Eberle Introduction To Stochastic Analysis

Introduction to Stochastic Analysis
Andreas Eberle
February 23, 2016

Contents
Contents 2
I Stochastic Processes 13
1 Brownian Motion 14
1.1 From Random Walks to Brownian Motion . . . . . . . . . . . . . . . . 15
Central Limit Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Brownian motion as a Lévy process. . . . . . . . . . . . . . . . . . . . 20
Brownian motion as a Markov process. . . . . . . . . . . . . . . . . . . 21
Wiener Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.2 Brownian Motion as a Gaussian Process . . . . . . . . . . . . . . . . . 27
Multivariate normals . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Gaussian processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.3 The Wiener-Lévy Construction . . . . . . . . . . . . . . . . . . . . . . 38
A first attempt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
The Wiener-Lévy representation of Brownian motion . . . . . . . . . . 41
Lévy’s construction of Brownian motion . . . . . . . . . . . . . . . . . 47
1.4 The Brownian Sample Paths . . . . . . . . . . . . . . . . . . . . . . . 52
Typical Brownian sample paths are nowhere differentiable . . . . . . . 52
Hölder continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
Law of the iterated logarithm . . . . . . . . . . . . . . . . . . . . . . . 56
Passage times . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
1.5 Strong Markov property and reflection principle . . . . . . . . . . . . . 60
2
CONTENTS 3
Maximum of Brownian motion . . . . . . . . . . . . . . . . . . . . . . 60

Strong Markov property for Brownian motion . . . . . . . . . . . . . . 63
A rigorous reflection principle . . . . . . . . . . . . . . . . . . . . . . 65
II Introduction to Stochastic Analysis 67
2 Martingales in discrete time 68

2.1 Definitions and examples . . . . . . . . . . . . . . . . . . . . . . . . . 68
Martingales and supermartingales . . . . . . . . . . . . . . . . . . . . 69
Some fundamental examples . . . . . . . . . . . . . . . . . . . . . . . 71
2.2 Doob Decomposition and Martingale Problem . . . . . . . . . . . . . . 78
Doob Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Conditional Variance Process . . . . . . . . . . . . . . . . . . . . . . . 79
Martingale problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
2.3 Gambling strategies and stopping times . . . . . . . . . . . . . . . . . 85
Martingale transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
Stopped Martingales . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Optional Stopping Theorems . . . . . . . . . . . . . . . . . . . . . . . 95
Wald’s identity for random sums . . . . . . . . . . . . . . . . . . . . . 99
2.4 Maximal inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
Doob’s inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
Lp inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
Hoeffding’s inequality . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3 Martingales in continuous time 107

3.1 Some fundamental martingales of Brownian Motion . . . . . . . . . . . 108
Filtrations generated by Brownian motion . . . . . . . . . . . . . . . . 108
Brownian Martingales . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
3.2 Optional Sampling and Optional Stopping . . . . . . . . . . . . . . . . 112
The Optional Sampling Theorem . . . . . . . . . . . . . . . . . . . . . 112
Ruin probabilities and passage times revisited . . . . . . . . . . . . . . 115
Exit laws and Dirichlet problem . . . . . . . . . . . . . . . . . . . . . 116
University of Bonn 2015/2016

4 CONTENTS
3.3 Maximal inequalities and the LIL . . . . . . . . . . . . . . . . . . . . . 118

Maximal inequalities in continuous time . . . . . . . . . . . . . . . . . 118
Application to LIL . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
4 Martingale Convergence Theorems 125

4.1 Convergence in L 2
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Martingales in L2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
The Convergence Theorem . . . . . . . . . . . . . . . . . . . . . . . . 126
Summability of sequences with random signs . . . . . . . . . . . . . . 128
L2 convergence in continuous time . . . . . . . . . . . . . . . . . . . . 129
4.2 Almost sure convergence of supermartingales . . . . . . . . . . . . . . 130
Doob’s upcrossing inequality . . . . . . . . . . . . . . . . . . . . . . . 131
Proof of Doob’s Convergence Theorem . . . . . . . . . . . . . . . . . 133
Examples and first applications . . . . . . . . . . . . . . . . . . . . . . 134
Generalized Borel-Cantelli Lemma . . . . . . . . . . . . . . . . . . . . 138
Upcrossing inequality and convergence theorem in continuous time . . 140
4.3 Uniform integrability and L1 convergence . . . . . . . . . . . . . . . . 141
Uniform integrability . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
Definitive version of Lebesgue’s Dominated Convergence Theorem . . 145
L1 convergence of martingales . . . . . . . . . . . . . . . . . . . . . . 147
Backward Martingale Convergence . . . . . . . . . . . . . . . . . . . . 150
5 Stochastic Integration w.r.t. Continuous Martingales 152

5.1 Defining stochastic integrals: A first attempt . . . . . . . . . . . . . . . 154
Riemann sum approximations . . . . . . . . . . . . . . . . . . . . . . 154
Itô integrals for continuous bounded integrands . . . . . . . . . . . . . 156
The Hilbert space M2c . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
Definition of Itô integral in M2c . . . . . . . . . . . . . . . . . . . . . . 161
5.2 Itô’s isometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
Predictable step functions . . . . . . . . . . . . . . . . . . . . . . . . . 162
Itô’s isometry for Brownian motion . . . . . . . . . . . . . . . . . . . . 165
Itô’s isometry for martingales . . . . . . . . . . . . . . . . . . . . . . . 167
Stochastic Analysis Andreas Eberle

CONTENTS 5
Definition of Itô integrals for square-integrable integrands . . . . . . . . 169

Identification of admissible integrands . . . . . . . . . . . . . . . . . . 171
5.3 Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
Local dependence on integrand and integrator . . . . . . . . . . . . . . 175
Itô integrals for locally square-integrable integrands . . . . . . . . . . . 176
Stochastic integrals as local martingales . . . . . . . . . . . . . . . . . 178
Approximation by Riemann-Itô sums . . . . . . . . . . . . . . . . . . 181
6 Itô’s formula and pathwise integrals 183

6.1 Stieltjes integrals and chain rule . . . . . . . . . . . . . . . . . . . . . 185
Lebesgue-Stieltjes integrals . . . . . . . . . . . . . . . . . . . . . . . . 185
The chain rule in Stieltjes calculus . . . . . . . . . . . . . . . . . . . . 188
6.2 Quadratic variation and Itô’s formula . . . . . . . . . . . . . . . . . . . 191
Quadratic variation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
Itô’s formula and pathwise integrals in R1 . . . . . . . . . . . . . . . . 194
The chain rule for anticipative integrals . . . . . . . . . . . . . . . . . 197
6.3 Itô’s formula for Brownian motion and martingales . . . . . . . . . . . 198
Quadratic variation of Brownian motion . . . . . . . . . . . . . . . . . 199
Itô’s formula for Brownian motion . . . . . . . . . . . . . . . . . . . . 201
Quadratic variation of continuous martingales . . . . . . . . . . . . . . 204
From continuous martingales to Brownian motion . . . . . . . . . . . . 207
6.4 Multivariate and time-dependent Itô formula . . . . . . . . . . . . . . . 209
Covariation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
Itô to Stratonovich conversion . . . . . . . . . . . . . . . . . . . . . . 211
Itô’s formula in Rd . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
Product rule, integration by parts . . . . . . . . . . . . . . . . . . . . . 214
Time-dependent Itô formula . . . . . . . . . . . . . . . . . . . . . . . 216
7 Brownian Motion and PDE 221

7.1 Dirichlet problem, recurrence and transience . . . . . . . . . . . . . . . 222
The Dirichlet problem revisited . . . . . . . . . . . . . . . . . . . . . . 222
Recurrence and transience of Brownian motion in Rd . . . . . . . . . . 223

6 CONTENTS
7.2 Boundary value problems, exit and occupation times . . . . . . . . . . 227

The stationary Feynman-Kac-Poisson formula . . . . . . . . . . . . . . 227
Poisson problem and mean exit time . . . . . . . . . . . . . . . . . . . 230
Occupation time density and Green function . . . . . . . . . . . . . . . 231
Stationary Feynman-Kac formula and exit time distributions . . . . . . 232
Boundary value problems in Rd and total occupation time . . . . . . . . 234
7.3 Heat equation and time-dependent FK formula . . . . . . . . . . . . . . 236
Brownian Motion with Absorption . . . . . . . . . . . . . . . . . . . . 237
Time-dependent Feynman-Kac formula . . . . . . . . . . . . . . . . . 240
Occupation times and arc-sine law . . . . . . . . . . . . . . . . . . . . 242
8 SDE: Explicit Computations 245

