36-754, Advanced Probability II

or
Almost None of the Theory of Stochastic
Processes

Cosma Shalizi

Spring 2007
Contents

Table of Contents i

Comprehensive List of Definitions, Lemmas, Propositions, Theo-

rems, Corollaries, Examples and Exercises xxiii

Preface 1

I Stochastic Processes in General 2

1 Basics 3
1.1 So, What Is a Stochastic Process? . . . . . . . . . . . . . . . . . 3
1.2 Random Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Building Processes 15
2.1 Finite-Dimensional Distributions . . . . . . . . . . . . . . . . . . 15
2.2 Consistency and Extension . . . . . . . . . . . . . . . . . . . . . 16

3 Building Processes by Conditioning 21

3.1 Probability Kernels . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Extension via Recursive Conditioning . . . . . . . . . . . . . . . 22
3.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

II One-Parameter Processes in General 26

4 One-Parameter Processes 27
4.1 One-Parameter Processes . . . . . . . . . . . . . . . . . . . . . . 27
4.2 Operator Representations of One-Parameter Processes . . . . . . 31
4.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

5 Stationary Processes 34
5.1 Kinds of Stationarity . . . . . . . . . . . . . . . . . . . . . . . . . 34

i
CONTENTS ii

5.2 Strictly Stationary Processes and Measure-Preserving Transfor-

mations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

6 Random Times 38
6.1 Reminders about Filtrations and Stopping Times . . . . . . . . . 38
6.2 Waiting Times . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
6.3 Kac’s Recurrence Theorem . . . . . . . . . . . . . . . . . . . . . 41
6.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

7 Continuity 46
7.1 Kinds of Continuity for Processes . . . . . . . . . . . . . . . . . . 46
7.2 Why Continuity Is an Issue . . . . . . . . . . . . . . . . . . . . . 48
7.3 Separable Random Functions . . . . . . . . . . . . . . . . . . . . 50
7.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

8 More on Continuity 52
8.1 Separable Versions . . . . . . . . . . . . . . . . . . . . . . . . . . 52
8.2 Measurable Versions . . . . . . . . . . . . . . . . . . . . . . . . . 57
8.3 Cadlag Versions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
8.4 Continuous Modifications . . . . . . . . . . . . . . . . . . . . . . 59

III Markov Processes 60

9 Markov Processes 61
9.1 The Correct Line on the Markov Property . . . . . . . . . . . . . 61
9.2 Transition Probability Kernels . . . . . . . . . . . . . . . . . . . 62
9.3 The Markov Property Under Multiple Filtrations . . . . . . . . . 65
9.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

10 Markov Characterizations 70
10.1 Markov Sequences as Transduced Noise . . . . . . . . . . . . . . 70
10.2 Time-Evolution (Markov) Operators . . . . . . . . . . . . . . . . 72
10.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

11 Markov Examples 78
11.1 Probability Densities in the Logistic Map . . . . . . . . . . . . . 78
11.2 Transition Kernels and Evolution Operators for the Wiener Process 80
11.3 Lévy Processes and Limit Laws . . . . . . . . . . . . . . . . . . . 82
11.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

12 Generators 88
12.1 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
CONTENTS iii

13 Strong Markov, Martingales 95

13.1 The Strong Markov Property . . . . . . . . . . . . . . . . . . . . 95
13.2 Martingale Problems . . . . . . . . . . . . . . . . . . . . . . . . . 96
13.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

14 Feller Processes 99
14.1 Markov Families . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
14.2 Feller Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
14.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

15 Convergence of Feller Processes 107

15.1 Weak Convergence of Processes with Cadlag Paths (The Sko-
rokhod Topology) . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
15.2 Convergence of Feller Processes . . . . . . . . . . . . . . . . . . . 109
15.3 Approximation of Ordinary Differential Equations by Markov
Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
15.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

16 Convergence of Random Walks 114

16.1 The Wiener Process is Feller . . . . . . . . . . . . . . . . . . . . 114
16.2 Convergence of Random Walks . . . . . . . . . . . . . . . . . . . 116
16.2.1 Approach Through Feller Processes . . . . . . . . . . . . . 117
16.2.2 Direct Approach . . . . . . . . . . . . . . . . . . . . . . . 119
16.2.3 Consequences of the Functional Central Limit Theorem . 120
16.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

IV Diffusions and Stochastic Calculus 123

17 Diffusions and the Wiener Process 124
17.1 Diffusions and Stochastic Calculus . . . . . . . . . . . . . . . . . 124
17.2 Once More with the Wiener Process and Its Properties . . . . . . 126
17.3 Wiener Measure; Most Continuous Curves Are Not Differentiable 127

18 Stochastic Integrals Preview 130

18.1 Martingale Characterization of the Wiener Process . . . . . . . . 130
18.2 A Heuristic Introduction to Stochastic Integrals . . . . . . . . . . 131

19 Stochastic Integrals and SDEs 133

19.1 Integrals with Respect to the Wiener Process . . . . . . . . . . . 133
19.2 Some EasyR Stochastic Integrals, with a Moral . . . . . . . . . . . 138
19.2.1 R dW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
19.2.2 W dW . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
19.3 Itô’s Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
19.3.1 Stratonovich Integrals . . . . . . . . . . . . . . . . . . . . 146
19.3.2 Martingale Representation . . . . . . . . . . . . . . . . . . 146
19.4 Stochastic Differential Equations . . . . . . . . . . . . . . . . . . 147
CONTENTS iv

19.5 Brownian Motion, the Langevin Equation, and Ornstein-Uhlenbeck

Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
19.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

20 More on SDEs 158

20.1 Solutions of SDEs are Diffusions . . . . . . . . . . . . . . . . . . 158
20.2 Forward and Backward Equations . . . . . . . . . . . . . . . . . 160
20.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

21 Small-Noise SDEs 164

21.1 Convergence in Probability of SDEs to ODEs . . . . . . . . . . . 165
21.2 Rate of Convergence; Probability of Large Deviations . . . . . . 166

22 Spectral Analysis and L2 Ergodicity 170

22.1 White Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
22.2 Spectral Representation of Weakly Stationary Procesess . . . . . 174
22.2.1 How the White Noise Lost Its Color . . . . . . . . . . . . 180
22.3 The Mean-Square Ergodic Theorem . . . . . . . . . . . . . . . . 181
22.3.1 Mean-Square Ergodicity Based on the Autocovariance . . 181
22.3.2 Mean-Square Ergodicity Based on the Spectrum . . . . . 183
22.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184

V Ergodic Theory 186

23 Ergodic Properties and Ergodic Limits 187
23.1 General Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
23.2 Dynamical Systems and Their Invariants . . . . . . . . . . . . . . 188
23.3 Time Averages and Ergodic Properties . . . . . . . . . . . . . . . 191
23.4 Asymptotic Mean Stationarity . . . . . . . . . . . . . . . . . . . 194
23.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197

24 The Almost-Sure Ergodic Theorem 198

25 Ergodicity and Metric Transitivity 204

25.1 Metric Transitivity . . . . . . . . . . . . . . . . . . . . . . . . . . 204
25.2 Examples of Ergodicity . . . . . . . . . . . . . . . . . . . . . . . 206
25.3 Consequences of Ergodicity . . . . . . . . . . . . . . . . . . . . . 207
25.3.1 Deterministic Limits for Time Averages . . . . . . . . . . 208
25.3.2 Ergodicity and the approach to independence . . . . . . . 208
25.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209

26 Ergodic Decomposition 210

26.1 Preliminaries to Ergodic Decompositions . . . . . . . . . . . . . . 210
26.2 Construction of the Ergodic Decomposition . . . . . . . . . . . . 212
26.3 Statistical Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . 216
26.3.1 Ergodic Components as Minimal Sufficient Statistics . . . 216
CONTENTS v

26.3.2 Testing Ergodic Hypotheses . . . . . . . . . . . . . . . . . 218

26.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219

27 Mixing 220
27.1 Definition and Measurement of Mixing . . . . . . . . . . . . . . . 221
27.2 Examples of Mixing Processes . . . . . . . . . . . . . . . . . . . . 223
27.3 Convergence of Distributions Under Mixing . . . . . . . . . . . . 223
27.4 A Central Limit Theorem for Mixing Sequences . . . . . . . . . . 225

VI Information Theory 227

28 Entropy and Divergence 228
28.1 Shannon Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
28.2 Relative Entropy or Kullback-Leibler Divergence . . . . . . . . . 231
28.2.1 Statistical Aspects of Relative Entropy . . . . . . . . . . . 232
28.3 Mutual Information . . . . . . . . . . . . . . . . . . . . . . . . . 234
28.3.1 Mutual Information Function . . . . . . . . . . . . . . . . 234

29 Rates and Equipartition 236

29.1 Information-Theoretic Rates . . . . . . . . . . . . . . . . . . . . . 236
29.2 Asymptotic Equipartition . . . . . . . . . . . . . . . . . . . . . . 239
29.2.1 Typical Sequences . . . . . . . . . . . . . . . . . . . . . . 242
29.3 Asymptotic Likelihood . . . . . . . . . . . . . . . . . . . . . . . . 243
29.3.1 Asymptotic Equipartition for Divergence . . . . . . . . . 243
29.3.2 Likelihood Results . . . . . . . . . . . . . . . . . . . . . . 243
29.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243

VII Large Deviations 245

30 Large Deviations: Basics 246
30.1 Large Deviation Principles: Main Definitions and Generalities . . 246
30.2 Breeding Large Deviations . . . . . . . . . . . . . . . . . . . . . . 250

31 IID Large Deviations 256

31.1 Cumulant Generating Functions and Relative Entropy . . . . . . 257
31.2 Large Deviations of the Empirical Mean in Rd . . . . . . . . . . . 260
31.3 Large Deviations of the Empirical Measure in Polish Spaces . . . 263
31.4 Large Deviations of the Empirical Process in Polish Spaces . . . 264

32 Large Deviations for Markov Sequences 266

32.1 Large Deviations for Pair Measure of Markov Sequences . . . . . 266
32.2 Higher LDPs for Markov Sequences . . . . . . . . . . . . . . . . . 270
CONTENTS vi

33 The Gärtner-Ellis Theorem 271

33.1 The Gärtner-Ellis Theorem . . . . . . . . . . . . . . . . . . . . . 271
33.2 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275

34 Freidlin-Wentzell Theory 276

34.1 Large Deviations of the Wiener Process . . . . . . . . . . . . . . 277
34.2 Large Deviations for SDEs with State-Independent Noise . . . . 282
34.3 Large Deviations for State-Dependent Noise . . . . . . . . . . . . 283
34.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284

Bibliography 284
Definitions, Lemmas,
Propositions, Theorems,
Corollaries, Examples and
Exercises

Chapter 1 Basic Definitions: Indexed Collections and Random

Functions 3
Definition 1 A Stochastic Process Is a Collection of Random Vari-
ables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Example 2 Random variables . . . . . . . . . . . . . . . . . . . . 4
Example 3 Random vector . . . . . . . . . . . . . . . . . . . . . . 4
Example 4 One-sided random sequences . . . . . . . . . . . . . . 4
Example 5 Two-sided random sequences . . . . . . . . . . . . . . 4
Example 6 Spatially-discrete random fields . . . . . . . . . . . . . 4
Example 7 Continuous-time random processes . . . . . . . . . . . 4
Example 8 Random set functions . . . . . . . . . . . . . . . . . . 4
Example 9 One-sided random sequences of set functions . . . . . 4
Example 10 Empirical distributions . . . . . . . . . . . . . . . . . 5
Definition 11 Cylinder Set . . . . . . . . . . . . . . . . . . . . . . 5
Definition 12 Product σ-field . . . . . . . . . . . . . . . . . . . . 6
Definition 13 Random Function; Sample Path . . . . . . . . . . . 6
Definition 14 Functional of the Sample Path . . . . . . . . . . . . 6
Definition 15 Projection Operator, Coordinate Map . . . . . . . . 6
Theorem 16 Product σ-field-measurability is equvialent to mea-
surability of all coordinates . . . . . . . . . . . . . . . . . 6
Definition 17 A Stochastic Process Is a Random Function . . . . 7
Corollary 18 Measurability of constrained sample paths . . . . . 7
Example 19 Random Measures . . . . . . . . . . . . . . . . . . . 7
Example 20 Point Processes . . . . . . . . . . . . . . . . . . . . . 7
Example 21 Continuous random processes . . . . . . . . . . . . . 7
Exercise 1.1 The product σ-field answers countable questions . . 7

vii
CONTENTS viii

Exercise 1.2 The product σ-field constrained to a given set of paths 7

Chapter 2 Building Infinite Processes from Finite-Dimensional

Distributions 15
Definition 22 Finite-dimensional distributions . . . . . . . . . . . 15
Theorem 23 Finite-dimensional distributions determine process
distributions . . . . . . . . . . . . . . . . . . . . . . . . . 16
Definition 24 Projective Family of Distributions . . . . . . . . . . 17
Lemma 25 FDDs Form Projective Families . . . . . . . . . . . . 17
Proposition 26 Randomization, transfer . . . . . . . . . . . . . . 18
Theorem 27 Daniell Extension Theorem . . . . . . . . . . . . . . 18
Proposition 28 Carathéodory Extension Theorem . . . . . . . . . 19
Theorem 29 Kolmogorov Extension Theorem . . . . . . . . . . . 19

Chapter 3 Building Infinite Processes from Regular Conditional

Probability Distributions 21
Definition 30 Probability Kernel . . . . . . . . . . . . . . . . . . 21
Definition 31 Composition of probability kernels . . . . . . . . . 22
Proposition 32 Set functions continuous at ∅ . . . . . . . . . . . . 23
Theorem 33 Ionescu Tulcea Extension Theorem . . . . . . . . . . 23
Exercise 3.1 Lomnick-Ulam Theorem on infinite product measures 25
Exercise 3.2 Measures of cylinder sets . . . . . . . . . . . . . . . 25

Chapter 4 One-Parameter Processes, Usually Functions of Time 27

Definition 34 One-Parameter Process . . . . . . . . . . . . . . . . 28
Example 35 Bernoulli process . . . . . . . . . . . . . . . . . . . . 28
Example 36 Markov models . . . . . . . . . . . . . . . . . . . . . 28
Example 37 “White Noise” (Not Really) . . . . . . . . . . . . . . 28
Example 38 Wiener Process . . . . . . . . . . . . . . . . . . . . . 29
Example 39 Logistic Map . . . . . . . . . . . . . . . . . . . . . . 29
Example 40 Symbolic Dynamics of the Logistic Map . . . . . . . 29
Example 41 IID Samples . . . . . . . . . . . . . . . . . . . . . . . 30
Example 42 Non-IID Samples . . . . . . . . . . . . . . . . . . . . 30
Example 43 Estimating Distributions . . . . . . . . . . . . . . . . 30
Example 44 Doob’s Martingale . . . . . . . . . . . . . . . . . . . 30
Example 45 The One-Dimensional Ising Model . . . . . . . . . . 30
Example 46 Text . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Example 47 Polymer Sequences . . . . . . . . . . . . . . . . . . . 31
Definition 48 Shift Operators . . . . . . . . . . . . . . . . . . . . 31
Exercise 4.1 Existence of proto-Wiener processes . . . . . . . . . 32
Exercise 4.2 Time-Evolution Semi-Group . . . . . . . . . . . . . . 32

Chapter 5 Stationary One-Parameter Processes 34

Definition 49 Strong Stationarity . . . . . . . . . . . . . . . . . . 34
Definition 50 Weak Stationarity . . . . . . . . . . . . . . . . . . . 35
Definition 51 Conditional (Strong) Stationarity . . . . . . . . . . 35
CONTENTS ix

Theorem 52 Stationarity is Shift-Invariance . . . . . . . . . . . . 35

Definition 53 Measure-Preserving Transformation . . . . . . . . . 36
Corollary 54 Measure-preservation implies stationarity . . . . . . 36
Exercise 5.1 Functions of Stationary Processes . . . . . . . . . . . 37
Exercise 5.2 Continuous Measure-Preserving Families of Trans-
formations . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Exercise 5.3 The Logistic Map as a Measure-Preserving Trans-
formation . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

Chapter 6 Random Times and Their Properties 38

Definition 55 Filtration . . . . . . . . . . . . . . . . . . . . . . . . 38
Definition 56 Adapted Process . . . . . . . . . . . . . . . . . . . 39
Definition 57 Stopping Time, Optional Time . . . . . . . . . . . 39
Definition 58 Fτ for a Stopping Time τ . . . . . . . . . . . . . . 39
Definition 59 Hitting Time . . . . . . . . . . . . . . . . . . . . . . 40
Example 60 Fixation through Genetic Drift . . . . . . . . . . . . 40
Example 61 Stock Options . . . . . . . . . . . . . . . . . . . . . . 40
Definition 62 First Passage Time . . . . . . . . . . . . . . . . . . 40
Definition 63 Return Time, Recurrence Time . . . . . . . . . . . 41
Proposition 64 Some Sufficient Conditions for Waiting Times to
be Weakly Optional . . . . . . . . . . . . . . . . . . . . . 41
Lemma 65 Some Recurrence Relations for Kac’s Theorem . . . . 42
Theorem 66 Recurrence in Stationary Processes . . . . . . . . . . 42
Corollary 67 Poincaré Recurrence Theorem . . . . . . . . . . . . 43
Corollary 68 “Nietzsche” . . . . . . . . . . . . . . . . . . . . . . . 43
Theorem 69 Kac’s Recurrence Theorem . . . . . . . . . . . . . . 44
Example 70 Counter-example for Kac’s Recurrence Theorem . . 45
Exercise 6.1 Weakly Optional Times and Right-Continuous Fil-
trations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Exercise 6.2 Discrete Stopping Times and Their σ-Algebras . . . 45
Exercise 6.3 Kac’s Theorem for the Logistic Map . . . . . . . . . 45

Chapter 7 Continuity of Stochastic Processes 46

Definition 71 Continuity in Mean . . . . . . . . . . . . . . . . . . 46
Definition 72 Continuity in Probability, Stochastic Continuity . . 47
Definition 73 Continuous Sample Paths . . . . . . . . . . . . . . 47
Definition 74 Cadlag . . . . . . . . . . . . . . . . . . . . . . . . . 47
Definition 75 Versions of a Stochastic Process, Stochastically Equiv-
alent Processes . . . . . . . . . . . . . . . . . . . . . . . . 47
Lemma 76 Versions Have Identical FDDs . . . . . . . . . . . . . 47
Definition 77 Indistinguishable Processes . . . . . . . . . . . . . . 48
Proposition 78 Continuous Sample Paths Have Measurable Extrema 48
Example 79 A Horrible Version of the proto-Wiener Process . . . 49
Definition 80 Separable Function . . . . . . . . . . . . . . . . . . 50
Lemma 81 Some Sufficient Conditions for Separability . . . . . . 50
Definition 82 Separable Process . . . . . . . . . . . . . . . . . . . 51
CONTENTS x

Exercise 7.1 Piecewise Linear Paths Not in the Product σ-Field . 51

Exercise 7.2 Indistinguishability of Right-Continuous Versions . . 51

Chapter 8 More on Continuity 52

Definition 83 Compactness, Compactification . . . . . . . . . . . 53
Proposition 84 Compact Spaces are Separable . . . . . . . . . . . 53
Proposition 85 Compactification is Possible . . . . . . . . . . . . 53
Example 86 Compactifying the Reals . . . . . . . . . . . . . . . . 53
Lemma 87 Alternative Characterization of Separability . . . . . . 54
Lemma 88 Confining Bad Behavior to a Measure Zero Event . . 54
Lemma 89 A Null Exceptional Set . . . . . . . . . . . . . . . . . 55
Theorem 90 Separable Versions, Separable Modification . . . . . 56
Corollary 91 Separable Modifications in Compactified Spaces . . 57
Corollary 92 Separable Versions with Prescribed Distributions Exist 57
Definition 93 Measurable random function . . . . . . . . . . . . . 58
Theorem 94 Exchanging Expectations and Time Integrals . . . . 58
Theorem 95 Measurable Separable Modifications . . . . . . . . . 58
Theorem 96 Cadlag Versions . . . . . . . . . . . . . . . . . . . . 58
Theorem 97 Continuous Versions . . . . . . . . . . . . . . . . . . 59
Definition 98 Modulus of continuity . . . . . . . . . . . . . . . . . 59
Lemma 99 Modulus of continuity and uniform continuity . . . . . 59
Definition 100 Hölder continuity . . . . . . . . . . . . . . . . . . 59
Theorem 101 Hölder-continuous versions . . . . . . . . . . . . . . 59

Chapter 9 Markov Processes 61

Definition 102 Markov Property . . . . . . . . . . . . . . . . . . . 61
Lemma 103 The Markov Property Extends to the Whole Future 61
Definition 104 Product of Probability Kernels . . . . . . . . . . . 62
Definition 105 Transition Semi-Group . . . . . . . . . . . . . . . 63
Theorem 106 Existence of Markov Process with Given Transition
Kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Definition 107 Invariant Distribution . . . . . . . . . . . . . . . . 64
Theorem 108 Stationarity and Invariance for Homogeneous Markov
Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
Definition 109 Natural Filtration . . . . . . . . . . . . . . . . . . 65
Definition 110 Comparison of Filtrations . . . . . . . . . . . . . . 65
Lemma 111 The Natural Filtration Is the Coarsest One to Which
a Process Is Adapted . . . . . . . . . . . . . . . . . . . . . 65
Theorem 112 Markovianity Is Preserved Under Coarsening . . . . 66
Example 113 The Logistic Map Shows That Markovianity Is Not
Preserved Under Refinement . . . . . . . . . . . . . . . . 66
Exercise 9.1 Extension of the Markov Property to the Whole Future 67
Exercise 9.2 Futures of Markov Processes Are One-Sided Markov
Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
Exercise 9.3 Discrete-Time Sampling of Continuous-Time Markov
Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
CONTENTS xi

Exercise 9.4 Stationarity and Invariance for Homogeneous Markov

Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
Exercise 9.5 Rudiments of Likelihood-Based Inference for Markov
Chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Exercise 9.6 Implementing the MLE for a Simple Markov Chain . 68
Exercise 9.7 The Markov Property and Conditional Independence
from the Immediate Past . . . . . . . . . . . . . . . . . . 68
Exercise 9.8 Higher-Order Markov Processes . . . . . . . . . . . . 69
Exercise 9.9 AR(1) Models . . . . . . . . . . . . . . . . . . . . . 69

Chapter 10 Alternative Characterizations of Markov Processes 70

Theorem 114 Markov Sequences as Transduced Noise . . . . . . . 70
Definition 115 Transducer . . . . . . . . . . . . . . . . . . . . . . 71
Definition 116 Markov Operator on Measures . . . . . . . . . . . 72
Definition 117 Markov Operator on Densities . . . . . . . . . . . 72
Lemma 118 Markov Operators on Measures Induce Those on
Densities . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
Definition 119 Transition Operators . . . . . . . . . . . . . . . . 73
Definition 120 L∞ Transition Operators . . . . . . . . . . . . . . 73
Lemma 121 Kernels and Operators . . . . . . . . . . . . . . . . . 74
Definition 122 Functional . . . . . . . . . . . . . . . . . . . . . . 74
Definition 123 Conjugate or Adjoint Space . . . . . . . . . . . . . 74
Proposition 124 Conjugate Spaces are Vector Spaces . . . . . . . 74
Proposition 125 Inner Product is Bilinear . . . . . . . . . . . . . 74
Example 126 Vectors in Rn . . . . . . . . . . . . . . . . . . . . . 74
Example 127 Row and Column Vectors . . . . . . . . . . . . . . 74
Example 128 Lp spaces . . . . . . . . . . . . . . . . . . . . . . . . 75
Example 129 Measures and Functions . . . . . . . . . . . . . . . 75
Definition 130 Adjoint Operator . . . . . . . . . . . . . . . . . . 75
Proposition 131 Adjoint of a Linear Operator . . . . . . . . . . . 75
Lemma 132 Markov Operators on Densities and L∞ Transition
Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
Theorem 133 Transition operator semi-groups and Markov pro-
cesses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
Lemma 134 Markov Operators are Contractions . . . . . . . . . . 76
Lemma 135 Markov Operators Bring Distributions Closer . . . . 76
Theorem 136 Invariant Measures Are Fixed Points . . . . . . . . 76
Exercise 10.1 Kernels and Operators . . . . . . . . . . . . . . . . 76
Exercise 10.2 L1 and L∞ . . . . . . . . . . . . . . . . . . . . . . 77
Exercise 10.3 Operators and Expectations . . . . . . . . . . . . . 77
Exercise 10.4 Bayesian Updating as a Markov Process . . . . . . 77
Exercise 10.5 More on Bayesian Updating . . . . . . . . . . . . . 77
CONTENTS xii

Chapter 11 Examples of Markov Processes 78

Definition 137 Processes with Stationary and Independent Incre-
ments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
Definition 138 Lévy Processes . . . . . . . . . . . . . . . . . . . . 82
Example 139 Wiener Process is Lévy . . . . . . . . . . . . . . . . 82
Example 140 Poisson Counting Process . . . . . . . . . . . . . . 82
Theorem 141 Processes with Stationary Independent Increments
are Markovian . . . . . . . . . . . . . . . . . . . . . . . . 83
Theorem 142 Time-Evolution Operators of Processes with Sta-
tionary, Independent Increments . . . . . . . . . . . . . . 83
Definition 143 Infinitely-Divisible Distributions and Random Vari-
ables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
Proposition 144 Limiting Distributions Are Infinitely Divisible . 84
Theorem 145 Infinitely Divisible Distributions and Stationary In-
dependent Increments . . . . . . . . . . . . . . . . . . . . 84
Corollary 146 Infinitely Divisible Distributions and Lévy Processes 84
Definition 147 Self-similarity . . . . . . . . . . . . . . . . . . . . . 85
Definition 148 Stable Distributions . . . . . . . . . . . . . . . . . 86
Theorem 149 Scaling in Stable Lévy Processes . . . . . . . . . . . 86
Exercise 11.1 Wiener Process with Constant Drift . . . . . . . . . 86
Exercise 11.2 Perron-Frobenius Operators . . . . . . . . . . . . . 86
Exercise 11.3 Continuity of the Wiener Process . . . . . . . . . . 86
Exercise 11.4 Independent Increments with Respect to a Filtration 86
Exercise 11.5 Poisson Counting Process . . . . . . . . . . . . . . 86
Exercise 11.6 Poisson Distribution is Infinitely Divisible . . . . . 87
Exercise 11.7 Self-Similarity in Lévy Processes . . . . . . . . . . 87
Exercise 11.8 Gaussian Stability . . . . . . . . . . . . . . . . . . . 87
Exercise 11.9 Poissonian Stability? . . . . . . . . . . . . . . . . . 87
Exercise 11.10 Lamperti Transformation . . . . . . . . . . . . . . 87

Chapter 12 Generators of Markov Processes 88

Definition 150 Infinitessimal Generator . . . . . . . . . . . . . . . 89
Definition 151 Limit in the L-norm sense . . . . . . . . . . . . . . 89
Lemma 152 Generators are Linear . . . . . . . . . . . . . . . . . 89
Lemma 153 Invariant Distributions of a Semi-group Belong to
the Null Space of Its Generator . . . . . . . . . . . . . . . 89
Lemma 154 Invariant Distributions and the Generator of the
Time-Evolution Semigroup . . . . . . . . . . . . . . . . . 89
Lemma 155 Operators in a Semi-group Commute with Its Generator 90
Definition 156 Time Derivative in Function Space . . . . . . . . . 90
Lemma 157 Generators and Derivatives at Zero . . . . . . . . . . 90
Theorem 158 The Derivative of a Function Evolved by a Semi-
Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Corollary 159 Initial Value Problems in Function Space . . . . . 91
Corollary 160 Derivative of Conditional Expectations of a Markov
Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
CONTENTS xiii

Definition 161 Resolvents . . . . . . . . . . . . . . . . . . . . . . 92

Definition 162 Yosida Approximation of Operators . . . . . . . . 92
Theorem 163 Hille-Yosida Theorem . . . . . . . . . . . . . . . . . 93
Corollary 164 Stochastic Approximation of Initial Value Problems 93
Exercise 12.1 Generators are Linear . . . . . . . . . . . . . . . . . 93
Exercise 12.2 Semi-Groups Commute with Their Generators . . . 94
Exercise 12.3 Generator of the Poisson Counting Process . . . . . 94

Chapter 13 The Strong Markov Property and Martingale Prob-

lems 95
Definition 165 Strongly Markovian at a Random Time . . . . . . 96
Definition 166 Strong Markov Property . . . . . . . . . . . . . . 96
Example 167 A Markov Process Which Is Not Strongly Markovian 96
Definition 168 Martingale Problem . . . . . . . . . . . . . . . . . 97
Proposition 169 Cadlag Nature of Functions in Martingale Problems 97
Lemma 170 Alternate Formulation of Martingale Problem . . . . 97
Theorem 171 Markov Processes Solve Martingale Problems . . . 97
Theorem 172 Solutions to the Martingale Problem are Strongly
Markovian . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Exercise 13.1 Strongly Markov at Discrete Times . . . . . . . . . 98
Exercise 13.2 Markovian Solutions of the Martingale Problem . . 98
Exercise 13.3 Martingale Solutions are Strongly Markovian . . . 98

Chapter 14 Feller Processes 99

Definition 173 Initial Distribution, Initial State . . . . . . . . . . 99
Lemma 174 Kernel from Initial States to Paths . . . . . . . . . . 100
Definition 175 Markov Family . . . . . . . . . . . . . . . . . . . . 100
Definition 176 Mixtures of Path Distributions (Mixed States) . . 100
Definition 177 Feller Process . . . . . . . . . . . . . . . . . . . . . 101
Definition 178 Contraction Operator . . . . . . . . . . . . . . . . 101
Definition 179 Strongly Continuous Semigroup . . . . . . . . . . 101
Definition 180 Positive Operator . . . . . . . . . . . . . . . . . . 101
Definition 181 Conservative Operator . . . . . . . . . . . . . . . . 102
Lemma 182 Continuous semi-groups produce continuous paths in
function space . . . . . . . . . . . . . . . . . . . . . . . . 102
Definition 183 Continuous Functions Vanishing at Infinity . . . . 102
Definition 184 Feller Semigroup . . . . . . . . . . . . . . . . . . . 102
Lemma 185 The First Pair of Feller Properties . . . . . . . . . . 103
Lemma 186 The Second Pair of Feller Properties . . . . . . . . . 103
Theorem 187 Feller Processes and Feller Semigroups . . . . . . . 103
Theorem 188 Generator of a Feller Semigroup . . . . . . . . . . . 103
Theorem 189 Feller Semigroups are Strongly Continuous . . . . . 103
Proposition 190 Cadlag Modifications Implied by a Kind of Mod-
ulus of Continuity . . . . . . . . . . . . . . . . . . . . . . 104
Lemma 191 Markov Processes Have Cadlag Versions When They
Don’t Move Too Fast (in Probability) . . . . . . . . . . . 104
CONTENTS xiv

Lemma 192 Markov Processes Have Cadlag Versions If They

Don’t Move Too Fast (in Expectation) . . . . . . . . . . . 104
Theorem 193 Feller Implies Cadlag . . . . . . . . . . . . . . . . . 105
Theorem 194 Feller Processes are Strongly Markovian . . . . . . 105
Theorem 195 Dynkin’s Formula . . . . . . . . . . . . . . . . . . . 106
Exercise 14.1 Yet Another Interpretation of the Resolvents . . . . 106
Exercise 14.2 The First Pair of Feller Properties . . . . . . . . . . 106
Exercise 14.3 The Second Pair of Feller Properties . . . . . . . . 106
Exercise 14.4 Dynkin’s Formula . . . . . . . . . . . . . . . . . . . 106
Exercise 14.5 Lévy and Feller Processes . . . . . . . . . . . . . . 106

Chapter 15 Convergence of Feller Processes 107

Definition 196 Convergence in Finite-Dimensional Distribution . 107
Lemma 197 Finite and Infinite Dimensional Distributional Con-
vergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
Definition 198 The Space D . . . . . . . . . . . . . . . . . . . . . 108
Definition 199 Modified Modulus of Continuity . . . . . . . . . . 108
Proposition 200 Weak Convergence in D(R+ , Ξ) . . . . . . . . . 108
Proposition 201 Sufficient Condition for Weak Convergence . . . 109
Definition 202 Closed and Closable Generators, Closures . . . . . 109
Definition 203 Core of an Operator . . . . . . . . . . . . . . . . . 109
Lemma 204 Feller Generators Are Closed . . . . . . . . . . . . . 109
Theorem 205 Convergence of Feller Processes . . . . . . . . . . . 110
Corollary 206 Convergence of Discret-Time Markov Processes on
Feller Processes . . . . . . . . . . . . . . . . . . . . . . . . 111
Definition 207 Pure Jump Markov Process . . . . . . . . . . . . . 111
Proposition 208 Pure-Jump Markov Processea and ODEs . . . . 112
Exercise 15.1 Poisson Counting Process . . . . . . . . . . . . . . 113
Exercise 15.2 Exponential Holding Times in Pure-Jump Processes 113
Exercise 15.3 Solutions of ODEs are Feller Processes . . . . . . . 113

Chapter 16 Convergence of Random Walks 114

Definition 209 Continuous-Time Random Walk (Cadlag) . . . . . 117
Lemma 210 Increments of Random Walks . . . . . . . . . . . . . 117
Lemma 211 Continuous-Time Random Walks are Pseudo-Feller . 117
Lemma 212 Evolution Operators of Random Walks . . . . . . . . 118
Theorem 213 Functional Central Limit Theorem (I) . . . . . . . 118
Lemma 214 Convergence of Random Walks in Finite-Dimensional
Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . 119
Theorem 215 Functional Central Limit Theorem (II) . . . . . . . 119
Corollary 216 The Invariance Principle . . . . . . . . . . . . . . . 120
Corollary 217 Donsker’s Theorem . . . . . . . . . . . . . . . . . . 121
Exercise 16.1 Example 167 Revisited . . . . . . . . . . . . . . . . 121
Exercise 16.2 Generator of the d-dimensional Wiener Process . . 121
Exercise 16.3 Continuous-time random walks are Markovian . . . 121
Exercise 16.4 Donsker’s Theorem . . . . . . . . . . . . . . . . . . 121
CONTENTS xv

Exercise 16.5 Diffusion equation . . . . . . . . . . . . . . . . . . . 122

Exercise 16.6 Functional CLT for Dependent Variables . . . . . . 122

Chapter 17 Diffusions and the Wiener Process 124

Definition 218 Diffusion . . . . . . . . . . . . . . . . . . . . . . . 124
Lemma 219 The Wiener Process Is a Martingale . . . . . . . . . 126
Definition 220 Wiener Processes with Respect to Filtrations . . . 127
Definition 221 Gaussian Process . . . . . . . . . . . . . . . . . . . 127
Lemma 222 Wiener Process Is Gaussian . . . . . . . . . . . . . . 127
Theorem 223 Almost All Continuous Curves Are Non-Differentiable128

Chapter 18 Stochastic Integrals with the Wiener Process 130

Theorem 224 Martingale Characterization of the Wiener Process 131

Chapter 19 Stochastic Differential Equations 133

Definition 225 Progressive Process . . . . . . . . . . . . . . . . . 133
Definition 226 Non-anticipating filtrations, processes . . . . . . . 134
Definition 227 Elementary process . . . . . . . . . . . . . . . . . 134
Definition 228 Mean square integrable . . . . . . . . . . . . . . . 134
Definition 229 S2 norm . . . . . . . . . . . . . . . . . . . . . . . . 134
Proposition 230 k·kS2 is a norm . . . . . . . . . . . . . . . . . . . 134
Definition 231 Itô integral of an elementary process . . . . . . . . 135
Lemma 232 Approximation of Bounded, Continuous Processes
by Elementary Processes . . . . . . . . . . . . . . . . . . . 135
Lemma 233 Approximation by of Bounded Processes by Bounded,
Continuous Processes . . . . . . . . . . . . . . . . . . . . 136
Lemma 234 Approximation of Square-Integrable Processes by
Bounded Processes . . . . . . . . . . . . . . . . . . . . . . 136
Lemma 235 Approximation of Square-Integrable Processes by El-
ementary Processes . . . . . . . . . . . . . . . . . . . . . . 136
Lemma 236 Itô Isometry for Elementary Processes . . . . . . . . 136
Theorem 237 Itô Integrals of Approximating Elementary Pro-
cesses Converge . . . . . . . . . . . . . . . . . . . . . . . . 137
Definition 238 Itô integral . . . . . . . . . . . . . . . . . . . . . . 138
Corollary 239 The Itô isometry . . . . . . . . . . . . . . . . . . . 138
Definition 240 Itô Process . . . . . . . . . . . . . . . . . . . . . . 141
Lemma 241 Itô processes are non-anticipating . . . . . . . . . . . 141
Theorem 242 Itô’s Formula in One Dimension . . . . . . . . . . . 141
Example 243 Section 19.2.2 summarized . . . . . . . . . . . . . . 145
Definition 244 Multidimensional Itô Process . . . . . . . . . . . . 145
Theorem 245 Itô’s Formula in Multiple Dimensions . . . . . . . . 146
Theorem 246 Representation of Martingales as Stochastic Inte-
grals (Martingale Representation Theorem) . . . . . . . . 147
Theorem 247 Martingale Characterization of the Wiener Process 147
Definition 248 Stochastic Differential Equation, Solutions . . . . 147
Lemma 249 A Quadratic Inequality . . . . . . . . . . . . . . . . . 148
CONTENTS xvi

Definition 250 Maximum Process . . . . . . . . . . . . . . . . . . 148

Definition 251 The Space QM(T ) . . . . . . . . . . . . . . . . . 149
Lemma 252 Completeness of QM(T ) . . . . . . . . . . . . . . . . 149
Proposition 253 Doob’s Martingale Inequalities . . . . . . . . . . 149
Lemma 254 A Maximal Inequality for Itô Processes . . . . . . . . 149
Definition 255 Picard operator . . . . . . . . . . . . . . . . . . . 150
Lemma 256 Solutions are fixed points of the Picard operator . . 150
Lemma 257 A maximal inequality for Picard iterates . . . . . . . 150
Lemma 258 Gronwall’s Inequality . . . . . . . . . . . . . . . . . . 151
Theorem 259 Existence and Uniquness of Solutions to SDEs in
One Dimension . . . . . . . . . . . . . . . . . . . . . . . . 151
Theorem 260 Existence and Uniqueness for Multidimensional SDEs153
Exercise 19.1 Basic Properties of the Itô Integral . . . . . . . . . 155
Exercise 19.2 Martingale Properties of the Itô Integral . . . . . . 156
Exercise 19.3 Continuity of the Itô Integral . . . . . . . . . . . . 156
Exercise 19.4 “The square of dW ” . . . . . . . . . . . . . . . . . 156
Exercise 19.5 Itô integrals of elementary processes do not depend
on the break-points . . . . . . . . . . . . . . . . . . . . . . 156
Exercise 19.6 Itô integrals are Gaussian processes . . . . . . . . . 156
Exercise 19.7 A Solvable SDE . . . . . . . . . . . . . . . . . . . . 156
Exercise 19.8 Building Martingales from SDEs . . . . . . . . . . 157
Exercise 19.9 Brownian Motion and the Ornstein-Uhlenbeck Pro-
cess . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
Exercise 19.10 Again with the martingale characterization of the
Wiener process . . . . . . . . . . . . . . . . . . . . . . . . 157

Chapter 20 More on Stochastic Differential Equations 158

Theorem 261 Solutions of SDEs are Non-Anticipating and Con-
tinuous . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
Theorem 262 Solutions of SDEs are Strongly Markov . . . . . . . 159
Theorem 263 SDEs and Feller Diffusions . . . . . . . . . . . . . . 159
Corollary 264 Convergence of Initial Conditions and of Processes 160
Example 265 Wiener process, heat equation . . . . . . . . . . . . 162
Example 266 Ornstein-Uhlenbeck process . . . . . . . . . . . . . 163
Exercise 20.1 Langevin equation with a conservative force . . . . 163

Chapter 21 Large Deviations for Small-Noise Stochastic Differen-

tial Equations 164
Theorem 267 Small-Noise SDEs Converge in Probability on No-
Noise ODEs . . . . . . . . . . . . . . . . . . . . . . . . . . 166
Lemma 268 A Maximal Inequality for the Wiener Process . . . . 167
Lemma 269 A Tail Bound for Maxwell-Boltzmann Distributions . 167
Theorem 270 Upper Bound on the Rate of Convergence of Small-
Noise SDEs . . . . . . . . . . . . . . . . . . . . . . . . . . 168
CONTENTS xvii

Chapter 22 Spectral Analysis and L2 Ergodicity 170

Proposition 271 Linearity of White Noise Integrals . . . . . . . . 172
Proposition 272 White Noise Has Mean Zero . . . . . . . . . . . 172
Proposition 273 White Noise and Itô Integrals . . . . . . . . . . . 173
Proposition 274 White Noise is Uncorrelated . . . . . . . . . . . 173
Proposition 275 White Noise is Gaussian and Stationary . . . . . 173
Definition 276 Autocovariance Function . . . . . . . . . . . . . . 174
Lemma 277 Autocovariance and Time Reversal . . . . . . . . . . 174
Definition 278 Second-Order Process . . . . . . . . . . . . . . . . 174
Definition 279 Spectral Representation, Power Spectrum . . . . . 174
Lemma 280 Regularity of the Spectral Process . . . . . . . . . . 175
Definition 281 Jump of the Spectral Process . . . . . . . . . . . . 175
Proposition 282 Spectral Representations of Weakly Stationary
Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
Definition 283 Orthogonal Increments . . . . . . . . . . . . . . . 176
Lemma 284 Orthogonal Spectral Increments and Weak Stationarity176
Definition 285 Spectral Function, Spectral Density . . . . . . . . 177
Theorem 286 Weakly Stationary Processes Have Spectral Functions177
Theorem 287 Existence of Weakly Stationary Processes with Given
Spectral Functions . . . . . . . . . . . . . . . . . . . . . . 179
Definition 288 Jump of the Spectral Function . . . . . . . . . . . 179
Lemma 289 Spectral Function Has Non-Negative Jumps . . . . . 179
Theorem 290 Wiener-Khinchin Theorem . . . . . . . . . . . . . . 179
Definition 291 Time Averages . . . . . . . . . . . . . . . . . . . . 181
Definition 292 Integral Time Scale . . . . . . . . . . . . . . . . . 182
Theorem 293 Mean-Square Ergodic Theorem (Finite Autocovari-
ance Time) . . . . . . . . . . . . . . . . . . . . . . . . . . 182
Corollary 294 Convergence Rate in the Mean-Square Ergodic
Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
Lemma 295 Mean-Square Jumps in the Spectral Process . . . . . 183
Lemma 296 The Jump in the Spectral Function . . . . . . . . . . 183
Lemma 297 Existence of L2 Limits for Time Averages . . . . . . 183
Lemma 298 The Mean-Square Ergodic Theorem . . . . . . . . . 184
Exercise 22.1 Mean-Square Ergodicity in Discrete Time . . . . . 184
Exercise 22.2 Mean-Square Ergodicity with Non-Zero Mean . . . 185
Exercise 22.3 Functions of Weakly Stationary Processes . . . . . 185
Exercise 22.4 Ergodicity of the Ornstein-Uhlenbeck Process? . . . 185

Chapter 23 Ergodic Properties and Ergodic Limits 187

Definition 299 Dynamical System . . . . . . . . . . . . . . . . . . 189
Lemma 300 Dynamical Systems are Markov Processes . . . . . . 189
Definition 301 Observable . . . . . . . . . . . . . . . . . . . . . . 189
Definition 302 Invariant Functions, Sets and Measures . . . . . . 189
Lemma 303 Invariant Sets are a σ-Algebra . . . . . . . . . . . . . 189
Lemma 304 Invariant Sets and Observables . . . . . . . . . . . . 190
Definition 305 Infinitely Often, i.o. . . . . . . . . . . . . . . . . . 190
CONTENTS xviii

Lemma 306 “Infinitely often” implies invariance . . . . . . . . . . 190

Definition 307 Invariance Almost Everywhere . . . . . . . . . . . 190
Lemma 308 Almost-invariant sets form a σ-algebra . . . . . . . . 190
Lemma 309 Invariance for simple functions . . . . . . . . . . . . 190
Definition 310 Time Averages . . . . . . . . . . . . . . . . . . . . 191
Lemma 311 Time averages are observables . . . . . . . . . . . . . 191
Definition 312 Ergodic Property . . . . . . . . . . . . . . . . . . . 191
Definition 313 Ergodic Limit . . . . . . . . . . . . . . . . . . . . 191
Lemma 314 Linearity of Ergodic Limits . . . . . . . . . . . . . . 192
Lemma 315 Non-negativity of ergodic limits . . . . . . . . . . . . 192
Lemma 316 Constant functions have the ergodic property . . . . 192
Lemma 317 Ergodic limits are invariantifying . . . . . . . . . . . 192
Lemma 318 Ergodic limits are invariant functions . . . . . . . . . 193
Lemma 319 Ergodic limits and invariant indicator functions . . . 193
Lemma 320 Ergodic properties of sets and observables . . . . . . 193
Lemma 321 Expectations of ergodic limits . . . . . . . . . . . . . 193
Lemma 322 Convergence of Ergodic Limits . . . . . . . . . . . . 194
Lemma 323 Cesàro Mean of Expectations . . . . . . . . . . . . . 194
Corollary 324 Replacing Boundedness with Uniform Integrability 194
Definition 325 Asymptotically Mean Stationary . . . . . . . . . . 195
Lemma 326 Stationary Implies Asymptotically Mean Stationary 195
Proposition 327 Vitali-Hahn Theorem . . . . . . . . . . . . . . . 195
Theorem 328 Stationary Means are Invariant Measures . . . . . . 195
Lemma 329 Expectations of Almost-Invariant Functions . . . . . 196
Lemma 330 Limit of Cesàro Means of Expectations . . . . . . . . 196
Lemma 331 Expectation of the Ergodic Limit is the AMS Expec-
tation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
Corollary 332 Replacing Boundedness with Uniform Integrability 197
Theorem 333 Ergodic Limits of Bounded Observables are Condi-
tional Expectations . . . . . . . . . . . . . . . . . . . . . . 197
Corollary 334 Ergodic Limits of Integrable Observables are Con-
ditional Expectations . . . . . . . . . . . . . . . . . . . . 197
Exercise 23.1 “Infinitely often” implies invariant . . . . . . . . . 197
Exercise 23.2 Invariant simple functions . . . . . . . . . . . . . . 197
Exercise 23.3 Ergodic limits of integrable observables . . . . . . . 197

Chapter 24 The Almost-Sure Ergodic Theorem 198

Definition 335 Lower and Upper Limiting Time Averages . . . . 199
Lemma 336 Lower and Upper Limiting Time Averages are Invariant199
Lemma 337 “Limits Coincide” is an Invariant Event . . . . . . . 199
Lemma 338 Ergodic properties under AMS measures . . . . . . . 199
Theorem 339 Birkhoff’s Almost-Sure Ergodic Theorem . . . . . . 199
Corollary 340 Birkhoff’s Ergodic Theorem for Integrable Observ-
ables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
CONTENTS xix

Chapter 25 Ergodicity and Metric Transitivity 204

Definition 341 Ergodic Systems, Processes, Measures and Trans-
formations . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
Definition 342 Metric Transitivity . . . . . . . . . . . . . . . . . . 205
Lemma 343 Metric transitivity implies ergodicity . . . . . . . . . 205
Lemma 344 Stationary Ergodic Systems are Metrically Transitive 205
Example 345 IID Sequences, Strong Law of Large Numbers . . . 206
Example 346 Markov Chains . . . . . . . . . . . . . . . . . . . . 206
Example 347 Deterministic Ergodicity: The Logistic Map . . . . 206
Example 348 Invertible Ergodicity: Rotations . . . . . . . . . . . 207
Example 349 Ergodicity when the Distribution Does Not Converge207
Theorem 350 Ergodicity and the Triviality of Invariant Functions 207
Theorem 351 The Ergodic Theorem for Ergodic Processes . . . . 208
Lemma 352 Ergodicity Implies Approach to Independence . . . . 208
Theorem 353 Approach to Independence Implies Ergodicity . . . 208
Exercise 25.1 Ergodicity implies an approach to independence . . 209
Exercise 25.2 Approach to independence implies ergodicity . . . . 209
Exercise 25.3 Invariant events and tail events . . . . . . . . . . . 209
Exercise 25.4 Ergodicity of ergodic Markov chains . . . . . . . . 209

Chapter 26 Decomposition of Stationary Processes into Ergodic

Components 210
Proposition 354 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
Proposition 355 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
Proposition 356 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
Definition 357 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
Proposition 358 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
Proposition 359 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
Proposition 360 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
Proposition 361 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
Proposition 362 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
Definition 363 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
Proposition 364 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
Definition 365 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
Proposition 366 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
Proposition 367 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
Proposition 368 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Lemma 369 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Lemma 370 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Lemma 371 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Theorem 372 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
Definition 373 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
Lemma 374 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
Theorem 375 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
Theorem 376 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
Corollary 377 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
CONTENTS xx

Corollary 378 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218

Exericse 26.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219

Chapter 27 Mixing 220

Definition 379 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
Lemma 380 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
Theorem 381 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
Definition 382 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
Lemma 383 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
Theorem 384 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
Example 385 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Example 386 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Example 387 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Example 388 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Lemma 389 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
Theorem 390 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
Theorem 391 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
Definition 392 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
Lemma 393 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
Definition 394 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
Definition 395 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
Theorem 396 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226

Chapter 28 Shannon Entropy and Kullback-Leibler Divergence 228

Definition 397 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
Lemma 398 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
Definition 399 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Definition 400 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Lemma 401 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Lemma 402 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Definition 403 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Lemma 404 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
Lemma 405 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
Definition 406 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
Lemma 407 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Lemma 408 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Corollary 409 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Definition 410 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Corollary 411 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
Definition 412 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
Proposition 413 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
Proposition 414 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
Definition 415 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
Theorem 416 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
CONTENTS xxi

Chapter 29 Entropy Rates and Asymptotic Equipartition 236

Definition 417 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
Definition 418 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
Lemma 419 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
Theorem 420 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
Theorem 421 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
Lemma 422 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
Example 423 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
Example 424 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
Definition 425 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
Definition 426 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
Lemma 427 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
Lemma 428 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
Lemma 429 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
Definition 430 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
Lemma 431 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
Lemma 432 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
Lemma 433 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
Lemma 434 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
Theorem 435 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
Theorem 436 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
Theorem 437 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
Exercise 29.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
Exercise 29.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244

Chapter 30 General Theory of Large Deviations 246

Definition 438 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
Definition 439 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
Lemma 440 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
Lemma 441 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
Definition 442 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
Lemma 443 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
Definition 444 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
Lemma 445 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
Lemma 446 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
Definition 447 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
Lemma 448 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
Lemma 449 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
Theorem 450 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
Theorem 451 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
Definition 452 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
Definition 453 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
Definition 454 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
Definition 455 Empirical Process Distribution . . . . . . . . . . . 252
Corollary 456 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
Corollary 457 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
CONTENTS xxii

Definition 458 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253

Theorem 459 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
Theorem 460 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
Theorem 461 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
Definition 462 Exponentially Equivalent Random Variables . . . 254
Lemma 463 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254

Chapter 31 Large Deviations for IID Sequences: The Return of

Relative Entropy 256
Definition 464 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
Definition 465 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
Lemma 466 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
Definition 467 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
Definition 468 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
Lemma 469 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
Theorem 470 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
Proposition 471 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
Proposition 472 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
Lemma 473 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
Lemma 474 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
Theorem 475 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
Corollary 476 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
Theorem 477 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265

Chapter 32 Large Deviations for Markov Sequences 266

Theorem 478 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
Corollary 479 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
Corollary 480 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
Corollary 481 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
Corollary 482 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
Theorem 483 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
Theorem 484 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270

Chapter 33 Large Deviations for Weakly Dependent Sequences via

the Gärtner-Ellis Theorem 271
Definition 485 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
Lemma 486 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
Lemma 487 Upper Bound in the Gärtner-Ellis Theorem for Com-
pact Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
Lemma 488 Upper Bound in the Gärtner-Ellis Theorem for Closed
Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
Definition 489 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
Lemma 490 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
Lemma 491 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
Definition 492 Exposed Point . . . . . . . . . . . . . . . . . . . . 273
Definition 493 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
CONTENTS xxiii

Lemma 494 Lower Bound in the Gärtner-Ellis Theorem for Open

Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
Theorem 495 Abstract Gärtner-Ellis Theorem . . . . . . . . . . . 274
Definition 496 Relative Interior . . . . . . . . . . . . . . . . . . . 274
Definition 497 Essentially Smooth . . . . . . . . . . . . . . . . . 274
Proposition 498 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
Theorem 499 Euclidean Gärtner-Ellis Theorem . . . . . . . . . . 274
Exercise 33.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
Exercise 33.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
Exercise 33.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
Exercise 33.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
Exercise 33.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275

Chapter 34 Large Deviations for Stochastic Differential Equations276

Definition 500 Cameron-Martin Spaces . . . . . . . . . . . . . . . 277
Lemma 501 Cameron-Martin Spaces are Hilbert . . . . . . . . . . 277
Definition 502 Effective Wiener Action . . . . . . . . . . . . . . . 278
Proposition 503 Girsanov Formula for Deterministic Drift . . . . 278
Lemma 504 Exponential Bound on the Probability of Tubes Around
Given Trajectories . . . . . . . . . . . . . . . . . . . . . . 278
Lemma 505 Trajectories are Rarely Far from Action Minima . . 279
Proposition 506 Compact Level Sets of the Wiener Action . . . . 280
Theorem 507 Schilder’s Theorem on Large Deviations of the Wiener
Process on the Unit Interval . . . . . . . . . . . . . . . . . 280
Corollary 508 Extension of Schilder’s Theorem to [0, T ] . . . . . 281
Corollary 509 Schilder’s Theorem on R+ . . . . . . . . . . . . . . 281
Definition 510 SDE with Small State-Independent Noise . . . . . 282
Definition 511 Effective Action under State-Independent Noise . 282
Lemma 512 Continuous Mapping from Wiener Process to SDE
Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
Lemma 513 Mapping Cameron-Martin Spaces Into Themselves . 283
Theorem 514 Freidlin-Wentzell Theorem with State-Independent
Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
Definition 515 SDE with Small State-Dependent Noise . . . . . . 283
Definition 516 Effective Action under State-Dependent Noise . . 284
Theorem 517 Freidlin-Wentzell Theorem for State-Dependent Noise284
Exercise 34.1 Cameron-Martin Spaces are Hilbert Spaces . . . . . 284
Exercise 34.2 Lower Bound in Schilder’s Theorem . . . . . . . . . 284
Exercise 34.3 Mapping Cameron-Martin Spaces Into Themselves 284
Preface: Stochastic
Processes in
Measure-Theoretic
Probability

This is intended to be a second course in stochastic processes (at least!); I am

going to assume you have all had a first course on stochastic processes, using
elementary probability theory, say our 36-703. You might then ask what the
added benefit is of taking this course, of re-studying stochastic processes within
the framework of measure-theoretic probability. There are a number of reasons
to do this.
First, the measure-theoretic framework allows us to greatly generalize the
range of processes we can consider. Topics like empirical process theory and
stochastic calculus are basically incomprehensible without the measure-theoretic
framework. Much of the impetus for developing measure-theoretic probability
in the first place came from the impossibility of properly handling continuous
random motion, especially the Wiener process, with only the tools of elementary
probability.
Second, even topics like Markov processes and ergodic theory, which can be
discussed without it, greatly benefit from measure-theoretic probability, because
it lets us establish important results which are beyond the reach of elementary
methods.
Third, many of the greatest names in twentieth century mathematics have
worked in this area, and the theories they have developed are profound, useful
and beautiful. Knowing them will make you a better person.

Definitions, lemmas, theorems, corollaries, examples, etc., are all numbered

together, consecutively across lectures. Exercises are separately numbered within
lectures.

1
 Part I

Stochastic Processes in
General

2
Chapter 1

Basic Definitions: Indexed

Collections and Random
Functions

Section 1.1 introduces stochastic processes as indexed collections

of random variables.
Section 1.2 builds the necessary machinery to consider random
functions, especially the product σ-field and the notion of sample
paths, and then re-defines stochastic processes as random functions
whose sample paths lie in nice sets.
This first chapter begins at the beginning, by defining stochastic processes.
Even if you have seen this definition before, it will be useful to review it.
We will flip back and forth between two ways of thinking about stochastic
processes: as indexed collections of random variables, and as random functions.
As always, assume we have a nice base probability space (Ω, F, P ), which is
rich enough that all the random variables we need exist.

1.1 So, What Is a Stochastic Process?

Definition 1 (A Stochastic Process Is a Collection of Random Vari-
ables) A stochastic process {Xt }t∈T is a collection of random variables Xt ,
taking values in a common measure space (Ξ, X ), indexed by a set T .
That is, for each t ∈ T , Xt (ω) is an F/X -measurable function from Ω to Ξ,
which induces a probability measure on Ξ in the usual way.
It’s sometimes more convenient to write X(t) in place of Xt . Also, when
S ⊂ T , Xs or X(S) refers to that sub-collection of random variables.
Example 2 (Random variables) Any single random variable is a (trivial)
stochastic process. (Take T = {1}, say.)

3
CHAPTER 1. BASICS 4

Example 3 (Random vector) Let T = {1, 2, . . . k} and Ξ = R. Then {Xt }t∈T

is a random vector in Rk .

Example 4 (One-sided random sequences) Let T = {1, 2, . . .} and Ξ be

some finite set (or R or C or Rk . . . ). Then {Xt }t∈T is a one-sided discrete
(real, complex, vector-valued, . . . ) random sequence. Most of the stochas-
tic processes you have encountered are probably of this sort: Markov chains,
discrete-parameter martingales, etc. Figures 1.2, 1.3, 1.4 and 1.5 illustrate
some one-sided random sequences.

Example 5 (Two-sided random sequences) Let T = Z and Ξ be as in

Example 4. Then {Xt }t∈T is a two-sided random sequence.

Example 6 (Spatially-discrete random fields) Let T = Zd and Ξ be as in

Example 4. Then {Xt }t∈T is a d-dimensional spatially-discrete random field.

Example 7 (Continuous-time random processes) Let T = R and Ξ = R.

Then {Xt }t∈T is a real-valued, continuous-time random process (or random mo-
tion or random signal). Figures 1.6 and 1.7 illustrate some of the possibilities.

Vector-valued processes are an obvious generalization.

Example 8 (Random set functions) Let T = B, the Borel field on the reals,
+
and Ξ = R , the non-negative extended reals. Then {Xt }t∈T is a random set
function on the reals.

The definition of random set functions on Rd is entirely parallel. Notice that

if we want not just a set function, but a measure or a probability measure,
this will imply various forms of dependence among the random variables in the
collection, e.g., a measure must respect countable additivity over disjoint sets.
We will return to this topic in the next section.

Example 9 (One-sided random sequences of set functions) Let T =

+
B × N and Ξ = R . Then {Xt }t∈T is a one-sided random sequence of set
functions.

Example 10 (Empirical distributions) Suppose Zi , = 1, 2, . . . are indepen-

dent, identically-distributed real-valued random variables. (We can see from Ex-
ample 4 that this is a one-sided real-valued random sequence.) For each Borel
set B and each n, define
n
1X
P̂n (B) = 1B (Zi )
n i=1
CHAPTER 1. BASICS 5

i.e., the fraction of the samples up to time n which fall into that set. This is
the empirical measure. P̂n (B) is a one-sided random sequence of set functions
— in fact, of probability measures. We would like to be able to say something
about how it behaves. It would be very reassuring, for instance, to be able to
show that it converges to the common distribution of the Zi (Figure 1.8).

1.2 Random Functions

X(t, ω) has two arguments, t and ω. For each fixed value of t, Xt (ω) is straight-
forward random variable. For each fixed value of ω, however, X(t) is a function
from T to Ξ — a random function. The advantage of the random function
perspective is that it lets us consider the realizations of stochastic processes as
single objects, rather than large collections. This isn’t just tidier; we will need
to talk about relations among the variables in the collection or their realiza-
tions, rather than just properties of individual variables, and this will help us
do so. In Example 10, it’s important that we’ve got random probability mea-
sures, rather than just random set functions, so we need to require that, e.g.,
P̂n (A ∪ B) = P̂n (A) + P̂n (B) when A and B are disjoint Borel sets, and this is
a relationship among the three random variables P̂n (A), P̂n (B) and P̂n (A ∪ B).
Plainly, working out all the dependencies involved here is going to get rather
tedious, so we’d like a way to talk about acceptable realizations of the whole
stochastic process. This is what the random functions notion will let us do.
We’ll make this more precise by defining a random function as a function-
valued random variable. To do this, we need a measure space of functions, and
a measurable mapping from (Ω, F, P ) to that function space. To get a measure
space, we need a carrier set and a σ-field on it. The natural set to use is ΞT ,
the set of all functions from T to Ξ. (We’ll see how to restrict this to just the
functions we want presently.) Now, how about the σ-field?

Definition 11 (Cylinder Set) Given an index set T and a collection of σ-

fields
Q Xt on spaces Ξt , t ∈ T . Pick any t ∈ T and any At ∈ Xt . Then At ×
s6=t s is a one-dimensional cylinder set.
Ξ

For any finite k, k−dimensional cylinder sets are defined similarly, and
clearly are the intersections of k different one-dimensional cylinder sets. To
see why they have this name, notice a cylinder, in Euclidean geometry, con-
sists of all the points where the x and y coordinates fall into a certain set
(the base),
Q leaving the z coordinate unconstrained. Similarly, a cylinder set
like At × s6=t Ξs consists of all the functions in ΞT where f (t) ∈ At , and are
otherwise unconstrained.

Definition 12 (Product σ-field) The product σ-field, ⊗t∈T Xt , is the σ-field

over ΞT generated by all the one-dimensional cylinder sets. If all the Xt are the
same, X , we write the product σ-field as X T .
CHAPTER 1. BASICS 6

The product σ-field is enough to let us define a random function, and is

going to prove to be almost enough for all our purposes.
Definition 13 (Random Function; Sample Path) A Ξ-valued random func-
tion on T is a map X : Ω 7→ ΞT which is F/X T -measurable. The realizations
of X are functions x(t) taking values in Ξ, called its sample paths.

N.B., it has become common to apply the term “sample path” or even just
“path” even in situations where the geometric analogy it suggests may be some-
what misleading. For instance, for the empirical distributions of Example 10,
the “sample path” is the measure P̂n , not the curves shown in Figure 1.8.
Definition 14 (Functional of the Sample Path) Let E, E be a measure-
space. A functional of the sample path is a mapping f : ΞT 7→ E which is
X T /E-measurable.
Examples of useful and common functionals include maxima, minima, sam-
ple averages, etc. Notice that none of these are functions of any one random
variable, and in fact their value cannot be determined from any part of the
sample path smaller than the whole thing.
Definition 15 (Projection Operator, Coordinate Map) A projection op-
erator or coordinate map πt is a map from ΞT to Ξ such that πt X = X(t).
The projection operators are a convenient device for recovering the individ-
ual coordinates — the random variables in the collection — from the random
function. Obviously, as t ranges over T , πt X gives us a collection of random vari-
ables, i.e., a stochastic process in the sense of our first definition. The following
lemma lets us go back and forth between the collection-of-variables, coordinate
view, and the entire-function, sample-path view.
Theorem 16 (Product σ-field-measurability is equvialent to measur-
ability of all coordinates) X is F/ ⊗t∈T Xt -measurable iff πt X is F/Xt -
measurable for every t.
Proof: This follows from the fact that the one-dimensional cylinder sets
generate the product σ-field. 
We have said before that we will want to constrain our stochastic processes
to have certain properties — to be probability measures, rather than just set
functions, or to be continuous, or twice differentiable, etc. Write the set of all
such functions in ΞT as U . Notice that U does not have to be an element of the
product σ-field, and in general is not. (We will consider some of the reasons for
this later.) As usual, by U ∩ X T we will mean the collection of all sets of the
form U ∩ C, where C ∈ X T . Notice that (U, U ∩ X T ) is a measure space. What
we want is to ensure that the sample path of our random function lies in U .
Definition 17 (A Stochastic Process Is a Random Function) A Ξ-valued
stochastic process on T with paths in U , U ⊆ ΞT , is a random function X : Ω 7→
U which is F/U ∩ X T -measurable.
CHAPTER 1. BASICS 7

Corollary 18 (Measurability of constrained sample paths) A function

X from Ω to U is F/U ∩ X T -measurable iff Xt is F/X -measurable for all t.

Proof: Because X(ω) ∈ U , F/U ∩ X T measurability works just the same

way as F/X T measurability (Exercise 1.2). So imitate the proof of Theorem
16. 

Example 19 (Random Measures) Let T = B d , the Borel field on Rd , and let

+
Ξ = R , the non-negative extended reals. Then ΞT is the class of set functions
on Rd . Let M be the class of such set functions which are also measures (i.e.,
which are countably additive and give zero on the null set). Then a random set
function X with realizations in M is a random measure.

Example 20 (Point Processes) Let X be a random measure, as in the previ-

ous example. If X(B) is a finite integer for every bounded Borel set B, then X
is a point process. If in addition X(r) ≤ 1 for every r ∈ Rd , then X is simple.
The Poisson process is a simple point process. See Figure 1.1.

Example 21 (Continuous random processes) Let T = R+ , Ξ = Rd , and

C(T ) the class of continuous functions from T to Ξ (in the usual topology). Then
a Ξ-valued random process on T with paths in C(T ) is a continuous random
process. The Wiener process, or Brownian motion, is an example. We will see
that most sample paths in C(T ) are not differentiable.

1.3 Exercises
Exercise
S 1.1 (The product σ-field answers countable questions) Let
D = S X S , where the union ranges over all countable subsets S of the index
set T . For any event D ∈ D, whether or not a sample path x ∈ D depends on
the value of xt at only a countable number of indices t.
1. Show that D is a σ-field.

2. Show that if A ∈ X T , then A ∈ X S for some countable subset S of T .

Exercise 1.2 (The product σ-field constrained to a given set of paths)

Let U ⊂ X T be a set of allowed paths. Show that
1. U ∩ X T is a σ-field on U ;

2. U ∩ X T is generated by sets of the form U ∩ B, where B ∈ X S for some

finite subset S of T .
CHAPTER 1. BASICS 8

Figure 1.1: Examples of point processes. The top row shows the dates of ap-
pearances of 44 genres of English novels (data taken from Moretti (2005)). The
bottom two rows show independent realizations of a Poisson process with the
same mean time between arrivals as the actual history. The number of tick-
marks falling within any measurable set on the horizontal axis determines an
integer-valued set function, in fact a measure.

AATGAAATAAAAAAAAACGAAAATAAAAAA
AAGGCCATTAAAGTTAAAATAATGAAAGGA
CAATGATTAGGACAATAACATACAAGTTAT
GGGGTTAATTAATGGTTAGGATGGGTTTTT
CCTTCAAAGTTAATGAAAAGTTAAAATTTA
TAAGTATTTGAAGCACAGCAACAACTAGGT

Figure 1.2: Examples of one-sided random sequences (Ξ = {A, C, G, T}). The

top line shows the first thirty bases of the first chromosome of the cellular
slime mold Dictyostelium discoideum (Eichinger et al., 2005), downloaded from
dictybase.org. The lower lines are independent samples from a simple Markov
model fitted to the full chromosome.
CHAPTER 1. BASICS 9

111111111111110110011011000001
011110110111111011110110110110
011110001101101111111111110000
011011011011111101111000111100
011011111101101100001111110111
000111111111111001101100011011

Figure 1.3: Examples of one-sided random sequences. These binary sequences

(Ξ = {0, 1}) are the first thirty steps from samples of a sofic process, one with
a finite number of underlying states which is nonetheless not a Markov chain
of any finite order. Can you figure out the rule specifying which sequences are
allowed and which are forbidden? (Hint: these are all samples from the even
process.)
CHAPTER 1. BASICS 10

●
● ●
●
●
2

● ● ●
●
● ●
● ●
● ● ● ●
●
● ● ●● ●●
● ●
●●
0
z

●
● ● ●
● ●
● ● ● ● ●
● ● ●
● ●
●
−2

● ●
● ●

●
−4

0 10 20 30 40 50

Figure 1.4: Examples of one-sided random sequences. These are linear Gaussian
random sequences, Xt+1 = 0.8Xt + Zt+1 , where the Zt are all i.i.d. N (0, 1), and
X1 = Z0 . Different shapes of dots represent different independent samples of
this autoregressive process. (The line segements are simply guides to the eye.)
This is a Markov sequence, but one with a continuous state-space.
CHAPTER 1. BASICS 11

1.0

● ● ●
● ● ● ● ● ●
●
●
● ● ●
●
0.8

●
●
0.6

● ●
● ●
●
●
x

●
0.4

● ●
●
● ● ●
●
0.2

●
● ●
● ● ●
● ● ●
● ●
0.0

● ● ●
●●●●

0 10 20 30 40 50

Figure 1.5: Nonlinear, non-Gaussian random sequences. Here X1 ∼ U (0, 1), i.e.,
uniformly distributed on the unit interval, and Xt+1 = 4Xt (1−Xt ). Notice that
while the two samples begin very close together, they rapidly separate; after a
few time-steps their locations are, in fact, effectively independent. We will study
both this approach to independence, known as mixing, and this example, known
as the logistic map, in some detail.
CHAPTER 1. BASICS 12

2
1
W

0
−1
−2

0.0 0.2 0.4 0.6 0.8 1.0

Figure 1.6: Continuous-time random processes. Shown are three samples from
the standard Wiener process, also known as Brownian motion, a Gaussian pro-
cess with independent increments and continuous trajectories. This is a central
part of the course, and actually what forced probability to be re-defined in terms
of measure theory.
CHAPTER 1. BASICS 13

3
2
1
X

0
−1
−2

0.0 0.2 0.4 0.6 0.8 1.0

Figure 1.7: Continuous-time random processes can have discontinuous trajecto-

ries. Here Xt = Wt +Jt , where Wt is a standard Wiener process, and Jt is piece-
wise constant (shown by the dashed lines), changing at t = 0.1, 0.2, 0.3, . . . 1.0.
The trajectory is discontinuous at t = 0.4, but continuous from the right there,
and there is a limit from the left. In fact, at every point the trajectory is con-
tinuous from the right and has a limit from the left. We will see many such
cadlag processeses.
CHAPTER 1. BASICS 14

1.0

1.0
● ●
●●
●
●●
●●
● ●●●
●
●●
●
●●
●
0.8

0.8
● ●●
●
●●
●●
●
●
●
●●
● ●●
●
●
●●
●
●
0.6

0.6
●
●
● ●
●
●
●
●
●
Fn(x)

Fn(x)
●
●
●
●
● ●●
●
●
●
●
●
●
0.4

0.4
●
●
● ●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●
●
●
0.2

0.2
●
●
● ●
●
●●
●
●
●
●
●
●
● ●●
●
●
●
●
●
●
0.0

0.0
●
●

0 1 2 3 4 5 6 0 2 4 6 8 10 12

x x
1.0

●● ● ● ●● ●
● ● ●
●
●●
●
●●
● ●●
●●●
● ● ●
●●
● ●
●
●●●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
0.8

●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
0.6

●
●
●
●
●
●
●
●
Fn(x)

●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.4

●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.2

●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0

●
●
●

0 5 10 15 20 25

Figure 1.8: Empirical cumulative distribution functions for successively larger

samples from the standard log-normal distribution (from left to right, n =
10, 100, 1000), with the theoretical CDF as the smooth dashed line. Because
the intervals of the form (−∞, a] are a generating class for the Borel σ-field, the
empirical C.D.F. suffices to represent the empirical distribution.
Chapter 2

Building Infinite Processes

from Finite-Dimensional
Distributions

Section 2.1 introduces the finite-dimensional distributions of a

stochastic process, and shows how they determine its infinite-dimensional
distribution.
Section 2.2 considers the consistency conditions satisfied by the
finite-dimensional distributions of a stochastic process, and the ex-
tension theorems (due to Daniell and Kolmogorov) which prove the
existence of stochastic processes with specified, consistent finite-
dimensional distributions.

2.1 Finite-Dimensional Distributions

So, we now have X, our favorite Ξ-valued stochastic process on T with paths
in U . Like any other random variable, it has a probability law or distribution,
which is defined over the entire set U . Generally, this is infinite-dimensional.
Since it is inconvenient to specify distributions over infinite-dimensional spaces
all in a block, we consider the finite-dimensional distributions.

Definition 22 (Finite-dimensional distributions) The finite-dimensional

distributions of X are the the joint distributions of Xt1 , Xt2 , . . . Xtn , t1 , t2 , . . . tn ∈
T , n ∈ N.

You will sometimes see “FDDs” and “fidis” as abbreviations for “finite-
dimensional distributions”. Please do not use “fidis”.
We can at least hope to specify the finite-dimensional distributions. But we
are going to want to ask a lot of questions about asymptotics, and global proper-
ties of sample paths, which go beyond any finite dimension, so you might worry

15
CHAPTER 2. BUILDING PROCESSES 16

that we’ll still need to deal directly with the infinite-dimensional distribution.
The next theorem says that this worry is unfounded; the finite-dimensional dis-
tributions specify the infinite-dimensional distribution (pretty much) uniquely.

Theorem 23 (Finite-dimensional distributions determine process dis-

tributions) Let X and Y be two Ξ-valued processes on T with paths in U . Then
X and Y have the same distribution iff all their finite-dimensional distributions
agree.

Proof: “Only if”: Since X and Y have the same distribution, applying
the any given set of coordinate mappings will result in identically-distributed
random vectors, hence all the finite-dimensional distributions will agree.
“If”: We’ll use the π-λ theorem.
Let C be the finite cylinder sets, i.e., all sets of the form

C = x ∈ ΞT |(xt1 , xt2 , . . . xtn ) ∈ B



where n ∈ N, B ∈ X n , t1 , t2 , . . . tn ∈ T . Clearly, this is a π-system, since it is

closed under intersection.
Now let L consist of all the sets L ∈ X T where P (X ∈ L) = P (Y ∈ L).
We need to show that this is a λ-system, i.e., that it (i) includes ΞT , (ii) is
closed under complementation, and (iii)
is closed under

monotone increasing
limits. (i) is clearly true: P X ∈ ΞT = P Y ∈ ΞT = 1. (ii) is true because
we’re looking at a probability: if L ∈ L, then P (X ∈ Lc ) = 1 − P (X ∈ L) =
1 − P (Y ∈ L) = P (Y ∈ Lc ). To see (iii), let Ln ↑ L be a monotone-increasing
sequence of sets in L, and recall that, for any measure, Ln ↑ L implies µLn ↑ µL.
So P (X ∈ Ln ) ↑ P (X ∈ L), P (Y ∈ Ln ) ↑ P (Y ∈ L), and (since P (X ∈ Ln ) =
P (Y ∈ Ln )), P (X ∈ Ln ) ↑ P (Y ∈ L) as well. A sequence cannot have two
limits, so P (X ∈ L) = P (Y ∈ L), and L ∈ L.
Since the finite-dimensional distributions match, P (X ∈ C) = P (Y ∈ C) for
all C ∈ C, which means that C ⊂ L. Also, from the definition of the product
σ-field, σ(C) = X T . Hence, by the π-λ theorem, X T ⊆ L. 
A note of caution is in order here. If X is a Ξ-valued process on T whose
paths are constrained to line in U , and Y is a similar process that is not so
constrained, it is nonetheless possible that X and Y agree in all their finite-
dimensional distributions. The trick comes if U is not, itself, an element of X T .
The most prominent instance of this is when Ξ = R, T = R, and the constraint
is continuity of the sample paths: we will see that U 6∈ B R . (This is the point
of Exercise 1.1.)

2.2 Consistency and Extension

The finite-dimensional distributions of a given stochastic process are related
to one another in the usual way of joint and marginal distributions. Take
some collection of indices t1 , t2 . . . tn ∈ T , and corresponding measurable sets
CHAPTER 2. BUILDING PROCESSES 17

B1 ∈ X1 , B2 ∈ X2 , . . . Bn ∈ Xn . Then, for any m > n, and any further indices

tn+1 , tn2 , . . . tm , it must be the case that

P (Xt1 ∈ B1 , Xt2 ∈ B2 , . . . Xtn ∈ Bn ) (2.1)

= P Xt1 ∈ B1 , Xt2 ∈ B2 , . . . Xtn ∈ Bn , Xtn+1 ∈ Ξ, Xtn+2 ∈ Ξ, . . . Xtm ∈ Ξ

This is going to get really awkward to write over and over, so let’s introduce
some simplifying notation. Fin(T ) will denote the class of all finite sub-sets
of our index set T , and likewise Denum(T ) all denumerable sub-sets. We’ll
indicate such sub-sets, for the moment, by capital letters like J, K, etc., and
extend the definition of coordinate maps (Definition 15) so that πJ maps from
ΞT to ΞJ in the obvious way, and πJK maps from ΞK to ΞJ , if J ⊂ K. If µ is
the measure for the whole process, then the finite-dimensional distributions are
{µJ |J ∈ Fin(T )}. Clearly, µJ = µ ◦ πJ −1 .

Definition 24 (Projective Family of Distributions) A family of distri-

butions µJ , J ∈ Denum(T ), is projective when for every J, K ∈ Denum(T ),
J ⊂ K implies

−1
µJ = µK ◦ πJK (2.2)

Such a family is also said to be consistent or compatible (with one another).

Lemma 25 (FDDs Form Projective Families) The finite-dimensional dis-

tributions of a stochastic process always form a projective family.

Proof: This is just the fact that we get marginal distributions by integrating
out some variables from the joint distribution. But, to proceed formally: Letting
J and K be finite sets of indices, J ⊂ K, we know that µK = µ ◦ πK −1 , that
µJ = µ ◦ πJ −1 and that πJ = πJK ◦ πK . Hence

−1
µJ = µ ◦ πJK ◦ πK (2.3)
−1

−1
= µ◦ πK ◦ πJK (2.4)

−1
= µK ◦ πJK (2.5)

as required. 
I claimed that the reason to care about finite-dimensional distributions is
that if we specify them, we specify the distribution of the whole process. Lemma
25 says that a putative family of finite dimensional distributions must be consis-
tent, if they are to let us specify a stochastic process. Theorem 23 says that there
can’t be more than one process distribution with all the same finite-dimensional
marginals, but it doesn’t guarantee that a given collection of consistent finite-
dimensional distributions can be extended to a process distribution — it gives
uniqueness but not existence. Proving the existence of an extension requires
some extra assumptions. Either we need to impose topological conditions on Ξ,
CHAPTER 2. BUILDING PROCESSES 18

or we need to ensure that all the finite-dimensional distributions can be related

through conditional probabilities. The first approach is due to Daniell and Kol-
mogorov, and will finish this lecture; the second is due to Ionescu-Tulcea, and
will begin the next.
We’ll start with Daniell’s theorem on the existence of random sequences, i.e.,
where the index set is the natural numbers, which uses mathematical induction
to extend the finite-dimensional family. To get there, we need a useful proposi-
tion about our ability to represent non-trivial random variables as functions of
uniform random variables on the unit interval.

