi'^K} = P{\\SJn  m j > s} < e""^ for all sufficiently large n.
Since Y,^=i Qxp( — nN) is finite, the BorelCantelh lemma imphes the strong law of large numbers: SJn ^ nip Pa.s. [Theorem A.5.3]. We may summarize (2.16) by saying that SJn converges exponentially to m^. Exponential convergence will be considered in Section II.6 for other random vectors.
40
II. Large Deviation Property and Asymptotics of Integrals
Level2. Let p be a Borel probability measure on R whose support is a finite set r = {xi,X2, . . . , x j , with x^ < X2 < ''' < x^. Thus if p^ = p{Xj}, then each pi is positive and Z[^i A = 1. ^ ( F ) denotes the closed subset of ^(U) consisting of probabihty measures v for which v{r} = 1. Let A2 be the set { v G ^ ( r ) : max v,— Pj > e}
where 0 < e
£. If P{^k]  Pp{^k] ^  e , then the proof is similar. We define the closed convex set
42
II. Large Deviation Property and Asymptotics of Integrals
Table II.2. Large Deviation Properties for Levels1, 2, and 3
Level
Random A.s. quantity limit
i
SJn
R.
Where large deviation property is proved Section VIL5
nip
Properties of the entropy function I^^^ convex
P
II^^ convex
Section VIII.2 (finite state space)
Pp
/;^> affine
Chapter IX (finite state space)
The interior of C is the set of PeJ^X^^) satisfying P { Z J  ^^{1^} > s. If PQ is any measure in int C, then Ip^\Po) is finite [Theorem IX.2.3], and since fp^^ is affine
inf Ij,'KP) = lf\P) = lim f/(v): v G J^(U'l
 xv(dx) = z\ .
For z = m , (2.21) is certainly valid since Ip^\m)
equals 0 and the in
*Theorem IL5.1 is proved in DonskerVaradhan (1976a, page 425). It will be proved in Chapter VIII [Theorems VIII.3.1, VIII.4.1] under additional hypotheses on p.
II.5. Contraction Principles
43
fimum on the righthand side is attained at v = p (Il^\p) = 0). For other z the righthand side may be oo unless there exists a probabihty measure v on U^ which is absolutely continuous with respect to p and satisfies ^^d xv(dx) = z. In order to determine those points z for which the terms in (2.21) are finite, consider again the definition (2.14) of/^^ The free energy function c^it) is convex and differentiable [Theorem VII.5.1]. Hence the supremum in (2.14) is attained ^t teW^ if and only if V,«t,zyc^{t))
=0
orVc^(t) = z.
Suppose that this equation has a solution t = t(z). The gradient of c^ can be calculated by differentiating under the integral sign; that is, —
Qxp0.xyp(dx)=
\
XiQxp(t,x}p(dx)
.rod
^^i Jud
and so VcJt) = z =
XQxp{t,xyp{dx)
^^dQxp{t,x}p(dx)'
Define p^ to be the Borel probabihty measure on IR"^ which is absolutely continuous with respect to p and which has RadonNikodym derivative^ (2.22)
^(x)=:exp^ dp ' ^^dQxp} is finite for all tsU^. In this case, with an = n
II.6. Large Deviation Property for Random Vectors and Exponential Convergence
cM =  log^{exp 0, then W„/a„^ZQ implies W„/a„^ZQ. The theorems in this section were motivated by large deviation results in statistical mechanics proved by Lanford (1973). In our formulation of these results for random vectors {W„}, the free energy function c^(t) plays a central role. If this function is differentiable for all /, then the distributions of {WJa^} have a large deviation property [Theorem II.6.1]. The differentiabihty of c^(t) at ^ = 0 is equivalent to the exponential convergence of {WJa„} to a constant [Theorem II.6.3]. These theorems will have many applications in subsequent chapters. Theorem II.6.1 will be used to derive the level1 theorem for i.i.d. random vectors [Theorem II.4.1] and the level2 and 3 theorems for i.i.d. random variables with a finite state space
50
II. Large Deviation Property and Asymptotics of Integrals
[Theorems II.4.3 and II.4.4]. Theorem II.6.3 will be applied to the study of phase transitions in statistical mechanics.
II.7. Varadhan's Theorem on the Asymptotics of Integrals The basic problems and techniques in the subject of large deviations have been inspired by a number of different areas. These include probability, analysis, statistics, information theory, and statistical mechanics. Its roots in analysis go back to Laplace, who devised a method for studying the asymptotic behavior of real integrals of the form ^^Qxp(nF(x))dx as « ^ oo.^ For example, consider the integral n\ = j'^x"e~^(ix, where nis a positive integer. Change variables to z = x/n, obtaining poo
n\ = rf^^
exp[ —^(z — \ogz)']dz.
Laplace says that the leading order asymptotic behavior of the integral is determined by the largest value of the integrand, in this case ^~" since z — logz has a minimum value of 1 at z = 1. In fact, it is not hard to prove that n\ ^ ^""*'^^"", in the sense that «~Mog(«!/w"^^) ^ —1 as « ^ oo. By making a quadratic approximation about z = 1 and suitably expanding, one derives Stirhng's formula n! = n"e~''^2nn(l + e„), where e„ is an asymptotic series in powers of l/n which converges to 0 as « ^ oo. Similar problems arise in probability theory. Let p be a Borel probabihty measure on IR"^ and X^,X2, ... i.i.d. random vectors with distribution p. Set 5„ = YJ=^ ^j' What is the leading order asymptotic behavior of the expectation (2.33)
Zi^) = £{exp[/2F(5„M/«)]},
where F is a bounded continuous function from [R to R? The special case where Z^, Z2, . . . are Gaussian A^(0,1) random variables is easy. For then SJn is Gaussian N(0,n~^), and so 1/2
2n
exp nfF(z)^zA']dz.
Since n^'^" ^ 1, Laplace tells us that (2.34)
limlogZ^^> = sup {F(z) 
\z^),
Thus Z^^^ grows or decays exponentially depending on the sign of the supremum. One may determine the leading order asymptotic behavior of Z^^^ for arbitrary i.i.d. random vectors A^^, Z2, . . . by applying the level1 large deviation property. Indeed, if g^^^ is the distribution of SJn, then
II.7. Varadhan's Theorem on the Asymptotics of Integrals
(2.35)
Z