8.1 Stochastic Calculus for Itô processes . . . . . . . . . . . . . . . . . . . 248
Stochastic integrals w.r.t. Itô processes . . . . . . . . . . . . . . . . . . 249
Calculus for Itô processes . . . . . . . . . . . . . . . . . . . . . . . . . 252
The Itô-Doeblin formula in R1 . . . . . . . . . . . . . . . . . . . . . . 255
Martingale problem for solutions of SDE . . . . . . . . . . . . . . . . 257
8.2 Stochastic growth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
Scale functions and exit distributions . . . . . . . . . . . . . . . . . . . 258
Recurrence and asymptotics . . . . . . . . . . . . . . . . . . . . . . . 259
Geometric Brownian motion . . . . . . . . . . . . . . . . . . . . . . . 261
Feller’s branching diffusion . . . . . . . . . . . . . . . . . . . . . . . . 262
Cox-Ingersoll-Ross model . . . . . . . . . . . . . . . . . . . . . . . . 264
8.3 Linear SDE with additive noise . . . . . . . . . . . . . . . . . . . . . . 265
Variation of constants . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
Solutions as Gaussian processes . . . . . . . . . . . . . . . . . . . . . 267
The Ornstein-Uhlenbeck process . . . . . . . . . . . . . . . . . . . . . 269
Change of time-scale . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
8.4 Brownian bridge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
Wiener-Lévy construction . . . . . . . . . . . . . . . . . . . . . . . . 275
Finite-dimensional distributions . . . . . . . . . . . . . . . . . . . . . 276
SDE for the Brownian bridge . . . . . . . . . . . . . . . . . . . . . . . 278

CONTENTS 7
8.5 Stochastic differential equations in Rn . . . . . . . . . . . . . . . . . . 280

Existence, uniqueness and stability . . . . . . . . . . . . . . . . . . . . 280
Itô processes driven by several Brownian motions . . . . . . . . . . . . 282
Multivariate Itô-Doeblin formula . . . . . . . . . . . . . . . . . . . . . 283
General Ornstein-Uhlenbeck processes . . . . . . . . . . . . . . . . . . 285
Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
9 Change of measure 286

9.1 Local and global densities of probability measures . . . . . . . . . . . . 286
Absolute Continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
From local to global densities . . . . . . . . . . . . . . . . . . . . . . . 289
Derivatives of monotone functions . . . . . . . . . . . . . . . . . . . . 293
Absolute continuity of infinite product measures . . . . . . . . . . . . . 294
9.2 Translations of Wiener measure . . . . . . . . . . . . . . . . . . . . . 298
The Cameron-Martin Theorem . . . . . . . . . . . . . . . . . . . . . . 299
Passage times for Brownian motion with constant drift . . . . . . . . . 304
9.3 Girsanov transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
Change of measure on filtered probability spaces . . . . . . . . . . . . 306
Girsanov’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
Novikov’s condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
9.4 Itô’s Representation Theorem and Option Pricing . . . . . . . . . . . . 312
Representation theorems for functions and martingales . . . . . . . . . 313
Application to option pricing . . . . . . . . . . . . . . . . . . . . . . . 316
Application to stochastic filtering . . . . . . . . . . . . . . . . . . . . . 317
III Stochastic Analysis 318
10 Lévy processes and Poisson point processes 313

10.1 Lévy processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
Characteristic exponents . . . . . . . . . . . . . . . . . . . . . . . . . 315
Basic examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
Compound Poisson processes . . . . . . . . . . . . . . . . . . . . . . . 318

8 CONTENTS
Examples with infinite jump intensity . . . . . . . . . . . . . . . . . . 321

10.2 Martingales and Markov property . . . . . . . . . . . . . . . . . . . . 325
Martingales of Lévy processes . . . . . . . . . . . . . . . . . . . . . . 325
Lévy processes as Markov processes . . . . . . . . . . . . . . . . . . . 326
10.3 Poisson random measures and Poisson point processes . . . . . . . . . 330
The jump times of a Poisson process . . . . . . . . . . . . . . . . . . . 330
The jumps of a Lévy process . . . . . . . . . . . . . . . . . . . . . . . 333
Poisson point processes . . . . . . . . . . . . . . . . . . . . . . . . . . 336
Construction of compound Poisson processes from PPP . . . . . . . . . 338
10.4 Stochastic integrals w.r.t. Poisson point processes . . . . . . . . . . . . 340
Elementary integrands . . . . . . . . . . . . . . . . . . . . . . . . . . 341
Lebesgue integrals . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
Itô integrals w.r.t. compensated Poisson point processes . . . . . . . . . 345
10.5 Lévy processes with infinite jump intensity . . . . . . . . . . . . . . . 347
Construction from Poisson point processes . . . . . . . . . . . . . . . . 347
The Lévy-Itô decomposition . . . . . . . . . . . . . . . . . . . . . . . 351
Subordinators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354
Stable processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
11 Transformations of SDE 360

11.1 Lévy characterizations and martingale problems . . . . . . . . . . . . . 362
Lévy’s characterization of Brownian motion . . . . . . . . . . . . . . . 364
Martingale problem for Itô diffusions . . . . . . . . . . . . . . . . . . 367
Lévy characterization of weak solutions . . . . . . . . . . . . . . . . . 371
11.2 Random time change . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
Continuous local martingales as time-changed Brownian motions . . . . 373
Time-change representations of stochastic integrals . . . . . . . . . . . 376
Time substitution in stochastic differential equations . . . . . . . . . . 376
One-dimensional SDE . . . . . . . . . . . . . . . . . . . . . . . . . . 379
11.3 Change of measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
Change of measure for Brownian motion . . . . . . . . . . . . . . . . . 384
Applications to SDE . . . . . . . . . . . . . . . . . . . . . . . . . . . 386

CONTENTS 9
Doob’s h-transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388

11.4 Path integrals and bridges . . . . . . . . . . . . . . . . . . . . . . . . . 389
Path integral representation . . . . . . . . . . . . . . . . . . . . . . . . 390
The Markov property . . . . . . . . . . . . . . . . . . . . . . . . . . . 392
Bridges and heat kernels . . . . . . . . . . . . . . . . . . . . . . . . . 393
SDE for diffusion bridges . . . . . . . . . . . . . . . . . . . . . . . . . 395
11.5 Large deviations on path spaces . . . . . . . . . . . . . . . . . . . . . 397
Support of Wiener measure . . . . . . . . . . . . . . . . . . . . . . . . 397
Schilder’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 399
Random perturbations of dynamical systems . . . . . . . . . . . . . . . 403
12 Extensions of Itô calculus 406

12.1 SDE with jumps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
Lp Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
Existence of strong solutions . . . . . . . . . . . . . . . . . . . . . . . 411
Non-explosion criteria . . . . . . . . . . . . . . . . . . . . . . . . . . 414
12.2 Stratonovich differential equations . . . . . . . . . . . . . . . . . . . . 416
Itô-Stratonovich formula . . . . . . . . . . . . . . . . . . . . . . . . . 416
Stratonovich SDE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 418
Brownian motion on hypersurfaces . . . . . . . . . . . . . . . . . . . . 419
Doss-Sussmann method . . . . . . . . . . . . . . . . . . . . . . . . . . 422
Wong Zakai approximations of SDE . . . . . . . . . . . . . . . . . . . 425
12.3 Stochastic Taylor expansions . . . . . . . . . . . . . . . . . . . . . . . 426
Itô-Taylor expansions . . . . . . . . . . . . . . . . . . . . . . . . . . . 426
12.4 Numerical schemes for SDE . . . . . . . . . . . . . . . . . . . . . . . 431
Strong convergence order . . . . . . . . . . . . . . . . . . . . . . . . . 432
Weak convergence order . . . . . . . . . . . . . . . . . . . . . . . . . 437
12.5 Local time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
Local time of continuous semimartingales . . . . . . . . . . . . . . . . 440
Itô-Tanaka formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444
12.6 Continuous modifications and stochastic flows . . . . . . . . . . . . . . 446
Continuous modifications of deterministic functions . . . . . . . . . . . 447

10 CONTENTS
Continuous modifications of random fields . . . . . . . . . . . . . . . . 449

Existence of a continuous flow . . . . . . . . . . . . . . . . . . . . . . 451
Markov property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
Continuity of local time . . . . . . . . . . . . . . . . . . . . . . . . . . 454
13 Variations of parameters in SDE 457

13.1 Variations of parameters in SDE . . . . . . . . . . . . . . . . . . . . . 458
Differentation of solutions w.r.t. a parameter . . . . . . . . . . . . . . . 459
Derivative flow and stability of SDE . . . . . . . . . . . . . . . . . . . 460
Consequences for the transition semigroup . . . . . . . . . . . . . . . . 464
13.2 Malliavin gradient and Bismut integration by parts formula . . . . . . . 467
Gradient and integration by parts for smooth functions . . . . . . . . . 468
Skorokhod integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
Definition of Malliavin gradient II . . . . . . . . . . . . . . . . . . . . 473
Product and chain rule . . . . . . . . . . . . . . . . . . . . . . . . . . 475
Clark-Ocone formula . . . . . . . . . . . . . . . . . . . . . . . . . . . 476
13.3 First applications to SDE . . . . . . . . . . . . . . . . . . . . . . . . . 477
13.4 Existence and smoothness of densities . . . . . . . . . . . . . . . . . . 477
14 Stochastic calculus for semimartingales with jumps 478