Proposition 26 (Randomization, transfer) Let X and X 0 be identically-

distributed random variables in a measurable space Ξ and Y a random variable
in a Borel space Υ. Then there exists a measurable function f : Ξ × [0, 1] 7→ Υ
such that L (X 0 , f (X 0 , Z)) = L (X, Y ), when Z is uniformly distributed on the
unit interval and independent of X 0 .

Proof: See Kallenberg, Theorem 6.10 (p. 112–113). 

Basically what this says is that if we have two random variables with a
certain joint distribution, we can always represent the pair by a copy of one of
the variables (X), and a transformation of an independent random number. It is
important that Υ be a Borel space here; the result, while very natural-sounding,
does not hold for arbitrary measurable spaces, because the proof relies on having
a regular conditional probability.

Theorem 27 (Daniell Extension Theorem) For Qneach n ∈ N, let Ξn be a

Borel space, and µn be a probability measure on i=1 Ξi . If the µn form a
projective family, then there exist random variables Xi : Ω 7→ Q∞Ξi , i ∈ N, such
that L (X1 , X2 , . . . Xn ) = µn for all n, Q
and a measure µ on i=1 Ξi such that
µn is equal to the projection of µ onto i = 1n Ξi .

Proof: For any fixed n, X1 , X2 , . . . Xn is just a random vector with distri-

bution µn , and we can always construct such an object. The delicate part here is
showing that, when we go to n+1, we can use the same random elements for the
first n coordinates. We’ll do this by using the representation-by-randomization
proposition just introduced, starting with an IID sequence of uniform random
variables on the unit interval, and then transforming them to get a sequence
of variables in the Ξi which have the right joint distribution. (This is like the
quantile transform trick for generating random variates.) The proof will go
inductively, so first we’ll take care of the induction step, and then go back to
reassure ourselves about the starting point.
Induction: Assume we already have X1 , X2 , . . . Xn such that L (X1 , X2 , . . . Xn ) =
µn , and that we have a Zn+1 ∼ U (0, 1) and independent of all the Xi to date.
As remarked, we can always get Y1 , Y2 , . . . Yn+1 such that L (Y1 , Y2 , . . . Yn+1 ) =
µn+1 . Because the µn form a projective family, L (Y1 , Y2 , . . . Yn ) = L (X1 , X2 , . . . Xn ).
Hence, by Proposition 26, there is a measurable f such that, if we set Xn+1 =
f (X1 , X2 , . . . Xn , Zn+1 ), then L (X1 , X2 , . . . Xn , Xn+1 ) = µn+1 .
CHAPTER 2. BUILDING PROCESSES 19

First step: We need there to be an X1 with distribution µ1 , and we need a

(countably!) unlimited supply of IID variables Z2 , Z3 , . . . all ∼ U (0, 1). But the
existence of X1 is just the existence of a random variable with a well-defined
distribution, which is unproblematic, and the existence of an infinite sequence of
IID uniform random variates is too. (See 36-752, or Lemma 3.21 in Kallenberg,
or Exercise 3.1.)
Finally, to convince yourself of the existence of the measure µ on the product
space, recall Theoerem 16. 
Remark: Kallenberg, Corollary 6.15, gives a somewhat more abstract version
of this theorem.
Daniell’s extension theorem works fine for one-sided random sequences, but
we often want to work with larger and more interesting index sets. For this
we need the full Kolmogorov extension theorem, where the index set T can be
completely arbitrary. This in turn needs the Carathéodory Extension Theorem,
which I re-state here for convenience.

Proposition 28 (Carathéodory Extension Theorem) Let µ be a non-

negative, finitely additive set function on a field C of subsets of some space
Ω. If µ is also countably additive, then it extends to a measure on σ(C), and, if
µ(Ω) < ∞, the extension is unique.

Proof: See 36-752 lecture notes (Theorem 50, Exercise 51), or Kallenberg,
Theorem 2.5, pp. 26–27. Note that “extension” here means extending from a
mere field to a σ-field, not from finite to infinite index sets. 

Theorem 29 (Kolmogorov Extension Theorem) Let Ξt , t ∈ T , be a col-

lection of Borel spaces, with σ-fields Xi , and let µJ , J ∈ Fin(T ), be a projective
family of finite-dimensional distributions on those spaces. Then there exist Ξt -
valued random variables Xt such that L (XJ ) = µJ for all J ∈ Fin(T ).

Proof: This will be easier to follow if we first consider the case there T
is countable, which is basically Theorem 27 again, and then the general case,
where we need Proposition 28.
Countable T : We can, by definition, put the elements of T in 1−1 correspon-
dence with theN elements of N. This in turn establishes N a bijection between the
∞
product space t∈T Ξt = ΞT and the sequence space i=1 Ξt . This bijection
also induces a projective family of distributions on finite sequences. The Daniell
Extension Theorem (27) gives us a measure on the sequence space, which the
bijection takes back to a measure on ΞT . To see that this µ does not depend on
the order in which we arranged T , notice that any two arrangements must give
identical results for any finite set J, and then use Theorem 23.
Uncountable T : For each countable K ⊂ T , the argument of the preceding
paragraph gives us a measure µK on ΞK . And, clearly, these µK themselves form
a projective family. Now let’s define a set function µ on the countable cylinder
sets, i.e., on the class D of sets of the form A × ΞT \K , for some K ∈ Denum(T )
and some A ∈ XK . Specifically, µ : D 7→ [0, 1], and µ(A × ΞT \K ) = µK (A).
We would like to use Carathéodory’s theorem to extend this set function to
CHAPTER 2. BUILDING PROCESSES 20

a measure on the product σ-algebra XT . First, let’s check that the countable
cylinder sets form a field: (i) ΞT ∈ D, clearly. (ii) The complement, in ΞT , of
a countable cylinder A × ΞT \K is another countable cylinder, Ac × ΞT \K . (iii)
The union of two countable cylinders B1 = A1 × ΞT \K1 and B2 = A2 × ΞT \K2
is another countable cylinder, since we can always write it as A × ΞT \K for
some A ∈ XK , where K = K1 ∪ K2 . Clearly, µ(∅) = 0, so we just need to
check that µ is countably additive. So consider any sequence of disjoint cylinder
Si × ΞT \Ki , for
sets B1 , B2 , . . .. Because they’re cylinder sets, each i, Bi = A
some Ki ∈ Denum(T ), and some Ai ∈ XKi . Now set K = i Ki ; this is a
countable union of countable sets, and so S itself countable.
S Furthermore, say
Ci = Ai × ΞK\Ki , so we can say that i Bi = ( i Ci ) × ΞT \K . With this
notation in place,
[ [
µ Bi = µK Ci (2.6)
i i
X
= µK Ci (2.7)
i
X
= µKi Ai (2.8)
i
X
= µBi (2.9)
i

where in the second line we’ve used the fact that µK is a probability measure
on ΞK , and so countably additive on sets like the Ci . This proves that µ is
countably additive, so by Proposition 28 it extends to a measure on σ(D), the
σ-field generated by the countable cylinder sets. But we know from Definition
12 that this σ-field is the product σ-field. Since µ(ΞT ) = 1, Proposition 28
further tells us that the extension is unique. 
Borel spaces are good enough for most of the situations we find ourselves
modeling, so the Daniell-Kolmogorov Extension Theorem (as it’s often known)
see a lot of work. Still, some people dislike having to make topological assump-
tions to solve probabilistic problems; it seems inelegant. The Ionescu-Tulcea
Extension Theorem provides a purely probabilistic solution, available if we can
write down the FDDs recursively, in terms of regular conditional probability
distributions, even if the spaces where the process has its coordinates are not
nice and Borel. Doing this properly will involve our revisiting and extending
some ideas about conditional probability, which you will have seen in 36-752, so
it will be deferred to the next lecture.
Chapter 3

Building Infinite Processes

from Regular Conditional
Probability Distributions

Section 3.1 introduces the notion of a probability kernel, which

is a useful way of systematizing and extending the treatment of
conditional probability distributions you will have seen in 36-752.
Section 3.2 gives an extension theorem (due to Ionescu Tulcea)
which lets us build infinite-dimensional distributions from a family
of finite-dimensional distributions. Rather than assuming topolog-
ical regularity of the space, as in Section 2.2, we assume that the
FDDs can be derived from one another recursively, through applying
probability kernels. This is the same as assuming regularity of the
appropriate conditional probabilities.

3.1 Probability Kernels

Definition 30 (Probability Kernel) A measure kernel from a measurable
+
space Ξ, X to another measurable space Υ, Y is a function κ : Ξ × Y 7→ R such
that
1. for any Y ∈ Y, κ(x, Y ) is X -measurable; and
2. for any x ∈ Ξ, κ(x, Y ) ≡ κx (Y ) is a measure on Υ, Y. We will write
Rthe integral of aR function f : Υ 7→ R, with respect to this measure, as
f (y)κ(x, dy), f (y)κx (dy), or, most compactly, κf (x).

If, in addition, κx is a probability measure on Υ, Y for all x, then κ is a prob-

ability kernel.

21
CHAPTER 3. BUILDING PROCESSES BY CONDITIONING 22

(Some authors use “kernel”, unmodified, to mean a probability kernel, and

some a measure kernel; read carefully.)
Notice that we can represent any distribution on Υ as a kernel where the first
argument is irrelevant: κ(x1 , Y ) = κ(x2 , Y ) for all x1 , x2 ∈ Ξ. The “kernels” in
kernel density estimation are probability kernels, as are the stochastic transition
matrices of Markov chains. (The kernels in support vector machines, however,
generally are not.)
From a previous probability course, like 36-752, you will remember the
measure-theoretic definition of the conditional probability of a set A, as the
conditional expectation of 1A , i.e., P (A|G) = E [1A (ω)|G]. (I have explicitly
included the dependence on ω to make it plain that this is a random set func-
tion.) You may also recall a potential bit of trouble, which is that even when
these conditional expectations exist for all measurable
S∞ sets A, P
it’s not necessarily
∞
true that they give us a measure, i.e., that P ( i=1 Ai |G) = i=1 P (Ai |G) and
all the rest of it. A version of the conditional probabilities for which they do
form a measure is said to be a regular conditional probability. Clearly, regular
conditional probabilities are all probability kernels. The ordinary rules for ma-
nipulating conditional probabilities suggest how we can define the composition
of kernels.
Definition 31 (Composition of probability kernels) Let κ1 be a kernel
from Ξ to Υ, and κ2 a kernel from Ξ × Υ to Γ. Then we define κ1 ⊗ κ2 as the
kernel from Ξ to Υ × Γ such that
Z Z
(κ1 ⊗ κ2 )(x, B) = κ1 (x, dy) κ2 (x, y, dz)1B (y, z)

for every measurable B ⊆ Υ × Γ (where z ranges over the space Γ).

Verbally, κ1 gives us a distribution on Υ, from any starting point x ∈ Ξ.
Given a pair of points (x, y) ∈ Ξ × Υ, κ2 gives a distribution on Γ. So their
composition says, basically, how to chain together conditional distributions,
given a starting point.

3.2 Extension via Recursive Conditioning

With the machinery of probability kernels in place, we are in a position to give
an alternative extension theorem, i.e., a different way of proving the existence of
stochastic processes with specified finite-dimensional marginal distributions. In
Section 2.2, we assumed some topological niceness in the sample spaces, namely
that they were Borel spaces. Here, instead, we will assume probabilistic niceness
in the FDDs themselves, namely that they can be obtained through composing
probability kernels. This is the same as assuming that they can be obtained
by chaining together regular conditional probabilities. The general form of this
result is attributed in the literature to Ionescu Tulcea.
Just as proving the Kolmogorov Extension Theorem needed a measure-
theoretic result, the Carathéodory Extension Theorem, our proof of the Ionescu
CHAPTER 3. BUILDING PROCESSES BY CONDITIONING 23

Tulcea Extension Theorem will require a different measure-theoretic result,

which is not, so far as I know, named after anyone. (It re-appears in the theory
of Markov operators, beginning in Section 10.2.)

Proposition 32 (Set functions continuous at ∅) Suppose µ is a finite, non-

negative, additive set function on a field A. If, for any sequence of sets An ∈ A,
An ↓ ∅ =⇒ µAn → 0, then
1. µ is countably additive on A, and

2. µ extends uniquely to a measure on σ(A).

Proof: Part (1) is a weaker version of Theorem F in Chapter 2, §9 of

Halmos (1950, p. 39). (When reading his proof, remember that every field of
sets is also a ring of sets.) Part (2) follows from part (1) and the Carathéodory
Extension Theorem (Proposition 28). 
With this preliminary out of the way, let’s turn to the main event.

Theorem 33 (Ionescu Tulcea Extension Theorem) Consider a sequence

of measurable spaces Ξn , XQ n , n ∈ N. Suppose that for each n, there exists a
n−1
probability kernel κn from i=1 Ξi to Ξn (taking κ1 to be a kernel insensitive
to its first argument, i.e., a probability measure). Then there exists a sequence
of random variables Xn , N n ∈ N, taking values in the corresponding Ξn , such
n
that L (X1 , X2 , . . . Xn ) = i=1 κi .

Proof: As before, we’ll be working with the cylinder sets, but now we’ll
make our life simpler if we consider cylinders where Nnthe base set rests in the
first n spaces Ξ1 , ...Ξn .QMore specifically, set Bn = i=1 Xi (these are the
S base
∞
sets), and Cn = Bn × i=n+1 Ξi (these are the cylinder sets), and C = n Cn .
C clearly contains all the finite cylinders, so it generates the product σ-field on
infinite sequences. We will use it as the field in Proposition 32. (Checking that
C is a field is entirely parallel to checking that the D appearing in the proof of
Theorem 29 was a field.)
Q∞each base set A ∈ Bn , let [A] be the corresponding cylinder, [A] =
For
A × i=n+1 Ξi . Notice that for every set C ∈ C, there is at least one A, in some
Bn , such that C = [A]. Now we define a set function µ on C.
n
!
O
µ([A]) = κi A (3.1)
i=1

(Checking that this is well-defined is left as an exercise, 3.2.) Clearly, this is a

finite, and finitely-additive, set function defined on a field. So to use Proposition
32, we just need to check continuity from above at ∅. Let An be any sequence
of sets such that [An ] ↓ ∅ and An ∈ Bn . (Any sequence of sets in C ↓ ∅ can be
massaged into this form.) We wish to show that µ([An ]) ↓ 0. We’ll get this to
CHAPTER 3. BUILDING PROCESSES BY CONDITIONING 24

work by considering functions which are (pretty much) conditional probabilities

for these sets:
n
!
O
pn|k = κi 1An , k ≤ n (3.2)
i=k+1
pn|n = 1An (3.3)
Two facts follow immediately from the definitions:
n
!
O
pn|0 = κi 1An = µ([An ]) (3.4)
i=1
pn|k = κk+1 pn|k+1 (3.5)
From the fact that the [An ] ↓ ∅, we know that pn+1|k ≤ pn|k , for all k. This
implies that limn pn|k = mk exists, for each k, and is approached from above.
Applied to pn|0 , we see from 3.5 that µ([An ]) → m0 . We would like m0 = 0.
Assume the contrary, that m0 > 0. From 3.5 and the dominated convergence
theorem, we can see that mk = κk+1 mk+1 . Hence if m0 > 0, κ1 m1 > 0, which
means (since that last expression is really an integral) that there is at least one
point x1 ∈ Ξ1 such that m1 (s1 ) > 0. Recursing our way down the line, we get
a sequence x = x1 , x2 , . . . ∈ ΞN such that mn (x1 , . . . xn ) > 0 for all n. But now
look what we’ve done: for each n,
0 < mn (x1 , . . . xn ) (3.6)
≤ pn|n (x1 , . . . xn ) (3.7)
= 1An (x1 , . . . xn ) (3.8)
= 1[An ] (x) (3.9)
x ∈ [An ] (3.10)
T
This is the same as saying that x ∈ n [An ]. But [An ] ↓ ∅, so there can be no
such x. Hence m0 = 0, meaning that µ([An ]) → 0, and µ is continuous at the
empty set.
Since µ is finite, finitely-additive, non-negative and continuous at ∅, by
Proposition 32 it extends uniquely to a measure on the product σ-field. 
Notes on the proof: It would seem natural that one could show m0 = 0
directly, rather than by contradiction, but I can’t think of a way to do it, and
every book I’ve consulted does it in exactly this way.
To appreciate the simplification made possible by the notion of probability
kernels, compare this proof to the one given by Fristedt and Gray (1997, §22.1).
Notice that the Daniell, Kolmogorov and Ionescu Tulcea Extension Theo-
rems all give sufficient conditions for the existence of stochastic processes, not
necessary ones. The necessary and sufficient condition for extending the FDDs
to a process probability measure is something called σ-smoothness. (See Pollard
(2002) for details.) Generally speaking, we will deal with processes which satisfy
both the Kolmogorov and the Ionescu Tulcea type conditions, e.g., real-valued
Markov process.
CHAPTER 3. BUILDING PROCESSES BY CONDITIONING 25

3.3 Exercises
Exercise 3.1 (Lomnick-Ulam Theorem on infinite product measures)
Let T be an uncountable index set, and (Ξt , Xt , µt ) a collection of probability
spaces. Show that there exist independent random variables Xt in Ξt with dis-
tributions µt . Hint: use the Ionescu Tulcea theorem on countable subsets of T ,
and then imitate the proof of the Kolmogorov extension theorem.

Exercise 3.2 (Measures of cylinder sets) In the proof of the Ionescu Tulcea
Theorem, we employed a set function on the finite cylinder sets, where the mea-
sure of an infinite-dimensional cylinder set [A] is taken to be the measure of its
finite-dimensional base set A. However, the same cylinder set can be specified
by different base sets, so it is necessary to show that Equation 3.1 has a unique
value on its right-hand side. In what follows, C is an arbitrary member of the
class C.
1. Show that, when A, B ∈ Bn , [A] = [B] iff A = B. That is, two cylinders
generated by bases of equal dimensionality are equal iff their bases are
equal.
2. Show that there is a smallest n such that C = [A] for an A ∈ Bn . Conclude
that the right-hand side of Equation 3.1 could be made well-defined if we
took n there to be this least possible n.

Qn m < n, A ∈ Bm , B ∈ Bn , and [A] = [B]. Show that

3. Suppose that
B = A × i=m+1 Ξi .

4. Continuing the situation in (iii), show that

m
! n
!
O O
κi A = κi B
i=1 i=1

Conclude that the right-hand side of Equation 3.1 is well-defined, as promised.

 Part II

One-Parameter Processes in
General

26
Chapter 4

One-Parameter Processes,
Usually Functions of Time

Section 4.1 defines one-parameter processes, and their variations

(discrete or continuous parameter, one- or two- sided parameter),
including many examples.
Section 4.2 shows how to represent one-parameter processes in
terms of “shift” operators.

We’ve been doing a lot of pretty abstract stuff, but the point of this is to
establish a common set of tools we can use across many different concrete situa-
tions, rather than having to build very similar, specialized tools for each distinct
case. Today we’re going to go over some examples of the kind of situation our
tools are supposed to let us handle, and begin to see how they let us do so. In
particular, the two classic areas of application for stochastic processes are dy-
namics (systems changing over time) and inference (conclusions changing as we
acquire more and more data). Both of these can be treated as “one-parameter”
processes, where the parameter is time in the first case and sample size in the
second.

4.1 One-Parameter Processes

The index set T isn’t, usually, an amorphous abstract set, but generally some-
thing with some kind of topological or geometrical structure. The number of
(topological) dimensions of this structure is the number of parameters of the
process.

Definition 34 (One-Parameter Process) A process whose index set T has

one dimension is a one-parameter process. A process whose index set has more
than one dimension is a multi-parameter process. A one-parameter process is

27
CHAPTER 4. ONE-PARAMETER PROCESSES 28

discrete or continuous depending on whether its index set is countable or un-

countable. A one-parameter process where the index set has a minimal element
it is one-sided, otherwise it is two-sided.

N is a one-sided discrete index set, Z a two-sided discrete index set, R+ (in-

cluding zero!) is a one-sided continuous index set, and R a two-sided continuous
index set.
Most of this course will be concerned with one-parameter processes, which
are intensely important in applications. This is because the one-dimensional
parameter is usually either time (when we’re doing dynamics) or sample size
(when we’re doing inference), or both at once. There are also some important
cases where the single parameter is space.

Example 35 (Bernoulli process) You all know this one: a one-sided infinite
sequence of independent, identically-distributed binary variables, where, for all
t, Xt = 1 with probability p.

Example 36 (Markov models) Markov chains are discrete-parameter stochas-

tic processes. They may be either one-sided or two-sided. So are Markov models
of order k, and hidden Markov models. Continuous-time Markov processes are,
naturally enough, continuous-parameter stochastic processes, and again may be
either one-sided or two-sided.

Instances of physical processes that may be represented by Markov models

include: the positions and velocities of the planets; the positions and velocities
of molecules in a gas; the pressure, temperature and volume of the gas (Keizer,
1987); the position and velocity of a tracer particle in a turbulent fluid flow; the
three-dimensional velocity field of a turbulent fluid (Frisch, 1995); the gene pool
of an evolving population (Fisher, 1958; Haldane, 1932; Gillespie, 1998). In-
stances of physical processes that may be represented by hidden Markov models
include: the spike trains of neurons (Tuckwell, 1989); the sonic wave-forms of
human speech (Rabiner, 1989); many economic and social time-series (Durbin
and Koopman, 2001); etc.

Example 37 (“White Noise” (Not Really)) For each t ∈ R+ , let Xt ∼

N (0, 1), all mutually independent of one another. This is a process with a one-
sided continuous parameter.

It would be character building, at this point, to convince yourself that the

process just described exists. (You will need the Kolmogorov Extension Theo-
rem, 29).

Example 38 (Wiener Process) Here T = R+ and Ξ = R. The Wiener

process is the continuous-parameter random process where
1. W (0) = 0,
CHAPTER 4. ONE-PARAMETER PROCESSES 29

2. for any three times, t1 < t2 < t3 , W (t2 ) − W (t1 ) and W (t3 ) − W (t2 ) are
independent (the “independent increments” property),
3. W (t2 ) − W (t1 ) ∼ N (0, t2 − t1 ) and
4. W (t, ω) is a continuous function of t for almost all ω.
We will spend a lot of time with the Wiener process, because it turns out to
play a role in the theory of stochastic processes analogous to that played by
the Gaussian distribution in elementary probability — the easily-manipulated,
formally-nice distribution delivered by limit theorems.

When we examine the Wiener process in more detail, we will see that it
almost never has a derivative. Nonetheless, in a sense which will be made
clearer when we come to stochastic calculus, the Wiener process can be regarded
as the integral over time of something very like white noise, as described in the
preceding example.

Example 39 (Logistic Map) Let T = N, Ξ = [0, 1], X(0) ∼ U (0, 1), and
X(t + 1) = aX(t)(1 − X(t)), a ∈ [0, 4]. This is called the logistic map. Notice
that all the randomness is in the initial value X(0); given the initial condition,
all later values X(t) are fixed. Nonetheless, this is a Markov process, and we will
see that, at least for certain values of a, it satisfies versions of the laws of large
numbers and the central limit theorem. In fact, large classes of deterministic
dynamical systems have such stochastic properties.

Example 40 (Symbolic Dynamics of the Logistic Map) Let X(t) be the

logistic map, as in the previous example, and let S(t) = 0 if X(t) ∈ [0, 0.5) and
S(t) = 1 if X(t) = [0.5, 1]. That is, we partition the state space of the logistic
map, and record which cell of the partition the original process finds itself in.
X(t) is a Markov process, but these “symbolic” dynamics are not necessarily
Markovian. We will want to know when functions of Markov processes are
themselves Markov. We will also see that there is a sense in which, Markovian
or not, this partition is exactly as informative as the original, continuous state
— that it is generating. Finally, when a = 4 in the logistic map, the symbol
sequence is actually a Bernoulli process, so that a deterministic function of a
completely deterministic dynamical system provides a model of IID randomness.

Here are some examples where the parameter is sample size.

Example 41 (IID PSamples) Let Xi , i ∈ N be samples from an IID distribu-

n
tion, and Zn = n1 i=1 Xi be the sample mean. Then Zn is a one-parameter
stochastic process. The point of the ordinary law of large numbers is to reassure
√ Zn → E [Xn ] a.s. The point of the central limit theorem is to reassure us
us that
that n(Zn − E [X]) has constant√average size, so that the sampling fluctuation
Zn − E [X] must be shrinking as n grows.
If Xi is the indicator of a set, this convergence means that the relative fre-
quency with which the set is occupied will converge on its true probability.
CHAPTER 4. ONE-PARAMETER PROCESSES 30

Example 42 (Non-IID Samples) Let Xi be a non-IID one-sided discrete-

parameter process, say a Markov chain, and again let Zn be its sample mean,
now often called its “time average”. The usual machinery of the law of large
numbers and the central limit theorem are now inapplicable, and someone who
has just taken 36-752 has, strictly speaking, no idea as to whether or not their
time averages will converge on expectations. Under the heading of ergodic theory,
we will see when this will happen. Since this is the situation with all interesting
time series, the application of statistical methods to situations where we cannot
contrive to randomize depends crucially on ergodic considerations.

Example 43 (Estimating Distributions) Recall Example 10, where we looked

at the sequence of empirical distributions P̂n for samples from an IID data-
source. We would like to be able to say that P̂n converges on P . The usual way
to do this, if our samples are of a real-valued random variable, is to consider the
empirical cumulative distribution function, Fn . For each n, this may be regarded
as a one-parameter random process (T = R, Ξ = [0, 1]), and the difficulty is to
show that this sequence
√ of random processes converges to F . The usual way is to
show that En ≡ n(Fn − F ), the empirical process, converges to a fixed process,
related to the Wiener process, depending only on the true distribution F . (See √
Figure 4.1.) Since the magnitude of the limiting process is constant, and n
grows, it follows that Fn − F must shrink. So theorizing even this elementary
bit of statistical inference really requires two doses of stochastic process theory,
one to get a grip on Fn at each n, and the other to get a grip on what happens
to Fn as n grows.

Example 44 (Doob’s Martingale) Let X be a random variable, and Fi ,

i ∈ N, a sequence of increasing σ-algebras (i.e. a filtration). Then Yi = E [X|Fi ]
is a one-sided discrete-parameter stochastic process, and in fact a martingale.
Martingales in general are one-parameter stochastic processes. Note that poste-
rior mean parameter estimates, in Bayesian inference, are instances of Doob’s
martingale.
Here are some examples where the one-dimensional parameter is neither time
nor sample size.
Example 45 (The One-Dimensional Ising Model) This system serves as
a toy model of magnetism in theoretical physics. Atoms sit evenly spaced on the
points of a regular, infinite, one-dimensional crystalline lattice. Each atom has
a magnetic moment, which is either pointing north (+1) or south (−1). Atoms
are more likely to point north if their neighbors point north, and vice-versa. The
natural index here is Z, so the parameter is discrete and two-sided.

Example 46 (Text) Text (at least in most writing systems!) can be repre-
sented by a sequence of discrete values at discrete, ordered locations. Since texts
can be arbitrarily long, but they all start somewhere, they are discrete-parameter,
one-sided processes. Or, more exactly, once we specify a distribution over se-
quences from the appropriate alphabet, we will have such a process.
CHAPTER 4. ONE-PARAMETER PROCESSES 31

Example 47 (Polymer Sequences) Similarly, DNA, RNA and proteins are

all heteropolymers — compounds in which distinct constituent chemicals (the
monomers) are joined in a sequence. Position along the sequence (chromosome,
protein) provides the index, and the nature of the monomer at that position the
value.

Linguists believe that no Markovian model (with finitely many states) can
capture human language (Chomsky, 1957; Pullum, 1991). Whether this is true
of DNA sequences is not known. In both cases, hidden Markov models are used
extensively Baldi and Brunak (2001); Charniak (1993), even if they can only be
approximately true of language.

4.2 Operator Representations of One-Parameter

Processes
Consider our favorite discrete-parameter process, say Xt . If we try to relate Xt
to its history, i.e., to the preceding values from the process, we will often get a
horribly complicated probabilistic expression. There is however an alternative,
which represents the dynamical part of any process as a remarkably simple
semi-group of operators.

Definition 48 (Shift Operators) Consider ΞT , T = N, = Z, = R+ or = R.

The shift-by-τ operator Στ , τ ≥ 0, maps ΞT into itself by shifting forward in
time: (Στ )x(t) = x(t + τ ). The collection of all shift operators is the shift
semi-group or time-evolution semi-group.

(A semi-group does not need to have an identity element, and one which
does is technically called a “monoid”. No one talks about the shift monoid or
time-evolution monoid, however.)
Before we had a Ξ-valued stochastic process X on T , i.e., our process was
a random function from T to Ξ. To extract individual random variables, we
used the projection operators πt , which took X to Xt . With the shift operators,
we simply have πt = π0 ◦ Σt . To represent the passage of time, then, we just
apply elements of this semi-group to the function space. Rather than having
complicated dynamics which gets us from one value to the next, by working with
shifts on function space, all of the complexity is shifted to the initial distribution.
This will prove to be extremely useful when we consider stationary processes in
the next lecture, and even more useful when, later on, we want to extend the
limit theorems from IID sequences to dependent processes.

4.3 Exercises
Exercise 4.1 (Existence of proto-Wiener processes) Use Theorem 29 and
the properties of Gaussian distributions to show that processes exist which satisfy
CHAPTER 4. ONE-PARAMETER PROCESSES 32

points (1)–(3) of Example 38 (but not necessarily continuity). You will want to
begin by finding a way to write down the FDDs recursively.

Exercise 4.2 (Time-Evolution Semi-Group) These are all very easy, but
worth the practice.

1. Verify that the time-evolution semi-group, as described, is a monoid, i.e.,

that it is closed under composition, that composition is associative, and
that there is an identity element. What, in fact, is the identity?
2. Can a one-sided process have a shift group, rather than just a semi-group?

3. Verify that πτ = π0 ◦ Στ .
t
4. Verify that, for a discrete-parameter process, Σt = (Σ1 ) , and so Σ1 gen-
erates the semi-group. (For this reason it is often abbreviated to Σ.)
CHAPTER 4. ONE-PARAMETER PROCESSES 33

Empirical process for 102 samples from a log−normal Empirical process for 103 samples from a log−normal

0.8

0.8
0.6

0.6
0.4

0.4
0.2
E102(x)

E103(x)

0.2
−0.2 0.0

0.0
−0.4
−0.6

0 5 10 15 20 0 5 10 15 20

x x

Empirical process for 104 samples from a log−normal Empirical process for 105 samples from a log−normal

0.8
0.4

0.6
0.2
E104(x)

E105(x)

0.4
0.0

0.2
−0.2

0.0
−0.4

−0.2

0 5 10 15 20 0 5 10 15 20

x x

Empirical process for 106 samples from a log−normal Empirical process for 106 samples from a log−normal
1.0

0.350
0.5
E106(x)

E106(x)

0.340
0.0

0.330
−0.5

0 5 10 15 20 5.000 5.002 5.004 5.006 5.008 5.010

x x

Figure 4.1: Empirical process for successively larger samples from a log-normal
distribution, mean of the log = 0.66, standard deviation of the log = 0.66. The
samples were formed by taking the first n variates from an over-all sample of
1,000,000 variates. The bottom right figure magnifies a small part of E106 to
show how spiky the paths are.
Chapter 5

Stationary One-Parameter
Processes

Section 5.1 describes the three main kinds of stationarity: strong,

weak, and conditional.
Section 5.2 relates stationary processes to the shift operators in-
troduced in the last chapter, and to measure-preserving transforma-
tions more generally.

5.1 Kinds of Stationarity

Stationary processes are those which are, in some sense, the same at different
times — slightly more formally, which are invariant under translation in time.
There are three particularly important forms of stationarity: strong or strict,
weak, and conditional.

Definition 49 (Strong Stationarity) A one-parameter process is strongly

stationary or strictly stationary when all its finite-dimensional distributions are
invariant under translation of the indices. That is, for all τ ∈ T , and all
J ∈ Fin(T ),
L (XJ ) = L (XJ+τ ) (5.1)

Notice that when the parameter is discrete, we can get away with just check-
ing the distributions of blocks of consecutive indices.

Definition 50 (Weak Stationarity) A one-parameter process is weakly sta-

tionary or second-order stationary when, for all t ∈ T ,

E [Xt ] = E [X0 ] (5.2)

and for all t, τ ∈ T ,

E [Xτ Xτ +t ] = E [X0 Xt ] (5.3)

34
CHAPTER 5. STATIONARY PROCESSES 35

At this point, you should check that a weakly stationary process has time-
invariant correlations. (We will say much more about this later.) You should
also check that strong stationarity implies weak stationarity. It will turn out
that weak and strong stationarity coincide for Gaussian processes, but not in
general.
Definition 51 (Conditional (Strong) Stationarity) A one-parameter pro-
cess is conditionally stationary if its conditional distributions are invariant un-
der time-translation: ∀n ∈ N, for every set of n + 1 indices t1 , . . . tn+1 ∈ T ,
ti < ti+1 , and every shift τ ,



L Xtn+1 |Xt1 , Xt2 . . . Xtn = L Xtn+1 +τ |Xt1 +τ , Xt2 +τ . . . Xtn +τ (5.4)
(a.s.).
Strict stationarity implies conditional stationarity, but the converse is not
true, in general. (Homogeneous Markov processes, for instance, are all con-
ditionally stationary, but most are not stationary.) Many methods which are
normally presented using strong stationarity can be adapted to processes which
are merely conditionally stationary.1
Strong stationarity will play an important role in what follows, because it
is the natural generaliation of the IID assumption to situations with dependent
variables — we allow for dependence, but the probabilistic set-up remains, in a
sense, unchanging. This will turn out to be enough to let us learn a great deal
about the process from observation, just as in the IID case.

5.2 Strictly Stationary Processes and Measure-

Preserving Transformations
The shift-operator representation of Section 4.2 is particularly useful for strongly
stationary processes.
Theorem 52 (Stationarity is Shift-Invariance) A process X with measure
µ is strongly stationary if and only if µ is shift-invariant, i.e., µ = µ ◦ Σ−1
τ for
all Στ in the time-evolution semi-group.
Proof: “If” (invariant distributions imply stationarity): For any finite col-
lection of indices J, L (XJ ) = µ ◦ πJ−1 (Lemma 25), and similarly L (XJ+τ ) =
−1
µ ◦ πJ+τ .
πJ+τ = πJ ◦ Σ τ (5.5)
−1 −1
πJ+τ = Σ−1
τ ◦ πJ (5.6)
−1 −1
µ ◦ πJ+τ = µ ◦ Σ−1
τ ◦ πJ (5.7)
L (XJ+τ ) = µ◦ πJ−1 (5.8)
= L (XJ ) (5.9)
1 For more on conditional stationarity, see Caires and Ferreira (2005).
CHAPTER 5. STATIONARY PROCESSES 36

“Only if”: The statement that µ = µ ◦ Σ−1 τ really means that, for any
set A ∈ X T , µ(A) = µ(Σ−1 τ A). Suppose A is a finite-dimensional cylinder
set. Then the equality holds, because all the finite-dimensional distributions
agree (by hypothesis). But this means that X and Στ X are two processes
with the same finite-dimensional distributions, and so their infinite-dimensional
distributions agree (Theorem 23), and the equality holds on all measurable sets
A. 
This can be generalized somewhat.

Definition 53 (Measure-Preserving Transformation) A measurable map-

ping F from a measurable space Ξ, X into itself preserves measure µ iff, ∀A ∈ X ,
d
µ(A) = µ(F −1 A), i.e., iff µ = µ ◦ F −1 . This is true just when F (X) = X, when
X is a Ξ-valued random variable with distribution µ. We will often say that
F is measure-preserving, without qualification, when the context makes it clear
which measure is meant.

Remark on the definition. It is natural to wonder why we write the defining

property as µ = µ ◦ F −1 , rather than µ = µ ◦ F . There is actually a subtle
difference, and the former is stronger than the latter. To see this, unpack the
statements, yielding respectively

∀A ∈ X , µ(A) = µ(F −1 (A)) (5.10)

∀A ∈ X , µ(A) = µ(F (A)) (5.11)

To see that Eq. 5.10 implies Eq. 5.11, pick any measurable set B, and then
apply 5.10 to F (B) (which is ∈ X , because F is measurable). To go the other
way, from 5.11 to 5.10, it would have to be the case that, ∀A ∈ X , ∃B ∈ X such
that A = F (B), i.e., every measurable set would have to be the image, under
F , of another measurable set. This is not necessarily the case; it would require,
for starters, that F be onto (surjective).
Theorem 52 says that every stationary process can be represented by a
measure-preserving transformation, namely the shift. Since measure-preserving
transformations arise in many other ways, however, it is useful to know about
the processes they generate.

Corollary 54 (Measure-preservation implies stationarity) If F is a measure-

preserving transformation on Ξ with invariant measure µ, and X is a Ξ-valued
random variable, L (X) = µ, then the sequence F n (X), n ∈ N is strongly sta-
tionary.

Proof: Consider shifting the sequence F n (X) by one: the nth term in the
shifted sequence is F n+1
(X) = F n (F (X)). But since L (F (X)) = L (X), by
hypothesis, L F n+1 (X) = L (F n (X)), and the measure is shift-invariant. So,
by Theorem 52, the process F n (X) is stationary.
CHAPTER 5. STATIONARY PROCESSES 37

5.3 Exercises
Exercise 5.1 (Functions of Stationary Processes) Use Corollary 54 to
show that if g is any measurable function on Ξ, then the sequence g(F n (X)) is
also stationary.

Exercise 5.2 (Continuous Measure-Preserving Families of Transfor-

mations) Let Ft , t ∈ R+ , be a semi-group of measure-preserving transforma-
tions, with F0 being the identity. Prove the analog of Corollary 54, i.e., that
Ft (X), t ∈ R+ , is a stationary process.

Exercise 5.3 (The Logistic Map as a Measure-Preserving Transforma-

tion) The logistic map with a = 4 is a measure-preserving
p transformation, and
the measure it preserves has the density 1/π x(1 − x) (on the unit interval).
1. Verify that this density is invariant under the action of the logistic map.
2. Simulate the logistic map with uniformly distributed X0 . What happens to
the density of Xt as t → ∞?
Chapter 6

Random Times and Their

Properties

Section 6.1 recalls the definition of a filtration (a growing col-

lection of σ-fields) and of “stopping times” (basically, measurable
random times).
Section 6.2 defines various sort of “waiting” times, including hit-
ting, first-passage, and return or recurrence times.
Section 6.3 proves the Kac recurrence theorem, which relates the
finite-dimensional distributions of a stationary process to its mean
recurrence times.

6.1 Reminders about Filtrations and Stopping

Times
You will have seen these in 36-752 as part of martingale theory, though their
application is more general, as we’ll see.

Definition 55 (Filtration) Let T be an ordered index set. A collection Ft , t ∈

T of σ-algebras is a filtration (with respect to this order) if it is non-decreasing,
i.e., f ∈ Ft implies f ∈ Fs forT all s > t. We generally abbreviate this filtration
by {Ft }. Define {F}t+ as s>t Fs . If {F}t+ = {Ft }, then {Ft } is right-
continuous.

Recall that we generally think of a σ-algebra as representing available infor-

mation — for any event f ∈ F, we can answer the question “did f happen?”
A filtration is a way of representing our information about a system growing
over time. ToT see what right-continuity is about, imagine it failed, which would
mean Ft ⊂ s>t Fs . Then there would have to be events which were detectable
at all times after t, but not at t itself, i.e., some sudden jump in our information
right after t. This is what right-continuity rules out.

38
CHAPTER 6. RANDOM TIMES 39

Definition 56 (Adapted Process) A stochastic process X on T is adapted

to a filtration {Ft } if ∀t, Xt is Ft -measurable. Any process is adapted to the
filtration it induces, σ {Xs : s ≤ t}. This natural filtration is written FtX .
A process being adapted to a filtration just means that, at every time, the
filtration gives us enough information to find the value of the process.
Definition 57 (Stopping Time, Optional Time) An optional time or a
stopping time, with respect to a filtration {Ft }, is a T -valued random variable
τ such that, for all t,
{ω ∈ Ω : τ (ω) ≤ t} ∈ Ft (6.1)
If Eq. 6.1 holds with < instead of ≤, then τ is weakly optional or a weak
stopping time.
Basically, all we’re doing here is defining what we mean by “a random time
at which something detectable happens”. That being the case, it is natural to
ask what information we have when that detectable thing happens.
Definition 58 (Fτ for a Stopping Time τ ) If τ is a {Ft } stopping time,
then the σ-algebra Fτ is given by
Fτ ≡ {A ∈ F : ∀t, A ∩ {ω : τ (ω) ≤ t} ∈ Ft } (6.2)
I admit that the definition of Fτ looks bizarre, and I won’t blame you if you
have to read it a few times to convince yourself it isn’t circular. Here is a simple

case where it makes sense. Let X be a one-sided process, and τ a discrete FtX
stopping time. Then
FτX = σ (X(t ∧ τ ) : t ≥ 0) (6.3)
That is, FτX is everything we know from observing X up to time τ . (This is
Exercise 6.2.) The convoluted-looking definition of Fτ carries this idea over
to the more general situation where τ is continuous and we don’t necessarily
have a single variable generating the filtration. A filtration lets us tell whether
some event A happened by the random time τ if simultaneously gives us enough
information to notice τ and A.
The process Y (t) = X(t ∧ τ ) is follows along with X up until τ , at which
point it becomes fixed in place. It is accordingly called an arrested, halted or
stopped version of the process. This seems to be the origin of the name “stopping
time”.

6.2 Waiting Times

“Waiting times” are particular kinds of optional kinds: how much time must
elapse before a given event happens, either from a particular starting point,
or averaging over all trajectories? Often, these are of particular interest in
themselves, and some of them can be related to other quantities of interest.
CHAPTER 6. RANDOM TIMES 40

Definition 59 (Hitting Time) Given a one-sided Ξ-valued process X, the

hitting time τB of a measurable set B ⊂ Ξ is the first time at which X(t) ∈ B;

τB = inf {t > 0 : Xt ∈ B} (6.4)

Example 60 (Fixation through Genetic Drift) Consider the variation in

a given locus (roughly, gene) in an evolving population. If there are k different
versions of the gene (“alleles”), the state of the population
Pcan be represented by
a vector X(t) ∈ Rk , where at each time Xi (t) ≥ 0 and i Xi (t) = 1. This set
is known as the k-dimensional probability simplex Sk . We say that a certain
allele has been fixed in the population or gone to fixation at t if Xi (t) = 1 for
some i, meaning that all members of the population have that version of the
gene. Fixation corresponds to X(t) ∈ V , where V consists of the vertices of
the simplex. An important question in evolutionary theory is how long it takes
the population to go to fixation. By comparing the actual rate of fixation to
that expected under a model of adaptively-neutral genetic drift, it is possible to
establish that some genes are under the influence of natural selection.

Gillespie (1998) is a nice introduction to population genetics, including this

problem among many others, using only elementary probability. More sophis-
ticated models treat populations as measure-valued stochastic processes.

Example 61 (Stock Options) A stock option1 is a legal instrument giving

the holder the right to buy a stock at a specified price (the strike price, c) before
a certain expiration date te . The point of the option is that, if you exercise it at
a time t when the price of the stock p(t) is above c, you can turn around and sell
the stock to someone else, making a profit of p(t) − c. When p(t) > c, the option
is said to be in money or above water. Options can themselves be sold, and the
value of an option depends on how much money it could be used to make, which
in turn depends on the probability that it will be “in money” before time te . An
important part of mathematical finance thus consists of problems of the form
“assuming prices p(t) follow a process distribution µ, what is the distribution of
hitting times of the set p(t) > c?”

While the financial industry is a major consumer of stochastics, and it has

a legitimate role to play in capitalist society, I do hope you will find something
more interesting to do with your new-found mastery of random processes, so I
will not give many examples of this sort. If you want much, much more, read
Shiryaev (1999).

Definition 62 (First Passage Time) When Ξ = R or Z, we call the hitting

time of the origin the time of first passage through the origin, and similarly for
other points.

1 Actually, this is just one variety of option (an “American call”), out of a huge variety. I

will not go into details.

CHAPTER 6. RANDOM TIMES 41

Definition 63 (Return Time, Recurrence Time) Fix a set B ∈ Ξ. Sup-

pose that X(t0 ) ∈ B. Then the return time or first return time of B is recur-
rence time of B is inf {t > t0 : X(t) ∈ B}, and the recurrence time θB is the
difference between the first return time and t0 .

Note 1: If I’m to be honest with you, I should admit that “return time”
and “recurrence time” are used more or less interchangeably in the literature to
refer to either the time coordinate of the first return (what I’m calling the return
time) or the time interval which elapses before that return (what I’m calling
the recurrence time). I will try to keep these straight here. Check definitions
carefully when reading papers!
Note 2: Observe that if we have a discrete-parameter process, and are in-
terested in recurrences of a finite-length sequence of observations w ∈ Ξk , we
can handle this situation by the device of working with the shift operator in
sequence space.
The question of whether any of these waiting times is optional (i.e., mea-
surable) must, sadly, be raised. The following result is generally enough for our
purposes.

Proposition 64 (Some Sufficient Conditions for Waiting Times to be

Weakly Optional) Let X be a Ξ-valued process on a one-sided parameter T ,
adapted to a filtration {Ft }, and let B be an arbitrary measurable set in Ξ. Then
τB is weakly {Ft }-optional under any of the following (sufficient) conditions,
and {Ft }-optional under the first two:
1. T is discrete.
2. T is R+ , Ξ is a metric space, B is closed, and X(t) is a continuous
function of t.
3. T is R+ , Ξ is a topological space, B is open, and X(t) is right-continuous
as a function of t.

Proof: See, for instance, Kallenberg, Lemma 7.6, p. 123. 

6.3 Kac’s Recurrence Theorem

For strictly stationary, discrete-parameter sequences, a very pretty theorem,
due to Mark Kac (1947), relates the probability of seeing a particular event to
the mean time between recurrences of the event. Throughout, we consider an
arbitrary Ξ-valued process X, subject only to the requirements of stationarity
and a discrete parameter.
Fix an arbitrary measurable set A ∈ Ξ with P (X1 ∈ A) > 0, and consider a
new process Y (t), where Yt = 1 if Xt ∈ A and Yt = 0 otherwise. By Exercise
5.1, Yt is also stationary. Thus P (X1 ∈ A, X2 6∈ A) = P (Y1 = 1, Y2 = 0). Let
us abbreviate P (Y1 = 0, Y2 = 0, . . . Yn1 = 0, Yn = 0) as wn ; this is the probabil-
ity of making n consecutive observations, none of which belong to the event
CHAPTER 6. RANDOM TIMES 42

A. Clearly, wn ≥ wn+1 . Similarly, let en = P (Y1 = 1, Y2 = 0, . . . Yn = 0) and

rn = P (Y1 = 1, Y2 = 0, . . . Yn = 1) — these are, respectively, the probabilities
of starting in A and not returning within n − 1 steps, and of starting in A
and returning for the first time after n − 2 steps. (Set e1 to P (Y1 = 1), and
w0 = e0 = 1.)

Lemma 65 (Some Recurrence Relations for Kac’s Theorem) The fol-

lowing recurrence relations hold among the probabilities wn , en and rn :

en = wn−1 − wn , n ≥ 1 (6.5)
rn = en−1 − en , n ≥ 2 (6.6)
rn = wn−2 − 2wn−1 + wn , n ≥ 2 (6.7)

Proof: To see the first equality, notice that

P (Y1 = 0, Y2 = 0, . . . Yn−1 = 0) (6.8)

= P (Y2 = 0, Y3 = 0, . . . Yn = 0)
= P (Y1 = 1, Y2 = 0, . . . Yn = 0) + P (Y1 = 0, Y2 = 0, . . . Yn = 0) (6.9)

using first stationarity and then total probability. To see the second equality,
notice that, by total probability,

P (Y1 = 1, Y2 = 0, . . . Yn−1 = 0) (6.10)

= P (Y1 = 1, Y2 = 0, . . . Yn−1 = 0, Yn = 0) + P (Y1 = 1, Y2 = 0, . . . Yn−1 = 0, Yn = 1)

The third relationship follows from the first two. 

Theorem 66 (Recurrence in Stationary Processes) Let X be a Ξ-valued

discrete-parameter stationary process. For any set A with P (X1 ∈ A) > 0, for
almost all ω such that X1 (ω) ∈ A, there exists a τ for which Xτ (ω) ∈ A.
∞
X
P (θA = k|X1 ∈ A) = 1 (6.11)
k=1

Proof: The event {θA = k, X1 ∈ A} is the same as the event {Y1 = 1, Y2 = 0, . . . Yk+1 = 1}.
Since P (X1 ∈ A) > 0, we can handle the conditional probabilities in an elemen-
tary fashion:
P (θA = k, X1 ∈ A)
P (θA = k|X1 ∈ A) = (6.12)
P (X1 ∈ A)
P (Y1 = 1, Y2 = 0, . . . Yk+1 = 1)
= (6.13)
P (Y1 = 1)
∞ P∞
k=1 P (Y1 = 1, Y2 = 0, . . . Yk+1 = 1)
X
P (θA = k|X1 ∈ A) = (6.14)
P (Y1 = 1)
k=1
P∞
k=2 rk
= (6.15)
e1
CHAPTER 6. RANDOM TIMES 43

Now consider the finite sums, and apply Eq. 6.7.

n
X n
X
rk = wk−2 − 2wk−1 + wk (6.16)
k=2 k=2
n−2
X n
X n−1
X
= wk + wk − 2 wk (6.17)
k=0 k=2 k=1
= w0 + wn − w1 − wn−1 (6.18)
= (w0 − w1 ) − (wn−1 − wn ) (6.19)
= e1 − (wn−1 − wn ) (6.20)

where the last line uses Eq. 6.6. Since wn−1 ≥ wn , there exists a limn wn , which
is ≥ 0 since every individual wn is. Hence limn wn−1 − wn = 0.
∞ P∞
k=2 rk
X
P (θA = k|X1 ∈ A) = (6.21)
e1
k=1
e1 − (wn−1 − wn )
= lim (6.22)
n→∞ e1
e1
= (6.23)
e1
= 1 (6.24)

which was to be shown. 

Corollary 67 (Poincaré Recurrence Theorem) Let F be a transformation

which preserves measure µ. Then for any set A of positive µ measure, for µ-
almost-all x ∈ A, ∃n ≥ 1 such that F n (x) ∈ A.

Proof: A direct application of the theorem, given the relationship between

stationary processes and measure-preserving transformations we established by
Corollary 54. 

Corollary 68 (“Nietzsche”) In the set-up of the previous theorem, if X1 (ω) ∈

A, then Xt ∈ A for infinitely many t (a.s.).

Proof: Repeated application of the theorem yields an infinite sequence of

times τ1 , τ2 , τ3 , . . . such that Xτi (ω) ∈ A, for almost all ω such that X1 (ω) ∈ A
in the first place. 
Now that we’ve established that once something happens, it will happen
again and again, we would like to know how long we have to wait between
recurrences.

Theorem 69 (Kac’s Recurrence Theorem) Continuing the previous nota-

tion, E [θA |X1 ∈ A] = 1/P (X1 ∈ A) if and only if limn wn = 0.
CHAPTER 6. RANDOM TIMES 44

Proof: “If”: Unpack the expectation:

∞
X P (Y1 = 1, Y2 = 0, . . . Yk+1 = 1)
E [θA |X1 ∈ A] = k (6.25)
P (Y1 = 1)
k=1
∞
1 X
= krk+1 (6.26)
P (X1 ∈ A)
k=1

so we just need to show that the last series above sums to 1. Using Eq. 6.7
again,
n
X n
X
krk+1 = k(wk−1 − 2wk + wk+1 ) (6.27)
k=1 k=1
Xn n
X n
X
= kwk−1 + kwk+1 − 2 kwk (6.28)
k=1 k=1 k=1
n−1
X n+1
X n
X
= (k + 1)wk + (k − 1)wk − 2 kwk (6.29)
k=0 k=2 k=1
= w0 + nwn+1 − (n + 1)wn (6.30)
= 1 − wn − n(wn − wn+1 ) (6.31)
We therefore wish to show that limn wn = 0 implies limn wn + n(wn − wn+1 ) =
0. By hypothesis, it is enough to show that limn n(wn − wn+1 ) = 0. The partial
sums on the left-hand side of Eq. 6.27 are non-decreasing, so wn +n(wn −wn+1 )
is non-increasing. Since it is also ≥ 0, the limit limn wn + n(wn − wn+1 ) exists.
Since wn → 0, wn − wn+1 ≤ wn must also go to zero; the only question is
whether it goes to zero fast enough. So consider
n
X n
X
lim wk − wk+1 (6.32)
n
k=1 k=1

Telescoping the sums again, this is limn w1 − wn+1 . Since limn+1 wn+1 =
limn wn = 0, the limit exists. But we can equally re-write Eq. 6.32 as
n
X ∞
X
lim wk − wk+1 = wn − wn+1 (6.33)
n
k=1 n=1

Since the sum converges, the individual terms wn −wn+1 must be o(n−1 ). Hence
limn n(wn − wn+1 ) = 0, as was to be shown.
“Only if”: From Eq. 6.31 in the “if” part, we see that the hypothesis is
equivalent to
1 = lim (1 − wn − n(wn − wn+1 )) (6.34)
n

Since wn ≥ wn+1 , 1 − wn − n(wn − wn+1 ) ≤ 1 − wn . We know from the proof of

Theorem 66 that limn wn exists, whether or not it is zero. If it is not zero, then
limn (1 − wn − n(wn − wn+1 )) ≤ 1 − limn wn < 1. Hence wn → 0 is a necessary
condition. 
CHAPTER 6. RANDOM TIMES 45

Example 70 (Counter-example for Kac’s Recurrence Theorem) One

might imagine that the condition wn → 0 in Kac’s Theorem is redundant, given
the assumption of stationarity. Here is a counter-example. Consider a ho-
mogeneous Markov chain on a finite space Ξ, which is partitioned into two
non-communicating components, Ξ1 and Ξ2 . Each component is, internally,
irreducible and aperiodic, so there will be an invariant measure µ1 supported
on Ξ1 , and another invariant measure µ2 supported on Ξ2 . But then, for any
s ∈ [0, 1], sµ1 + (1 − s)µ2 will also be invariant. (Why?) Picking A ⊂ Ξ2 gives
limn wn = s, the probability that the chain begins in the wrong component to
ever reach A.

Kac’s Theorem turns out to be the foundation for a fascinating class of

methods for learning the distributions of stationary processes, and for “univer-
sal” prediction and data compression. There is also an interesting interaction
with large deviations theory. This subject is one possibility for further discus-
sion at the end of the course. Whether or not we get there, let me recommend
some papers in a footnote.2

6.4 Exercises
Exercise 6.1 (Weakly Optional Times and Right-Continuous Filtra-
tions) Show that a random time τ is weakly {Ft }-optional iff it is {F}t+ -
optional.

Exercise 6.2 (Discrete Stopping Times and Their σ-Algebras) Prove

Eq. 6.3. Does the corresponding statement hold for two-sided processes?

Exercise 6.3 (Kac’s Theorem for the Logistic Map) First, do Exercise
5.3. Then, using the same code, suitably modified, numerically check Kac’s
Theorem for the logistic map with a = 4. Pick any interval I ⊂ [0, 1] you like,
but be sure not to make it too small.
1
1. Generate n initial points in I, according to the invariant measure √ .
π x(1−x)
For each point xi , find the first t such that F t (xi ) ∈ I, and take the mean
over the sample. What happens to this space average as n grows?
2. Generate a single point x0 in I, according to the invariant measure. Iterate
it T times. Record the successive times t1 , t2 , . . . at which F t (x0 ) ∈ I, and
find the mean of ti − ti−1 (taking t0 = 0). What happens to this time
average as T grows?

2 “How Sampling Reveals a Process” (Ornstein and Weiss, 1990); Algoet (1992); Kon-

toyiannis et al. (1998).

Chapter 7

Continuity of Stochastic
Processes

Section 7.1 describes the leading kinds of continuity for stochastic

processes, which derive from the modes of convergence of random
variables. It also defines the idea of versions of a stochastic process.
Section 7.2 explains why continuity of sample paths is often prob-
lematic, and why we need the whole “paths in U ” song-and-dance.
As an illustration, we consider a Gausssian process which is close to
the Wiener process, except that it’s got a nasty non-measurability.
Section 7.3 introduces separable random functions.

7.1 Kinds of Continuity for Processes

Continuity is a convergence property: a continuous function is one where con-
vergence of the inputs implies convergence of the outputs. But we have several
kinds of convergence for random variables, so we may expect to encounter several
kinds of continuity for random processes. Note that the following definitions are
stated broadly enough that the index set T does not have to be one-dimensional.

Definition 71 (Continuity in Mean) h A stochastic iprocess X is continuous

2
in the mean at t0 if t → t0 implies E |X(t) − X(t0 )| → 0. X is continuous
in the mean if this holds for all t0 .

It would, of course, be more natural to refer to this as “continuity in mean

square”, or even “continuity in L2 ”, and one can define continuity in Lp for
arbitrary p.

Definition 72 (Continuity in Probability, Stochastic Continuity) X is

P
continuous in probability at t0 if t → t0 implies X(t) → X(t0 ). X is continuous
in probability or stochastically continuous if this holds for all t0 .

46
CHAPTER 7. CONTINUITY 47

Note that neither Lp -continuity nor stochastic continuity says that the indi-
vidual sample paths, themselves, are continuous.

Definition 73 (Continuous Sample Paths) A process X is continuous at t0

if, for almost all ω, t → t0 implies X(t, ω) → X(t0 , ω). A process is continuous
if, for almost all ω, X(·, ω) is a continuous function.

Obviously, continuity of sample paths implies stochastic continuity and Lp -

continuity.
A weaker pathwise property than strict continuity, frequently used in prac-
tice, is the combination of continuity from the right with limits from the left.
This is usually known by the term “cadlag”, abbreviating the French phrase
“continues à droite, limites à gauche”; “rcll” is an unpronounceable synonym.

Definition 74 (Cadlag) A sample function x on a well-ordered set T is cadlag

if it is continuous from the right and limited from the left at every point. That
is, for every t0 ∈ T , t ↓ t0 implies x(t) → x(t0 ), and for t ↑ t0 , limt↑t0 x(t)
exists, but need not be x(t0 ). A stochastic process X is cadlag if almost all its
sample paths are cadlag.

As we will see, it will not be easy to show that our favorite random processes
have any of these desirable properties. What will be easy will be to show that
they are, in some sense, easily modified into ones which do have good regularity
properties, without loss of probabilistic content. This is made more precise
by the notion of versions of a stochastic process, related to that of versions of
conditional probabilities.

Definition 75 (Versions of a Stochastic Process, Stochastically Equiv-

alent Processes) Two stochastic processes X and Y with a common index set
T are called versions of one another if

∀t ∈ T, P (ω : X(t, ω) = Y (t, ω)) = 1

Such processes are also said to be stochastically equivalent.

Lemma 76 (Versions Have Identical FDDs) If X and Y are versions of

one another, they have the same finite-dimensional distributions.

Proof: Clearly it will be enough to show that P (XJ = YJ ) = 1 for arbitrary

finite collections of indices J. Pick any such collection J = {t1 , t2 , . . . tj }. Then


P (XJ = YJ ) = P Xt1 = Yt1 , . . . Xtj = Ytj (7.1)
!
[
= 1−P Xti 6= Yti (7.2)
ti ∈J
X
≥ 1− P (Xti 6= Yti ) (7.3)
ti ∈J
= 1 (7.4)
CHAPTER 7. CONTINUITY 48

using only finite sub-additivity. 

There is a stronger notion of similarity between processes than that of ver-
sions, which will sometimes be useful.

Definition 77 (Indistinguishable Processes) Two stochastic processes X

and Y are indistinguishable, or equivalent up to evanescence, when

P (ω : ∀t, X(t, ω) = Y (t, ω)) = 1

Notice that saying X and Y are indistinguishable means that their sample
paths are equal almost surely, while saying they are versions of one another
means that, at any time, they are almost surely equal. Indistinguishable pro-
cesses are versions of one another, but not necessarily the reverse. (Look at
where the quantifier and the probability statements go.) However, if T = Rd ,
then any two right-continuous versions of the same process are indistinguishable
(Exercise 7.2).

7.2 Why Continuity Is an Issue

In many situations, we want to use stochastic processes to model dynamical
systems, where we know that the dynamics are continuous in time (i.e. the
index set is R, or maybe R+ or [0, T ] for some real T ).1 This means that we
ought to restrict the sample paths to be continuous functions; in some cases we’d
even want them to be differentiable, or yet more smooth. As well as being a
matter of physical plausibility or realism, it is also a considerable mathematical
convenience, as the following shows.

Proposition 78 (Continuous Sample Paths Have Measurable Extrema)

Let X(t, ω) be a real-valued continuous-parameter process with continuous sam-
ple paths. Then on any finite interval I, M (ω) ≡ supt∈I X(t, ω) and m(ω) ≡
inf t∈I X(t, ω) are measurable random variables.

Proof: It’ll be enough to prove this for the supremum function M ; the proof
for m is entirely parallel. First, notice that M (ω) must be finite, because the
sample paths X(·, ω) are continuous functions, and continuous functions are
bounded on bounded intervals. Next, notice that M (ω) > a if and only if
X(t, ω) > a for some t ∈ I. But then, by continuity, there will be some rational
t0 ∈ I ∩ Q such that X(t0 , ω) > a; countably many, in fact.2 Hence
[
{ω : M (ω) > a} = {ω : X(t, ω) > a}
t∈I∩Q

1 Strictly speaking, we don’t really know that space-time is a continuum, but the discretiza-

tion, if there is one, is so fine that it might as well be.

2 Continuity means that we can pick a δ such that, for all t0 within δ of t, X(t0 , ω) is within
1
2
(X(t, ω) − a) of X(t, ω). And there are countably many rational numbers within any real
interval. — There is nothing special about the rational numbers here; any countable, dense
subset of the real numbers would work as well.
CHAPTER 7. CONTINUITY 49

Since, for each t, X(t, ω) is a random variable, the sets in the union on the
right-hand side are all measurable, and the union of a countable collection of
measurable sets is itself measurable. Since intervals of the form (a, ∞) generate
the Borel σ-field on the reals, we have shown that M (ω) is a measurable function
from Ω to the reals, i.e., a random variable. 
Continuity raises some very tricky technical issues. The product σ-field is
the usual way of formalizing the notion that what we know about a stochas-
tic process are values observed at certain particular times. What we saw in
Exercise 1.1 is that “the product σ-field answers countable questions”: for any
measurable set A, whether x(·, ω) ∈ A depends only on the value of x(t, ω) at
countably many indices t. It follows that the class of all continuous sample
paths is not product-σ-field measurable, because x(·, ω) is continuous at t iff
x(tn , ω) → x(t, ω) along every sequence tn → t, and this is involves the value of
the function at uncountably many coordinates. It is further true that the class
of differentiable functions is not product σ-field measurable. For that matter,
neither is the class of piecewise linear functions! (See Exercise 7.1.)
You might think that, on the basis of Theorem 23, this should not really be
much of an issue: that even if the class of continuous sample paths (say) isn’t
strictly measurable, it could be well-approximated by measurable sets, and so
getting the finite-dimensional distributions right is all that matters. This would
make the theory of stochastic processes in continuous time much simpler, but
unfortunately it’s not quite the case. Here is an example to show just how bad
things can get, even when all the finite-dimensional distributions agree.3

Example 79 (A Horrible Version of the proto-Wiener Process) Exam-

ple 38 defined the Wiener process by four requirements: starting at the origin,
independent increments, a Gaussian distribution of increments, and continu-
ity of sample paths. Take a Wiener process W (t, ω) and consider M (ω) ≡
supt∈[0,1] W (t, ω), its supremum over the unit interval. By the preceding propo-
sition, we know that M is a measurable random variable. But we can construct
a version of W for which the supremum is not measurable.
For starters, assume that Ω can be partitioned into an uncountable collection
of disjoint measurable sets, one for each t ∈ [0, 1]. (This can be shown as an
exercise in real analysis.) Select any non-measurable real-valued function B(ω),
so long as B(ω) > M (ω) for all ω. (There are uncountably many suitable
functions.) Set W ∗ (t, ω) = W (t, ω) if ω 6∈ Ωt , and = B(ω) if ω ∈ Ωt . Now,
at every t, P (W (t, ω) = W ∗ (t, ω)) = 1. W ∗ is a version of W , and all their
finite-dimensional distributions agree. But, for every ω, there is a t such that
W ∗ (t, ω) = B(ω) > supt W (t, ω), so supt W ∗ (t, ω) = B(ω), which by design is
non-measurable.

Why care about this example? Two reasons. First, and technically, we’re
going to want to take suprema of the Wiener process a lot when we deal with
large deviations theory, and with the approximation of discrete processes by
continuous ones. Second, nothing in the example really depended on starting
3I stole this example from Pollard (2002, p. 214).
CHAPTER 7. CONTINUITY 50

from the Wiener process; any process with continuous sample paths would have
worked as well. So controlling the finite-dimensional distributions is not enough
to guarantee that a process has measurable extrema, which would lead to all
kinds of embarrassments. We could set up what would look like a reasonable
stochastic model, and then be unable to say things like “with probability p, the
{ temperature/ demand for electricity/ tracking error/ interest rate } will not
exceed r over the course of the year”.
Fundamentally, the issues with continuity are symptoms of a deeper problem.
The reason the supremum function is non-measurable in the example is that it
involves uncountably many indices. A countable collection of ill-behaved sets of
measure zero is a set of measure zero, and may be ignored, but an uncountable
collection of them can have probability 1. Fortunately, there are standard ways
of evading these measure-theoretic issues, by showing that one can always find
random functions which not only have prescribed finite-dimensional distribu-
tions (what we did in Lectures 2 and 3), but also are regular enough that we
can take suprema, or integrate over time, or force them to be continuous. This
hinges on the notion of separability for random functions.

7.3 Separable Random Functions

The basic idea of a separable random function is one whose properties can be
handled by dealing only with a countable, dense subset, just as, in the proof
of Proposition 78, we were able to get away with only looking at X(t) at only
rational values of t. Because a space with a countable, dense subset is called a
“separable” space, we will call such functions “separable functions”.

Definition 80 (Separable Function) Let Ξ and T be metric spaces, and D

be a countable, dense subset of T . A function x : T 7→ Ξ is D-separable or
separable with respect to D if, f orallt ∈ T , there exists a sequence ti ∈ D such
that ti → t and x(ti ) → x(t).

Lemma 81 (Some Sufficient Conditions for Separability) The following

conditions are sufficient for separability:
1. T is countable.
2. x is continuous.
3. T is well-ordered and x is right-continuous.

Proof: (1) Take the separating set to be T itself. (2) Pick any countable
dense D. By density, for every t there will be a sequence ti ∈ D such that ti → t.
By continuity, along any sequence converging to t, x(ti ) → t. (3) Just like (2),
only be sure to pick the ti > t. (You can do this, again, for any countable dense
D.) 
CHAPTER 7. CONTINUITY 51

Definition 82 (Separable Process) A Ξ-valued process X on T is separable

with respect to D if D is a countable, dense subset of T , and there is a measure-
zero set N ⊂ Ω such that for every ω 6∈ N , X(·, ω) is D-separable. That is,
X(·, ω) is almost surely D-separable.

We cannot easily guarantee that a process is separable. What we can easily

do is go from one process, which may or may not be separable, to a separa-
ble process with the same finite-dimensional distributions. This is known as
a separable modification of the original process. Combined with the extension
theorems (Theorems 27, 29 and 33), this tells that we can always construct a
separable process with desired finite-dimensional distributions. We shall there-
fore feel entitled to assume that our processes are separable, without further
ado. The proofs of the existence of separable and continuous versions of gen-
eral processes are, however, somewhat involved, and so postponed to the next
lecture.

7.4 Exercises
Exercise 7.1 (Piecewise Linear Paths Not in the Product σ-Field)
Consider real-valued functions on the unit interval (i.e., Ξ = R, T = [0, 1],
X = B). The product σ-field is thus B [0,1] . In many circumstances, it would be
useful to constrain sample paths to be piece-wise linear functions of the index.
Let PL([0, 1]) denote this class of functions. Use the argument of Exercise 1.1
to show that PL([0, 1]) 6∈ B [0,1] .

Exercise 7.2 (Indistinguishability of Right-Continuous Versions) Show

that, if X and Y are versions of one another, with index set Rd , and both are
right-continuous, then they are indistinguishable.
Chapter 8

More on Continuity

Section 8.1 constructs separable modifications of reasonable but

non-separable random functions, and explains how separability re-
lates to non-denumerable properties like continuity.
Section 8.2 constructs versions of our favorite one-parameter pro-
cesses where the sample paths are measurable functions of the pa-
rameter.
Section 8.3 gives conditions for the existence of cadlag versions.
Section 8.4 gives some criteria for continuity, and for the existence
of “continuous modifications” of discontinuous processes.

Recall the story so far: last time we saw that the existence of processes with
given finite-dimensional distributions does not guarantee that they have desir-
able and natural properties, like continuity, and in fact that one can construct
discontinuous versions of processes which ought to be continuous. We therefore
need extra theorems to guarantee the existence of continuous versions of pro-
cesses with specified FDDs. To get there, we will first prove the existence of
separable versions. This will require various topological conditions on both the
index set T and the value space Ξ.
In the interest of space (or is it time?), Section 8.1 will provide complete and
detailed proofs. The other sections will simply state results, and refer proofs
to standard sources, mostly Gikhman and Skorokhod (1965/1969). (They in
turn follow Doob (1953), but are explicit about what he regarded as obvious
generalizations and extensions, and they cost about $20, whereas Doob costs
$120 in paperback.)

8.1 Separable Versions

We can show that separable versions of our favorite stochastic processes exist
under quite general conditions, but first we will need some preliminary results,
living at the border between topology and measure theory. This starts by re-
calling some facts about compact spaces.

52
CHAPTER 8. MORE ON CONTINUITY 53

Definition 83 (Compactness, Compactification) A set A in a topological

space Ξ is compact if every covering of A by open sets contains a finite sub-cover.
Ξ is a compact space if it is itself a compact set. If Ξ is not compact, but is a
sub-space of a compact topological space Ξ̃, the super-space is a compactification
of Ξ.

Proposition 84 (Compact Spaces are Separable) Every compact space is

separable.
Proof: A standard result of analysis. Interestingly, it relies on the axiom
of choice. 
Proposition 85 (Compactification is Possible) Every non-compact topo-
logical space Ξ is a sub-space of some compact topological space Ξ̃, i.e., every
non-compact topological space can be compactified.

Example 86 (Compactifying the Reals) The real numbers R are not com-
pact: they have no finite covering by open intervals (or other open sets). The
extended reals, R ≡ R ∪ +∞ ∪ −∞, are compact, since intervals of the form
(a, ∞] and [−∞, a) are open. This is a two-point compactification of the reals.
There is also a one-point compactification, with a single point at ±∞, but this
has the undesirable property of making big negative and positive numbers close
to each other.
Recall that a random function is separable if its value at any arbitrary in-
dex can be determined almost surely by examining its values on some fixed,
countable collection of indices. The next lemma states an alternative charac-
terization of separability. The lemma after that gives conditions under which a
weaker property holds — the almost-sure determination of whether X(t, ω) ∈ B,
for a specific t and set B, by the behavior of X(tn , ω) at countably many tn .
The final lemma extends this to large collections of sets, and then the proof of
the theorem puts all the parts together.
Lemma 87 (Alternative Characterization of Separability) Let T be a
separable set, Ξ a compact metric space, and D a countable dense subset of T .
Define V as the class of all open balls in T centered at points in D and with
rational radii. For any G ⊂ T , let
!
[
R(G, ω) ≡ closure X(t, ω) (8.1)
t∈G∩D
\
R(t, ω) ≡ R(S, ω) (8.2)
S: S∈V, t∈S

Then X(t, ω) is D-separable if and only if there exists a set N ⊂ Ω such that
ω 6∈ N ⇒ ∀t, X(t, ω) ∈ R(t, ω) (8.3)
and P (N ) = 0.
CHAPTER 8. MORE ON CONTINUITY 54

Proof: Roughly speaking, R(t, ω) is what we’d think the range of the
function would be, in the vicinity of t, if it we went just by what it did at points
in the separating set D. The actual value of the function falling into this range
(almost surely) is necessary and sufficient for the function to be separable. But
let’s speak less roughly.
“Only if”: Since X(t, ω) is D-separable, for almost all ω, for any t there
is some sequence tn ∈ D such that tn → t and X(tn , ω) → X(t, ω). For any
ball S centered at t, there is some N such that tS n ∈ S if n ≥ N . Hence the
values of x(tn ) are eventually confined to the set t∈S∩D X(t, ω). Recall that
the closure of a set A consists of the points x such that, for some sequence
xn ∈ A, Sxn → x. As X(t
n , ω) → X(t, ω), it must be the case that X(t, ω) ∈
closure t∈S∩D X(t, ω) . Since this applies to all S, X(t, ω) must be in the
intersection of all those closures, hence X(t, ω) ∈ R(t, ω) — unless we are on
one of the probability-zero bad sample paths, i.e., unless ω ∈ N .
“If”: Assume that, with probability 1, X(t, ω) ∈ R(t, ω). Thus, for any
S ∈ V , we know that there exists a sequence of points tn ∈ S ∩ D such that
X(tn , ω) → X(t, ω). However, this doesn’t say that tn → t, which is what we
need for separability. We will now build such a sequence. Consider a series
of spheres Sk ∈ V such that (i) every point in Sk is within a distance 2−k of
(k)
t and (ii) Sk+1 ⊂ Sk . For each Sk , there is a sequence tn ∈ Sk such that
(k) (k)
X(tn , ω) → X(t, ω). In fact, for any m > 0, |X(tn , ω) − X(t, ω)| < 2−m if
n ≥ N (k, m), for some N (k, m). Our final sequence of indices ti then consists of
(1) (2)
the following points: tn for n from N (1, 1) to N (1, 2); tn for n from N (2, 2)
(k)
to N (2, 3); and in general tn for n from N (k, k) to N (k, k + 1). Clearly, ti → t,
and X(ti , ω) → X(t, ω). Since every ti ∈ D, we have shown that X(t, ω) is
D-separable. 

Lemma 88 (Confining Bad Behavior to a Measure Zero Event) Let T

be a separable index set, Ξ a compact space, X a random function from T to Ξ,
and B be an arbitrary Borel set of Ξ. Then there exists a denumerable set of
points tn ∈ T such that, for any t ∈ T , the set
∞
!
\
N (t, B) ≡ {ω : X(t, ω) 6∈ B} ∩ {ω : X(tn , ω) ∈ B} (8.4)
n=1

has probability 0.

Proof: We proceed recursively. The first point, t1 , can be whatever we

like. Suppose t1 , t2 , . . . tn are already found, and define the following:
n
\
Mn ≡ {ω : X(tk , ω) ∈ B} (8.5)
k=1
Ln (t) ≡ Mn ∩ {ω : X(t, ω) 6∈ B} (8.6)
pn ≡ sup P (Ln (t)) (8.7)
t
CHAPTER 8. MORE ON CONTINUITY 55

Mn is the set where the random function, evaluated at the first n indices, gives a
value in our favorite set; it’s clearly measurable. Ln (t), also clearly measurable,
gives the collection of points in Ω where, if we chose t for the next point in
the collection, this will break down. pn is the worst-case probability of this
happening. For each t, Ln+1 (t) ⊆ Ln (t), so pn+1 ≤ pn . Suppose pn = 0;
then we’ve found the promised denumerable sequence, and we’re done. Suppose
instead that pn > 0. Pick any t such that P (Ln (t)) ≥ 12 pn , and call it tn+1 .
(There has to be such a point, or else pn wouldn’t be the supremum.) Now notice
that L1 (t2 ), L2 (t3 ), . . . Ln (tn+1 ) are all mutually exclusive, but not necessarily
jointly exhaustive. So

1 = P (Ω) (8.8)
!
[
≥ P Ln (tn+1 ) (8.9)
n
X
= P (Ln (tn+1 )) (8.10)
n
X1
≥ pn > 0 (8.11)
n
2

so pn → 0 as n → ∞.
We saw that Ln (t) is a monotone-decreasing sequence of sets, for each t,
so a limiting set exists, and in fact limn Ln (t) = N (t, B). So, by monotone
convergence,
 
P (N (t, B)) = P lim Ln (t) (8.12)
n
= lim P (Ln (t)) (8.13)
n
≤ lim pn (8.14)
n
= 0 (8.15)

as was to be shown. 

Lemma 89 (A Null Exceptional Set) Let B0 be any countable class of Borel

sets in Ξ, and B the closure of B0 under countable intersection. Under the
hypotheses of the previous lemma, there is a denumerable sequence tn such that,
for every t ∈ T , there exists a set N (t) ⊂ Ω with P (N (t)) = 0, and, for all
B ∈ B,
∞
!
\
{ω : X(t, ω) 6∈ A} ∩ {ω : X(tn , ω) ∈ A} ⊆ N (t) (8.16)
n=1

Proof: For each B ∈ B0 , construct the sequence of indices as in the previous

lemma. Since there only countably many sets in B, if we take the union of all of
these sequences, we will get another countable sequence, call it tn . Then we have
that, ∀B ∈ B0 , ∀t ∈ T , P (X(tn , ω) ∈ B, n ≥ 1, X(t, ω) 6∈ B) = 0. Take this set
CHAPTER 8. MORE ON CONTINUITY 56

S
to be N (t, B), and define N (t) ≡ B∈B0 N (t, B). Since N (t) is a countable
union of probability-zero events, it is itself a probability-zero event. Now, take
any B ∈ B, and any B0 ∈ B0 such that B ⊆ B0 . Then
∞
!
\
{X(t, ω) 6∈ B0 } ∩ {X(tn , ω) ∈ B} (8.17)
n=1
∞
!
\
⊆ {X(t, ω) 6∈ B0 } ∩ {X(tn , ω) ∈ B0 }
n=1
⊆ N (t) (8.18)
T (k) (k)
Since B = k B0 for some sequence of sets B0 ∈ B0 , it follows (via De
Morgan’s laws and the distributive law) that
∞ n o
(k)
[
{X(t, ω) 6∈ B} = X(t, ω) 6∈ B0 (8.19)
k=1

∞
!
\
{X(t, ω) 6∈ B} ∩ {X(tn , ω) ∈ B}
n=1
∞ n ∞
!
o
(k)
[ \
= X(t, ω) 6∈ B0 ∩ {X(tn , ω) ∈ B} (8.20)
k=1 n=1
[∞
⊆ N (t) (8.21)
n=1
= N (t) (8.22)

which was to be shown. 

Theorem 90 (Separable Versions, Separable Modification) Suppose that

Ξ is a compact metric space and T is a separable metric space. Then, for any
Ξ-valued stochastic process X on T , there exists a separable version X̃. This is
called a separable modification of X.