Semimartingales in discrete time . . . . . . . . . . . . . . . . . . . . . 479
Semimartingales in continuous time . . . . . . . . . . . . . . . . . . . 480
14.1 Finite variation calculus . . . . . . . . . . . . . . . . . . . . . . . . . . 483
Lebesgue-Stieltjes integrals revisited . . . . . . . . . . . . . . . . . . . 484
Product rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485
Chain rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488
Exponentials of finite variation functions . . . . . . . . . . . . . . . . . 490
14.2 Stochastic integration for semimartingales . . . . . . . . . . . . . . . . 498
Integrals with respect to bounded martingales . . . . . . . . . . . . . . 498
Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 504
Integration w.r.t. semimartingales . . . . . . . . . . . . . . . . . . . . . 507
14.3 Quadratic variation and covariation . . . . . . . . . . . . . . . . . . . . 508

CONTENTS 11
Covariation and integration by parts . . . . . . . . . . . . . . . . . . . 509

Quadratic variation and covariation of local martingales . . . . . . . . . 511
Covariation of stochastic integrals . . . . . . . . . . . . . . . . . . . . 517
The Itô isometry for stochastic integrals w.r.t. martingales . . . . . . . . 520
14.4 Itô calculus for semimartingales . . . . . . . . . . . . . . . . . . . . . 521
Integration w.r.t. stochastic integrals . . . . . . . . . . . . . . . . . . . 521
Itô’s formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 522
Application to Lévy processes . . . . . . . . . . . . . . . . . . . . . . 525
Burkholder’s inequality . . . . . . . . . . . . . . . . . . . . . . . . . . 528
14.5 Stochastic exponentials and change of measure . . . . . . . . . . . . . 530
Exponentials of semimartingales . . . . . . . . . . . . . . . . . . . . . 530
Change of measure for Poisson point processes . . . . . . . . . . . . . 533
Change of measure for Lévy processes . . . . . . . . . . . . . . . . . . 536
Change of measure for general semimartingales . . . . . . . . . . . . . 539
14.6 General predictable integrands . . . . . . . . . . . . . . . . . . . . . . 539
Definition of stochastic integrals w.r.t. semimartingales . . . . . . . . . 541
Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 543
Properties of the stochastic integral . . . . . . . . . . . . . . . . . . . . 544
A Conditional expectations 549

A.1 Conditioning on discrete random variables . . . . . . . . . . . . . . . . 549
Conditional expectations as random variables . . . . . . . . . . . . . . 549
Characteristic properties of conditional expectations . . . . . . . . . . . 550
A.2 General conditional expectations . . . . . . . . . . . . . . . . . . . . . 551
The factorization lemma . . . . . . . . . . . . . . . . . . . . . . . . . 551
Conditional expecations given σ-algebras . . . . . . . . . . . . . . . . 553
Properties of conditional expectations . . . . . . . . . . . . . . . . . . 555
A.3 Conditional expectation as best L2 -approximation . . . . . . . . . . . . 557
Jensen’s inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . 558
Conditional expectation as best L2 -prediction value . . . . . . . . . . . 559
Existence of conditional expectations . . . . . . . . . . . . . . . . . . 561

12 CONTENTS
Index 564
Bibliography 567

Part I
Stochastic Processes
13
Chapter 1
Brownian Motion
This introduction to stochastic analysis starts with an introduction to Brownian motion.

Brownian Motion is a diffusion process, i.e. a continuous-time Markov process (Bt )t≥0
with continuous sample paths t 7→ Bt (ω). In fact, it is the only nontrivial continuous-
time process that is a Lévy process as well as a martingale and a Gaussian process. A
rigorous construction of this process has been carried out first by N. Wiener in 1923.
Already about 20 years earlier, related models had been introduced independently for
financial markets by L. Bachelier [Théorie de la spéculation, Ann. Sci. École Norm.
Sup. 17, 1900], and for the velocity of molecular motion by A. Einstein [Über die von
der molekularkinetischen Theorie der Wärme geforderte Bewegung von in ruhenden
Flüssigkeiten suspendierten Teilchen, Annalen der Physik 17, 1905].
It has been a groundbreaking approach of K. Itô to construct general diffusion processes
from Brownian motion, cf. [. . . ]. In classical analysis, the solution of an ordinary dif-
ferential equation x′ (t) = f (t, x(t)) is a function, that can be approximated locally for
t close to t0 by the linear function x(t0 ) + f (t0 , x(t0 )) · (t − t0 ). Similarly, Itô showed,
that a diffusion process behaves locally like a linear function of Brownian motion – the
connection being described rigorously by a stochastic differential equation (SDE).
The fundamental rôle played by Brownian motion in stochastic analysis is due to the
central limit Theorem. Similarly as the normal distribution arises as a universal scal-
ing limit of standardized sums of independent, identically distributed, square integrable
14
1.1. FROM RANDOM WALKS TO BROWNIAN MOTION 15
random variables, Brownian motion shows up as a universal scaling limit of Random

Walks with square integrable increments.
1.1 From Random Walks to Brownian Motion

To motivate the definition of Brownian motion below, we first briefly discuss discrete-
time stochastic processes and possible continuous-time scaling limits on an informal
level.
A standard approach to model stochastic dynamics in discrete time is to start from a se-
quence of random variables η1 , η2 , . . . defined on a common probability space (Ω, A, P ).
The random variables ηn describe the stochastic influences (noise) on the system. Often
they are assumed to be independent and identically distributed (i.i.d.). In this case the
collection (ηn ) is also called a white noise, whereas a colored noise is given by depen-
dent random variables. A stochastic process Xn , n = 0, 1, 2, . . . , taking values in Rd is
then defined recursively on (Ω, A, P ) by
Xn+1 = Xn + Φn+1 (Xn , ηn+1 ), n = 0, 1, 2, . . . . (1.1.1)
Here the Φn are measurable maps describing the random law of motion. If X0 and
η1 , η2 , . . . are independent random variables, then the process (Xn ) is a Markov chain
with respect to P .
Now let us assume that the random variables ηn are independent and identically dis-
tributed taking values in R, or, more generally, Rd . The easiest type of a nontrivial
Pn
stochastic dynamics as described above is the Random Walk Sn = ηi which satisfies
i=1
Sn+1 = Sn + ηn+1 for n = 0, 1, 2, . . . .
Since the noise random variables ηn are the increments of the Random Walk (Sn ), the
law of motion (1.1.1) in the general case can be rewritten as
Xn+1 − Xn = Φn+1 (Xn , Sn+1 − Sn ), n = 0, 1, 2, . . . . (1.1.2)
This equation is a difference equation for (Xn ) driven by the stochastic process (Sn ).

16 CHAPTER 1. BROWNIAN MOTION
Our aim is to carry out a similar construction as above for stochastic dynamics in con-
tinuous time. The stochastic difference equation (1.1.2) will then eventually be replaced
by a stochastic differential equation (SDE). However, before even being able to think
about how to write down and make sense of such an equation, we have to identify a
continuous-time stochastic process that takes over the rôle of the Random Walk. For
this purpose, we first determine possible scaling limits of Random Walks when the time
steps tend to 0. It will turn out that if the increments are square integrable and the size
of the increments goes to 0 as the length of the time steps tends to 0, then by the Central
Limit Theorem there is essentially only one possible limit process in continuous time:
Brownian motion.
Central Limit Theorem

Suppose that Yn,i : Ω → Rd , 1 ≤ i ≤ n < ∞, are identically distributed, square-
integrable random variables on a probability space (Ω, A, P ) such that Yn,1 , . . . , Yn,n
are independent for each n ∈ N. Then the rescaled sums
n
1 X
√ (Yn,i − E[Yn,i ])
n i=1
converge in distribution to a multivariate normal distribution N(0, C) with covariance

matrix
(k) (l)
Ckl = Cov[Yn,i , Yn,i ].
To see, how the CLT determines the possible scaling limits of Random Walks, let us
consider a one-dimensional Random Walk
n
X
Sn = ηi , n = 0, 1, 2, . . . ,
i=1
on a probability space (Ω, A, P ) with independent increments ηi ∈ L2 (Ω, A, P ) nor-

malized such that
E[ηi ] = 0 and Var[ηi ] = 1. (1.1.3)
Plotting many steps of the Random Walk seems to indicate that there is a limit process
with continuous sample paths after appropriate rescaling:

4 10
2 5
5 10 15 20 20 40 60 80 100
−2 −5
50 100
25 50
500 1000 5000 10000

−25 −50
To see what appropriate means, we fix a positive integer m, and try to define a rescaled
(m)
Random Walk St (t = 0, 1/m, 2/m, . . .) with time steps of size 1/m by
(m)
Sk/m = c m · Sk (k = 0, 1, 2, . . .)
for some constants cm > 0. If t is a multiple of 1/m, then
(m)
Var[St ] = c2m · Var[Smt ] = c2m · m · t.
(m)
Hence in order to achieve convergence of St as m → ∞, we should choose cm
(m)
proportional to m−1/2 . This leads us to define a continuous time process (St )t≥0 by
(m) 1
St (ω) := √ Smt (ω) whenever t = k/m for some integer k,
m
k−1 k

and by linear interpolation for t ∈ ,
m m
.