Proof: Let D be a countable dense subset of T , and V the class of open

spheres of rational radius centered at points in D. Any open subset of T is
a union of countably many sets from V , which is itself countable. Similarly,
let C be a countable dense subset of Ξ, and let B0 consist of the complements
of spheres centers at points in D with rational radii, and (as in the previous
lemma) let B be the closure of B0 under countable intersection. Every closed
set in Ξ belongs to B.1 For every S ∈ V , consider the restriction of X(t, ω) to
t ∈ S, and apply Lemma 89 to the random function X(t, ω) to get a sequence
of indices I(S) ⊂ T , and, for every t ∈ S, a measure-zero set NS (t) ⊂ Ω where
1 You show this.
CHAPTER 8. MORE ON CONTINUITY 57

S S
things can go wrong. Set I = S∈V I(S) and N (t) = S∈V NS (t). Because
V is countable, I is still a countable set of indices, and N (t) is still of measure
zero. I is going to be our separating set, and we’re going to show that we have
uncountably many sets N (t) won’t be a problem.
Define X̃(t, ω) = X(t, ω) if t ∈ I or ω 6∈ N (t) — if we’re at a time in the
separating set, or we’re at some other time but have avoided the bad set, we’ll
just copy our original random function. What to do otherwise, when t 6∈ I and
ω ∈ N (t)? Construct R(t, ω), as in the proof of Lemma 87, and let X̃(t, ω) take
any value in this set. Since R(t, ω) depends only on the value of the function
at indices in the separating set, it doesn’t matter whether we build it from
X or from X̃. In fact, for all t and ω, X̃(t,n ω) ∈ R(t, ω), so, o by Lemma 87,
X̃(t, ω) is separable. Finally, for every t, X̃(t, ω) = X(t, ω) ⊆ N (t), so ∀t,
 
P X̃(t) = X(t) , and X̃ is a version of X (Definition 75). 

Corollary 91 (Separable Modifications in Compactified Spaces) If the

situation is as in the previous theorem, but Ξ is not compact, there exists a
separable version of X in any compactification Ξ̃ of Ξ.

Proof: Because Ξ is a sub-space of any of its compactifications Ξ̃, X is also

a process with values in Ξ̃.2 Since Ξ̃ is compact, X has a separable modification
X̃ with values in Ξ̃, but (with probability 1) X̃(t) ∈ Ξ. 

Corollary 92 (Separable Versions with Prescribed Distributions Ex-

ist) Let Ξ be a complete, compact metric space, T a separable index set, and
µJ , J ∈ Fin(T ) a projective family of probability distributions. Then there is a
separable stochastic process with finite-dimensional distributions given by µJ .

Proof: Combine Theorem 90 with the Kolmogorov Extension Theorem 29.

(By Proposition 84, a complete, compact metric space is a complete, separable
metric space.) 

8.2 Measurable Versions

It would be nice for us if X(t) is a measurable function of t, because we are
going to want to write down things like
Z t=b
X(t)dt
t=a

and have them mean something. Irritatingly, this will require another modifi-
cation.
2 If you want to be really picky, define a 1-1 function h : Ξ 7→ Ξ̃ taking points to their

counterparts. Then X and h−1 (X) are indistinguishable. Do I need to go on?

CHAPTER 8. MORE ON CONTINUITY 58

Definition 93 (Measurable random function) Let T, T , τ be a measurable

space, its σ-field and a measure defined thereon. A random function X on T
with values in Ξ, X is measurable if X : T × Ω 7→ Ξ is T^× F/X measurable,
where T × F is the product σ-field on T × Ω, and T^ × F its completion by the
null sets of the product measure τ × P.

It would seem more natural to simply define measurable processes by saying

that the sample paths are measurable, that X(·, ω) is a T -measurable function of
t for P-almost-all ω. This would be like our definition of almost-sure continuity.
However, Definition 93 implies this version, via Fubini’s Theorem, and facilitates
the proofs of the two following theorems.

Theorem 94 (Exchanging Expectations and Time Integrals) If X(t) is

measurable, and E [X(t)] is integrable (with respect to the measure τ on T ), then
for any set I ∈ T ,
Z Z 
E [X(t)] τ (dt) = E X(t)τ (dt) (8.23)
I I

Proof: This is just Fubini’s Theorem! 

Theorem 95 (Measurable Separable Modifications) Suppose that T and

Ξ are both compact. If X(t, ω) is continuous in probability at τ -almost-all t,
then it has a version which is both separable and measurable, its measurable
separable modification.

Proof: See Gikhman and Skorokhod (1965/1969, ch. IV, sec. 3, thm. 1,
p. 157). 

8.3 Cadlag Versions

Theorem 96 (Cadlag Versions) Let X be a separable random process with
T = [a, b] ⊆ R, and Ξ a complete metric space with metric ρ. Suppose that X(t)
is continuous in probability on T , and there are real constants p, q, C ≥ 0, r > 1
such that, for any three indices t1 < t2 < t3 ∈ T ,
r
E [ρp (X(t1 ), X(t2 ))ρq (X(t2 ), X(t3 ))] ≤ C|t3 − t1 | (8.24)

Then there is a version of X whose sample paths are cadlag (a.s.).

Proof: Combine Theorem 1 and Theorem 3 of Gikhman and Skorokhod

(1965/1969, ch. IV, sec. 4, pp. 159–169). 
CHAPTER 8. MORE ON CONTINUITY 59

8.4 Continuous Modifications

Theorem 97 (Continuous Versions) Let X be a separable stochastic process
with T = [a, b] ⊆ R, and Ξ a complete metric space with metric ρ. Suppose that
there are constants C, p > 0, r > 1 such that, for any t1 < t2 ∈ T ,
r
E [ρp (X(t1 ), X(t2 ))] ≤ C|t2 − t1 | (8.25)
Then X(t) has a continuous version.
Proof: See Gikhman and Skorokhod (1965/1969, ch. IV, sec. 5, thm. 2,
p. 170), and the first remark following the theorem. 
A slightly more refined result requires two preliminary definitions.
Definition 98 (Modulus of continuity) For any function x from a metric
space T, d to a metric space Ξ, ρ, the modulus of continuity is the function
mx (r) : R+ 7→ R+ given by
mx (r) = sup {ρ(x(s), x(t)) : s, t ∈ T, d(s, t) ≤ r} (8.26)

Lemma 99 (Modulus of continuity and uniform continuity) x is uni-

formly continuous if and only if its modulus of continuity → 0 as r → 0.
Proof: Obvious from Definition 98 and the definition of uniform continuity.
Definition 100 (Hölder continuity) Continuing the notation of Definition
98, we say that x is Hölder-continuous with exponent c if there are positive
constants c, γ such that mx (r) ≤ γrc for all sufficiently small r; i.e., mx (r) =
O(rc ). If this holds on every bounded subset of T , then the function is locally
Hölder-continuous.

Theorem 101 (Hölder-continuous versions) Let T be Rd and Ξ a complete

metric space with metric ρ. If there are constants p, q, γ > 0, such that, for any
t1 , t 2 ∈ T ,
d+q
E [ρp (X(t1 ), X(t2 ))] ≤ γ|t1 − t2 | (8.27)
then X has a continuous version X̃, and almost all sample paths of X̃ are locally
Hölder-continuous for any exponent between 0 and q/p exclusive.
Proof: See Kallenberg, theorem 3.23 (pp. 57–58). Note that part of Kallen-
berg’s proof is a restricted case of what we’ve already done in proving the exis-
tence of a separable version! 
This lecture, the last, and even a lot of the one before have all been pretty
hard and abstract. As a reward for our labor, however, we now have a collec-
tion of very important tools — operator representations, filtrations and optional
times, recurrence times, and finally existence theorems for continuous processes.
These are the devices which will let us take the familiar theory of elementary
Markov chains, with finitely many states in discrete time, and produce the gen-
eral theory of Markov processes with continuously-many states and/or continu-
ous time. The next lecture will begin this work, starting with the operators.
 Part III

Markov Processes

60
Chapter 9

Markov Processes

This chapter begins our study of Markov processes.

Section 9.1 is mainly “ideological”: it formally defines the Markov
property for one-parameter processes, and explains why it is a nat-
ural generalization of both complete determinism and complete sta-
tistical independence.
Section 9.2 introduces the description of Markov processes in
terms of their transition probabilities and proves the existence of
such processes.
Section 9.3 deals with the question of when being Markovian
relative to one filtration implies being Markov relative to another.

9.1 The Correct Line on the Markov Property

The Markov property is the independence of the future from the past, given the
present. Let us be more formal.

Definition 102 (Markov Property) A one-parameter process X is a Markov

process with respect to a filtration {Ft } when Xt is adapted to the filtration, and,
|=

for any s > t, Xs is independent of Ft given Xt , Xs Ft |Xt . If no filtration is

mentioned, it may be assumed to be the natural one generated by X. If X is also
conditionally stationary, then it is a time-homogeneous (or just homogeneous)
Markov process.

Lemma 103 (The Markov Property Extends to the Whole Future) Let
Xt+ stand for the collection of Xu , u > t. If X is Markov, then Xt+ Ft |Xt .
|=

Proof: See Exercise 9.1. 

There are two routes to the Markov property. One is the path followed by
Markov himself, of desiring to weaken the assumption of strict statistical inde-
pendence between variables to mere conditional independence. In fact, Markov

61
CHAPTER 9. MARKOV PROCESSES 62

specifically wanted to show that independence was not a necessary condition for
the law of large numbers to hold, because his arch-enemy claimed that it was,
and used that as grounds for believing in free will and Christianity.1 It turns
out that all the key limit theorems of probability — the weak and strong laws of
large numbers, the central limit theorem, etc. — work perfectly well for Markov
processes, as well as for IID variables.
The other route to the Markov property begins with completely deterministic
systems in physics and dynamics. The state of a deterministic dynamical system
is some variable which fixes the value of all present and future observables.
As a consequence, the present state determines the state at all future times.
However, strictly deterministic systems are rather thin on the ground, so a
natural generalization is to say that the present state determines the distribution
of future states. This is precisely the Markov property.
Remarkably enough, it is possible to represent any one-parameter stochastic
process X as a noisy function of a Markov process Z. The shift operators give
a trivial way of doing this, where the Z process is not just homogeneous but
actually fully deterministic. An equally trivial, but slightly more probabilistic,
approach is to set Zt = Xt− , the complete past up to and including time t. (This
is not necessarily homogeneous.) It turns out that, subject to mild topological
conditions on the space X lives in, there is a unique non-trivial representation
where Zt = (Xt− ) for some function , Zt is a homogeneous Markov process,
|=

and Xu σ({Xt , t ≤ u})|Zt . (See Knight (1975, 1992); Shalizi and Crutchfield
(2001).) We may explore such predictive Markovian representations at the end
of the course, if time permits.

9.2 Transition Probability Kernels

The most obvious way to specify a Markov process is to say what its transition
probabilities are. That is, we want to know P (Xs ∈ B|Xt = x) for every s > t,
x ∈ Ξ, and B ∈ X . Probability kernels (Definition 30) were invented to let us
do just this. We have already seen how to compose such kernels; we also need
to know how to take their product.

Definition 104 (Product of Probability Kernels) Let µ and ν be two prob-

ability kernels from Ξ to Ξ. Then their product µν is a kernel from Ξ to Ξ,
defined by
Z
(µν)(x, B) ≡ µ(x, dy)ν(y, B) (9.1)

= (µ ⊗ ν)(x, Ξ × B) (9.2)

1 I am not making this up. See Basharin et al. (2004) for a nice discussion of the origin of

Markov chains and of Markov’s original, highly elegant, work on them. There is a translation
of Markov’s original paper in an appendix to Howard (1971), and I dare say other places as
well.
CHAPTER 9. MARKOV PROCESSES 63

Intuitively, all the product does is say that the probability of starting at
the point x and landing in the set B is equal the probability of first going to
y and then ending in B, integrated over all intermediate points y. (Strictly
speaking, there is an abuse of notation in Eq. 9.2, since the second kernel in a
composition ⊗ should be defined over a product space, here Ξ × Ξ. So suppose
we have such a kernel ν 0 , only ν 0 ((x, y), B) = ν(y, B).) Finally, observe that if
µ(x, ·) = δx , the delta function at x, then (µν)(x, B) = ν(x, B), and similarly
that (νµ)(x, B) = ν(x, B).

Definition 105 (Transition Semi-Group) For every (t, s) ∈ T × T , s ≥ t,

let µt,s be a probability kernel from Ξ to Ξ. These probability kernels form a
transition semi-group when
1. For all t, µt,t (x, ·) = δx .
2. For any t ≤ s ≤ u ∈ T , µt,u = µt,s µs,u .
A transition semi-group for which ∀t ≤ s ∈ T , µt,s = µ0,s−t ≡ µs−t is homoge-
neous.

As with the shift semi-group, this is really a monoid (because µt,t acts as
the identity).
The major theorem is the existence of Markov processes with specified tran-
sition kernels.

Theorem 106 (Existence of Markov Process with Given Transition

Kernels) Let µt,s be a transition semi-group and νt a collection of distributions
on a Borel space Ξ. If

νs = νt µt,s (9.3)

then there exists a Markov process X such that ∀t,

L (Xt ) = νt (9.4)

and ∀t1 ≤ t2 ≤ . . . ≤ tn ,

L (Xt1 , Xt2 . . . Xtn ) = νt1 ⊗ µt1 ,t2 ⊗ . . . ⊗ µtn−1 ,tn (9.5)

Conversely, if X is a Markov process with values in Ξ, then there exist distri-

butions νt and a transition kernel semi-group µt,s such that Equations 9.4 and
9.3 hold, and

P (Xs ∈ B|Ft ) = µt,s a.s. (9.6)

Proof: (From transition kernels to a Markov process.) For any finite set of
times J = {t1 , . . . tn } (in ascending order), define a distribution on ΞJ as

νJ ≡ νt1 ⊗ µt1 ,t2 ⊗ . . . ⊗ µtn−1 ,tn (9.7)

CHAPTER 9. MARKOV PROCESSES 64

It is easily checked, using point (2) in the definition of a transition kernel semi-
group (Definition 105), that the νJ form a projective family of distributions.
Thus, by the Kolmogorov Extension Theorem (Theorem 29), there exists a
stochastic process whose finite-dimensional distributions are the νJ . Now pick
a J of size n, and two sets, B ∈ X n−1 and C ∈ X .

P (XJ ∈ B × C) = νJ (B × C) (9.8)
= E [1B×C (XJ )] (9.9)
 
= E 1B (XJ\tn )µtn−1 ,tn (Xtn−1 , C) (9.10)

Set {Ft } to be the natural filtration, σ({Xu , u ≤ s}). If A ∈ Fs for some s ≤ t,

then by the usual generating class arguments we have

P Xt ∈ C, Xs− ∈ A = E [1A µs,t (Xs , C)]

(9.11)
P (Xt ∈ C|Fs ) = µs,t (Xs , C) (9.12)
|=

i.e., Xt Fs |Xs , as was to be shown.

(From the Markov property to the transition kernels.) From the Markov
property, for any measurable set C ∈ X , P (Xt ∈ C|Fs ) is a function of Xs
alone. So define the kernel µs,t by µs,t (x, C) = P (Xt ∈ C|Xs = x), with a pos-
sible measure-0 exceptional set from (ultimately) the Radon-Nikodym theorem.
(The fact that Ξ is Borel guarantees the existence of a regular version of this
conditional probability.) We get the semi-group property for these kernels thus:
pick any three times t ≤ s ≤ u, and a measurable set C ⊆ Ξ. Then

µt,u (Xt , C) = P (Xu ∈ C|Ft ) (9.13)

= P (Xu ∈ C, Xs ∈ Ξ|Ft ) (9.14)
= (µt,s ⊗ µs,u )(Xt , Ξ × C) (9.15)
= (µt,s µs,u )(Xt , C) (9.16)

The argument to get Eq. 9.3 is similar. 

Note: For one-sided discrete-parameter processes, we could use the Ionescu-
Tulcea Extension Theorem 33 to go from a transition kernel semi-group to a
Markov process, even if Ξ is not a Borel space.

Definition 107 (Invariant Distribution) Let X be a homogeneous Markov

process with transition kernels µt . A distribution ν on Ξ is invariant when, ∀t,
ν = νµt , i.e.,
Z
(νµt )(B) ≡ ν(dx)µt (x, B) (9.17)

= ν(B) (9.18)

ν is also called an equilibrium distribution.

CHAPTER 9. MARKOV PROCESSES 65

The term “equilibrium” comes from statistical physics, where however its
meaning is a bit more strict, in that “detailed balance” must also be satisified:
for any two sets A, B ∈ X ,
Z Z
ν(dx)1A µt (x, B) = ν(dx)1B µt (x, A) (9.19)

i.e., the flow of probability from A to B must equal the flow in the opposite
direction. Much confusion has resulted from neglecting the distinction between
equilibrium in the strict sense of detailed balance and equilibrium in the weaker
sense of invariance.
Theorem 108 (Stationarity and Invariance for Homogeneous Markov
Processes) Suppose X is homogeneous, and L (Xt ) = ν, where ν is an invari-
ant distribution. Then the process Xt+ is stationary.
Proof: Exercise 9.4. 

9.3 The Markov Property Under Multiple Fil-

trations
Definition 102 specifies what it is for a process to be Markovian relative to a
given filtration {Ft }. The question arises of when knowing that X Markov with
respect to one filtration {Ft } will allow us to deduce that it is Markov with
respect to another, say {Gt }.
To begin with, let’s introduce a little notation.
Definition 109
 (Natural
Filtration) The natural filtration for a stochastic
process X is FtX ≡ σ({Xu , u ≤ t}). Every process X is adapted to its natural
filtration.

Definition 110 (Comparison of Filtrations) A filtration {Gt } is finer than

or more refined than or a refinement of {Ft }, {Ft } ≺ {Gt }, if, for all t, Ft ⊆ Gt ,
and at least sometimes the inequality is strict. {Ft } is coarser or less fine than
{Gt }. If {Ft } ≺ {Gt } or {Ft } = {Gt }, we write {Ft }  {Gt }.

Lemma 111 (The Natural Filtration Is the Coarsest  One to Which a

Process Is Adapted) If X is adapted to {Gt }, then FtX  {Gt }.
Proof: For each t, Xt is Gt measurable. But FtX is, by construction,
the smallest σ-algebra with respect to which Xt is measurable, so, for every t,
FtX ⊆ Gt , and the result follows. 
Theorem 112 (Markovianity Is Preserved Under Coarsening) If X is
Markovian with respect to {Gt }, then it is Markovian with respect to any coarser
filtration to which it is adapted, and in particular with respect to its natural
filtration.
CHAPTER 9. MARKOV PROCESSES 66

Proof: Use the smoothing property of conditional expectations: For any

two σ-fields H ⊂ K and random variable Y , E [Y |H] = E [E [Y |K] |H] a.s. So,
if {Ft } is coarser than {Gt }, and X is Markovian with respect to the latter, for
any function f ∈ L1 and time s > t,

E [f (Xs )|Ft ] = E [E [f (Xs )|Gt ] |Ft ] a.s. (9.20)

= E [E [f (Xs )|Xt ] |Ft ] (9.21)
= E [f (Xs )|Xt ] (9.22)

|=
The next-to-last line uses the fact that Xs Gt |Xt , because X is Markovian
with respect to {Gt }, and this in turn implies that conditioning Xs , or any
function thereof, on Gt is equivalent to conditioning on Xt alone. (Recall that
Xt is Gt -measurable.) The last line uses the facts that (i) E [f (Xs )|Xt ] is a
function Xt , (ii) X is adapted to {Ft }, so Xt is Ft -measurable, and (iii) if Y
is F-measurable, then E [Y |F] = Y . Since this holds for all f ∈ L1 , it holds
in particular for 1A , where A is any measurable set, and this established the
conditional independence which constitutes the Markov property. Since (Lemma
111) the natural filtration is the coarsest filtration to which X is adapted, the
remainder of the theorem follows. 
The converse is false, as the following example shows.

Example 113 (The Logistic Map Shows That Markovianity Is Not

Preserved Under Refinement) We revert to the symbolic dynamics of the
logistic map, Examples p 39 and 40. Let S1 be distributed on the unit inter-
val with density 1/π s(1 − s), and let Sn = 4Sn−1 (1 − Sn−1 ). Finally, let
Xn = 1[0.5,1.0] (Sn ). It can be shown that the Xn are a Markov process with
respect to their natural filtration; in fact, with respect to that filtration, they are
independent and identically distributed
Bernoulli variables with probability of
success 1/2. However, P Xn+1 |FnS , Xn 6= P (Xn+1 |Xn ), since X  n+1 is adeter-

ministic function of Sn . But, clearly, FnX ⊂ FnS for each n, so FtX ≺ FtS .

The issue can be illustrated with graphical models (Spirtes et al., 2001; Pearl,
1988). A discrete-time Markov process looks like Figure 9.1a. Xn blocks all the
pasts from the past to the future (in the diagram, from left to right), so it
produces the desired conditional independence. Now let’s add another variable
which actually drives the Xn (Figure 9.1b). If we can’t measure the Sn variables,
just the Xn ones, then it can still be the case that we’ve got the conditional
independence among what we can see. But if we can see Xn as well as Sn —
which is what refining the filtration amounts to — then simply conditioning on
Xn does not block all the paths from the past of X to its future, and, generally
speaking, we will lose the Markov property. Note that knowing Sn does block all
paths from past to future — so this remains a hidden Markov model. Markovian
representation theory is about finding conditions under which we can get things
to look like Figure 9.1b, even if we can’t get them to look like Figure 9.1a.
CHAPTER 9. MARKOV PROCESSES 67

S1 S2 S3 ...

X1 X2 X3 ...

X1 X2 X3 ...

a b

Figure 9.1: (a) Graphical model for a Markov chain. (b) Refining the filtration,
say by conditioning on an additional random variable, can lead to a failure of
the Markov property.

9.4 Exercises
Exercise 9.1 (Extension of the Markov Property to the Whole Fu-
ture) Prove Lemma 103.

Exercise 9.2 (Futures of Markov Processes Are One-Sided Markov

Processes) Show that if X is a Markov process, then, for any t ∈ T , Xt+ is a
one-sided Markov process.

Exercise 9.3 (Discrete-Time Sampling of Continuous-Time Markov

Processes) Let X be a continuous-parameter Markov process, and tn a count-
able set of strictly increasing indices. Set Yn = Xtn . Is Yn a Markov process?
If X is homogeneous, is Y also homogeneous? Does either answer change if
tn = nt for some constant interval t > 0?

Exercise 9.4 (Stationarity and Invariance for Homogeneous Markov

Processes) Prove Theorem 108.

Exercise 9.5 (Rudiments of Likelihood-Based Inference for Markov

Chains) (This exercise presumes some knowledge of sufficient statistics and
exponential families from theoretical statistics.)
Assume T = N, Ξ is a finite set, and X a homogeneous Markov Ξ-valued
Markov chain. Further assume that X1 is constant, = x1 , with probability 1.
(This last is not an essential restriction.) Let pij = P (Xt+1 = j|Xt = i).
1. Show that pij fixes the transition kernel, and vice versa.
2. Write the probability of a sequence X1 = x1 , X2 = x2 , . . . Xn = xn , for
short X1n = xn1 , as a function of pij .
3. Write the log-likelihood ` as a function of pij and xn1 .
CHAPTER 9. MARKOV PROCESSES 68

4. Define nij to be the number of times t such that xt = i and xt+1 = j.

Similarly define ni as the number of times t such that xt = i. Write the
log-likelihood as a function of pij and nij .
5. Show that the nij are sufficient statistics for the parameters pij .
6. Show that the distribution has the form of a canonical exponential family,
with sufficient statistics nij , by finding the natural parameters. (Hint: the
natural parameters are transformations of the pij .) Is the family of full
rank?
7. Find the maximum likelihood estimators, for either the pij parameters or
for the natural
P parameters. (Hint: Use Lagrange multipliers to enforce the
constraint j nij = ni .)
The classic book by Billingsley (1961) remains an excellent source on statistical
inference for Markov chains and related processes.

Exercise 9.6 (Implementing the MLE for a Simple Markov Chain)

(This exercise continues the previous one.)
Set Ξ = {0, 1}, x1 = 0, and
 
0.75 0.25
p0 =
0.25 0.75

1. Write a program to simulate sample paths from this Markov chain.

2. Write a program to calculate the maximum likelihood estimate p̂ of the
transition matrix from a sample path.
3. Calculate 2(`(p̂) − `(p0 )) for many independent sample paths of length
n. What happens to the distribution as n → ∞? (Hint: see Billingsley
(1961), Theorem 2.2.)

Exercise 9.7 (The Markov Property and Conditional Independence

from the Immediate Past) Let X be a Ξ-valued discrete-parameter random
|=

process. Suppose that, for all t, Xt−1 Xt+1 |Xt . Either prove that X is a
Markov process, or provide a counter-example. You may assume that Ξ is a
Borel space if you find that helpful.

Exercise 9.8 (Higher-Order Markov Processes) A discrete-time process

X is a k th order Markov process with respect to a filtration {Ft } when
|=

Xt+1 Ft |σ(Xt , Xt−1 , . . . Xt−k+1 ) (9.23)

for some finite integer k. For any Ξ-valued discrete-time process X, define the k-
(k)
block process Y (k) as the Ξk -valued process where Yt = (Xt , Xt+1 , . . . Xt+k−1 ).
CHAPTER 9. MARKOV PROCESSES 69

1. Prove that if X is Markovian to order k, it is Markovian to any order

l > k. (For this reason, saying that X is a k th order Markov process
conventionally means that k is the smallest order at which Eq. 9.23 holds.)
2. Prove that X is k th -order Markovian if and only if Y (k) is Markovian.

The second result shows that studying on the theory of first-order Markov pro-
cesses involves no essential loss of generality. For a test of the hypothesis that
X is Markovian of order k against the alternative that is Markovian of order
l > k, see Billingsley (1961). For recent work on estimating the order of a
Markov process, assuming it is Markovian to some finite order, see the elegant
paper by Peres and Shields (2005).

Exercise 9.9 (AR(1) Models) A first-order autoregressive model, or AR(1)

model, is a real-valued discrete-time process defined by an evolution equation of
the form
X(t) = a0 + a1 X(t − 1) + (t)
where the innovations (t) are independent and identically distributed, and in-
dependent of X(0). A pth -order autoregressive model, or AR(p) model, is cor-
respondingly defined by
p
X
X(t) = a0 + ai X(t − i) + (t)
i=1

Finally an AR(p) model in state-space form is an Rp -valued process defined by

 
a1 a2 . . . ap−1 ap
 1 0 ... 0 0 
 
~  0 1 ... 0 0  ~ (t − 1) + (t)~e1
Y (t) = a0~e1 +  Y
 .. .. .. .. .. 
 . . . . . 
0 0 ... 1 0

where ~e1 is the unit vector along the first coordinate axis.

1. Prove that AR(1) models are Markov for all choices of a0 and a1 , and all
distributions of X(0) and .

2. Give an explicit form for the transition kernel of an AR(1) in terms of

the distribution of .
3. Are AR(p) models Markovian when p > 1? Prove or give a counter-
example.
~ is a Markov process, without using Exercise 9.8. (Hint:
4. Prove that Y
What is the relationship between Yi (t) and X(t − i + 1)?)
Chapter 10

Alternative
Characterizations of
Markov Processes

This lecture introduces two ways of characterizing Markov pro-

cesses other than through their transition probabilities.
Section 10.1 describes discrete-parameter Markov processes as
transformations of sequences of IID uniform variables.
Section 10.2 describes Markov processes in terms of measure-
preserving transformations (Markov operators), and shows this is
equivalent to the transition-probability view.

10.1 Markov Sequences as Transduced Noise

A key theorem says that discrete-time Markov processes can be viewed as the
result of applying a certain kind of filter to pure noise.

Theorem 114 (Markov Sequences as Transduced Noise) Let X be a one-

sided discrete-parameter process taking values in a Borel space Ξ. X is Markov
iff there are measurable functions fn : Ξ × [0, 1] 7→ Ξ such that, for IID random
variables Zn ∼ U (0, 1), all independent of X1 , Xn+1 = fn (Xn , Zn ) almost
surely. X is homogeneous iff fn = f for all n.

Proof: Kallenberg, Proposition 8.6, p. 145. Notice that, in order to get the
“only if” direction to work, Kallenberg invokes what we have as Proposition 26,
which is where the assumptions that Ξ is a Borel space comes in. You should
verify that the “if” direction does not require this assumption. 
Let us stick to the homogeneous case, and consider the function f in some-
what more detail.

70
CHAPTER 10. MARKOV CHARACTERIZATIONS 71

In engineering or computer science, a transducer is an apparatus — really, a

function — which takes a stream of inputs of one kind and produces a stream
of outputs of another kind.

Definition 115 (Transducer) A (deterministic) transducer is a sextuple hΣ, Υ, Ξ, f, h, s0 i

where Σ, Υ and Ξ are, respectively, the state, input and output spaces, f : Σ ×
Ξ 7→ Σ is the state update function or state transition function, h : Σ × Υ 7→ Ξ
is the measurement or observation function, and s0 ∈ Σ is the starting state.
(We shall assume both f and h are always measurable.) If h does not depend
on its state argument, the transducer is memoryless. If f does not depend on
its state argument, the transducer is without after-effect.

It should be clear that if a memoryless transducer is presented with IID

inputs, its output will be IID as well. What Theorem 114 says is that, if
we have a transducer with memory (so that h depends on the state) but is
without after-effect (so that f does not depend on the state), IID inputs will
produce Markovian outputs, and conversely any reasonable Markov process can
be represented in this way. Notice that if a transducer is without memory,
we can replace it with an equivalent with a single state, and if it is without
after-effect, we can identify Σ and Ξ.
Notice also that the two functions f and h determine a transition func-
tion where we use the input to update the state: g : Σ × Υ 7→ Σ, where
g(s, y) = f (s, h(s, y)). Thus, if the inputs are IID and uniformly distributed,
then (Theorem 114) the successive states of the transducer are always Marko-
vian. The question of which processes can be produced by noise-driven transduc-
ers is this intimately bound up with the question of Markovian representations.
While, as mentioned, quite general stochastic processes can be put in this form
(Knight, 1975, 1992), it is not necessarily possible to do this with a finite in-
ternal state space Σ, even when Ξ is finite. The distinction between finite and
infinite Σ is crucial to theoretical computer science, and we might come back to
it later, but
Two issues suggest themselves in connection with this material. One is
whether, given a two-sided process, we can pull the same trick, and represent a
Markovian X as a transformation of an IID sequence extending into the infinite
past. (Remember that the theorem is for one-sided processes, and starts with
an initial X1 .) This is more subtle than it seems at first glance, or even than it
seemed to Norbert Wiener when he first posed the question (Wiener, 1958); for
a detailed discussion, see Rosenblatt (1971), and, for recent set of applications,
Wu (2005). The other question is whether the same trick can be pulled in
continuous time; here much less is known.
CHAPTER 10. MARKOV CHARACTERIZATIONS 72

10.2 Time-Evolution (Markov) Operators

Let’s look again at the evolution of the one-dimensional distributions for a
Markov process:

νs = νt µt,s (10.1)
Z
νs (B) = νt (dx)µt,s (x, B) (10.2)

The transition kernels define linear operators taking probability measures on Ξ

to probability measures on Ξ. This can be abstracted.

Definition 116 (Markov Operator on Measures) Take any measurable

space Ξ, X . A Markov operator on measures is an operator M which takes
finite measures on this space to other finite measures on this space such that
1. M is linear, i.e., for any a1 , a2 ∈ [0, 1] and any two measures µ1 , µ2 ,

M (a1 µ1 + a2 µ2 ) = a1 M µ1 + a2 M µ2 (10.3)

2. M is norm-preserving, i.e., M µ(Ξ) = µ(Ξ).

(In particular P must take probability measures to probability measures.)

Definition 117 (Markov Operator on Densities) Take any probability space

Ξ, X , µ, and let L1 be as usual the class of all µ-integrable generalized functions
on Ξ. A linear operator P : L1 7→ L1 is a Markov operator on densities when:
1. If f ≥ 0 (a.e. µ), P f ≥ 0 (a.e. µ).
2. If f ≥ 0 (a.e. µ), kP f k = kf k.

By “a Markov operator” I will often mean a Markov operator on densities,

with the reference measure µ being some suitable uniform distribution on Ξ.
However, the general theory applies to operators on measures.

Lemma 118 (Markov Operators on Measures Induce Those on Densi-

ties) Let M be a Markov operator on measures. If M takes measures absolutely
continuous with respect to µ to measures absolutely continuous with respect to
µ, i.e., it preserves domination by µ, then it induces an almost-unique Markov
operator P on densities with respect to µ.

Proof: Let f be a function which is inR L1 (µ) and is non-negative (µ-a.e.).

If f ≥ 0 µ a.e., the set function νf (A) = A f (x)dµ is a finite measure which
is absolutely continuous with respect to µ. (Why is it finite?) By hypothesis,
then, M νf is another finite measure which is absolutely continuous with respect
to µ, and νf (Ξ) = M νf (Ξ). Hence, by the Radon-Nikodym
R theorem, there is an
L1 (µ) function, call it P f , such that M νf (A) = A P f (x)dµ. (“Almost unique”
CHAPTER 10. MARKOV CHARACTERIZATIONS 73

refers to the possibility of replacing P f with another versionR of dM νf /dµ.)

In particular, P f (x)R ≥ 0 for µ-almost-all x, and so kP f k = Ξ |P f (x)|dµ =
M νf (Ξ) = νf (Ξ) = Ξ |f (x)|dµ = kf k. Finally, the linearity of the operator on
densities follows directly from the linearity of the operator on measures and the
linearity of integration. If f is sometimes negative, apply the reasoning above
to f + and f − , its positive and negative parts, and then use linearity again.
R Recall from Definition 30 that, for an arbitrary kernel κ, κf (x) is defined as
f (y)κ(x, dy). Applied to our transition kernels, this suggests another kind of
operator.

Definition 119 (Transition Operators) Take any measurable space Ξ, X ,

and let B(Ξ) be the class of bounded measurable functions. An operator K :
B(Ξ) 7→ B(Ξ) is a transition operator when:
1. K is linear
2. If f ≥ 0 (a.e. µ), Kf ≥ 0 (a.e. µ)
3. K1Ξ = 1Ξ .
4. If fn ↓ 0, then Kfn ↓ 0.

Definition 120 (L∞ Transition Operators) For a probability space Ξ, X , µ,

an L∞ -transition operator is an operator on L∞ (µ) satisfying points (1)–(4) of
Definition 119.

Note that every function in B(Ξ) is in L∞ (µ) for each µ, so the definition
of an L∞ transition operator is actually stricter than that of a plain transition
operator. This is unlike the case with Markov operators, where the L1 version
is weaker than the unrestricted version.

Lemma 121 (Kernels and Operators) Every probability kernel κ from Ξ to

Ξ induces a Markov operator M on measures,

M ν = νκ (10.4)

and every Markov operator M on measures induces a probability kernel κ,

κ(x, B) = M δx (B) (10.5)

Similarly, every transition probability kernel induces a transition operator K on

functions,
Kf (x) = κf (x) (10.6)
and every transition operator K induces a transition probability kernel,

κ(x, B) = K1B (x) (10.7)

CHAPTER 10. MARKOV CHARACTERIZATIONS 74

Proof: Exercise 10.1. 

Now we need some basic concepts from functional analysis; see, e.g., Kol-
mogorov and Fomin (1975) for background.

Definition 122 (Functional) A functional is a map g : V 7→ R, that is, a

real-valued function of a function. A functional g is
• linear when g(af1 + bf2 ) = ag(f1 ) + bg(f2 );

• continuous when fn → f implies g(fn ) → f ;

• bounded by M when |g(f )| ≤ M for all f ∈ V ;
• bounded when it is bounded by M for some M ;
• non-negative when g(f ) ≥ 0 for all f ;
etc.

Definition 123 (Conjugate or Adjoint Space) The conjugate space or ad-

joint space of a vector space V is the space V † of its continuous linear function-
als. For f ∈ V and g ∈ V † , hf, gi denotes g(f ). This is sometimes called the
inner product.

Proposition 124 (Conjugate Spaces are Vector Spaces) For every V , V †

is also a vector space.

Proposition 125 (Inner Product is Bilinear) For any a, b, c, d ∈ R, any

f1 , f2 ∈ V and any g1 , g2 ∈ V † ,

haf1 + bf2 , cg1 + dg2 i = achf1 , g1 i + adhf1 , g2 i + bchf2 , g1 i + bdhf2 , g2 i (10.8)

Proof: Follows from the fact that V † is a vector space, and each gi is a
linear operator. 
You are already familiar with an example of a conjugate space.

Example 126 (Vectors in Rn ) The vector spaceP Rn is self-conjugate. If g(~x)

n
is a continuous linear function of ~x, then g(~x) = i=1 yi xi for some real con-
stants yi , which means g(~x) = ~y · ~x.

Here is the simplest example where the conjugate space is not equal to the
original space.

Example 127 (Row and Column Vectors) The space of row vectors is con-
jugate to the space of column vectors, since every continuous linear functional
of a column vector x takes the form of y T x for some other column vector y.
CHAPTER 10. MARKOV CHARACTERIZATIONS 75

Example 128 (Lp spaces) The function spaces Lp (µ) and Lq (µ) are conjugate
to each other, when 1/p + 1/q = 1, and the inner product is defined through
Z
hf, gi ≡ f gdµ (10.9)

In particular, L1 and L∞ are conjugates.

Example 129 (Measures and Functions) The space of Cb (Ξ) of bounded,

continuous functions on Ξ and the spaces M(Ξ, X ) of finite measures on Ξ are
conjugates, with inner product
Z
hµ, f i = f dµ (10.10)

Definition 130 (Adjoint Operator) For conjugate spaces V and V † , the

adjoint operator, O† , to an operator O on V is an operator on V † such that

hOf, gi = hf, O† gi (10.11)

for all f ∈ V, g ∈ V † .

Proposition 131 (Adjoint of a Linear Operator) If O is a continuous

linear operator on V , then its adjoint O† exists and is linear.

Lemma 132 (Markov Operators on Densities and L∞ Transition Op-

erators) Every Markov operator P on densities induces an L∞ transition op-
erator U on essentially-bounded functions, and vice versa.

Proof: Exercise 10.2. 

Clearly, if κ is part of a transition kernel semi-group, then the collection of
induced Markov operators and transition operators also form semi-groups.

Theorem 133 (Transition operator semi-groups and Markov processes)

Let X be a Markov process with transition kernels µt,s , and let Kt,s be the cor-
responding semi-group of transition operators. Then for any f ∈ B(Ξ),

E [f (Xs )|Ft ] = (Kt,s f )(Xt ) (10.12)

Conversely, let X be any stochastic process, and let Kt,s be a semi-group of tran-
sition operators such that Equation 10.12 is valid (a.s.). Then X is a Markov
process.
CHAPTER 10. MARKOV CHARACTERIZATIONS 76

Proof: Exercise 10.3. 

Remark. The proof works because the expectations of all B(Ξ) functions
together determine a probability measure. (Recall that P (B) = E [1B ], and
indicators are bounded everywhere.) If we knew of another collection of func-
tions which also sufficed to determine a measure, then linear operators on that
collection would work just as well, in the theorem, as do the transition oper-
ators, which by definition apply to all of B(Ξ). In particular, it is sometimes
possible to define operators only on much smaller, more restricted collections of
functions, which can have technical advantages. See Ethier and Kurtz (1986,
ch. 4, sec. 1) for details.
The next two lemmas are useful in establishing asymptotic results.

Lemma 134 (Markov Operators are Contractions) For any Markov op-
erator P and f ∈ L1 ,

kP f k ≤ kf k (10.13)

Proof (after Lasota and Mackey (1994, prop. 3.1.1, pp. 38–39)): First,
+
notice that (P f (x)) ≤ P f + (x), because
+ +
(P f (x)) = (P f + − P f − ) = max (0, P f + − P f − ) ≤ max (0, P f + ) = P f +
−
Similarly (P f (x)) ≤ P f − (x). Therefore |P f | ≤ P |f |, and then the statement
follows by integration. 

Lemma 135 (Markov Operators Bring Distributions Closer) For any

Markov operator, and any f, g ∈ L1 , kP n f − P n gk is non-increasing.

Proof: By linearity, kP n f − P n gk = kP n (f − g)k. By the definition of P n ,

kP (f − g)k = kP P n−1 (f − g)k. By the contraction property (Lemma 134),
n

kP P n−1 (f − g)k ≤ kP n−1 (f − g)k = kP n−1 f − P n−1 gk (by linearity again). 

Theorem 136 (Invariant Measures Are Fixed Points) A probability mea-

sure ν is invariant for a homogeneous Markov process iff it is a fixed point of
all the Markov operators, Mt ν = ν.

Proof: Clear from the definitions! 

10.3 Exercises
Exercise 10.1 (Kernels and Operators) Prove Lemma 121. Hints: 1. You
will want to use the fact that 1A ∈ B(Ξ) for all measurable sets A. 2. In going
back and forth between transition kernels and transition operators, you may find
Proposition 32 helpful.

Exercise 10.2 (L1 and L∞ ) Prove Lemma 132.

CHAPTER 10. MARKOV CHARACTERIZATIONS 77

Exercise 10.3 (Operators and Expectations) Prove Theorem 133. Hint:

in showing that a collection of operators determines a Markov process, try using
mathematical induction on the finite-dimensional distributions.

Exercise 10.4 (Bayesian Updating as a Markov Process) Consider a

simple version of Bayesian learning, where the set of hypotheses Θ is finite, and,
for each θ ∈ Θ, fθ (x) is a probability density on Ξ with respect to a common
dominating measure, µ say, and the Ξ-valued data X1 , X2 , . . . are all IID, both
under the hypotheses and in reality. Given a prior probability vector π0 on Θ,
the posterior πn is defined via Bayes’s rule:
Qn
π0 (θ) i=1 fθ (Xi )
πn (θ) = P Qn
θ∈Θ π0 (θ) i=1 fθ (Xi )

1. Prove that the random sequence π1 , π2 , . . . is adapted to {Ft } if X is

adapted.
2. Prove that the sequence of posterior distributions is Markovian with respect
to its natural filtration.
3. Is this still a Markov process if X is not IID? If the hypotheses θ do not
model X as IID?
4. When, if ever, is the Markov process homogeneous? (If it is sometimes
homogeneous, you may give either necessary or sufficient conditions, as
you find easier.)

Exercise 10.5 (More on Bayesian Updating) Consider a more complicated

version of Bayesian updating. Let T be one-sided, H be a Θ-valued random
variable, and {Gt } be any filtration. Assume that πt = L (H|Gt ) is a regular
conditional probability distribution on Θ for all t and all ω. As before, π0 is the
prior. Prove that πt is Markovian with respect to {Gt }. (Hint: E [E [Y |Gt ] |Gs ] =
E [X|Gs ] a.s., when s ≤ t and Y ∈ L1 so the expectations exist.)
Chapter 11

Examples of Markov
Processes

Section 11.1 looks at the evolution of densities under the action

of the logistic map; this shows how deterministic dynamical systems
can be brought under the sway of the theory we’ve developed for
Markov processes.
Section 11.2 finds the transition kernels for the Wiener process,
as an example of how to manipulate such things.
Section 11.3 generalizes the Wiener process example to other
processes with stationary and independent increments, and in doing
so uncovers connections to limits of sums of IID random variables
and to self-similarity.

11.1 Probability Densities in the Logistic Map

Let’s revisit the first part of Exercise 5.3, from the point of view of what we now
know about Markov processes. The exercise asks us to show that the density
√ 1 is invariant under the action of the logistic map with a = 4.
π x(1−x)
Let’s write the mapping as F (x) = 4x (1 −x). Solving √ a simple quadratic
√
equation gives us the fact that F −1 (x) is the set 12 1 − 1 − x , 12 1 + 1 − x .

Notice, for later

 use, that√ the
 two solutions√ add
up to 1. Notice also that
F −1 ([0, x]) = 0, 12 1 − 1 − x ∪ 12 1 + 1 − x , 1 . Now we consider the

78
CHAPTER 11. MARKOV EXAMPLES 79

cumulative distribution function of Xn+1 , P (Xn+1 ≤ x).

P (Xn+1 ≤ x)
= P (Xn+1 ∈ [0, x]) (11.1)
= P Xn ∈ F −1 ([0, x])


(11.2)
√ √
    
1
1

= P Xn ∈ 0, 1− 1−x ∪ 1 + 1 − x ,1 (11.3)
2 2
Z 12 (1−√1−x) Z 1
= ρn (y) dy + √
ρn (y) dy (11.4)
2 (1+ 1−x)
1
0

where ρn is the density of Xn . So we have an integral equation for the evolution

of the density,
√
Z x Z 1
2 (1− 1−x) Z 1
ρn+1 (y) dy = ρn (y) dy + √
ρn (y) dy (11.5)
0 0 1
2 (1+ 1−x)

This sort of integral equation is complicated to solve directly. Instead, take

the derivative of both sides with respect to x; we can do this through the
fundamental theorem of calculus. On the left hand side, this will just give
ρn+1 (x), the density we want.

ρn+1 (x) (11.6)

√
d
Z 1
2 (1− 1−x) d
Z 1
= ρn (y) dy +
√
ρn (y) dy
dx 0 (1+ 1−x) dx 1
2

√ √
   
1
d 1

= ρn 1− 1−x 1− 1−x (11.7)
2 dx 2
√ √
   
1
d 1

−ρn 1+ 1−x 1+ 1−x
2 dx 2
√ √
    
1 1
1

= √ ρn 1 − 1 − x + ρn 1+ 1−x (11.8)
4 1−x 2 2

Notice that this defines a linear operator taking densities to densities. (You
should verify the linearity.) In fact, this is a Markov operator, by the terms of
Definition 117. Markov operators of this sort, derived from deterministic maps,
are called Perron-Frobenius or Frobenius-Perron operators, and accordingly de-
noted by P . Thus an invariant density is a ρ∗ such that ρ∗ = P ρ∗ . All the
CHAPTER 11. MARKOV EXAMPLES 80

1
problem asks us to do is to verify that √ is such a solution.
π x(1−x)

√
 
1
ρ∗


1− 1−x (11.9)
2
−1/2
√ √
  
1 1
1

= 1− 1−x 1− 1− 1−x
π 2 2
√ √
−1/2
 
1 1
1
= 1− 1−x 1+ 1−x (11.10)
π 2 2
2
= √ (11.11)
π x

Since ρ∗ (x) = ρ∗ (1 − x), it follows that

√
 
1 1
P ρ∗ ρ∗


= 2 √ 1− 1−x (11.12)
4 1−x 2
1
= p (11.13)
π x (1 − x)
= ρ∗ (11.14)

as desired.
By Lemma 135, for any distribution ρ, kP n ρ − P n ρ∗ k is a non-increasing
function of n. However, P n ρ∗ = ρ∗ , so the iterates of any distribution, under
the map, approach the invariant distribution monotonically. It would be very
handy if we could show that any initial distribution ρ eventually converged on
ρ∗ , i.e. that kP n ρ − ρ∗ k → 0. When we come to ergodic theory, we will see
conditions under which such distributional convergence holds, as it does for the
logistic map, and learn how such convergence in distribution is connected to
both pathwise convergence properties, and to the decay of correlations.

11.2 Transition Kernels and Evolution Opera-

tors for the Wiener Process
We have previously defined the Wiener process (Examples 38 and 79) as the
real-valued process on R+ with the following properties:
1. W (0) = 0;
|=

2. For any three times t1 ≤ t2 ≤ t3 , W (t3 ) − W (t2 ) W (t2 ) − W (t1 ) (inde-

pendent increments);
3. For any two times t1 ≤ t2 , W (t2 ) − W (t1 ) ∼ N (0, t2 − t1 ) (Gaussian
increments);
4. Continuous sample paths (in the sense of Definition 73).
CHAPTER 11. MARKOV EXAMPLES 81

In Exercise 4.1, you showed that a process satisfying points (1)–(3) above
exists. To show that any such process has a version with continuous sample
paths, invoke Theorem 97 (Exercise 11.3). Here, however, we will show that
it is a homogeneous Markov process, and find both the transition kernels and
the evolution operators. Markovianity follows from the independent increments
property (2), while homogeneity and the form of the transition operators comes
from the Gaussian assumption (3).
First, here’s how the Gaussian increments property gives us the transition
probabilities:
P (W (t2 ) ∈ B|W (t1 ) = w1 ) = P (W (t2 ) − W (t1 ) ∈ B − w1 ) (11.15)
Z
1 − u 2
= du p e 2(t2 −t1 )(11.16)
B−w1 2π(t 2 − t 1 )
(w −w1 )2
Z
1 − 2
= dw2 p e 2(t2 −t1 ) (11.17)
B 2π(t2 − t1 )
≡ µt1 ,t2 (w1 , B) (11.18)
Since this evidently depends only on t2 − t1 , and not the individual times, the
process is homogeneous. Now, assuming homogeneity, we can find the time-
evolution operators for well-behaved observables:
Kt f (w) = E [f (Wt + s)|Ws = w] (11.19)
Z
= f (u)µt (w, du) (11.20)
Z
1 − (u−w)2
= f (u) √ e 2t du (11.21)
2πt
h √ i
= E f (w + tZ) (11.22)

where Z is a standard Gaussian random variable independent of Ws .

To show that W (t) is a Markov process, we must show that, for any finite
collection of times t1 ≤ t2 ≤ . . . ≤ tk ,
L (Wt1 , Wt2 , . . . Wtk ) = µt1 ,t2 µt2 ,t3 . . . µtk−1 ,tk (11.23)
Let’s just go through the k = 3 case, as the others are fundamentally similar,
but with more notation. Notice that W (t3 ) − W (t1 ) = (W (t3 ) − W (t2 )) +
(W (t2 ) − W (t1 )). Because increments are independent, then, W (t3 ) − W (t1 ) is
the sum of two independent random variables, W (t3 )−W (t2 ) and W (t2 )−W (t1 ).
The distribution of W (t3 ) − W (t1 ) is then the convolution of distributions of
W (t3 ) − W (t2 ) and W (t2 ) − W (t1 ). Those are N (0, t3 − t2 ) and N (0, t2 − t1 )
respectively. The convolution of two Gaussian distributions is a third Gaus-
sian, summing their parameters, so according to this argument, we must have
W (t3 ) − W (t1 ) ∼ N (0, t3 − t1 ). But this is precisely what we should have, by
the Gaussian-increments property. Since the trick we used above to get the
transition kernel from the increment distribution can be applied again, we con-
clude that µt1 ,t2 µt2 ,t3 = µt1 ,t3 . The same trick applies when k > 3. Therefore
(Theorem 106), W (t) is a Markov process, with respect to its natural filtration.
CHAPTER 11. MARKOV EXAMPLES 82

11.3 Lévy Processes and Limit Laws

Let’s think a bit more about the trick we’ve just pulled with the Wiener process.
We wrote the time evolution operators in terms of a random variable,
h √ i
Kt f (w) = E f (w + tZ) (11.24)
√
Notice that tZ ∼ N (0, t), i.e., it is the increment of the process over an interval
of time t. This can be generalized to other processes of a similar sort. These
are one natural way of generalizing the idea of a sum of IID random variables.

Definition 137 (Processes with Stationary and Independent Incre-

ments) A stochastic process X is a stationary, independent-increments process,
or has stationary, independent increments when
1. The increments are independent; for any collection of indices t1 ≤ t2 ≤
. . . ≤ tk , with k finite, the increments X(ti ) − X(ti−1 ) are all jointly
independent.
2. The increments are stationary; for any τ > 0,

L (X(t2 ) − X(t1 ), X(t3 ) − X(t2 ), . . . X(tk ) − X(tk−1 )) (11.25)

= L (X(t2 + τ ) − X(t1 + τ ), X(t3 + τ ) − X(t2 + τ ), . . . X(tk + τ ) − X(tk−1 + τ ))

Definition 138 (Lévy Processes) A Lévy process is a process with station-

ary and independent increments and cadlag sample paths, which start at zero,
X(0) = 0.

Example 139 (Wiener Process is Lévy) The Wiener process is a Lévy

process (since it has not just cadlag but continuous sample paths).

Example 140 (Poisson Counting Process) The Poisson counting process

with intensity λ is the integer-valued process N on R+ where N (0) = 0, N (t) ∼
Poisson(λt), and independent increments. It defines a Poisson point process
(in the sense of Example 20) by assigning the interval [0, t] the measure N (t),
the measure extending to other Borel sets in the usual way. Conversely, any
Poisson process defines a counting process in this sense. N can be shown to be
continuous in probability, and so, via Theorem 96, to have a cadlag modification.
(Exercise 11.5.) This is a Lévy process.

It is not uncommon to see people writing just “processes with independent

increments” when they mean “stationary and independent increments”.

Theorem 141 (Processes with Stationary Independent Increments are

Markovian) Any process with stationary, independent increments is Markovian
with respect to its natural filtration, and homogeneous in time.
CHAPTER 11. MARKOV EXAMPLES 83

Proof: The joint distribution for any finite collection of times factors into
a product of incremental distributions:

L (X(t1 ), X(t2 ), . . . X(tk )) (11.26)

= L (X(t1 )) L (X(t2 ) − X(t1 )) . . . L (X(tk ) − X(tk−1 ))

Define µt (x, B) = P (X(t1 + t) − X(t1 ) + x ∈ B). This is a probability ker-

nel. Moreover, it satisfies the semi-group property, since X(t3 ) − X(t1 ) =
X(t3 ) − X(t2 ) + X(t2 ) + X(t1 ) (and similarly for more indices). Thus the
µt , so defined, are transition probability kernels (Definition 105). By Theorem
106, X is Markovian with respect to its natural filtration. 
Remark: The assumption that the increments are stationary is not necessary
to prove Markovianity. What property of X does it then deliver?

Theorem 142 (Time-Evolution Operators of Processes with Station-

ary, Independent Increments) If X has stationary and independent incre-
ments, then its time-evolution operators Kt are given by

Kt f (x) = E [f (x + Zt )] (11.27)

where L (Zt ) = L (X(t) − X(0)), the increment from 0 to t.

Proof: Exactly parallel to the Wiener process case.

Kt f (x) = E [f (Xt )|F0 ] (11.28)

= E [f (Xt )|X0 = x] (11.29)
Z
= f (y)µt (x, dy) (11.30)

Since µt (x, B) = P (X(t1 + t) − X(t1 ) + x ∈ B), the theorem follows. 

Notice that so far everything has applied equally to discrete or to continuous
time. For discrete time, we can chose the distribution of increments over a single
time-step, L (X(1) − X(0)), to be essentially whatever we like; stationarity and
independence then determine the distribution of all other increments. For con-
tinuous time, however, our choices are more constrained, and in an interesting
way.
By the semi-group property of the time-evolution operators, we must have
Ks Kt = Kt+s for all times s and t. Applying Theorem 142, it must be the case
that

Kt+s f (x) = E [f (x + Zt+s )] (11.31)

= Ks E [f (x + Zt )] (11.32)
= E [f (x + Zs + Zt )] (11.33)
|=

where Zs Zt . That is, the distribution of Zt+s must be the convolution of

the distributions of Zt and Zs . Let us in particular pick s = t; this implies
L (Z2t ) = L (Zt ) ? L (Zt ), where ? indicates convolution. Writing νt for L (Zt ),
CHAPTER 11. MARKOV EXAMPLES 84

another way to phrase our conclusion is that νt = νt/2 ? νt/2 . Clearly, this
argument could be generalized to relate νt to νt/n for any n:
?n
νt = νt/n (11.34)

where the superscript ?n indicates the n-fold convolution of the distribution

with itself. Equivalently, for each t, and for all n,
n
d
X
Zt = Di (11.35)
i=1

where the IID Di ∼ νt/n . This is a very curious-looking property with a name.

Definition 143 (Infinitely-Divisible Distributions and Random Vari-

ables) A probability distribution ν is infinitely divisible when, for each n, there
exists a νn such that ν = νn?n . A random variable Z is infinitely divisible when,
(n) d P (n)
for each n, there are n IID random variables Di such that Z = i Di , i.e.,
when its distribution is infinitely divisible.

Clearly, if ν is an infinitely divisible distribution, then it can be obtained as

(n)
the limiting distribution of a sum of IID random variables Di . (To converge,
the individual terms must be going to zero in probability as n grows.) More
remarkably, the converse is also true:

Proposition 144 (Limiting Distributions Are Infinitely Divisible) If ν

is the limiting distribution of a sequence of IID sums, then it is an infinitely-
divisible distribution.

Proof: See, for instance, Kallenberg (2002, Theorem 15.12). 

Remark: This should not be obvious. If we take larger and larger sums of
(centered, standardized) IID Bernoulli variables, we obtain a Gaussian as the
limiting distribution, but at no time is the Gaussian the distribution of any of
the finite sums. That is, the IID sums in the definition of infinite divisibility
are not necessarily at all the same as those in the proposition.

Theorem 145 (Infinitely Divisible Distributions and Stationary In-

dependent Increments) If X is a process with stationary and independent
increments, then for all t, X(t) − X(0) is infinitely divisible.

Proof: See the remarks before the definition of infinite divisibility. 

Corollary 146 (Infinitely Divisible Distributions and Lévy Processes)

If ν is an infinitely divisible distribution, there exists a Lévy process where
X(1) ∼ ν, and this determines the other distributions.

Proof: If X(1) ∼ ν, then by Eq. 11.34, X(n) ∼ ν ∗n for all integer n.

∗n
Conversely, the distribution of X(1/n) is also fixed, since ν = L (X(1/n)) ,
which means that the characteristic function of ν is equal to the nth power of
CHAPTER 11. MARKOV EXAMPLES 85

that of X(1/n); inverting the latter then gives the desired distribution. Because
d
increments are independent, X(n+1/m) = X(n)+X(1/m), hence L (X(1)) fixes
the increment distribution for all rational time-intervals. By continuity, however,
this also fixes it for time intervals of any length. Since the distribution of
increments and the condition X(0) = 0 fix the finite-dimensional distributions, it
remains only to show that the process is cadlag, which can be done by observing
that it is continuous in probability, and then using Theorem 96. 
We have established a correspondence between the limiting distributions of
IID sums, and processes with stationary independent increments. There are
many possible applications of this correspondence; one is to reduce problems
about limits of IID sums to problems about certain sorts of Markov process,
and vice versa. Another is to relate discrete and continuous time processes. For
0 ≤ t ≤ 1, set
bntc
(n) 1 X
Xt = √ Yi (11.36)
n i=0
where Y0 = 0 but otherwise the Yi are IID with mean 0 and variance 1. Then
each X (n) is a Markov process in discrete time, with stationary, independent
increments, and its time-evolution operator is
  
[nt]
1 X
Kn f (x) = E f x + √ Yi  (11.37)
n i=1

√
As n grows, the normalized sum approaches
√ tZ, with Z a standard Gaussian
random variable. (Where does the t come from?) Since the time-evolution
operators determine the finite-dimensional distributions of a Markov process
(Theorem 133), this suggests that in some sense X (n) should be converging
in distribution on the Wiener process W . This is in fact true, but we will
have to go deeper into the structure of the operators concerned before we can
make it precise, as a first form of the functional central limit theorem. Just
as the infinitely divisible distributions are the limits of IID sums, processes
with stationary and independent increments are the limits of the corresponding
random walks.
There is a yet further sense which the Gaussian distribution is a special kind
of limiting distribution, which is reflected in another curious property of the
Wiener process. Gaussian random variables are not only infinitely divisible, but
they can be additively decomposed into more Gaussians. Distributions which
can be infinitely divided into others of the same kind will be called “stable”
(under convolution or averaging). “Of the same kind” isn’t very mathematical;
here is a more precise expression of the idea.

Definition 147 (Self-similarity) A process is self-similar if, for all t, there

d
exists a measurable h(t), the scaling function, such that h(t)X(1) = X(t).
CHAPTER 11. MARKOV EXAMPLES 86

Definition 148 (Stable Distributions) A distribution ν is stable if, for any

Lévy process X where X(1) ∼ ν, X is self-similar.

Theorem 149 (Scaling in Stable Lévy Processes) In a self-similar Lévy

process, the scaling function h(t) must take the form tα for some real α, the
index of stability.

Proof: Exercise 11.7. 

It turns out that there are analogs to the functional central limit theorem
for a broad range of self-similar processes, especially ones with stable increment
distributions; Embrechts and Maejima (2002) is a good short introduction to
this topic, and some of its statistical implications.

11.4 Exercises
Exercise 11.1 (Wiener Process with Constant Drift) Consider a process
X(0) which, like the Wiener process, has X(0) = 0 and independent increments,
but where X(t2 ) − X(t1 ) ∼ N (a(t2 − t1 ), σ 2 (t2 − t1 )). a is called the drift rate
and σ 2 the diffusion constant. Show that X(t) is a Markov process, following
the argument for the standard Wiener process (a = 0, σ 2 = 1) above. Do such
processes have continuous modifications for all (finite) choices of a and σ 2 ? If
so, prove it; if not, give at least one counter-example.

Exercise 11.2 (Perron-Frobenius Operators) Verify that P defined in the

section on the logistic map above is a Markov operator.

Exercise 11.3 (Continuity of the Wiener Process) Show that the Wiener
process has continuous sample paths, using its finite-dimensional distributions
and Theorem 97.

Exercise 11.4 (Independent Increments with Respect to a Filtration)

Let X be adapted to a filtration {Ft }. Then X has independent increments with
respect to {Ft } when Xt − Xs is independent of Fs for all s ≤ t. Show that X
is Markovian with respect to {Ft }, by analogy with Theorem 141.

Exercise 11.5 (Poisson Counting Process) Consider the Poisson counting

process N of Example 140.
1. Prove that N is continuous in probability.
2. Prove that N has a cadlag modification. Hint: Theorem 96 is one way,
but there are others.
CHAPTER 11. MARKOV EXAMPLES 87

Exercise 11.6 (Poisson Distribution is Infinitely Divisible) Prove, di-

rectly from the definition, that every Poisson distribution is infinitely divisible.
What is the corresponding stationary, independent-increment process?

Exercise 11.7 (Self-Similarity in Lévy Processes) Prove Theorem 149.

Exercise 11.8 (Gaussian Stability) Find the index of stability a standard

Gaussian distribution.

Exercise 11.9 (Poissonian Stability?) Is the standard Poisson distribution

stable? If so, prove it, and find the index of stability. If not, prove that it is
not.

Exercise 11.10 (Lamperti Transformation) Prove Lamperti’s Theorem: If

Y is a strictly stationary process and α > 0, then X(t) = tα Y (log t) is self-
similar, and if X is self-similar with index of stability α, then Y (t) = e−αt X(et )
is strictly stationary. These operations are called the Lamperti transformations,
after Lamperti (1962).
Chapter 12

Generators of Markov
Processes

This lecture is concerned with the infinitessimal generator of a

Markov process, and the sense in which we are able to write the evo-
lution operators of a homogeneous Markov process as exponentials
of their generator.
Take our favorite continuous-time homogeneous Markov process, and con-
sider its semi-group of time-evolution operators Kt . They obey the relationship
Kt+s = Kt Ks . That is, composition of the operators corresponds to addition of
their parameters, and vice versa. This is reminiscent of the exponential func-
tions on the reals, where, for any k ∈ R, k (t+s) = k t k s . In the discrete-parameter
t
case, in fact, Kt = (K1 ) , where integer powers of operators are defined in the
obvious way, through iterated composition, i.e., K 2 f = K ◦ (Kf ). It would
be nice if we could extend this analogy to continuous-parameter Markov pro-
cesses. One approach which suggests itself is to notice that, for any k, there’s
another real number g such that k t = etg , and that etg has a nice representation
involving integer powers of g:
∞
X (tg)i
etg =
i=0
i!

The strategy this suggests is to look for some other operator G such that
∞ i i
tG
X tG
Kt = e ≡
i=0
i!

Such an operator G is called the generator of the process, and the purpose of this
chapter is to work out the conditions under which this analogy can be carried
through.
In the exponential function
 case, we notice that g can be extracted by taking
d tg 
the derivative at zero: dt e t=0 = g. This suggests the following definition.

88
CHAPTER 12. GENERATORS 89

Definition 150 (Infinitessimal Generator) Let Kt be a continuous-parameter

semi-group of linear operators on L, where L is a normed linear vector space,
e.g., Lp for some 1 ≤ p ≤ ∞. Say that a function f ∈ L belongs to Dom(G) if
the limit
Kh f − K0 f
lim ≡ Gf (12.1)
h↓0 h
exists in an L-norm sense (Definition 151). The operator G defined through Eq.
12.1 is called the infinitessimal generator of the semi-group Kt .

Definition 151 (Limit in the L-norm sense) Let L be a normed vector

space. We say that a sequence of elements fn ∈ L has a limit f in L when

lim kfn − f k = 0 (12.2)

n→∞

This definition extends in the natural way to continuously-indexed collections of

elements.

Lemma 152 (Generators are Linear) For every semi-group of homogeneous

transition operators Kt , the generator G is a linear operator.

Proof: Exercise 12.1. 

Lemma 153 (Invariant Distributions of a Semi-group Belong to the

Null Space of Its Generator) If ν is an invariant distribution of a semi-group
of Markov operators Mt with generator G, then Gν = 0.

Proof: Since ν is invariant, Mt ν = ν for all t (Theorem 136). Hence

Mt ν − ν = Mt ν − M0 ν = 0, so, applying the definition of the generator (Eq.
12.1), Gν = 0. 
Remark: The converse assertion, that Gν = 0 implies ν is invariant under
Mt , requires extra conditions.
There is a conjugate version of this lemma.

Lemma 154 (Invariant Distributions and the Generator of the Time-

Evolution Semigroup) If ν is an invariant distribution of a Markov process,
and the time-evolution semi-group Kt is generated by G, then, ∀f ∈ Dom(G),
νGf = 0.

Proof: Since ν is invariant, νKt = ν for all t, hence νKt f = νf for all
t ≥ 0 and all f . Since taking expectations with respect to a measure is a linear
operator, ν(Kt f − f ) = 0, and obviously then (Eq. 12.1) νGf = 0. 
Remark: Once again, νGf = 0 for all f is not enough, in itself, to show that
ν is an invariant measure.
You will usually see the definition of the generator written with f instead
of K0 f , but I chose this way of doing it to emphasize that G is, basically,
CHAPTER 12. GENERATORS 90

the derivative at zero, that G = dK/dt|t=0 . Recall, from calculus, that the
d t
exponential function can k t be defined by the fact that dt k ∝ k t (and e can
be defined as the k such that the constant of proportionality is 1). As part of
our program, we will want to extend this differential point of view. The next
lemma builds towards it, by showing that if f ∈ Dom(G), then Kt f is too.

Lemma 155 (Operators in a Semi-group Commute with Its Genera-

tor) If G is the generator of the semi-group Kt , and f is in the domain of G,
then Kt and G commute, for all t:
Kt0 f − Kt f
Kt Gf = lim (12.3)
t0 →tt0 − t
= GKt f (12.4)

Proof: Exercise 12.2. 

Definition 156 (Time Derivative in Function Space) For every t ∈ T , let

u(t, x) be a function in L. When the limit

u(t, x) − u(t0 , x)
u0 (t0 , x) = lim (12.5)
t→t0 t − t0
exists in the L sense, then we say that u0 (t0 ) is the time derivative or strong
derivative of u(t) at t0 .

Lemma 157 (Generators and Derivatives at Zero) Let Kt be a homo-

geneous semi-group of operators with generator G. Let u(t) = Kt f for some
f ∈ Dom(G). Then u(t) is differentiable at t = 0, and its derivative there is
Gf .

Proof: Obvious from the definitions. 

Theorem 158 (The Derivative of a Function Evolved by a Semi-Group)

Let Kt be a homogeneous semi-group of operators with generator G, and let
u(t, x) = (Kt f )(x), for fixed f ∈ Dom(G). Then u0 (t) exists for all t, and is
equal to Gu(t).

Proof: Since f ∈ Dom(G), Kt Gf exists, but then, by Lemma 155, Kt Gf =

GKt f = Gu(t), so u(t) ∈ Dom(G) for all t. Now let’s consider the time deriva-
tive of u(t) at some arbitrary t0 , working from above:

(u(t) − u(t0 ) Kt−t0 u(t0 ) − u(t0 )

= (12.6)
t − t0 t − t0
Kh u(t0 ) − u(t0 )
= (12.7)
h
Taking the limit as h ↓ 0, we get that u0 (t0 ) = Gu(t0 ), which exists, because
u(t0 ) ∈ Dom(G). 
CHAPTER 12. GENERATORS 91

Corollary 159 (Initial Value Problems in Function Space) u(t) = Kt f ,

f ∈ Dom(G), solves the initial value problem u(0) = f , u0 (t) = Gu(t).
Proof: Immediate from the theorem. 
Remark: Such initial value problems are sometimes called Cauchy problems,
especially when G takes the form of a differential operator.
Corollary 160 (Derivative of Conditional Expectations of a Markov
Process) Let X be a homogeneous Markov process whose time-evolution op-
erators are Kt , with generator G. If f Dom(G), then its condition expectation
E [f (Xs )|Xt ] has strong derivative Gu(t).
Proof: An immediate application of the theorem.
We are now almost ready to state the sense in which Kt is the result of
exponentiating G. This is given by the remarkable Hille-Yosida theorem, which
in turn involves a family of operators related to the time-evolution operators,
the “resolvents”, again built by analogy to the exponential functions, and to
Laplace transforms.
Recall that the Laplace transform of a function f : R 7→ R is another func-
tion, f˜, defined by Z ∞
f˜(λ) ≡ e−λt f (t)dt
0
for positive λ. Laplace transforms arise in many contexts (linear systems theory,
integral equations, etc.), one of which is moment-generating functions in basic
probability theory. If Y is a real-valued random variable with probability law
P , then the moment-generating function is
Z Z
 λY 
MY (λ) ≡ E e = e dP = eλy p(y)dy
λy

when the density in the last expression exists. You may recall, from this context,
that the distributions of well-behaved random variables are completely specified
by their moment-generating functions; this is actually a special case of a more
general result about when functions are uniquely described by their Laplace
transforms, i.e., when f can be expressed uniquely in terms of f˜. This is im-
portant to us, because it turns out that the Laplace transform, so to speak, of a
semi-group of operators is better-behaved than the semi-group itself, and we’ll
want to say when we can use the Laplace transform to recover the semi-group.
The analogy with exponential functions, again, is a key. Notice that, for any
positive constant λ,
Z ∞
1
e−λt etg dt = (12.8)
t=0 λ − g
from which we could recover g, as that value of λ for which the Laplace transform
is singular. In our analogy, we will want to take the Laplace transform of the
semi-group. Just as f˜(λ) is another real number, the Laplace transform of a
semi-group of operators is going to be another operator. We will want that to
be the inverse operator to λ − G.
CHAPTER 12. GENERATORS 92

Definition 161 (Resolvents) Given a continuous-parameter homogeneous semi-

group Kt , for each λ > 0, the resolvent operator or resolvent Rλ is the Laplace
transform of Kt : for every f ∈ L,
Z ∞
(Rλ f )(x) ≡ e−λt (Kt f )(x)dt (12.9)
t=0

Remark 1: Think of Kt as a function from the real numbersRto the linear

∞
operators on L. Symbolically, its Laplace transform would be 0 e−λt Kt dt.
The equation above just fills in the content of that symbolic expression.
Remark 2: The name “resolvent”, like some of the other ideas an termi-
nology of operator semi-groups, comes from the theory of integral equations;
invariant densities (when they exist) are solutions of homogeneous linear Fred-
holm integral equations of the second kind. Rather than pursue this connection,
or even explain what that phrase means, I will refer you to the classic treatment
of integral equations by Courant and Hilbert (1953, ch. 3), which everyone else
seems to follow very closely.
Remark 3: When the function f is a value (loss, benefit, utility, ...) function,
(Kt f )(x) is the expected value at time t when starting the process in state x.
(Rλ f )(x) can be thought of as the net present expected value when starting at
x and applying a discount rate λ.

Definition 162 (Yosida Approximation of Operators) The Yosida ap-

proximation to a semi-group Kt with generator G is given by
(λ) (λ)
Kt ≡ etG (12.10)
G(λ) ≡ λ(λRλ − I) = λGRλ (12.11)

The domain of G(λ) is all of L, not just Dom(G).

Theorem 163 (Hille-Yosida Theorem) Let G be a linear operator on some

linear subspace D of L. G is the generator of a continuous semi-group of con-
tractions Kt if and only if
1. D is dense in L;
2. For every f ∈ L and λ > 0, there exists a unique g ∈ D such that λg−Gg =
f;

3. For every g ∈ D and positive λ, kλg − Ggk ≥ λkgk.

Under these conditions, the resolvents of Kt are given by Rλ = (λI − G)−1 , and
Kt is the limit of the Yosida approximations as λ → ∞:

Kt f = lim Ktλ f , ∀f ∈ L (12.12)

λ→∞
CHAPTER 12. GENERATORS 93

Proof: See Kallenberg, Theorem 19.11. 

Remark 1: Other good sources on the Hille-Yosida theorem include Ethier
and Kurtz (1986, sec. 1.2), and of course Hille’s own book (Hille, 1948, ch. XII).
I have not read Yosida’s original work.
Remark 2: The point of condition (1) in the theorem is that for any f ∈ L,
we can chose fn ∈ D such that fn → f . Then, even if Gf is not directly defined,
Gfn exists for each n. Because G is linear and therefore continuous, Gfn goes
to a limit, which we can chose to write as Gf . Similarly for G2 , G3 , etc. Thus
we can define etG f as
∞ j j
X t G fn
lim
n→∞
j=0
j!

assuming all of the sums inside the limit converge, which is yet to be shown.
Remark 3: The point of condition (2) in the theorem is that, when it holds,
−1
(λI − G) is well-defined, i.e., there is an inverse to the operator λI − G. This
is, recall, what we would like the resolvent to be, if the analogy with exponential
functions is to hold good.
Remark 4: If we start from the semi-group Kt and obtain its generator G, the
theorem tells us that it satisfies properties (1)–(3). If we start from an operator
G and see that it satsifies (1)–(3), the theorem tells us that it generates some
semi-group. It might seem, however, that it doesn’t tell us how to construct that
semi-group, since the Yosida approximation involves the resolvent Rλ , which is
defined in terms of the semi-group, creating an air of circularity. In fact, when
−1
we start from G, we decree Rλ to be (λI − G) . The Yosida approximations
then is defined in terms of G and λ alone.

Corollary 164 (Stochastic Approximation of Initial Value Problems)

Let u(0) = f , u0 (t) = Gu(t) be an initial value problem in L. Then a stochastic
approximation to u(t) can be founded by taking
n
1X
û(t) = f (Xi (t)) (12.13)
n i=1

where the Xi are independent copies of the Markov process corresponding to the
semi-group Kt generated by G, with initial condition Xi (0) = x.

Proof: Combine corollaries. 

12.1 Exercises
Exercise 12.1 (Generators are Linear) Prove Lemma 152.

Exercise 12.2 (Semi-Groups Commute with Their Generators) Prove

Lemma 155.
CHAPTER 12. GENERATORS 94

1. Prove Equation 12.3, restricted to t0 ↓ t instead of t0 → t. Hint: Write Tt

in terms of an integral over the corresponding transition kernel, and find
a reason to exchange integration and limits.
2. Show that the limit as t0 ↑ t also exists, and is equal to the limit from
above. Hint: Re-write the quotient inside the limit so it only involves
positive time-differences.
3. Prove Equation 12.4.

Exercise 12.3 (Generator of the Poisson Counting Process) Find the

generator of the time-evolution operators of the Poisson counting process (Ex-
ample 140).
Chapter 13

The Strong Markov

Property and Martingale
Problems

Section 13.1 introduces the strong Markov property — indepen-

dence of the past and future conditional on the state at random
(optional) times. It includes an example of a Markov process which
is not strongly Markovian.
Section 13.2 describes “the martingale problem for Markov pro-
cesses”, explains why it would be nice to solve the martingale prob-
lem, and how solutions are strong Markov processes.

13.1 The Strong Markov Property

A process is Markovian, with respect to a filtration {Ft }, if for any fixed time t,
the future of the process is independent of Ft given Xt . This is not necessarily
the case for a random time τ , because there could be subtle linkages between
the random time and the evolution of the process. If these can be ruled out, we
have a strong Markov process.

Definition 165 (Strongly Markovian at a Random Time) Let X be a

Markov process with respect to a filtration {Ft }, with transition kernels µt,s and
evolution operators Kt,s . Let τ be an {Ft }-optional time which is almost surely
finite. Then X is strongly Markovian at τ when either of the two following
(equivalent) conditions hold

P (Xt+τ ∈ B|Fτ ) = µτ,τ +t (Xτ , B) (13.1)

E [f (Xτ +t )|Fτ ] = (Kτ,τ +t f )(Xτ ) (13.2)

for all t ≥ 0, B ∈ X and bounded measurable functions f .

95
CHAPTER 13. STRONG MARKOV, MARTINGALES 96

Definition 166 (Strong Markov Property) If X is Markovian with respect

to {Ft }, and strongly Markovian at every {Ft }-optional time which is almost
surely finite, then it is a strong Markov process with respect to {Ft }.
If the index set T is discrete, then the strong Markov property is implied
by the ordinary Markov property (Definition 102). If time is continuous, this
is not necessarily the case. It is generally true that, if X is Markov and τ
takes on only countably many values, X is strongly Markov at τ (Exercise 13.1).
In continuous time, however, the Markov property does not imply the strong
Markov property.
Example 167 (A Markov Process Which Is Not Strongly Markovian)
(After Fristedt and Gray (1997, pp. 626–627).) We will construct an R2 -valued
Markov process on [0, ∞) which is not strongly Markovian. Begin by defining
the following map from R to R2 :

 (w, 0) w≤0
f (w) = (sin w, 1 − cos w) 0 < w < 2π (13.3)
(w − 2π, 0) w ≥ 2π


When w is less than zero or above 2π, f (w) moves along the x axis of the plane;
in between, it moves along a circle of radius 1, centered at (0, 1), which it enters
and leaves at the origin. Notice that f is invertible everywhere except at the
origin, which is ambiguous between w = 0 and w = 2π.
Let X(t) = f (W (t) + π), where W (t) is a standard Wiener process. At
all t, P (W (t) + π = 0) = P (W (t) + π = 2π) = 0, so, with probability 1, X(t)
can be inverted to get W (t). Since W (t) is a Markov
process, it follows that
P (X(t + h) ∈ B|X(t) = x) = P X(t + h) ∈ B|FtX almost surely, i.e., X is
Markov. Now consider τ = inf t X(t) = (0, 0), the hitting time of the origin.
This is clearly an F X -optional time, and equally clearly almost surely finite,
because, with probability 1, W (t) will leave the interval (−π, π) within a finite
time. But, equally clearly, the future behavior of X will be very different if it hits
the origin because W = π or because W = −π, which cannot be determined just
from X. Hence, there is at least one optional time at which X is not strongly
Markovian, so X is not a strong Markov process.
Since we often want to condition on the state of the process at random times,
we would like to find conditions under which a process is strongly Markovian
for all optional times.

13.2 Martingale Problems

One approach to getting strong Markov processes is through martingales, and
more specifically through what is known as the martingale problem.
Notice the following consequence of Theorem 158:
Z t
Kt f (x) − f (x) = Ks Gf (x)ds (13.4)
0
CHAPTER 13. STRONG MARKOV, MARTINGALES 97

for any t ≥ 0 and f ∈ Dom(G). The relationship between Kt f and the condi-
tional expectation of f suggests the following definition.