(m)
St
√
m
1 2 t
m
Figure 1.1: Rescaling of a Random Walk.
Clearly,
(m)
E[St ] = 0 for all t ≥ 0,
and
(m) 1
Var[St ]Var[Smt ] = t
=
m
whenever t is a multiple of 1/m. In particular, the expectation values and variances for a
fixed time t do not depend on m. Moreover, if we fix a partition 0 ≤ t0 < t1 < . . . < tn
such that each ti is a multiple of 1/m, then the increments
(m) (m) 1
Sti+1 − Sti = √ Smti+1 − Smti , i = 0, 1, 2, . . . , n − 1, (1.1.4)
m
(m)
of the rescaled process (St )t≥0 are independent centered random variables with vari-
ances ti+1 − ti . If ti is not a multiple of 1/m, then a corresponding statement holds
approximately with an error that should be negligible in the limit m → ∞. Hence, if
(m)
the rescaled Random Walks (St )t≥0 converge in distribution to a limit process (Bt )t≥0 ,
then (Bt )t≥0 should have independent increments Bti+1 −Bti over disjoint time intervals
with mean 0 and variances ti+1 − ti .
It remains to determine the precise distributions of the increments. Here the Central
Limit Theorem applies. In fact, we can observe that by (1.1.4) each increment
mti+1
(m) (m) 1 X
Sti+1 − Sti = √ ηk
m k=mt +1
i

of the rescaled process is a rescaled sum of m · (ti+1 − ti ) i.i.d. random variables

with mean 0 and variance 1. Therefore, the CLT implies that the distributions of the
increments converge weakly to a normal distribution:
(m) (m) D
Sti+1 − Sti −→ N(0, ti+1 − ti ).
Hence if a limit process (Bt ) exists, then it should have independent, normally dis-
tributed increments.
Our considerations motivate the following definition:
Definition (Brownian Motion).
(1). Let a ∈ R. A continuous-time stochastic process Bt : Ω → R, t ≥ 0, defined on

a probability space (Ω, A, P ), is called a Brownian motion (starting in a) if and
only if
(a) B0 (ω) = a for each ω ∈ Ω.

(b) For any partition 0 ≤ t0 < t1 < . . . < tn , the increments Bti+1 − Bti are
independent random variables with distribution
Bti+1 − Bti ∼ N(0, ti+1 − ti ).
(c) P -almost every sample path t 7→ Bt (ω) is continuous.

(1) (d)
(2). An Rd -valued stochastic process Bt (ω) = (Bt (ω), . . . , Bt (ω)) is called a mul-
ti-dimensional Brownian motion if and only if the component processes
(1) (d)
(Bt ), . . . , (Bt ) are independent one-dimensional Brownian motions.
Thus the increments of a d-dimensional Brownian motion are independent over disjoint
time intervals and have a multivariate normal distribution:
Bt − Bs ∼ N(0, (t − s) · Id ) for any 0 ≤ s ≤ t.

Remark. (1). Continuity: Continuity of the sample paths has to be assumed sepa-
rately: If (Bt )t≥0 is a one-dimensional Brownian motion, then the modified pro-
cess (Bet )t≥0 defined by Be0 = B0 and
et
B = Bt · I{Bt ∈R\Q} for t > 0
has almost surely discontinuous paths. On the other hand, it satisfies (a) and (b)
et , . . . , B
since the distributions of (B etn ) and (Bt , . . . , Btn ) coincide for all n ∈ N
1 1
and t1 , . . . , tn ≥ 0.
(2). Spatial Homogeneity: If (Bt )t≥0 is a Brownian motion starting at 0, then the
translated process (a + Bt )t≥0 is a Brownian motion starting at a.
(3). Existence: There are several constructions and existence proofs for Brownian mo-
tion. In Section 1.3 below we will discuss in detail the Wiener-Lévy construction
of Brownian motion as a random superposition of infinitely many deterministic
paths. This explicit construction is also very useful for numerical approximations.
A more general (but less constructive) existence proof is based on Kolmogorov’s
extension Theorem, cf. e.g. [Klenke].
(4). Functional Central Limit Theorem: The construction of Brownian motion as

a scaling limit of Random Walks sketched above can also be made rigorous.
Donsker’s invariance principle is a functional version of the central limit The-
(m)
orem which states that the rescaled Random Walks (St ) converge in distribu-
tion to a Brownian motion. As in the classical CLT the limit is universal, i.e., it
does not depend on the distribution of the increments ηi provided (1.1.3) holds,
cf. Section ??.
Brownian motion as a Lévy process.

The definition of Brownian motion shows in particular that Brownian motion is a Lévy
process, i.e., it has stationary independent increments (over disjoint time intervals). In
fact, the analogues of Lévy processes in discrete time are Random Walks, and it is rather
obvious, that all scaling limits of Random Walks should be Lévy processes. Brownian

motion is the only Lévy process Lt in continuous time with paths such that E[L1 ] =
0 and Var[L1 ] = 1. The normal distribution of the increments follows under these
assumptions by an extension of the CLT, cf. e.g. [Breiman: Probability]. A simple
example of a Lévy process with non-continuous paths is the Poisson process. Other
examples are α-stable processes which arise as scaling limits of Random Walks when
the increments are not square-integrable. Stochastic analysis based on general Lévy
processes has attracted a lot of interest recently.
Let us now consider consider a Brownian motion (Bt )t≥0 starting at a fixed point a ∈
Rd , defined on a probability space (Ω, A, P ). The information on the process up to time
t is encoded in the σ-algebra
FtB = σ(Bs | 0 ≤ s ≤ t)
generated by the process. The independence of the increments over disjoint intervals
immediately implies:
Lemma 1.1. For any 0 ≤ s ≤ t, the increment Bt − Bs is independent of FsB .
Proof. For any partition 0 = t0 ≤ t1 ≤ . . . ≤ tn = s of the interval [0, s], the increment
Bt − Bs is independent of the σ-algebra
σ(Bt1 − Bt0 , Bt2 − Bt1 , . . . , Btn − Btn−1 )
generated by the increments up to time s. Since

k
X
Btk = Bt0 + (Bti − Bti−1 )
i=1
and Bt0 is constant, this σ-algebra coincides with σ(Bt0 , Bt1 , . . . , Btn ). Hence Bt − Bs
is independent of all finite subcollections of (Bu | 0 ≤ u ≤ s) and therefore independent
of FsB .
Brownian motion as a Markov process.

As a process with stationary increments, Brownian motion is in particular a time-homo-
geneous Markov process. In fact, we have:

Theorem 1.2 (Markov property). A Brownian motion (Bt )t≥0 in Rd is a time-homo-

geneous Markov process with transition densities

−d/2 |x − y|2
pt (x, y) = (2πt) · exp − , t > 0, x, y ∈ Rd ,
2t
i.e., for any Borel set A ⊆ Rd and 0 ≤ s < t,

ˆ
B
P [Bt ∈ a | Fs ] = pt−s (Bs , y) dy P -almost surely.
A
Proof. For 0 ≤ s < t we have Bt = Bs + (Bt − Bs ) where Bs is FsB -measurable, and

Bt − Bs is independent of FsB by Lemma 1.1. Hence
P [Bt ∈ A | FsB ](ω) = P [Bs (ω) + Bt − Bs ∈ A] = N(Bs (ω), (t − s) · Id )[A]

|y − Bs (ω)|2
ˆ
−d/2
= (2π(t − s)) · exp − dy P -almost surely.
2(t − s)
A
Remark (Heat equation as backward equation and forward equation). The tran-
sition function of Brownian motion is the heat kernel in Rd , i.e., it is the fundamental
solution of the heat equation
∂u 1
= ∆u.
∂t 2
More precisely, pt (x, y) solves the initial value problem
∂ 1
pt (x, y) = ∆x pt (x, y) for any t > 0, x, y ∈ Rd ,
∂t 2
ˆ (1.1.5)
lim pt (x, y)f (y) dy = f (x) for any f ∈ Cb (Rd ), x ∈ Rd ,
tց0
Pd ∂2
where ∆x = 2
denotes the action of the Laplace operator on the x-variable. The
i=1 ∂xi
equation (1.1.5) can be viewed as a version of Kolmogorov’s backward equation for

Brownian motion as a time-homogeneous Markov process, which states that for each
t > 0, y ∈ Rd and f ∈ Cb (Rd ), the function
ˆ
v(s, x) = pt−s (x, y)f (y) dy
solves the terminal value problem
∂v 1
(s, x) = − ∆x v(s, x) for s ∈ [0, t), lim v(s, x) = f (x). (1.1.6)
∂s 2 sրt
Note that by the Markov property, v(s, x) = (pt−s f )(x) is a version of the conditional
expectation E[f (Bt ) | Bs = x]. Therefore, the backward equation describes the depen-
dence of the expectation value on starting point and time.
By symmetry, pt (x, y) also solves the initial value problem
∂ 1
pt (x, y) = ∆y pt (x, y) for any t > 0, and x, y ∈ Rd ,
∂t 2
ˆ (1.1.7)
lim g(x)pt (x, y) dx = g(y) for any g ∈ Cb (Rd ), y ∈ Rd .
tց0
The equation (1.1.7) is a version of Kolmogorov’s forward equation, stating that for
g ∈ Cb (Rd ), the function u(t, y) = g(x)pt (x, y) dx solves
´
∂u 1
(t, y) = ∆y u(t, y) for t > 0, lim u(t, y) = g(y). (1.1.8)
∂t 2 tց0
The forward equation describes the forward time evolution of the transition densities
pt (x, y) for a given starting point x.
The Markov property enables us to compute the marginal distributions of Brownian

motion:
Corollary 1.3 (Finite dimensional marginals). Suppose that (Bt )t≥0 is a Brownian
motion starting at x0 ∈ Rd defined on a probability space (Ω, A, P ). Then for any

n ∈ N and 0 = t0 < t1 < t2 < . . . < tn , the joint distribution of Bt1 , Bt2 , . . . , Btn is
absolutely continuous with density
fBt1 ,...,Btn (x1 , . . . , xn ) = pt1 (x0 , x1 )pt2 −t1 (x1 , x2 )pt3 −t2 (x2 , x3 ) · · · ptn −tn−1 (xn−1 , xn )
n n
!
Y 1 X |x i − xi−1 | 2
= (2π(ti − ti−1 ))−d/2 · exp − (. 1.1.9)
i=1
2 i=1 ti − ti−1
Proof. By the Markov property and induction on n, we obtain
P [Bt1 ∈ A1 , . . . , Btn ∈ An ]
= E[P [Btn ∈ An | FtBn−1 ] ; Bt1 ∈ A1 , . . . , Btn−1 ∈ An−1 ]
= E[ptn −tn−1 (Btn−1 , An ) ; Bt1 ∈ A1 , . . . , Btn−1 ∈ An−1 ]
ˆ ˆ
= ··· pt1 (x0 , x1 )pt2 −t1 (x1 , x2 ) · · ·
A1 An−1
·ptn−1 −tn−2 (xn−2 , xn−1 )ptn −tn−1 (xn−1 , An ) dxn−1 · · · dx1

ˆ ˆ Y n
!
= ··· pti −ti−1 (xn−1 , xn ) dxn · · · dx1
A1 An i=1
for all n ≥ 0 and A1 , . . . , An ∈ B(Rd ).
Remark (Brownian motion as a Gaussian process). The corollary shows in particular

that Brownian motion is a Gaussian process, i.e., all the marginal distributions in (1.1.9)
are multivariate normal distributions. We will come back to this important aspect in the
next section.
Wiener Measure
The distribution of Brownian motion could be considered as a probability measure on
the product space (Rd )[0,∞) consisting of all maps x : [0, ∞) → Rd . A disadvantage
of this approach is that the product space is far too large for our purposes: It contains
extremely irregular paths x(t), although at least almost every path of Brownian motion
is continuous by definition. Actually, since [0, ∞) is uncountable, the subset of all

continuous paths is not even measurable w.r.t. the product σ-algebra on (Rd )[0,∞) .
Instead of the product space, we will directly consider the distribution of Brownian
motion on the continuous path space C([0, ∞), Rd). For this purpose, we fix a Brownian
motion (Bt )t≥0 starting at x0 ∈ Rd on a probability space (Ω, A, P ), and we assume that
every sample path t 7→ Bt (ω) is continuous. This assumption can always be fulfilled by
modifying a given Brownian motion on a set of measure zero. The full process (Bt )t≥0
can then be interpreted as a single path-space valued random variable (or a "random
path").
Rd
x0
ω Ω
t
B(ω)
Figure 1: B : Ω → C([0, ∞), Rd ), B(ω) = (Bt (ω))t≥0 .
Figure 1.2: B : Ω → C([0, ∞), Rd), B(ω) = (Bt (ω))t≥0 .
We endow the space of continuous paths x : [0, ∞) → Rd with the σ-algebra
B = σ(Xt | t ≥ 0)
generated by the coordinate maps
Xt : C([0, ∞), Rd) → Rd , Xt (x) = xt , t ≥ 0.
Note that we also have

B = σ(Xt | t ∈ D)
for any dense subset D of [0, ∞), because Xt = lim Xs for each t ∈ [0, ∞) by con-
s→t
tinuity. Furthermore, it can be shown that B is the Borel σ-algebra on C([0, ∞), Rd)
endowed with the topology of uniform convergence on finite intervals.

Theorem 1.4 (Distribution of Brownian motion on path space). The map B : Ω →

C([0, ∞), Rd) is measurable w.r.t. the σ-algebras A/B. The distribution P ◦ B −1 of B
is the unique probability measure µx0 on (C([0, ∞), Rd), B) with marginals

µx0 {x ∈ C([0, ∞), Rd) : xt1 ∈ A1 , . . . , xtn ∈ An } (1.1.10)
n n
!
Y ˆ ˆ
1 X |xi − xi−1 |2
= (2π(ti − ti−1 ))−d/2 · · · exp − dxn · · · dx1
i=1
2 i=1 ti − ti−1
A1 An
for any n ∈ N, 0 < t1 < . . . < tn , and A1 , . . . , An ∈ B(Rd ).
Definition. The probability measure µx0 on the path space C([0, ∞), Rd) determined
by (1.1.10) is called Wiener measure (with start in x0 ).
Remark (Uniqueness in distribution). The Theorem asserts that the path space distri-
bution of a Brownian motion starting at a given point x0 is the corresponding Wiener
measure. In particular, it is uniquely determined by the marginal distributions in (1.1.9).
Proof of Theorem 1.4. For n ∈ N, 0 < t1 < . . . < tn , and A1 , . . . , An ∈ B(Rd ), we

have
B −1 ({Xt1 ∈ A1 , . . . , Xtn ∈ An }) = {ω : Xt1 (B(ω)) ∈ A1 , . . . , Xtn (B(ω)) ∈ An }

= {Bt1 ∈ A1 , . . . , Btn ∈ An } ∈ A.
Since the cylinder sets of type {Xt1 ∈ A1 , . . . , Xtn ∈ An } generate the σ-algebra B, the
map B is A/B-measurable. Moreover, by corollary 1.3, the probabilities
P [B ∈ {Xt1 ∈ A1 , . . . , Xtn ∈ An }] = P [Bt1 ∈ A1 , . . . , Btn ∈ An ],
are given by the right hand side of (1.1.10). Finally, the measure µx0 is uniquely deter-
mined by (1.1.10), since the system of cylinder sets as above is stable under intersections
and generates the σ-algebra B.

1.2. BROWNIAN MOTION AS A GAUSSIAN PROCESS 27
Definition (Canonical model for Brownian motion.). By (1.1.10), the coordinate pro-
cess
Xt (x) = xt , t ≥ 0,
on C([0, ∞), Rd) is a Brownian motion starting at x0 w.r.t. Wiener measure µx0 . We
refer to the stochastic process (C([0, ∞), Rd), B, µx0 , (Xt )t≥0 ) as the canonical model
for Brownian motion starting at x0 .
1.2 Brownian Motion as a Gaussian Process

We have already verified that Brownian motion is a Gaussian process, i.e., the finite
dimensional marginals are multivariate normal distributions. We will now exploit this
fact more thoroughly.
Multivariate normals
Let us first recall some basics on normal random vectors:
Definition. Suppose that m ∈ Rn is a vector and C ∈ Rn×n is a symmetric non-

negative definite matrix. A random variable Y : Ω → Rn defined on a probability
space (Ω, A, P ) has a multivariate normal distribution N(m, C) with mean m and
covariance matrix C if and only if its characteristic function is given by
1
E[eip·Y ] = eip·m− 2 p·Cp for any p ∈ Rn . (1.2.1)
If C is non-degenerate, then a multivariate normal random variable Y is absolutely

continuous with density

−1/2 1 −1
fY (x) = (2π det C) exp − (x − m) · C (x − m) .
2

A degenerate normal distribution with vanishing covariance matrix is a Dirac measure:
N(m, 0) = δm .
Differentiating (1.2.1) w.r.t. p shows that for a random variable Y ∼ N(m, C), the
mean vector is m and Ci,j is the covariance of the components Yi and Yj . Moreover, the
following important facts hold:
Theorem 1.5 (Properties of normal random vectors).
(1). A random variable Y : Ω → Rn has a multivariate normal distribution if and

only if any linear combination
n
X
p·Y = pi Y i , p ∈ Rn ,
i=1
of the components Yi has a one dimensional normal distribution.
(2). Any affine function of a normally distributed random vector Y is again normally
distributed:
Y ∼ N(m, C) =⇒ AY + b ∼ N(Am + b, ACA⊤ )
for any d ∈ N, A ∈ Rd×n and b ∈ Rd .
(3). If Y = (Y1 , . . . , Yn ) has a multivariate normal distribution, and the components

Y1 , . . . , Yn are uncorrelated random variables, then Y1 , . . . , Yn are independent.
Proof. (1). follows easily from the definition.
(2). For Y ∼ N(m, C), A ∈ Rd×n and b ∈ Rd we have

⊤ p)·Y
E[eip·(AY +b) ] = eip·bE[ei(A ]
⊤ p)·m− 1 (A⊤ p)·CA⊤ p
= eip·bei(A 2
1 ⊤
= eip·(Am+b)− 2 p·ACA for any p ∈ Rd ,
i.e., AY + b ∼ N(Am + b, ACA⊤ ).