Definition 168 (Martingale Problem) Let Ξ be a Polish space, D a class

of bounded, continuous, real-valued functions on Ξ, and G an operator from D
to bounded, measurable functions on Ξ. A Ξ-valued stochastic process on R+ is
a solution to the martingale problem for G and D if, for all f ∈ D,
Z t
f (Xt ) − Gf (Xs )ds (13.5)
0
 X

is a martingale with respect to Ft , the natural filtration of X.

Proposition 169 (Cadlag Nature of Functions in Martingale Prob-

lems) Suppose X is a cadlag solution to the martingale problem for G, D. Then
for any f ∈ D, the stochastic process given by Eq. 13.5 is also cadlag.

Proof: Follows from the assumption that f is continuous. 

Lemma 170 (Alternate Formulation of Martingale Problem) X is a

solution to the martingale problem for G, D if and only if, for all t, s ≥ 0,
Z t+s 
X X
 
E f (Xt+s )|Ft − E Gf (Xu )du|Ft = f (Xt ) (13.6)
t

Proof: Take the definition of a martingale and re-arrange the terms in Eq.
13.5. 
Martingale problems are important because of the two following theorems
(which can both be refined considerably).

Theorem 171 (Markov Processes Solve Martingale Problems) Let X

be a homogeneous Markov process with generator G and cadlag sample paths,
and let D be the continuous functions in Dom(G). Then X solves the martingale
problem for G, D.

Proof: Exercise 13.2. 

Theorem 172 (Solutions to the Martingale Problem are Strongly Marko-

vian) Suppose that for each x ∈ Ξ, there is a unique cadlag solution to the
martingale problem for G, D such that X0 = x. Then the collection of these
solutions is a homogeneous strong Markov family X, and the generator is equal
to G on D.

Proof: Exercise 13.3. 

The main use of Theorem 171 is that it lets us prove convergence of some
functions of Markov processes, by showing that they can be cast into the form of
Eq. 13.5, and then applying the martingale convergence devices. The other use
is in conjunction with Theorem 172. We will often want to show that a sequence
CHAPTER 13. STRONG MARKOV, MARTINGALES 98

of Markov processes converges on a limit which is, itself, a Markov process. One
approach is to show that the terms in the sequence solve martingale problems
(via Theorem 171), argue that then the limiting process does too, and finally
invoke Theorem 172 to argue that the limiting process must itself be strongly
Markovian. This is often much easier than showing directly that the limiting
process is Markovian, much less strongly Markovian. Theorem 172 itself is often
a convenient way of showing that the strong Markov property holds.

13.3 Exercises
Exercise 13.1 (Strongly Markov at Discrete Times) Let X be a homoge-
neous Markov process with respect to a filtration {Ft } and τ be an {Ft }-optional
time. Prove that if P (τ < ∞) = 1, and τ takes on only countably many values,
then X is strongly Markovian at τ . (Note: the requirement that X be homo-
geneous can be lifted, but requires some more technical machinery I want to
avoid.)

Exercise 13.2 (Markovian Solutions of the Martingale Problem) Prove

Theorem 171. Hints: Use Lemma 170, bounded convergence, and Theorem 158.

Exercise 13.3 (Martingale Solutions are Strongly Markovian) Prove

Theorem 172. Hint: use the Optional Sampling Theorem (from 36-752, or from
chapter 7 of Kallenberg).
Chapter 14

Feller Processes

Section 14.1 makes explicit the idea that the transition kernels
of a Markov process induce a kernel over sample paths, mostly to
fix notation for later use.
Section 14.2 defines Feller processes, which link the cadlag and
strong Markov properties.

14.1 Markov Families

We have been fairly cavalier about the idea of a Markov process having a par-
ticular initial state or initial distribution, basically relying on our familiarity
with these ideas from elementary courses on stochastic processes. For future
purposes, however, it is helpful to bring this notions formally within our general
framework, and to fix some notation.

Definition 173 (Initial Distribution, Initial State) Let Ξ be a Borel space

with σ-field X , T be a one-sided index set, and µt,s be a collection of Markovian
transition kernels on Ξ. Then the Markov process with initial distribution ν,
Xν , is the Markov process whose finite-dimensional distributions are given by
the action of µt,s on ν. That is, for all 0 ≤ t1 ≤ t2 ≤ . . . ≤ tn ,

Xν (0), Xν (t1 ), Xν (t2 ), . . . Xν (tn ) ∼ ν ⊗ µ0,t1 ⊗ µt1 ,t2 ⊗ . . . ⊗ µtn−1 ,t(14.1)

n

If ν = δ(x − a), the delta distribution at a, then we write Xa and call it the
Markov process with initial state a.

The existence of processes with given initial distributions and initial states
is a trivial consequence of Theorem 106, our general existence result for Markov
processes.

Lemma 174 (Kernel from Initial States to Paths) For every initial state
x, there is a probability distribution Px on ΞT , X T . The function Px (A) : Ξ ×
X T → [0, 1] is a probability kernel.

99
CHAPTER 14. FELLER PROCESSES 100

Proof: The initial state fixes all the finite-dimensional distributions, so

the existence of the probability distribution follows from Theorem 23. The fact
that Px (A) is a kernel is a straightforward application of the definition of kernels
(Definition 30). 

Definition 175 (Markov Family) The Markov family corresponding to a

given set of transition kernels µt,s is the collection of all Px .

That is, rather than thinking of a different stochastic process for each initial
state, we can simply think of different distributions over the path space ΞT .
This suggests the following definition.

Definition 176 (Mixtures of Path Distributions (Mixed States)) For

a given initial distribution ν on Ξ, we define a distribution on the paths in a
Markov family as, ∀A ∈ X T ,
Z
Pν (A) ≡ Px (A)ν(dx) (14.2)
Ξ

In physical contexts, we sometimes refer to distributions ν as mixed states,

as opposed to the pure states x, because the path-space distributions induced by
the former are mixtures of the distributions induced by the latter. You should
check that the distribution over paths given by a Markov process with initial
distribution ν, according to Definition 173, agrees with that given by Definition
176.

14.2 Feller Processes

Working in the early 1950s, Feller showed that, by imposing very reasonable
conditions on the semi-group of evolution operators corresponding to a homo-
geneous Markov process, one could obtain very powerful results about the near-
continuity of sample paths (namely, the existence of cadlag versions), about the
strong Markov property, etc. Ever since, processes with such nice semi-groups
have been known as Feller processes, or sometimes as Feller-Dynkin processes,
in recognition of Dynkin’s work in extending Feller’s original approach. (This
is also recognized in the name of Dynkin’s Formula, Theorem 195 below.) Un-
fortunately, to first order there are as many definitions of a Feller semi-group as
there are books on Markov processes. I am going to try to follow Kallenberg as
closely as possible, because his version is pretty clearly motivated, and you’ve
already got it.

Definition 177 (Feller Process) A continuous-time homogeneous Markov

family X is a Feller process when, for all x ∈ Ξ,
d
∀t, y → x ⇒ Xy (t) → Xx (t) (14.3)
P
t → 0 ⇒ Xx (t) → x (14.4)
CHAPTER 14. FELLER PROCESSES 101

Remark 1: The first property basically says that the dynamics are a smooth
function of the initial state. Recall1 that if we have an ordinary differential
equation, dx/dt = f (x), and the function f is reasonably well-behaved, the ex-
istence and uniqueness theorem tells us that there is a function x(t, x0 ) satisfying
the equation, and such that x(0, x0 ) = x0 . Moreover, x(t, x0 ) is a continuous
function of x0 for all t. The first Feller property is one counterpart of this for
stochastic processes. This is a very natural assumption to make in physical or
social modeling, that very similar initial conditions should lead to very similar
developments.
Remark 2: The second property imposes a certain amount of smoothness
on the trajectories themselves, and not just how they depend on the initial
state. It’s a pretty well-attested fact about the world that teleportation does
not take place, and that its state changes in a reasonably continuous manner
(“natura non facit saltum”, as they used to say). However, the second Feller
property does not require that the sample paths actually be continuous. We
will see below that they are, in general, merely cadlag. So a certain limited
amount of teleportation, or salti, is allowed after all. We do not think this
actually happens, but it is a convenience when using a discrete set of states to
approximate a continuous variable.
In developing the theory of Feller processes, we will work mostly with the
time-evolution operators, acting on observables, rather than the Markov oper-
ators, acting on distributions. This is traditional in this part of the theory, as
it seems to let us get away with less technical
R machinery, in effect because the
norm supx |f (x)| is stronger than the norm |f (x)|dx. Of course, because the
two kinds of operators are adjoint, you can work out everything for the Markov
operators, if you want to.
As usual, we warm up with some definitions. The first two apply to operators
on any normed linear space L, which norm we generically write as k · k. The
second two apply specifically when L is a space of real-valued functions on some
Ξ, such as Lp , p from 1 to ∞ inclusive.

Definition 178 (Contraction Operator) An operator A is an L-contraction

when kAf k ≤ kf k.

Definition 179 (Strongly Continuous Semigroup) A semigroup of oper-

ators At is strongly continuous in the L sense on a set of functions D when,
∀f ∈ D

lim kAt f − f k = 0 (14.5)

t→0

Definition 180 (Positive Operator) An operator A on a function space L

is positive when f ≥ 0 a.e. implies Af ≥ 0 a.e.

1 Or read Arnol’d (1973), if memory fails you.

CHAPTER 14. FELLER PROCESSES 102

Definition 181 (Conservative Operator) An operator A is conservative

when A1Ξ = 1Ξ .
In these terms, our earlier Markov operators are linear, positive, conservative
contractions, either on L1 (µ) (for densities) or M(Ξ) (for measures).
Lemma 182 (Continuous semi-groups produce continuous paths in
function space) If At is a strongly continuous semigroup of linear contrac-
tions on L, then, for each f ∈ L, At f is a continuous function of t.
Proof: Continuity here means that limt0 →t kAt0 f − At k = 0 — we are using
the norm k · k to define our metric in function space. Consider first the limit
from above:
kAt+h f − At f k = kAt (Ah f − f )k (14.6)
≤ k|Oh f − f k (14.7)
since the operators are contractions. Because they are strongly continuous,
kAh f − f k can be made smaller than any  > 0 by taking h sufficiently small.
Hence limh↓0 At+h f exists and is At f . Similarly, for the limit from below,
kAt−h f − At f k = kAt f − At−h f k (14.8)
= kAt−h (Ah f − f )k (14.9)
≤ kAh f − f k (14.10)
using the contraction property again. So limh↓0 At−h f = At f , also, and we can
just say that limt0 →t At0 f = At f . 
Remark: The result actually holds if we just assume strong continuity, with-
out contraction, but the proof isn’t so pretty; see Ethier and Kurtz (1986, ch.
1, corollary 1.2, p. 7).
There is one particular function space L we will find especially interesting.
Definition 183 (Continuous Functions Vanishing at Infinity) Let Ξ be
a locally compact and separable metric space. The class of functions C0 will
consist of functions f : Ξ 7→ R which are continuous and for which kxk → ∞
implies f (x) → 0. The norm on C0 is supx |f (x)|.

Definition 184 (Feller Semigroup) A semigroup of linear, positive, conser-

vative contraction operators Kt is a Feller semigroup if, for every f ∈ C0 and
x ∈ Ξ, (Definition 183),
Kt f ∈ C0 (14.11)
lim Kt f (x) = f (x) (14.12)
t→0

Remark: Some authors omit the requirement that Kt be conservative. Also,

this is just the homogeneous case, and one can define inhomogeneous Feller
semigroups. However, the homogeneous case will be plenty of work enough for
us!
You can guess how Feller semi-groups relate to Feller processes.
CHAPTER 14. FELLER PROCESSES 103

Lemma 185 (The First Pair of Feller Properties) Eq. 14.11 holds if and
only if Eq. 14.3 does.

Proof: Exercise 14.2. 

Lemma 186 (The Second Pair of Feller Properties) Eq. 14.12 holds if
and only if Eq. 14.4 does.

Proof: Exercise 14.3. 

Theorem 187 (Feller Processes and Feller Semigroups) A Markov pro-

cess is a Feller process if and only if its evolution operators form a Feller semi-
group.

Proof: Combine Lemmas 185 and 186. 

Feller semigroups in continuous time have generators, as in Chapter 12. In
fact, the generator is especially useful for Feller semigroups, as seen by this
theorem.

Theorem 188 (Generator of a Feller Semigroup) If Kt and Ht are Feller

semigroups with generator G, then Kt = Ht .

Proof: Because Feller semigroups consist of contractions, the Hille-Yosida

−1
Theorem 163 applies, and, for every positive λ, the resolvent Rλ = (λI − G) .
Hence, if Kt and Ht have the same generator, they have the same resolvent
operators. But this means that, for every f ∈ C0 and x, Kt f (x) and Ht f (x) have
the same Laplace transforms. Since, by Eq. 14.12 Kt f (x) and Ht f (x) are both
right-continuous, their Laplace transforms are unique, so Kt f (x) = Ht f (x). 

Theorem 189 (Feller Semigroups are Strongly Continuous) Every Feller

semigroup Kt with generator G is strongly continuous on Dom(G).

Proof: From Corollary 159, we have, as seen in Chapter 13, for all t ≥ 0,
Z t
Kt f − f = Ks Gf ds (14.13)
0

Clearly, the right-hand side goes to zero as t → 0. 

The two most important properties of Feller processes is that they are cadlag
(or, rather, always have cadlag versions), and that they are strongly Markovian.
First, let’s look at the cadlag property. We need a result which I really should
have put in Chapter 8.

Proposition 190 (Cadlag Modifications Implied by a Kind of Modulus

of Continuity) Let Ξ be a locally compact, separable metric space with metric
ρ, and let X be a separable Ξ-valued stochastic process on T . For given , δ > 0,
define α(, δ) to be

P ω : ρ(X(s, ω), X(t, ω)) ≥ , ω ∈ Γ|FsX (14.14)

inf sup
Γ∈FsX : P(Γ)=1 s,t∈T : s≤t≤s+δ
CHAPTER 14. FELLER PROCESSES 104

If, for all ,

lim α(, δ) = 0 (14.15)

δ→0

then X has a cadlag version.

Proof: Combine Theorem 2 and Theorem 3 of Gikhman and Skorokhod

(1965/1969, Chapter IV, Section 4). 

Lemma 191 (Markov Processes Have Cadlag Versions When They

Don’t Move Too Fast (in Probability)) Let X be a separable homogeneous
Markov process. Define

α(, δ) = sup P (ρ(Xx (t), x) ≥ ) (14.16)

t∈T : 0≤t≤δ; x∈Ξ

If, for every  > 0,

lim α(, δ) = 0 (14.17)

δ→0

then X has a cadlag version.

Proof: The α in this lemma is clearly the α in the preceding proposition

(190), using the fact that X is Markovian with respect to its natural filtration
(Theorem 112) and homogeneous. 

Lemma 192 (Markov Processes Have Cadlag Versions If They Don’t

Move Too Fast (in Expectation)) A separable homogeneous Markov process
X has a cadlag version if

lim sup E [ρ(Xx (t), x)] = 0 (14.18)

δ↓0 x∈Ξ, 0≤t≤δ

Proof: Start with the Markov inequality.

E [ρ(Xx (t), x)]

∀x, t > 0,  > 0, P (ρ(Xx (t), x) ≥ ) ≤ (14.19)

E [ρ(Xx (t), x)]
∀x, δ > 0,  > 0, sup P (ρ(Xx (t), x) ≥ ) ≤ sup (14.20)
0≤t≤δ 0≤t≤δ 
1
∀δ > 0,  > 0, sup P (ρ(Xx (δ), x) ≥ ) ≤ sup E [ρ(Xx (δ),(14.21)
x)]
x, 0≤t≤δ  x, 0≤t≤δ

Taking the limit as δ ↓ 0, we have, for all  > 0,

1
lim α(, δ) ≤ lim sup E [ρ(Xx (δ), x)] = 0 (14.22)
δ→0  δ↓0 x, 0≤t≤δ

So the preceding lemma (191) applies. 

CHAPTER 14. FELLER PROCESSES 105

Theorem 193 (Feller Implies Cadlag) Every Feller process X has a cadlag
version.

Proof: First, by the usual arguments, we can get a separable version of X.

Next, we want to show that the last lemma is satisfied. Notice that, because Ξ
is compact, limx ρ(xn , x) = 0 if and only if fk (xn ) → fk (x), for all fk in some
countable dense subset of the continuous functions on the state space.2 Since the
Feller semigroup is strongly continuous on the domain of its generator (Theorem
189), and that domain is dense in C0 by the Hille-Yosida Theorem (163), we can
pick our fk to be in this class. The strong continuity is with respect to the C0
norm, so supx |Kt f (x) − Ks f (x)| = supx |Ks (Kt−s f (x) − f (x))| → 0 as t − s →
0, for every f ∈ C0 . But supx |Kt f (x) − Ks f (x)| = supx E [|f (Xx (t)) − f (Xx (s))|].
So supx, 0≤t≤δ E [|f (Xx (t)) − f (x)|] → 0 as δ → 0. Now Lemma 192 applies. 
Remark: Kallenberg (Theorem 19.15, p. 379) gives a different proof, using
the existence of cadlag paths for certain kinds of supermartingales, which he
builds using the resolvent operator. This seems to be the favored approach
among modern authors, but obscures, somewhat, the work which the Feller
properties do in getting the conclusion.

Theorem 194 (Feller Processes are Strongly Markovian) Any Feller

process X is strongly Markovian with respect to F X+ , the right-continuous ver-
sion of its natural filtration.

Proof: The strong Markov property holds if and only if, for all bounded,
continuous functions f , t ≥ 0 and F X+ -optional times τ ,

E f (X(τ + t))|FτX+ = Kt f (X(τ ))

 
(14.23)

We’ll show this holds for arbitrary, fixed choices of f , t and τ . First, we discretize
time, to exploit the fact that the Markov and strong Markov properties coincide
for discrete parameter processes. For every h > 0, set

τh ≡ inf {u ≥ τ : u = kh, k ∈ N} (14.24)

u

Now τh is almost surely finite (because τ is), and τh → τ a.s. as h → 0. We

construct the discrete-parameter sequence Xh (n) = X(nh), n ∈ N. This is a
Markov sequence with respect to the natural filtration, i.e., for every bounded
continuous f and m ∈ N,

E f (Xh (n + m))|FnX
 
= Kmh f (Xh (n)) (14.25)

Since the Markov and strong Markov properties coincide for Markov sequences,
we can now assert that

E f (X(τh + mh))|FτXh
 
= Kmh f (X(τh )) (14.26)
2 Roughly speaking, if f (x ) → f (x) for all continuous functions f , it should be obvious
n
that there is no way to avoid having xn → x. Picking a countable dense subset of functions
is still enough.
CHAPTER 14. FELLER PROCESSES 106

Since τh ≥ τ , FτX ⊆ FτXh . Now pick any set B ∈ FτX+ and use smoothing:
E [f (X(τh + t))1B ] = E [Kt f (X(τh ))1B ] (14.27)
E [f (X(τ + t))1B ] = E [Kt f (X(τ ))1B ] (14.28)
where we let h ↓ 0, and invoke the fact that X(t) is right-continuous (Theorem
193) and Kt f is continuous. Since this holds for arbitrary B ∈ FτX+ , and
Kt f (X(τ )) has to be FτX+ -measurable, we have that
E f (X(τ + t))|FτX+ = Kt f (X(τ ))
 
(14.29)
as required. 
Here is a useful consequence of Feller property, related to the martingale-
problem properties we saw last time.
Theorem 195 (Dynkin’s Formula) Let X be a Feller process with generator
G. Let α and β be two almost-surely-finite F-optional times, α ≤ β. Then, for
every continuous f ∈ Dom(G),
"Z #
β
E [f (X(β)) − f (X(α))] = E Gf (X(t))dt (14.30)
α

Proof: Exercise 14.4. 

Remark: A large number of results very similar to Eq. 14.30 are also called
“Dynkin’s formula”. For instance, Rogers and Williams (1994, ch. III, sec. 10,
pp. 253–254) give that name to three different equations. Be careful about what
people mean!

14.3 Exercises
Exercise 14.1 (Yet Another Interpretation of the Resolvents) Consider
again a homogeneous Markov process with transition kernel µt . Let τ be an
exponentially-distributed random variable with rate λ, independent of X. Show
that E [Kτ f (x)] = λRλ f (x).

Exercise 14.2 (The First Pair of Feller Properties) Prove Lemma 185.
Hint: you may use the fact that, for measures, νt → ν if and only if νt f → νf ,
for every bounded, continuous f .

Exercise 14.3 (The Second Pair of Feller Properties) Prove Lemma 186.

Exercise 14.4 (Dynkin’s Formula) Prove Theorem 195.

Exercise 14.5 (Lévy and Feller Processes) Is every Lévy process a Feller
process? If yes, prove it. If not, provide a counter-example, and try to find a
sufficient condition for a Lévy process to be Feller.
Chapter 15

Convergence of Feller
Processes

This chapter looks at the convergence of sequences of Feller pro-

cesses to a limiting process.
Section 15.1 lays some ground work concerning weak convergence
of processes with cadlag sample paths.
Section 15.2 states and proves the central theorem about the
convergence of sequences of Feller processes.
Section 15.3 examines a particularly important special case, the
approximation of ordinary differential equations by pure-jump Markov
processes.

15.1 Weak Convergence of Processes with Cad-

lag Paths (The Skorokhod Topology)
Recall that a sequence of random variables X1 , X2 , . . . converges in distribution
d
on X, or weakly converges on X, Xn → X, if and only if E [f (Xn )] → E [f (X)],
for all bounded, continuous functions f . This is still true when Xn are ran-
dom functions, i.e., stochastic processes, only now the relevant functions f are
functionals of the sample paths.

Definition 196 (Convergence in Finite-Dimensional Distribution) Ran-

fd
dom processes Xn on T converge in finite-dimensional distribution on X, Xn →
d
X, when, ∀J ∈ Fin(T ), Xn (J) → X(J).

Lemma 197 (Finite and Infinite Dimensional Distributional Conver-

gence) Convergence in finite-dimensional distribution is necessary but not suf-
ficient for convergence in distribution.

107
CHAPTER 15. CONVERGENCE OF FELLER PROCESSES 108

Proof: Necessity is obvious: the coordinate projections πt are continuous

functionals of the sample path, so they must converge if the distributions con-
verge. Insufficiency stems from the problem that, even if a sequence of Xn all
have sample paths in some set U , the limiting process might not: recall our
example (79) of the version of the Wiener process with unmeasurable suprema.


Definition 198 (The Space D) By D(T, Ξ) we denote the space of all cadlag
functions from T to Ξ. By default, D will mean D(R+ , Ξ).

D admits of multiple topologies. For most purposes, the most convenient one
is the Skorokhod topology, a.k.a. the J1 topology or the Skorokhod J1 topology,
which makes D(Ξ) a complete separable metric space when Ξ is itself complete
and separable. (See Appendix A2 of Kallenberg.) For our purposes, we need
only the following notion and propositions.

Definition 199 (Modified Modulus of Continuity) The modified modulus

of continuity of a function x ∈ D(T, Ξ) at time t ∈ T and scale h > 0 is given
by

w(x, t, h) ≡ inf max sup ρ(x(s), x(r)) (15.1)

(Ik ) k r,s∈Ik

where the infimum is over partitions of [0, t) into half-open intervals whose length
is at least h (except possibly for the last one). Because x is cadlag, for fixed x
and t, w(x, t, h) → 0 as h → 0.

Proposition 200 (Weak Convergence in D(R+ , Ξ)) Let Ξ be a complete,

separable metric space. Then a sequence of random functions X1 , X2 , . . . ∈
D(R+ , Ξ) converges in distribution to X ∈ D if and only if
i The set Tc = {t ∈ T : X(t) = X(t− )} has a countable dense subset T0 ,
and the finite-dimensional distributions of the Xn converge on those of X
on T0 .
ii For every t,

lim lim sup E [w(Xn , t, h) ∧ 1] = 0 (15.2)

h→0 n→∞

Proof: See Kallenberg, Theorem 16.10, pp. 313–314. 

Proposition 201 (Sufficient Condition for Weak Convergence) The fol-

lowing three conditions are all equivalent, and all imply condition (ii) in Propo-
sition 200.
1. For any sequence of a.s.-finite F Xn -optional times τn and positive con-
stants hn → 0,
P
ρ(Xn (τn ), Xn (τn + hn )) → 0 (15.3)
CHAPTER 15. CONVERGENCE OF FELLER PROCESSES 109

2. For all t > 0, for all

lim lim sup sup E [ρ(Xn (σ), Xn (τ )) ∧ 1] = 0 (15.4)

h→0 n→∞ σ,τ

where σ and τ are F Xn -optional times σ, τ ≤ t, with σ ≤ τ ≤ τ + h.

3. For all t > 0,

lim lim sup sup sup E [ρ(Xn (τ ), Xn (τ + h)) ∧ 1] = 0 (15.5)

δ→0 n→∞ τ ≤t 0≤h≤δ

where the supremum in τ runs over all F Xn -optional times ≤ t.

Proof: See Kallenberg, Theorem 16.11, pp. 314–315. 

15.2 Convergence of Feller Processes

We need some technical notions about generators.

Definition 202 (Closed and Closable Generators,  Closures) A linear op-

erator A on a Banach space B is closed if its graph — f, g ∈ B 2 : f ∈ Dom(A), g = Af
— is a closed set. An operator is closable if the closure of its graph is a function
(and not just a relation). The closure of a closable operator is that function.

Notice, by the way, that because A is linear, it is closable iff fn → 0 and

Afn → g implies g = 0.

Definition 203 (Core of an Operator) Let A be a closed linear operator on

a Banach space B. A linear subspace D ⊆ Dom(A) is a core of A if the closure
of A restricted to D is, again A.

The idea of a core is that we can get away with knowing how the operator
works on a linear subspace, which is often much easier to deal with, rather than
controlling how it acts on its whole domain.

Lemma 204 (Feller Generators Are Closed) The generator of every Feller
semigroup is closed.

Proof: We need to show that the graph of G contains all of its limit points,
that is, if fn ∈ Dom(G) converges (in L∞ ) on f , and Gfn → g, then f ∈ Dom(G)
and Gf = g. First we show that f ∈ Dom(G).

lim (I − G)fn = lim fn − lim Gfn (15.6)

n→∞ n n
= f −g (15.7)
CHAPTER 15. CONVERGENCE OF FELLER PROCESSES 110

−1
But (I − G) = R1 . Since this is a bounded linear operator, we can exchange
applying the inverse and taking the limit, i.e.,

R1 lim (I − G)fn = R1 (f − g) (15.8)

n
lim R1 (I − G)fn = R1 (f − g) (15.9)
n
lim fn = R1 (f − g) (15.10)
n
f = R1 (f − g) (15.11)

Since the range of the resolvents is contained in the domain of the generator,
f ∈ Dom(G). We can therefore say that f − g = (I − G)f , which implies that
Gf = g. Hence, the graph of G contains all its limit points, and G is closed. 

Theorem 205 (Convergence of Feller Processes) Let Xn be a sequence

of Feller processes with semigroups Kn,t and generators Gn , and X be another
Feller process with semigroup Kt and a generator G containing a core D. Then
the following are equivalent.
1. If f ∈ D, there exists a sequence of fn ∈ Dom(Gn ) such that kfn − f k∞ →
0 and kAn fn − Af k∞ → 0.
2. For every t > 0, Kn,t f → Kt f for every f ∈ C0
3. kKn,t f − Kt f k∞ → 0 for each f ∈ C0 , uniformly in t for bounded positive
t
d d
4. If Xn (0) → X(0) in Ξ, then Xn → X in D.

Proof: See Kallenberg, Theorem 19.25, p. 385. 

Remark: The important versions of the property above are the second —
convergence of the semigroups — and the fourth — converge in distribution
of the processes. The other two are there to simplify the proof. The way the
proof works is to first show that conditions (1)–(3) are all equvialent, i.e., that
convergence of the operators in the semi-group implies the apparently-stronger
conditions about uniform convergence as well as convergence on the core. Then
one establishes the equivalence between the second condition and the fourth. To
go from convergence in distribution to convergence of conditional expectations
is fairly straightforward; to go the other way involves using both the first and
third condition, and the first part of Proposition 201. This last step uses the
Feller proporties to bound, in expectation, the amount by which the process can
move in a small amount of time.

Corollary 206 (Convergence of Discret-Time Markov Processes on

Feller Processes) Let X be a Feller process with semigroup Kt , generator
G and core D as in Theorem 205. Let hn be a sequence of positive real con-
stants converging (not necessarily monotonically) to zero. Let Yn be a sequence
of discrete-time Markov processes with evolution operators Hn . Finally, let
CHAPTER 15. CONVERGENCE OF FELLER PROCESSES 111

bt/h c
Xn (t) ≡ Yn (bt/hn c), with corresponding semigroup Kn,t = Hn n , and gener-
ator An = (1/hn )(Hn − I). Even though Xn is not in general a homogeneous
Markov process, the conclusions of Theorem 205 remain valid.

Proof: Kalleneberg, Theorem 19.28, pp. 387–388. 

Remark 1: The basic idea is to show that the process Xn is close (in the
Skorokhod-topology sense) to a Feller process X̃n , whose generator is An . One
d
then shows that X̃n converges on X, using Theorem 205, and that X̃n → Xn .
Remark 2: Even though Kn,t in the corollary above is not, strictly speaking,
the time-evolution operator of Xn , because Xn is not a Markov process, it is a
conditional expectation operator. Much more general theorems can be proved
on when non-Markov processes converge on Markov processes, using the idea
that Kn,t → Kt . See Kurtz (1975).

15.3 Approximation of Ordinary Differential Equa-

tions by Markov Processes
The following result, due to Kurtz (1970, 1971), is essentially an application of
Theorem 205.
First, recall that continuous-time, discrete-state Markov processes work es-
sentially like a combination of a Poisson process (giving the time of transitions)
with a Markov chain (giving the state moved to on transitions). This can be
generalized to continuous-time, continuous-state processes, of what are called
“pure jump” type.

Definition 207 (Pure Jump Markov Process) A continuous-parameter

Markov process is a pure jump process when its sample paths are piece-wise
constant. For each state, there is an exponential distribution of times spent in
that state, whose parameter is denoted λ(x), and a transition probability kernel
or exit distribution µ(x, B).

Observe that pure-jump Markov processes always have cadlag sample paths.
Also observe that the average amount of time the process spends in state x, once
it jumps there, is 1/λ(x). So the time-average “velocity”, i.e., rate of change,
starting from x, Z
λ(x) (y − x)µ(x, dy)
Ξ

Proposition 208 (Pure-Jump Markov Processea and ODEs) Let Xn be

a sequence of pure-jump Markov processes with state spaces Ξn , holding time
parameters λn and transition probabilities µn . Suppose that, for all n Ξn is a
Borel-measurable subset of Rk for some k. Let Ξ be another measurable subset
of Rk , on which there exists a function F (x) such that |F (x) − F (y)| ≤ M |x − y|
for some constant M . Suppose all of the following conditions holds.
CHAPTER 15. CONVERGENCE OF FELLER PROCESSES 112

1. The time-averaged rate of change is always finite:

Z
sup sup λn (x) |y − x|µn (x, dy) < ∞ (15.12)
n x∈Ξn ∩Ξ Ξn

2. There exists a positive sequence n → 0 such that

Z
lim sup λn (x) |y − x|µn (x, dy) = 0 (15.13)
n→∞ x∈Ξn ∩Ξ |y−x|>

3. The worst-case difference between F (x) and the time-averaged rates of

change goes to zero:
 Z 
 
lim sup F (x) − λn (x) (y − x)µn (x, dy) = 0 (15.14)

n→∞ x∈Ξn ∩Ξ

Let X(s, x0 ) be the solution to the initial-value problem where the differential is
given by F , i.e., for each 0 ≤ s ≤ t,
∂
X(s, x0 ) = F (X(s, x0 )) (15.15)
∂s
X(0, x0 ) = x0 (15.16)

and suppose there exists an η > 0 such that, for all n,

 
k
Ξn ∩ y ∈ R : inf |y − X(s, x0 )| ≤ η ⊆ Ξ (15.17)
0≤s≤t

Then lim Xn (0) = x0 implies that, for every δ > 0,

 
lim P sup |Xn (s) − X(s, x0 )| > δ = 0 (15.18)
n→∞ 0≤s≤t

The first conditions on the Xn basically make sure that they are Feller
processes. The subsequent ones make sure that the mean time-averaged rate of
change of the jump processes converges on the instantaneous rate of change of
the differential equation, and that, if we’re sufficiently close to the solution of
the differential equation in Rk , we’re not in some weird way outside the relevant
domains of definition. Even though Theorem 205 is about weak convergence,
converging in distribution on a non-random object is the same as converging in
probability, which is how we get uniform-in-time convergence in probability for
a conclusion.
There are, broadly speaking, two kinds of uses for this result. One kind is
practical, and has to do with justifying convenient approximations. If n is large,
we can get away with using an ODE instead of the noisy stochastic scheme, or
alternately we can use stochastic simulation to approximate the solutions of ugly
ODEs. The other kind is theoretical, about showing that the large-population
limit behaves deterministically, even when the individual behavior is stochastic
and strongly dependent over time.
CHAPTER 15. CONVERGENCE OF FELLER PROCESSES 113

15.4 Exercises
Exercise 15.1 (Poisson Counting Process) Show that the Poisson counting
process is a pure jump Markov process.

Exercise 15.2 (Exponential Holding Times in Pure-Jump Processes)

Prove that pure-jump Markov processes have exponentially-distributed holding
times, i.e., that if X is a Markov process with piecewise-constant sample paths,
and x = inf t > 0X(t) 6= x, that x |X(0) = x is exponentially distributed.

Exercise 15.3 (Solutions of ODEs are Feller Processes) Let F : Rd 7→

Rd be a sufficiently smooth vector field that the ordinary differential equation
dx/dt = F (x) as a unique solution for every initial condition x0 . Prove that the
set of solutions forms a Feller process.
Chapter 16

Convergence of Random
Walks

This lecture examines the convergence of random walks to the

Wiener process. This is very important both physically and statis-
tically, and illustrates the utility of the theory of Feller processes.
Section 16.1 finds the semi-group of the Wiener process, shows
it satisfies the Feller properties, and finds its generator.
Section 16.2 turns random walks into cadlag processes, and gives
a fairly easy proof that they converge on the Wiener process.

16.1 The Wiener Process is Feller

Recall that the Wiener process W (t) is defined by starting at the origin, by
independent increments over non-overlapping intervals, by the Gaussian distri-
bution of increments, and by continuity of sample paths (Examples 38 and 79).
The process is homogeneous, and the transition kernels are (Section 11.2)
Z
1 − (w2 −w1 )2
µt (w1 , B) = dw2 √ e 2t (16.1)
B 2πt
dµt (w1 , w2 ) 1 − (w2 −w1 )2
= √ e 2t (16.2)
dλ 2πt
where the second line gives the density of the transition kernel with respect to
Lebesgue measure.
Since the kernels are known, we can write down the corresponding evolution
operators:
Z
1 − (w2 −w1 )2
Kt f (w1 ) = dw2 f (w2 ) √ e 2t (16.3)
2πt
We saw in Section 11.2 that the kernels have the semi-group property, so
(Lemma 121) the evolution operators do too.

114
CHAPTER 16. CONVERGENCE OF RANDOM WALKS 115

Let’s check that {Kt } , t ≥ 0 is a Feller semi-group. The first Feller property
d
is easier to check in its probabilistic form, that, for all t, y → x implies Wy (t) →
Wx (t). The distribution of Wx (t) is just N (x, t), and it is indeed true that y → x
implies N (y, t) → N (x, t). The second Feller property can be checked in its
semi-group form: as t → 0, µt (w1 , B) approaches δ(w −w1 ), so limt→0 Kt f (x) =
f (x). Thus, the Wiener process is a Feller process. This implies that it has
cadlag sample paths (Theorem 193), but we already knew that, since we know
it’s continuous. What we did not know was that the Wiener process is not just
Markov but strong Markov, which follows from Theorem 194.
To find the generator of {Kt } , t ≥ 0, it will help to re-write it in an equivalent
form, as
h √ i
Kt f (w) = E f (w + Z t) (16.4)

where Z is an independent N (0, 1) random variable. (We saw that this was
equivalent in Section 11.2.) Now let’s pick an f ∈ C0 which is also twice
continuously differentiable, i.e., f ∈ C0 ∩ C 2 . Look at Kt f (w) − f (w), and
apply Taylor’s theorem, expanding around w:
h √ i
Kt f (w) − f (w) = E f (w + Z t) − f (w) (16.5)
h √ i
= E f (w + Z t) − f (w) (16.6)
√ √
 
1
= E Z tf 0 (w) + tZ 2 f 00 (w) + R(Z t) (16.7)
2
√ 0 f 00 (w)  2  h √ i
= tf (w)E [Z] + t E Z + E R(Z t) (16.8)
2
 2
Recalling that E [Z] = 0, E Z = 1,
 √ 
Kt f (w) − f (w) 1 00 E R(Z t)
lim = f (w) + lim (16.9)
t↓0 t 2 t↓0 t
√
So, we need to investigate the behavior of the remainder term R(Z t).
We know from Taylor’s theorem that
√ tZ 2 1
Z √
R(Z t) = du f 00 (w + uZ t) − f 00 (w) (16.10)
2 0
(16.11)

Since f ∈ C0 ∩C 2 , we know that f 00 ∈ C0 . Therefore, f 00 is uniformly continuous,

and has a modulus of continuity,

m(f 00 , h) = sup |f 00 (x) − f 0 0(y)| (16.12)

x,y: |x−y|≤h
CHAPTER 16. CONVERGENCE OF RANDOM WALKS 116

which goes to 0 as h ↓ 0. Thus

 √  tZ 2 √
R(Z t) ≤ m(f 00 , Z t) (16.13)

√  2
 √
R(Z t) Z 2 m(f 00 , Z t)
lim ≤ lim (16.14)
t→0 t t→0 2
= 0 (16.15)
Plugging back in to Equation 16.9,
 √ 
1 00 E R(Z t)
Gf (w) = f (w) + lim (16.16)
2 t↓0 t
1 00
= f (w) (16.17)
2
2
That is, G = 12 dwd
2 , one half of the Laplacian. We have shown this only for

C0 ∩ C 2 , but this is clearly a linear subspace of C0 , and, since C 2 is dense in

C, it is dense in C0 , i.e., this is a core for the generator. Hence the generator is
d2
really the extension of 21 dw 2 to the whole of C0 , but this is too cumbersome to

repeat all the time, so we just say it’s the Laplacian.

16.2 Convergence of Random Walks

Let X1 , X2 , . . . be a sequence of IID variables
Pn with mean 0 and variance 1. The
random walk process Sn is then just i=1 Xi . It is a discrete-time Markov
process, and consequently also a strong Markov process. Imagine each step of
the walk takes some time h, and imagine this time interval becoming smaller
and smaller. Then, between any two times t1 and t2 , the number of steps of
the random walk will be about t2 −t h , which will go to infinity. The increment
1

of the random walk from t1 to t2 will then be a sum of an increasingly large

number of IID random variables, and by the central limit theorem will approach
a Gaussian distribution. Moreover, if we look at the interval of time from t2 to
t3 , we will see another Gaussian, but all of the random-walk steps going into
it will be independent of those going into our first interval. So, we expect that
the random walk will in some sense come to look like the Wiener process, no
matter what the exact distribution of the X1 . (We saw some of this in Section
11.3.) Let’s consider this in more detail.
Definition 209 (Continuous-Time Random Walk (Cadlag)) Let X1 , X2 , . . .
be an IID sequence of real-valued random variables with mean 0 and variance 1.
Define S(m) as X1 + X2 . . . + Xm , and X0 = 0. The corresponding continuous-
time random walks (CTRWs) are the processes
bntc
1 X
Yn (t) ≡ Xi (16.18)
n1/2 i=0
−1/2
= n S(bntc) (16.19)
CHAPTER 16. CONVERGENCE OF RANDOM WALKS 117

Remark 1: You should verify (Exercise 16.3) that continuous-time random

walks are inhomogeneous Markov processes with cadlag sample paths.
Remark 2: We will later see CTRWs with continuous sample paths, obtained
by linear interpolation, but these, with their piece-wise constant sample paths,
will do for now.
As has been hinted at repeatedly, random walks converge in distribution on
the Wiener process. There are, broadly speaking, two ways to show this. One is
to use the Feller process machinery of Chapter 15, and apply Corollary 206. The
other is to directly manipulate the criteria for convergence of cadlag processes.
Both lead to the same conclusion.

16.2.1 Approach Through Feller Processes

The processes Yn are not homogeneously Markovian, though the discrete-time
processes n−1/2 S(m) are. Nonetheless, we can find the equivalent of their evo-
lution operators, and show that they converge on the evolution operators of the
Wiener process. First, let’s establish a nice property of the increments of Yn .

Lemma 210 (Increments of Random Walks) For a continuous-time ran-

dom walk, for all n,

Yn (t + h) − Yn (t) = n−1/2 S 0 (bn(t + h)c − bntc) (16.20)

Pm
where S 0 (m) ≡ i=0 Xi0 and the IID sequence X 0 is an independent copy of X.

Proof: By explicit calculation. For any n, for any t and any h > 0,
bn(t+h)c
1 X
Yn (t + h) = Xi (16.21)
n1/2 i=0
bn(t+h)c
X
= Yn (t) + n−1/2 Xi (16.22)
i=bntc+1
bn(t+h)c−bntc
X
= Yn (t) + n−1/2 Xi0 (16.23)
i=0

using, in the last line, the fact that the Xi are IID. Eq. 16.20 follows from the
definition of S(m). 

Lemma 211 (Continuous-Time Random Walks are Pseudo-Feller) Ev-

ery continuous-time random walk has the Feller properties (but is not a homo-
geneous Markov process).

Proof: This is easiest using the process/probabilistic forms (Definition 177)

of the Feller properties.
CHAPTER 16. CONVERGENCE OF RANDOM WALKS 118

To see that the first property (Eq. 14.3) holds, we need the distribution of
Yn,y (t), that is, the distribution of the state of Yn at time t, when started from
state y (rather than 0). By Lemma 210,
d
Yn,y (t) = y + n−1/2 S 0 (bntc) (16.24)
d
Clearly, as y → x, y + n−1/2 S 0 (bntc) → x + n−1/2 S 0 (bntc), so the first Feller
property holds.
To see that the second Feller property holds, observe that, for each n, for all
t, and for all ω, Yn (t + h, ω) = Yn (t, ω) if 0 ≥ h < 1/n. This sure convergence
implies almost-sure convergence, which implies convergence in probability. 
Lemma 212 (Evolution Operators of Random Walks) The “evolution”
(i.e., conditional expectation) operators of the random walk Yn , Kn,t , are given
by
Kn,t f (y) = E [f (y + Yn0 (t))] (16.25)
where Yn0 is an independent copy of Yn .
Proof: Substitute Lemma 210 into the definition of the evolution operator.
Kn,h f (y) ≡ E [f (Yn (t + h))|Yn (t) = y] (16.26)
= E [f (Yn (t + h) + Yn (t) − Yn (t)) |Yn (t) = y] (16.27)
h   i
= E f n−1/2 S 0 (bntc) + Yn (t)|Yn (t) = y (16.28)
h i
= E f (y + n−1/2 S 0 (bntc)) (16.29)

In words, the transition operator simply takes the expectation over the incre-
ments of the random walk, just as with the Wiener process. Finally, substitution
of Yn0 (t) for n−1/2 S 0 (bntc) is licensed by Eq. 16.19. 
d
Theorem 213 (Functional Central Limit Theorem (I)) Yn → W in D.
Proof: Apply Theorem 206. Clause (4) of the theorem says that if any of
d d
the other three clauses are satisfied, and Yn (0) → W (0) in R, then Yn → W
in D. Clause (2) is that Kn,t → Kt for all t > 0. That is, for any t > 0, and
f ∈ C0 , Kn,t f → Kt f as n → ∞. Pick any such t and f and consider Kn,t f .
By Lemma 212,
h i
Kn,t f (y) = E f (y + n−1/2 S 0 (bntc)) (16.30)

−1/2
As n → ∞, n−1/2 → t1/2 bntc . Since the Xi (and so the Xi0 ) have variance
−1/2 d
1, the central limit theorem applies, and bntc S 0 (bntc) → N (0, 1), say Z.
d √
Consequently n−1/2 S 0 (bntc → tZ. Since f ∈ C0 , it is continuous and bounded,
hence, by the definition of convergence in distribution,
h  i h  √ i
E f y + n−1/2 S 0 (bntc) → E f y + tZ (16.31)
CHAPTER 16. CONVERGENCE OF RANDOM WALKS 119
 √ 
But E f (y + tZ) = Kt f (y), the time-evolution operator of the Wiener pro-
cess applied to f at y. Since the evolution operators of the random walks con-
verge on those of the Wiener process, and since their initial conditions match,
d
by the theorem Yn → W in D. 

16.2.2 Direct Approach

The alternate approach to the convergence of random walks works directly with
the distributions, avoiding the Feller properties. It is not quite so slick, but
provides a comparatively tractable example of how general results about con-
vergence of stochastic processes go.
We want to find the limiting distribution of Yn as n → ∞. First of all,
we should convince ourselves that a limit distribution exists. But this is not
too hard. For any fixed t, Yn (t) approaches a Gaussian distribution by the
central limit theorem. For any fixed finite collection of times t1 ≤ t2 . . . ≤ tk ,
Yn (t1 ), Yn (t2 ), . . . Yn (tk ) approaches a limiting distribution if Yn (t1 ), Yn (t2 ) −
Yn (t1 ), . . . Yn (tk ) − Yn (tk−1 ) does, but that again will be true by the (multivari-
ate) central limit theorem. Since the limiting finite-dimensional distributions
exist, some limiting distribution exists (via Theorem 23). It remains to convince
ourselves that this limit is in D, and to identify it.

Lemma 214 (Convergence of Random Walks in Finite-Dimensional

fd
Distribution) Yn → W .

Proof: For all n, Yn (0) = 0 = W (0). For any t2 > t1 ,

 
bnt2 c
1 X
L (Yn (t2 ) − Yn (t1 )) = L  √ Xi  (16.32)
n
i=bnt1 c
d
→ N (0, t2 − t1 ) (16.33)
= L (W (t2 ) − W (t1 )) (16.34)

Finally, for any three times t1 < t2 < t3 , Yn (t3 ) − Yn (t2 ) and Yn (t2 ) − Yn (t1 )
are independent for sufficiently large n, because they become sums of disjoint
collections of independent random variables. The same applies to large groups
of times. Thus, the limiting distribution of Yn starts at the origin and has
independent Gaussian increments. Since these properties determine the finite-
fd
dimensional distributions of the Wiener process, Yn → W . 
d
Theorem 215 (Functional Central Limit Theorem (II)) Yn → W .
fd
Proof: By Proposition 200, it is enough to show that Yn → W , and that
any of the properties in Proposition 201 hold. The lemma took care of the
finite-dimensional convergence, so we can turn to the second part. A sufficient
CHAPTER 16. CONVERGENCE OF RANDOM WALKS 120

P
condition is property (1) in the latter theorem, that |Yn (τn + hn ) − Yn (τn )| → 0
for all finite optional times τn and any sequence of positive constants hn → 0.
|Yn (τn + hn ) − Yn (τn )| = n−1/2 |S(bnτn + nhn c) − S(bnτn c)| (16.35)
d
= n−1/2 |S(bnhn c) − S(0)| (16.36)
= n−1/2 |S(bnhn c)| (16.37)
 
bnhn c 
X
= n−1/2 
 
Xi  (16.38)
 i=0 
To see that this converges in probability to zero, we will appeal to Chebyshev’s
inequality: if Zi have common mean 0 and variance σ 2 , then, for every positive
,
m  !
X  mσ 2
Zi  >  ≤ (16.39)
 
P 

i=1
 2
√
Here we have Zi = Xi / n, so σ 2 = 1/n, and m = bnhn c. Thus
  bnhn c
P n−1/2 |S(bnhn c)| >  ≤ (16.40)
n2
As 0 ≤ bnhn c /n ≤ hn , and hn → 0, the bounding probability must go to zero
P
for every fixed . Hence n−1/2 |S(bnhn c)| → 0. 

16.2.3 Consequences of the Functional Central Limit The-

orem
Corollary 216 (The Invariance Principle) Let X1 , X2 , . . . be IID random
variables with mean µ and variance σ 2 . Then
bntc
1 X Xi − µ d
Yn (t) ≡ √ → W (t) (16.41)
n i=0 σ

Proof: (Xi − µ)/σ has mean 0 and variance 1, so Theorem 215 applies. 
This result is called “the invariance principle”, because it says that the
limiting distribution of the sequences of sums depends only on the mean and
variance of the individual terms, and is consequently invariant under changes
which leave those alone. Both this result and the previous one are known as the
“functional central limit theorem”, because convergence in distribution is the
same as convergence of all bounded continuous functionals of the sample path.
Another name is “Donsker’s Theorem”, which is sometimes associated however
with the following corollary of Theorem 215.
Corollary 217 (Donsker’s Theorem) Let Yn (t) and W (t) be as before, but
restrict the index set T to the unit interval [0, 1]. Let f be any function from
d
D([0, 1]) to R which is measurable and a.s. continuous at W . Then f (Yn ) →
f (W ).
CHAPTER 16. CONVERGENCE OF RANDOM WALKS 121

Proof: Exercise. 
This version is especially important for statistical purposes, as we’ll see a
bit later.

16.3 Exercises
Exercise 16.1 (Example 167 Revisited) Go through all the details of Ex-
ample 167.
 
1. Show that FtX ⊆ FtW for all t, and that FtX ⊂ FtW .

2. Show that τ = inf t X(t) = (0, 0) is a FtX -optional time, and that it is
finite with probability 1.
3. Show that X is Markov with respect to both its natural filtration and the
natural filtration of the driving Wiener process.
4. Show that X is not strongly Markov at τ .
5. Which, if either, of the Feller properties does X have?

Exercise 16.2 (Generator of the d-dimensional Wiener Process) Con-

sider a d-dimensional Wiener process, i.e., an Rd -valued process where each
coordinate is an independent Wiener process. Find the generator.

Exercise 16.3 (Continuous-time random walks are Markovian) Show

that every continuous-time random walk (as per Definition 209) is an inhomogeneous
Markov process, with cadlag sample paths.

Exercise 16.4 (Donsker’s Theorem) Prove Donsker’s Theorem (Corollary

217).

Exercise 16.5 (Diffusion equation) The partial differential equation

1 ∂ 2 f (x, t) ∂f (x, t)
=
2 ∂x2 ∂t
is called the diffusion equation. From our discussion of initial value problems
in Chapter 12 (Corollary 159 and related material), it is clear that the function
f (x, t) solves the diffusion equation with initial condition f (x, 0) if and only if
f (x, t) = Kt f (x, 0), where Kt is the evolution operator of the Wiener process.
−1/2 − x2
1. Take f (x, 0) = (2π10−4 ) e 2·10−4 . f (x, t) can be found analytically;
do so.
CHAPTER 16. CONVERGENCE OF RANDOM WALKS 122

2. Estimate f (x, 10) over the interval [−5, 5] stochastically. Use the fact that
Kt f (x) = E [f (W (t))|W (0) = x], and that random walks converge on the
Wiener process. (Be careful that you scale your random walks the right
way!) Give an indication of the error in this estimate.
3. Can you find an analytical form for f (x, t) if f (x, 0) = 1[−0.5,0.5] (x)?

4. Find f (x, 10), with the new initial conditions, by numerical integration on
the domain [−10, 10], and compare it to a stochastic estimate.

Exercise 16.6 (Functional CLT for Dependent Variables) Let Xi , i =

1, 2, . . ., be a weakly stationary but dependent sequence of real-valued random
variables, with mean 0 and standard deviation 1. (Note that any weakly-stationary
sequence with finite variance can be normalized into this form.) Let Yn be the
corresponding continuous-time random walk, i.e.,
bntc
1 X
Yn (t) = Xi
n1/2 i=0

Suppose that, despite their interdependence, the Xi still obey the central limit
theorem,
n
1 X d
Xi → N (0, 1)
n1/2 i=1
Are these conditions enough to prove a functional central limit theorem, that
d
Yn → W ? If so, prove it. If not, explain what the problem is, and suggest an
additional sufficient condition on the Xi .
 Part IV

Diffusions and Stochastic

Calculus

123
Chapter 17

Diffusions and the Wiener

Process

Section 17.1 introduces the ideas which will occupy us for the
next few lectures, the continuous Markov processes known as diffu-
sions, and their description in terms of stochastic calculus.
Section 17.2 collects some useful properties of the most important
diffusion, the Wiener process.
Section 17.3 shows, first heuristically and then more rigorously,
that almost all sample paths of the Wiener process don’t have deriva-
tives.

17.1 Diffusions and Stochastic Calculus

So far, we have looked at Markov processes in general, and then paid particular
attention to Feller processes, because the Feller properties are very natural con-
tinuity assumptions to make about stochastic models and have very important
consequences, especially the strong Markov property and cadlag sample paths.
The natural next step is to go to Markov processes with continuous sample
paths. The most important case, overwhelmingly dominating the literature, is
that of diffusions.

Definition 218 (Diffusion) A stochastic process X adapted to a filtration F

is a diffusion when it is a strong Markov process with respect to F, homogeneous
in time, and has continuous sample paths.

Remark: Having said that, I should confess that some authors don’t insist
that diffusions be homogeneous, and some even don’t insist that they be strong
Markov processes. But this is the general sense in which the term is used.
Diffusions matter to us for several reasons. First, they are very natural
models of many important systems — the motion of physical particles (the

124
CHAPTER 17. DIFFUSIONS AND THE WIENER PROCESS 125

source of the term “diffusion”), fluid flows, noise in communication systems,

financial time series, etc. Probabilistic and statistical studies of time-series data
thus need to understand diffusions. Second, many discrete Markov models have
large-scale limits which are diffusion processes: these are important in physics
and chemistry, population genetics, queueing and network theory, certain as-
pects of learning theory1 , etc. These limits are often more tractable than more
exact finite-size models. (We saw a hint of this in Section 15.3.) Third, many
statistical-inferential problems can be described in terms of diffusions, most
prominently ones which concern goodness of fit, the convergence of empirical
distributions to true probabilities, and nonparametric estimation problems of
many kinds.
The easiest way to get at diffusions is to through the theory of stochas-
tic differential equations; the most important diffusions can be thought of as,
roughly speaking, the result of adding a noise term to the right-hand side of a
differential equation. A more exact statement is that, just as an autonomous
ordinary differential equation
dx
= f (x), x(t0 ) = x0 (17.1)
dt
has the solution Z t
x(t) = x0 + f (x(s))ds (17.2)
t0

a stochastic differential equation

dX dY
= f (X) + g(X) , X(t0 ) = x0 a.s. (17.3)
dt dt
where X and Y are stochastic processes, is solved by
Z Z
X(t) = x0 + f (X(s))ds + g(X(s))dY (17.4)
R
where g(X, t)dY is a stochastic integral. It turns out that, properly con-
structed, this sort of integral, and so this sort of stochastic differential equation,
makes sense even when dY /dt does not make sense as any sort of ordinary
derivative, so that the more usual way of writing an SDE is

dX = f (X)dt + g(X)dY, X(t0 ) = x0 a.s. (17.5)

even though this seems to invoke infinitessimals, which don’t exist.2

1 Specifically, discrete-time reinforcement learning converges to the continuous-time repli-

cator equation of evolutionary theory.

2 Some people, like Ethier and Kurtz (1986), prefer to talk about stochastic integral equa-

tions, rather than stochastic differential equations, because things like 17.5 are really short-
hands for “find an X such that Eq. 17.4 holds”, and objects like dX don’t really make much
sense on their own. There’s a certain logic to this, but custom is overwhelmingly against
them.
CHAPTER 17. DIFFUSIONS AND THE WIENER PROCESS 126

The fully general theory of stochastic calculus considers integration with re-
spect to a very broad range of stochastic processes, but the original case, which
is still the most important, is integration with respect to the Wiener process,
which corresponds to driving a system with white noise. In addition to its
many applications in all the areas which use diffusions, the theory of integra-
tion against the Wiener process occupies a central place in modern probability
theory; I simply would not be doing my job if this course did not cover it.
We therefore begin our study of diffusions and stochastic calculus by reviewing
some of the properties of the Wiener process — which is also the most important
diffusion process.

17.2 Once More with the Wiener Process and

Its Properties
To review, the standard Wiener process W (t) is defined by
1. W (0) = 0,
2. centered Gaussian increments with linearly-growing variance, L (W (t2 ) − W (t1 )) =
N (0, t2 − t1 ),
3. independent increments and
4. continuity of sample paths.
We have seen that it is a homogeneous Markov process (Section 11.2), and
in fact (Section 16.1) a Feller process, and therefore a strong Markov process,
whose generator (Equation 16.16) is 21 ∇2 . By Definition 218, W is a diffusion.
This section proves a few more useful properties.

Lemma 219 (The Wiener Process Is a Martingale) The Wiener process

is a martingale with respect to its natural filtration.

Proof: This follows directly from the Gaussian increment property:

E W (t + h)|FtX
 
= E [W (t + h)|W (t)] (17.6)
= E [W (t + h) − W (t) + W (t)|W (t)] (17.7)
= E [W (t + h) − W (t)|W (t)] + W (t) (17.8)
= 0 + W (t) = W (t) (17.9)

where the first line uses the Markov property of W , and the last line the Gaussian
increments property. 

Definition 220 (Wiener Processes with Respect to Filtrations) If W (t, ω)

is adapted to a filtration {Ft } and is an {Ft }-filtration, it is an {Ft } Wiener
process or {Ft } Brownian motion.
CHAPTER 17. DIFFUSIONS AND THE WIENER PROCESS 127

It seems natural to speak of the Wiener process as a Gaussian process. This

motivates the following definition.

Definition 221 (Gaussian Process) A real-valued stochastic process is Gaus-

sian when all its finite-dimensional distributions are multivariate Gaussian dis-
tributions.

Lemma 222 (Wiener Process Is Gaussian) The Wiener process is a Gaus-

sian process.

Proof: Pick any k times t1 < t2 < . . . < tk . Then the increments
W (t1 ) − W (0), W (t2 ) − W (t1 ), W (t3 ) − W (t2 ), . . . W (tk ) − W (tk−1 ) are in-
dependent Gaussian random variables. If X and Y are independent Gaus-
sians, then X, X + Y is a multivariate Gausssian, so (recursively) W (t1 ) −
W (0), W (t2 )−W (0), . . . W (tk )−W (0) has a multivariate Gaussian distribution.
Since W (0) = 0, the Gaussian distribution property follows. Since t1 , . . . tk were
arbitrary, as was k, all the finite-dimensional distributions are Gaussian. 
Just as the distribution of a Gaussian random variable is determined by
its mean and covariance, the distribution of a Gaussian process is determined
by its mean over time, E [X(t)], and its covariance function, cov (X(s), X(t)).
(You might find it instructive to prove this without looking at Lemma 13.1 in
Kallenberg.) Clearly, E [W (t)] = 0, and, taking s ≤ t without loss of generality,

cov (W (s), W (t)) = E [W (s)W (t)] − E [W (s)] E [W (t)] (17.10)

= E [(W (t) − W (s) + W (s))W (s)] (17.11)
= E [(W (t) − W (s))W (s)] + E [W (s)W (s)] (17.12)
= E [W (t) − W (s)] E [W (s)] + s (17.13)
= s (17.14)

17.3 Wiener Measure; Most Continuous Curves

Are Not Differentiable
We can regard the Wiener process as establishing a measure on the space C(R+ )
of continuous real-valued functions; this is one of the considerations which led
Wiener to it (Wiener, 1958)3 . This will be important when we want to do
statistical inference for stochastic processes. All Bayesian methods, and most
frequentist ones, will require us to have a likelihood for the model θ given data
x, fθ (x), but likelihoods are really Radon-Nikodym derivatives, fθ (x) = dνdµ (x)
θ

with respect to some reference measure µ. When our sample space is Rd , we

generally use Lebesgue measure as our reference measure, since its support is
3 The early chapters of this book form a wonderfully clear introduction to Wiener measure,

starting from prescriptions on the measures of finite-dimensional cylinders and building from
there, deriving the incremental properties we’ve started with as consequences.
CHAPTER 17. DIFFUSIONS AND THE WIENER PROCESS 128

the whole space, it treats all points uniformly, and it’s reasonably normalizable.
Wiener measure will turn out to play a similar role when our sample space is
C.
A mathematically important question, which will also turn out to matter
to us very greatly when we try to set up stochastic differential equations, is
whether, under this Wiener measure, most curves are differentiable. If, say,
almost all curves were differentiable, then it would be easy to define dW/dt.
Unfortunately, this is not the case; almost all curves are nowhere differentiable.
There is an easy heuristic argument to this conclusion. W (t) is a Gaussian,
whose variance is t. If we look at the ratio in a derivative
W (t + h) − W (t)
(t + h) − t

the numerator has variance h and the denominator is the constant h, so the
ratio has variance 1/h, which goes to infinity as h → 0. In other words, as we
look at the curve of W (t) on smaller and smaller scales, it becomes more and
more erratic, and the slope finally blows up into a completely unpredictable
quantity. This is basically the shape of the more rigorous argument as well.

Theorem 223 (Almost All Continuous Curves Are Non-Differentiable)

With probability 1, W (t) is nowhere-differentiable.

Proof: Assume, by way of contradiction, that W (t) is differentiable at t0 .

Then
W (t, ω) − W (t0 , ω)
lim (17.15)
t→t0 t − t0
must exist, for some set of ω of positive measure. Call its supposed value
W 0 (t0 , ω). That is, for every  > 0, we must have some δ such that t0 − δ ≤ t ≤
t0 + δ implies  
 W (t, ω) − W (t0 , ω) 0

 − W (t0 , ω) ≤  (17.16)
 t − t0
Without loss of generality, take t > t0 . Then W (t, ω) − W (t0 , ω) is independent
of W (t0 , ω) and has a Gaussian distribution with mean zero and variance t − t0 .
1
Therefore the differential ratio is N (0, t−t0
). The quantity inside the absolute
value sign in Eq. 17.16 is thus Gaussian with distribution N (−W 0 (t0 ), t−t 1
0
).
The probability that it exceeds any  is therefore always positive, and in fact
can be made arbitrarily large by taking t sufficiently close to t0 . Hence, with
probability 1, there is no point of differentiability. 
Continuous curves which are nowhere differentiable are odd-looking beasts,
but we’ve just established that such “pathological” cases are in fact typical, and
non-pathological ones vanishingly rare in C. What’s worse, in the functional
central limit theorem (215), we obtained W as the limit of piecewise constant,
and so piecewise differentiable, random functions. We could even have linearly
interpolated between the points of the random walk, and those random functions
CHAPTER 17. DIFFUSIONS AND THE WIENER PROCESS 129

would also have converged in distribution on W .4 The continuous, almost-

everywhere-differentiable curves form a subset of C, and now we have a sequence
of measures which give them probability 1, converging on Wiener measure, which
gives them probability 0. This sounds like trouble, especially if we want to use
Wiener measure as a reference measure in likelihoods, because it sounds like
lots of interesting measures, which do produce differentiable curves, will not be
absolutely continuous...5

4 Could we have used quadratic spline interpolation to get sample paths with continuous
d
first derivatives, and still have had → W ?
5 Doing Exercises 1.1 and 7.1, if you haven’t already, should help you resolve this paradox.
Chapter 18

A First Look at Stochastic

Integrals with the Wiener
Process

Section 18.1 touches briefly on the martingale characterization

of the Wiener process.
Section 18.2 gives a heuristic introduction to stochastic integrals,
via Euler’s method for approximating ordinary integrals.

18.1 Martingale Characterization of the Wiener

Process
Because the Wiener process is a Lévy process (Example 139), it is self-similar
d
in the sense of Definition 147. That is, for any a > 0, W (at) = a1/2 W (t). In
fact, if we define a new process Wa through Wa (t, ω) = a−1/2 W (at, ω), then
Wa is itself a Wiener process. Thus the whole process is self-similar. This is
only one of several sorts of self-similarities in the Wiener process. Another is
sometimes called spatial homogeneity: Wτ , defined through Wτ (t, ω) = W (t +
τ, ω) − W (τ, omega) is also a Wiener process. That is, if we “re-zero” to the
state of the Wiener process W (τ ) at an arbitrary time τ , the new process looks
just like the old process. −W (t), obviously, is also a Wiener process.
2
properties is the fact that W (t) − t is a martingale with
Relatedto these
W
respect to Ft . (This is easily shown with a little algebra.) What is more
surprising is that this is enough to characterize the Wiener process.

Theorem 224 (Martingale Characterization of the Wiener Process) If

M (t) is a continuous martingale, and M 2 (t) − t is also a martingale, then M (t)
is a Wiener process.

130
CHAPTER 18. STOCHASTIC INTEGRALS PREVIEW 131

There are some very clean proofs of this theorem1 — but they require us to
use stochastic calculus! Doob (1953, pp. 384ff) gives a proof which does not,
however. The details of his proof are messy, but the basic idea is to get the
central limit theorem to apply, using the martingale property of M 2 (t) − t to
get the variance to grow linearly with time and to get independent increments,
and then seeing that between any two times t1 and t2 , we can fit arbitrarily
many little increments so we can use the CLT.
We will return to this result as an illustration of the stochastic calculus
(Theorem 247).

18.2 A Heuristic Introduction to Stochastic In-

tegrals
Euler’s method is perhaps the most basic method for numerically approximating
Rb
integrals. If we want to evaluate I(x) ≡ a x(t)dt, then we pick n intervals of
time, with boundaries a = t0 < t1 < . . . tn = b, and set
n
X
In (x) = x (ti−1 ) (ti − ti−1 )
i=1

Then In (x) → I(x), if x is well-behaved and the length of the largest interval
R t=b
→ 0. If we want to evaluate t=a x(t)dw, where w is another function of t, the
natural thing to do is to get the derivative of w, w0 , replace the integrand by
x(t)w0 (t), and perform the integral with respect to t. The approximating sums
are then
Xn
x (ti−1 ) w0 (ti−1 ) (ti − ti−1 ) (18.1)
i=1

Alternately, we could, if w(t) is nice enough, approximate the integral by

n
X
x (ti−1 ) (w (ti ) − w (ti−1 )) (18.2)
i=1

even if w0 doesn’t exist.

(You may be more familiar with using Euler’s method to solve ODEs, dx/dt =
f (x). Then one generally picks a ∆t, and iterates

x(t + ∆t) = x(t) + f (x)∆t (18.3)

from the initial condition x(t0 ) = x0 , and uses linear interpolation to get a
continuous, almost-everywhere-differentiable curve. Remarkably enough, this
converges on the actual solution as ∆t shrinks (Arnol’d, 1973).)
Let’s try to
R carry all this over to random functions of time X(t) and W (t).
The integral X(t)dt is generally not a problem — we just find a version of X
1 See especially Ethier and Kurtz (1986, Theorem 5.2.12, p. 290).
CHAPTER 18. STOCHASTIC INTEGRALS PREVIEW 132

R
with measurable sample paths (Section 8.2). X(t)dW is also comprehensible
if dW/dt exists (almost surely). Unfortunately, we’ve seen that this is not the
case for the Wiener process, which (as you can tell from the W ) is what we’d
really like to use here. So we can’t approximate the integral with a sum like Eq.
18.1. But there’s nothing preventing us from using one like Eq. 18.2, since that
only demands increments of W . So what we would like to say is that
Z t=b n
X
X(t)dW ≡ lim X (ti−1 ) (W (ti ) − W (ti−1 )) (18.4)
t=a n→∞
i=1

This is a crude-but-workable approach to numerically evaluating stochastic in-

tegrals, and apparently how the first stochastic integrals were defined, back in
the 1920s. Notice that it is going to make the integral a random variable, i.e.,
a measurable function of ω. Notice also that I haven’t said anything yet which
should lead you to believe that the limit on the right-hand side exists, in any
sense, or that it is independent of the choice of partitions a = t0 < t1 < . . . tn b.
The next chapter will attempt to rectify this.
(When it comes to the SDE dX = f (X)dt + g(X)dW , the counterpart of
Eq. 18.3 is
X(t + ∆t) = X(t) + f (X(t))∆t + g(X(t))∆W (18.5)
where ∆W = W (t+∆t)−W (t), and again we use linear interpolation in between
the points, starting from X(t0 ) = x0 .)
Chapter 19

Stochastic Integrals and

Stochastic Differential
Equations

Section 19.1 gives a rigorous construction for the Itô integral of

a function with respect to a Wiener process.
Section 19.2 gives two easy examples of Itô integrals. The second
one shows that there’s something funny about change of variables,
or if you like about the chain rule.
Section 19.3 explains how to do change of variables in a stochastic
integral, also known as “Itô’s formula”.
Section 19.4 defines stochastic differential equations.
Section 19.5 sets up a more realistic model of Brownian motion,
leading to an SDE called the Langevin equation, and solves it to get
Ornstein-Uhlenbeck processes.

19.1 Integrals with Respect to the Wiener Pro-

cess
The drill by now should be familiar: first we define integrals of step functions,
then we approximate more general classes of functions by these elementary
functions. We need some preliminary technicalities.

Definition 225 (Progressive Process) A continuous-parameter stochastic

process X adapted to a filtration {Gt } is progressively measurable or progressive
when X(s, ω), 0 ≤ s ≤ t, is always measurable with respect to Bt × Gt , where Bt
is the Borel σ-field on [0, t].

If X has continuous sample paths, for instance, then it is progressive.

133
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 134

Definition 226 (Non-anticipating filtrations, processes) Let W be a stan-

dard Wiener process, {Ft } the right-continuous completion of the natural filtra-
tion of W , and G any σ-field independent of {Ft }. Then the non-anticipating
filtrations are the ones of the form σ(Ft ∪ G), 0 ≤ t < ∞. A stochastic process
X is non-anticipating if it is adapted to some non-anticipating filtration.

The idea of the definition is that if X is non-anticipating, we allow it to

depend on the history of W , and possibly some extra, independent random
stuff, but none of that extra information is of any use in predicting the future
development of W , since it’s independent.

Definition 227 (Elementary process) A progressive, non-anticipating pro-

cess X is elementary if there exist an increasing sequence of times ti , starting
at zero and tending to infinity, such that X(t) = X(tn ) if t ∈ [tn , tn+1 ), i.e., if
X is a step-function of time.

Remark: It is sometimes convenient to allow the break-points of the elemen-

tary process to be optional random times. We won’t need this for our purposes,
however.

Definition 228 (Mean square integrable)

hR i A random process X is mean-
b 2
square-integrable from a to b if E a X (t)dt is finite. The class of all such
processes will be written S2 [a, b].

Notice that if X is bounded on [a, b], in the sense that |X(t)| ≤ M with
probability 1 for all a ≤ t ≤ b, then X is square-integrable from a to b.

Definition 229 (S2 norm) The norm of a process X ∈ S2 [a, b] is its root-
mean-square time integral:
 "Z #1/2
 b 
2
kXkS2 ≡ E X (t)dt  (19.1)
 
 a 

Proposition 230 (k·kS2 is a norm) k·kS2 is a semi-norm on S2 [a, b]; it is a

full norm if processes such that X(t) − Y (t) = 0 a.s., for Lebesgue-almost-all t,
are identified. Like any norm, it induces a metric on S2 [a, b], and by “a limit
in S2 ” we will mean a limit with respect to this metric. As a normed space, it
is complete, i.e. every Cauchy-convergent sequence has a limit in the S2 sense.

Proof: Recall that a semi-norm is a function from a vector space to the real
numbers such that, for any vector X and any scalar a, kaXk = |a|kXk, and,
for any two vectors X and Y , kX + Y k ≤ kXk + kY k. The root-mean-square
time integral kXkS2 clearly has both properties. To be a norm and not just a
semi-norm, we need in addition that kXk = 0 if and only if X = 0. This is not
true for random processes, because the process which is zero at irrational times
t ∈ [a, b] but 1 at rational times in the interval also has semi-norm 0. However,
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 135

by identifying two processes X and Y if X(t) − Y (t) = 0 a.s. for almost all t,
we get a norm. This is exactly analogous to the way the L2 norm for random
variables is really only a norm on equivalence classes of variables which are equal
almost always. The proof that the space with this norm is complete is exactly
analogous to the proof that the L2 space of random variables is complete, see
e.g. Lemma 1.31 in Kallenberg (2002, p. 15). 

Definition 231 (Itô integral of an elementary process) If X is an el-

ementary, progressive, non-anticipative process, square-integrable from a to b,
then its Itô integral from a to b is
Z b X
X(t)dW ≡ X(ti ) (W (ti+1 ) − W (ti )) (19.2)
a i≥0

where the ti are as in Definition 227, truncated below by a and above by b.

Notice that this is basically a Riemann-Stieltjes integral. It’s a random

variable, but we don’t have to worry about the existence of a limit. Now we set
about approximating more general sorts of processes by elementary processes.

Lemma 232 (Approximation of Bounded, Continuous Processes by

Elementary Processes) Suppose X is progressive, non-anticipative, bounded
on [a, b], and has continuous sample paths. Then there exist bounded elementary
processes Xn , Itô-integrable on [a, b], such that X is the S2 [a, b] limit of Xn , i.e.,

lim kX − Xn kS2 = 0 (19.3)

n→∞
"Z #
b
2
lim E (X − Xn ) dt = 0 (19.4)
n→∞ a

Proof: Set
∞
X
Xn (t) ≡ X(i2−n )1[i/2n ,(i+1)/2n ) (t) (19.5)
i=0

This is clearly elementary, bounded and square-integrable on [a, b]. Moreover,

Rb 2
for fixed ω, a (X(t, ω) − Xn (t, ω)) dt → 0, since X(t, ω) is continuous. So the
expectation of the time-integral goes to zero by bounded convergence. 
Remark: There is nothing special about using intervals of length 2−n . Any
division of [a, b] into sub-intervals would do, provided the width of the largest
sub-interval shrinks to zero.

Lemma 233 (Approximation by of Bounded Processes by Bounded,

Continuous Processes) Suppose X is progressive, non-anticipative, and bounded
on [a, b]. Then there exist progressive, non-anticipative processes Xn which are
bounded and continuous on [a, b], and have X as their S2 limit,

lim kX − Xn kS2 = 0 (19.6)

n→∞
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 136

Proof: Let M be the bound on the absolute value of X. For each n,

pick a probability density fn (t) on R whose support is confined to the interval
(−1/n, 0). Set
Z t
Xn (t) ≡ fn (s − t)X(s)ds (19.7)
0
Xn (t) is then a sort of moving average of X, over the interval (t−1/n, t). Clearly,
it’s continuous, bounded, progressively measurable, and non-anticipative. More-
over, for each ω,
Z b
2
lim (Xn (t, ω) − X(t, ω)) dt = 0 (19.8)
n→∞ a

because of the way we’ve set up fn and Xn . By bounded convergence, this

implies "Z #
b
2
lim E (X − Xn ) dt = 0 (19.9)
n→∞ a

which is equivalent to Eq. 19.6. 

Lemma 234 (Approximation of Square-Integrable Processes by Bounded
Processes) Suppose X is progressive, non-anticipative, and square-integrable
on [a, b]. Then there exist a sequence of random processes Xn which are pro-
gressive, non-anticipative and bounded on [a, b], which have X as their limit in
S2 .
Proof: Set Xn (t) = (−n ∨ X(t)) ∧ n. This has the desired properties, and
the result follows from dominated (not bounded!) convergence. 
Lemma 235 (Approximation of Square-Integrable Processes by Ele-
mentary Processes) Suppose X is progressive, non-anticipative, and square-
integrable on [a, b]. Then there exist a sequence of bounded elementary processes
Xn with X as their limit in S2 .
Proof: Combine Lemmas 232, 233 and 234. 
This lemma gets its force from the following result.
Lemma 236 (Itô Isometry for Elementary Processes) Suppose X is as
in Definition 231, and in addition bounded on [a, b]. Then
 !2  "Z #
Z b b
2 2
E X(t)dW  = E X (t)dt = kXkS2 (19.10)
a a

Proof: Set ∆Wi = W (ti+1 ) − W (ti ). Notice that ∆Wj is independent of

X(ti )X(tj )∆Wi ifhi < j, because
i of the non-anticipation properties of X. On
2
the other hand, E (∆Wi ) = ti+1 − ti , by the linear variance of the increments
of W . So
E [X(ti )X(tj )∆Wj ∆Wi ] = E X 2 (ti ) (ti+1 − ti )δij
 
(19.11)
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 137

Substituting Eq. 19.2 into the left-hand side of Eq. 19.10,

 !2   
Z b X
E X(t)dW  = E  X(ti )X(tj )∆Wj ∆Wi  (19.12)
a i,j
X
= E [X(ti )X(tj )∆Wj ∆Wi ] (19.13)
i,j
X
E X 2 (ti ) (ti+1 − ti )
 
= (19.14)
i
" #
X
2
= E X (ti )(ti+1 − ti ) (19.15)
i
"Z #
b
2
= E X (t)dt (19.16)
a

where the last step uses the fact that X 2 must also be elementary. 

Theorem 237 (Itô Integrals of Approximating Elementary Processes

Converge) Let X and Xn be as in Lemma 235. Then the sequence In (X) ≡
Z b
Xn (t)dW (19.17)
a

has a limit in L2 . Moreover, this limit is the same for any such approximating
sequence Xn .

Proof: For each Xn , In (X(ω)) is an S2 function of ω, by the fact that Xn

is square-integrable and Lemma 236. Now, the Xn are converging on X, in the
sense that
kX − Xn kS2 → 0 (19.18)
Since (Proposition 230) the space S2 [a, b] is complete, every convergent sequence
is also Cauchy, and so, for every  > 0, there exists an n such that

kXn+k − Xn kS2 <  (19.19)

for every positive k. Since Xn and Xn+k are both elementary processes, their
difference is also elementary, and we can apply Lemma 236 to it. That is, for
every  > 0, there is an n such that
 !2 
Z b
E (Xn+k (t) − Xn (t))dW  <  (19.20)
a

for all k. (Chose the  in Eq. 19.19 to be the square root of the  in Eq. 19.20.)
But this is to say that In (X) is a Cauchy sequence in L2 , therefore it has a
limit, which is also in L2 (because L2 is complete). If Yn is another sequence
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 138

of approximations of X by elementary processes, parallel arguments show that

the Itô integrals of Yn are a Cauchy sequence, and that for every  > 0, there
exist m and n such that kXn+k − Ym+l kS2 ≤ , hence the integrals of Yn must
be converging on the same limit as the integrals of Xn . 

Definition 238 (Itô integral) Let X be progressive, non-anticipative and square-

integrable on [a, b]. Then its Itô integral is
Z b Z b
X(t)dW ≡ lim Xn (t)dW (19.21)
a n a

taking the limit in L2 , with Xn as in Lemma 235. We will say that X is Itô-
integrable on [a, b].

Corollary 239 (The Itô isometry) Eq. 19.10 holds for all Itô-integrable X.