(3). If Y1 , . . . , Yn are uncorrelated, then the covariance matrix Ci,j = Cov[Yi , Yj ] is a

diagonal matrix. Hence the characteristic function
n
Y
1 1 2
E[eip·Y ] = eip·m− 2 p·Cp = eimk pk − 2 Ck,k pk
k=1
is a product of characteristic functions of one-dimensional normal distributions.

Since a probability measure on Rn is uniquely determined by its characteristic
function, it follows that the adjoint distribution of Y1 , . . . , Yn is a product measure,
i.e. Y1 , . . . , Yn are independent.
If Y has a multivariate normal distribution N(m, C) then for any p, q ∈ Rn , the random
variables p · Y and q · Y are normally distributed with means p · m and q · m, and
covariance
n
X
Cov[p · Y, q · Y ] = pi Ci,j qj = p · Cq.
i,j=1
In particular, let {e1 , . . . , en } ⊆ Rn be an orthonormal basis consisting of eigenvectors

of the covariance matrix C. Then the components ei · Y of Y in this basis are uncor-
related and therefore independent, jointly normally distributed random variables with
variances given by the corresponding eigenvectors λi :
Cov[ei · Y, ej · Y ] = λi δi,j , 1 ≤ i, j ≤ n. (1.2.2)
Correspondingly, the contour lines of the density of a non-degenerate multivariate nor-

mal distribution N(m, C) are ellipsoids with center at m and principal axes of length
√
λi given by the eigenvalues ei of the covariance matrix C.

Figure !1.3: Level!!lines of the density of a normal random vector Y ∼

1 1 1
N , .
2 −1 1
Conversely, we can generate a random vector Y with distribution N(m, C) from i.i.d.
standard normal random variables Z1 , . . . , Zn by setting
n p
X
Y = m+ λi Zi ei . (1.2.3)
i=1
More generally, we have:
Corollary 1.6 (Generating normal random vectors). Suppose that C = UΛU ⊤ with
a matrix U ∈ Rn×d , d ∈ N, and a diagonal matrix Λ = diag(λ1 , . . . , λd ) ∈ Rd×d with
nonnegative entries λi . If Z = (Z1 , . . . , Zd ) is a random vector with i.i.d. standard
normal random components Z1 , . . . , Zd then
Y = UΛ1/2 Z + m
has distribution N(m, C).

Proof. Since Z ∼ N(0, Id ), the second assertion of Theorem 1.5 implies
Y ∼ N(m, UΛU ⊤ ).
Choosing for U the matrix (e1 , . . . , en ) consisting of the orthonormal eigenvectors

e1 , . . . , en of C, we obtain (1.2.3) as a special case of the corollary. For computational
purposes it is often more convenient to use the Cholesky decomposition
C = LL⊤
of the covariance matrix as a product of a lower triangular matrix L and the upper
triangular transpose L⊤ :
Algorithm 1.7 (Simulation of multivariate normal random variables).
Given: m ∈ Rn , C ∈ Rn×n symmetric and non-negative definite.
Output: Sample y ∼ N(m, C).
(1). Compute the Cholesky decomposition C = LL⊤ .
(2). Generate independent samples z1 , . . . , zn ∼ N(0, 1) (e.g. by the Box-Muller

method).
(3). Set y := Lz + m.
Gaussian processes
Let I be an arbitrary index set, e.g. I = N, I = [0, ∞) or I = Rn .
Definition. A collection (Yt )t∈I of random variables Yt : Ω → Rd defined on a proba-

bility space (Ω, A, P ) is called a Gaussian process if and only if the joint distribution
of any finite subcollection Yt1 , . . . , Ytn with n ∈ N and t1 , . . . , tn ∈ I is a multivariate
normal distribution.

The distribution of a Gaussian process (Yt )t∈I on the path space RI or C(I, R) endowed
with the σ-algebra generated by the maps x 7→ xt , t ∈ I, is uniquely determined by
the multinormal distributions of finite subcollections Yt1 , . . . , Ytn as above, and hence
by the expectation values
m(t) = E[Yt ], t ∈ I,
and the covariances
c(s, t) = Cov[Ys , Yt ], s, t ∈ I.
A Gaussian process is called centered, if m(t) = 0 for any t ∈ I.
Example (AR(1) process). The autoregressive process (Yn )n=0,1,2,... defined recur-
sively by Y0 ∼ N(0, v0 ),
Yn = αYn−1 + εηn for n ∈ N,
with parameters v0 > 0, α, ε ∈ R, ηn i.i.d. ∼ N(0, 1), is a centered Gaussian process.

The covariance function is given by
c(n, n + k) = v0 + ε2 n for any n, k ≥ 0 if α = 1,
and

k 2n ε2 2n
c(n, n + k) = α · α v0 + (1 − α ) · for n, k ≥ 0 otherwise.
1 − α2
This is easily verified by induction. We now consider some special cases:
α = 0: In this case Yn = εηn . Hence (Yn ) is a white noise, i.e., a sequence of inde-
pendent normal random variables, and
Cov[Yn , Ym ] = ε2 · δn,m for any n, m ≥ 1.
P
n
α = 1: Here Yn = Y0 + ε ηi , i.e., the process (Yn ) is a Gaussian Random Walk, and
i=1
Cov[Yn , Ym ] = v0 + ε2 · min(n, m) for any n, m ≥ 0.
We will see a corresponding expression for the covariances of Brownian motion.

α < 1: For α < 1, the covariances Cov[Yn , Yn+k ] decay exponentially fast as k → ∞.
ε2
If v0 = 1−α2
, then the covariance function is translation invariant:
ε2 α k
c(n, n + k) = for any n, k ≥ 0.
1 − α2
Therefore, in this case the process (Yn ) is stationary, i.e., (Yn+k )n≥0 ∼ (Yn )n≥0 for all
k ≥ 0.
Brownian motion is our first example of a nontrivial Gaussian process in continuous

time. In fact, we have:
Theorem 1.8 (Gaussian characterization of Brownian motion). A real-valued stoch-

astic process (Bt )t∈[0,∞) with continuous sample paths t 7→ Bt (ω) and B0 = 0 is a
Brownian motion if and only if (Bt ) is a centered Gaussian process with covariances
Cov[Bs , Bt ] = min(s, t) for any s, t ≥ 0. (1.2.4)
Proof. For a Brownian motion (Bt ) and 0 = t0 < t1 < . . . < tn , the increments Bti −
Bti−1 , 1 ≤ i ≤ n, are independent random variables with distribution N(0, ti − ti−1 ).
Hence,
n
O
(Bt1 − Bt0 , . . . , Btn − Btn−1 ) ∼ N(0, ti − ti−1 ),
i=1
which is a multinormal distribution. Since Bt0 = B0 = 0, we see that
 
1 0 0 ... 0 0
 
  1 1 0 . . . 0 0  
Bt1   Bt1 − Bt0
 . .. 
 .    .. 
 ..  =   . 
   ..   
 . 
 
Btn 1 1 1 . . . 1 0 Btn − Btn−1
 
1 1 1 ... 1 1
also has a multivariate normal distribution, i.e., (Bt ) is a Gaussian process. Moreover,
since Bt = Bt − B0 , we have E[Bt ] = 0 and
Cov[Bs , Bt ] = Cov[Bs , Bs ] + Cov[Bs , Bt − Bs ] = Var[Bs ] = s

for any 0 ≤ s ≤ t, i.e., (1.2.4) holds.

Conversely, if (Bt ) is a centered Gaussian process satisfying (1.2.4), then for any 0 =
t0 < t1 < . . . < tn , the vector (Bt1 − Bt0 , . . . , Btn − Btn−1 ) has a multivariate normal
distribution with
E[Bti − Bti−1 ] = E[Bti ] − E[Bti−1 ] = 0, and
Cov[Bti − Bti−1 , Btj − Btj−1 ] = min(ti , tj ) − min(ti , tj−1 )

− min(ti−1 , tj ) + min(ti−1 , tj−1 )
= (ti − ti−1 ) · δi,j for any i, j = 1, . . . , n.
Hence by Theorem 1.5 (3), the increments Bti − Bti−1 , 1 ≤ i ≤ n, are independent with
distribution N(0, ti − ti−1 ), i.e., (Bt ) is a Brownian motion.
Symmetries of Brownian motion
A first important consequence of the Gaussian characterization of Brownian motion are

several symmetry properties of Wiener measure:
Theorem 1.9 (Invariance properties of Wiener measure). Let (Bt )t≥0 be a Brown-
ian motion starting at 0 defined on a probability space (Ω, A, P ). Then the following
processes are again Brownian motions:
(1). (−Bt )t≥0 (Reflection invariance)
(2). (Bt+h − Bh )t≥0 for any h ≥ 0 (Stationarity)
(3). (a−1/2 Bat )t≥0 for any a > 0 (Scale invariance)
et )t≥0 defined by
(4). The time inversion (B
e0 = 0,
B et = t · B1/t
B for t > 0.