Proof: Obvious from the approximation by elementary processes and Lemma

236.
This would be a good time to do Exercises 19.1, 19.2 and 19.3.

19.2 Some Easy Stochastic Integrals, with a Moral

R
19.2.1 dW
Let’s start with the easiest possible stochastic integral:
Z b
dW (19.22)
a

i.e., the Itô integral of the function which is always 1, 1R+ (t). If this is any
kind of integral at all, it should be W — more exactly, because this is a definite
Rb
integral, we want a dW = W (b) − W (a). Mercifully, this works. Pick any set
of time-points ti we like, and treat 1 as an elementary function with those times
as its break-points. Then, using our definition of the Itô integral for elementary
functions,
Z b X
dW = W (ti+1 ) − W (ti ) (19.23)
a ti
= W (b) − W (a) (19.24)

as required. (This would be a good time to convince yourself that adding extra
break-points to an elementary function doesn’t change its integral (Exercise
19.5.)
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 139
R
19.2.2 W dW
R
Tradition dictates that the next example be W dW . First, we should con-
vince ourselves that W (t) is Itô-integrable: it’s clearly measurable and non-
anticipative, but is it square-integrable? Yes; by Fubini’s theorem,
Z t  Z t
W 2 (s)ds E W 2 (s) ds
 
E = (19.25)
0 0
Z t
= sds (19.26)
0

which is clearly finite on finite intervals [0, t]. So, this integral should exist.
Now, if the ordinary rules for change of variables held — equivalent, if the
chain-rule worked the usual way — we could say that W dW = 21 d(W 2 ), so
Rt
W dW = 12 dW 2 , and we’d expect 0 W dW = 12 W 2 (t). But, alas, this can’t
R R

be right. To see why, take the expectation: it’d be 12 t. But we know that it has
to be zero, and it has to be a martingale in t, and this is neither. A bone-head
would try to fix this byR subtracting off the non-martingale part, i.e., a bone-
t
head would guess that 0 W dW = 12 W 2 (t) − t/2. Annoyingly, in this case the
bone-head is correct. The demonstration is fundamentally straightforward, if
somewhat long-winded.
To begin, we need to approximate W by elementary functions. For each n,
P2n −1
let ti = i 2tn , 0 ≤ i ≤ 2n − 1. Set φn (t) = i=0 W (ti )1[ti ,ti+1 ) . Let’s check that
this converges to W (t) as n → ∞:
Z t "2n −1 Z #
2
 X ti+1 2
E (φn (s) − W (s)) ds = E (W (ti ) − W (s)) ds (19.27)
0 i=0 ti
n
−1
2X Z ti+1 
2
= E (W (ti ) − W (s)) ds (19.28)
i=0 ti
n
−1 Z
2X ti+1 h i
2
= E (W (ti ) − W (s)) ds(19.29)
i=0 ti
n
2X −1 Z ti+1
= (s − ti )ds (19.30)
i=0 ti
n
2X −1 Z 2−n
= sds (19.31)
i=0 0
n
2X −1  2 2−n
t
= (19.32)
i=0
2 0
n
2X −1
= 2−2n−1 (19.33)
i=0
−n−1
= 2 (19.34)
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 140

which → 0 as n → ∞. Hence
Z t Z t
W (s)dW = lim φn (s)dW (19.35)
0 n 0
n
2X −1
= lim W (ti )(W (ti+1 ) − W (ti )) (19.36)
n
i=0
n
2X −1
= lim W (ti )∆W (ti ) (19.37)
n
i=0

where ∆W (ti ) ≡ W (ti+1 ) − W (ti ), because I’m getting tired of writing both
subscripts. Define ∆W 2 (ti ) similarly. Since W (0) = 0 = W 2 (0), we have that
X
W (t) = ∆W (ti ) (19.38)
i
X
2
W (t) = ∆W 2 (ti ) (19.39)
i

Now let’s re-write ∆W 2 in such a way that we get a W ∆W term, which is what
we want to evaluate our integral.
∆W 2 (ti ) = W 2 (ti+1 ) − W 2 (ti ) (19.40)
2 2
= (∆W (ti ) + W (ti )) − W (ti ) (19.41)
2 2 2
= (∆W (ti )) + 2W (ti )∆W (ti ) + W (ti ) − W (ti ) (19.42)
2
= (∆W (ti )) + 2W (ti )∆W (ti ) (19.43)
This looks promising, because it’s got W ∆W in it. Plugging in to Eq. 19.39,
X 2
W 2 (t) = (∆W (ti )) + 2W (ti )∆W (ti ) (19.44)
i
X 1 2 1X 2
W (ti )∆W (ti ) = W (t) − (∆W (ti )) (19.45)
i
2 2 i

Now, it is possible to show (Exercise 19.4) that

n
2X −1
2
lim (∆W (ti )) = t (19.46)
n
i=0

in L2 , so we have that
n
Z t 2X −1
W (s)dW = lim W (ti )∆W (ti ) (19.47)
0 n
i=0
n
2X−1
1 2 2
= W (t) − lim (∆W (ti )) (19.48)
2 n
i=0
1 2 t
= W (t) − (19.49)
2 2
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 141

as required.
Clearly, something weird is going on here, and it would be good to get to the
bottom of this. At the very least, we’d like to be able to use change of variables,
so that we can find functions of stochastic integrals.

19.3 Itô’s Formula

R
Integrating W dW has taught us two things: first, we want to avoid evaluating
Itô integrals directly from the definition; and, second, there’s something funny
about change of variables in Itô integrals. A central result of stochastic calculus,
known as Itô’s formula, gets us around both difficulties, by showing how to write
functions of stochastic integrals as, themselves, stochastic integrals.

Definition 240 (Itô Process) If A is a non-anticipating measurable process,

B is Itô-integrable,
R t and X0Ris an L2 random variable independent of W , then
t
X(t) = X0 + 0 A(s)ds + 0 B(s)dW is an Itô process. This is equivalently
written dX = Adt + BdW .

Lemma 241 (Itô processes are non-anticipating) Every Itô process is

non-anticipating.

Proof: Clearly, the non-anticipating processes are closed under linear oper-
ations, so it’s enough to show that the three components of any Itô process are
|=
non-anticipating. But a process which is always equal to X R 0 W (t) is clearly
non-anticipating. Similarly, since A(t) is non-anticipating, A(s)ds is too: its
natural filtration is smaller than that of A, so it cannot provide more infor-
mation about W (t), and A is, by assumption,
R non-anticipating. Finally, Itô
integrals are always non-anticipating, so B(s)dW is non-anticipating. 

Theorem 242 (Itô’s Formula in One Dimension) Suppose X is an Itô

process with dX = Adt + BdW . Let f (t, x) : R+ × R 7→ R be a function
with continuous partial time derivative ∂f
∂t , and continuous second partial space
∂2f
derivative, ∂x2 . Then F (t) = f (t, X(t)) is an Itô process, and

∂f ∂f 1 ∂2f
dF = (t, X(t))dt + (t, X(t))dX + B 2 (t) 2 (t, X(t))dt (19.50)
∂t ∂x 2 dx
That is,

F (t) − F (0) = (19.51)

Z t Z t
1 2 ∂2f

∂f ∂f ∂f
(s, X(s)) + A(s) (s, X(s)) + B (s) 2 (s, X(s)) dt + B(s) (s, X(s))dW
0 ∂t ∂x 2 ∂x 0 ∂x

Proof: I will suppose first of all that f , and its partial derivatives appear-
ing in Eq. 19.50, are all bounded. (You can show that the general case of C 2
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 142

functions can be uniformly approximated by functions with bounded deriva-

tives.) I will further suppose that A and B are elementary processes, since in
the last chapter we saw how to use them to approximate general Itô-integrable
functions. (If you are worried about the interaction of all these approximations
and simplifications, I commend your caution, and suggest you step through the
proof in the general case.)
For each n, let ti = i 2tn , as in the last section. Define ∆ti ≡ ti+1 − ti ,
∆X(ti ) = X(ti+1 ) − X(ti ), etc. Thus
n
2X −1
F (t) = f (t, X(t)) = f (0, X(0)) + ∆f (ti , X(ti )) (19.52)
i=0
Now we’ll approximate the increments of F by a Taylor expansion:
n
2X −1
∂f
F (t) = f (0, X(0)) + ∆ti (19.53)
i=0
∂t
n
2X −1
∂f
+ ∆X(ti )
i=0
∂x
n
2 −1
1 X ∂2f 2
+ (∆ti )
2 i=0 ∂t2
n
2X −1
∂2f
+ ∆ti ∆X(ti )
i=0
∂t∂x
n
2 −1
1 X ∂2f 2
+ (∆X(ti ))
2 i=0 ∂x2
n
2X −1
+ Ri
i=0
Because the derivatives are bounded, all the remainder terms Ri are of third
order,
 
3 2 2 3
Ri = O (∆ti ) + ∆X(ti )(∆ti ) + (∆X(ti )) ∆ti + (∆X(ti )) (19.54)
We will come back to showing that the remainders are harmless, but for now
let’s concentrate on the leading-order components of the Taylor expansion.
First, as n → ∞,
n
2X −1 Z t
∂f ∂f
∆ti → ds (19.55)
i=0
∂t 0 ∂t
n
2X −1 Z t
∂f ∂f
∆X(ti ) → dX (19.56)
i=0
∂x 0 ∂x
Z t Z t
∂f ∂f
≡ A(s)dt + B(s)dW (19.57)
0 ∂x 0 ∂x
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 143

[You can use the definition in the last line to build up a theory of stochastic
integrals with respect to arbitrary Itô processes, not just Wiener processes.]
n
2X −1 t
∂2f ∂2f
Z
2
(∆ti ) → 0 ds = 0 (19.58)
i=0
∂t2 0 ∂t2

Next, since A and B are (by assumption) elementary,

n n
2X −1 2X −1
∂2f 2 ∂2f 2 2
(∆X(ti )) = A (ti ) (∆ti ) (19.59)
i=0
∂x2 i=0
∂x2
n
2X −1
∂2f
+2 A(ti )B(ti )∆ti ∆W (ti )
i=0
∂x2
n
2X −1
∂2f 2 2
+ B (ti )(∆W (ti ))
i=0
∂x2

2
The first term on the right-hand side, in (∆t) , goes to zero as n increases.
2 P ∂2f 2
Since A is square-integrable and ∂∂xf2 is bounded, ∂x2 A (ti )∆ti converges to
a finite value as ∆t → 0, so multiplying by another factor ∆t, as n → ∞, gives
zero. (This is the same argument as the one for Eq. 19.58.) Similarly for the
second term, in ∆t∆X:
n
2X −1 t
∂2f ∂2f
Z
t
lim 2
A(ti )B(ti )∆ti ∆W (ti ) = lim n A(s)B(s)dW (19.60)
n
i=0
∂x n 2 0 ∂x2

because A and B are elementary and the partial derivative is bounded. Now
apply the Itô isometry:
" Z
t 2 2 # "Z 
t 2 #
t ∂ f t2 ∂2f 2 2
E A(s)B(s)dW = 2n E A (s)B (s)ds
2n 0 ∂x2 2 0 ∂x2

The time-integral on the right-hand side is finite, since A and B are square-
integrable and the partial derivative is bounded, and so, as n grows, both sides
go to zero. But this means that, in L2 ,
n
2X −1
∂2f
A(ti )B(ti )∆ti ∆W (ti ) → 0 (19.61)
i=0
∂x2

2
The third term, in (∆X) , does not vanish, but rather converges in L2 to a time
integral:
n
2X −1 t
∂2f 2 ∂2f 2
Z
2
B (ti )(∆W (ti )) → B (s)ds (19.62)
i=0
∂x2 0 ∂x2
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 144

You will prove this in Exercise 19.4.

The mixed partial derivative term has no counterpart in Itô’s formula, so it
needs to go away.
n n
2X −1 2X −1
∂2f ∂2f h 2
i
∆ti ∆X(ti ) = A(ti )(∆ti ) + B(ti )∆ti ∆W (ti ) (19.63)
i=0
∂t∂x i=0
∂t∂x

n
2X −1
∂2f 2
A(ti )(∆ti ) → 0 (19.64)
i=0
∂t∂x
n
2X −1
∂2f
B(ti )∆ti ∆W (ti ) → 0 (19.65)
i=0
∂t∂x

where the argument for Eq. 19.65 is the same as that for Eq. 19.58, while that
for Eq. 19.65 follows the pattern of Eq. 19.61.
Let us, as promised, dispose of the remainder terms, given by Eq. 19.54,
re-stated here for convenience:
 
3 2 2 3
Ri = O (∆t) + ∆X(∆t) + (∆X) ∆t + (∆X) (19.66)
Taking ∆X = A∆t + B∆W , expanding the powers of ∆X, and using the fact
that everything is inside a big O to let us group together terms with the same
powers of ∆t and ∆W , we get
 
3 2 2 3
Ri = O (∆t) + ∆W (∆t) + (∆W ) ∆t + (∆W ) (19.67)
2
From our previous uses of Exercise 19.4, it’s clear that in the limit (∆W ) terms
will behave like ∆t terms, so
 
3 2 2
Ri = O (∆t) + ∆W (∆t) + (∆t) + ∆W ∆t (19.68)
2
Now, by our previous arguments, the sum of terms which are O((∆t) ) → 0, so
the first three terms all go to zero; similarly we have seen that a sum of terms
which are O(∆W ∆T ) → 0. We may conclude that the sum of the remainder
terms goes to 0, in L2 , as n increases.
Putting everything together, we have that
Z t Z t
1 ∂2f

∂f ∂f ∂f
F (t) − F (0) = + A + B 2 2 dt + BdW (19.69)
0 ∂t ∂x 2 ∂x 0 ∂x

exactly as required. This completes the proof, under the stated restrictions on
f , A and B; approximation arguments extend the result to the general case. 
Remark 1. Our manipulations in the course of the proof are often summa-
rized in the following multiplication rules for differentials: dtdt = 0, dW dt = 0,
dtdW = 0, and, most important of all,
dW dW = dt
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 145

This last is of course related to the linear growth of the variance of the increments
of the Wiener process. If we used a different driving noise term, it would be
replaced by the corresponding rule for the growth of that noise’s variance.
Remark 2. Re-arranging Itô’s formula a little yields
∂f ∂f 1 ∂2f
dF = dt + dX + B 2 2 dt (19.70)
∂t ∂x 2 ∂x
The first two terms are what we expect from the ordinary rules of calculus; it’s
2
the third term which is new and strange. Notice that it disappears if ∂∂xf2 = 0.
When we come to stochastic differential equations, this will correspond to state-
independent noise.
Remark 3. One implication of Itô’s formula is that Itô processes are closed
under the application of C 2 mappings.
R
Example 243 (Section 19.2.2 summarized) The integral W dW is now
trivial. Let X(t) = W (t) (by setting A = 0, B = 1 in the definition of an Itô
process), and f (t, x) = x2 /2. Applying Itô’s formula,

∂f ∂f 1 ∂2f
dF = dt + dW + dt (19.71)
∂t ∂x 2 ∂x2
1 1
dW 2 = W dW + dt (19.72)
Z2 Z 2 Z
1 1
dW 2 = W dW + dt (19.73)
2 2
Z t
1 2 t
W (s)dW = W (t) − (19.74)
0 2 2
All of this extends naturally to higher dimensions.

Definition 244 (Multidimensional Itô Process) Let A by an n-dimensional

vector of non-anticipating processes, B an n × m matrix of Itô-integrable pro-
cesses, and W an m-dimensional Wiener process. Then
Z t Z t
X(t) = X(0) + A(s)ds + B(s)dW (19.75)
0 0
dX = A(t)dt + B(t)dW (19.76)

is an n-dimensional Itô process.

Theorem 245 (Itô’s Formula in Multiple Dimensions) Let X(t) be an

n-dimensional Itô process, and let f (t, x) : R+ × Rn 7→ Rm have a continuous
partial time derivative and continuous second partial space derivatives. Then
F (t) = f (t, X(t)) is an m-dimensional Itô process, whose k th component Fk is
given by
∂gk ∂gk 1 ∂ 2 gk
dFk = dt + dXi + dXi dXj (19.77)
∂t ∂xi 2 ∂Xi ∂Xj
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 146

summing over repeated indices, with the understanding that dWi dWj = δij dt,
dWi dt = dtdWi = dtdt = 0.

Proof: Entirely parallel to the one-dimensional case, only with even more
algebra. 
It is also possible to define Wiener processes and stochastic integrals on
arbitrary curved manifolds, but this would take us way, way too far afield.

19.3.1 Stratonovich Integrals

It is possible to make the extra term in Eq. 19.70 go away, and have stochastic
differentials which work just like the ordinary ones. This corresponds to making
stochastic integrals limits of sums of the form
X  ti+1 + ti 
X ∆W (ti )
i
2

rather than the Itô sums we are using,

X
X(ti )∆W (ti )
i

That is, we could evade the Itô formula if we evaluated our test function in
the middle of intervals, rather than at their beginnning. This leads to what are
called Stratonovich integrals. However, while Stratonovich integrals give simpler
change-of-variable formulas, they have many other inconveniences: they are not
martingales, for instance, and the nice connections between the form of an SDE
and its generator, which we will see and use in the next chapter, go away.
Fortunately, every Stratonovich SDE can be converted into an Itô SDE, and
vice versa, by adding or subtracting the appropriate noise term.

19.3.2 Martingale Representation

One property of the Itô integral is that it is always a continuous square-integrable
martingale. Remarkably enough, the converse is also true. In the interest of
time, I omit the proof of the following theorem; there is one using only tools
we’ve seen so far in Øksendal (1995, ch. 4), but it builds up from some auxiliary
results.

Theorem 246 (Representation of Martingales as Stochastic Integrals

(Martingale Representation
 Theorem)) Let M (t) be a continuous martin-
gale, with E M 2 (t) < ∞ for all t ≥ 0. Then there exists a unique process
M 0 (t), Itô-integrable for all finite positive t, such that
Z t
M (t) = E [M (0)] + M 0 (t)dW a.s. (19.78)
0
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 147

A direct consequence of this and Itô’s formula is the promised martingale

characterization of the Wiener process.

Theorem 247 (Martingale Characterization of the Wiener Process)

If M is continuous martingale, and M 2 − t is also a martingale, then M is a
Wiener process.

Proof: Since M is a continuous martingale, by Theorem 246 there is a

unique process M 0 such that dM = M 0 dw. That is, M is an Itô process with
A = 0, B = M 0 . Now define f (t, x) = x2 − t. Clearly, M 2 (t) − t = f (t, M ) ≡
Y (t). Since f is smooth, we can apply Itô’s formula (Theorem 242):

∂f ∂f 1 ∂2f
dY = dt + dM + B 2 2 dt (19.79)
∂t ∂x 2 ∂x
= −dt + 2M M 0 dW + (M 0 )2 dt (19.80)

Since Y is itself a martingale, dY = Y 0 dW , and this is the unique representation

as an Itô process, hence the dt terms must cancel. Therefore

0 = −1 + (M 0 (t))2 (19.81)
±1 = M 0 (t) (19.82)
d
Since −W is also a Wiener process, it follows that M = W (plus a possible
additive). 

19.4 Stochastic Differential Equations

Definition 248 (Stochastic Differential Equation, Solutions) Let a(x) :
Rn 7→ Rn and b(x) : Rn 7→ Rnm be measurable functions (vector and matrix val-
ued, respectively), W an m-dimensional Wiener process, and X0 an L2 random
variable in Rn , independent of W . Then an Rn -valued stochastic process X on
R+ is a solution to the autonomous stochastic differential equation

dX = a(X)dt + b(X)dW, X(0) = X0 (19.83)

when, with probability 1, it is equal to the corresponding Itô process,

Z t Z s
X(t) = X0 + a(X(s))ds + b(X(s))dW a.s. (19.84)
0 0

The a term is called the drift, and the b term the diffusion.

Remark 1: A given process X can fail to be a solution either because it

happens not to agree with Eq. 19.84, or, perhaps more seriously, because the
integrals on the right-hand side don’t even exist. This can, in particular, hap-
pen if b(X(t)) is anticipating. For a fixed choice of Wiener process, there are
circumstances where otherwise reasonable SDEs have no solution, for basically
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 148

this reason — the Wiener process is constructed in such a way that the class of
Itô processes is impoverished. This leads to the idea of a weak solution to Eq.
19.83, which is a pair X, W such that W is a Wiener process, with respect to
the appropriate filtration, and X then is given by Eq. 19.84. I will avoid weak
solutions in what follows.
Remark 2: In a non-autonomous SDE, the coefficients would be explicit
functions of time, a(t, X)dt + b(t, X)dW . The usual trick for dealing with
non-autonomous n-dimensional ODEs is turn them into autonomous n + 1-
dimensional ODEs, making xn+1 = t by decreeing that xn+1 (t0 ) = t0 , x0n+1 = 1
(Arnol’d, 1973). This works for SDEs, too: add time as an extra variable with
constant drift 1 and constant diffusion 0. Without loss of generality, therefore,
I’ll only consider autonomous SDEs.
Let’s now prove the existence of unique solutions to SDEs. First, recall how
we do this for ordinary differential equations. There are several approaches,
most of which carry over to SDEs, but one of the most elegant is the “method
of successive approximations”, or “Picard’s method” (Arnol’d, 1973, SS30–31)).
To construct a solution to dx/dt = f (x), x(0) = x0 , this approach uses functions
Rt
xn (t), with xn+1 (t) = x0 + 0 f (xn (s)ds, starting with x0 (t) = x0 . That is, there
is an operator P such that xn+1 = P xn , and x solves the ODE iff it is a fixed
point of the operator. Step 1 is to show that the sequence xn is Cauchy on finite
intervals [0, T ]. Step 2 uses the fact that the space of continuous functions is
complete, with the topology of uniform convergence of compact sets — which,
for R+ , is the same as uniform convergence on finite intervals. So, xn has a
limit. Step 3 is to show that the limit point must be a fixed point of P , that
is, a solution. Uniqueness is proved by showing that there cannot be more than
one fixed point.
Before plunging in to the proof, we need some lemmas: an algebraic triviality,
a maximal inequality for martingales, a consequent maximal inequality for Itô
processes, and an inequality from real analysis about integral equations.
2
Lemma 249 (A Quadratic Inequality) For any real numbers a and b, (a + b) ≤
2a2 + 2b2 .
2
Proof: No matter what a and b are, a2 , b2 , and (a − b) are non-negative,
so
2
(a − b) ≥ 0 (19.85)
a + b2 − 2ab ≥ 0
2
(19.86)
a2 + b2 ≥ 2ab (19.87)
2
2a2 + 2b2 ≥ a2 + 2ab + b2 = (a + b) (19.88)

Definition 250 (Maximum Process) Given a stochastic process X(t), we

define its maximum process X ∗ (t) as sup0≤s≤t |X(s)|.

Remark: Example 79 was of course designed with malice aforethought.

CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 149

Definition 251 (The Space QM(T )) Let QM(T ), T > 0, be the space of all
non-anticipating processes, square-integrable on [0, T ], with norm kXkQM(T ) ≡
kX ∗ (T )k2 .
Technically, this is only a norm on equivalence classes of processes, where
the equivalence relation is “is a version of”. You may make that amendment
mentally as you read what follows.
Lemma 252 (Completeness of QM(T )) QM(T ) is a complete normed space
for each T .
Proof: Identical to the usual proof that Lp spaces are complete, see, e.g.,
Lemma 1.31 of Kallenberg (2002, p. 15). 
Proposition 253 (Doob’s Martingale Inequalities) If M (t) is a continu-
ous martingale, then, for all p ≥ 1, t ≥ 0 and  > 0,
p
E [|M (t)| ]
P (M ∗ (t) ≥ ) ≤ (19.89)
p
kM ∗ (t)kp ≤ qkM (t)kp (19.90)

where q −1 + p−1 = 1. In particular, for p = q = 2,

h i
2
E (M ∗ (t)) ≤ 4E M 2 (t)
 

Proof: See Propositions 7.15 and 7.16 in Kallenberg (pp. 128 and 129). 
These can be thought of as versions of the Markov inequality, only for mar-
tingales. They accordingly get used all the time.
Lemma 254 (A Maximal Inequality for Itô Processes) Let X(t) be an
Rt Rt
Itô process, X(t) = X0 + 0 A(s)ds + 0 B(s)dW . Then there exists a constant
C, depending only on T , such that, for all t ∈ [0, T ],
 Z t 
2  2 2 2
kXkQM(t) ≤ C E X0 + E A (s) + B (s)ds (19.91)
0

Proof: Clearly,
Z t Z s 
X ∗ (t) ≤
 
|X0 | + |A(s)|ds + sup  B(s)dW  (19.92)
0 0≤s≤t 0
Z t 2  Z s  2
2
(X ∗ (t)) ≤ 2X02 + 2 B(s0 )dW  (19.93)
 
|A(s)|ds + 2 sup 
0 0≤s≤t 0

by Lemma 249. By Jensen’s inequality1 ,

Z t 2 Z t
|A(s)|ds ≤ t A2 (s)ds (19.94)
0 0
1 Remember that Lebesgue measure isn’t a probability measure on [0, t], but 1 ds is a
t
probability measure, so we can apply Jensen’s inequality to that. This is where the t on the
right-hand side will come from.
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 150

Rt
Writing I(t) for 0
B(s)dW , and noticing
h that itiis a martingale, we have, from
∗ 2  2 
Doob’s inequality (Proposition 253), E (I (t)) ≤ 4E I (t) . But, from Itô’s
hR i
  t
isometry (Corollary 239), E I 2 (t) = E 0 B 2 (s)ds . Putting all the parts
together, then,
h i  Z t Z t 
2
E (X ∗ (t)) ≤ 2E X02 + 2E t A2 (s)ds + B 2 (s)ds (19.95)
 
0 0

and the conclusion follows, since t ≤ T . 

Remark: The lemma also holds for multidimensional Itô processes, and for
powers greater than two (though then the Doob inequality needs to be replaced
by a different one: see Rogers and Williams (2000, Ch. V, Lemma 11.5, p. 129)).

Definition 255 (Picard operator) Given an SDE dX = a(X)dt + b(X)dW

with initial condition X0 , the corresponding integral operator PX0 ,a,b is defined
for all Itô processes Y as
Z t Z t
PX0 ,a,b Y (t) = X0 + a(Y (s))ds + b(Y (s))dW (19.96)
0 0

Lemma 256 (Solutions are fixed points of the Picard operator) Y is a

solution of dX = a(X)dt + b(X)dW , X(0) = X0 , if and only if PX0 ,a,b Y = Y
a.s.

Proof: Obvious from the definitions. 

Lemma 257 (A maximal inequality for Picard iterates) If a and b are

uniformly Lipschitz continuous, with constants Ka and KB , then, for some
positive D depending only on T , Ka and Kb ,
Z t
2 2
kPX0 ,a,b X − PX0 ,a,b Y kQM(t) ≤ D kX − Y kQM(s) ds (19.97)
0

Proof: Since the SDE is understood to be fixed, abbreviate PX0 ,a,b by P .

Let X and Y be any two Itô processes. We want to find the QM(t) norm of
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 151

PX − PY .

|P X(t) − P Y (t)| (19.98)

Z t Z t 
 
=  a(X(s)) − a(Y (s))dt + b(X(s)) − b(Y (s))dW 
0 0
Z t Z t
≤ |a(X(s)) − a(Y (s))| ds + |b(X(s)) − b(Y (s))| dW (19.99)
0 0
Z t Z t
≤ Ka |X(s) − Y (s)| ds + Kb |X(s) − Y (s)| dW (19.100)
0 0
2
kP X − P Y kQM(t) (19.101)
Z t 
2
≤ C(Ka2 + Kb2 )E |X(s) − Y (s)| ds
0
Z t
2
≤ C(Ka2 + Kb2 )t kX − Y kQM(s) ds (19.102)
0

which, as t ≤ T , completes the proof. 

Lemma 258 (Gronwall’s Inequality) If f is continuous function on [0, T ]

Rt
such that f (t) ≤ c1 + c2 0 f (s)ds, then f (t) ≤ c1 ec2 t .

Proof: See Kallenberg, Lemma 21.4, p. 415. 

Theorem 259 (Existence and Uniquness of Solutions to SDEs in One

Dimension) Let X0 , a, b and W be as in Definition 248, and let a and b
be uniformly Lipschitz continuous. Then there exists a square-integrable, non-
anticipating X(t) which solves dX = a(X)dt + b(X)dW with initial condition
X0 , and this solution is unique (almost surely).

Proof: I’ll first prove existence, along with square-integrability, and then
uniqueness. That X is non-anticipating follows from the fact that it is an Itô
process (Lemma 241). For concision, abbreviate PX0 ,a,b by P .
As with ODEs, iteratively construct approximate solutions. Fix a T > 0,
and, for t ∈ [0, T ], set

X0 (t) = X0 (19.103)
Xn+1 (t) = P Xn (t) (19.104)

The first step is showing that Xn is Cauchy in QM(T ). Define φn (t) ≡

2 2
kXn+1 − Xn kQM(t) . Notice that φn (t) = kP Xn − P Xn−1 kQM(t) , and that,
for each n, φn (t) is non-decreasing in t (because of the supremum embedded in
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 152

its definition). So, using Lemma 257,

Z t
2
φn (t) ≤ D kXn − Xn−1 kQM(s) ds (19.105)
0
Z t
≤ D φn−1 (s)ds (19.106)
0
Z t
≤ D φn−1 (t)ds (19.107)
0
= Dtφn−1 (0) (19.108)
D n tn
≤ φ0 (t) (19.109)
n!
n n
D t
≤ φ0 (T ) (19.110)
n!
Since, for any constant c, cn /n! → 0, to get the successive approximations to be
Cauchy, we just need to show that φ0 (T ) is finite, using Lemma 254.
2
φ0 (T ) = kPX0 ,a,b, X0 − X0 kQM(T ) (19.111)
Z t Z t 2

=
a(X 0 )ds + b(X 0 )dW
(19.112)
0 0 QM(T )
"Z #
T
≤ CE a2 (X0 ) + b2 (X0 )ds (19.113)
0

≤ CT E a2 (X0 ) + b2 (X0 )
 
(19.114)
Because a and b are Lipschitz, this will be finite if X0 has a finite second moment,
which, by assumption, it does. So Xn is a Cauchy sequence in QM(T ), which
is a complete space, so Xn has a limit in QM(T ), call it X:
lim kX − Xn kQM(T ) = 0 (19.115)
n→∞

The next step is to show that X is a fixed point of the operator P . This is
true because P X is also a limit of the sequence Xn .
2 2
kP X − Xn+1 kQM(T ) = kP X − P Xn kQM(T ) (19.116)
2
≤ DT kX − Xn kQM(T ) (19.117)

which → 0 as n → ∞ (by Eq. 19.115). So P X is the limit of Xn+1 , which means

it is the limit of Xn , and, since X is also a limit of Xn and limits are unique,
P X = X. Thus, by Lemma 256, X is a solution.
To prove uniqueness, suppose that there were another solution, Y . By
Lemma 256, P Y = Y as well. So, with Lemma 257,
2 2
kX − Y kQM(t) = kP X − P Y kQM(t) (19.118)
Z t
2
≤ D kX − Y kQM(s) ds (19.119)
0
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 153

So, from Gronwall’s inequality (Lemma 258), we have that kX − Y kQM(t) ≤ 0

for all t, implying that X(t) = Y (t) a.s. 
Remark: For an alternative approach, based on Euler’s method (rather than
Picard’s), see Fristedt and Gray (1997, §33.4). It has a certain appeal, but it
also involves some uglier calculations. For a side-by-side comparison of the two
methods, see Lasota and Mackey (1994).

Theorem 260 (Existence and Uniqueness for Multidimensional SDEs)

Theorem 259 also holds for multi-dimensional stochastic differential equations,
provided a and b are uniformly Lipschitz in the appropriate Euclidean norms.

Proof: Entirely parallel to the one-dimensional case, only with more alge-
bra. 
The conditions on the coefficients can be reduced to something like “locally
Lipschitz up to a stopping time”, but it does not seem profitable to pursue this
here. See Rogers and Williams (2000, Ch. V, Sec. 12).

19.5 Brownian Motion, the Langevin Equation,

and Ornstein-Uhlenbeck Processes
The Wiener process is not a realistic model of Brownian motion, because it
implies that Brownian particles do not have well-defined velocities, which is
absurd. Setting up a more realistic model will eliminate this absurdity, and
illustrate how SDEs can be used as models.2 I will first need to summarize
classical mechanics in one paragraph.
Classical mechanics starts with Newton’s laws of motion. The zeroth law,
implicit in everything, is that the laws of nature are differential equations in
position variables with respect to time. The first law says that they are not
first-order differential equations. The second law says that they are second-order
differential equations. The usual trick for higher-order differential equations is
to introduce supplementary variables, so that we have a higher-dimensional
system of first-order differential equations. The supplementary variable here is
momentum. Thus, for particle i, with mass mi ,

d~xi p~i
= (19.120)
dt mi
d~pi F (x, p, t)
= (19.121)
dt mi
constitute the laws of motion. All the physical content comes from specifying
the force function F (x, p, t). We will consider only autonomous systems, so we
do not need to deal with forces which are explicit functions of time. Newton’s
2 See Selmeczi et al. (2006) for an account of the physicalm theory of Brownian motion,

including some of its history and some fascinating biophysical applications. Wax (1954)
collects classic papers on this and related subjects, still very much worth reading.
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 154

third law says that total momentum is conserved, when all bodies are taken into
account.
Consider a large particle of (without loss of generality) mass 1, such as a
pollen grain, sitting in a still fluid at thermal equilibrium. What forces act on
it? One is drag. At a molecular level, this is due to the particle colliding with
the molecules (mass m) of the fluid, whose average momentum is zero. This
typically results in momentum being transferred from the pollen to the fluid
molecules, and the amount of momentum lost by the pollen is proportional to
what it had, i.e., one term in d~ p/dt is −γ~p. In addition, however, there will be
fluctuations, which will be due to the fact that the fluid molecules are not all at
rest. In fact, because the fluid is at equilibrium, the momenta of the molecules
will follow a Maxwell-Boltzmann distribution,
p2
−3/2 − 12 molec
f (~
pmolec ) = (2πmkB T ) e mkB T

where which is a zero-mean Gaussian with variance mkB T . Tracing this through,
we expect that, over short time intervals in which the pollen grain nonetheless
collides with a large number of molecules, there will be a random impulse (i.e.,
random change in momentum) which is Gaussian, but uncorrelated over shorter
sub-intervals (by the functional CLT). That is, we would like to write
p = −γ~
d~ pdt + DIdW (19.122)
where D is the diffusion constant, I is the 3 × 3 identity matrix, and W of
course is the standard three-dimensional Wiener process. This is known as the
Langevin equation in the physics literature, as this model was introduced by
Langevin in 1907 as a correction to Einstein’s 1905 model of Brownian motion.
(Of course, Langevin didn’t use Wiener processes and Itô integrals, which came
much later, but the spirit was the same.) If you like time-series models, you
might recognize this as a continuous-time version of an mean-reverting AR(1)
model, which explains why it also shows up as an interest rate model in financial
theory.
We can consider each component of the Langevin equation separately, be-
cause they decouple, and solve them easily with Itô’s formula:
d(eγt p) = Deγt dW (19.123)
Z t
eγt p(t) = p0 + D eγs dW (19.124)
0
Z t
p(t) = p0 e−γt + D e−γ(t−s) dW (19.125)
0
We will see in the next chapter a general method of proving that solutions of
equations like 19.122 are Markov processes; for now, you can either take that
on faith, or try to prove it yourself.
Assuming p0 is itself Gaussian, with mean 0 and variance σ 2 , then (using
Exercise 19.6), p~ always has mean zero, and the covariance is
D2  −γ|s−t| 
p(t), p~(s)) = σ 2 e−γ(s+t) +
cov (~ e − e−γ(s+t) (19.126)
2γ
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 155

If σ 2 = D2 /2γ, then the covariance is a function of |s − t| alone, and the pro-

cess is weakly stationary. Such a solution of Eq. 19.122 is known as a stationary
Ornstein-Uhlenbeck process. (Ornstein and Uhlenbeck provided the Wiener pro-
cesses and Itô integrals.)
Weak stationarity, and the fact that the Ornstein-Uhlenbeck process is Marko-
vian, allow us to say that the distribution N (0, D2 /2γ) is invariant. Now, if
the Brownian particle began in equilibrium, we expect its energy to have a
Maxwell-Boltzmann distribution, which means that its momentum has a Gaus-
sian distribution, and the variance is (as with the fluid molecules) kB T . Thus,
kB T = D2 /2γ, or D2 = 2γkb T . This is an example of what the physics literature
calls a fluctuation-dissipation relation, since one side of the equation involves
the magnitude of fluctuations (the diffusion coefficient D) and the other the re-
sponse to fluctuations (the frictional damping coefficient γ). Such relationships
turn out to hold quite generally at or near equilibrium, and are often summa-
rized by the saying that “systems respond to forcing just like fluctuations”. (Cf.
19.125.)
Oh, and that story I told you before about Brownian particles following
Wiener processes? It’s something of a lie told to children, or at least to proba-
bility theorists, but see Exercise 19.9.
For more on the physical picture of Brownian motion, fluctuation-dissipation
relations, and connections to more general thermodynamic processes in and out
of equilibrium, see Keizer (1987).3

19.6 Exercises
Exercise 19.1 (Basic Properties of the Itô Integral) Prove the following,
first for elementary Itô-integrable processes, and then in general.
1. Z c Z b Z c
X(t)dW = X(t)dW + X(t)dW (19.127)
a a b
almost surely.
2. If c is any real constant, then, almost surely,
Z b Z b Z b
(cX(t) + Y (t))dW = c XdW + Y (t)dW (19.128)
a a a

Exercise 19.2 (Martingale Properties of the Itô Integral) Suppose X is

Itô-integrable on [a, b]. Show that
Z t
Ix (t) ≡ X(s)dW (19.129)
a

a ≤ t ≤ b, is a martingale. What is E[Ix (t)]?

3 Be warned that he perversely writes the probability of event A conditional on event B as

P (B|A), not P (A|B).

CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 156

Exercise 19.3 (Continuity of the Itô Integral) Show that Ix (t) has (a
modification with) continuous sample paths.

Exercise 19.4 (“The square of dW ”) Use the notation of Section 19.2 here.
P 2
1. Show that i (∆W (ti )) converges on t (in L2 ) as n grows. Hint: Show
that the terms in the sum are IID, and that their variance shrinks suffi-
ciently fast as n grows. (You will need the fourth moment of a Gaussian
distribution.)
2. If X(t) is measurable and non-anticipating, show that
n
2X −1 Z t
2
lim X(ti )(∆W (ti )) = X(s)ds (19.130)
n 0
i=0

in L2 .

Exercise 19.5 (Itô integrals of elementary processes do not depend

on the break-points) Let X and Y be two elementary processes which are
Rb Rb
versions of each other. Show that a XdW = a Y dW a.s.

Exercise 19.6 (Itô integrals are Gaussian processes) For any fixed, non-
Rt
random cadlag function f on R+ , let If (t) = 0 f (s)dW .
1. Show that E [If (t)] = 0 for all t.
R t∧s
2. Show cov (If (t), If (s)) = 0 f 2 (u)du.
3. Show that If (t) is a Gaussian process.

Exercise 19.7 (A Solvable SDE) Consider

1 p
dX = Xdt + 1 + X 2 dW (19.131)
2
1. Show that there is a unique solution for every initial value X(0) = x0 .
2. It happens (you do not have to show this) that, for fixed x0 , the the solution
has the form X(t) = φ(W (t)), where φ is a C 2 function. Use Itô’s formula
to find the first two derivatives of φ, and then solve the resulting second-
order ODE to get φ.
3. Verify that, with the φ you found in the previous part, φ(W (t)) solves Eq.
19.131 with initial condition X(0) = x0 .
CHAPTER 19. STOCHASTIC INTEGRALS AND SDES 157

Exercise 19.8 (Building Martingales from SDEs) Let X be an Itô process

given by dX = Adt + BdW , and f any C 2 function. Use Itô’s formula to prove
that Z t
1 ∂2f

∂f
f (X(t)) − f (X(0)) − A + B 2 2 dt
0 ∂x 2 ∂x
is a martingale.

Exercise 19.9 (Brownian Motion and the Ornstein-Uhlenbeck Pro-

cess) Consider a Brownian particle whose momentum follows a stationary Ornstein-
Uhlenbeck process, in one spatial dimension (for simplicity). Assume that its
Rt
initial position x(0) is fixed at the origin, and then x(t) = 0 p(t)dt. Show that
as D → ∞ and D/γ → 1, the distribution of x(t) converges to a standard
Wiener process. Explain why this limit is a physically reasonable one.

Exercise 19.10 (Again with the martingale characterization of the

Wiener process) Try to prove Theorem 247, starting from the integral repre-
sentation of M 2 − t and using Itô’s lemma to get the integral representation of
M.
Chapter 20

More on Stochastic
Differential Equations

Section 20.1 shows that the solutions of SDEs are diffusions, and
how to find their generators. Our previous work on Feller processes
and martingale problems pays off here. Some other basic properties
of solutions are sketched, too.
Section 20.2 explains the “forward” and “backward” equations
associated with a diffusion (or other Feller process). We get our
first taste of finding invariant distributions by looking for stationary
solutions of the forward equation.

For the rest of this lecture, whenever I say “an SDE”, I mean “an SDE
satisfying the requirements of the existence and uniqueness theorem”, either
Theorem 259 (in one dimension) or Theorem 260 (in multiple dimensions). And
when I say “a solution”, I mean “a strong solution”. If you are really curious
about what has to be changed to accommodate weak solutions, see Rogers and
Williams (2000, ch. V, sec. 16–18).

20.1 Solutions of SDEs are Diffusions

Solutions of SDEs are diffusions: i.e., continuous, homogeneous strong Markov
processes.

Theorem 261 (Solutions of SDEs are Non-Anticipating and Contin-

uous) The solution of an SDE is non-anticipating, and has a version with
continuous sample paths. If X(0) = x is fixed, then X(t) is FtW -adapted.

Proof: Every solution is an Itô process, so it is non-anticipating by Lemma

241. The adaptation for non-random initial conditions follows similarly. (Infor-
mally: there’s nothing else for it to depend on.) In the proof of the existence

158
CHAPTER 20. MORE ON SDES 159

of solutions, each of the successive approximations is continuous, and we bound

the maximum deviation over time, so the solution must be continuous too. 

Theorem 262 (Solutions of SDEs are Strongly Markov) Let Xx be the

process solving a one-dimensional SDE with non-random initial condition X(0) =
x. Then Xx forms a homogeneous strong Markov family.

Proof: By Exercise 19.8, for every C 2 function f ,

Z t
∂2f

∂f 1 2
f (X(t))−f (X(0))− a(X(s)) (X(s)) + b (X(s)) 2 (X(s)) ds (20.1)
0 ∂x 2 ∂x

is a martingale. Hence, for every x0 , there is a unique, continuous solution to

∂ ∂2
the martingale problem with operator G = a(x) ∂x + 12 b2 (x) ∂x2 and function

class D = C 2 . Since the process is continuous, it is also cadlag. Therefore,

by Theorem 172, X is a homogeneous strong Markov family, whose generator
equals G on C 2 . 
Similarly, for a multi-dimensional SDE, where a is a vector and b is a matrix,
the generator extends1 ai (x)∂i + 12 (bbT )ij (x)∂ij
2
. Notice that the coefficients are
outside the differential operators.
Remark: To see what it is like to try to prove this without using our more
general approach, read pp. 103–114 in Øksendal (1995).

Theorem 263 (SDEs and Feller Diffusions) The processes which solve
SDEs are all Feller diffusions.

Proof: Theorem 262 shows that solutions are homogeneous strong Markov
processes, the previous theorem shows that they are continuous (or can be made
so), and so by Definition 218, solutions are diffusions. For them to be Feller,
d P
we need (i) for every t > 0, Xy (t) → Xx (t) as y → x, and (ii) Xx (t) → x as
t → 0. But, since solutions are a.s. continuous, Xx (t) → x with probability 1,
automatically implying convergence in probability, so (ii) is automatic.
1 Here,and elsewhere, I am going to freely use the Einstein conventions for vector calculus:
repeated indices in a term indicate that you should sum over those indices, ∂i abbreviates
∂ 2 means ∂2 ∂
∂x
, ∂ij ∂x ∂x
, etc. Also, ∂t ≡ ∂t .
i i j
CHAPTER 20. MORE ON SDES 160

To get (i), prove convergence in mean square (i.e. in L2 ), which implies

convergence in distribution.
h i
2
E |Xx (t) − Xy (t)| (20.2)
" Z t Z t 2 #
 
= E x − y +
 a(Xx (s)) − a(Xy (s))ds + b(Xx (s)) − b(Xy (s))dW 
0 0
"Z 2 #
2  t 
≤ |x − y| + E  a(Xx (s)) − a(Xy (s))ds (20.3)
0
"Z
t 2 #
 
+E   b(Xx (s)) − b(Xy (s))dW 
0
"Z 2 #
2  t 
= |x − y| + E  a(Xx (s)) − a(Xy (s))ds (20.4)
0
Z t h i
2
+ E |b(Xx (s)) − b(Xy (s))| ds
0
Z t h i
2 2
≤ |x − y| + K E |Xx (s) − Xy (s)| ds (20.5)
0

for some K ≥ 0, using the Lipschitz properties of a and b. So, by Gronwall’s

Inequality (Lemma 258),
h i
2 2
E |Xx (t) − Xy (t)| ≤ |x − y| eKt (20.6)

This clearly goes to zero as y → x, so Xy (t) → Xx (t) in L2 , which implies

convergence in distribution. 

Corollary 264 (Convergence of Initial Conditions and of Processes)

For a given SDE, convergence in distribution of the initial condition implies
d d
convergence in distribution of the trajectories: if Y → X0 , then XY → XX0 .

Proof: For every initial condition, the generator of the semi-group is the
same (Theorem 262, proof). Since the process is Feller for every initial condition
(Theorem 263), and a Feller semi-group is determined by its generator (Theorem
188), the process has the same evolution operator for every initial condition.
Hence, condition (ii) of Theorem 205 holds. This implies condition (iv) of that
theorem, which is the stated convergence. 

20.2 Forward and Backward Equations

You will often seen probabilists, and applied stochastics people, write about
“forward” and “backward” equations for Markov processes, sometimes with
the eponym “Kolmogorov” attached. We have already seen a version of the
CHAPTER 20. MORE ON SDES 161

“backward” equation for Markov processes, with semi-group Kt and generator

G, in Theorem 158:
∂t Kt f (x) = GKt f (x) (20.7)
Let’s unpack this a little, which will help see where the “backwards” comes from.
First, remember that the operators Kt are really just conditional expectation:

∂t E [f (Xt )|X0 = x] = GE [f (Xt )|X0 = x] (20.8)

Next, turn the expectations into integrals with respect to the transition proba-
bility kernels: Z Z
∂t µt (x, dy)f (y) = G µt (x, dy)f (y) (20.9)

Finally, assume that there is some reference measure λ  µt (x, ·), for all t ∈ T
and x ∈ Ξ. Denote the corresponding transition densities by κt (x, y).
Z Z
∂t dλκt (x, y)f (y) = G dλκt (x, y)f (y) (20.10)
Z Z
dλf (y)∂t κt (x, y) = dλf (y)Gκt (x, y) (20.11)
Z
dλf (y) [∂t κt (x, y) − Gκt (x, y)] = 0 (20.12)

Since this holds for arbitrary nice test functions f ,

∂t κt (x, y) = Gκt (x, y) (20.13)

The operator G alters the way a function depends on x, the initial state. That is,
this equation is about how the transition density κ depends on the starting point,
“backwards” in time. Generally, we’re in a position to know κ0 (x, y) = δ(x − y);
what we want, rather, is κt (x, y) for some positive value of t. To get this, we
need the “forward” equation.
We obtain this from Lemma 155, which asserts that GKt = Kt G.
Z
∂t dλκt (x, y)f (y) = Kt Gf (x) (20.14)
Z
= dλκt (x, y)Gf (y) (20.15)

Notice that here, G is altering the dependence on the y coordinate, i.e. the state
at time t, not the initial state at time 0. Writing the adjoint2 operator as G† ,
Z Z
∂t dλκt (x, y)f (y) = dλG† κt (x, y)f (y) (20.16)

∂t κt (x, y) = G† κt (x, y) (20.17)

2 Recall that, in a vector space with an inner product, such as L , the adjoint of an operator
2
A is another operator, defined through hf, Agi = hA† f, gi. Further recall that L2 is an inner-
product space, where hf, gi = E [f (X)g(X)].
CHAPTER 20. MORE ON SDES 162

N.B., G† is acting on the y-dependence of the transition density, i.e., it says

how the probability density is going to change going forward from t.
In the physics literature, this is called the Fokker-Planck equation, because
Fokker and Planck discovered it, at least in the special case of Langevin-type
equations, in 1913, about 20 years before Kolmogorov’s work on Markov pro-
cesses. Notice that, writing νt for the distribution of Xt , νt = ν0 µt . Assuming
νt has density ρt w.r.t. λ, one can get, by integrating the forward equation over
space,
∂t ρt (x) = G† ρt (x) (20.18)
and this, too, is sometimes called the “Fokker-Planck equation”.
We saw, in the last section, that a diffusion process solving an equation with
drift terms ai (x) and diffusion terms bij (x) has the generator

1
Gf (x) = ai (x)∂i f (x) + (bbT )ij (x)∂ij
2
f (x) (20.19)
2
You can show — it’s an exercise in vector calculus, integration by parts, etc. —
that the adjoint to G is the differential operator
1 2 T
G† f (x) = −∂i ai (x)f (x) + ∂ij (bb )ij (x)f (x) (20.20)
2
Notice that the space-dependence of the SDE’s coefficients now appears inside
the derivatives. Of course, if a and b are independent of x, then they simply
pull outside the derivatives, giving us, in that special case,
1
G† f (x) = −ai ∂i f (x) + (bbT )ij ∂ij
2
f (x) (20.21)
2
Let’s interpret this physically, imagining a large population of independent
tracer particles wandering around the state space Ξ, following independent
copies of the diffusion process. The second derivative term is easy: diffusion
tends to smooth out the probability density, taking probability mass away from
maxima (where f 00 < 0) and adding it to minima. (Remember that bbT is posi-
tive semi-definite.) If ai is positive, then particles tend to move in the positive
direction along the ith axis. If ∂i ρ is also positive, this means that, on average,
the point x sends more particles up along the axis than wander down, against
the gradient, so the density at x will tend to decline.

Example 265 (Wiener process, heat equation) Notice that (for diffusions
produced by SDEs) G† = G when a = 0 and b is constant over the state space.
This is the case with Wiener processes, where G = G† = 12 ∇2 . Thus, the heat
equation holds both for the evolution of observable functions of the Wiener pro-
cess, and for the evolution of the Wiener process’s density. You should convince
yourself that there is no non-negative integrable ρ such that Gρ(x) = 0.
CHAPTER 20. MORE ON SDES 163

Example 266 (Ornstein-Uhlenbeck process) For the one-dimensional Ornstein-

Uhlenbeck process, the generator may be read off from the Langevin equation,
1
Gf (p) = −γp∂p f (p) + D2 ∂pp
2
f (p)
2
and the Fokker-Planck equation becomes
1 2
∂t ρ(p) = γ∂p (pρ(p)) + D2 ∂pp f (p)
2
It’s easily checked that ρ(p) = N (0, D2 /2γ) gives ∂t ρ = 0. That is, the long-run
invariant distribution can be found as a stationary solution of the Fokker-Planck
equation. See also Exercise 20.1.

20.3 Exercises
Exercise 20.1 (Langevin equation with a conservative force) A conser-
vative force is one derived from an external potential, i.e., there is a function
φ(x) giving energy, and F (x) = −dφ/dx. The equations of motion for a body
subject to a conservative force, drag, and noise read
p
dx = dt (20.22)
m
dp = −γpdt + F (x)dt + σdW (20.23)

1. Find the corresponding forward (Fokker-Planck) equation.

2. Find a stationary density for this equation, at least up to normalization
constants. Hint: use separation of variables, i.e., ρ(x, p) = u(x)v(p).
You should be able to find the normalizing constant for the momentum
density v(p), but not for the position density u(x). (Its general form should
however be familiar from theoretical statistics: what is it?)

3. Show that your stationary solution reduces to that of the Ornstein-Uhlenbeck

process, if F (x) = 0.
Chapter 21

Large Deviations for

Small-Noise Stochastic
Differential Equations

This lecture is at once the end of our main consideration of dif-

fusions and stochastic calculus, and a first taste of large deviations
theory. Here we study the divergence between the trajectories pro-
duced by an ordinary differential equation, and the trajectories of the
same system perturbed by a small amount of uncorrelated Gaussian
noise (“white noise”; see Sections 22.1 and 22.2.1).
Section 21.1 establishes that, in the small noise limit, the SDE’s
trajectories converge in probability on the ODE’s trajectory. This
uses Feller-process convergence.
Section 21.2 upper bounds the rate at which the probability of
large deviations goes to zero as the noise vanishes. The methods are
elementary, but illustrate deeper themes to which we will recur once
we have the tools of ergodic and information theory.

In this chapter, we will use the results we have already obtained about
SDEs to give a rough estimate of a basic problem, frequently arising in practice1
namely taking a system governed by an ordinary differential equation and seeing
how much effect injecting a small amount of white noise has. More exactly,
we will put an upper bound on the probability that the perturbed trajectory
goes very far from the unperturbed trajectory, and see the rate at which this
1 For applications in statistical physics and chemistry, see Keizer (1987). For applications

in signal processing and systems theory, see Kushner (1984). For applications in nonparamet-
ric regression and estimation, and also radio engineering (!) see Ibragimov and Has’minskii
(1979/1981). The last book is especially recommended for those who care about the connec-
tions between stochastic process theory and statistical inference, but unfortunately expound-
ing the results, or even just the problems, would require a too-long detour through asymptotic
statistical theory.

164
CHAPTER 21. SMALL-NOISE SDES 165

probability goes to zero as the amplitude  of the noise shrinks; this will be
2
O(e−C ). This will be our first illustration of a large deviations calculation. It
will be crude, but it will also introduce some themes to which we will return
(inshallah!) at greater length towards the end of the course. Then we will see
that the major improvement of the more refined tools is to give a lower bound
to match the upper bound we will calculate now, and see that we at least got
the logarithmic rate right.
I should say before going any further that this example is shamelessly ripped
off from Freidlin and Wentzell (1998, ch. 3, sec. 1, pp. 70–71), which is the
book on the subject of large deviations for continuous-time processes.

21.1 Convergence in Probability of SDEs to ODEs

To begin with, consider an unperturbed ordinary differential equation:
d
x(t) = a(x(t)) (21.1)
dt
x(0) = x0 ∈ Rd (21.2)

Assume that a is uniformly Lipschitz-continuous (as in the existence and unique-

ness theorem for ODEs, and more to the point for SDEs). Then, for the given,
non-random initial condition, there exists a unique continuous function x which
solves the ODE.
Now, for  > 0, consider the SDE

dX = a(X )dt + dW (21.3)

where W is a standard d-dimensional Wiener process, with non-random ini-

tial condition X (0) = x0 . Theorem 260 clearly applies, and consequently so
does Theorem 263, meaning X is a Feller diffusion with generator G f (x) =
2
ai (x)∂i f 0 (x) + 2 ∇2 f (x).
Write X0 for the deterministic solution of the ODE.
d
Our first assertion is that X → X0 as  → 0. Notice that X0 is a Feller
2
process , whose generator is G0 = ai (x)∂i . We can apply Theorem 205 on
convergence of Feller processes. Take the class of functions with bounded second
derivatives. This is clearly a core for G0 , and for every G . For every function
f in this class,

2 2


kG f − G0 f k∞ = a
i i∂ f (x) + ∇ f (x) − a ∂
i i f (x) (21.4)
2
∞
2 ∇2 f (x)

= ∞
(21.5)
2
d
2 You can amuse yourself by showing this. Remember that X (t) → X (t) is equivalent to
y x
E [f (Xt )|X0 = y] → E [f (Xt )|X0 = x] for all bounded continuous f , and the solution of an
ODE depends continuously on its initial condition.
CHAPTER 21. SMALL-NOISE SDES 166

which goes to zero as  → 0. But this is condition (i) of the convergence theorem,
which is equivalent to condition (iv), that convergence in distribution of the
initial condition implies convergence in distribution of the whole trajectory.
Since the initial condition is the same non-random point x0 for all , we have
d P
X → X0 as  → 0. In fact, since X0 is non-random, we have that X → X0 .
That last assertion really needs some consideration of metrics on the space of
continuous random functions to make sense (see Appendix A2 of Kallenberg),
but once that’s done, the upshot is
Theorem 267 (Small-Noise SDEs Converge in Probability on No-Noise
ODEs) Let ∆ (t) = |X (t) − X0 (t)|. For every T > 0, δ > 0,
 
lim P sup ∆ (t) > δ = 0 (21.6)
→0 0≤t≤T

Or, using the maximum-process notation, for every T > 0,

∗ P
∆(T ) → 0 (21.7)
Proof: See above. 
This is a version of the weak law of large numbers, and nice enough in its
own way. One crucial limitation, however, is that it tells us nothing about the
rate of convergence. That is, it leaves us clueless about how big the noise can
be, while still leaving us in the small-noise limit. If the rate of convergence were,
say, O(1/100 ), then this would not be very useful. (In fact, if the convergence
were that slow, we should be really suspicious of numerical solutions of the
unperturbed ODE.)

21.2 Rate of Convergence; Probability of Large

Deviations
Large deviations theory is essentially a study of rates of convergence in prob-
abilistic limit theorems. Here, we will estimate the rate of convergence: our
methods will be crude, but it will turn out that even more refined estimates
won’t change the rate, at least not by more than log factors.
Let’s go back to the difference between the perturbed and unperturbed tra-
jectories, going through our now-familiar procedure.
Z t
X (t) − X0 (t) = [a(X (s)) − a(X0 (s))] ds + W (t) (21.8)
0
Z t
∆ (t) ≤ |a(X (s)) − a(X0 (s))| ds + |W (t)| (21.9)
0
Z t
≤ Ka ∆ (s)ds + |W (t)| (21.10)
0
Z t
∆∗ (T ) ≤ W ∗ (T ) + Ka ∆∗ (s)ds (21.11)
0
CHAPTER 21. SMALL-NOISE SDES 167

Applying Gronwall’s Inequality (Lemma 258),

∆∗ (T ) ≤ W ∗ (T )eKa T (21.12)

The only random component on the RHS is the supremum of the Wiener process,
so we’re in business, at least once we take on two standard results, one about
the Wiener process itself, the other just about multivariate Gaussians.

Lemma 268 (A Maximal Inequality for the Wiener Process) For a

standard Wiener process, P (W ∗ (t) > a) = 2P (|W (t)| > a).

Proof: Proposition 13.13 (pp. 256–257) in Kallenberg. 

Lemma 269 (A Tail Bound for Maxwell-Boltzmann Distributions) If

Z is a d-dimensional standard Gaussian (i.e., mean 0 and covariance matrix
I), then
2
2z d−2 e−z /2
P (|Z| > z) ≤ d/2 (21.13)
2 Γ(d/2)
for sufficiently large z.
qP
d 2
Proof: Each component of Z, Zi ∼ N (0, 1). So |Z| = i=1 Zi has the
density function (see, e.g., Cramér (1945, sec. 18.1, p. 236))
2 z2
d−1 − 2σ
f (z) = z e 2
2d/2 σ d Γ(d/2)

This is the d-dimensional Maxwell-Boltzmann distribution, sometimes called the

2 2
 |Z| is χ-distributed with d degrees of freedom. Notice
χ-distribution, because
2 2
that P (|Z| ≥ z) = P |Z| ≥ z , so we will be able to solve this problem in
 
2
terms of the χ2 distribution. Specifically, P |Z| ≥ z 2 = Γ(d/2, z 2 /2)/Γ(d/2),
where Γ(r, a) is the upper incomplete gamma function. For said function, for
every r, Γ(r, a) ≤ ar−1 e−a for sufficiently large a (Abramowitz and Stegun,
1964, Eq. 6.5.32, p. 263). Hence (for sufficiently large z)
 
2
P (|Z| ≥ z) = P |Z| ≥ z 2 (21.14)
Γ(d/2, z 2 /2)
= (21.15)
Γ(d/2)

d/2−1 1−d/2 −z2 /2
z2 2 e
≤ (21.16)
Γ(d/2)
2
2z d−2 e−z /2
= (21.17)
2d/2 Γ(d/2)


CHAPTER 21. SMALL-NOISE SDES 168

Theorem 270 (Upper Bound on the Rate of Convergence of Small-

Noise SDEs) In the limit as  → 0, for every δ > 0, T > 0,

log P (∆∗ (T ) > δ) ≤ O(−2 ) (21.18)

Proof: Start by directly estimating the probability of the deviation, using

preceding lemmas.

δe−Ka T
 
∗ ∗
P (∆ (T ) > δ) ≤ P |W | (T ) > (21.19)

δe−Ka T
 
= 2P |W (T )| > (21.20)

 2 −2Ka T d/2−1
4 δ e δ 2 e−2Ka T
≤ d/2 2
e− 22 (21.21)
2 Γ(d/2) 

if  is sufficiently small, so that −1 is sufficiently large to apply Lemma 269.

Now take the log and multiply through by 2 :

2 log P (∆∗ (T ) > δ) (21.22)

 
4 d
≤ 2 log d/2 + 2 − 1 log δ 2 e−2Ka T − 2 log  − δ 2 e−2Ka T
 
2 Γ(d/2) 2

lim 2 log P (∆∗ (T ) > δ) ≤ −δ 2 e−2Ka T (21.23)

↓0

since 2 log  → 0, and the conclusion follows. 

Notice several points.
1. Here  gauges the size of the noise, and we take a small noise limit. In many
forms of large deviations theory, we are concerned with large-sample (N →
∞) or long-time (T → ∞) limits. In every case, we will identify some
asymptotic parameter, and obtain limits on the asymptotic probabilities.
There are deviations inequalities which hold non-asymptotically, but they
have a different flavor, and require different machinery. (Some people are
made uncomfortable
√ by an 2 rate, and prefer to write the SDE dX =
a(X)dt + dW so as to avoid it. I don’t get this.)
2. The magnitude of the deviation δ does not change as the noise becomes
small. This is basically what makes this a large deviations result. There
is also a theory of moderate deviations.
3. We only have an upper bound. This is enough to let us know that the
probability of large deviations becomes exponentially small. But we might
be wrong about the rate — it could be even faster than we’ve estimated.
In this case, however, it’ll turn out that we’ve got at least the order of
magnitude correct.
CHAPTER 21. SMALL-NOISE SDES 169

4. We also don’t have a lower bound on the probability, which is something

that would be very useful in doing reliability analyses. It will turn out
that, under many circumstances, one can obtain a lower bound on the
probability of large deviations, which has the same asymptotic dependence
on  as the upper bound.

5. Suppose we’re right about the rate (which, it will turn out, we are), and
it holds both from above and below. It would be nice to be able to say
something like
−2
P (∆∗ (T ) > δ) → C1 (δ, T )e−C2 (δ,T ) (21.24)

rather than
2 log P (∆∗ (T ) > δ) → −C2 (δ, T ) (21.25)
The difficulty with making an assertion like 21.24 is that the large devia-
tion probability actually converges on any function which goes to asymp-
totically to zero! So, to extract the actual rate of dependence, we need to
get a result like 21.25.
More generally, one consequence of Theorem 270 is that SDE trajectories
which are far from the trajectory of the ODE have exponentially small proba-
bilities. The vast majority of the probability will be concentrated around the
unperturbed trajectory. Reasonable sample-path functionals can therefore be
well-approximated by averaging their value over some small (δ) neighborhood of
the unperturbed trajectory. This should sound very similar to Laplace’s method
for the evaluate of asymptotic integrals in Euclidean space, and in fact one of
the key parts of large deviations theory is an extension of Laplace’s method to
infinite-dimensional function spaces.
In addition to this mathematical content, there is also a close connection
to the principle of least action in physics. In classical mechanics, the system
follows the trajectory of least action, the “action” of a trajectory being the
integral of the kinetic minus potential energy along that path. In quantum
mechanics, this is no longer an axiom but a consequence of the dynamics: the
action-minimizing trajectory is the most probable one, and large deviations from
it have exponentially small probability. Similarly, the theory of large deviations
can be used to establish quite general stochastic principles of least action for
Markovian systems.3

3 For a fuller discussion, see Eyink (1996),Freidlin and Wentzell (1998, ch. 3).
Chapter 22

Spectral Analysis and L2

Ergodicity

Section 22.1 makes sense of the idea of white noise. This forms
the bridge from the ideas of Wiener integrals, in the previous lec-
tures, and spectral and ergodic theory, which we will pursue here.
Section 22.2 introduces the spectral representation of weakly sta-
tionary processes, and the central Wiener-Khinchin theorem con-
necting autocovariance to the power spectrum. Subsection 22.2.1
explains why white noise is “white”.
Section 22.3 gives our first classical ergodic result, the “mean
square” (L2 ) ergodic theorem for weakly stationary processes. Sub-
section 22.3.1 gives an easy proof of a sufficient condition, just using
the autocovariance. Subsection 22.3.2 gives a necessary and suffi-
cient condition, using the spectral representation.

Any reasonable real-valued function x(t) of time, t ∈ R, has a Fourier trans-

form, that is, we can write
Z ∞
1
x̃(ν) = dteiνt x(t)
2π −∞

which can usually be inverted to recover the original function,

Z ∞
x(t) = dνe−iνt x̃(ν)
−∞

This one example of an “analysis”, in the original sense of resolving into parts,
of a function into a collection of orthogonal basis functions. (You can find the
details in any book on Fourier analysis, as well as the varying conventions on
where the 2π goes, which side gets the e−iνt , the constraints on x̃ which arise
from the fact that x is real, etc.)

170
CHAPTER 22. SPECTRAL ANALYSIS AND L2 ERGODICITY 171

There are various reasons to prefer the trigonometric basis functions eiνt
over other possible choices. One is that they are invariant under translation
in time, which just changes phases1 . This suggests that the Fourier basis will
be particularly useful when dealing with time-invariant systems. For stochas-
tic processes, however, time-invariance is stationarity. This suggests that there
should be some useful way of doing Fourier analysis on stationary random func-
tions. In fact, it turns out that stationary and even weakly-stationary processes
can be productively Fourier-transformed. This is potentially a huge topic, es-
pecially when it’s expanded to include representing random functions in terms
of (countable) series of orthogonal functions. The spectral theory of random
functions connects Fourier analysis, disintegration of measures, Hilbert spaces
and ergodicity. This lecture will do no more than scratch the surface, covering,
in succession, white noise, the basics of the spectral representation of weakly-
stationary random functions and the fundamental Wiener-Khinchin theorem
linking covariance functions to power spectra, why white noise is called “white”,
and the mean-square ergodic theorem.
Good sources, if you want to go further, are the books of Bartlett (1955,
ch. 6) (from whom I’ve stolen shamelessly), the historically important and in-
spiring Wiener (1949, 1961), and of course Doob (1953). Loève (1955, ch. X) is
highly edifying, particular his discussion of Karhunen-Loève transforms, and the
associated construction of the Wiener process as a Fourier series with random
phases.

22.1 White Noise

Scientists and engineers are often uncomfortable with the SDEs in the way
probabilists write them, because they want to divide through by dt and have
the result mean something. The trouble, of course, is that dW/dt does not,
in any ordinary sense, exist. They, however, are often happier ignoring this
inconvenient fact, and talking about “white noise” as what dW/dt ought to be.
This is not totally crazy. Rather, one can define ξ ≡ dW/dt as a generalized
derivative, one whose value at any given time is a random real linear functional,
rather than a random real number. Consequently, it only really makes sense in
integral expressions (like the solutions of SDEs!), but it can, in many ways, be
formally manipulated like an ordinary function.
One way to begin to make sense of this is to start with a standard Wiener
process W (t), and a C 1 non-random function u(t), and to use integration by
1 If t 7→ t + τ , then x̃(ν) 7→ eiντ x̃(ν).
CHAPTER 22. SPECTRAL ANALYSIS AND L2 ERGODICITY 172

parts:
d dW du
(uW ) = u + W (22.1)
dt dt dt
= u(t)ξ(t) + u̇(t)W (t) (22.2)
Z t Z t
d
(uW )ds = u̇(s)W (s) + u(s)ξ(s)ds (22.3)
0 dt 0
Z t Z t
u(t)W (t) − u(0)W (0) = u̇(s)W (s)ds + u(s)ξ(s)ds (22.4)
0 0
Z t Z t
u(s)ξ(s)ds ≡ u(t)W (t) − u̇(s)W (s)ds (22.5)
0 0

We can take the last line to define ξ, and time-integrals within which it appears.
Notice that the terms on the RHS are well-defined without the Itô calculus: one
is just a product of two measurable random variables, and the other is the time-
integral of a continuous random function. With this definition, we can establish
some properties of ξ.

Proposition 271 (Linearity of White Noise Integrals) ξ(t) is a linear

functional:
Z t Z t Z t
(a1 u1 (s) + a2 u2 (s))ξ(s)ds = a1 u1 (s)ξ(s)ds + a2 u2 (s)ξ(s)ds (22.6)
0 0 0

Proof:
Z t
(a1 u1 (s) + a2 u2 (s))ξ(s)ds (22.7)
0
Z t
= (a1 u1 (t) + a2 u2 (t))W (t) − (a1 u̇1 (s) + a2 u̇2 (s))W (s)ds
0
Z t Z t
= a1 u1 (s)ξ(s)ds + a2 u2 (s)ξ(s)ds (22.8)
0 0

Proposition 272 (White Noise Has Mean Zero) For all t, E [ξ(t)] = 0.

Proof:
Z t Z t 
u(s)E [ξ(s)] ds = E u(s)ξ(s)ds (22.9)
0 0
 Z t 
= E u(t)W (t) − u̇(s)W (s)ds (22.10)
0
Z t
= E [u(t)W (t)] − u̇(s)E [W (t)] ds (22.11)
0
= 0−0=0 (22.12)
CHAPTER 22. SPECTRAL ANALYSIS AND L2 ERGODICITY 173

Rt
Proposition 273 (White Noise and Itô Integrals) For all u ∈ C 1 , 0
u(s)ξ(s)ds =
Rt
0
u(s)dW .

Proof: Apply Itô’s formula to the function f (t, W ) = u(t)W (t):

d(uW )
= W (t)u̇(t)dt + u(t)dW (22.13)
Z t Z t
u(t)W (t) = u̇(s)W (s)ds + u(t)dW (22.14)
0 0
Z t Z t
u(t)dW = u(t)W (t) − u̇(s)W (s)ds (22.15)
0 0
Z t
= u(s)ξ(s)ds (22.16)
0


This could be used to extend the definition of white-noise integrals to any
Itô-integrable process.

Proposition 274 (White Noise is Uncorrelated) ξ has delta-function co-

variance: cov (ξ(t1 ), ξ(t2 )) = δ(t1 − t2 ).

Proof: Since E [ξ(t)] = 0, we just need to show that E [ξ(t1 )ξ(t2 )] = δ(t1 −
t2 ). Remember (Eq. 17.14 on p. 127) that E [W (t1 )W (t2 )] = t1 ∧ t2 .
Z tZ t
u(t1 )u(t2 )E [ξ(t1 )ξ(t2 )] dt1 dt2 (22.17)
0 0
Z t Z t 
= E u(t1 )ξ(t1 )dt1 u(t2 )ξ(t2 )dt2 (22.18)
0 0
"Z
t 2 #
= E u(t1 )ξ(t1 )dt1 (22.19)
0
Z t Z t
E u2 (t1 ) dt1 = u2 (t1 )dt1
 
= (22.20)
0 0

using the preceding proposition, the Itô isometry, and the fact that u is non-
random. But
Z tZ t Z t
u(t1 )u(t2 )δ(t1 − t2 )dt1 dt2 = u2 (t1 )dt1 (22.21)
0 0 0

so δ(t1 − t2 ) = E [ξ(t1 )ξ(t2 )] = cov (ξ(t1 ), ξ(t2 )). 

Proposition 275 (White Noise is Gaussian and Stationary) ξ is a strongly

stationary Gaussian process.

Proof: To show that it is Gaussian, use Exercise 19.6. The mean is constant
for all times, and the covariance depends only on |t1 −t2 |, so it satisfies Definition
50 and is weakly stationary. But a weakly stationary Gaussian process is also
strongly stationary. 
CHAPTER 22. SPECTRAL ANALYSIS AND L2 ERGODICITY 174

22.2 Spectral Representation of Weakly Station-

ary Procesess
This section will only handle spectral representations of real- and complex-
valued one-parameter processes in continuous time. Generalizations to vector-
valued and multi-parameter processes are straightforward; handling discrete
time is actually in some ways more irritating, because of limitations on allowable
frequencies of Fourier components (to the range from −π to π).

Definition 276  (Autocovariance

 Function) Suppose that, for all t ∈ T , X
is real and E X 2 (t) is finite. Then Γ(t1 , t2 ) ≡ E [X(t1 )X(t2 )]−E [X(t1 )] E [X(t2 )]
is the autocovariance function of the process. If the process is weakly station-
ary, so that Γ(t,
 t + τ ) = Γ(0, τ) for all  t, τ , write Γ(τ ). If X(t) ∈ C, then
Γ(t1 , t2 ) ≡ E X † (t1 )X(t2 ) − E X † (t1 ) E [X(t2 )], where † is complex conjuga-
tion.

Lemma 277 (Autocovariance and Time Reversal) If X is real and weakly

stationary, then Γ(τ ) = Γ(−τ ); if X is complex and weakly stationary, then
Γ(τ ) = Γ† (−τ ).

Proof: Direct substitution into the definitions. 

Remarks on terminology. It is common, when only dealing with one stochas-
tic process, to drop the qualifying “auto” and just speak of the covariance func-
tion; I probably will myself. It is also common (especially in the time series
literature) to switch to the (auto)correlation function, i.e., to normalize by the
standard deviations. Finally, be warned that the statistical physics literature
(e.g. Forster, 1975) uses “correlation function” to mean E [X(t1 )X(t2 )], i.e., the
uncentered mixed second moment. This is a matter of tradition, not (despite
appearances) ignorance.

Definition 278 (Second-Order

h i Process) A complex-valued process X is sec-
2
ond order when E |X| (t) < ∞ for all t.

Definition 279 (Spectral Representation, Power Spectrum) A real-valued

process X on T has a complex-valued spectral process X̃, if it has a spectral
representation: Z ∞
X(t) ≡ e−iνt dX̃(ν) (22.22)
−∞
 2 
The power spectrum V (ν) ≡ E X̃(ν) .
 

Remark 1. The name “power spectrum” arises because this is proportional to

the amount of power (energy per unit time) carried by oscillations of frequency
≤ ν, at least in a linear system.
CHAPTER 22. SPECTRAL ANALYSIS AND L2 ERGODICITY 175

Remark 2. Notice that the only part of the right-hand side of Equation
22.22 which depends on t is the integrand, e−iνt , which just changes the phase
of each Fourier component deterministically. Roughly speaking, for a fixed ω the
amplitudes of the different Fourier components in X(t, ω) are fixed, and shifting
forward in time just involves changing their phases. (Making this simple is why
we have to allow X̃ to have complex values.)
The spectral representation is another stochastic integral, like the Itô integral
we saw in Section 19.1. There, the measure of the time interval [t1 , t2 ] was given
by the increment of the Wiener process, W (t2 )−W (t1 ). Here, for any frequency
interval [ν1 , ν2 ], the increment X̃(ν2 ) − X̃(ν1 ) defines a random set function
(admittedly, a complex-valued one). Since those intervals are a generating class
for the Borel σ-field, the usual arguments extend this set functionRuniquely to a
random complex measure on R, B. When we write something like G(ν)dX̃(ν),
we mean an integral with respect to this measure.
Rather than dealing directly with this measure, we can, as with the Itô inte-
gral, use approximation by elementary processes. That is, we should interpret
Z ∞
G(t, ν)dX̃(ν)
−∞

as the L2 limit of sums

∞
X
G(t, ν1 )(X̃(νi+1 ) − X̃(νi ))
νi =−∞

as sup νi+1 − νi goes to zero. This limit can be shown to exist in pretty much
exactly the same way we showed the corresponding limit for Itô integrals to
exist.2

Lemma 280 (Regularity of the Spectral Process) When it exists, X̃(ν)

has right and left limits at every point ν, and limits as ν → ±∞.

Proof: See Loève (1955, §34.4). You can prove this yourself, however, using
the material on characteristic functions in 36-752. 

Definition 281 (Jump of the Spectral Process) The jump of the spec-
tral process at ν is the difference between the right- and left- hand limits at ν,
∆X̃(ν) ≡ X̃(ν + 0) − X̃(ν − 0).

Remark 1: As usual, X̃(ν+0) ≡ limh↓0 X̃(ν + h), and X̃(ν−0) ≡ limh↓0 X̃(ν − h).
Remark 2: Some people call the set of points at which the jump is non-
zero the “spectrum”. This usage comes from functional analysis, but seems
needlessly confusing in the present context.

Proposition 282 (Spectral Representations of Weakly Stationary Pro-

cesses) Every weakly-stationary process has a spectral representation.
2 There is a really excellent discussion of such stochastic integrals, and L stochastic calculus
2
more generally, in Loève (1955, §34).
CHAPTER 22. SPECTRAL ANALYSIS AND L2 ERGODICITY 176

Proof: See Loève (1955, §34.4), or Bartlett (1955, §6.2). 

The following property will be very important for us, since when the spectral
process has it, many nice consequences follow.

Definition 283 (Orthogonal Increments) A one-parameter random func-

tion (real or complex) has orthogonal increments if, for t1 ≤ t2 ≤ t3 ≤ t4 ∈ T ,
the covariance of the increment from t1 to t2 and the increment from t3 to t4 is
always zero:
  † 
E X̃(ν4 ) − X̃(ν3 ) X̃(ν2 ) − X̃(ν1 ) =0 (22.23)

Lemma 284 (Orthogonal Spectral Increments and Weak Stationarity)

The spectral process of a second-order process has orthogonal increments if and
only if the process is weakly stationary.

i Proof: Assume, without loss of generality, that E [X(t)] = 0, so

h Sketch
E X̃(ν) = 0. “If”: Pick any arbitrary t. We can write, using the fact that
X(t) = X † (t) for real-valued processes,

Γ(τ ) = Γ(t, t + τ ) (22.24)

= E X † (t)X(t + τ )
 
(22.25)
Z ∞ Z ∞ 
iν1 t −iν2 (t+τ ) †
= E e e dX̃ν1 dX̃ν2 (22.26)
−∞ −∞
" #
XX
it(ν1 −ν2 ) −iν2 τ †
= E lim e e ∆X̃ (ν1 )∆X̃(ν2 ) (22.27)
∆ν→0
ν1 ν2
XX h i
= lim eit(ν1 −ν2 ) e−iν2 τ E ∆X̃ † (ν1 )∆X̃(ν2 ) (22.28)
∆ν→0
ν1 ν2

where ∆X̃(ν) = X̃(ν+∆ν)−X̃(ν). Since t was arbitrary, everyhterm on the righti

must be independent of t. When ν1 = ν2 , eit(ν1 −ν2 ) = 1, so E ∆X̃ † (ν)∆X̃(ν)
h i
is unconstrained. If ν1 6= ν2 , however, we must have E ∆X̃ † (ν1 )∆X̃(ν2 ) = 0,
which is to say (Definition 283) we must have orthogonal increments.
“Only if”: if the increments are orthogonal, then clearly the steps of the
argument can be reversed to conclude that Γ(t1 , t2 ) depends only on t2 − t1 . 

Definition 285 (Spectral Function, Spectral Density) The spectral func-

tion of a weakly stationary process is the function S(ν) appearing in the spectral
representation of its autocovariance:
Z ∞
Γ(τ ) = e−iντ dSν (22.29)
−∞
CHAPTER 22. SPECTRAL ANALYSIS AND L2 ERGODICITY 177

Remark. Some people prefer to talk about the spectral function as the
Fourier transform of the autocorrelation function, rather than of the autoco-
variance. This has the advantage that the spectral function turns out to be
a normalized cumulative distribution function (see Theorem 286 immediately
below), but is otherwise inconsequential.
Theorem 286 (Weakly Stationary Processes Have Spectral Functions)
The spectral function exists for every weakly stationary process, if Γ(τ ) is con-
tinuous. Moreover, S(ν) ≥ 0, S is non-decreasing, S(−∞) = 0, S(∞) = Γ(0),
and limh↓0 S(ν + h) and limh↓0 S(ν − h) exist for every ν.
Proof: Usually, by an otherwise-obscure result in Fourier analysis called
Bochner’s theorem. A more direct proof is due to Loève. Assume, without loss
of generality, that E [X(t)] = 0.
Start by defining
Z T /2
1
HT (ν) ≡ √ eiνt X(t)dt (22.30)
T −T /2
and define fT (ν) through H:
h i
2πfT (ν) ≡ E HT (ν)HT† (ν) (22.31)
" Z #
1 T /2 T /2 iνt1
Z
= E e X(t1 )e−iνt2 X † (t2 )dt1 dt2 (22.32)
T −T /2 −T /2
1 T /2 T /2 iν(t1 −t2 )
Z Z
= e E [X(t1 )X(t2 )] dt1 dt2 (22.33)
T −T /2 −T /2
1 T /2 T /2 iν(t1 −t2 )
Z Z
= e Γ(t1 − t2 )dt1 dt2 (22.34)
T −T /2 −T /2
Z T  
|τ |
= 1− Γ(τ )eiντ dτ (22.35)
−T T
Recall that Γ(τ ) defines a non-negative quadratic form, meaning that
X
a†s at Γ(t − s) ≥ 0
s,t

for any sets of times and any complex numbers at . This will in particular work if
the complex numbers lie on the unit circle and can be written eiνt . This means
that integrals Z Z
eiν(t1 −t2 ) Γ(t1 − t2 )dt1 dt2 ≥ 0 (22.36)

so fT (ν) ≥ 0.
Define φT (τ ) as the integrand in Eq. 22.35, so that
Z ∞
1
fT (ν) = φT (τ )eiντ dτ (22.37)
2π −∞
CHAPTER 22. SPECTRAL ANALYSIS AND L2 ERGODICITY 178

which is recognizable as a proper Fourier transform. Now pick some N > 0 and
massage the equation so it starts to look like an inverse transform.
Z ∞
−iνt 1
fT (ν)e = φT (τ )eiντ e−iνt dτ (22.38)
2π −∞
  Z ∞  
|ν| 1 |ν|
1− fT (ν)e−iνt = φT (τ )eiντ e−iνt 1 − dτ (22.39)
N 2π −∞ N

Integrating over frequencies,

Z N  
|ν|
1− fT (ν)e−iνt dν (22.40)
−N N
Z N Z ∞  
1 iντ −iνt |ν|
= φT (τ )e e 1− dτ dν
−N 2π −∞ N
Z ∞  2
1 sin N (τ − t)/2
= φT (τ ) N dτ (22.41)
2π −∞ N (τ − t)/2

For fixed N , it is easy to verify that

Z ∞  2
sin N (τ − t)/2
N dt = 1
−∞ N (τ − t)/2

and that  2
sin N (τ − t)/2
lim N =N
t→τ N (τ − t)/2
On the other hand, if τ 6= t,
 2
sin N (τ − t)/2
lim N =0
N →∞ N (τ − t)/2

uniformly over any bounded interval in t. (You might find it instructive to try
plotting this function; you will need to be careful near the origin!) In other
words, this is a representation of the Dirac delta function, so that
Z ∞  2
sin N (τ − t)/2
lim φT (τ ) N dτ = φT (τ )
N →∞ −∞ N (τ − t)/2

and in fact the convergence is uniform.  

|ν|
Turning to the other side of Equation 22.41, 1 − N fT (ν) ≥ 0, so

N  
|ν|
Z
1− fT (ν)e−iνt dν
−N N

is like a characteristic function of a distribution, up to, perhaps, an over-all

normalizing factor, which will be (given the right-hand side) φT (0) = Γ(0) > 0.
Since Γ(τ ) is continuous, φT (τ ) is too, and so, as N → ∞, the right-hand
CHAPTER 22. SPECTRAL ANALYSIS AND L2 ERGODICITY 179

side converges uniformly on φT (t), but a uniform limit of characteristic func-

tions is still a characteristic function. Thus φT (t), too, can be obtained from
a characteristic function. Finally, since Γ(t) is the uniform limit of φT (t) on
every bounded interval, Γ(t) has a characteristic-function representation of the
stated form. This allows us to further conclude that S(ν) is real-valued, non-
decreasing, S(−∞) = 0 and S(∞) = Γ(0), and has both right and left limits
everywhere. 
There is a converse, with a cute constructive proof.

Theorem 287 (Existence of Weakly Stationary Processes with Given

Spectral Functions) Let S(ν) be any function with the properties described
at the end of Theorem 286. Then there is a weakly stationary process whose
autocovariance is of the form given in Eq. 22.29.

Proof: Define σ 2 = Γ(0), F (ν) = S(ν)/σ 2 . Now F (ν) is a properly normal-

ized cumulative distribution function. Let N be a random variable distributed
i(Φ−N t)
according to F , and Φ ∼U (0,  independent of N . Set X(t) ≡ σe
2π) be
 −iN .
iΦ t
Then E [X(t)] = σE e E e = 0. Moreover,
h i
E X † (t1 )X(t2 ) = σ 2 E e−i(Φ−N t1 ) ei(Φ−N t2 )
 
(22.42)
h i
= σ 2 E e−iN (t1 −t2 ) (22.43)
Z ∞
= σ2 e−iν(t1 −t2 ) dFν (22.44)
−∞
= Γ(t1 − t2 ) (22.45)

Definition 288 (Jump of the Spectral Function) The jump of the spectral
function at ν, ∆S(ν), is S(ν + 0) − S(ν − 0).

Lemma 289 (Spectral Function Has Non-Negative Jumps) ∆S(ν) ≥ 0.

Proof: Obvious from the fact that S(ν) is non-decreasing. 

Theorem 290 (Wiener-Khinchin Theorem) If X is a weakly stationary

process, then its power spectrum is equal to its spectral function.
 2 
V (ν) ≡ E X̃(ν) = S(ν) (22.46)
 

Proof: Assume, without loss of generality, that E [X(t)] = 0. Substitute

the spectral representation of X into the autocovariance, using Fubini’s theorem
CHAPTER 22. SPECTRAL ANALYSIS AND L2 ERGODICITY 180

to turn a product of integrals into a double integral.

Γ(τ ) = E [X(t)X(t + τ )] (22.47)

= E X † (t)X(t + τ )
 
(22.48)
Z ∞ Z ∞ 
= E e−i(t+τ )ν1 eitν2 dX̃ν1 dX̃ν2 (22.49)
−∞ −∞
Z ∞ Z ∞ 
= E e−it(ν1 −ν2 ) e−iτ ν2 dX̃ν1 dX̃ν2 (22.50)
−∞ −∞
Z ∞Z ∞ h i
= e−it(ν1 −ν2 ) e−iτ ν2 E dX̃ν1 dX̃ν2 (22.51)
−∞ −∞

using the fact that integration and expectation commute to (formally)h bring the i
expectation inside the integral. Since X̃ has orthogonal increments, E dX̃ν†1 dX̃ν2 =
0 unless ν1 = ν2 . This turns the double integral into a single integral, and kills
the e−it(ν1 −ν2 ) factor, which had to go away because t was arbitrary.
Z ∞ h i
Γ(τ ) = e−iτ ν E d(X̃ν† X̃ν ) (22.52)
−∞
Z ∞
= e−iτ ν dVν (22.53)
−∞
R∞
using the definition of the power spectrum. Since Γ(τ ) = −∞ e−iτ ν dVν , it
follows that Sν and Vν differ by a constant, namely the value of V (−∞), which
can be chosen to be zero without affecting the spectral representation of X. 

22.2.1 How the White Noise Lost Its Color

Why is white noise, as defined in Section 22.1, called “white”? The answer is
easy, given the Wiener-Khinchin relation in Theorem 290.
Recall from Proposition 274 that the autocovariance function of white noise
is δ(t1 − t2 ). Recall from general analysis that one representation of the delta
function is the following Fourier integral:
Z ∞
1
δ(t) = dνeiνt
2π −∞
(This can be “derived” from inserting the definition of the Fourier transform
into the inverse Fourier transform, among other, more respectable routes.) Ap-
1
pealing then to the theorem, S(ν) = 2π for all ν. That is, there is equal power at
all frequencies, just as white light is composed of light of all colors (frequencies),
mixed with equal intensity.
Relying on this analogy, there is an elaborate taxonomy red, pink, black,
brown, and other variously-colored noises, depending on the shape of their power
spectra. The value of this terminology has honestly never been very clear to
me, but the curious reader is referred to the (very fun) book of Schroeder (1991)
and references therein.
CHAPTER 22. SPECTRAL ANALYSIS AND L2 ERGODICITY 181

22.3 The Mean-Square Ergodic Theorem

Ergodic theorems relate functionals calculated along individual sample paths
RT
(say, the time average, T −1 0 dtX(t), or the maximum attained value) to func-
tionals calculated over the whole distribution (say, the expectation, E [X(t)], or
the expected maximum). The basic idea is that the two should be close, and they
should get closer the longer the trajectory we use, because in some sense any
one sample path, carried far enough, is representative of the whole distribution.
Since there are many different kinds of functionals, and many different modes of
stochastic convergence, there are many different kinds of ergodic theorem. The
classical ergodic theorems say that time averages converge on expectations3 , ei-
ther in Lp or a.s. (both implying convergence in distribution or in probability).
The traditional centerpiece of ergodic theorem is Birkhoff’s “individual” ergodic
theorem, asserting a.s. convergence. We will see its proof, but it will need a lot
of preparatory work, and it requires strict stationarity. By contrast, the L2 , or
“mean square”, ergodic theorem, attributed to von Neumann4 is already in our
grasp, and holds for weakly stationary processes.
We will actually prove it twice, once with a fairly transparent sufficient condi-
tion, and then again with a more complicated necessary-and-sufficient condition.
The more complicated proof will wait until next lecture.

22.3.1 Mean-Square Ergodicity Based on the Autocovari-

ance
First, the easy version, which gives an estimate of the rate of convergence.
(What I say here is ripped off from the illuminating discussion in (Frisch, 1995,
sec. 4.4, especially pp. 49–50).)

Definition 291 (Time Averages) When X is a one-sided, continuous-parameter

random process, we say that its time average between times T1 and T2 is X(T1 , T2 ) ≡
−1 R T2
(T2 − T1 ) T1
dtX(t). When we only mention one time argument, by default
the time average is from 0 to T , X(T ) ≡ X(0, T ).

(Only considering time averages starting from zero involves no loss of gen-
erality for weakly stationary processes: why?)

Definition 292 (Integral Time Scale) The integral time scale of a weakly-
stationary random process is
R∞
dτ |Γ(τ )|
τint ≡ 0 (22.54)
Γ(0)
3 Proverbially: “time averages converge on space averages”, the space in question being

the state space Ξ; or “converge on phase averages”, since physicists call certain kinds of state
space “phase space”.
4 See von Plato (1994, ch. 3) for a fascinating history of the development of ergodic theory

through the 1930s, and its place in the history of mathematical probability.
CHAPTER 22. SPECTRAL ANALYSIS AND L2 ERGODICITY 182

Notice that τint does, indeed, have units of time.

As a particular example, suppose that Γ(τ ) = Γ(0)e−τ /A , where the constant
A is known as the autocorrelation time. Then simple calculus shows that τint =
A.
Theorem 293 (Mean-Square Ergodic Theorem (Finite Autocovariance
Time)) Let X(t) be a weakly stationary process with E [X(t)] = 0. If τint < ∞,
L
then X(T ) →2 0 as T → ∞.
Proof: Use Fubini’s theorem to to the square of the integral into a double
integral, and then bring the expectation inside it:
 !2  " #
Z T Z TZ T
1 1
E  dtX(t)  = E dt1 dt2 X(t1 )X(t2 ) (22.55)
T 0 T2 0 0
Z TZ T
1
= dt1 dt2 E [X(t1 )X(t2 )] (22.56)
T2 0 0
Z TZ T
1
= dt1 dt2 Γ(t1 − t2 ) (22.57)
T2 0 0
Z T Z t1
2
= dt 1 dτ Γ(τ ) (22.58)
T2 0 0
Z T Z ∞
2
≤ dt 1 dτ |Γ(τ )| (22.59)
T2 0 0
2 ∞
Z
= dτ |Γ(τ )| (22.60)
T 0
Since the integral in the final inequality is Γ(0)τint , which is finite, everything
must go to zero as T → ∞. 
Remark. From the proof, we can see that the rate of convergence of the
2
mean-square of X(T ) 2 is (at least) O(1/T ). This would give a root-mean-
√
square (rms) convergence rate of O(1/ T ), which is what the naive statistician
who ignored inter-temporal dependence would expect from the central limit
theorem. (This ergodic theorem says nothing about the form of the distribution
of X(T ) for large T . We will see that, under some circumstances, it is Gaussian,
but that needs stronger assumptions [forms of “mixing”] than we have imposed.)
The naive statistician wouldexpect that the mean-square time average would go
like Γ(0)/T , since Γ(0) = E X 2 (t) = Var [X(t)]. The proportionality constant
R∞
is instead 0 dτ |Γ(τ )|. This is equal to the naive guess for white noise, and for
other collections of IID variables, but not in the general case. This leads to the
following
Corollary 294 (Convergence Rate in the Mean-Square Ergodic The-
orem) Under the conditions of Theorem 293,
  τint
Var X(T ) ≤ 2Var [X(0)] (22.61)
T
CHAPTER 22. SPECTRAL ANALYSIS AND L2 ERGODICITY 183

  2  
Proof: Since X(t) is centered, E X(T ) = 0, and X(T ) 2 = Var X(T ) .
Everything else follows from re-arranging the bound in the proof of Theorem
293, Definition 292, and the fact that Γ(0) = Var [X(0)]. 
As a consequence of the corollary, if T  τint , then the variance of the time
average is negigible compared to the variance at any one time.

22.3.2 Mean-Square Ergodicity Based on the Spectrum

Let’s warm up with some lemmas of a technical nature. The first relates the
jumps of the spectral process X̃(ν) to the jumps of the spectral function S(ν).
Lemma 295 (Mean-Square
 Jumps in the Spectral Process) For a weakly
2 
stationary process, E ∆X̃(ν) = ∆S(ν).
 

Proof: This follows directly from the Wiener-Khinchin relation (Theorem

290). 
Lemma 296 (The Jump in the Spectral Function) The jump of the spec-
tral function at ν is given by
1 T
Z
∆S(ν) = lim Γ(τ )eiντ dτ (22.62)
T →∞ T 0

Proof: This is a basic inversion result for characteristic functions. It should

become plausible by thinking of this as getting the Fourier transform of Γ as T
grows. 
Lemma 297 (Existence of L2 Limits for Time Averages) If X is weakly
stationary, then for any real f , eif t X(T ) converges in L2 to ∆X̃(f ).
Proof: Start by looking at the squared modulus of the time average for
finite time.
 Z 2
1 T 
eif t X(t)dt (22.63)
 
T 0


Z TZ T
1
= e−if (t1 −t2 ) X † (t1 )X(t2 )dt1 dt2
T2 0 0
1 T T −if (t1 −t2 ) ∞ iν1 t1
Z Z Z Z ∞
= e e dX̃ν1 e−iν2 t2 dX̃ν2 (22.64)
T2 0 0 −∞ −∞
Z TZ ∞ Z TZ ∞
1 it1 (f −ν1 )
= dt 1 d X̃ ν1 e dt2 dX̃ν2 e−it2 (f −ν2 ) (22.65)
T 2 0 −∞ 0 −∞

L
As T → ∞, these integrals pick out ∆X̃(f ) and ∆X̃ † (f ). So, eif t X(T ) →2
∆X̃(f ). 
Notice that the limit provided by the lemma is a random quantity. What’s
really desired, in most applications, is convergence to a deterministic limit,
which here would mean convergence (in L2 ) to zero.
CHAPTER 22. SPECTRAL ANALYSIS AND L2 ERGODICITY 184

Lemma 298 (The Mean-Square Ergodic Theorem) If X is weakly sta-

tionary, and E [X(t)] = 0, then X(t) converges in L2 to 0 iff
Z T
lim T −1 dτ Γ(τ ) = 0 (22.66)
0

L
Proof: Taking f = 0 in Lemma 297, X(T ) →2 ∆X̃(0), the jump in the
spectral function at zero. Let’s show that the (i) expectation of this jump is
zero, and that (ii) its variance is given by the integral expression on the LHS of
L  
Eq. 22.66. For (i), because X(T ) h→2 Y , we i know that E X(T ) → E [Y ]. But
 
E X(T ) = E [X](T ) = 0. So E ∆X̃(0) = 0. For (ii), Lemma 295, plus the
h i
fact that E ∆X̃(0) = 0, shows that the variance is equal to the jump in the
spectrum at 0. But, by Lemma 296 with ν = 0, that jump is exactly the LHS
of Eq. 22.66. 
Remark 1: Notice that if the integral time is finite, then the integral condi-
tion on the autocovariance is automatically satisfied, but not vice versa, so the
hypotheses here are strictly weaker than in Theorem 293.
Remark 2: One interpretation of the theorem is that the time-average is con-
verging on the zero-frequency component of the spectral process. Intuitively,
all the other components are being killed off by the time-averaging, because the
because time-integral of a sinusoidal function is bounded, but the denominator
in a time average is unbounded. The only part left is the zero-frequency com-
ponent, whose time integral can also grow linearly with time. If there is a jump
at 0, then this has finite variance; if not, not.
Remark 3: Lemma 297 establishes the L2 convergence of time-averages of
the form
1 T if t
Z
e X(t)dt
T 0
for any real f . Specifically, from Lemma 295, the mean-square of this variable
is converging on the jump in the spectrum at f . Multiplying X(t) by eif t makes
the old frequency f component the new frequency zero component, so it is the
surviving term. While the ergodic theorem itself only needs the f = 0 case, this
result is useful in connection with estimating spectra from time series (Doob,
1953, ch. X, §7).

22.4 Exercises
Exercise 22.1 (Mean-Square Ergodicity in Discrete Time) It is often
convenient to have a mean-square ergodic theorem for discrete-time sequences
rather than continuous-time processes. If the dt in the definition of X is re-
interpreted as counting measure on N, rather than Lebesgue measure on R+ ,
does the proof of Theorem 293 remain valid? (If yes, say why; if no, explain
where the argument fails.)
CHAPTER 22. SPECTRAL ANALYSIS AND L2 ERGODICITY 185

Exercise 22.2 (Mean-Square Ergodicity with Non-Zero Mean) State

and prove a version of Theorem 293 which does not assume that E [X(t)] = 0.

Exercise 22.3 (Functions of Weakly Stationary Processes) Suppose X is

a weakly stationary process, and f is a measurable function such that kf (X0 )k2 <
∞. Is f (X) a weakly stationary process? (If yes, prove it; if not, give a counter-
example.)

Exercise 22.4 (Ergodicity of the Ornstein-Uhlenbeck Process?) Sup-

pose the Ornstein-Uhlenbeck process is has its invariant distribution as its initial
distribution, and is therefore weakly stationary. Does Theorem 293 apply?
 Part V

Ergodic Theory

186
Chapter 23

Ergodic Properties and

Ergodic Limits

Section 23.1 gives a general orientation to ergodic theory, which

we will study in discrete time.
Section 23.2 introduces dynamical systems and their invariants,
the setting in which we will prove our ergodic theorems.
Section 23.3 considers time averages, defines what we mean for
a function to have an ergodic property (its time average converges),
and derives some consequences.
Section 23.4 defines asymptotic mean stationarity, and shows
that, with AMS dynamics, the limiting time average is equivalent to
conditioning on the invariant sets.

23.1 General Remarks

To begin our study of ergodic theory, let us consider a famous1 line from Gne-
denko and Kolmogorov (1954, p. 1):
In fact, all epistemological value of the theory of probability is
based on this: that large-scale random phenomena in their collective
action create strict, nonrandom regularity.

Now, this is how Gnedenko and Kolmogorov introduced their classic study of the
limit laws for independent random variables, but most of the random phenomena
we encounter around us are not independent. Ergodic theory is a study of
how large-scale dependent random phenomena nonetheless create nonrandom
regularity. The classical limit laws for IID variables X1 , X2 , . . . assert that,
1 Among mathematical scientists, anyway.

187
CHAPTER 23. ERGODIC PROPERTIES AND ERGODIC LIMITS 188

under the right conditions, sample averages converge on expectations,

n
1X
Xi → E [Xi ]
n i=1

where the sense of convergence can be “almost sure” (strong law of large num-
bers), “Lp ” (pth mean), “in probability” (weak law), etc., depending on the
hypotheses we put on the Xi . One meaning of this convergence is that suffi-
ciently large random samples are representative of the entire population — that
the sample mean makes a good estimate of E [X].
The ergodic theorems, likewise, assert that for dependent sequences X1 , X2 , . . .,
time averages converge on expectations
t
1X
Xi → E [X∞ |I]
t i=1

where X∞ is some limiting random variable, or in the most useful cases a non-
random variable, and I is a σ-field representing some sort of asymptotic in-
formation. Once again, the mode of convergence will depend on the kind of
hypotheses we make about the random sequence X. Once again, the interpre-
tation is that a single sample path is representative of an entire distribution
over sample paths, if it goes on long enough. The IID laws of large numbers
are, in fact, special cases of the corresponding ergodic theorems.
Section 22.3 proved a mean-square (L2 ) ergodic theorem for weakly sta-
tionary continuous-parameter processes.2 The next few chapters, by contrast,
will develop ergodic theorems for non-stationary discrete-parameter processes.3
This is a little unusual, compared to most probability books, so let me say a
word or two about why. (1) Results we get will include stationary processes as
special cases, but stationarity fails for many applications where ergodicity (in a
suitable sense) holds. So this is more general and more broadly applicable. (2)
Our results will all have continuous-time analogs, but the algebra is a lot cleaner
in discrete time. (3) Some of the most important applications (for people like
you!) are to statistical inference and learning with dependent samples, and to
Markov chain Monte Carlo, and both of those are naturally discrete-parameter
processes. We will, however, stick to continuous state spaces.

23.2 Dynamical Systems and Their Invariants

It is a very remarkable fact — but one with deep historical roots (von Plato,
1994, ch. 3) — that the way to get regular limits for stochastic processes is
to first turn them into irregular deterministic dynamical systems, and then let
averaging smooth away the irregularity. This section will begin by laying out
dynamical systems, and their invariant sets and functions, which will be the
foundation for what follows.
2 Can you identify X∞ and I for this case?
3 In doing so, I’m ripping off Gray (1988), especially chapters 6 and 7.
CHAPTER 23. ERGODIC PROPERTIES AND ERGODIC LIMITS 189

Definition 299 (Dynamical System) A dynamical system consists of a mea-

surable space Ξ, a σ-field X on Ξ, a probability measure µ defined on X , and a
mapping T : Ξ 7→ Ξ which is X /X -measurable.
Remark: Measure-preserving transformations (Definition 53) are special cases
of dynamical systems. Since (Theorem 52) every strongly stationary process can
be represented by a measure-preserving transformation, namely the shift (Def-
inition 48), the theory of ergodicity for dynamical systems which we’ll develop
is easily seen to include the usual ergodic theory of strictly-stationary processes
as a special case. Thus, at the cost of going to the infinite-dimensional space of
sample paths, we can always make it the case that the time-evolution is com-
pletely deterministic, and the only stochastic component to the process is its
initial condition.
Lemma 300 (Dynamical Systems are Markov Processes) Let Ξ, X , µ, T
be a dynamical system. Let L (X1 ) = µ, and define Xt = T t−1 X1 . Then the
Xt form a Markov process, with evolution operator K defined through Kf (x) =
f (T x).
Proof: For every x ∈ Ξ and B ∈ X , define κ(x, B) ≡ 1B (T x). For fixed
x, this is clearly a probability measure (specifically, the δ measure at T x).
For fixed B, this is a measurable function of x, because T is a measurable
mapping. Hence, κ(x, B) is a probability kernel. So, by Theorem 106, the Xt
form a Markov process. By definition, E [f (X1 )|X0 = x] = Kf (x). But the
expectation is in this case just f (T x). 
Notice that, as a consequence, there is a corresponding operator, call it U ,
which takes signed measures (defined over X ) to signed measures, and specifi-
cally takes probability measures to probability measures.
Definition 301 (Observable) A function f : Ξ 7→ R which is B/X measurable
is an observable of the dynamical system Ξ, X , µ, T .
Pretty much all of what follows would work if the observables took values in
any real or complex vector space, but that situation can be built up from this
one.
Definition 302 (Invariant Functions, Sets and Measures) A function is
invariant, under the action of a dynamical system, if f (T x) = f (x) for all
x ∈ Ξ, or equivalently if Kf = f everywhere. An event B ∈ X is invariant if
its indicator function is an invariant function. A measure ν is invariant if it is
preserved by T , i.e. if ν(C) = ν(T −1 C) for all C ∈ X , equivalently if U ν = ν.

Lemma 303 (Invariant Sets are a σ-Algebra) The class I of all measurable
invariant sets in Ξ forms a σ-algebra.
Proof: Clearly, Ξ is invariant. The other properties of a σ-algebra follow
because set-theoretic operations (union, complementation, etc.) commute with
taking inverse images. 
CHAPTER 23. ERGODIC PROPERTIES AND ERGODIC LIMITS 190

Lemma 304 (Invariant Sets and Observables) An observable is invariant

if and only if it is I-measurable. Consequently, I is the σ-field generated by the
invariant observables.
−1
Proof: “If”: Pick any Borel set B. Since f = f ◦T , f −1 (B) = (f ◦ T ) (B) =
T −1 f −1 B. Hence f −1 (B) ∈ I. Since the inverse image of every Borel set is
in I, f is I-measurable. “Only if”: Again, pick any Borel set B. By assump-
−1
tion, f −1 (B) ∈ I, so f −1 (B) = T −1 f −1 (B) = (f ◦ T ) (B), so the inverse
image of under T f of any Borel set is an invariant set, implying that f ◦ T
−1
is I-measurable. Since, for every B, f −1 (B) = (f ◦ T ) (B), we must have
f ◦ T = f . The consequence follows. 

Definition 305 (Infinitely Often, i.o.) For any set C ∈ X , the set C in-
finitely often, Ci.o. , consists of all those points in Ξ whose trajectories visit C
infinitely often, Ci.o. ≡ lim supt T −t C.

Lemma 306 (“Infinitely often” implies invariance) For every C ∈ X ,

Ci.o. is invariant.

Proof: Exercise 23.1. 

Definition 307 (Invariance Almost Everywhere) A measurable function

is invariant µ-a.e., or almost invariant, when

µ {x ∈ Ξ|∀n, f (x) = K n f (x)} = 1 (23.1)

A measurable set is invariant µ-a.e., when its indicator function is almost in-
variant.

Remark 1: Some of the older literature annoyingly calls these objects totally
invariant.
Remark 2: Invariance implies invariance µ-almost everywhere, for any µ.

Lemma 308 (Almost-invariant sets form a σ-algebra) The almost-invariant

sets form a σ-field, I 0 , and an observable is almost invariant if and only if it is
measurable with respect to this field.

Proof: Entirely parallel to that for the strict invariants. 

Let’s close this section with a simple lemma, which will however be useful
in approximation-by-simple-function arguments in building up expectations.

Lemma
Pm 309 (Invariance for simple functions) A simple function, f (x) =
k=1 am 1Ck (x), is invariant if and only if all the sets Ck ∈ I. Similarly, a
simple function is almost invariant iff all the defining sets are almost invariant.

Proof: Exercise 23.2. 

CHAPTER 23. ERGODIC PROPERTIES AND ERGODIC LIMITS 191

23.3 Time Averages and Ergodic Properties

For convenience, let’s re-iterate the definition of a time average. The notation
differs here a little from that given earlier, in a way which helps with discrete
time.
Definition 310 (Time Averages) The time-average of an observable f is the
real-valued function
t−1
1X
At f (x) ≡ f (T i x) (23.2)
t i=0
where At is the operator taking functions to their time-averages.

Lemma 311 (Time averages are observables) For every t, the time-average
of an observable is an observable.
Proof: The class of measurable functions is closed under finite iterations
of arithmetic operations. 
Definition 312 (Ergodic Property) An observable f has the ergodic prop-
erty when At f (x) converges as t → ∞ for µ-almost-all x. An observable has
the mean ergodic property when At f (x) converges in L1 (µ), and similarly for
the other Lp ergodic properties. If for some class of functions D, every f ∈ D
has an ergodic property, then the class D has that ergodic property.
Remark. Notice that what is required for f to have the ergodic property is
that almost every initial point has some limit for its time average,
n  o
µ x ∈ Ξ ∃r ∈ R : lim At f (x) = r = 1 (23.3)

t→∞

This does not mean that there is some common limit for almost every initial
point, n  o
∃r ∈ R : µ x ∈ Ξ  lim At f (x) = r = 1 (23.4)

t→∞
Similarly, a class of functions has the ergodic property if all of their time averages
converge; they do not have to converge uniformly.
Definition 313 (Ergodic Limit) If an observable f has the ergodic property,
define Af (x) to be the limit of At f (x) where that exists, and 0 elsewhere. The
domain of A consists of all and only the functions with ergodic properties.
Observe that
t
1X n
Af (x) = lim K f (x) (23.5)
t→∞ t
n=0
That is, A is the limit of an arithmetic mean of conditional expectations. This
suggests that it should itself have many of the properties of conditional expec-
tations. In fact, under a reasonable condition, we will see that Af = E [f |I],
expectation conditional on the σ-algebra of invariant sets. We’ll check first that
A has the properties we’d want from a conditional expectation.
CHAPTER 23. ERGODIC PROPERTIES AND ERGODIC LIMITS 192

Lemma 314 (Linearity of Ergodic Limits) A is a linear operator, and its

domain is a linear space.

Proof: If c is any real number, then At cf (x) = cAt f (x), and so clearly, if
the limit exists, Acf (x) = cAf (x). Similarly, At (f + g)(x) = At f (x) + At g(x),
so if f and g both have ergodic properties, then so does f +g, and A(f +g)(x) =
Af (x) + Ag(x). 

Lemma 315 (Non-negativity of ergodic limits) If f ∈ DomA, and, for all

n ≥ 0, f T n ≥ 0 a.e., then Af (x) ≥ 0 a.e.

Proof: The event Af (x) < 0 is a sub-event of n {f (T n (x)) < 0}. Since
S
the union of a countable collection of measure zero events has measure zero,
Af (x) ≥ 0 almost everywhere. 
Notice that our hypothesis is that f T n ≥ 0 a.e. for all n, not just that f ≥ 0.
The reason we need the stronger assumption is that the transformation T might
map every point to the bad set of f ; the lemma guards against that. Of course,
if f (x) ≥ 0 for all, and not just almost all, x, then the bad set is non-existent,
and Af ≥ 0 follows automatically.

Lemma 316 (Constant functions have the ergodic property) The con-
stant function 1 has the ergodic property. Consequently, so does every other
constant function.

Proof: For every n, 1(T n x) = 1. Hence At 1(x) = 1 for all t, and so

A1(x) = 1. Extension to other constants follows by linearity (Lemma 314). 
Remember that for any time-evolution operator K, K1 = 1.

Lemma 317 (Ergodic limits are invariantifying) If f ∈ Dom(A), then,

for all n, f ◦ T n is too, and Af (x) = Af ◦ T n (x). Or, AK n f (x) = Af (x).

Proof: Start with n = 1, and show that the discrepancy goes to zero.
t
1X
K i+1 f (x) − K i f (x)


AKf (x) − Af (x) = lim (23.6)
t t i=0
1
K t f (x) − f (x)


= lim (23.7)
t t
Pt−1
Since Af (x) exists a.e., we know that the series t−1 i=0 K i f (x) converges
−1 −1
a.e., implying that (t + 1) K t f (x) → 0 a.e.. But t−1 = t+1 t (t + 1) , and for
−1 −1
large t, t + 1/t < 2. Hence (t + 1) K t f (x) ≤ t−1 K t f (x) ≤ 2(t + 1) K t f (x),
implying that t−1 K t f (x) itself goes to zero (a.e.). Similarly, t−1 f (x) must go
to zero. Thus, overall, we have AKf (x) = Af (x) a.e., and Kf (x) ∈ Dom(A).


Lemma 318 (Ergodic limits are invariant functions) If f ∈ Dom(A),

then Af is an invariant, and I-measurable.
CHAPTER 23. ERGODIC PROPERTIES AND ERGODIC LIMITS 193

Proof: Af exists, so (previous lemma) AKf exists and is equal to Af (al-

most everywhere). But AKf (x) = Af (T x), by definition, hence Af is invariant,
i.e., KAf = AKf = Af . Measurability follows from Lemma 304. 

Lemma 319 (Ergodic limits and invariant indicator functions) If f ∈

Dom(A), and B is any set in I, then A(1B (x)f (x)) = 1B (x)Af (x).

Proof: For every n, 1B (T n x)f (T n x) = 1B (x)f (T n x), since x ∈ B iff

n
T x ∈ B. So, for all finite t, At (1B (x)f (x)) = 1B (x)At f (x), and the lemma
follows by taking the limit. 

Lemma 320 (Ergodic properties of sets and observables) All indicator

functions of measurable sets have ergodic properties if and only if all bounded
observables have ergodic properties.

Proof: A standard approximation-by-simple-functions argument, as in the

construction of Lebesgue integrals. 

Lemma 321 (Expectations of ergodic limits) Let f be bounded and have

the ergodic property. Then Af is µ-integrable, and E [Af (X)] = E [f (X)], where
L (X) = µ.

Proof: Since f is bounded, it is integrable. Hence At f is bounded, too,

for all t, and At f (X) is an integrable random variable. A sequence of bounded,
integrable random variables is uniformly integrable. Uniform integrability, plus
the convergence At f (x) → Af (x) for µ-almost-all x, gives us that E [Af (X)]
exists and is equal to lim E [At f (X)] via Fatou’s lemma. (See e.g., Theorem 117
in the notes to 36-752.)
Now use the invariance of Af , i.e., the fact that Af (X) = Af (T X) µ-a.s.

0 = E [Af (T X)] − E [Af (X)] (23.8)

t−1 t−1
1 X 1X
= lim E [K n f (T X)] − lim E [K n f (X)] (23.9)
t n=0 t n=0
t−1
1X
= lim E [K n f (T X)] − E [K n f (X)] (23.10)
t n=0
t−1
1 X  n+1
f (X) − E [K n f (X)]

= lim E K (23.11)
t n=0
1
E K t f (X) − E [f (X))
  
= lim (23.12)
t
Hence
t−1
1X
E [Af ] = lim E [K n f (X)] = E [f (X)] (23.13)
t n=0
as was to be shown. 
CHAPTER 23. ERGODIC PROPERTIES AND ERGODIC LIMITS 194

Lemma 322 (Convergence of Ergodic Limits) If f is as in Lemma 321,

then At f → f in L1 (µ).

Proof: From Lemma 321, lim E [At f (X)] = E [f (X)]. Since the variables
At f (X) are uniformly integrable (as we saw in the proof of that lemma), it
follows (Proposition 4.12 in Kallenberg, p. 68) that they also converge in L1 (µ).


Lemma 323 (Cesàro Mean of Expectations) Let f be as in Lemmas 321

and 322, and B ∈ X be an arbitrary measurable set. Then
t−1
1X
lim E [1B (X)K n f (X)] = E [1B (X)f (X)] (23.14)
t→∞ t
n=0

where L (X) = µ.

Proof: Let’s write out the expectations explicitly as integrals.

Z t−1 Z

 1X n

 f (x)dµ − K f (x)dµ (23.15)
 
 B t n=0 B 
Z 
t−1
 1X n 
=  f (x) − K f (x)dµ
 
 B t n=0 
Z 
 
=  f (x) − At f (x)dµ (23.16)
ZB
≤ |f (x) − At f (x)| dµ (23.17)
ZB
≤ |f (x) − At f (x)| dµ (23.18)

= kf − At f kL1 (µ) (23.19)

But (previous lemma) these functions converge in L1 (µ), so the limit of the
norm of their difference is zero. 
Boundedness is not essential.

Corollary 324 (Replacing Boundedness with Uniform Integrability)

Lemmas 321, 322 and 323 hold for any integrable observable f ∈ Dom(A),
bounded or not, provided that At f is a uniformly integrable sequence.

Proof: Examining the proofs shows that the boundedness of f was impor-
tant only to establish the uniform integrability of At f . 

23.4 Asymptotic Mean Stationarity

Next, we come to an important concept which will prove to be necessary and
sufficient for the most important ergodic properties to hold.
CHAPTER 23. ERGODIC PROPERTIES AND ERGODIC LIMITS 195

Definition 325 (Asymptotically Mean Stationary) A dynamical system

is asymptotically mean stationary when, for every C ∈ X , the limit
t−1
1X
m(C) ≡ lim µ(T −n C) (23.20)
t→∞ t
n=0

exists, and the set function m is its stationary mean.

Remark 1: It might’ve been more logical to call this “asymptotically measure
stationary”, or something, but I didn’t make up the names...
Remark 2: Symbolically, we can write
t−1
1X n
m = lim U µ
t→∞ t
n=0

where U is the operator taking measures to measures. This leads us to the next
proposition.
Lemma 326 (Stationary Implies Asymptotically Mean Stationary) If
a dynamical system is stationary, i.e., T is preserves the measure µ, then it is
asymptotically mean stationary, with stationary mean µ.
Proof: If T preserves µ, then for every measurable set, µ(C) = µ(T −1 C).
Hence every term in the sum in Eq. 23.20 is µ(C), and consequently the limit
exists and is equal to µ(C). 
Proposition 327 (Vitali-Hahn Theorem) If mt are a sequence of probabil-
ity measures on a common σ-algebra X , and m(C) is a set function such that
limt mt (C) = m(C) for all C ∈ X , then m is a probability measure on X .
Proof: This is a standard result from measure theory. 
Theorem 328 (Stationary Means are Invariant Measures) If a dynam-
ical system is asymptotically mean stationary, then its stationary mean is an
invariant probability measure.
Pt−1
Proof: For every t, let mt (C) = 1t n=0 µ(T −n (C)). Then mt is a linear
combination of probability measures, hence a probability measure itself. Since,
for every C ∈ X , lim mt (C) = m(C), by Definition 325, Proposition 327 says
that m(C) is also a probability measure. It remains to check invariance.
m(C) − m(T −1 C) (23.21)
t−1 t−1
1X 1 X
= lim µ(T −n (C)) − lim µ(T −n (T −1 C))
t n=0 t n=0
t−1
1X
= lim µ(T −n−1 C) − µ(T −n C) (23.22)
t n=0
1
µ(T −t C) − µ(C)


= lim (23.23)
t
CHAPTER 23. ERGODIC PROPERTIES AND ERGODIC LIMITS 196

Since the probability measure of any set is at most 1, the difference between
two probabilities is at most 1, and so m(C) = m(T −1 C), for all C ∈ X . But
this means that m is invariant under T (Definition 53). 
Remark: Returning to the symbolic manipulations, if µ is AMS with sta-
tionary mean m, then U m = m (because m is invariant), and so we can write
µ = m + (µ − m), knowing that µ − m goes to zero under averaging. Speaking
loosely (this can be made precise, at the cost of a fairly long excursion) m is
an eigenvector of U (with eigenvalue 1), and µ − m lies in an orthogonal di-
rection, along which U is contracting, so that, under averaging, it goes away,
leaving only m, which is like the projection of the original measure µ on to the
invariant manifold of U .
The relationship between an AMS measure µ and its stationary mean m
is particularly simple on invariant sets: they are equal there. A slightly more
general theorem is actually just as easy to prove, however, so we’ll do that.
Lemma 329 (Expectations of Almost-Invariant Functions) If µ is AMS
with limit m, and f is an observable which is invariant µ-a.e., then Eµ [f ] =
Em [f ].
Proof: Let C be any almost invariant set. Then, for any t, C and T −t C
differ by, at most, a set of µ-measure 0, so that µ(C) = µ(T −t C). The definition
of the stationary mean (Equation 23.20) then gives µ(C) = m(C), or Eµ [1C ] =
Em [1C ], i.e., the result holds for indicator functions. By Lemma 309, this then
extends to simple functions. The usual arguments then take us to all functions
which are measurable with respect to I 0 , the σ-field of almost-invariant sets,
but this (Lemma 308) is the class of all almost-invariant functions. 
Lemma 330 (Limit of Cesàro Means of Expectations) If µ is AMS with
stationary mean m, and f is a bounded observable,
lim Eµ [At f ] = Em [f ] (23.24)
t→∞

Proof: By Eq. 23.20, this must hold when f is an indicator function. By

the linearity of At and of expectations, it thus holds for simple functions, and
so for general measurable functions, using boundedness to exchange limits and
expectations where necessary. 
Lemma 331 (Expectation of the Ergodic Limit is the AMS Expecta-
tion) If f is a bounded observable in Dom(A), and µ is AMS with stationary
mean m, then Eµ [Af ] = Em [f ].
Proof: From Lemma 323, Eµ [Af ] = limt→∞ Eµ [At f ]. From Lemma 330,
the latter is Em [f ]. 
Remark: Since Af is invariant, we’ve got Eµ [Af ] = Em [Af ], from Lemma
329, but that’s not the same.
Corollary 332 (Replacing Boundedness with Uniform Integrability)
Lemmas 330 and 331 continue to hold if f is not bounded, but At f is uniformly
integrable (µ).
CHAPTER 23. ERGODIC PROPERTIES AND ERGODIC LIMITS 197

Proof: As in Corollary 324. 

Theorem 333 (Ergodic Limits of Bounded Observables are Condi-
tional Expectations) If µ is AMS, with stationary mean m, and the dynamics
have ergodic properties for all the indicator functions, then, for any bounded ob-
servable f ,
Af = Em [f |I] (23.25)
with probability 1 under both µ and m.
Proof: Begin by showing this holds for the indicator functions of measur-
able sets, i.e., when f = 1C for arbitrary measurable C. By assumption, this
function has the ergodic property.
By Lemma 318, A1C is an invariant function. Pick any set B ∈ I, so that
1B is also invariant. By Lemma 319, A(1B 1C ) = 1B A1C , which is invariant
(as a product of invariant functions). So Lemma 329 gives
Eµ [1B A1C ] = Em [1B A1C ] (23.26)
while Lemma 331 says
Eµ [A(1B 1C )] = Em [1B 1C ] (23.27)
Since the left-hand sides are equal, the right-hand sides must be equal as well,
so
m(B ∩ C) = Em [1B 1C ] (23.28)
= Em [1B A1C ] (23.29)
Since this holds for all invariant sets B ∈ I, we conclude that A1C must be a
version of the conditional probability m(C|I).
From Lemma 320, every bounded observable has the ergodic property. The
proof above then extends naturally to arbitrary bounded f . 
Corollary 334 (Ergodic Limits of Integrable Observables are Condi-
tional Expectations) Equation 23.25 continues to hold if At f are uniformly
µ-integrable, or f is m-integrable.
Proof: Exercise 23.3. 

23.5 Exercises
Exercise 23.1 (“Infinitely often” implies invariant) Prove Lemma 306.

Exercise 23.2 (Invariant simple functions) Prove Lemma 309.

Exercise 23.3 (Ergodic limits of integrable observables) Prove Corollary

334.
Chapter 24

The Almost-Sure Ergodic

Theorem

This chapter proves Birkhoff’s ergodic theorem, on the almost-

sure convergence of time averages to expectations, under the as-
sumption that the dynamics are asymptotically mean stationary.
This is not the usual proof of the ergodic theorem, as you will find in e.g.
Kallenberg. Rather, it uses the AMS machinery developed in the last lecture,
following Gray (1988, sec. 7.2), in turn following Katznelson and Weiss (1982).
The central idea is that of “blocking”: break the infinite sequence up into non-
overlapping blocks, show that each block is well-behaved, and conclude that
the whole sequence is too. This is a very common technique in modern ergodic
theory, especially among information theorists. In pure probability theory, the
usual proof of the ergodic theorem uses a result called the “maximal ergodic
lemma”, which is clever but somewhat obscure, and doesn’t seem to generalize
well to non-stationary processes: see Kallenberg, ch. 10.
We saw, at the end of the last chapter, that if time-averages converge in the
long run, they converge on conditional expectations. Our work here is showing
that they (almost always) converge. We’ll do this by showing that their lim infs
and lim sups are (almost always) equal. This calls for some preliminary results
about the upper and lower limits of time-averages.
Definition 335 (Lower and Upper Limiting Time Averages) For any
observable f , define the lower and upper limits of its time averages as, respec-
tively,
Af (x) ≡ lim inf At f (x) (24.1)
t→∞
Af (x) ≡ lim sup At f (x) (24.2)
t→∞

Define Lf as the set of x where the limits coincide:

 
Lf ≡ x Af (x) = Af (x) (24.3)

198
CHAPTER 24. THE ALMOST-SURE ERGODIC THEOREM 199

Lemma 336 (Lower and Upper Limiting Time Averages are Invari-
ant) For each obsevable f , Af and Af are invariant functions.

Proof: Use our favorite trick, and write At f (T x) = t+1 t At+1 f (x) − f (x)/t.
Clearly, the lim sup and lim inf of this expression will equal the lim sup and
lim inf of At+1 f (x), which is the same as that of At f (x). 

Lemma 337 (“Limits Coincide” is an Invariant Event) For each f , the

set Lf is invariant.

Proof: Since Af and Af are both invariant, they are both measurable with
respect to I (Lemma 304), so the set of x such that Af (x) = Af (x) is in I,
therefore it is invariant (Definition 303). 

Lemma 338 (Ergodic properties under AMS measures) An observable

f has the ergodic property with respect to an AMS measure µ if and only if it
has it with respect to the stationary limit m.

Proof: By Lemma 337, Lf is an invariant set. But then, by Lemma 329,

m(Lf ) = µ(Lf ). (Take f = 1Lf in the lemma.) f has the ergodic property with
respect to µ iff µ(Lf ) = 1, so f has the ergodic property with respect to µ iff it
has it with respect to m. 

Theorem 339 (Birkhoff ’s Almost-Sure Ergodic Theorem) If a dynam-

ical system is AMS with stationary mean m, then all bounded observables have
the ergodic property, and with probability 1 (under both µ and m),

Af = Em [f |I] (24.4)

for all f ∈ L1 (m).

Proof: From Theorem 333 and its corollaries, it is enough to prove that all
L1 (m) observables have ergodic properties to get Eq. 24.4. From Lemma 338, it
is enough to show that the observables have ergodic properties in the stationary
system Ξ, X , m, T . (Accordingly, all expectations in the rest of this proof will
be with respect to m.) Since any observable can be decomposed into its positive
and negative parts, f = f + − f − , assume, without loss of generality,
 that
 f is
positive. Since Af ≥ Af everywhere, it suffices to show that E Af − Af ≤ 0.
 
This in turn will follow from E Af ≤ E [f ] ≤ E [Af ]. (Since f is bounded, the
integrals exist.)  
We’ll prove that E Af ≤ E [f ], by showing that the time average comes
close to its lim sup, but from above (in the mean). Proving that E [Af ] ≥ E [f ]
will be entirely parallel.
Since f is bounded, we may assume that f ≤ M everywhere.
For every  > 0, for every x there exists a finite t such that

At f (x) ≥ f (x) −  (24.5)

CHAPTER 24. THE ALMOST-SURE ERGODIC THEOREM 200

This is because f is the limit of the least upper bounds. (You can see where
this is going already — the time-average has to be close to its lim sup, but close
from above.)
Define t(x, ) to be the smallest t such that f (x) ≤  + At f (x). Then, since
f is invariant, we can add from from time 0 to time t(x, ) − 1 and get:
t(x,)−1 t(x,)−1
X X
K n f (x) + t(x, ) ≥ K n f (x) (24.6)
n=0 n=0

Define BN ≡ {x|t(x, ) ≥ N }, the set of “bad” x, where the sample average fails
to reach a reasonable () distance of the lim sup before time N . Because t(x, )
is finite, m(BN ) goes to zero as N → ∞. Chose a N such that m(BN ) ≤ /M ,
and, for the corresponding bad set, drop the subscript. (We’ll see why this level
is important presently.)
We’ll find it convenient to not deal directly with f , but with a related func-
tion which is better-behaved on the bad set B. Set f˜(x) = M when x ∈ B,
and = f (x) elsewhere. Similarly, define t̃(x, ) to be 1 if x ∈ B, and t(x, )
elsewhere. Notice that t̃(x, ) ≤ N for all x. Something like Eq. 24.6 still holds
for the nice-ified function f˜, specifically,
t̃(x,)−1 t̃(x,)−1
X X
K n f (x) ≤ K n f˜(x) + t̃(x, ) (24.7)
n=0 n=0

If x ∈ B, this reduces to f (x) ≤ M + , which is certainly true because f (x) ≤

M . If x 6∈ B, it will follow from Eq. 24.6, provided that T n x 6∈ B, for all
n ≤ t̃(x, ) − 1. To see that this, in turn, must be true, suppose that T n x ∈ B
for some such n. Because (we’re assuming) n < t(x, ), it must be the case that

An f (x) < f (x) −  (24.8)

Otherwise, t(x, ) would not be the first time at which Eq. 24.5 held true. Sim-
ilarly, because T n x ∈ B, while x 6∈ B, t(T n x, ) > N ≥ t(x, ), and so

At(x,)−n f (T n x) < f (x) −  (24.9)

Combining the last two displayed equations,

At(x,) f (x) < f (x) −  (24.10)

contradicting the definition of t(x, ). Consequently, there can be no n < t(x, )
such that T n x ∈ B.
We are now ready to consider the time average AL f over a stretch of time
of some considerable length L. We’ll break the time indices over which we’re
averaging into blocks, each block ending when T t x hits B again. We need to
make sure that L is sufficiently large, and it will turn out that L ≥ N/(/M )
suffices, so that N M/L ≤ . The end-points of the blocks are defined recursively,
CHAPTER 24. THE ALMOST-SURE ERGODIC THEOREM 201

starting with b0 = 0, bk+1 = bk + t̃(T bk x, ). (Of course the bk are implicitly
dependent on x and  and N , but suppress that for now, since these are constant
through the argument.) The number of completed blocks, C, is the largest k
such that L − 1 ≥ bk . Notice that L − bC ≤ N , because t̃(x, ) ≤ N , so if
L − bC > N , we could squeeze in another block after bC , contradicting its
definition.
Now let’s examine the sum of the lim sup over the trajectory of length L.
L−1
X C
X bk
X L−1
X
K n f (x) = K n f (x) + K n f (x) (24.11)
n=0 k=1 n=bk−1 n=bC

For each term in the inner sum, we may assert that

t̃(T bk x,)−1 t̃(T bk x,)−1
X X
n bk
K f (T x) ≤ K n f˜(T bk x) + t̃(T bk x, ) (24.12)
n=0 n=0

on the strength of Equation 24.7, so, returning to the over-all sum,

L−1
X C
X k −1
bX L−1
X
K n f (x) ≤ K n f˜(x) + (bk − bk−1 ) + K n f (x)
(24.13)
n=0 k=1 n=bk−1 n=bC
C −1
bX L−1
X
= bC + K n f˜(x) + K n f (x) (24.14)
n=0 n=bC
C −1
bX L−1
X
≤ bC + K n f˜(x) + M (24.15)
n=0 n=bC
C −1
bX
≤ bC + M (L − 1 − bC ) + K n f˜(x) (24.16)
n=0
C −1
bX
≤ bC + M (N − 1) + K n f˜(x) (24.17)
n=0
L−1
X
≤ L + M (N − 1) + K n f˜(x) (24.18)
n=0

where the last step, going from bC to L, uses the fact that both  and f˜ are
non-negative. Taking expectations of both sides,
"L−1 # " L−1
#
X X
n n˜
E K f (X) ≤ E L + M (N − 1) + K f (X) (24.19)
n=0 n=0
L−1
X L−1
X h i
E K n f (X) ≤ L + M (N − 1) + E K n f˜(X)
 
(24.20)
n=0 n=0
h i
LE f (x) ≤ L + M (N − 1) + LE f˜(X)
 
(24.21)
CHAPTER 24. THE ALMOST-SURE ERGODIC THEOREM 202

using the fact that f is invariant on the left-hand side, and that m is stationary
on the other. Now divide both sides by L.

M (N − 1) h i
+ E f˜(X)
 
E f (x) ≤  + (24.22)
hL i
≤ 2 + E f˜(X) (24.23)
h i
since M N/L ≤ . Now let’s bound E f˜(X) in terms of E [f ]:

h i Z
E f˜ = f˜(x)dm (24.24)
Z Z
= f˜(x)dm + f˜(x)dm (24.25)
Bc B
Z Z
= f (x)dm + M dm (24.26)
Bc B
Z
≤ E [f ] + M dm (24.27)
B
= E [f ] + M m(B) (24.28)

≤ E [f ] + M (24.29)
M
= E [f ] +  (24.30)

using the definition of f˜ in Eq. 24.26, the non-negativity of f in Eq. 24.27, and
the bound on m(B) in Eq. 24.29. Substituting into Eq. 24.23,
 
E f ≤ E [f ] + 3 (24.31)

Since  can be made arbitrarily small, we conclude that

 
E f ≤ E [f ] (24.32)

as was to be shown. 
The proof of E f ≥ E [f ] proceeds in parallel, only the nice-ified function
f˜ is set equal
 to 0 on the bad
 set.  
Since E f ≥ E [f ] ≥ E f , we have that E f − f ≥ 0. Since however it is
always true that f −f ≥ 0, we may conclude that f −f = 0 m-almost everywhere.
Thus m(Lf ) = 1, i.e., the time average converges m-almost everywhere. Since
this is an invariant event, it has the same measure under µ and its stationary
limit m, and so the time average converges µ-almost-everywhere as well. By
Corollary 334, Af = Em [f |I], as promised. 

Corollary 340 (Birkhoff ’s Ergodic Theorem for Integrable Observ-

ables) Under the assumptions of Theorem 339, all L1 (m) functions have ergodic
properties, and Eq. 24.4 holds a.e. m and µ.
CHAPTER 24. THE ALMOST-SURE ERGODIC THEOREM 203

Proof: We need merely show that the ergodic properties hold, and then
the equation follows. To do so, define f M (x) ≡ f (x) ∧ M , an upper-limited
version of the lim sup. Reasoning
 entirely
 parallel to the proof of Theorem 339
leads to the conclusion that E f M ≤ E [f ]. Then let M → ∞, and apply the
 
monotone convergence theorem to conclude that E f ≤ E [f ]; the rest of the
proof goes through as before. 
Chapter 25

Ergodicity and Metric

Transitivity

Section 25.1 explains the ideas of ergodicity (roughly, there is

only one invariant set of positive measure) and metric transivity
(roughly, the system has a positive probability of going from any-
where to anywhere), and why they are (almost) the same.
Section 25.2 gives some examples of ergodic systems.
Section 25.3 deduces some consequences of ergodicity, most im-
portantly that time averages have deterministic limits (§25.3.1), and
an asymptotic approach to independence between events at widely
separated times (§25.3.2), admittedly in a very weak sense.

25.1 Metric Transitivity

Definition 341 (Ergodic Systems, Processes, Measures and Transfor-
mations) A dynamical system Ξ, X , µ, T is ergodic, or an ergodic system or an
ergodic process when µ(C) = 0 or µ(C) = 1 for every T -invariant set C. µ is
called a T -ergodic measure, and T is called a µ-ergodic transformation, or just
an ergodic measure and ergodic transformation, respectively.
Remark: Most authorities require a µ-ergodic transformation to also be
measure-preserving for µ. But (Corollary 54) measure-preserving transforma-
tions are necessarily stationary, and we want to minimize our stationarity as-
sumptions. So what most books call “ergodic”, we have to qualify as “stationary
and ergodic”. (Conversely, when other people talk about processes being “sta-
tionary and ergodic”, they mean “stationary with only one ergodic component”;
but of that, more later.
Definition 342 (Metric Transitivity) A dynamical system is metrically tran-
sitive, metrically indecomposable, or irreducible when, for any two sets A, B ∈
X , if µ(A), µ(B) > 0, there exists an n such that µ(T −n A ∩ B) > 0.

204
CHAPTER 25. ERGODICITY AND METRIC TRANSITIVITY 205

Remark: In dynamical systems theory, metric transitivity is contrasted with

topological transitivity: T is topologically transitive on a domain D if for any
two open sets U, V ⊆ D, the images of U and V remain in D, and there is
an n such that T n U ∩ V 6= ∅. (See, e.g., Devaney (1992).) The “metric”
in “metric transitivity” refers not to a distance function, but to the fact that
a measure is involved. Under certain conditions, metric transitivity in fact
implies topological transitivity: e.g., if D is a subset of a Euclidean space and
µ has a positive density with respect to Lebesgue measure. The converse is not
generally true, however: there are systems which are transitive topologically but
not metrically.
A dynamical system is chaotic if it is topologically transitive, and it contains
dense periodic orbits (Banks et al., 1992). The two facts together imply that a
trajectory can start out arbitrarily close to a periodic orbit, and so remain near
it for some time, only to eventually find itself arbitrarily close to a different
periodic orbit. This is the source of the fabled “sensitive dependence on ini-
tial conditions”, which paradoxically manifests itself in the fact that all typical
trajectories look pretty much the same, at least in the long run. Since metric
transitivity generally implies topological transitivity, there is a close connection
between ergodicity and chaos; in fact, most of the well-studied chaotic systems
are also ergodic (Eckmann and Ruelle, 1985), including the logistic map. How-
ever, it is possible to be ergodic without being chaotic: the one-dimensional
rotations with irrational shifts are, because there periodic orbits do not exist,
and a fortiori are not dense.

Lemma 343 (Metric transitivity implies ergodicity) If a dynamical sys-

tem is metrically transitive, then it is ergodic.

Proof: By contradiction. Suppose there was an invariant set A whose µ-

measure was neither 0 nor 1; then Ac is also invariant, and has strictly positive
measure. By metric transitivity, for some n, µ(T −n A∩Ac ) > 0. But T −n A = A,
and µ(A ∩ Ac ) = 0. So metrically transitive systems are ergodic. 
There is a partial converse.

Lemma 344 (Stationary Ergodic Systems are Metrically Transitive)

If a dynamical system is ergodic and stationary, then it is metrically transitive.
S∞
Proof: Take any µ(A), µ(B) > 0. Let Aever ≡ n=0 T −n A — the union of
A with all its pre-images. This set contains its pre-images, T −1 Aever ⊆ Aever ,
since if x ∈ T −n A, T −1 x ∈ T −n−1 A. The T
sequence of pre-images is thus non-
∞ S∞
increasing, and so tends to a limiting set, n=1 k=n T −k A = Ai.o. , the set of
points which not only visit A eventually, but visit A infinitely often. This is an
invariant set (Lemma 306), so by ergodicity it has either measure 0 or measure
1. By the Poincaré recurrence theorem (Corollaries 67 and 68), since µ(A) > 0,
µ(Ai.o. ) = 1. Hence, for any B, µ(Ai.o. ∩ B) = µ(B). But this means that, for
some n, µ(T −n A ∩ B) > 0, and the process is metrically transitive. 
CHAPTER 25. ERGODICITY AND METRIC TRANSITIVITY 206

25.2 Examples of Ergodicity

Example 345 (IID Sequences, Strong Law of Large Numbers) Every
IID sequence is ergodic. This is because the Kolmogorov 0-1 law states that every
tail event has either probability 0 or 1, and (Exercise 25.3) every invariant event
is a tail event. The strong law of large numbers is thus a two-line corollary of
the Birkhoff ergodic theorem.

Example 346 (Markov Chains) In the elementary theory of Markov chains,

an ergodic chain is one which is irreducible, aperiodic and positive recurrent.
To see that such a chain corresponds to an ergodic process in the present sense,
look at the shift operator on the sequence space. For consistency of notation, let
S1 , S2 , . . . be the values of the Markov chain in Σ, and X be the semi-infinite
sequence in sequence space Ξ, with shift operator T , and distribution µ over
sequences. µ is the product of an initial distribution ν ∼ S1 and the Markov-
family kernel. Now, “irreducible” means that one goes from every state to every
other state with positive probability at some lag, i.e., for every s1 , s2 ∈ Σ, there
is an n such that P (Sn = s2 |S1 = s1 ) > 0. But, writing [s] for the cylinder set in
Ξ with base s, this means that, for every [s1 ], [s2 ], µ(T −n [s2 ]∩[s1 ]) > 0, provided
µ([s1 ]) > 0. The Markov property of the S chain, along with positive recurrence,
can be used to extend this to all finite-dimensional cylinder sets (Exercise 25.4),
and so, by a generating-class argument, to all measurable sets.

Example 347 (Deterministic Ergodicity: The Logistic Map) We have

seen that the logistic map, T x = 4x(1−x), has an invariant density (with respect
to Lebesgue measure). It has an infinite collection of invariant sets, but the only
invariant interval is the whole state space [0, 1] — any smaller interval is not
invariant. From this, it is easy to show that all the invariant sets either have
measure 0 or measure 1 — they differ from ∅ or from [0, 1] by only a countable
collection of points. Hence, the invariant measure is ergodic. Notice, too, that
the Lebesgue measure on [0, 1] is ergodic, but not invariant.

Example 348 (Invertible Ergodicity: Rotations) Let Ξ = [0, 1), T x =

x + φ mod 1, and let µ be the Lebesgue measure on Ξ. (This corresponds to
a rotation, where the angle advances by 2πφ radians per unit time.) Clearly,
T preserve µ. If φ is rational, then, for any x, the sequence of iterates will
visit only finitely many points, and the process is not ergodic, because one can
construct invariant sets whose measure is neither 0 nor 1. (You may construct
such a set by taking any one of the periodic orbits, and surrounding its points
by internals of sufficiently small, yet positive, width.) If, on the other hand, φ
is irrational, then T n x never repeats, and it is easy to show that the process is
ergodic, because it is metrically transitive. Nonetheless, T is invertible.
This example (suitably generalized to multiple coordinates) is very important
in physics, because many mechanical systems can be represented in terms of
CHAPTER 25. ERGODICITY AND METRIC TRANSITIVITY 207

“action-angle” variables, the speed of rotation of the angular variables being set
by the actions, which are conserved, energy-like quantities. See Mackey (1992);
Arnol’d and Avez (1968) for the ergodicity of rotations and its limitations, and
Arnol’d (1978) for action-angle variables. Astonishingly, the result for the one-
dimensional case was proved by Nicholas Oresme in the 14th century (von Plato,
1994).

Example 349 (Ergodicity when the Distribution Does Not Converge)

Ergodicity does not ensure a uni-directional evolution of the distribution. (Some
people (Mackey, 1992) believe this has great bearing on the foundations of ther-
modynamics.) For a particularly extreme example, which also illustrates why
elementary Markov chain theory insists on aperiodicity, consider the period-two
deterministic chain, where state A goes to state B with probability 1, and vice
versa. Every sample path spends just much time in state A as in state B, so
every time average will converge on Em [f ], where m puts equal probability on
both states. It doesn’t matter what initial distribution we use, because they are
all ergodic (the only invariant sets are the whole space and the empty set, and
every distribution gives them probability 1 and 0, respectively). The uniform
distribution is the unique stationary distribution, but other distributions do not
approch it, since U 2n ν = ν for every integer n. So, At f → Em [f ] a.s., but
L (Xn ) 6→ m. We will see later that aperiodicity of Markov chains connects to
“mixing” properties, which do guarantee stronger forms of distributional con-
vergence.

25.3 Consequences of Ergodicity

The most basic consequence of ergodicity is that all invariant functions are
constant almost everywhere; this in fact characterizes ergodicity. This in turn
implies that time-averages converge to deterministic, rather than random, limits.
Another important consequence is that events widely separated in time become
nearly independent, in a somewhat funny-looking sense.

Theorem 350 (Ergodicity and the Triviality of Invariant Functions)

A T transformation is µ-ergodic if and only if all T -invariant observables are
constant µ-almost-everywhere.

Proof: “Only if”: Because invariant observables are I-measurable (Lemma

304), the pre-image under an invariant observable f of any Borel set B is an
invariant set. Since every invariant set has µ-probability 0 or 1, the probability
that f (x) ∈ B is either 0 or 1, hence f is constant with probability 1. “If”: The
indicator function of an invariant set is an invariant function. If all invariant
functions are constant µ-a.s., then for any A ∈ I, either 1A (x) = 0 or 1A (x) = 1
for µ-almost all x, which is the same as saying that either µ(A) = 0 or µ(A) = 1,
as required. 
CHAPTER 25. ERGODICITY AND METRIC TRANSITIVITY 208

25.3.1 Deterministic Limits for Time Averages

Theorem 351 (The Ergodic Theorem for Ergodic Processes) Suppose
µ is AMS, with stationary mean m, and T -ergodic. Then, almost surely,
lim At f (x) = Em [f ] (25.1)
t→∞

for µ- and m- almost all x, for any L1 (m) observable f .

Proof: Because every invariant set has µ-probability 0 or 1, it likewise has
m-probability 0 or 1 (Lemma 329). Hence, Em [f ] is a version of Em [f |I]. Since
At f is also a version of Em [f |I] (Corollary 340), they are equal almost surely.

An important consequence is the following. Suppose St is a strictly sta-
tionary random sequence. Let Φt (S) = f (St+τ1 , St+τ2 , . . . St+τn ) for some fixed
collection of shifts τn . Then Φt is another strictly stationary random sequence.
Every strictly stationary random sequence can be represented by a measure-
preserving transformation (Theorem 52), where X is the sequence S1 , S2 , . . ., the
mapping T is just the shift, and the measure µ is the infinite-dimensional mea-
sure of the original stochastic process. Thus Φt = φ(Xt ), for some measurable
function φ. If the measure is ergodic, and E [Φ] is finite, then the time-average
of Φ converges almost surely to its expectation. In particular,
P∞ let Φt = St St+τ .
Then, assuming the mixed moments are finite, t−1 t=1 St St+τ → E [St St+τ ]
almost surely, and so the sample covariance converges on the true covariance.
More generally, for a stationary ergodic process, if the n-point correlation func-
tions exist, the sample correlation functions converge a.s. on the true correlation
functions.

25.3.2 Ergodicity and the approach to independence

Lemma 352 (Ergodicity Implies Approach to Independence) If µ is
T -ergodic, and µ is AMS with stationary mean m, then
t−1
1X
lim µ(B ∩ T −n C) = µ(B)m(C) (25.2)
t→∞ t
n=0

for any measurable events B, C.

Proof: Exercise 25.1. 
Theorem 353 (Approach to Independence Implies Ergodicity) Suppose
X is generated by a field F. Then an AMS measure µ, with stationary mean
m, is ergodic if and only if, for all F ∈ F,
t−1
1X
lim µ(F ∩ T −n F ) = µ(F )m(F ) (25.3)
t→∞ t
n=0

i.e., iff Eq. 25.2 holds, taking B = C = F ∈ F.

Proof: “Only if”: Lemma 352. “If”: Exercise 25.2. 
CHAPTER 25. ERGODICITY AND METRIC TRANSITIVITY 209

25.4 Exercises
Exercise 25.1 (Ergodicity implies an approach to independence) Prove
Lemma 352.

Exercise 25.2 (Approach to independence implies ergodicity) Prove

the “if ” part of Theorem 353.

Exercise 25.3 (Invariant events and tail events) Prove that every invari-
ant event is a tail event. Does the converse hold?

Exercise 25.4 (Ergodicity of ergodic Markov chains) Complete the argu-

ment in Example 346, proving that ergodic Markov chains are ergodic processes
(in the sense of Definition 341).
Chapter 26

Decomposition of
Stationary Processes into
Ergodic Components

This chapter is concerned with the decomposition of asymptotically-

mean-stationary processes into ergodic components.
Section 26.1 introduces some preliminary ideas about convex
mixtures of invariant distributions, and in particular characterizes
ergodic distributions as extremal points.
Section 26.2 shows how to write the stationary distribution as a
mixture of distributions, each of which is stationary and ergodic, and
each of which is supported on a distinct part of the state space. This
is connected to ideas in nonlinear dynamics, each ergodic component
being a different basin of attraction.
Section 26.3 lays out some connections to statistical inference:
ergodic components can be seen as minimal sufficient statistics, and
lead to powerful tests.

26.1 Preliminaries to Ergodic Decompositions

It is always the case, with a dynamical system, that if x lies within some invariant
set A, then all its future iterates stay within A as well. In general, therefore, one
might expect to be able to make some predictions about the future trajectory
by knowing which invariant sets the initial condition lies within. An ergodic
process is one where this is actually not possible. Because all invariants sets
have probability 0 or 1, they are all independent of each other, and indeed of
every other set. Therefore, knowing which invariant sets x falls into is completely
uninformative about its future behavior. In the more general non-ergodic case,
a limited amount of prediction is however possible on this basis, the limitations

210
CHAPTER 26. ERGODIC DECOMPOSITION 211

being set by the way the state space breaks up into invariant sets of points with
the same long-run average behavior — the ergodic components. Put slightly
differently, the long-run behavior of an AMS system can be represented as a
mixture of stationary, ergodic distributions, and the ergodic components are, in
a sense, a minimal parametrically sufficient statistic for this distribution. (They
are not in generally predictively sufficient.)
The idea of an ergodic decomposition goes back to von Neumann, but was
considerably refined subsequently, especially by the Soviet school, who seem to
have introduced most of the talk of predictions, and all of the talk of ergodic
components as minimal sufficient statistics. Our treatment will follow Gray
(1988, ch. 7), and Dynkin (1978).

Proposition 354 Any convex combination of invariant probability measures is

an invariant probability measure.

Proof: Let µ1 and µ2 be two invariant probability measures. It is elementary

that for every 0 ≤ a ≤ 1, ν ≡ aµ1 + (1 − a)µ2 is a probability measure. Now
consider the measure under ν of the pre-image of an arbitrary measurable set
B ∈ X:

ν(T −1B) = aµ1 (T −1 B) + (1 − a)µ2 (T −1 B) (26.1)

= aµ1 (B) + (1 − a)µ2 (B) (26.2)
= ν(B) (26.3)

so ν is also invariant. 

Proposition 355 If µ1 and µ2 are invariant ergodic measures, then either µ1 =

µ2 , or they are singular, meaning that there is a set B on which µ1 (B) = 0,
µ2 (B) = 1.

Proof: Suppose µ1 6= µ2 . Then there is at least one set C where µ1 (C) 6=

µ2 (C). Because both µi are stationary and ergodic, At 1C (x) converges to µi (C)
for µi -almost-all x. So the set
n o
x| lim At 1C (x) = µ2 (C)
t

has a µ2 measure of 1, and a µ1 measure of 0 (since, by hypothesis, µ1 (C) 6=

µ2 (C). 

Proposition 356 Ergodic invariant measures are extremal points of the convex
set of invariant measures, i.e., they cannot be written as combinations of other
invariant measures.

Proof: By contradiction. That is, suppose µ is ergodic and invariant, and

that there were invariant measures ν and λ, and an a ∈ (0, 1), such that µ =
aν + (1 − a)λ. Let C be any invariant set; then µ(C) = 0 or µ(C) = 1.
Suppose µ(C) = 0. Then, because a is strictly positive, it must be the case that
CHAPTER 26. ERGODIC DECOMPOSITION 212

ν(C) = λ(C) = 0. If µ(C) = 1, then C c is also invariant and has µ-measure 0,

so ν(C c ) = λ(C c ) = 0, i.e., ν(C) = λ(C) = 1. So ν and λ would both have to
be ergodic, with the same support as µ. But then (Proposition 355 preceeding)
λ = ν = µ. 
Remark: The converse is left as an exercise (26.1).

26.2 Construction of the Ergodic Decomposi-

tion
In the last section, we saw that the stationary distributions of a given dynamical
system form a convex set, with the ergodic distributions as the extremal points.
A standard result in convex analysis is that any point in a convex set can
be represented as a convex combination of the extremal points. Thus, any
stationary distribution can be represented as a mixture of stationary and ergodic
distributions. We would like to be able to determine the weights used in the
mixture, and even more to give them some meaningful stochastic interpretation.
Let’s begin by thinking about the effective distribution we get from taking
time-averages starting from a given point. For every measurable set B, and
every finite t, At 1B (x) is a well-defined measurable function. As B ranges over
the σ-field X , holding x and t fixed, we get a set function, and one which,
moreover, meets the requirements for being a probability measure. Suppose we
go further and pass to the limit.

Definition 357 (Long-Run Distribution) The long-run distribution start-

ing from the point x is the set function λ(x), defined through λ(x, B) = limt At 1B (x),
when the limit exists for all B ∈ X . If λ(x) exists, x is an ergodic point. The
set of all ergodic points is E.

Notice that whether or not λ(x) exists depends only on x (and T and X );
the initial distribution has nothing to do with it. Let’s look at some properties
of the long-run distributions. (The name “ergodic point” is justified by one of
them, Proposition 359.)

Proposition 358 If x ∈ E, then λ(x) is a probability distribution.

Proof: For every t, the set function given by At 1B (x) is clearly a probability
measure. Since λ(x) is defined by passage to the limit, the Vitali-Hahn Theorem
(Proposition 327) says λ(x) must be as well. 

Proposition 359 If x ∈ E, then λ(x) is an ergodic distribution.

Proof: For every invariant set I, 1I (T n x) = 1I (x) for all n. Hence A1I (x)
exists and is either 0 or 1. This means λ(x) assigns every invariant set either
probability 0 or probability 1, so by Definition 341 it is ergodic. 

Proposition 360 If x ∈ E, then λ(x) is an invariant function of x, i.e., λ(x) =

λ(T x).
CHAPTER 26. ERGODIC DECOMPOSITION 213

Proof: By Lemma 317, A1B (x) = A1B (T x), when the appropriate limit exists.
Since, by assumption, it does in this case, for every measurable set λ(x, B) =
λ(T x, B), and the set functions are thus equal. 

Proposition 361 If x ∈ E, then λ(x) is a stationary distribution.

Proof: For all B and x, 1T −1 B (x) = 1B (T x). So λ(x, T −1 B) = λ(T x, B).

Since, by Proposition 360, λ(T x, B) = λ(x, B), it finally follows that λ(x, B) =
λ(x, T −1 B), which proves that λ(x) is an invariant distribution. 

Proposition 362 If x ∈ E and f ∈ L1 (λ(x)), then limt At f (x) exists, and is

equal to Eλ(x) [f ].

Proof: This is true, by the definition of λ(x), for the indicator functions of
all measurable sets. Thus, by linearity of At and of expectation, it is true for
all simple functions. Standard arguments then let us pass to all the functions
integrable with respect to the long-run distribution. 
At this point, you should be tempted to argue as follows. If µ is an AMS
distribution with stationary mean m, then Af (x) = Em [f |I] for almost all x.
So, it’s reasonable to hope that m is a combination of the λ(x), and yet further
that
Af (x) = Eλ(x) [f ]
for µ-almost-all x. This is basically true, but will take some extra assumptions
to get it to work.

Definition 363 (Ergodic Component) Two ergodic points x, y ∈ E belong

to the same ergodic component when λ(x) = λ(y). We will write the ergodic
components as Ci , and the function mapping x to its ergodic component as φ(x).
φ(x) is not defined if x is not an ergodic point. By a slight abuse of notation,
we will write λ(Ci , B) for the common long-run distribution of all points in Ci .

Obviously, the ergodic components partition the set of ergodic points. (The
partition is not necessarily countable, and in some important cases, such as
that of Hamiltonian dynamical systems in statistical mechanics, it must be
uncountable (Khinchin, 1949).) Intuitively, they form the coarsest partition
which is still fully informative about the long-run distribution. It’s also pretty
clear that the partition is left alone with the dynamics.

Proposition 364 For all ergodic points x, φ(x) = φ(T x).

Proof: By Lemma 360, λ(x) = λ(T x), and the result follows. 
Notice that I have been careful not to say that the ergodic components are
invariant sets, because we’ve been using that to mean sets which are both left
along by the dynamics and are measurable, i.e. members of the σ-field X , and
we have not established that any ergodic component is measurable, which in
turn is because we have not established that λ(x) is a measurable function.
CHAPTER 26. ERGODIC DECOMPOSITION 214

Let’s look a little more closely at the difficulty. If B is a measurable set,

then At 1B (x) is a measurable function. If the limit exists, then A1B (x) is also
a measurable function, and consequently the set {y : A1B (y) = A1B (x)} is a
measurable set. Then
\
φ(x) = {y : A1B (x) = A1B (y)} (26.4)
B∈X

gives the ergodic component to which x belongs. The difficulty is that the
intersection is over all measurable sets B, and there are generally an uncountable
number of them (even if Ξ is countable!), so we have no guarantee that the
intersection of uncountably many measurable sets is measurable. Consequently,
we can’t say that any of the ergodic components is measurable.
The way out, as so often in mathematics, is to cheat; or, more politely,
to make an assumption which is strong enough to force open an exit, but not
so strong that we can’t support it or verify it1 What we will assume is that
there is a countable collection of sets G such that λ(x) = λ(y) if and only if
λ(x, G) = λ(y, G) for every G ∈ G. Then the intersection in Eq. 26.4 need only
run over the countable class G, rather than all of X , which will be enough to
reassure us that φ(x) is a measurable set.

Definition 365 (Countable Extension Space) A measurable space Ω, F is

a countable extension space when there is a countable field G of sets in Ω such
that F = σ(G), i.e., G is the generating field of the σ-field, and any normalized,
non-negative, finitely-additive set function on G has a unique extension to a
probability measure on F.

The reason the countable extension property is important is that it lets us

get away with just checking properties of measures on a countable class (the
generating field G). Here are a few important facts about countable extension
spaces; proofs, along with a much more detailed treatment of the general theory,
are given by Gray (1988, chs. 2 and 3), who however calls them “standard”
spaces.

Proposition 366 Every countable space is a countable extension space.

Proposition 367 Every Borel space is a countable extension space.

Remember that finite-dimensional Euclidean spaces are Borel spaces.