Proof. The proofs of (1), (2) and (3) are left as an exercise to the reader. To show (4),
et1 , . . . , B
we first note that for each n ∈ N and 0 ≤ t1 < . . . < tn , the vector (B etn ) has a
multivariate normal distribution since it is a linear transformation of (B1/t1 , . . . , B1/tn ),
(B0 , B1/t2 , . . . , B1/tn ) respectively. Moreover,
et ] = 0
E[B for any t ≥ 0,
es , B
Cov[B et ] = st · Cov[B1/s , B1/t ]
1 1
= st · min( , ) = min(t, s) for any s, t > 0, and
s t
e e
Cov[B0 , Bt ] = 0 for any t ≥ 0.
et )t≥0 is a centered Gaussian process with the covariance function of Brownian

Hence (B
motion. By Theorem 1.8, it only remains to show that P -almost every sample path
et (ω) is continuous. This is obviously true for t > 0. Furthermore, since the finite
t 7→ B
et )t≥0 and (Bt )t≥0 are multivariate normal
dimensional marginals of the processes (B
distributions with the same means and covariances, the distributions of (Bet )t≥0 and
(Bt )t≥0 on the product space R(0,∞) endowed with the product σ-algebra generated by
the cylinder sets agree. To prove continuity at 0 we note that the set
 
 

x : (0, ∞) → R limxt = 0
 tց0 
t∈Q
is measurable w.r.t. the product σ-algebra on R(0,∞) . Therefore,

   
et = 0
P lim B = P lim Bt = 0 = 1.
tց0 tց0
t∈Q t∈Q
et is almost surely continuous for t > 0, we can conclude that outside a set of
Since B
measure zero,
es |
sup |B = sup es |
|B −→ 0 as t ց 0,
s∈(0,t) s∈(0,t)∩Q
et is almost surely continuous at 0 as well.

i.e., t 7→ B

Remark (Long time asymptotics versus local regularity, LLN). The time inversion
invariance of Wiener measure enables us to translate results on the long time asymp-
totics of Brownian motion (t ր ∞) into local regularity results for Brownian paths
et ) at 0 is equiva-
(t ց 0) and vice versa. For example, the continuity of the process (B
lent to the law of large numbers:

1
P lim Bt = 0 = P lim sB1/s = 0 = 1.
t→∞ t sց0
At first glance, this looks like a simple proof of the LLN. However, the argument is based
on the existence of a continuous Brownian motion, and the existence proof requires
similar arguments as a direct proof of the law of large numbers.
Wiener measure as a Gaussian measure, path integral heuristics
Wiener measure (with start at 0) is the unique probability measure µ on the continuous
path space C([0, ∞), Rd) such that the coordinate process
Xt : C([0, ∞), Rd) → Rd , Xt (x) = xt ,
is a Brownian motion starting at 0. By Theorem 1.8, Wiener measure is a centered

Gaussian measure on the infinite dimensional space C([0, ∞), Rd), i.e., for any n ∈ N
and t1 , . . . , tn ∈ R+ , (Xt1 , . . . , Xtn ) is normally distributed with mean 0. We now "de-
rive" a heuristic representation of Wiener measure that is not mathematically rigorous
but nevertheless useful:
Fix a constant T > 0. Then for 0 = t0 < t1 < . . . < tn ≤ T , the distribution of
(Xt1 , . . . , Xtn ) w.r.t. Wiener measure is
n
! n
1 1 X |xti − xti−1 |2 Y
µt1 ,...,tn (dxt1 , . . . , dxtn ) = exp − dxti ,
Z(t1 , . . . , tn ) 2 i=1 ti − ti−1 i=1
(1.2.5)
where Z(t1 , . . . , tn ) is an appropriate finite normalization constant, and x0 := 0. Now
(k) (k) (k)
choose a sequence (τk )k∈N of partitions 0 = t0 < t1 < . . . < tn(k) = T of the interval

(k) (k)
[0, T ] such that the mesh size max |ti+1 − ti | tends to zero. Taking informally the limit
i
in (1.2.5), we obtain the heuristic asymptotic representation
 
ˆT 2 Y
1
1 dx 
µ(dx) = exp − dt δ (dx ) dxt (1.2.6)
2 dt
0 0
Z∞
0 t∈(0,T ]
for Wiener measure on continuous paths x : [0, T ] → Rd with a "normalizing constant"

Z∞ . Trying to make the informal expression (1.2.6) rigorous fails for several reasons:
(k) (k)
• The normalizing constant Z∞ = lim Z(t1 , . . . , tn(k) ) is infinite.
k→∞

´T dx 2
• The integral dt is also infinite for µ-almost every path x, since typical
0 dt
paths of Brownian motion are nowhere differentiable, cf. below.
Q
• The product measure dxt can be defined on cylinder sets but an extension to
t∈(0,T ]
the σ-algebra generated by the coordinate maps on C([0, ∞), Rd) does not exist.
Hence there are several infinities involved in the informal expression (1.2.6). These
infinities magically balance each other such that the measure µ is well defined in contrast
to all of the factors on the right hand side.
In physics, R. Feynman introduced correspondingly integrals w.r.t. "Lebesgue measure

on path space", cf. e.g. the famous Feynman Lecture notes [...], or Glimm and Jaffe [ ...
].
Although not mathematically rigorous, the heuristic expression (1.2.5) can be a very
useful guide for intuition. Note for example that (1.2.5) takes the form
µ(dx) ∝ exp(−kxk2H /2) λ(dx), (1.2.7)
1/2
where kxkH = (x, x)H is the norm induced by the inner product
ˆT
dx dy
(x, y)H = dt (1.2.8)
dt dt
0

of functions x, y : [0, T ] → Rd vanishing at 0, and λ is a corresponding "infinite-

dimensional Lebesgue measure" (which does not exist!). The vector space
dx
H = {x : [0, T ] → Rd : x(0) = 0, x is absolutely continuous with ∈ L2 }
dt
is a Hilbert space w.r.t. the inner product (1.2.8). Therefore, (1.2.7) suggests to consider
Wiener measure as a standard normal distribution on H. It turns out that this idea can
be made rigorous although not as easily as one might think at first glance. The difficulty
is that a standard normal distribution on an infinite-dimensional Hilbert space does not
exist on the space itself but only on a larger space. In particular, we will see in the next
sections that Wiener measure µ can indeed be realized on the continuous path space
C([0, T ], Rd), but µ-almost every path is not contained in H!
Remark (Infinite-dimensional standard normal distributions). The fact that a stan-

dard normal distribution on an infinite dimensional separable Hilbert space H can not
be realized on the space H itself can be easily seen by contradiction: Suppose that µ
is a standard normal distribution on H, and en , n ∈ N, are infinitely many orthonormal
vectors in H. Then by rotational symmetry, the balls

1
Bn = x ∈ H : kx − en kH < , n ∈ N,
2
should all have the same measure. On the other hand, the balls are disjoint. Hence by
σ-additivity,
∞
X h[ i
µ[Bn ] = µ Bn ≤ µ[H] = 1,
n=1
and therefore µ[Bn ] = 0 for all n ∈ N. A scaling argument now implies
µ[{x ∈ H : kx − hk ≤ khk/2}] = 0 for all h ∈ H,
and hence µ ≡ 0.
1.3 The Wiener-Lévy Construction

In this section we discuss how to construct Brownian motion as a random superposi-
tion of deterministic paths. The idea already goes back to N. Wiener, who constructed

1.3. THE WIENER-LÉVY CONSTRUCTION 39
Brownian motion as a random Fourier series. The approach described here is slightly
different and due to P. Lévy: The idea is to approximate the paths of Brownian mo-
tion on a finite time interval by their piecewise linear interpolations w.r.t. the sequence
of dyadic partitions. This corresponds to a development of the Brownian paths w.r.t.
Schauder functions ("wavelets") which turns out to be very useful for many applica-
tions including numerical simulations.
Our aim is to construct a one-dimensional Brownian motion Bt starting at 0 for t ∈

[0, 1]. By stationarity and independence of the increments, a Brownian motion defined
for all t ∈ [0, ∞) can then easily be obtained from infinitely many independent copies
of Brownian motion on [0, 1]. We are hence looking for a random variable
B = (Bt )t∈[0,1] : Ω −→ C([0, 1])
defined on a probability space (Ω, A, P ) such that the distribution P ◦ B −1 is Wiener

measure µ on the continuous path space C([0, 1]).
A first attempt
Recall that µ0 should be a kind of standard normal distribution w.r.t. the inner product
ˆ1
dx dy
(x, y)H = dt (1.3.1)
dt dt
0
on functions x, y : [0, 1] → R. Therefore, we could try to define

∞
X
Bt (ω) := Zi (ω)ei (t) for t ∈ [0, 1] and ω ∈ Ω, (1.3.2)
i=1
where (Zi )i∈N is a sequence of independent standard normal random variables, and
(ei )i∈N is an orthonormal basis in the Hilbert space
H = {x : [0, 1] → R | x(0) = 0, x is absolutely continuous with (x, x)H < ∞}.