Proposition 368 A countable product of countable extension spaces is a count-

able extension space.
1 For instance, we could just assume that uncountable intersections of measurable sets

are measurable, but you will find it instructive to try to work out the consequences of this
assumption, and to examine whether it holds for the Borel σ-field B — say on the unit interval,
to keep things easy.
CHAPTER 26. ERGODIC DECOMPOSITION 215

The last proposition is important for us: if Σ is a countable extension space,

it means that Ξ ≡ ΣN is too. So if we have a discrete- or Euclidean- valued
random sequence, we can switch to the sequence space, and still appeal to
generating-class arguments based on countable fields. Without further ado,
then, let’s assume that Ξ, the state space of our dynamical system, is a countable
extension space, with countable generating field G.
Lemma 369 x ∈ E iff limt At 1G (x) converges for every G ∈ G.
Proof: “If”: A direct consequence of Definition 365, since the set function
A1G (x) extends to a unique measure. “Only if”: a direct consequence of Defi-
nition 357, since every member of the generating field is a measurable set. 
Lemma 370 The set of ergodic points is measurable: E ∈ X .
Proof: For each G ∈ G, the set of x where At 1G (x) converges is measurable,
because G is a measurable set. The set where those relative frequencies converge
for all G ∈ G is the intersection of countably many measurable sets, hence itself
measurable. This set is, exactly, the set of ergodic points (Lemma 369). 
Lemma 371 All the ergodic components are measurable sets, and φ(x) is a
measurable function. Thus, all Ci ∈ I.
Proof: For each G, the set {y : λ(y, G) = λ(x, G)} is measurable. So their
intersection over all G ∈ G is also measurable. But, by the countable extension
property, this intersection is precisely the set {y : λ(y) = λ(x)}. So the ergodic
components are measurable sets, and, since φ−1 (Ci ) = Ci , φ is measurable.
Since we have already seen that T −1 Ci = Ci , and now that Ci ∈ X , we may
say that Ci ∈ I. 
Remark: Because Ci is a (measurable) invariant set, λ(x, Ci ) = 1 for every
x ∈ Ci . However, it does not follow that there might not be a smaller set, also
with long-run measure 1, i.e., there might be a B ⊂ Ci such that λ(x, B) = 1.
For an extreme example, consider the uniform contraction on R, with T x = ax
for some 0 ≤ a ≤ 1. Every trajectory converges on the origin. The only ergodic
invariant measure the the Dirac delta function. Every point belongs to a single
ergodic component.
More generally, if a little roughly2 , the ergodic components correspond to
the dynamical systems idea of basins of attraction, while the support of the
long-run distributions corresponds to the actual attractors. Basins of attraction
typically contain points which are not actually parts of the attractor.
Theorem 372 (Ergodic Decomposition of AMS Processes) Suppose Ξ, X
is a countable extension space. If µ is an asymptotically mean stationary mea-
sure on Ξ, with stationary mean m, then µ(E) = m(E) = 1, and, for any
f ∈ L1 (m), and µ- and m- almost all x,
Af (x) = Eλ(x) [f ] = Em [f |I] (26.5)
2Idon’t want to get into subtleties arising from the dynamicists tendency to define things
topologically, rather than measure-theoretically.
CHAPTER 26. ERGODIC DECOMPOSITION 216

so that Z
m(B) = λ(x, B)dµ(x) (26.6)

Proof: For every set G ∈ G, At 1G (x) converges for µ- and m- almost all
x (Theorem 339). Since there are only countably many G, the set on which
they all converge also has probability 1; this set is E. Since (Proposition 362)
Af (x) = Eλ(x) [f ], and (Theorem 339 again) Af (x) = Em [f |I] a.s., we have
that Eλ(x) [f ] = Em [f |I] a.s.
Now let f = 1B . As we know (Lemma 331), Eµ [A1B (X)] = Em [1B (X)] =
m(B). But, for each x, A1B (x) = λ(x, B), so m(B) = Eµ [λ(X, B)]. 
In words, we have decomposed the stationary mean m into the long-run
distributions of the ergodic components, with weights given by the fraction of
the initial measure µ falling into each component. Because of Propositions 354
and 356, we may be sure that by mixing stationary ergodic measures, we obtain
an ergodic measure, and that our decomposition is unique.

26.3 Statistical Aspects

26.3.1 Ergodic Components as Minimal Sufficient Statis-
tics
The connection between sufficient statistics and ergodic decompositions is a very
pretty one. First, recall the idea of parametric statistical sufficiency.3

Definition 373 (Sufficiency, Necessity) Let P be a class of probability mea-

sures on a common measurable space Ω, F, indexed by a parameter θ. A σ-field
S ⊆ F is parametrically sufficient for θ, or just sufficient, when Pθ (A|S) =
Pθ0 (A|S) for all θ, θ0 . That is, all the distributions in P have the same distri-
bution, conditional on S. A random variable such that S = σ(S) is called a
sufficient statistic. A σ-field is necessary (for the parameter θ) if it is a sub-
σ-field of every sufficient σ-field; a necessary statistic is defined similarly. A
σ-field which is both necessary and sufficient is minimal sufficient.

Remark: The idea of sufficiency originates with Fisher; that of necessity, so

far as I can work out, with Dynkin. This definition (after Dynkin (1978)) is
based on what ordinary theoretical statistics texts call the “Neyman factoriza-
tion criterion” for sufficiency. We will see all these concepts again when we do
information theory.
3 There is also a related idea of predictive statistical sufficiency, which we unfortunately

will not be able to get to. Also, note that most textbooks on theoretical statistics state things
in terms of random variables and measurable functions thereof, rather than σ-fields, but this
is the more general case (Blackwell and Girshick, 1954).
CHAPTER 26. ERGODIC DECOMPOSITION 217

Lemma 374 S is sufficient for θ if and only if there exists an F-measurable

function λ(ω, A) such that

Pθ (A|S) = λ(ω, A) (26.7)

almost surely, for all θ.

Proof: Nearly obvious. “Only if”: since the conditional probability exists,
there must be some such function (it’s a version of the conditional probabil-
ity), and since all the conditional probabilities are versions of one another, the
function cannot depend on θ. “If”: In this case, we have a single function
which is a version of all the conditional probabilities, so it must be true that
Pθ (A|S) = Pθ0 (A|S). 

Theorem 375 If a process on a countable extension space is asymptotically

mean stationary, then φ is a minimal sufficient statistic for its long-run distri-
bution.

Proof: The set of distributions P is now the set of all long-run distributions
generated by the dynamics, and θ is an index which tracks them all unambigu-
ously. We need to show both sufficiency and necessity. Sufficiency: The σ-field
generated by φ is the one generated by the ergodic components, σ({Ci }). (Be-
cause the Ci are mutually exclusive, this is a particularly simple σ-field.) Clearly,
Pθ (A|σ({Ci })) = λ(φ(x), A) for all x and θ, so (Lemma 374), φ is a sufficient
statistic. Necessity: Follows from the fact that a given ergodic component con-
tains all the points with a given long-run distribution. Coarser σ-fields will not,
therefore, preserve conditional probabilities. 
This theorem may not seem particularly exciting, because there isn’t, neces-
sarily, anything whose distribution matches the long-run distribution. However,
it has deeper meaning under two circumstances when λ(x) really is the asymp-
totic distribution of random variables.
1. If Ξ is really a sequence space, so that X = S1 , S2 , S3 , . . ., then λ(x)
really is the asymptotic marginal distribution of the St , conditional on
the starting point.
2. Even if Ξ is not a sequence space, if stronger conditions than ergodicity
known as “mixing”, “asymptotic stability”, etc., hold, there are reason-
able senses in which L (Xt ) does converge, and converges on the long-run
distribution.4
In both these cases, knowing the ergodic component thus turns out to be neces-
sary and sufficient for knowing the asymptotic distribution of the observables.
(Cf. Corollary 378 below.)
4 Lemma 352 already gave us a kind of distributional convergence, but it is of a very

weak sort, known as “convergence in Cesàro mean”, which was specially invented to handle
sequences which are not convergent in normal senses! We will see that there is a direct
correspondence between levels of distributional convergence and levels of decay of correlations.
CHAPTER 26. ERGODIC DECOMPOSITION 218

26.3.2 Testing Ergodic Hypotheses

Finally, let’s close with an application to hypothesis testing, inspired by Badino
(2004).

Theorem 376 Let Ξ, X be a measurable space, and let µ0 and µ1 be two infinite-
dimensional distributions of one-sided, discrete-parameter strictly-stationary Σ-
valued stochastic processes, i.e., µ0 and µ1 are distributions on ΞN , X N , and
they are invariant under the shift operator. If they are also ergodic under the
shift, then there exists a sequence of sets Rt ∈ X t such that µ0 (Rt ) → 0 while
µ1 (Rt ) → 1.

Proof: By Proposition 355, there exists a set R ∈ X N such that µ0 (R) = 0,

µ1 (R) = 1. So we just need to approximate B by sets which are defined on
the first t observations in such a way that µi (Rt ) → µi (R). If Rt ↓ R, then
monotone convergence will give us the necessary convergence of probabilities.
Here is a construction with cylinder sets5 that gives us the necessary sequence
of approximations. Let
Y∞
Rt ≡ R ∪ Ξt (26.8)
n=t+1

Clearly, Rt forms a non-increasing sequence, so it converges to a limit, which

equally clearly must be R. Hence µi (Rt ) → µi (R) = i. 
Remark: “R” is for “rejection”. Notice that the regions Rt will in general
depend on the actual sequence X1 , X2 , . . . Xt ≡ X1t , and not necessarily be
permutation-invariant. When we come to the asymptotic equipartition theorem
in information theory, we will see a more explicit way of constructing such tests.

Corollary 377 Let H0 be “Xi are IID with distribution p0 ” and H1 be “Xi are
IID with distribution p1 ”. Then, as t → ∞, there exists a sequence of tests of
H0 against H1 whose size goes to 0 while their power goes to 1.

Proof: Let µ0 be the product measure induced by p0 , and µ1 the product

measure induced p1 , and apply the previous theorem. 

Corollary 378 If X is a strictly stationary (one-sided) random sequence whose

shift representation has countably-many ergodic components, then there exists a
sequence of functions φt , each Xt -measurable, such that φt (X1t ) converges on the
ergodic component with probability 1.

Proof: From Theorem 52, we can write X1t = π1:t U , for a sequence-valued
random variable U , using the projection operators of Chapter 2. For each
ergodic component, by Theorem 376, there exists a sequence of sets Rt,i such
that P (X1t ∈ Rt,i ) → 1 if U ∈ Ci , and goes to zero otherwise. Let φ(X1t ) be the
5 Introduced in Chapters 2 and 3. It’s possible to give an alternative construction using the

Hilbert space of all square-integrable random variables, and then projecting onto the subspace
of those which are X t measurable.
CHAPTER 26. ERGODIC DECOMPOSITION 219

set of all Ci for which X1t ∈ Rt,i . By Theorem 372, U is in some component with
probability 1, and, since there are only countably many ergodic components,
with probability 1 X1t will eventually leave all but one of the Rt,i . The remaining
one is the ergodic component. 

26.4 Exercises
Exercise 26.1 Prove the converse to Proposition 356: every extermal point of
the convex set of invariant measures is an ergodic measure.
Chapter 27

Mixing

A stochastic process is mixing if its values at widely-separated

times are asymptotically independent.
Section 27.1 defines mixing, and shows that it implies ergodicity.
Section 27.2 gives some examples of mixing processes, both de-
terministic and non-deterministic.
Section 27.3 looks at the weak convergence of distributions pro-
duced by mixing, and the resulting decay of correlations.
Section 27.4 defines strong mixing, and the “mixing coefficient”
which measures it. It then states, but does not prove, a central limit
theorem for strongly mixing sequences. (The proof would demand
first working through the central limit theorem for martingales.)

For stochastic processes, “mixing” means “asymptotically independent”:

that is, the statistical dependence between X(t1 ) and X(t2 ) goes to zero as
|t1 − t2 | increases. To make this precise, we need to specify how we measure the
dependence between X(t1 ) and X(t2 ). The most common and natural choice
(first used by Rosenblatt, 1956) is the total variation distance between their
joint distribution and the product of their marginal distributions, but there are
other ways of measuring such “decay of correlations”1 . Under all reasonable
choices, IID processes are, naturally enough, special cases of mixing processes.
This suggests that many of the properties of IID processes, such as laws of
large numbers and central limit theorems, should continue to hold for mixing
processes, at least if the approach to independence is sufficiently rapid. This in
turn means that many statistical methods originally developed for the IID case
will continue to work when the data-generating process is mixing; this is true
both of parametric methods, such as linear regression, ARMA models being
mixing (Doukhan, 1995, sec. 2.4.1), and of nonparametric methods like kernel
prediction (Bosq, 1998). Considerations of time will prevent us from going into
1 The term is common, but slightly misleading: lack of correlation, in the ordinary

covariance-normalized-by-standard-deviations sense, implies independence only in special

cases, like Gaussian processes. Nonetheless, see Theorem 391.

220
CHAPTER 27. MIXING 221

the purely statistical aspects of mixing processes, but the central limit theorem
at the end of this chapter will give some idea of the flavor of results in this area:
much like IID results, only with the true sample size replaced by an effective
sample size, with a smaller discount the faster the rate of decay of correlations.

27.1 Definition and Measurement of Mixing

Definition 379 (Mixing) A dynamical system Ξ, X , µ, T is mixing when, for
any A, B ∈ X ,

lim |µ(A ∩ T −t B) − µ(A)µ(T −t B)| = 0 (27.1)

t→∞

Lemma 380 If µ is T -invariant, mixing is equivalent to

lim µ(A ∩ T −t B) = µ(A)µ(B) (27.2)

t→∞

Proof: By stationarity, µ(T −t B) = µ(B), so µ(A)µ(T −t B) = µ(A)µ(B). The

result follows. 

Theorem 381 Mixing implies ergodicity.

Proof: Let A be any invariant set. By mixing, limt µ(T −t A ∩ A) = µ(T −t A)µ(A).
But T −t A = A for every t, so we have lim µ(A) = µ2 (A), or µ(A) = µ2 (A). This
can only be true if µ(A) = 0 or
mu(A) = 1, i.e., only if µ is T -ergodic. 
Everything we have established about ergodic processes, then, applies to
mixing processes.

Definition 382 A dynamical system is asymptotically stationary, with station-

ary limit m, when limt µ(T −t A) = m(A) for all A ∈ X .

Lemma 383 An asymptotically stationary system is mixing iff

lim µ(A ∩ T −t B) = µ(A)m(B) (27.3)

t→∞

for all A, B ∈ X .

Proof: Directly from the fact that in this case m(B) = limt T −t B. 

Theorem 384 Suppose G is a π-system, and µ is an asymptotically stationary

measure. If
lim µ(A ∩ T −t B) − µ(A)µ(T −t B) = 0
 
(27.4)
t

for all A, B ∈ G, then it holds for all pairs of sets in σ(G). If σ(G) = X , then
the process is mixing.
CHAPTER 27. MIXING 222

Proof(after Durrett, 1991, Lemma 6.4.3): Via the π-λ theorem, of course. Let
ΛA be the class of all B such that the equation holds, for a given A ∈ G. We
need to show that ΛA really is a λ-system.
Ξ ∈ ΛA is obvious. T −t Ξ = Ξ so µ(A ∩ Ξ) = µ(A) = µ(A)µ(Ξ).
Closure under complements. Let B1 and B2 be two sets in ΛA , and assume
B1 ⊂ B2 . Because set-theoretic operations commute with taking inverse images,
T −t (B2 \ B1 ) = T −t B2 \ T −t B1 . Thus

0 ≤ |µ A ∩ T −t (B2 \ B1 ) − µ(A)µ(T −t (B2 \ B1 ))|

(27.5)
−t −t −t −t
= |µ(A ∩ T B2 ) − µ(A ∩ T B1 ) − µ(A)µ(T B2 ) + µ(A)µ(T B1 )|
≤ |µ(A ∩ T −t B2 ) − µ(A)µ(T −t B2 )| (27.6)
−t −t
+|µ(A ∩ T B1 ) − µ(A)µ(T B1 )|

Taking limits of both sides, we get that lim |µ (A ∩ T −t (B2 \ B1 )) − µ(A)µ(T −t (B2 \ B1 ))| =
0, so that B2 \ B1 ∈ ΛA .
Closure under monotone limits: Let Bn be any monotone increasing sequence
in ΛA , with limit B. Thus, µ(Bn ) ↑ µ(B), and at the same time m(Bn ) ↑ m(B),
where m is the stationary limit of µ. Using Lemma 383, it is enough to show
that
lim µ(A ∩ T −t B) = µ(A)m(B) (27.7)
t

Since Bn ⊂ B, we can always use the following trick:

µ(A ∩ T −t B) = µ(A ∩ T −t Bn ) + µ(A ∩ T −t (B \ Bn )) (27.8)

lim µ(A ∩ T −t B) = µ(A)m(Bn ) + lim µ(A ∩ T −t (B \ Bn )) (27.9)
t t

For any  > 0, µ(A)m(Bn ) can be made to come within  of µ(A)m(B) by

taking n sufficiently large. Let us now turn our attention to the second term.

0 ≤ lim µ(A ∩ T −t (B \ Bn )) = lim µ(T −t (B \ Bn )) (27.10)

t t
= lim µ(T −t B \ T −t Bn ) (27.11)
t
= lim µ(T −t B) − lim µ(T −t Bn ) (27.12)
t t
= m(B) − m(Bn ) (27.13)

which again can be made less than any positive  by taking n large. So, for
sufficiently large n, limt µ(A ∩ T −t B) is always within 2 of µ(A)m(B). Since 
can be made arbitrarily small, we conclude that limt µ(A ∩ T −t B) = µ(A)m(B).
Hence, B ∈ ΛA .
We conclude, from the π − λ theorem, that Eq. 27.4 holds for all A ∈ G and
all B ∈ σ(G). The same argument can be turned around for A, to show that
Eq. 27.4 holds for all pairs A, B ∈ σ(G). If G generates the whole σ-field X ,
then clearly Definition 379 is satisfied and the process is mixing. 
CHAPTER 27. MIXING 223

27.2 Examples of Mixing Processes

Example 385 (IID Sequences) IID sequences are mixing from Theorem 384,
applied to finite-dimensional cylinder sets.

Example 386 (Ergodic Markov Chains) Another application of Theorem

384 shows that ergodic Markov chains are mixing.

Example 387 (Irrational Rotations of the Circle are Not Mixing) Irrational
rotations of the circle, T x = x + φ mod 1, φ irrational, are ergodic (Example
348), and stationary under the Lebesgue measure. They are not, however, mix-
ing. Recall that T t x is dense in the unit interval, for arbitrary initial x. Because
it is dense, there is a sequence tn such that tn φ mod 1 goes to 1/2. Now let
A = [0, 1/4]. Because T maps intervals to intervals (of equal length), it follows
that T −tn A becomes an interval disjoint from A, i.e., µ(A ∩ T −tn A) = 0. But
mixing would imply that µ(A∩T −tn A) → 1/16 > 0, so the process is not mixing.

Example 388 (Deterministic, Reversible Mixing: The Cat Map) Here

2
Ξ = [0, 1) , X are the appropriate Borel sets, µ is Lebesgue measure on the
square, and T x = (x1 + x2 , x1 + 2x2 ) mod 1. This is known as the cat map. It
is a deterministic, invertible transformation, but it can be shown that it is actu-
ally mixing. (For a proof, which uses Theorem 390, the Fibonacci numbers and
a clever trick with Fourier transforms, see Lasota and Mackey (1994, example
4.4.3, pp. 77–78).) The origins of the name lie with a figure in Arnol’d and
Avez (1968), illustrating the mixing action of the map by successively distorting
an image of a cat.

27.3 Convergence of Distributions Under Mix-

ing
To show how distributions converge (weakly) under mixing, we need to recall
some properties of Markov operators. Remember that, for a Markov process,
the time-evolution operator for observables, K, was defined through Kf (x) =
E [f (X1 )|X0 = x]. Remember also that it induces an adjoint operator for the
evolution of distributions, taking signed measures to signed measures, through
the intermediary of the transition kernel. We can view the measure-updating
operator U as a linear operator on L1 (µ), which takes non-negative µ-integrable
functions to non-negative µ-integrable functions, and probability densities to
probability densities. Since dynamical systems are Markov processes, all of this
remains valid; we have K defined through Kf (x) = f (T x), and U through the
adjoint relationship, Eµ [f (X)Kg(X)] = E [U f (X)g(X)] µ, where g ∈ L∞ and
f ∈ L1 (µ). These relations continue to remain valid for powers of the operators.
CHAPTER 27. MIXING 224

Lemma 389 In any Markov process, U n d converges weakly to 1, for all initial
probability densities d, if and only if U n f converges weakly to Eµ [f ], for all
initial L1 functions f , i.e. Eµ [U n f (X)g(X)] → Eµ [f (X)] Eµ [g(X)] for all
bounded, measurable g.

Proof: “If”: If d is a probability density with respect to µ, then Eµ [d] = 1.

“Only if”: Re-write an arbitrary f ∈ L1 (µ) as the difference of its positive
and negative parts, f = f + − f − . A positive f is a re-scaling of some density,
f = cd for constant c = Eµ [f ] and a density d. Through the linearity of U and
its powers,

lim U t f = lim U t f + − lim U t f − (27.14)

= Eµ f + lim U t d+ − Eµ f − lim U t d−
   
(27.15)
= Eµ f + − Eµ f −
   
(27.16)
= Eµ f + − f − = Eµ [f ]
 
(27.17)

using the linearity of expectations at the last step. 

Theorem 390 A T -invariant probability measure µ is T -mixing if and only if

any initial probability measure ν << µ converges weakly to µ under the action
of T , i.e., iff, for all bounded, measurable f ,

EU t ν [f (X)] → Eµ [f (X)] (27.18)

Proof: Exercise. The way to go is to use the previous lemma, of course. With
that tool, one can prove that the convergence holds for indicator functions, and
then for simple functions, and finally, through the usual arguments, for all L1
densities.

Theorem 391 (Decay of Correlations) A stationary system is mixing if and

only if
lim cov (f (X0 ), g(Xt )) = 0 (27.19)
t→∞

for all bounded observables f , g.

Proof: Exercise, from the fact that convergence in distribution implies con-
vergence of expectations of all bounded measurable functions. 
It is natural to ask what happens if U t ν → µ not weakly but strongly. This
is known as asymptotic stability or (especially in the nonlinear dynamics liter-
ature) exactness. Remarkably enough, it is equivalent to the requirement that
µ(T t A) → 1 whenever µ(A) > 0. (Notice that for once the expression involves
images rather than pre-images.) There is a kind of hierarchy here, where differ-
ent levels of convergence of distribution (Cesáro, weak, strong) match different
sorts of ergodicity (metric transitivity, mixing, exactness). For more details, see
Lasota and Mackey (1994).
CHAPTER 27. MIXING 225

27.4 A Central Limit Theorem for Mixing Se-

quences
Notice that I say “a central limit theorem”, rather than “the central limit the-
orem”. In the IID case, the necessary and sufficient condition for the CLT is
well-known (you saw it in 36-752) and reasonably comprehensible. In the mixing
case, a necessary and sufficient condition is known2 , but not commonly used,
because quite opaque and hard to check. Rather, the common practice is to
rely upon a large set of distinct sufficient conditions. Some of these, it must be
said, are pretty ugly, but they are more susceptible of verification.
Recall the notation that Xt− consists of the entire past of the process, in-
cluding Xt , and Xt+ its entire future.
Definition 392 (Mixing Coefficients) For a stochastic process Xt , define
the strong-, Rosenblatt- or α- mixing coefficients as
α(t1 , t2 ) = sup |P (A ∩ B) − P (A) P (B)| : A ∈ σ(Xt−1 ), B ∈ σ(Xt+2 ) (27.20)


If the system is conditionally stationary, then α(t1 , t2 ) = α(t2 , t1 ) = α(|t1 −

t2 |) ≡ α(τ ). If α(τ ) → 0, then the process is strong-mixing or α-mixing. If
α(τ ) = O(e−bτ ) for some b > 0, the process is exponentially mixing, b is the
mixing rate, and 1/b is the mixing time. If α(τ ) = O(τ −k ) for some k > 0,
then the process is polynomially mixing.
Notice that α(t1 , t2 )
is just the total variation distance between the joint distri-
bution, L Xt−1 , Xt+2 , and the product of the marginal distributions, L Xt−1 ×

L Xt+2 . Thus, it is a natural measure of the degree to which the future of

the system depends on its past. However, there are at least four other mixing
coefficients (β, φ, ψ and ρ) regularly used in the literature. Since any of these
others going to zero implies that α goes to zero, we will stick with α-mixing, as
in Rosenblatt (1956).
Also notice that if Xt is a Markov process (e.g., a dynamical system) then
the Markov property tells us that we only need to let the supremum run over
measurable sets in σ(Xt1 ) and σ(Xt2 ).
Lemma 393 If a dynamical system is α-mixing, then it is mixing.
Proof: α is the supremum of the quantity appearing in the definition of mixing.

Notation: For the remainder of this section,
n
X
Sn ≡ Xn (27.21)
k=1
σn2 ≡ Var [Sn ] (27.22)
S[nt]
Yn (t) ≡ (27.23)
σn
2 Doukhan (1995, p. 47) cites Jakubowski and Szewczak (1990) as the source, but I have

not verified the reference.

CHAPTER 27. MIXING 226

where n is any positive integer, and t ∈ [0, 1].

Definition 394 Xt obeys the central limit theorem when

Sn d
√ → N (0, 1) (27.24)
σ n

for some positive σ.

Definition 395 Xt obeys the functional central limit theorem or the invariance
principle when
d
Yn → W (27.25)
where W is a standard Wiener process on [0, 1], and the convergence is in the
Skorokhod topology of Sec. 15.1.

Theorem 396 (Central Limit Theorem for α-Mixing Sequences) Let Xt

be a stationary sequence with E [Xt ] = 0. Suppose X is α-mixing, and that for
some δ > 0
h i
2+δ
E |Xt | ≤ ∞ (27.26)
∞
X δ
α 2+δ (n) ≤ ∞ (27.27)
n=0

Then
∞
σn2 h
2
i X
lim = E |X1 | + 2 E [X1 Xk ] ≡ σ 2 (27.28)
n→∞ n
k=1

If σ 2 > 0, moreover, Xt obeys both the central limit theorem with variance σ 2 ,
and the functional central limit theorem.

Proof: Complicated, and based on a rather technical central limit theorem for
martingale difference arrays. See Doukhan (1995, sec. 1.5), or, for a simplified
presentation, Durrett (1991, sec. 7.7).  √
For the rate of convergence of of L (Sn / n) to a Gaussian distribution, in
the total variation metric, see Doukhan (1995, sec. 1.5.2), summarizing sev-
eral works. Polynomially-mixing sequences converge polynomially in n, and
exponentially-mixing sequences converge exponentially.
There are a number of results on central limit theorems and functional cen-
tral limit theorems for deterministic dynamical systems. A particularly strong
one was recently proved by Tyran-Kamińska (2005), in a friendly paper which
should be accessible to anyone who’s followed along this far, but it’s too long
for us to do more than note its existence.
 Part VI

Information Theory

227
Chapter 28

Shannon Entropy and

Kullback-Leibler
Divergence

Section 28.1 introduces Shannon entropy and its most basic prop-
erties, including the way it measures how close a random variable is
to being uniformly distributed.
Section 28.2 describes relative entropy, or Kullback-Leibler di-
vergence, which measures the discrepancy between two probability
distributions, and from which Shannon entropy can be constructed.
Section 28.2.1 describes some statistical aspects of relative entropy,
especially its relationship to expected log-likelihood and to Fisher
information.
Section 28.3 introduces the idea of the mutual information shared
by two random variables, and shows how to use it as a measure of
serial dependence, like a nonlinear version of autocovariance (Section
28.3.1).

Information theory studies stochastic processes as sources of information,

or as models of communication channels. It appeared in essentially its modern
form with Shannon (1948), and rapidly proved to be an extremely useful mathe-
matical tool, not only for the study of “communication and control in the animal
and the machine” (Wiener, 1961), but more technically as a vital part of prob-
ability theory, with deep connections to statistical inference (Kullback, 1968),
to ergodic theory, and to large deviations theory. In an introduction that’s so
limited it’s almost a crime, we will do little more than build enough theory to
see how it can fit in with the theory of inference, and then get what we need
to progress to large deviations. If you want to learn more (and you should!),
the deservedly-standard modern textbook is Cover and Thomas (1991), and a
good treatment, at something more like our level of mathematical rigor, is Gray

228
CHAPTER 28. ENTROPY AND DIVERGENCE 229

(1990).1

28.1 Shannon Entropy

The most basic concept of information theory is that of the entropy of a random
variable, or its distribution, often called Shannon entropy to distinguish it from
the many other sorts. This is a measure of the uncertainty or variability asso-
ciated with the random variable. Let’s start with the discrete case, where the
variable takes on only a finite or countable number of values, and everything is
easier.

Definition 397 (Shannon Entropy (Discrete Case)) The Shannon entropy,

or just entropy, of a discrete random variable X is
X
H[X] ≡ − P (X = x) log P (X = x) = −E [log P (X)] (28.1)
x

when the sum exists. Entropy has units of bits when the logarithm has base 2,
and nats when it has base e.
The joint entropy of two random variables, H[X, Y ], is the entropy of their
joint distribution.
The conditional entropy of X given Y , H[X|Y ] is
X X
H[X|Y ] ≡ P (Y = y) P (X = x|Y = y) log P (X = x|Y = y)(28.2)
y x
= −E [log P (X|Y )] (28.3)
= H[X, Y ] − H[Y ] (28.4)

Here are some important properties of the Shannon entropy, presented with-
out proofs (which are not hard).
1. H[X] ≥ 0
2. H[X] = 0 iff ∃x0 : X = x0 a.s.
3. If X can take on n < ∞ different values (with positive probability), then
H[X] ≤ log n. H[X] = log n iff X is uniformly distributed.
4. H[X]+H[Y ] ≥ H[X, Y ], with equality iff X and Y are independent. (This
comes from the logarithm in the definition.)
1 Remarkably, almost all of the post-1948 development has been either amplifying or refining

themes first sounded by Shannon. For example, one of the fundamental results, which we
will see in the next chapter, is the “Shannon-McMillan-Breiman theorem”, or “asymptotic
equipartition property”, which says roughly that the log-likelihood per unit time of a random
sequence converges to a constant, characteristic of the data-generating process. Shannon’s
original version was convergence in probability for ergodic Markov chains; the modern form
is almost sure convergence for any stationary and ergodic process. Pessimistically, this says
something about the decadence of modern mathematical science; optimistically, something
about the value of getting it right the first time.
CHAPTER 28. ENTROPY AND DIVERGENCE 230

5. H[X, Y ] ≥ H[X].
6. H[X|Y ] ≥ 0, with equality iff X is a.s. constant given Y , for almost all
Y.
7. H[X|Y ] ≤ H[X], with equality iff X is independent of Y . (“Conditioning
reduces entropy”.)
8. H[f (X)] ≤ H[X], for any measurable function f , with equality iff f is
invertible.
The first three properties can be summarized by saying that H[X] is max-
imized by a uniform distribution, and minimized, to zero, by a degenerate one
which is a.s. constant. We can then think of H[X] as the variability of X,
something like the log of the effective number of values it can take on. We can
also think of it as how uncertain we are about X’s value.2 H[X, Y ] is then how
much variability or uncertainty is associated with the pair variable X, Y , and
H[Y |X] is how much uncertainty remains about Y once X is known, averag-
ing over Y . Similarly interpretations follow for the other properties. The fact
that H[f (X)] = H[X] if f is invertible is nice, because then f just relabels the
possible values, meshing nicely with this interpretation.
A simple consequence of the above results is particularly important for later
use.

Lemma 398 (Chain Rule for Shannon Entropy) Let X1 , X2 , . . . Xn be discrete-

valued random variables on a common probability space. Then
n
X
H[X1 , X2 , . . . Xn ] = H[X1 ] + H[Xn |X1 , . . . Xn−1 ] (28.5)
i=2

Proof: From the definitions, it is easily seen that H[X2 |X1 ] = H[X2 , X1 ] −
H[X1 ]. This establishes the chain rule for n = 2. A simple argument by
induction does the rest. 
For non-discrete random variables, it is necessary to introduce a reference
measure, and many of the nice properties go away.

Definition 399 (Shannon Entropy (General Case)) The Shannon entropy

of a random variable X with distribution µ, with respect to a reference measure
ρ, is  
dµ
Hρ [X] ≡ −Eµ log (28.6)
dρ
2 This line of reasoning is sometimes supplemented by saying that we are more “surprised”

to find that X = x the less probable that event is, supposing that surprise should go as the
log of one over that probability, and defining entropy as expected surprise. The choice of
the logarithm, rather than any other increasing function, is of course retroactive, though one
might cobble together some kind of psychophysical justification, since the perceived intensity
of a sensation often grows logarithmically with the physical magnitude of the stimulus. More
dubious, to my mind, is the idea that there is any surprise at all when a fair coin coming up
heads.
CHAPTER 28. ENTROPY AND DIVERGENCE 231

when µ << ρ. Joint and conditional entropies are defined similarly. We will
also write Hρ [µ], with the same meaning. This is sometimes called differential
entropy when ρ is Lebesgue measure on Euclidean space, especially R, and then
is written h(X) or h[X].

It remains true, in the general case, that Hρ [X|Y ] = Hρ [X, Y ] − Hρ [Y ], pro-

vided all of the entropies are finite. The chain rule remains valid, conditioning
still reduces entropy, and the joint entropy is still ≤ the sum of the marginal
entropies, with equality iff the variables are independent. However, depending
on the reference measure, Hρ [X] can be negative; e.g., if ρ is Lebesgue measure
and L (X) = δ(x), then Hρ [X] = −∞.

28.2 Relative Entropy or Kullback-Leibler Di-

vergence
Some of the difficulties associated with Shannon entropy, in the general case,
can be evaded by using relative entropy.

Definition 400 (Relative Entropy, Kullback-Leibler Divergence) Given

two probability distributions, ν << µ, the relative entropy of ν with respect to
µ, or the Kullback-Leibler divergence of ν from µ, is
 
dν
D(µkν) = −Eµ log (28.7)
dµ

If ν is not absolutely continuous with respect to µ, then D(µkν) = ∞.

Lemma 401 D(µkν) ≥ 0, with equality iff ν = µ almost everywhere (µ).

h i h i
dν dν
Proof: From Jensen’s inequality, Eµ log dµ ≤ log Eµ dµ = log 1 = 0. The
second part follows from the conditions for equality in Jensen’s inequality. 

Lemma 402 (Divergence and Total Variation) For any two distributions,
1 2
D(µkν) ≥ 2 ln 2 kµ − νk1 .

Proof: Algebra. See, e.g., Cover and Thomas (1991, Lemma 12.6.1, pp. 300–
301). 

Definition 403 The conditional relative entropy, D(µ(Y |X)kν(Y |X)) is

 
dν(Y |X)
D(µ(Y |X)kν(Y |X)) ≡ −Eµ log (28.8)
dµ(Y |X)

Lemma 404 (Chain Rule for Relative Entropy) D(µ(X, Y )kν(X, Y )) =

D(µ(X)kν(X)) + D(µ(Y |X)kν(Y |X))
CHAPTER 28. ENTROPY AND DIVERGENCE 232

Proof: Algebra. 
Shannon entropy can be constructed from the relative entropy.

Lemma 405 The Shannon entropy of a discrete-valued random variable X,

with distribution µ, is
H[X] = log n − D(µkυ) (28.9)
where n is the number of values X can take on (with positive probability), and
υ is the uniform distribution over those values.

Proof: Algebra. 
A similar result holds for the entropy of a variable which takes values in a
finite subset, of volume V , of a Euclidean space, i.e., Hλ [X] = log V − D(µkυ),
where λ is Lebesgue measure and υ is the uniform probability measure on the
range of X.

28.2.1 Statistical Aspects of Relative Entropy

From Lemma 402, “convergence in relative entropy”, D(µkνn ) → 0 as n → ∞,
implies convergence in the total variation (L1 ) metric. Because of Lemma 401,
we can say that KL divergence has some of the properties of a metric on the
space of probability distribution: it’s non-negative, with equality only when the
two distributions are equal (a.e.). Unfortunately, however, it is not symmetric,
and it does not obey the triangle inequality. (This is why it’s the KL divergence
rather than the KL distance.) Nonetheless, it’s enough like a metric that it can
be used to construct a kind of geometry on the space of probability distributions,
and so of statistical models, which can be extremely useful. While we will not
be able to go very far into this information geometry3 , it will be important to
indicate a few of the connections between information-theoretic notions, and
the more usual ones of statistical theory.

Definition 406 (Cross-entropy) The cross-entropy of ν and µ, Q(µkν), is

 
dν
Qρ (µkν) ≡ −Eµ log (28.10)
dρ
where ν is absolutely continuous with respect to the reference measure ρ. If the
domain is discrete, we will take the reference measure to be uniform and drop
the subscript, unless otherwise noted.

Lemma 407 Suppose ν and µ are the distributions of two probability models,
and ν << µ. Then the cross-entropy is the expected negative log-likelihood of
the model corresponding to ν, when the actual distribution is µ. The actual
or empirical negative log-likelihood of the model corresponding to ν is Qρ (νkη),
where η is the empirical distribution.
3 See Kass and Vos (1997) or Amari and Nagaoka (1993/2000). For applications to sta-

tistical inference for stochastic processes, see Taniguchi and Kakizawa (2000). For an easier
general introduction, Kulhavý (1996) is hard to beat.
CHAPTER 28. ENTROPY AND DIVERGENCE 233

Proof: Obvious from the definitions. 

Lemma 408 If ν << µ << ρ, then Qρ (µkν) = Hρ [µ] + D(µkν).
Proof: By the chain rule for densities,
dν dµ dν
= (28.11)
dρ dρ dµ
dν dµ dν
log = log + log (28.12)
dρ dρ dµ
     
dν dµ dν
Eµ log = Eµ log + Eµ log (28.13)
dρ dρ dµ
The result follows by applying the definitions. 
Corollary 409 (Gibbs’s Inequality) Qρ (µkν) ≥ Hρ [µ], with equality iff ν =
µ a.e.
Proof: Insert the result of Lemma 401 into the preceding proposition. 
The statistical interpretation of the proposition is this: The log-likelihood
of a model, leading to distribution ν, can be broken into two parts. One is the
divergence of ν from µ; the other just the entropy of µ, i.e., it is the same for all
models. If we are considering the expected log-likelihood, then µ is the actual
data-generating distribution. If we are considering the empirical log-likelihood,
then µ is the empirical distribution. In either case, to maximize the likelihood
is to minimize the relative entropy, or divergence. What we would like to do, as
statisticians, is minimize the divergence from the data-generating distribution,
since that will let us predict future values. What we can do is minimize diver-
gence from the empirical distribution. The consistency of maximum likelihood
methods comes down, then, to finding conditions under which a shrinking di-
vergence from the empirical distribution guarantees a shrinking divergence from
the true distribution.4
Definition 410 Let θ ∈ Rk , k < ∞, be the parameter indexing a set M of
statistical models, where for every θ, νθ << ρ, with densities pθ . Then the
Fisher information matrix is
  
∂ log pθ ∂ log pθ
Iij (θ) ≡ Eνθ (28.14)
dθi dθj

Corollary 411 The Fisher information matrix is equal to the Hessian (second
partial derivative) matrix of the relative entropy:
∂2
Iij (θ0 ) = D(νθ0 kνθ ) (28.15)
∂θi ∂θj
4 If we did have a triangle inequality, then we could say D(µkν) ≤ D(µkη) + D(ηkν), and

it would be enough to make sure that both the terms on the RHS went to zero, say by some
combination of maximizing the likelihood in-sample, so D(ηkν) is small, and ergodicity, so
that D(µkη) is small. While, as noted, there is no triangle inequality, under some conditions
this idea is roughly right; there are nice diagrams in Kulhavý (1996).
CHAPTER 28. ENTROPY AND DIVERGENCE 234

Proof: It is a classical
h 2result (see,
i e.g., Lehmann and Casella (1998, sec. 2.6.1))
∂
that Iij (θ) = −Eνθ ∂θi ∂θj log pθ . The present result follows from this, Lemma
407, Lemma 408, and the fact that Hρ [νθ0 ] is independent of θ. 

28.3 Mutual Information

Definition 412 (Mutual Information) The mutual information between two
random variables, X and Y , is the divergence of the product of their marginal
distributions from their actual joint distribution:
I[X; Y ] ≡ D(L (X, Y ) kL (X) × L (Y )) (28.16)
Similarly, the mutual information among n random variables X1 , X2 , . . . Xn is
n
Y
I[X1 ; X2 ; . . . ; Xn ] ≡ D(L (X1 , X2 , . . . Xn ) k L (Xi )) (28.17)
i=1

the divergence of the product distribution from the joint distribution.

Proposition 413 I[X; Y ] ≥ 0, with equality iff X and Y are independent.

Proof: Directly from Lemma 401. 
Proposition 414 If all the entropies involved are finite,
I[X; Y ] = H[X] + H[Y ] − H[X, Y ] (28.18)
= H[X] − H[X|Y ] (28.19)
= H[Y ] − H[Y |X] (28.20)
so I[X; Y ] ≤ H[X] ∧ H[Y ].
Proof: Calculation. 
This leads to the interpretation of the mutual information as the reduction
in uncertainty or effective variability of X when Y is known, averaging over their
joint distribution. Notice that in the discrete case, we can say H[X] = I[X; X],
which is why H[X] is sometimes known as the self-information.

28.3.1 Mutual Information Function

Just as with the autocovariance function, we can define a mutual information
function for one-parameter processes, to serve as a measure of serial dependence.
Definition 415 (Mutual Information Function) The mutual information
function of a one-parameter stochastic process X is
ι(t1 , t2 ) ≡ I[Xt1 ; Xt2 ] (28.21)
which is symmetric in its arguments. If the process is stationary, it is a function
of |t1 − t2 | alone.
CHAPTER 28. ENTROPY AND DIVERGENCE 235

Notice that, unlike the autocovariance function, ι includes nonlinear depen-

dencies between Xt1 and Xt2 . Also notice that ι(τ ) = 0 means that the two
variables are strictly independent, not just uncorrelated.

Theorem 416 A stationary process is mixing if ι(τ ) → 0.

Proof: Because then the total variation distance between the joint distribution,
L (Xt1 Xt2 ), and the product of the marginal distributions, L (Xt1 ) L (Xt2 ), is
being forced down towards zero, which implies mixing (Definition 379). 
Chapter 29

Entropy Rates and

Asymptotic Equipartition

Section 29.1 introduces the entropy rate — the asymptotic en-

tropy per time-step of a stochastic process — and shows that it is
well-defined; and similarly for information, divergence, etc. rates.
Section 29.2 proves the Shannon-MacMillan-Breiman theorem,
a.k.a. the asymptotic equipartition property, a.k.a. the entropy
ergodic theorem: asymptotically, almost all sample paths of a sta-
tionary ergodic process have the same log-probability per time-step,
namely the entropy rate. This leads to the idea of “typical” se-
quences, in Section 29.2.1.
Section 29.3 discusses some aspects of asymptotic likelihood, us-
ing the asymptotic equipartition property, and allied results for the
divergence rate.

29.1 Information-Theoretic Rates

Definition 417 (Entropy Rate) The entropy rate of a random sequence X
is
h(X) ≡ lim Hρ [X1n ]n (29.1)
n

when the limit exists.

Definition 418 (Limiting Conditional Entropy) The limiting conditional

entropy of a random sequence X is

h0 (X) ≡ lim Hρ [Xn |X1n−1 ] (29.2)

n

when the limit exists.

236
CHAPTER 29. RATES AND EQUIPARTITION 237

Lemma 419 For a stationary sequence, Hρ [Xn |X1n−1 ] is non-increasing in n.

Moreover, its limit exists if X takes values in a discrete space.

Proof: Because “conditioning reduces entropy”, Hρ [Xn+1 |X1n ] ≤ H[Xn+1 |X2n ].

By stationarity, Hρ [Xn+1 |X2n ] = Hρ [Xn |X1n−1 ]. If X takes discrete values,
then conditional entropy is non-negative, and a non-increasing sequence of non-
negative real numbers always has a limit. 
Remark: Discrete values are a sufficient condition for the existence of the
limit, not a necessary one.
We now need a natural-looking, but slightly technical, result from real anal-
ysis.

Theorem 420 (Cesàro)

Pn For any sequence of real numbers an → a, the se-
quence bn = n−1 i=1 an also converges to a.

Proof: For every  > 0, there is an N () such that |an − a| <  whenever
n > N (). Now take bn and break it up into two parts, one summing the terms
below N (), and the other the terms above.
 n

 
 −1 X
lim |bn − a| = lim n ai − a (29.3)

n n  
i=1
n
X
≤ lim n−1 |ai − a| (29.4)
n
i=1
 
N ()
X
≤ lim n−1  |ai − a| + (n − N ()) (29.5)
n
i=1
 
N ()
X
≤ lim n−1  |ai − a| + n (29.6)
n
i=1
N ()
X
=  + lim n−1 |ai − a| (29.7)
n
i=1
=  (29.8)

Since  was arbitrary, lim bn = a. 

Theorem 421 (Entropy Rate) For a stationary sequence, if the limiting con-
ditional entropy exists, then it is equal to the entropy rate, h(X) = h0 (X).

Proof: Start with the chain rule to break the joint entropy into a sum of
conditional entropies, use Lemma 419 to identify their limit as h]prime (X), and
CHAPTER 29. RATES AND EQUIPARTITION 238

then use Cesàro’s theorem:

1
h(X) = lim Hρ [X1n ] (29.9)
n n
n
1X
= lim Hρ [Xi |X1i−1 ] (29.10)
n n
i=1
= h0 (X) (29.11)

as required. 
Because h(X) = h0 (X) for stationary processes (when both limits exist), it is
not uncommon to find what I’ve called the limiting conditional entropy referred
to as the entropy rate.

Lemma 422 For a stationary sequence h(X) ≤ H[X1 ], with equality iff the
sequence is IID.

Proof: Conditioning reduces entropy, unless the variables are independent, so

H[Xn |X1n−1 ] < H[Xn ], unless Xn X1n−1 . For this to be true of all n, which
|=

is what’s needed for h(X) = H[X1 ], all the values of the sequence must be
independent of each other; since the sequence is stationary, this would imply
that it’s IID. 

Example 423 (Markov Sequences) If X is a stationary Markov sequence,

then
Pn h(X) = t−1 Hρ [X2 |X1 ], because, by the chain rule, Hρ [X1n ] = Hρ [X1 ] +
t=2 Hρ [Xt |X1 ]. By the Markov property, however, Hρ [Xt |X1t−1 ] = Hρ [Xt |Xt−1 ],
which by stationarity is Hρ [X2 |X1 ]. Thus, Hρ [X1n ] = Hρ [X1 ]+(n−1)Hρ [X2 |X1 ].
Dividing by n and taking the limit, we get Hρ [X1n ] = Hρ [X2 |X1 ].

Example 424 (Higher-Order Markov Sequences) If X is a k th order Markov

sequence, then the same reasoning as before shows that h(X) = Hρ [Xk+1 |X1k ]
when X is stationary.

Definition 425 (Divergence Rate) The divergence rate or relative entropy

rate of the infinite-dimensional distribution Q from the infinite-dimensional dis-
tribution P , d(P kQ), is
"  !#
dP 
d(P kQ) = lim EP log (29.12)
n dQ σ(X 0 )
−n

if all the finite-dimensional distributions of Q dominate all the finite-dimensional

distributions of P . If P and Q have densities, respectively p and q, with respect
to a common reference measure, then
" #
−1
p(X0 |X−n )
d(P kQ) = lim EP log −1 (29.13)
n q(X0 |X−n )
CHAPTER 29. RATES AND EQUIPARTITION 239

29.2 The Shannon-McMillan-Breiman Theorem

or Asymptotic Equipartition Property
This is a central result in information theory, acting as a kind of ergodic theorem
for the entropy. That is, we want to say that, for almost all ω,
1 1
− log P (X1n (ω)) → lim E [− log P (X1n )] = h(X)
n n n

At first, it looks like we should be able to make a nice time-averaging argument.

We can always factor the joint probability,
n
1 1X
log P (X1n ) = log P Xt |X1t−1


n n t=1


with the understanding that P X1 |X10 = P (X1 ). This looks rather like the
sort of Cesàro average that we became familiar with in ergodic theory. The
problem is, there we were averaging f (T t ω) for a fixed function f . This is not
the case here, because we are conditioning on long and longer stretches of the
past. There’s no problem if the sequence is Markovian, because then the remote
past is irrelevant, by the Markov property, and we can just condition on a fixed-
length stretch of the past, so we’re averaging a fixed function shifted in time.
(This is why Shannon’s original argument was for Markov chains.) The result
nonetheless more broadly, but requires more subtlety than might otherwise be
thought. Breiman’s original proof of the general case was fairly involved1 , re-
quiring both martingale theory, and a sort of dominated convergence theorem
for ergodic time averages. (You can find a simplified version of his argument
in Kallenberg, at the end of chapter 11.) We will go over the “sandwiching”
argument of Algoet and Cover (1988), which is, to my mind, more transparent.
The idea of the sandwich argument is to show that, for large n, −n−1 log P (X1n )
must lie between an upper bound, hk , obtained by approximating the sequence
by a Markov process of order k, and a lower bound, which will be shown to be
h. Once we establish that hk ↓ h, we will be done.

Definition 426 (Markov Approximation) For each k, define the order k

Markov approximation to X by
n

Y t−1
µk (X1n ) X1k


=P P Xt |Xt−k (29.14)
t=k+1

µk is the distribution of a stationary Markov process of order k, where the

distribution of X1k+1 matches that of the original process.
1 Notoriously, the proof in his original paper was actually invalid, forcing him to publish a

correction.
CHAPTER 29. RATES AND EQUIPARTITION 240

Lemma 427 For each k, the entropy rate of the order k Markov approximation
is is equal to H[Xk+1 |X1k ].
Proof: Under the approximation (but not under the original distribution of X),
H[X1n ] = H[X1k ]+(n−k)H[Xk+1 |X1k ], by the Markov property and stationarity
(as in Examples 423 and 424). Dividing by n and taking the limit as n → ∞
gives the result. 
t
Lemma 428 If X is a stationary two-sided sequence, then Yt = f (X−∞ ) de-
fines a stationary sequence, for any measurable f . If X is also ergodic, then Y
is ergodic too.
Proof: Because X is stationary, it can be represented as a measure-preserving
t d t
shift on sequence space. Because it is measure-preserving, θX−∞ = X−∞ , so
d
Y (t) = Y (t + 1), and similarly for all finite-length blocks of Y . Thus, all of the
finite-dimensional distributions of Y are shift-invariant, and these determine the
infinite-dimensional distribution, so Y itself must be stationary.
To see that Y must be ergodic if X is ergodic, recall that a random sequence
is ergodic iff its corresponding shift dynamical system is ergodic. A dynamical
system is ergodic iff all invariant functions are a.e. constant (Theorem 350).
Because the Y sequence is obtained by applying a measurable function to the
X sequence, a shift-invariant function of the Y sequence is a shift-invariant
function of the X sequence. Since the latter are all constant a.e., the former are
too, and Y is ergodic. 
Lemma 429 If X is stationary and ergodic, then, for every k,
 
1 n
P lim − log µk (X1 (ω)) = hk = 1 (29.15)
n n
i.e., − n1 log µk (X1n (ω)) converges a.s. to hk .
Proof: Start by factoring the approximating Markov measure in the way sug-
gested by its definition:
n
1 n 1 k

1 X t−1


− log µk (X1 ) = − log P X1 − log P Xt |Xt−k (29.16)
n n n
t=k+1

t−1
As n grows, n1 log P X1k → 0, for every fixed k. On the other hand, − log P Xt |Xt−k

is a measurable function of the past of the process, and since X is stationary

and ergodic, it, too, is stationary and ergodic (Lemma 428). So
n
1 1 X t−1
log µk (X1n )


− → − log P Xt |Xt−k (29.17)
n n
t=k+1
a.s.
E − log P Xk+1 |X1k


→ (29.18)
= hk (29.19)
by Theorem 351. 
CHAPTER 29. RATES AND EQUIPARTITION 241

Definition 430 The infinite-order approximation to the entropy rate of a discrete-

valued stationary process X is
−1


h∞ (X) ≡ E − log P X0 |X−∞ (29.20)

Lemma 431 If X is stationary and ergodic, then

1
lim − log P X1n |X−∞0


= h∞ (29.21)
n n
almost surely.
Proof: Via Theorem 351 again, as in Lemma 429. 
Lemma 432 For a stationary, ergodic, finite-valued random sequence, hk (X) ↓
h∞ (X).
Proof: By the martingale convergence theorem, for every x0 ∈ Ξ,

a.s.
P X0 = x0 |Xn−1 → P X0 = x0 |X∞ −1


(29.22)
Since Ξ is finite, the probability of any point in Ξ is between 0 and 1 inclusive,
and p log p is bounded and continuous. So we can apply bounded convergence
to get that
" #
X
−1 −1



hk = E − P X0 = x0 |X−k log P X0 = x0 |X−k (29.23)
x0
" #
X
−1 −1



→ E − P X0 = x0 |X−∞ log P X0 = x0 |X−∞ (29.24)
x0
= h∞ (29.25)
Lemma 433 h∞ (X) is the entropy rate of X, i.e. h∞ (X) = h(X).
Proof: Clear from Theorem 421 and the definition of conditional entropy. 
We are almost ready for the proof, but need one technical lemma first.
Lemma 434 If Rn ≥ 0, E [Rn ] ≤ 1 for all n, then
1
lim sup log Rn ≤ 0 (29.26)
n n
almost surely.
Proof: Pick any  > 0.
 
1
P log Rn ≥  = P (Rn ≥ en ) (29.27)
n
E [Rn ]
≤ (29.28)
en
−n
≤ e (29.29)
−n
P
by Markov’s inequality. Since ne ≤ ∞, by the Borel-Cantelli lemma,
lim supn n−1 log Rn ≤ . Since  was arbitrary, this concludes the proof. 
CHAPTER 29. RATES AND EQUIPARTITION 242

Theorem 435 (Asymptotic Equipartition Property) For a stationary, er-

godic, finite-valued random sequence X,
1
− log P (X1n ) → h(X) a.s. (29.30)
n
Proof: For every k, µk (X1n )/P (X1n ) ≥ 0, and E [µk (X1n )/P (X1n )] ≤ 1. Hence,
by Lemma 434,
1 µk (X1n )
lim sup log ≤0 (29.31)
n n P (X1n )
a.s. Manipulating the logarithm,
1 1
lim sup log µk (X1n ) ≤ − lim sup − log P (X1n ) (29.32)
n n n n
1 1
From Lemma 429, lim supn n log µk (X1n ) = limn n log µk (X1n ) = −hk (X), a.s.
Hence, for each k,
1
hk (X) ≥ lim sup − log P (X1n ) (29.33)
n n
almost surely.

A similar manipulation of P (X1n ) /P X1n |X−∞
0
gives

1
h∞ (X) ≤ lim inf − log P (X1n ) (29.34)
n n
a.s.
As hk ↓ h∞ , it follows that the liminf and the limsup of the normalized log
likelihood must be equal almost surely, and so equal to h∞ , which is to say to
h(X). 
Why is this called the AEP? Because, to within an o(n) term, all sequences
of length n have the same log-likelihood (to within factors of o(n), if they have
positive probability at all. In this sense, the likelihood is “equally partitioned”
over those sequences.

29.2.1 Typical Sequences

Let’s turn the result of the AEP around. For large n, the probability of a
given sequence is either approximately 2−nh or approximately zero2 . To get
the total probability to sum up to one, there need to be about 2nh sequences
with positive probability. If the size of the alphabet is s, then the fraction
of sequences which are actually exhibited is 2n(h−log s) , an increasingly small
fraction (as h ≤ log s). Roughly speaking, these are the typical sequences, any
one of which, via ergodicity, can act as a representative of the complete process.
2 Of course that assumes using base-2 logarithms in the definition of entropy.
CHAPTER 29. RATES AND EQUIPARTITION 243

29.3 Asymptotic Likelihood

29.3.1 Asymptotic Equipartition for Divergence
Using methods analogous to those we employed on the AEP for entropy, it is
possible to prove the following.

Theorem 436 Let P be an asymptotically mean-stationary distribution, with

stationary mean P , with ergodic component function φ. Let M be a homoge-
neous finite-order Markov process, whose finite-dimensional distributions dom-
inate those of P and P ; denote the densities with respect to M by p and p,
respectively. If limn n−1 log p(X1n ) is an invariant function P -a.e., then
1 a.s.
− log p(X1n (ω)) → d(P φ(ω) kM ) (29.35)
n
where P φ(ω) is the stationary, ergodic distribution of the ergodic component.

Proof: See Algoet and Cover (1988, theorem 4), Gray (1990, corollary 8.4.1).
Remark. The usual AEP is in fact a consequence of this result, with the
appropriate reference measure. (Which?)

29.3.2 Likelihood Results

It is left as an exercise for you to obtain the following result, from the AEP for
relative entropy, Lemma 408 and the chain rules.

Theorem 437 Let P be a stationary and ergodic data-generating process, whose

entropy rate, with respect to some reference measure ρ, is h. Further let M be a
finite-order Markov process which dominates P , whose density, with respect to
the reference measure, is m. Then
1
− log m(X1n ) → h + d(P kM ) (29.36)
n
P -almost surely.

29.4 Exercises
Exercise 29.1 Markov approximations are maximum-entropy approximations.
(You may assume that the process X takes values in a finite set.)
a Prove that µk , as defined in Definition 426, gets the distribution of se-
quences of length k + 1 correct, i.e., for any set A ∈ X k+1 , ν(A) =
k+1


P X1 ∈ A .
b Prove that µk0 , for any any k 0 > k, also gets the distribution of length
k + 1 sequences right.
CHAPTER 29. RATES AND EQUIPARTITION 244

c In a slight abuse of notation, let H[ν(X1n )] stand for the entropy of a se-
quence of length n when distributed according to ν. Show that H[µk (X1n )] ≥
H[µk0 (X1n )] if k 0 > k. (Note that the n ≤ k case is easy!)
d Is it true that that if ν is any other measure which gets the distribution
of sequences of length k + 1 right, then H[µk (X1n )] ≥ H[ν(X1n )]? If yes,
prove it; if not, find a counter-example.

Exercise 29.2 Prove Theorem 437.

 Part VII

Large Deviations

245
Chapter 30

General Theory of Large

Deviations

A family of random variables follows the large deviations princi-

ple if the probability of the variables falling into “bad” sets, repre-
senting large deviations from expectations, declines exponentially in
some appropriate limit. Section 30.1 makes this precise, using some
associated technical machinery, and explores a few consequences.
The central one is Varadhan’s Lemma, for the asymptotic evalua-
tion of exponential integrals in infinite-dimensional spaces.
Having found one family of random variables which satisfy the
large deviations principle, many other, related families do too. Sec-
tion 30.2 lays out some ways in which this can happen.

As the great forensic statistician C. Chan once remarked, “Improbable events

permit themselves the luxury of occurring” (reported in Biggers, 1928). Large
deviations theory, as I have said, studies these little luxuries.

30.1 Large Deviation Principles: Main Defini-

tions and Generalities
Some technicalities:

Definition 438 (Level Sets) For any real-valued function f : Ξ 7→ R, the

level sets are the inverse images of intervals from −∞ to c inclusive, i.e., all
sets of the form {x ∈ Ξ : f (x) ≤ c}.

Definition 439 (Lower Semi-Continuity) A real-valued function f : Ξ 7→

R is lower semi-continuous if xn → x implies lim inf f (xn ) ≥ f (x).

246
CHAPTER 30. LARGE DEVIATIONS: BASICS 247

Lemma 440 A function is lower semi-continuous iff either of the following

equivalent properties hold.
i For all x ∈ Ξ, the infimum of f over increasingly small open balls centered
at x approaches f (x):

lim inf f (y) = f (x) (30.1)

δ→0 y: d(y,x)<δ

ii f has closed level sets.

Proof: A character-building exercise in real analysis, left to the reader. 

Lemma 441 A lower semi-continuous function attains its minimum on every

non-empty compact set, i.e., if C is compact and 6= ∅, there is an x ∈ C such
that f (x) = inf y∈C f (y).

Proof: Another character-building exercise in real analysis. 

Definition 442 (Logarithmic Equivalence) Two sequences of positive real

numbers an and bn are logarithmically equivalent, an ' bn , when
1
lim (log an − log bn ) = 0 (30.2)
n→∞ n
Similarly, for continuous parameterizations by  > 0, a ' b when

lim  (log a − log b ) = 0 (30.3)

→0

Lemma 443 (“Fastest rate wins”) For any two sequences of positive num-
bers, (an + bn ) ' an ∨ bn .

Proof: A character-building exercise in elementary analysis. 

Definition 444 (Large Deviation Principle) A parameterized family of ran-

dom variables, X ,  > 0, taking values in a metric space Ξ with Borel σ-field
X , obeys a large deviation principle with rate 1/, or just obeys an LDP, when,
for any set B ∈ X ,

− inf J(x) ≤ lim inf  log P (X ∈ B) ≤ lim sup  log P (X ∈ B) ≤ − inf J(x)
x∈intB →0 →0 x∈clB
(30.4)
for some non-negative function J : Ξ 7→ [0, ∞], its raw rate function. If J is
lower semi-continuous, it is just a rate function. If J is lower semi-continuous
and has compact level sets, it is a good rate function.1 By a slight abuse of
notation, we will write J(B) = inf x∈B J(x).
1 Sometimes what Kallenberg and I are calling a “good rate function” is just “a rate func-

tion”, and our “rate function” gets demoted to “weak rate function”.
CHAPTER 30. LARGE DEVIATIONS: BASICS 248

Remark: The most common choices of  are 1/n, in sample-size or discrete

sequence problems, or ε2 , in small-noise problems (as in Chapter 21).
Lemma 445 (Uniqueness of Rate Functions) If X obeys the LDP with
raw rate function J, then it obeys the LDP with a unique rate function J 0 .
Proof: First, show that a raw rate function can always be replaced by a lower
semi-continuous function, i.e. a non-raw (cooked?) rate function. Then, show
that non-raw rate functions are unique.
For any raw rate function J, define J 0 (x) = lim inf y→x J(x). This is clearly
lower semi-continuous, and J 0 (x) ≤ J(x). However, for any open set B, inf x∈B J 0 (x) =
inf x∈B J(x), so J and J 0 are equivalent for purposes of the LDP.
Now assume that J is a lower semi-continuous rate function, and suppose
that K 6= J was too; without loss of generality, assume that J(x) > K(x) at
some point x. We can use semi-continuity to find an open neighborhood B
of x such that J(clB) > K(x). But, substituting into Eq. 30.4, we obtain a
contradiction:
− K(x) ≤ −K(B) (30.5)
≤ lim inf  log P (X ∈ B) (30.6)
→0
≤ −J(clB) (30.7)
≤ −K(x) (30.8)
Hence there can be no such rate function K, and J is the unique rate function.

Lemma 446 If X obeys an LDP with rate function J, then J(x) = 0 for some
x.
Proof: Because P (X ∈ Ξ) = 1, we must have J(Ξ) = 0, and rate functions
attain their infima. 
Definition 447 A Borel set B is J-continuous, for some rate function J, when
J(intB) = J(clB).

Lemma 448 If X satisfies the LDP with rate function J, then for every J-
continuous set B,
lim  log P (X ∈ B) = −J(B) (30.9)
→0

Proof: By J-continuity, the right and left hand extremes of Eq. 30.4 are equal,
so the limsup and the liminf sandwiched between them are equal; consequently
the limit exists. 
Remark: The obvious implication is that, for small , P (X ∈ B) ≈ ce−J(B)/ ,
which explains why we say that the LDP has rate 1/. (Actually, c need not be
constant, but it must be at least o(), i.e., it must go to zero faster than  itself
does.)
There are several equivalent ways of defining the large deviation principle.
The following is especially important, because it is often simplifies proofs.
CHAPTER 30. LARGE DEVIATIONS: BASICS 249

Lemma 449 X obeys the LDP with rate 1/ and rate function J(x) if and
only if

lim sup  log P (X ∈ C) ≤ −J(C) (30.10)

→0
lim inf  log P (X ∈ O) ≥ −J(O) (30.11)
→0

for every closed Borel set C and every open Borel set O ⊂ Ξ.

Proof: “If”: The closure of any set is closed, and the interior of any set is
open, so Eqs. 30.10 and 30.11 imply

lim sup  log P (X ∈ clB) ≤ −J(clB) (30.12)

→0
lim inf  log P (X ∈ intB) ≥ −J(intB) (30.13)
→0

but P (X ∈ B) ≤ P (X ∈ clB) and P (X ∈ B) ≥ P (X ∈ intB), so the LDP
holds. “Only if”: every closed set is equal to its own closure, and every open set
is equal to its own interior, so the upper bound in Eq. 30.4 implies Eq. 30.10,
and the lower bound Eq. 30.11. 
A deeply important consequence of the LDP is the following, which can be
thought of as a version of Laplace’s method for infinite-dimensional spaces.

Theorem 450 (Varadhan’s Lemma) If X are random variables in a metric

space Ξ, obeying an LDP with rate 1/ and rate function J, and f : Ξ 7→ R is
continuous and bounded from above, then
h i
Λf ≡ lim  log E ef (X )/ = sup f (x) − J(x) (30.14)
→0 x∈Ξ

Proof: We’ll find the limsup and the liminf, and show that they are both
sup f (x) − J(x).
First the limsup. Pick an arbitrary positive integer n. Because f is contin-
uous and bounded above, there exist finitelyS closed sets, call them B1 , . . . Bm ,
such that f ≤ −n on the complement of i Bi , and within each Bi , f varies by
at most 1/n. Now
h i
lim sup  log E ef (X )/ (30.15)
h i
≤ (−n) ∨ max lim sup  log E ef (X )/ 1Bi (X )
i≤m
≤ (−n) ∨ max sup f (x) − inf J(x) (30.16)
i≤m x∈Bi x∈Bi

≤ (−n) ∨ max sup f (x) − J(x) + 1/n (30.17)

i≤m x∈Bi

≤ (−n) ∨ sup f (x) − J(x) + 1/n (30.18)

x∈Ξ
CHAPTER 30. LARGE DEVIATIONS: BASICS 250
 
Letting n → ∞, we get lim sup  log E ef (X )/ = sup f (x) − J(x).
To get the liminf, pick any x ∈ Xi and an arbitrary ball of radius δ around
it, Bδ,x . We have
h i h i
lim inf  log E ef (X )/ ≥ lim inf  log E ef (X )/ 1Bδ,x (X ) (30.19)
≥ inf f (y) − inf J(y) (30.20)
y∈Bδ,x y∈Bδ,x

≥ inf f (y) − J(x) (30.21)

y∈Bδ,x

Since δ was arbitrary, we can let it go to zero, so (by continuity of f ) inf y∈Bδ,x f (y) →
f (x), or h i
lim inf  log E ef (X )/ ≥ f (x) − J(x) (30.22)

Since this holds for arbitrary x, we can replace the right-hand side by a supre-
mum over all x. Hence sup f (x) − J(x) is both the liminf and the limsup.   
Remark: The implication of Varadhan’s lemma is that, for small , E ef (X )/ ≈
−1
c()e (supx∈Ξ f (x)−J(x)) , where c() = o(). So, we can replace the exponential
integral with its value at the extremal points, at least to within a multiplicative
factor and to first order in the exponent.
An important, if heuristic, consequence of the LDP is that “Highly im-
probable events tend to happen in the least improbable way”. Let us con-
sider two events B ⊂ A, and suppose that P (X ∈ A) > 0 for all . Then
P (X ∈ B|X ∈ A) = P (X ∈ B) /P (X ∈ A). Roughly speaking, then, this
conditional probability will vanish exponentially, with rate J(A) − J(B). That
is, even if we are looking at an exponentially-unlikely large deviation, the vast
majority of the probability is concentrated around the least unlikely part of the
event. More formal statements of this idea are sometimes known as “conditional
limit theorems” or “the Gibbs conditioning principle”.

30.2 Breeding Large Deviations

Often, the easiest way to prove that one family of random variables obeys a
large deviations principle is to prove that another, related family does.

Theorem 451 (Contraction Principle) If X , taking values in a metric space

Ξ, obeys an LDP, with rate  and rate function J, and f : Ξ 7→ Υ is a continu-
ous function from that metric space to another, then Y = f (X ) also obeys an
LDP, with rate  and raw rate function K(y) = J(f −1 (y)). If J is a good rate
function, then so is K.
CHAPTER 30. LARGE DEVIATIONS: BASICS 251

Proof: Since f is continuous, f −1 takes open sets to open sets, and closed sets
to closed sets. Pick any closed C ⊂ Υ. Then
lim sup  log P (f (X ) ∈ C) (30.23)
→0
lim sup  log P X ∈ f −1 (C)


=
→0
≤ −J(f −1 (C)) (30.24)
= − inf−1
J(x) (30.25)
x∈f (C)

= − inf inf J(x) (30.26)

y∈C x∈f −1 (y)

= − inf K(y) (30.27)

y∈C

as required. The argument for open sets in Υ is entirely parallel, establishing

that K, as defined, is a raw rate function. By Lemma 445, K can be modified
to be lower semi-continuous without affecting the LDP, i.e., we can make a rate
function from it. If J is a good rate function, then it has compact level sets.
But continuous functions take compact sets to compact sets, so K = J ◦ f −1
will also have compact level sets, i.e., it will also be a good rate function. 
There are a bunch of really common applications of the contraction principle,
relating the large deviations at one level of description to those at coarser levels.
To make the most frequent set of implications precise, let’s recall a couple of
definitions.
Definition 452 (Empirical Mean) If X1 , . . . Xn are random
Pn variables in a
common vector space Ξ, their empirical mean is X n ≡ n1 i=1 Xi .
We have already encountered this as the sample average or, in ergodic theory,
the finite time average. (Notice that nothing is said about the Xi being IID, or
even having a common expectation.)
Definition 453 (Empirical Distribution) Let X1 , . . . Xn be random variables
in a common measurable space Ξ (notPn necessarily a vector or metric space). The
empirical distribution is P̂n ≡ n1 i=1 δXi , where δx is the probability measure
that puts all its probability on the point x, i.e., δx (B) = 1B (x). P̂n is a ran-
dom variable taking values in P (Ξ), the space of all probability measures on Ξ.
(Cf. Example 10 in chapter 1 and Example 43 in chapter 4.) P (Ξ) is a met-
ric space under any of several distances, and a complete separable metric space
(i.e., Polish) under, for instance, the total variation metric.

Definition 454 (Finite-Dimensional Empirical Distributions) For each

k, the k-dimensional empirical distribution is
n
1X
P̂nk ≡ δ(Xi ,Xi+1 ,...Xi+k ) (30.28)
n i=1

where the addition of indices for the delta
function is to be done modulo n, i.e.,
P̂32 = 13 δ(X1 ,X2 ) + δ(X2 ,X3 ) + δ(X3 ,X1 ) . P̂nk takes values in P Ξk .


CHAPTER 30. LARGE DEVIATIONS: BASICS 252

Definition 455 (Empirical Process Distribution) With a finite sequence

of random variables X1 , . . . Xn , the empirical process is the periodic, infinite
random sequence X̃n as the repetition of the sample without limit, i.e., X̃n (i) =
Xi mod n . If T is the shift operator on the sequence space, then the empirical
process distribution is
n−1
1X
P̂n∞ ≡ δ i (30.29)
n i−0 T X̃n

P̂n∞ takes values in

the space of infinite-dimensional distributions for one-sided
sequences, P ΞN . (In fact, it is always a stationary distribution, because by
construction it is invariant under the shift T .)

Be careful not to confuse this “empirical process” with the quite distinct
“empirical process” of Example 43.

Corollary 456 The following chain of implications hold:

i If the empirical process distribution obeys an LDP, so do all the finite-
dimensional distributions.
ii If the n-dimensional distribution obeys an LDP, all m < n dimensional
distributions do.
iii If any finite-dimensional distribution obeys an LDP, the empirical distri-
bution does.
iv If the empirical distribution obeys an LDP, the empirical mean does.

Proof: In each case, we obtain the lower-level statistic from the higher-level
one by applying a continuous function, hence the contraction principle applies.
For the distributions, the continuous functions are the projection operators of
Chapter 2. 

Corollary 457 (“Tilted” LDP) In set-up of Theorem 450, let µ = L (X ).

Define the probability measures µf, via
 
E ef (X )/ 1B (X )
µf, (B) ≡   (30.30)
E ef (X )/

Then Y ∼ µf, obeys an LDP with rate 1/ and rate function

JF (x) = −(f (x) − J(x)) + sup f (y) − J(y) (30.31)

y∈Ξ
CHAPTER 30. LARGE DEVIATIONS: BASICS 253
 
Proof: Define a set function F (B) = E ef (X )/ 1B (X ) ; then µf, (B) =
F (B)/F (Ξ). From Varadhan’s Lemma, we know that F (Ξ) has asymptotic
logarithm supy∈Ξ f (y) − J(y), so it is just necessary to show that

lim sup  log F (B) ≤ sup f (x) − J(x) (30.32)

 x∈clB
lim inf  log F (B) ≥ sup f (x) − J(x) (30.33)
 x∈intB

which can be done by imitating the proof of Varadhan’s Lemma itself. 

Remark: “Tilting” here refers to some geometrical analogy which, in all
honesty, has never made any sense to me.
Because the LDP is about exponential decay of probabilities, it is not sur-
prising that several ways of obtaining it require a sort of exponential bound on
the dispersion of the probability measure.

Definition 458 (Exponentially Tight) The parameterized family of random

variables X ,  > 0, is exponentially tight if, for every finite real M , there exists
a compact set C ⊂ Ξ such that

lim sup  log P (X 6∈ C) ≤ −M (30.34)

→0

The first use of exponential tightness is a converse to the contraction prin-

ciple: a high-level LDP is implied by the combination of a low-level LDP and
high-level exponential tightness.

Theorem 459 (Inverse Contraction Principle) If X are exponentially tight,

f is continuous and injective, and Y = f (X ) obeys an LDP with rate function
K, then X obeys an LDP with a good rate function J(x) = K(f (x)).

Proof: See Kallenberg, Theorem 27.11 (ii). Notice, by the way, that the proof
of the upper bound on probabilities (i.e. that lim sup  log P (X ∈ B) ≤ −J(B)
for closed B ⊆ Ξ) does not depend on exponential tightness, just the continuity
of f . Exponential tightness is only needed for the lower bound. 

Theorem 460 (Bryc’s Theorem) If X are exponentially tight, and, for all
bounded continuous f , the limit
h i
Λf ≡ lim  log E ef (X /) (30.35)
→0

exists, then X obeys the LDP with good rate function

J(x) ≡ sup f (x) − Λf (30.36)

f

where the supremum extends over all bounded, continuous functions.

CHAPTER 30. LARGE DEVIATIONS: BASICS 254

Proof: See Kallenberg, Theorem 27.10, part (ii). 

Remark: This is a converse to Varadhan’s Lemma.

Theorem 461 (Projective Limit Theorem (Dawson-Gärtner)) Let Ξ1 , Ξ2 , . . .

be a countable sequence of metric spaces, and let X be a random sequence from
this space. If, for every n, Xn = πn X obeys the LDP with good rate function
Jn , then X obeys the LDP with good rate function

J(x) ≡ sup Jn (πn x) (30.37)

n

Proof: See Kallenberg, Theorem 27.12. 

Definition 462 (Exponentially Equivalent Random Variables) Two fam-

ilies of random variables, X and Y , taking values in a common metric space,
are exponentially equivalent when, for all positive δ,

lim  log P (d(X , Y ) > δ) = −∞ (30.38)

→0

Lemma 463 If X and Y are exponentially equivalent, one of them obeys the
LDP with a good rate function J iff the other does as well.

Proof: It is enough to prove that the LDP for X implies the LDP for Y , with
the same rate function. (Draw a truth-table if you don’t believe me!) As usual,
first we’ll get the upper bound, and then the lower.
Pick any closed set C, and let Cδ be its closed δ neighborhood, i.e., Cδ =
{x : ∃y ∈ C, d(x, y) ≤ δ}. Now

P (Y ∈ Cδ ) ≤ P (X ∈ Cδ ) + P (d(X , Y ) > δ) (30.39)

Using Eq. 30.38 from Definition 462, the LDP for X , and Lemma 443

lim sup  log P (Y ∈ C) (30.40)

≤ lim sup  log P (X ∈ Cδ ) +  log P (d(X , Y ) > δ)
≤ lim sup  log P (X ∈ Cδ ) ∨ lim sup  log P (d(X , Y ) > δ) (30.41)
≤ −J(Cδ ) ∨ −∞ (30.42)
= −J(Cδ ) (30.43)

Since J is a good rate function, we have J(Cδ ) ↑ J(C) as δ ↓ 0; since δ was

arbitrary to start with,

lim sup  log P (Y ∈ C) ≤ −J(C) (30.44)

As usual, to obtain the lower bound on open sets, pick any open set O and any
point x ∈ O. Because O is open, there is a δ > 0 such that, for some open
CHAPTER 30. LARGE DEVIATIONS: BASICS 255

neighborhood U of x, not only is U ⊂ O, but Uδ ⊂ O. In which case, we can

say that
P (X ∈ U ) ≤ P (Y ∈ O) + P (d(X , Y ) > h) (30.45)
Proceeding as for the upper bound,

− J(x) ≤ −J(U ) (30.46)

≤ lim inf  log P (X ∈ U ) (30.47)
≤ lim inf  log P (Y ∈ O) ∨ lim sup  log P (d(X , Y ) > δ)(30.48)
= lim inf  log P (Y ∈ O) (30.49)

(Notice that the initial arbitrary choice of δ has dropped out.) Taking the
supremum over all x gives −J(O) ≤ lim inf  log P (Y ∈ O), as required. 
Chapter 31

Large Deviations for IID

Sequences: The Return of
Relative Entropy

Section 31.1 introduces the exponential version of the Markov in-

equality, which will be our major calculating device, and shows how
it naturally leads to both the cumulant generating function and the
Legendre transform, which we should suspect (correctly) of being the
large deviations rate function. We also see the reappearance of rela-
tive entropy, as the Legendre transform of the cumulant generating
functional of distributions.
Section 31.2 proves the large deviations principle for the empir-
ical mean of IID sequences in finite-dimensional Euclidean spaces
(Cramér’s Theorem).
Section 31.3 proves the large deviations principle for the empiri-
cal distribution of IID sequences in Polish spaces (Sanov’s Theorem),
using Cramér’s Theorem for a well-chosen collection of bounded con-
tinuous functions on the Polish space, and the tools of Section 30.2.
Here the rate function is the relative entropy.
Section 31.4 proves that even the infinite-dimensional empirical
process distribution of an IID sequence in a Polish space obeys the
LDP, with the rate function given by the relative entropy rate.

The usual approach in large deviations theory is to establish an LDP for

some comparatively tractable basic case through explicit calculations, and then
use the machinery of Section 30.2 to extend it to LDPs for more complicated
cases. This chapter applies this strategy to IID sequences.

256
CHAPTER 31. IID LARGE DEVIATIONS 257

31.1 Cumulant Generating Functions and Rela-

tive Entropy
Suppose the only inequality we knew in probability theory was Markov’s inequal-
ity, P (X ≥ a) ≤ E [X] /a when X ≥ 0. How might we extract an exponential
probability bound from it? Well, for any real-valued

 tX variable,
 ta etX is
 positive, so
tX ta
we can say that P (X ≥ a) = P e ≥ e ≤E e /e . E etX is of course
the moment generating function of X. It has the nice property that addition of
independent random variablesleads to multiplication  of their
  moment generat-

|=
 
ing functions, as E et(X1 +X2 ) = E etX1 etX2 = E etX1 E etX2 if X1 X2 .
If X1 , X2 , . . . are IID, then we can get a deviation bound for their sample mean
X n through the moment generating function:
n
!

X
P Xn ≥ a = P Xi ≥ na
i=1

n
≤ e−nta E etX1


P Xn ≥ a
1
≤ −ta + log E etX1

 
log P X n ≥ a
n
≤ inf −ta + log E etX1
 
t
≤ − sup ta − log E etX1
 
t
   
This suggests that the functions log E etX and sup ta − log E etX will be
useful to us. Accordingly, we encapsulate them in a pair of definitions.

Definition 464 (Cumulant Generating Function) The cumulant generat-

ing function of a random variable X in Rd is a function Λ : Rd 7→ R,

Λ(t) ≡ log E et·X

 
(31.1)

Definition 465 (Legendre Transform) The Legendre transform of a real-

valued function f on Rd is another real-valued function on Rd ,

f ∗ (x) ≡ sup t · x − f (t) (31.2)

t∈Rd

The definition of cumulant generating functions and their Legendre trans-

forms can be extended to arbitrary spaces where some equivalent of the inner
product (a real-valued form, bilinear in its two arguments) makes sense; f and
f ∗ then must take arguments from the complementary spaces.
Legendre transforms are particularly important in convex analysis1 , since
convexity is preserved by taking Legendre transforms. If f is not convex initially,
1 See Rockafellar (1970), or, more concisely, Ellis (1985, ch. VI).
CHAPTER 31. IID LARGE DEVIATIONS 258

then f ∗∗ is (in one dimension) something like the greatest convex lower bound
on f ; made precise, this statement even remains true in higher dimensions. I
make these remarks because of the following fact:

Lemma 466 The cumulant generating function Λ(t) is convex.

Proof: Simple calculation, using Hölder’s inequality in one step:

h i
Λ(at + bu) = log E e(at+bu)X (31.3)
= log E eatX ebuX
 
(31.4)
h
a uX
b i
= log E etX e (31.5)

a  buX 
b
≤ log E etX

E e (31.6)
= aΛ(t) + bΛ(u) (31.7)

which proves convexity. 

Our previous result, then, is easily stated: if the Xi are IID in R, then

P X n ≥ a ≤ Λ∗ (a)


(31.8)

where Λ∗ (a) is the Legendre transform of the cumulant generating function of

the Xi . This elementary fact is, surprisingly enough, the foundation of the large
deviations principle for empirical means.
The notion of cumulant generating functions can be extended to probability
measures, and this will be useful when dealing with large deviations of empiri-
cal distributions. The definitions follow the pattern one would expect from the
complementarity between probability measures and bounded continuous func-
tions.

Definition 467 (Cumulant Generating Functional) Let X be a random

variable on a metric space Ξ, with distribution µ, and let Cb (Ξ) be the class of all
bounded, continuous, real-valued functions on Ξ. Then the cumulant-generating
functional Λ : Cb (Ξ) 7→ R is
h i
Λ(f ) ≡ log E ef (X) (31.9)

Definition 468 The Legendre transform of a real-valued functional F on Cb (Ξ)

is
F ∗ (ν) ≡ sup Eν [f ] − Λ(f ) (31.10)
f ∈Cb (Ξ)

where ν ∈ P (Ξ), the set of all probability measures on Ξ.

Lemma 469 (Donsker and Varadhan) The Legendre transform of the cu-
mulant generating functional is the relative entropy:

Λ∗ (ν) = D (νkµ) (31.11)

CHAPTER 31. IID LARGE DEVIATIONS 259

Proof: First of all, notice that the supremum in Eq. 31.10 can be taken over
all bounded measurable functions, not just functions in Cb , since Cb is dense.
This will let us use indicator functions and simple functions in the subsequent
argument.
If ν 6 µ, then D (νkµ) = ∞. But then there is also a set, call it B, with
µ(B) = 0, ν(B) > 0. Take fn = n1B . Then Eν [fn ] − Λ(fn ) = nν(B) − 0, which
can be made arbitrarily large by taking n arbitrarily large, hence the supremum
in Eq. 31.10 is ∞.
If ν  µ, then show that D (νkµ) ≤ Λ∗ (ν) and D (νkµ) ≥ Λ∗ (ν), so they
must be equal. To get the first inequality, start with the observation then
dν dν
dµ exists, so set f = log dµ , which is measurable. Then D (νkµ) is Eν [f ] −
log Eµ e . If f is bounded, this shows that D (νkµ) ≤ Λ∗ (ν). If f is not
 f

bounded, approximate
 it by a sequence of bounded, measurable functions fn
with Eµ efn → 1 and Eν [fn ] → Eν [fn ], again concluding that D (νkµ) ≤
Λ∗ (ν).
To go the other way, first consider the special case where X is finite, and so
generated by a partition, with cells B1 , . . . Bn . Then all measurable functions
are simple functions, and Eν [f ] − Λ(f ) is
n
X n
X
g(f ) = fi ν(Bi ) − log efi µ(Bi ) (31.12)
i=1 i=1

Now, g(f ) is concave on all the fi , and

∂g(f ) 1
= ν(Bi ) − Pn f µ(Bi )efi (31.13)
∂fi i=1 e i µ(B )
i

Setting this equal to zero,

ν(Bi ) 1
= Pn fi
efi (31.14)
µ(Bi ) i=1 µ(B i )e
ν(Bi )
log = fi (31.15)
µ(Bi )

gives the maximum value of g(f ). (Remember that 0 log 0 = 0.) But then
g(f ) = D (νkµ). So Λ∗ (ν) ≤ D (νkµ) when the σ-algebra is finite. In the
general case, consider the
 case where f is a simple function. Then σ(f ) is finite,
and Eν [f ] − log Eµ ef ≤ D (νkµ) follows by the finite case and smoothing.
Finally, if f is not simple, but
 is bounded and measurable,
  there is a simple h
such that Eν [f ] − log Eµ ef ≤ Eν [h] − log Eµ eh , so

sup Eν [f ] − log Eµ ef ≤ D (νkµ)

 
(31.16)
f ∈Cb (Ξ)

which completes the proof. 

CHAPTER 31. IID LARGE DEVIATIONS 260

31.2 Large Deviations of the Empirical Mean in

Rd
Historically, the oldest and most important result in large deviations is that the
empirical mean of an IID sequence of real-valued random variables obeys a large
deviations principle with rate n; the oldest version of this proposition goes back
to Harald Cramér in the 1930s, and so it is known as Cramér’s theorem, even
though the modern version, which is both more refined technically and works in
arbitrary finite-dimensional Euclidean spaces, is due to Varadhan in the 1960s.

Theorem 470 (Cramér’s Theorem) If Xi are IID random variables in Rd ,

and Λ(t) < ∞ for all t ∈ Rd , then their empirical mean obeys an LDP with rate
n and good rate function Λ∗ (x).

Proof: The proof has three parts. First, the upper bound for closed sets;
second, the lower bound for open sets, under an additional assumption on Λ(t);
third and finally, lifting of the assumption on Λ by means of a perturbation
argument (related to Lemma 463).
To prove the upper bound for closed sets, we first prove the upper bound
for sufficiently small balls around arbitrary points. Then, we take our favorite
closed set, and divide it into a compact part close to the origin, which we can
cover by a finite number of closed balls, and a remainder which is far from the
origin and of low probability.
First the small balls of low probability. Because Λ∗ (x) = supu u · x − Λ(u),
for any  > 0, we can find some u such that u · x − Λ(x) > min 1/, Λ∗ (x) − .
(Otherwise, Λ∗ (x) would not be the least upper bound.) Since u · x is contin-
uous in x, it follows that there exists some open ball B of positive radius,
centered on x, within which u · y − Λ(x) > min 1/, Λ∗ (x) − , or u · y >
Λ(x) + min 1/, Λ∗ (x) − . Now use the exponential Markov inequality to get
h i
≤ E eu·nX n −n inf y∈B u·y


P Xn ∈ B (31.17)
∗
≤ e−n(min  ,Λ (x)−)
1
(31.18)

which is small. To get the the compact set near the origin of high probability,
use the exponential decay of the probability at large kxk. Since Λ(t) < ∞ for
all t, Λ∗ (x) → ∞ as kxk → ∞. So, using (once again) the exponential Markov
inequality, for every  > 0, there must exist an r > 0 such that
1
1
log P X n > r ≤ − (31.19)
n 
for all n.
Now pick your favorite closed measurable set C ∈ B d . Then C∩{x : kxk ≤ r}
is compact, and I can cover it by m balls B1 , . . . Bm , with centers x1 , . . . xm ,
of the sort built in the previous paragraph. So I can apply a union bound to
CHAPTER 31. IID LARGE DEVIATIONS 261


P X n ∈ C , as follows.


P Xn ∈ C (31.20)



= P X n ∈ C ∩ {x : kxk ≤ r} + P X n ∈ C ∩ {x : kxk > r}
m
!
[

≤ P Xn ∈ Bi + P X n > r (31.21)
i=1
m
!
X


≤ P X n ∈ Bi + P X n > r (31.22)
i=1
m
!
∗
−n(min 1
) 1
X
≤ e  ,Λ (xi )− + e−n  (31.23)
i=1
∗
≤ (m + 1)e−n(min  ,Λ (C)−)
1
(31.24)

with Λ∗ (C) = inf x∈C Λ∗ (x), as usual. So if I take the log, normalize, and go to
the limit, I have
1 1
− min , Λ∗ (C) − 


lim sup log P X n ∈ C ≤ (31.25)
n n 
≤ −Λ∗ (C) (31.26)

since  was arbitrary to start with, and I’ve got the upper bound for closed sets.
To get the lower bound for open sets, pick your favorite open set O ∈ B d ,
and your favorite x ∈ O. Suppose, for the moment, that Λ(t)/ktk → ∞ as
ktk → ∞. (This is the growth condition mentioned earlier, which we will left at
the end of the proof.) Then, because Λ(t) is smooth, there is some u such that
∇Λ(u) = x. (You will find it instructive to draw the geometry here.) Now let
Yi be a sequence of IID random variables, whose probability law is given by
 
E euX 1B (X)
= e−Λ(u) E euX 1B (X)
 
P (Yi ∈ B) = (31.27)
E [euX ]

It is not hard to show, by manipulating the cumulant generating functions, that

ΛY (t) = ΛX (t + u) − ΛX (u), and consequently that E [Yi ] = x. I construct these
Y to allow me to pull the following trick, which works if  > 0 is sufficiently
small that the first inequality holds (and I can always chose small enough ):



P X n ∈ O ≥ P X n − x <  (31.28)
h i
= enΛ(u) E e−nuY n 1{y : ky − xk < }(Y n ) (31.29)
≥ enΛ(u)−nu·x−nkuk P Y n − x < 


(31.30)
CHAPTER 31. IID LARGE DEVIATIONS 262


By the strong law of large numbers, P Y n − x <  → 1 for all , so

1

lim inf log P X n ∈ O ≥ Λ(u) − u · x − kuk (31.31)
n
∗
≥ −Λ (x) − kuk (31.32)
≥ −Λ∗ (x) (31.33)
≥ − inf Λ∗ (x) = −Λ(O) (31.34)
x∈O

as required. This proves the LDP, as required, if Λ(t)/ktk → ∞ as ktk → ∞.

Finally, to lift the last-named restriction (which, remember, only affected the
lower bound for open sets), introduce a sequence Zi of IID standard Gaussian
variables, i.e. Zi ∼ N (0, I), which are completely independent of the Xi . It is
2
easily calculated that the cumulant generating function of the Zi is ktk /2, so
that Z n satisfies the LDP. Another easy calculation shows that Xi + σZi has
2 2
cumulant generating function ΛX (t) + σ2 ktk , which again satisfies the previous
∗ ∗
condition. Since ΛX+σZ ≥ ΛX , ΛX ≥ ΛX+σZ . Now, once again pick any open
set O, and any point x ∈ O, and an  sufficiently small that all points within a
distance 2 of x are also in O. Since the LDP applies to X + σZ,

P X n + σZ n − x ≤  ≥ −Λ∗X+σZ (x)


(31.35)
∗
≥ −ΛX (x) (31.36)

On the other hand, basic probability manipulations give

P X n + σZ n − x ≤  ≤ P X n ∈ O + P σ Z n ≥  (31.37)



≤ 2 max P X n ∈ O , P σ Z n ≥  (31.38)

Taking the liminf of the normalized log of both sides,

1

lim inf log P X n + σZ n − x ≤  (31.39)
n
1



≤ lim inf log max P X n ∈ O , P σ Z n ≥ 
n
2
 
1

≤ lim inf log P X n ∈ O ∨ − 2 (31.40)
n 2σ
(31.41)

Since σ was arbitrary, we can let it go to zero, and obtain

1
−Λ∗X (x)


lim inf log P X n ∈ O ≥ (31.42)
n
≥ −Λ∗X (O) (31.43)

as required. 
CHAPTER 31. IID LARGE DEVIATIONS 263

31.3 Large Deviations of the Empirical Measure

in Polish Spaces
The Polish space setting is, apparently, more general than Rd , but we will
represent distributions on the Polish space in terms of the expectation of a
separating set of functions, and then appeal to the Euclidean result.

Proposition 471 Any Polish space S can be represented as a Borel subset of

N
a compact metric space, namely [0, 1] ≡ M .

Proof: See, for instance, Appendix A of Kallenberg. 

Strictly speaking, there should be a function mapping points from S to
points in M . However, since this is an embedding, I will silently omit it in what
follows.

Proposition 472 Cb (M ) has a countable dense separating set F = f1 , f2 , . . ..

Proof: See Kallenberg again. 

Because F is separating, to specify a probability distribution on K is equiv-
alent to specifying the expectation value of all the functions in F. Write f1d (X)
to abbreviate the d-dimensional vector (f1 (X), f2 (X), . . . fd (X)), and f1∞ (X)
to abbreviate the corresponding infinite-dimensional vector.

Lemma 473 Empirical means are expectations with respect to empirical mea-
sure. That is, let f be a real-valued measurable function and Yi = f (Xi ). Then
Y n = EP̂n [f (X)].

Proof: Direct calculation.

n
1X
Yn ≡ f (Xi ) (31.44)
n i=1
n
1X
= Eδ [f (X)] (31.45)
n i=1 Xi
≡ EP̂n [f (X)] (31.46)

Lemma 474 Let Xi be a sequence of IID random variables in a Polish space

Ξ. For each d, the sequence of vectors (EP̂n [f1 ] , . . . EP̂n [fd ]) obeys the LDP
with rate n and good rate function Jd .

Proof: For each d, the sequence of vectors (f1 (Xi ), . . . fd (Xi )) are IID, so, by
Cramér’s Theorem (470), their empirical mean obeys the LDP with rate n and
good rate function
h d i
Jd (x) = sup t · x − log E et·f1 (X) (31.47)
t∈Rd
CHAPTER 31. IID LARGE DEVIATIONS 264

But, by Lemma 473, the empirical means are expectations over the empirical
distributions, so the latter must also obey the LDP, with the same rate and rate
function. 
Notice, incidentally, that the fact that the fi ∈ F isn’t relevant for the proof
of the lemma; it will however be relevant for the proof of the theorem.

Theorem 475 (Sanov’s Theorem) Let Xi , i ∈ N, be IID random variables

in a Polish space Ξ, with common probability measure µ. Then the empirical dis-
tributions P̂n obey an LDP with rate n and good rate function J(ν) = D (νkµ).

Proof: Combining Lemma 474 and Theorem 461, we see that EP̂n [f1∞ (X)]
obeys the LDP with rate n and good rate function

J(x) = sup Jd (πd x) (31.48)

d
h d i
= sup sup t · πd x − log E et·f1 (X) (31.49)
d t∈Rd

Since P (M ) is compact (so all random sequences in it are exponentially

tight), and the mapping from ν ∈ P (M ) to Eν [f1∞ ] ∈ RN is continuous, apply
the inverse contraction principle (Theorem 459) to get that P̂n satisfies the LDP
with good rate function

J(ν) = J(Eν [f1∞ ]) (31.50)

h d i
sup sup t · Eν f1d − log Eµ et·f1 (X)
 
= (31.51)
d t∈Rd
= sup Eν [f ] − Λ(f ) (31.52)
f ∈spanF
= sup Eν [f ] − Λ(f ) (31.53)
f ∈Cb (M )

= D (νkµ) (31.54)

Notice however that this is an LDP in the space P (M ), not in P (Ξ). However,
the embedding taking P (Ξ) to P (M ) is continuous, and it is easily verified (see
Lemma 27.17 in Kallenberg) that P̂n is exponentially tight in P (Ξ), so another
application of the inverse contraction principle says that P̂n must obey the LDP
in the restricted space P (Ξ), and with the same rate. 

31.4 Large Deviations of the Empirical Process

in Polish Spaces
A fairly straightforward modification of the proof for Sanov’s theorem estab-
lishes a large deviations principle for the finite-dimensional empirical distribu-
tions of an IID sequence.
CHAPTER 31. IID LARGE DEVIATIONS 265

Corollary 476 Let Xi be an IID sequence in a Polish space Ξ, with common

measure µ. Then, for every finite positive integer k, the k-dimensional empirical
distribution P̂nk , obeys an LDP

with rate n and good rate function Jk (ν) =
D (νkπk−1 ν ⊗ µ) if ν ∈ P Ξk is shift invariant, and J(ν) = ∞ otherwise.

This leads to the following important generalization.

Theorem 477 If Xi are IID in a Polish space, with a common measure µ, then
the empirical process distribution P̂n∞ obeys an LDP with rate n and good rate
function J∞ (ν) = d(νkµ∞ ), the relative entropy rate, if ν is a shift-invariant
probability measure, and = ∞ otherwise.

Proof: By Corollary 476 and the projective limit theorem 461, P̂n∞ obeys an
LDP with rate n and good rate function

J∞ (ν) = sup Jk (πk ν) = sup D (πk νkπk−1 ν ⊗ µ) (31.55)

k k

But, applying the chain rule for relative entropy (Lemma 404),

D (πn νkµn ) = D (πn νkπn−1 ν ⊗ µ) + D πn−1 νkµn−1

(31.56)
Xn
= D (πk νkπk−1 ν ⊗ µ) (31.57)
k=1
n
1 1X
lim D (πn νkµn ) = lim D (πk νkπk−1 ν ⊗ µ) (31.58)
n n
k=1
= sup D (πk νkπk−1 ν ⊗ µ) (31.59)
k

But lim n−1 D (πn νkµn ) is the relative entropy rate, d(νkµ∞ ), and we’ve already
identified the right-hand side as the rate function. 
The strength of Theorem 477 lies in the fact that, via the contraction prin-
ciple (Theorem 451), it implies that the LDP holds for any continuous function
of the empirical process distribution. This in particular includes the finite-
dimensional distributions, the empirical mean, functions of finite-length trajec-
tories, etc. Moreover, Theorem 451 also provides a means to calculate the rate
function for all these quantities.
Chapter 32

Large Deviations for

Markov Sequences

This chapter establishes large deviations principles for Markov

sequences as natural consequences of the large deviations principles
for IID sequences in Chapter 31. (LDPs for continuous-time Markov
processes will be treated in the chapter on Freidlin-Wentzell theory.)
Section 32.1 uses the exponential-family representation of Markov
sequences to establish an LDP for the two-dimensional empirical dis-
tribution (“pair measure”). The rate function is a relative entropy.
Section 32.2 extends the results of Section 32.1 to other observ-
ables for Markov sequences, such as the empirical process and time
averages of functions of the state.

For the whole of this chapter, let X1 , X2 , . . . be a homogeneous Markov se-

quence, taking values in a Polish space Ξ, with transition probability kernel µ,
and initial distribution ν and invariant distribution ρ. If Ξ is not discrete, we
will assume that ν and ρ have densities n and r with respect to some reference
measure, and that µ(x, dy) has density m(x, y) with respect to that same ref-
erence measure, for all x. (LDPs can be proved for Markov sequences without
such density assumptions — see, e.g., Ellis (1988) — but the argument is more
complicated.)

32.1 Large Deviations for Pair Measure of Markov

Sequences
It is perhaps not sufficiently appreciated that Markov sequences form expo-
nential families (Billingsley, 1961; Küchler and Sørensen, 1997). Suppose Ξ is

266
CHAPTER 32. LARGE DEVIATIONS FOR MARKOV SEQUENCES 267

discrete. Then
t−1
Y
X1n xt1


P = = ν(x1 ) µ(xi , xi+1 ) (32.1)
i=1
Pt−1
log µ(xi ,xi+1 )
= ν(x1 )e i=1 (32.2)
P t
= ν(x1 )e x,y∈Ξ2 Tx,y (x1 ) log µ(x,y) (32.3)
where Tx,y (xt1 ) counts the number of times the state y follows the state x in the
sequence xt1 , i.e., it gives the transition counts. What we have just established is
that the Markov chains on Ξ with a given initial distribution form an exponential
family, whose natural sufficient statistics are the transition counts, and whose
natural parameters are the logarithms of the transition probabilities.
(If Ξ is not continuous, but we make the density assumptions mentioned at
the beginning of this chapter, we can write
t−1
Y
p X1t (xt1 ) = n(x1 ) m(xi , xi+1 ) (32.4)
i=1
dT (xt1 ) log m(x,y)
R
= n(x1 )e Ξ2 (32.5)
1
where now T (xt1 ) puts probability mass n−1 at x, y for every i such that xi = x,
xi+1 = y.)
We can use this exponential family representation to establish the following
basic theorem.
Theorem 478 Let Xi be a Markov sequence obeying the assumptions set out
at the beginning of this chapter, and furthermore that µ(x, y)/ρ(y) is bounded
above (in the discrete-state case) or that m(x, y)/r(y) is bounded above (in
the continuous-state case). Then the two-dimensional empirical distribution
(“pair measure”) P̂t2 obeys an LDP with rate n and with rate function J2 (ψ) =
D (ψkπ1 ψ × µ) if ν is shift-invariant, J(ν) = ∞ otherwise.
Proof: I will just give the proof for the discrete case, since the modifications
for the continuous case are straightforward (given the assumptions made about
densities), largely a matter of substituting Roman letters for Greek ones.
First, modify the representation of the probabilities in Eq. 32.3 slightly, so
that it refers directly to P̂t2 (as laid down in Definition 454), rather than to the
transition counts.
ν(x1 ) t Px,y∈Ξ P̂t2 (x,y) log µ(x,y)
P X1t = xt1


= e (32.6)
µ(xt , x1 )
ν(x1 ) nEP̂ 2 [log µ(X,Y )]
= e t (32.7)
µ(xt , x1 )
Now construct a sequence of IID variables Yi , all distributed according to ρ, the
invariant measure of the Markov chain:
nE 2 [log ρ(Y )]
P Y1t = y1t


= e P̂t (32.8)
CHAPTER 32. LARGE DEVIATIONS FOR MARKOV SEQUENCES 268

The ratio of these probabilities is the Radon-Nikodym derivative:

dPX t ν(x1 ) tEP̂ 2 [t log µ(X,Y )
ρ(Y ) ]
(x1 ) = e n (32.9)
dPY µ(xt , x1 )
(In the continuous-Ξ case, the derivative is the ratio of the densities with respect
to the common reference measure,
h andi the principle is the same.) Introducing
µ(X,Y ) 2
the functional F (ν) = Eν log ρ(Y ) , the derivative is equal to O(1)etF (P̂t ) ,
and our initial assumption amounts to saying that F is not just continuous
(which it must be) but bounded from above.
Now introduce Qt,X , the distribution of the empirical pair measure P̂t2 un-
der the Markov process, and Qt,Y , the distribution of P̂t2 for the IID samples
produced by Yi . From Eq. 32.9,
Z
1 
2
 1
log P P̂t ∈ B = log dQt,X (ψ) (32.10)
t t
ZB
1 dQt,X
= log dQt,Y (ψ) (32.11)
t B dQt,Y
  Z
1 1
= O + log etF (ψ) dQt,Y (ψ) (32.12)
t t B

It is thus clear that

Z
1   1
lim inf log P P̂t2 ∈ B = lim inf log etF (ψ) dQt,Y (ψ) (32.13)
t t B
Z
1   1
lim sup log P P̂t2 ∈ B = lim sup log etF (ψ) dQt,Y (ψ) (32.14)
t t B

Introduce a (final) proxy random sequence, also taking values in P (() Ξ2 ), call
R tF (ψ)
it Zt , with P (Zt ∈ B) = B e dQt,Y (ψ). We know (Corollary 476) that,
under Qt,Y , the empirical pair measure satisfies an LDP with rate t and good
rate function JY = D (ψkπ1 ψ ⊗ ρ), so by Corollary 457, Zt satisfies an LDP
with rate t and good rate function

JF (ψ) = −(F (ψ) − JY (ψ)) + sup F (ζ) − JY (ζ) (32.15)

ζ∈P(Ξ2 )

A little manipulation turns this into

JF (ψ) = D (ψkπ1 ψ ⊗ µ) − inf D (ζkπ1 ζ ⊗ µ) (32.16)

ζ∈P(Ξ2 )

and the infimum is clearly zero. Since this is the rate function Zt , in view of
Eqs. 32.13 and 32.14 it is also the rate function for P̂n2 , which we have agreed
to call J2 . 
Remark 1: The key to making this work is the assumption that F is bounded
from above. This can fail if, for instance, the process is not ergodic, although
usually in that case one can rescue the general idea by some kind of ergodic
decomposition.
CHAPTER 32. LARGE DEVIATIONS FOR MARKOV SEQUENCES 269

Remark 2: The LDP for the pair measure of an IID sequence can now
be seen to be a special case of the LDP for the pair measure of a Markov
sequence. The same is true, generally speaking, of all the other LDPs for IID
and Markov sequences. Calculations are almost always easier for the IID case,
however, which permits us to give explicit formulae for the rate functions of
empirical means and empirical distributions unavailable (generally speaking) in
the Markovian case.

Corollary 479 The minima of the rate function J2 are the invariant distribu-
tions.

Proof: The rate function is D (ψkπ1 ψ ⊗ µ). Since relative entropy is ≥ 0, and
equal to zero iff the two distributions are equal (Lemma 401), we get a minimum
of zero in the rate function iff ψ = π1 ψ ⊗ µ, or ψ = ρ2 , for some ρ ∈ P (Ξ) such
that ρµ = ρ. Conversely, if ψ is of this form, then J2 (ψ) = 0. 

Corollary 480 The empirical distribution P̂t obeys an LDP with rate t and
good rate function

J1 (ψ) = inf D (ζkπ1 ζ ⊗ µ) (32.17)

ζ∈P(Ξ2 ):π1 ζ=ψ

Proof: This is a direct application of the Contraction Principle (Theorem 451),

as in Corollary 456. 
Remark: Observe that if ψ is invariant under the action of the Markov chain,
then J1 (ψ) = 0 by a combination of the preceding corollaries. This is good,
because we know from ergodic theory that the empirical distribution converges
on the invariant distribution for an ergodic Markov chain. In fact, in view of
1 2
Lemma 402, which says that D (ψkρ) ≥ 2 ln 2 kψ − ρk1 , the probability that the
empirical distribution differs from the invariant distribution ρ by more than δ,
2
in total variation distance, goes down like O(e−tδ /2 ).

Corollary 481 If Theorem 478 holds, then time averages of observables, At f ,

obey a large deviations principle with rate function

J0 (a) = inf D (ζkπ1 ζ ⊗ µ) (32.18)

ζ∈P(Ξ2 ): Eπ1 ζ [f (X)]

Proof: Another application the Contraction Principle, as in Corollary 456. 

Remark: Observe that if a = Eρ [f (X)], with ρ invariant, then the J0 (a) = 0.
Again, it is reassuring to see that large deviations theory is compatible with
ergodic theory, which tells us to expect the almost-sure convergence of At f on
Eρ [f (X)].

Corollary 482 If Xi are from a Markov sequence of order k + 1, then, under

conditions analogous to Theorem 478, the k + 1-dimensional empirical distribu-
tion P̂tk+1 obeys an LDP with rate t and good rate function

D (νkπk−1 ν ⊗ µ) (32.19)
CHAPTER 32. LARGE DEVIATIONS FOR MARKOV SEQUENCES 270

Proof: An obvious extension of the argument for Theorem 478, using the
appropriate exponential-family representation of the higher-order process. 
Whether all exponential-family stochastic processes (Küchler and Sørensen,
1997) obey LDPs is an interesting question; I’m not sure if anyone knows the
answer.

32.2 Higher LDPs for Markov Sequences

In this section, I assume without further comment that the Markov sequence X
obeys the LDP of Theorem 478.

Theorem 483 For all k ≥ 2, the finite-dimensional empirical distribution P̂tk

obeys an LDP

with rate t and good rate function Jk (ψ) = D (ψkπk−1 ψ ⊗ µ), if
ψ ∈ P Ξk is shift-invariant, and = ∞ otherwise.

Proof: The case k = 2 is just Theorem  478. However, if k ≥ 3, the argument

preceding that theorem shows that P P̂tk ∈ B depends only on π2 P̂tk , the pair
measure implied by the k-dimensional distribution, so the proof of that theorem
can be adapted to apply to P̂tk , in conjunction with Corollary 476, establishing
the LDP for finite-dimensional distributions of IID sequences. The identification
of the rate function follows the same argument, too. 

Theorem 484 The empirical process distribution obeys an LDP with rate t
and good rate function J∞ (ψ) = d(ψkρ), with ρ here standing for the stationary
process distribution of the Markov sequence.

Proof: Entirely parallel to the proof of Theorem 477, with Theorem 483 sub-
stituting for Corollary 476. 
Consequently, any continuous function of the empirical process distribution
has an LDP.
Chapter 33

Large Deviations for

Weakly Dependent
Sequences via the
Gärtner-Ellis Theorem

This chapter proves the Gärtner-Ellis theorem, establishing an

LDP for not-too-dependent processes taking values in topological
vector spaces. Most of our earlier LDP results can be seen as con-
sequences of this theorem.

33.1 The Gärtner-Ellis Theorem

The Gärtner-Ellis theorem is a powerful result which establishes the existence
of a large deviation principle for processes where the cumulant generating func-
tion tends towards a well-behaved limit, implying not-too-strong dependence
between successive values. (Exercise 33.5 clarifies the meaning of “too strong”.)
It will imply our LDPs for IID and Markovian sequences. I could have started
with it, but its proof, as you’ll see, is pretty technical, and so it seemed better
to use the more elementary arguments of the preceding chapters.
To fix notation, Ξ will be a real topological vector space, and Ξ∗ will be its
dual space, of continuous linear functions Ξ 7→ R. (If Ξ = Rd , we can identify
Ξ and Ξ∗ by means of the inner product. In differential geometry, on the other
hand, Ξ might be a space of tangent vectors, and Ξ∗ the corresponding one-
forms.) X will be a family of Ξ-valued random variables, parameterized by
 > 0. Refer to Definitions 464 and 465 in Section 31.1 for the definition of the
cumulant generating function and its Legendre transform (respectively), which
I will denote by Λ : Ξ∗ 7→ R and Λ∗ : Ξ 7→ R.
The proof of the Gärtner-Ellis theorem goes through a number of lemmas.

271
CHAPTER 33. THE GÄRTNER-ELLIS THEOREM 272

Basically, the upper large deviation bound holds under substantially weaker
conditions than the lower bound does, and it’s worthwhile having the partial
results available to use in estimates even if the full large deviations principle
does not apply.

Definition 485 The upper-limiting cumulant generating function is

Λ (t) ≡ lim sup Λ (t/) (33.1)

→0

∗
and its Legendre transform is written Λ (x).

The point of this is that the limsup always exists, whereas the limit doesn’t,
necessarily. But we can show that the limsup has some reasonable properties,
and in fact it’s enough to give us an upper bound.
∗
Lemma 486 Λ (t) is convex, and Λ (x) is a convex rate function.

Proof: The proof of the convexity of Λ (t) follows the proof in Lemma 466,
∗
and the convexity of Λ (t) by passing to the limit. To etablish Λ (x) as a
rate function, we need it to be non-negative and lower semi-continuous. Since
∗
Λ (0) = 0 for all , Λ (0) = 0. This in turn implies that Λ (x) ≥ 0. Since the
latter is the supremum of a class of continuous functions, namely t(x) − Λ (t),
it must be lower semi-continuous. Finally, its convexity is implied by its being
a Legendre transform. 

Lemma 487 (Upper Bound in Gärtner-Ellis Theorem: Compact Sets)

For any compact set K ⊂ Ξ,
∗
lim sup  log P (X ∈ K) ≤ −Λ (K) (33.2)
→0

Proof: Entirely parallel to the proof of the upper bound in Cramér’s Theorem
(470), up through the point where closed sets are divided into a compact part
and a remainder far from the origin, of exponentially-small probability. Because
K is compact, we can proceed as though the remainder is empty. 

Lemma 488 (Upper Bound in Gärtner-Ellis Theorem: Closed Sets) If

the family of distributions L (X ) are exponentially tight, then for all closed
C ⊂ Ξ,
∗
lim sup  log P (X ∈ C) ≤ −Λ (C) (33.3)
→0

Proof: Exponential tightness, by definition (458), will let us repeat Theorem

470 trick of dividing closed sets into a compact part, and an exponentially-
vanishing non-compact part. 
CHAPTER 33. THE GÄRTNER-ELLIS THEOREM 273

Definition 489 The limiting cumulant generating function is

Λ(t) ≡ lim Λ (t/) (33.4)

→0

when the limit exists. Its domain of finiteness D ≡ {t ∈ Ξ∗ : Λ(t) < inf ty}. Its
limiting Legendre transform is Λ∗ , with domain of finiteness D∗ .

Lemma 490 If Λ(t) exists, then it is convex, Λ∗ (x) is a convex rate function,
and Eq. 33.2 applies to the latter. If in addition the process is exponentially
tight, then Eq. 33.3 holds for Λ∗ (x).

Proof: Because, if a limit exists, it is equal to the limsup. 

Lemma 491 If Ξ = Rd , then 0 ∈ intD is sufficient for exponential tightness.

Proof: Exercise. 
Unfortunately, this is not good enough to get exponential tightness in arbi-
trary vector spaces.
∗
Definition 492 (Exposed Point) A point x ∈ Ξ is exposed for Λ (·) when
∗ ∗
there is a t ∈ Ξ∗ such that Λ (y) − Λ (x) > t(y − x) for all y 6= x. t is the
exposing hyper-plane for x.
∗
In R1 , a point x is exposed if the curve Λ (y) lies strictly above the line of slope
∗ ∗
t through the point (x, Λ (x)). Similarly, in Rd , the Λ (y) surface must lie
∗
strictly above the hyper-plane passing through (x, Λ (x)) with surface normal
∗
t. Since Λ (y) is convex, we could generally arrange this by making this the
tangent hyper-plane, but we do not, yet, have any reason to think that the
∗
tangent is well-defined. — Obviously, if Λ(t) exists, we can replace Λ (·) by
∗
Λ (·) in the definition and the rest of this paragraph.

Definition 493 An exposed point x ∈ Ξ with exposing hyper-plane t is nice,

x ∈ N , if
lim Λ (t/) (33.5)
→0

exists, and, for some r > 1,

Λ (rt) < ∞ (33.6)

Note: Most books on large deviations do not give this property any particular
name.

Lemma 494 (Lower Bound in Gärtner-Ellis Theorem) If the X are ex-

ponentially tight, then, for any open set O ⊆ Ξ,
∗
lim inf  log P (X ∈ O) ≥ −Λ (O ∩ N ) (33.7)
→0
CHAPTER 33. THE GÄRTNER-ELLIS THEOREM 274

Proof: If we pick any nice, exposed point x ∈ O, we can repeat the proof of
the lower bound from Cramér’s Theorem (470). In fact, the point of Definition
493 is to ensure this. Taking the supremum over all such x gives the lemma. 

Theorem 495 (Abstract Gärtner-Ellis Theorem) If the X are exponen-

∗ ∗
tially tight, and Λ (O ∩ E) = Λ (O) for all open sets O ⊆ Ξ, then X obey an
∗
LDP with rate 1/ and good rate function Λ (x).

Proof: The large deviations upper bound is Lemma 488. The large deviations
lower bound is implied by Lemma 494 and the additional hypothesis of the
theorem. 
Matters can be made a bit more concrete in Euclidean space (the original
home of the theorem), using, however, one or two more bits of convex analysis.

Definition 496 (Relative Interior) The relative interior of a non-empty and

convex set A is

rintA ≡ {x ∈ A : ∀y ∈ A, ∃δ > 0, x − δ(y − x) ∈ A} (33.8)

Notice that intA ⊆ rintA, since the latter, in some sense, doesn’t care about
points outside of A.

Definition 497 (Essentially Smooth) Λ is essentially smooth if (i) intD 6=

∅, (ii) Λ is differentiable on intD, and (iii) Λ is “steep”, meaning that if D has
a boundary, ∂D, then limt→∂D k∇Λ(t)k = ∞.

This definition was introduced to exploit the following theorem of convex

analysis.

Proposition 498 If Λ is essentially smooth, then Λ∗ is continuous on rintD∗ ,

and rintD∗ ⊆ N , the set of exposed, nice points.

Proof: D∗ is non-empty, because there is at least one point x0 where Λ∗ (x0 ) =

0. Moreover, it is a convex set, because Λ∗ is a convex function (Lemma 490),
so it has a relative interior. Now appeal to Rockafellar (1970, Corollary 26.4.1).

Remark: You might want to try to prove that Λ∗ is continuous on rintD∗ .

Theorem 499 (Euclidean Gärtner-Ellis Theorem) Suppose X , taking val-

ues in Rd , are exponentially tight, and that Λ(t) exists and is essentially smooth.
Then X obey an LDP with rate 1/ and good rate function Λ∗ (x).

Proof: From the abstract Gärtner-Ellis Theorem 495, it is enough to show

that Λ∗ (O ∩ N ) = Λ∗ (O), for any open set O. That is, we want

inf Λ∗ (x) = inf Λ∗ (x) (33.9)

x∈O∩N x∈O
CHAPTER 33. THE GÄRTNER-ELLIS THEOREM 275

Since it’s automatic that

inf Λ∗ (x) ≥ inf Λ∗ (x) (33.10)

x∈O∩N x∈O

what we need to show is that

inf Λ∗ (x) ≤ inf Λ∗ (x) (33.11)

x∈O∩N x∈O

In view of Proposition 498, it’s really just enough to show that Λ∗ (O∩rintD∗ ) ≤
Λ∗ (O). This is trivial when the intersection O ∩ D∗ is empty, so assume it
isn’t, and pick any x in that intersection. Because O is open, and because
of the definition of the relative interior, we can pick any point y ∈ rintD∗ ,
and, for sufficiently small δ, δy + (1 − δ)x ∈ O ∩ rintD∗ . Since Λ∗ is convex,
Λ∗ (δy + (1 − δ)x) ≤ δΛ∗ (y) + (1 − δ)Λ∗ (x). Taking the limit as δ → 0,

Λ∗ (O ∩ rintD∗ ) ≤ Λ∗ (x) (33.12)

and the claim follows by taking the infimum over x.

33.2 Exercises
Exercise 33.1 Show that Cramér’s Theorem (470), is a special case of the
Euclidean Gärtner-Ellis Theorem (499).

Exercise 33.2 Let Z1 , Z2 , . . . be IID random variables in a discrete space, and

let X1 , X2 , . . . be the empirical distributions they generate. Use the Gärtner-
Ellis Theorem to re-prove Sanov’s Theorem (475). Can you extend this to the
case where the Zi take values in an arbitrary Polish space?

Exercise 33.3 Let Z1 , Z2 , . . . be values from a stationary ergodic Markov chain

(i.e. the state space is discrete). Repeat the previous exercise for the pair mea-
sure. Again, can the result be extended to arbitrary Polish state spaces?

Exercise
P∞ 33.4 Let Zi be real-valued mean-zero stationary
Pt Gaussian variables,
−1
with i=−∞ |cov (Z0 , Zi ) | < ∞. Let Xt = t i=1 Zi . Show that these
2
time
P∞ averages obey an LDP with rate t and rate function x /2Γ, where Γ =
i=−∞ cov (Z0 , Zi ). (Cf. the mean-square ergodic theorem 293 of chapter 22.)
If Zi are not Gaussian but are weakly stationary, find an additional hypothesis
such that the time averages still obey an LDP.

Exercise 33.5 (Too-Strong Dependence) Let Xi = X, a random variable

in R, for all i. Show that the Gärtner-Ellis Theorem fails, unless X is degener-
ate.
Chapter 34

Large Deviations for

Stochastic Differential
Equations

This last chapter revisits large deviations for stochastic differen-

tial equations in the small-noise limit, first raised in Chapter 21.
Section 34.1 establishes the LDP for the Wiener process (Schilder’s
Theorem).
Section 34.2 proves the LDP for stochastic differential equations
where the driving noise is independent of the state of the process.
Section 34.3 states the corresponding result for SDEs when the
noise is state-dependent, and gestures in the direction of the proof.

In Chapter 21, we looked at how the diffusions X which solve the SDE

dX = a(X )dt + dW, X (0) = x0 (34.1)

converge on the trajectory x0 (t) solving the ODE

dx
= a(x(t)), x(0) = x0 (34.2)
dt
in the “small noise” limit,  → 0. Specifically, Theorem 270 gave a (fairly crude)
upper bound on the probability of deviations:
 
lim 2 log P sup ∆ (t) > δ ≤ −δ 2 e−2Ka T (34.3)
→0 0≤t≤T

where Ka depends on the Lipschitz coefficient of the drift function a. The the-
ory of large deviations for stochastic differential equations, known as Freidlin-
Wentzell theory for its original developers, shows that, using the metric implicit

276
CHAPTER 34. FREIDLIN-WENTZELL THEORY 277

in the left-hand side of Eq. 34.3, the family of processes X obey a large devia-
tions principle with rate −2 , and a good rate function.
(The full Freidlin-Wentzell theory actually goes somewhat further than just
SDEs, to consider small-noise perturbations of dynamical systems of many sorts,
perturbations by Markov processes (rather than just white noise), etc. Time
does not allow us to consider the full theory (Freidlin and Wentzell, 1998), or
its many applications to nonparametric estimation (Ibragimov and Has’minskii,
1979/1981), systems analysis and signal processing (Kushner, 1984), statistical
mechanics (Olivieri and Vares, 2005), etc.)
As in Chapter 31, the strategy is to first prove a large deviations principle
for a comparatively simple case, and then transfer it to more subtle processes
which can be represented as appropriate functionals of the basic case. Here,
the basic case is the Wiener process W (t), with t restricted to the unit interval
[0, 1].

34.1 Large Deviations of the Wiener Process

We start with a standard d-dimensional Wiener process W , and consider its di-
lation by a factor , X (t) = W (t). There are a number of ways of establishing
that X obeys a large deviation principle as  → 0. One approach (see Dembo
and Zeitouni (1998, ch. 5) starts with establishing an LDP for continuous-time
random walks, ultimately based on the Gärtner-Ellis Theorem, and then show-
ing that the convergence of such processes to the Wiener process (the Functional
Central Limit Theorem, Theorem 215 of Chapter 16) is sufficiently fast that the
LDP carries over. However, this approach involves a number of surprisingly
tricky topological issues, so I will avoid it, in favor of a more probabilistic path,
marked out by Freidlin and Wentzell (Freidlin and Wentzell, 1998, sec. 3.2).
Until further notice, kwk∞ will denote the supremum norm in the space of
continuous curves over the unit interval, C([0, 1], Rd ).

Definition 500 (Cameron-Martin Spaces) The Cameron-Martin space HT

consists of all continuous sample paths x ∈ C([0, T ], Rd ) where x(0) = 0, x is
absolutely continuous, and its Radon-Nikodym derivative ẋ is square-integrable.

Lemma 501 (Cameron-Martin Spaces are Hilbert) Cameron-Martin spaces

RT 2
are Hilbert spaces, with norm kxkCM = 0 |ẋ(t)| dt.

Proof: An exercise (34.1) in verifying that the axioms of a Hilbert space

are satisfied. 

Definition 502 (Effective Wiener Action) The effective Wiener action of

an continuous function x ∈ C([0, t], Rd ) is
1 2
JT (x) ≡ kxkCM (34.4)
2
CHAPTER 34. FREIDLIN-WENTZELL THEORY 278

if x ∈ HT , and ∞ otherwise. In particular,

1 1
Z
2
J1 (x) ≡ |ẋ(t)| dt (34.5)
2 0
For every j > 0, let LT (j) = {x : JT (x) ≤ j}.

Proposition 503 (Girsanov Formula for Deterministic Drift) Fix a func-

tion f ∈ H1 , and let Y = X − f . Then L (Y ) = ν is absolutely continuous
with respect to L (X ) = µ , and the Radon-Nikodym derivative is

1 1
 Z Z 1 
dν 1 2
(w) = exp − ẇ(t) · dW − 2 |ẇ(t)| dt (34.6)
dµ  0 2 0
Proof: This is a special case of Girsanov’s Theorem. See Corollary 18.25
on p. 365 of Kallenberg, or, more transparently perhaps, the relevant parts of
Liptser and Shiryaev (2001, vol. I).
Lemma 504 (Exponential Bound on the Probability of Tubes Around
Given Trajectories) For any δ, γ, K > 0, there exists an 0 > 0 such that, if
 < 0 ,
J1 (x)+γ
P (kX − xk∞ ≤ δ) ≥ e− 2 (34.7)
provided x(0) = 0 and J1 (x) < K.
Proof: Using Proposition 503,
P (kX − xk∞ ≤ δ) = P (kY − 0k∞ ≤ δ) (34.8)
Z
dν
= (w)dµ (w) (34.9)
kwk∞ <δ dµ
Z
J1 (x) 1 1
R
= e− 2 e−  0 ẋ·dW dµ (w) (34.10)
kwk∞ <δ

From Lemma 268 in Chapter 21, we can see that P (kW k∞ < δ) → 1 as  → 0.
So, if  is sufficiently small, P (kW k∞ < δ) ≥ 3/4. Now, applying Chebyshev’s
inequality to the integrand,
√ !
1 1
Z
2 2p
P − ẋ · dW ≤ − J1 (x) (34.11)
 0 
 Z 1  √ !
1  2 2p
≤ P  ẋ · dW  ≤ J1 (x) (34.12)
 0 
 2 
R1
2 E 0
ẋ · dW
≤ (34.13)
82 J1 (x)
R1 2
0
|ẋ| dt 1
= = (34.14)
8J1 (x) 4
CHAPTER 34. FREIDLIN-WENTZELL THEORY 279

using the Itô isometry (Corollary 239). Thus

 1R1 √ √  3
2 2
P e−  0 ẋ·dW ≥ e−  J1 (x) ≥ (34.15)
4

1 − 2√2 √J1 (x)

Z R1
1
e−  0
ẋ·dW
dµ (w) > e  (34.16)
kwk∞ <δ 2
√ √
1 − J1 (x) 2 2
P (kX − xk∞ ≤ δ) > e 2 −  J1 (x) (34.17)
2
where the second term in the exponent can be made less than any desired γ by
taking  small enough. 

Lemma 505 (Trajectories are Rarely Far from Action Minima) For
every j > 0, δ > 0, let U (j, δ) be the open δ neighborhood of L1 (j), i.e., all the
trajectories coming within δ of a trajectory whose action is less than or equal to
j. Then for any γ > 0, there is an 0 > 0 such that, if  < 0 and
s−γ
P (X 6∈ U (j, δ)) ≤ e− 2 (34.18)

Proof: Basically, approximating the Wiener process by a continuous piecewise-

linear function, and showing that the approximation is sufficiently fine-grained.
Chose a natural number n, and let Yn, (t) be the piecewise linear random func-
tion which coincides with X at times 0, 1/n, 2/n, . . . 1, i.e.,
 
[tn]
Yn, (t) = X ([tn]/n) + t − X ([tn + 1]/n) (34.19)
n

We will see that, for large enough n, this is exponentially close to X . First,
though, let’s bound the probability in Eq. 34.18.

P (X ∈
6 U (j, δ))


= P X 6∈ U (j, δ), kX − Yn, k∞ < δ


+P X 6∈ U (j, δ), kX − Yn, k∞ ≥ δ (34.20)



≤ P X 6∈ U (j, δ), kX − Yn, k∞ < δ + P kX − Yn, k∞ ≥ δ (34.21)


≤ P (J1 (Yn, ) > j) + P kX − Yn, k∞ ≥ δ (34.22)

J1 (Yn, ) can be gotten at from the increments of the Wiener process:

n
2 X 2
J1 (Yn, ) = n |W (i/n) − W ((i − 1)/n)| (34.23)
2 i=1
dn
2 X
= ξi (34.24)
2 i=1
CHAPTER 34. FREIDLIN-WENTZELL THEORY 280

where the ξi have the χ2 distribution with one degree of freedom. Using our
results on such distributions and their sums in Ch. 21, it is not hard to show
that, for sufficiently small ,
1 − j−γ
P (J1 (Yn, ) > j) ≤ e 2 (34.25)
2
To estimate the probability that the distance between X and Yn, reaches or
exceeds δ, start with the independent-increments property of X , and the fact
that the two processes coincide when t = i/n.


P kX − Yn, k∞ ≥ δ
Xn  
≤ P max |X (t) − Yn, (t)| ≥ δ (34.26)
(i−1)/n≤t≤i/n
i=1
 
= nP max |X (t) − Yn, (t)| ≥ δ (34.27)
0≤t≤1/n
 
= nP max |W (t) − nW (1/n)| ≥ δ (34.28)
0≤t≤1/n
 
δ
≤ nP max |W (t)| ≥ (34.29)
0≤t≤1/n 2
 
δ
≤ 4dnP W1 (1/n) ≥ (34.30)
2d
2d − nδ2 22
≤ 4dn √ e 8d  (34.31)
δ 2πn


again freely using our calculations from Ch. 21. If n > 4d2 j/δ 2 , then P kX − Yn, k∞ ≤
j−γ
1 − 2
2e , and we have overall
j−γ
P (X 6∈ U (j, δ)) ≤ e− 2 (34.32)

as required. 

Proposition 506 (Compact Level Sets of the Wiener Action) The Cameron-
Martin norm has compact level sets.

Proof: See Kallenberg, Lemma 27.7, p. 543. 

Theorem 507 (Schilder’s Theorem on Large Deviations of the Wiener

Process on the Unit Interval) If W is a d-dimensional Wiener process on
the unit interval, then X = W obeys an LDP on C([0, 1], Rd ), with rate −2
and good rate function J1 (x), the effective Wiener action over [0, 1].

Proof: It is easy to show that Lemma 504 implies the large deviation lower
bound for open sets. (Exercise 34.2.) The tricky part is the upper bound. Pick
any closed set C and any γ > 0. Let s = J1 (C) − γ. By Lemma 506, the set
CHAPTER 34. FREIDLIN-WENTZELL THEORY 281

K = L1 (s) = {x : J1 (x) ≤ s} is compact. By construction, C ∩ K = ∅. So

δ = inf x∈C,y∈K kx − yk∞ > 0. Let U be the closed δ-neighborhood of K. Then
use Lemma 505

P (X ∈ C) ≤ P (X 6∈ U ) (34.33)

− s−γ
≤ e 2 (34.34)
J (C)−2γ
− 1 2
≤ e (34.35)
J1 (C) − 2γ
log P (X ∈ C) ≤ − (34.36)
2
2
 log P (X ∈ C) ≤ −J1 (C) − 2γ (34.37)
lim sup 2 log P (X ∈ C) ≤ −J1 (C) − 2γ (34.38)
→0

Since γ was arbitrary, this completes the proof. 

Remark: The trick used here, about establishing results like Lemmas 505
and 504, and then using compact level sets to prove large deviations, works
more generally. See Theorem 3.3 in Freidlin and Wentzell (1998, sec. 3.3).

Corollary 508 (Extension of Schilder’s Theorem to [0, T ]) Schilder’s the-

orem remains true for Wiener processes on [0, T ], for all T > 0, with rate
function JT , the effective Wiener action on [0, T ].

√ Proof: If W is a Wiener process on [0, 1], then, for every T , S(W ) =

T W (t/T ) is a Wiener process on [0, T ]. (Show this!) Since the mapping S
is continuous from C([0, 1], Rd ) to C([0, T ], Rd ), by the Contraction Principle
(Theorem 451) the family S(W ) obey an LDP with rate −2 and good rate
function JT (x) = J1 (S −1 (x)). (Notice that S is invertible, so S −1 (x) is a
function, not a set of functions.) Since x ∈ H1 iff y = S(x) ∈ HT , it’s easy to
check that for such, ẏ(t) = T −1/2 ẋ(t/T ), meaning that
Z T Z T
2 2 dt
kykCM = |ẏ(t)| dt = |ẋ(t/T )| = kxkCM (34.39)
0 0 T
which completes the proof. 

Corollary 509 (Schilder’s Theorem on R+ ) Schilder’s theorem remains

true for Wiener processes on R+ , with good rate function J∞ given by the ef-
fective Wiener action on R+ ,
1 ∞
Z
2
J∞ (x) ≡ |ẋ(t)| dt (34.40)
2 0
if x ∈ H∞ , J∞ (x) = ∞ otherwise.

Proof: For each natural number n, let πn x be the restriction of x to the

interval [0, n]. By Corollary 508, each of them obeys an LDP with rate function
1 n 2
R
2 0 |ẋ(t)| dt. Now apply the projective limit theorem (461) to get that J∞ (x) =
supn Jn (x), which is clearly Eq. 34.40, as the integrand is non-negative. 
CHAPTER 34. FREIDLIN-WENTZELL THEORY 282

34.2 Large Deviations for SDEs with State-Independent

Noise
Having established an LDP for the Wiener process, it is fairly straightforward
to get an LDP for stochastic differential equations where the driving noise is
independent of the state of the diffusion process.
Definition 510 (SDE with Small State-Independent Noise) An SDE
with small state-independent noise is a stochastic differential equation of the
form
dX = a(X )dt + dW (34.41)
X (0) = 0 (34.42)
where a : Rd 7→ Rd is uniformly Lipschitz continuous.
Notice that any non-random initial condition x0 can be handled by a simple
change of coordinates.
Definition 511 (Effective Action under State-Independent Noise) The
effective action of a trajectory x ∈ H∞ is
1 t
Z
2
J(x) ≡ |ẋ(t) − a(x(t))| dt (34.43)
2 0
and = ∞ if x ∈ C \ H∞ .

Lemma 512 (Continuous Mapping from Wiener Process to SDE So-

lutions) The map F : C(R+ , Rd ) 7→ C(R+ , Rd ) given by
Z t
x(t) = w(t) + a(x(s))ds (34.44)
0

when x = F (w) is continuous.

Proof: This goes rather in the same manner as the proof of existence
and uniqueness for SDEs (Theorem 260). For any w1 , w2 ∈ C(R+ , Rd ), set
x1 = F (w1 ), x2 = F (w2 ). From the Lipschitz property of a,
Z t
|x1 (t) − x2 (t)| ≤ kw1 − w2 k + Ka |x1 (s) − x2 (s)| ds (34.45)
0

(writing |y(t)| for the norm of Euclidean vectors y, and kxk for the supremum
norm of continuous curves). By Gronwall’s Inequality (Lemma 258), then,
kx1 − x2 k ≤ kw1 − w2 k eKa T (34.46)
on every interval [0, T ]. So we can make sure that kx1 − x2 k is less than any
desired amount by making sure that kw1 − w2 k is sufficiently small, and so F
is continuous. 
CHAPTER 34. FREIDLIN-WENTZELL THEORY 283

Lemma 513 (Mapping Cameron-Martin Spaces Into Themselves) If

w ∈ H∞ , then x = F (w) is in H∞ .

Proof: Exercise 34.3. 

Theorem 514 (Freidlin-Wentzell Theorem with State-Independent Noise)

The Itô processes X of Definition 510 obey the large deviations principle with
rate −2 and good rate function given by the effective action J(x).

Proof: For every , X = F (W ). Corollary 509 tells us that W obeys

the large deviation principle with rate −2 and good rate function J∞ . Since
(Lemma 512) F is continuous, by the Contraction Principle (Theorem 451) X
also obeys the LDP, with rate given by J∞ (F −1 (x)). If F −1 (x) ∩ H∞ = ∅,
this is ∞. On the other hand, if F −1 (x) does contain curves in H∞ , then
J∞ (F −1 (x)) = J∞ (F −1 (x) ∩ H∞ ). By Lemma 513, this implies that x ∈ H∞ ,
too. For Rany curve w ∈ F −1 (x) ∩ H∞ , ẋ = ẇ + a(x), or ẇ = ẋ − a(x).
∞ 2
J∞ (w) = 0 |ẋ − a(x)| dt is however the effective action of the trajectory x
(Definition 511). 

34.3 Large Deviations for State-Dependent Noise

If the diffusion term in the SDE does depend on the state of the process, one
obtains a very similar LDP to the results in the previous section. However, the
approach must be modified: the mapping from W to X , while still measurable,
is no longer necessarily continuous, so we can’t use the contraction principle as
before.

Definition 515 (SDE with Small State-Dependent Noise) An SDE with

small state-dependent noise is a stochastic differential equation of the form

dX = a(X )dt + b(X )dW (34.47)

X (0) = 0 (34.48)

where a and b are uniformly Lipschitz continuous, and b is non-singular.

Definition 516 (Effective Action under State-Dependent Noise) The

effective action of a trajectory x ∈ H∞ is given by
Z ∞
J(x) ≡ L(x(t), ẋ(t))dt (34.49)
0

where
1 −1
L(q, p) = (pi − ai (q)) Bij (q) (pj − aj (q)) (34.50)
2
and
B(q) = b(q)bT (q) (34.51)
with J(x) = ∞ if x ∈ C \ H∞ .
CHAPTER 34. FREIDLIN-WENTZELL THEORY 284

Theorem 517 (Freidlin-Wentzell Theorem for State-Dependent Noise)

The processes X obey a large deviations principle with rate −2 and good rate
function equal to the effective action.

Proof: Considerably more complicated. (See, e.g., Dembo and Zeitouni

(1998, sec. 5.6, pp. 213–220).) The essence, however, is to consider an approx-
imating Itô process Xn , where a(Xt ) and b(Xt ) are replaced in Eq. 34.47 by
a(Xn ([tn]/n)) and b(Xn ([tn]/n)). Here the mapping from W to Xn is continu-
ous, so it’s not too hard to show that the latter obey an LDP with a reasonable
rate function, and also that they’re exponentially equivalent (in n) to X . 

34.4 Exercises
Exercise 34.1 (Cameron-Martin Spaces are Hilbert Spaces) Prove Lemma
501.

Exercise 34.2 (Lower Bound in Schilder’s Theorem) Prove that Lemma

504 implies the large deviations lower bound for open sets.

Exercise 34.3 (Mapping Cameron-Martin Spaces Into Themselves)

Prove Lemma 513. Hint: Use Gronwall’s Inequality (Lemma
T∞ 258) again to
show that F maps HT into HT , and then show that H∞ = n=1 Hn .
Bibliography

Abramowitz, Milton and Irene A. Stegun (eds.) (1964). Handbook of Mathe-

matical Functions. Washington, D.C.: National Bureau of Standards. URL
http://www.math.sfu.ca/∼cbm/aands/.
Algoet, Paul (1992). “Universal Schemes for Prediction, Gambling
and Portfolio Selection.” The Annals of Probability, 20: 901–
941. URL http://links.jstor.org/sici?sici=0091-1798%28199204%
2920%3A2%3C901%3AUSFPGA%3E2.0.CO%3B2-J. Correction, The Annals of
Probability, 23 (1995): 474–478, JSTOR.
Algoet, Paul H. and Thomas M. Cover (1988). “A Sandwich Proof of the
Shannon-McMillan-Breiman Theorem.” The Annals of Probability, 16: 899–
909. URL http://links.jstor.org/sici?sici=0091-1798%28198804%
2916%3A2%3C899%3AASPOTS%3E2.0.CO%3B2-M.
Amari, Shun-ichi and Hiroshi Nagaoka (1993/2000). Methods of Informa-
tion Geometry. Providence, Rhode Island: American Mathematical Society.
Translated by Daishi Harada. As Joho Kika no Hoho, Tokyo: Iwanami Shoten
Publishers.
Arnol’d, V. I. (1973). Ordinary Differential Equations. Cambridge, Mas-
sachusetts: MIT Press. Trans. Richard A. Silverman from Obyknovennye
differentsial’nye Uravneniya.
— (1978). Mathematical Methods of Classical Mechanics. Berlin: Springer-
Verlag. First published as Matematicheskie metody klassicheskoĭ mekhaniki,
Moscow: Nauka, 1974.
Arnol’d, V. I. and A. Avez (1968). Ergodic Problems of Classical Mechanics.
Mathematical Physics Monograph Series. New York: W. A. Benjamin.
Badino, Massimiliano (2004). “An Application of Information Theory to the
Problem of the Scientific Experiment.” Synthese, 140: 355–389. URL http:
//philsci-archive.pitt.edu/archive/00001830/.

Baldi, Pierre and Søren Brunak (2001). Bioinformatics: The Machine Learning
Approach. Cambridge, Massachusetts: MIT Press, 2nd edn.

285
BIBLIOGRAPHY 286

Banks, J., J. Brooks, G. Cairns, G. Davis and P. Stacy (1992). “On Devaney’s
Definition of Chaos.” American Mathematical Monthly, 99: 332–334.
Bartlett, M. S. (1955). An Introduction to Stochastic Processes, with Special
Reference to Methods and Applications. Cambridge, England: Cambridge
University Press.

Basharin, Gely P., Amy N. Langville and Valeriy A. Naumov (2004). “The
Life and Work of A. A. Markov.” Linear Algebra and its Applica-
tions, 386: 3–26. URL http://decision.csl.uiuc.edu/∼meyn/pages/
Markov-Work-and-life.pdf.
Biggers, Earl Derr (1928). Behind That Curtain. New York: Grosset and
Dunlap.
Billingsley, Patrick (1961). Statistical Inference for Markov Processes. Chicago:
University of Chicago Press.
Blackwell, David and M. A. Girshick (1954). Theory of Games and Statistical
Decisions. New York: Wiley.
Bosq, Denis (1998). Nonparametric Statistics for Stochastic Processes: Estima-
tion and Prediction. Berlin: Springer-Verlag, 2nd edn.
Caires, S. and J. A. Ferreira (2005). “On the Non-parametric Prediction of
Conditionally Stationary Sequences.” Statistical Inference for Stochastic Pro-
cesses, 8: 151–184. Correction, vol. 9 (2006), pp. 109–110.
Charniak, Eugene (1993). Statistical Language Learning. Cambridge, Mas-
sachusetts: MIT Press.
Chomsky, Noam (1957). Syntactic Structures. The Hauge: Mouton.
Courant, Richard and David Hilbert (1953). Methods of Mathematical Physics.
New York: Wiley.
Cover, Thomas M. and Joy A. Thomas (1991). Elements of Information Theory.
New York: Wiley.

Cramér, Harald (1945). Mathematical Methods of Statistics. Uppsala: Almqvist

and Wiksells.
Dembo, Amir and Ofer Zeitouni (1998). Large Deviations Techniques and Ap-
plications. New York: Springer Verlag, 2nd edn.
Devaney, Robert L. (1992). A First Course in Chaotic Dynamical Systems:
Theory and Experiment. Reading, Mass.: Addison-Wesley.
Doob, Joseph L. (1953). Stochastic Processes. Wiley Publications in Statistics.
New York: Wiley.
BIBLIOGRAPHY 287

Doukhan, Paul (1995). Mixing: Properties and Examples. New York: Springer-
Verlag.
Durbin, James and Siem Jam Koopman (2001). Time Series Analysis by State
Space Methods. Oxford: Oxford University Press.

Durrett, Richard (1991). Probability: Theory and Examples. Belmont, Califor-

nia: Duxbury.
Dynkin, E. B. (1978). “Sufficient statistics and extreme points.” Annals
of Probability, 6: 705–730. URL http://links.jstor.org/sici?sici=
0091-1798%28197810%296%3A5%3C705%3ASSAEP%3E2.0.CO%3B2-D.

Eckmann, Jean-Pierre and David Ruelle (1985). “Ergodic Theory of Chaos and
Strange Attractors.” Reviews of Modern Physics, 57: 617–656.
Eichinger, L., J. A. Pachebat, G. Glockner et al. (2005). “The genome of the
social amoeba Dictyostelium discoideum.” Nature, 435: 43–57.
Ellis, Richard S. (1985). Entropy, Large Deviations, and Statistical Mechanics.
Berlin: Springer-Verlag.
— (1988). “Large Deviations for the Empirical Measure of a Markov Chain
with an Application to the Multivariate Empirical Measure.” The Annals
of Probability, 16: 1496–1508. URL http://links.jstor.org/sici?sici=
0091-1798%28198810%2916%3A4%3C1496%3ALDFTEM%3E2.0.CO%3B2-S.
Embrechts, Paul and Makoto Maejima (2002). Selfsimilar processes. Princeton,
New Jersey: Princeton University Press.
Ethier, Stewart N. and Thomas G. Kurtz (1986). Markov Processes: Charac-
terization and Convergence. New York: Wiley.
Eyink, Gregory L. (1996). “Action principle in nonequilbrium statistical dy-
namics.” Physical Review E , 54: 3419–3435.
Fisher, Ronald Aylmer (1958). The Genetical Theory of Natural Selection. New
York: Dover, 2nd edn. First edition published Oxford: Clarendon Press, 1930.
Forster, Dieter (1975). Hydrodynamic Fluctuations, Broken Symmetry, and
Correlation Functions. Reading, Massachusetts: Benjamin Cummings.
Freidlin, M. I. and A. D. Wentzell (1998). Random Perturbations of Dynami-
cal Systems. Berlin: Springer-Verlag, 2nd edn. First edition first published
as Fluktuatsii v dinamicheskikh sistemakh pod deistviem malykh sluchainykh
vozmushchenii, Moscow: Nauka, 1979.
Frisch, Uriel (1995). Turbulence: The Legacy of A. N. Kolmogorov . Cambridge,
England: Cambridge University Press.
BIBLIOGRAPHY 288

Fristedt, Bert and Lawrence Gray (1997). A Modern Approach to Probability

Theory. Probability Theory and Its Applications. Boston: Birkhäuser.
Gikhman, I. I. and A. V. Skorokhod (1965/1969). Introduction to the Theory
of Random Processes. Philadelphia: W. B. Saunders. Trans. Richard Silver-
man from Vvedenie v teoriiu slucainikh protessov, Moscow: Nauka; reprinted
Mineola, New York: Dover, 1996.
Gillespie, John H. (1998). Population Genetics: A Concise Guide. Baltimore:
Johns Hopkins University Press.
Gnedenko, B. V. and A. N. Kolmogorov (1954). Limit Distributions for Sums of
Independent Random Variables. Cambridge, Massachusetts: Addison-Wesley.
Translated from the Russian and annotated by K. L. Chung, with an Ap-
pendix by J. L. Doob.
Gray, Robert M. (1988). Probability, Random Processes, and Ergodic Properties.
New York: Springer-Verlag. URL http://ee-www.stanford.edu/∼gray/
arp.html.
— (1990). Entropy and Information Theory. New York: Springer-Verlag. URL
http://www-ee.stanford.edu/∼gray/it.html.
Haldane, J. B. S. (1932). The Causes of Evolution. Longmans, Green and Co.
Reprinted Princeton: Princeton University Press, 1990.
Halmos, Paul R. (1950). Measure Theory. New York: Van Nostrand.
Hille, Einer (1948). Functional Analysis and Semi-Groups. New York: American
Mathematical Society.
Howard, Ronald A. (1971). Markov Models, vol. 1 of Dynamic Probabilistic
Systems. New York: Wiley.
Ibragimov, I. A. and R. Z. Has’minskii (1979/1981). Statistical Estimation:
Asymptotic Theory. New York: Springer-Verlag. Tranlated by Samuel Kotz
from Asimptoticheskaia teoriia ostsenivaniia, Moscow: Nauka.
Jakubowski, A. and Z. S. Szewczak (1990). “A normal convergence criterion
for strongly mixing stationary sequences.” In Limit Theorems in Probabil-
ity and Statistics (I. Berkes and Endre Csáki and Pal Révesz, eds.), vol. 57
of Colloquia mathematica Societatis Janos Bolyai , pp. 281–292. New York:
North-Holland. Proceedings of the Third Hungarian Colloquium on Limit
Theorems in Probability and Statistics, held in Pécs, Hungary, July 3-7, 1989
and sponsored by the Bolyai János Mathematical Society.
Kac, Mark (1947). “On the Notion of Recurrence in Discrete Stochastic Pro-
cesses.” Bulletin of the American Mathematical Society, 53: 1002–1010.
Reprinted in Kac (1979), pp. 231–239.
BIBLIOGRAPHY 289

— (1979). Probability, Number Theory, and Statistical Physics: Selected Papers.

Cambridge, Massachusetts: MIT Press. Edited by K. Baclawski and M. D.
Donsker.
Kallenberg, Olav (2002). Foundations of Modern Probability. New York:
Springer-Verlag, 2nd edn.

Kass, Robert E. and Paul W. Vos (1997). Geometrical Foundations of Asymp-

totic Inference. Wiley Series in Probability and Statistics. New York: Wiley.
Katznelson, I. and B. Weiss (1982). “A simple proof of some ergodic theorems.”
Israel Journal of Mathematics, 42: 291–296.

Keizer, Joel (1987). Statistical Thermodynamics of Nonequilibrium Processes.

New York: Springer-Verlag.
Khinchin, Aleksandr Iakovlevich (1949). Mathematical Foundations of Statisti-
cal Mechanics. New York: Dover Publications. Translated from the Russian
by G. Gamow.
Knight, Frank B. (1975). “A Predictive View of Continuous Time Processes.”
Annals of Probability, 3: 573–596. URL http://links.jstor.org/sici?
sici=0091-1798%28197508%293%3A4%3C573%3AAPVOCT%3E2.0.CO%3B2-N.
— (1992). Foundations of the Prediction Process. Oxford: Clarendon Press.
Kolmogorov, A. N. and S. V. Fomin (1975). Introductory Real Analysis. New
York: Dover. Translated and edited by Richard A. Silverman.
Kontoyiannis, I., P. H. Algoet, Yu. M. Suhov and A. J. Wyner (1998). “Non-
parametric entropy estimation for stationary processes and random fields,
with applications to English text.” IEEE Transactions on Information The-
ory, 44: 1319–1327. URL http://www.dam.brown.edu/people/yiannis/
PAPERS/suhov2.pdf.
Küchler, Uwe and Michael Sørensen (1997). Exponential Families of Stochastic
Processes. Berlin: Springer-Verlag.
Kulhavý, Rudolf (1996). Recursive Nonlinear Estimation: A Geometric Ap-
proach, vol. 216 of Lecture Notes in Control and Information Sciences. Berlin:
Springer-Verlag.
Kullback, Solomon (1968). Information Theory and Statistics. New York: Dover
Books, 2nd edn.

Kurtz, Thomas G. (1970). “Solutions of Ordinary Differential Equations as

Limits of Pure Jump Markov Processes.” Journal of Applied Probability, 7:
49–58. URL http://links.jstor.org/sici?sici=0021-9002%28197004%
297%3A1%3C49%3ASOODEA%3E2.0.CO%3B2-P.
BIBLIOGRAPHY 290

— (1971). “Limit Theorems for Sequences of Jump Markov Processes Ap-

proximating Ordinary Differential Processes.” Journal of Applied Probabil-
ity, 8: 344–356. URL http://links.jstor.org/sici?sici=0021-9002%
28197106%298%3A2%3C344%3ALTFSOJ%3E2.0.CO%3B2-M.
— (1975). “Semigroups of Conditioned Shifts and Approximation of
Markov Processes.” The Annals of Probability, 3: 618–642. URL
http://links.jstor.org/sici?sici=0091-1798%28197508%293%3A4%
3C618%3ASOCSAA%3E2.0.CO%3B2-C.
Kushner, Harold J. (1984). Approximation and Weak Convergence Methods for
Random Processes, with Applications to Stochastic Systems Theory. Cam-
bridge, Massachusetts: MIT Press.
Lamperti, John (1962). “Semi-Stable Stochastic Processes.” Trans-
actions of the American Mathematical Society, 104: 62–78. URL
http://links.jstor.org/sici?sici=0002-9947%28196207%29104%3A1%
3C62%3ASSP%3E2.0.CO%3B2-I.
Lasota, Andrzej and Michael C. Mackey (1994). Chaos, Fractals, and Noise:
Stochastic Aspects of Dynamics. Berlin: Springer-Verlag. First edition, Prob-
abilistic Properties of Deterministic Systems, Cambridge University Press,
1985.
Lehmann, E. L. and George Casella (1998). Theory of Point Estimation.
Springer Texts in Statistics. Berlin: Springer-Verlag, 2nd edn.
Liptser, Robert S. and Albert N. Shiryaev (2001). Statistics of Random Pro-
cesses. Berlin: Springer-Verlag, 2nd edn. Two volumes. Trans. A. B. Aries.
First English edition 1977–1978. First published as Statistika sluchaĭnykh
protessov, Moscow: Nauka, 1974.
Loève, Michel (1955). Probability Theory. New York: D. Van Nostrand Com-
pany, 1st edn.
Mackey, Michael C. (1992). Time’s Arrow: The Origins of Thermodynamic
Behavior . Berlin: Springer-Verlag.
Moretti, Franco (2005). Graphs, Maps, Trees: Abstract Models for a Literary
History. London: Verso.
Øksendal, Bernt (1995). Stochastic Differential Equations: An Introduction with
Applications. Berlin: Springer-Verlag, 4th edn.
Olivieri, Enzo and Maria Eulália Vares (2005). Large Deviations and Metasta-
bility. Cambridge, England: Cambridge University Press.
Ornstein, Donald S. and Benjamin Weiss (1990). “How Sampling
Reveals a Process.” Annals of Probability, 18: 905–930. URL
http://links.jstor.org/sici?sici=0091-1798%28199007%2918%3A3%
3C905%3AHSRAP%3E2.0.CO%3B2-J.
BIBLIOGRAPHY 291

Pearl, Judea (1988). Probabilistic Reasoning in Intelligent Systems. New York:

Morgan Kaufmann.
Peres, Yuval and Paul Shields (2005). “Two New Markov Order Estimators.”
E-print, arxiv.org. URL http://arxiv.org/abs/math.ST/0506080.

Pollard, David (2002). A User’s Guide to Measure Theoretic Probability. Cam-

bridge, England: Cambridge University Press.
Pullum, Geoffrey K. (1991). The Great Eskimo Vocabulary Hoax, and Other
Irreverent Essays on the Study of Language. Chicago: University of Chicago
Press.

Rabiner, Lawrence R. (1989). “A Tutorial on Hidden Markov Models and Se-

lected Applications in Speech Recognition.” Proceedings of the IEEE , 77:
257–286. URL http://dx.doi.org/10.1109/5.18626.
Rockafellar, R. T. (1970). Convex Analysis. Princeton: Princeton University
Press.
Rogers, L. C. G. and David Williams (1994). Foundations, vol. 1 of Diffu-
sions, Markov Processes and Martingales. New York: John Wiley, 2nd edn.
Reprinted Cambridge, England: Cambridge University Press, 2000.
— (2000). Itô Calculus, vol. 2 of Diffusions, Markov Processes and Martingales.
Cambridge, England: Cambridge University Press, 2nd edn.
Rosenblatt, Murray (1956). “A Central Limit Theorem and a Strong Mixing
Condition.” Proceedings of the National Academy of Sciences (USA), 42:
43–47. URL http://www.pnas.org/cgi/reprint/42/1/43.
— (1971). Markov Processes. Structure and Asymptotic Behavior . Berlin:
Springer-Verlag.
Schroeder, Manfred (1991). Fractals, Chaos, Power Laws: Minutes from an
Infinite Paradise. San Francisco: W. H. Freeman.
Selmeczi, David, Simon F. Tolic-Norrelykke, Erik Schaeffer, Peter H. Hage-
dorn, Stephan Mosler, Kirstine Berg-Sorensen, Niels B. Larsen and Henrik
Flyvbjerg (2006). “Brownian Motion after Einstein: Some new applica-
tions and new experiments.” In Controlled Nanoscale Motion in Biological
and Artificial Systems (H. Linke and A. Mansson, eds.), vol. 131 of Pro-
ceedings of the Nobel Symposium, p. forthcoming. Berlin: Springer. URL
http://arxiv.org/abs/physics/0603142.

Shalizi, Cosma Rohilla and James P. Crutchfield (2001). “Computational Me-

chanics: Pattern and Prediction, Structure and Simplicity.” Journal of Sta-
tistical Physics, 104: 817–879. URL http://arxiv.org/abs/cond-mat/
9907176.
BIBLIOGRAPHY 292

Shannon, Claude E. (1948). “A Mathematical Theory of Communication.” Bell

System Technical Journal , 27: 379–423. URL http://cm.bell-labs.com/
cm/ms/what/shannonday/paper.html. Reprinted in Shannon and Weaver
(1963).
Shannon, Claude E. and Warren Weaver (1963). The Mathematical Theory of
Communication. Urbana, Illinois: University of Illinois Press.
Shiryaev, Albert N. (1999). Essentials of Stochastic Finance: Facts, Models,
Theory. Singapore: World Scientific. Trans. N. Kruzhilin.
Spirtes, Peter, Clark Glymour and Richard Scheines (2001). Causation, Predic-
tion, and Search. Cambridge, Massachusetts: MIT Press, 2nd edn.
Taniguchi, Masanobu and Yoshihide Kakizawa (2000). Asymptotic Theory of
Statistical Inference for Time Series. Berlin: Springer-Verlag.
Tuckwell, Henry C. (1989). Stochastic Processes in the Neurosciences. Philadel-
phia: SIAM.
Tyran-Kamińska, Marta (2005). “An Invariance Principle for Maps with Poly-
nomial Decay of Correlations.” Communications in Mathematical Physics,
260: 1–15. URL http://arxiv.org/abs/math.DS/0408185.
von Plato, Jan (1994). Creating Modern Probability: Its Mathematics, Physics
and Philosophy in Historical Perspective. Cambridge, England: Cambridge
University Press.
Wax, Nelson (ed.) (1954). Selected Papers on Noise and Stochastic Processes.
Dover.
Wiener, Norbert (1949). Extrapolation, Interpolation, and Smoothing of Station-
ary Time Series: With Engineering Applications. Cambridge, Massachusetts:
The Technology Press of the Massachusetts Institute of Technology.
— (1958). Nonlinear Problems in Random Theory. Cambridge, Massachusetts:
The Technology Press of the Massachusetts Institute of Technology.
— (1961). Cybernetics: Or, Control and Communication in the Animal and the
Machine. Cambridge, Massachusetts: MIT Press, 2nd edn. First edition New
York: Wiley, 1948.
Wu, Wei Biao (2005). “Nonlinear system theory: Another look at dependence.”
Proceedings of the National Academy of Sciences (USA), 102: 14150–14154.