(1.3.3)
However, the resulting series approximation does not converge in H:

Theorem 1.10. Suppose (ei )i∈N is a sequence of orthonormal vectors in a Hilbert space
H and (Zi )i∈N is a sequence of i.i.d. random variables with P [Zi 6= 0] > 0. Then the
P
∞
series Zi (ω)ei diverges with probability 1 w.r.t. the norm on H.
i=1
Proof. By orthonormality and by the law of large numbers,

2
Xn n
X

Zi (ω)ei = Zi (ω)2 −→ ∞

i=1 H i=1
P -almost surely as n → ∞.
The Theorem again reflects the fact that a standard normal distribution on an infinite-
dimensional Hilbert space can not be realized on the space itself.
To obtain a positive result, we will replace the norm

 1  12
ˆ 2

kxkH =  dx dt
dt
0
on H by the supremum norm
kxksup = sup |xt |,

t∈[0,1]
and correspondingly the Hilbert space H by the Banach space C([0, 1]). Note that the
supremum norm is weaker than the H-norm. In fact, for x ∈ H and t ∈ [0, 1], the
Cauchy-Schwarz inequality implies
t 2
ˆ ˆt

|xt | 2
= x′s ds ≤ t· |x′s |2 ds ≤ kxk2H ,

0 0
and therefore
kxksup ≤ kxkH for any x ∈ H.

There are two choices for an orthonormal basis of the Hilbert space H that are of par-
ticular interest: The first is the Fourier basis given by
√
2
e0 (t) = t, en (t) = sin(πnt) for n ≥ 1.
πn
With respect to this basis, the series in (1.3.2) is a Fourier series with random coeffi-
cients. Wiener’s original construction of Brownian motion is based on a random Fourier
series. A second convenient choice is the basis of Schauder functions ("wavelets") that
has been used by P. Lévy to construct Brownian motion. Below, we will discuss Lévy’s
construction in detail. In particular, we will prove that for the Schauder functions, the
series in (1.3.2) converges almost surely w.r.t. the supremum norm towards a contin-
uous (but not absolutely continuous) random path (Bt )t∈[0,1] . It is then not difficult to
conclude that (Bt )t∈[0,1] is indeed a Brownian motion.
The Wiener-Lévy representation of Brownian motion

Before carrying out Lévy’s construction of Brownian motion, we introduce the Schauder
functions, and we show how to expand a given Brownian motion w.r.t. this basis of
function space. Suppose we would like to approximate the paths t 7→ Bt (ω) of a Brow-
nian motion by their piecewise linear approximations adapted to the sequence of dyadic
partitions of the interval [0, 1].
An obvious advantage of this approximation over a Fourier expansion is that the values
of the approximating functions at the dyadic points remain fixed once the approximating

partition is fine enough. The piecewise linear approximations of a continuous function

on [0, 1] correspond to a series expansion w.r.t. the base functions
e(t) = t , and
en,k (t) = 2−n/2 e0,0 (2n t − k), n = 0, 1, 2, . . . , k = 0, 1, 2, . . . , 2n − 1, , where


t for t ∈ [0, 1/2]


e0,0 (t) = min(t, 1 − t)+ = 1−t for t ∈ (1/2, 1] .



0 for t ∈ R \ [0, 1]
1 e(t)
en,k (t)
−(1+n/2)
2
k · 2−n (k + 1)2−n 1

e0,0 (t)
0.5
The functions en,k (n ≥ 0, 0 ≤ k < 2n ) are called Schauder functions. It is rather

obvious that piecewise linear approximation w.r.t. the dyadic partitions corresponds to
the expansion of a function x ∈ C([0, 1]) with x(0) = 0 in the basis given by e(t)
and the Schauder functions. The normalization constants in defining the functions en,k
have been chosen in such a way that the en,k are orthonormal w.r.t. the H-inner product
introduced above.
Definition. A sequence (ei )i∈N of vectors in an infinite-dimensional Hilbert space H is

called an orthonormal basis (or complete orthonormal system) of H if and only if
(1). Orthonormality: (ei , ej ) = δij for any i, j ∈ N, and
(2). Completeness: Any h ∈ H can be expressed as

∞
X
h = (h, ei )H ei .
i=1
Remark (Equivalent characterizations of orthonormal bases). Let ei , i ∈ N, be

orthonormal vectors in a Hilbert space H. Then the following conditions are equivalent:
(1). (ei )i∈N is an orthonormal basis of H.
(2). The linear span

( )
k
X

span{ei | i ∈ N} = ci ei k ∈ N, c1 , . . . , ck ∈ R

i=1

is a dense subset of H.
(3). There is no element x ∈ H, x 6= 0, such that (x, ei )H = 0 for every i ∈ N.
(4). For any element x ∈ H, Parseval’s relation
∞
X
kxk2H = (x, ei )2H (1.3.4)
i=1
holds.
(5). For any x, y ∈ H,
∞
X
(x, y)H = (x, ei )H (y, ei )H . (1.3.5)
i=1
For the proofs we refer to any book on functional analysis, cf. e.g. [Reed and Simon:
Methods of modern mathematical physics, Vol. I].
Lemma 1.11. The Schauder functions e and en,k (n ≥ 0, 0 ≤ k < 2n ) form an or-
thonormal basis in the Hilbert space H defined by (1.3.3).
Proof. By definition of the inner product on H, the linear map d/dt which maps an
absolutely continuous function x ∈ H to its derivative x′ ∈ L2 (0, 1) is an isometry
from H onto L2 (0, 1), i.e.,
(x, y)H = (x′ , y ′)L2 (0,1) for any x, y ∈ H.
The derivatives of the Schauder functions are the Haar functions
e′ (t) ≡ 1,
e′n,k (t) = 2n/2 (I[k·2−n,(k+1/2)·2−n ) (t) − I[(k+1/2)·2−n ,(k+1)·2−n ) (t)) for a.e. t.

e′ (t) e′n,k (t)

1 2−n/2
(k + 1)2−n
1 k · 2−n 1
−2−n/2
It is easy to see that these functions form an orthonormal basis in L2 (0, 1). In fact,
orthonormality w.r.t. the L2 inner product can be verified directly. Moreover, the linear
span of the functions e′ and e′n,k for n = 0, 1, . . . , m and k = 0, 1, . . . , 2n − 1 consists of
all step functions that are constant on each dyadic interval [j · 2−(m+1) , (j + 1) · 2−(m+1) ).
An arbitrary function in L2 (0, 1) can be approximated by dyadic step functions w.r.t.
the L2 norm. This follows for example directly from the L2 martingale convergence
Theorem, cf. ... below. Hence the linear span of e′ and the Haar functions e′n,k is dense
in L2 (0, 1), and therefore these functions form an orthonormal basis of the Hilbert space
L2 (0, 1). Since x 7→ x′ is an isometry from H onto L2 (0, 1), we can conclude that e and
the Schauder functions en,k form an orthonormal basis of H.
The expansion of a function x : [0, 1] → R in the basis of Schauder functions can now
be made explicit. The coefficients of a function x ∈ H in the expansion are
ˆ1 ˆ1
′ ′
(x, e)H = x e dt = x′ dt = x(1) − x(0) = x(1)
0 0

ˆ1 ˆ1
(x, en,k )H = x′ e′n,k dt = 2 n/2
x′ (t)e′0,0 (2n t − k) dt
0 0

n/2 1 −n −n −n 1 −n
= 2 (x((k + ) · 2 ) − x(k · 2 )) − (x((k + 1) · 2 ) − x((k + ) · 2 )) .
2 2
Theorem 1.12. Let x ∈ C([0, 1]). Then the expansion

∞ 2X
−1 n
X
x(t) = x(1)e(t) − 2n/2 ∆n,k x · en,k (t),
n=0 k=0

−n 1 −n 1 −n −n
∆n,k x = (x((k + 1) · 2 ) − x((k + ) · 2 )) − (x((k + ) · 2 ) − x(k · 2 ))
2 2
holds w.r.t. uniform convergence on [0, 1]. For x ∈ H the series also converges w.r.t.
the stronger H-norm.
Proof. It can be easily verified that by definition of the Schauder functions, for each
m ∈ N the partial sum
m 2X
−1 n
X
(m)
x (t) := x(1)e(t) − 2n/2 ∆n,k x · en,k (t) (1.3.6)
n=0 k=0
is the polygonal interpolation of x(t) w.r.t. the (m+1)-th dyadic partition of the interval
[0, 1]. Since the function x is uniformly continuous on [0, 1], the polygonal interpola-
tions converge uniformly to x. This proves the first statement. Moreover, for x ∈ H,
the series is the expansion of x in the orthonormal basis of H given by the Schauder
functions, and therefore it also converges w.r.t. the H-norm.
Applying the expansion to the paths of a Brownian motions, we obtain:
Corollary 1.13 (Wiener-Lévy representation). For a Brownian motion (Bt )t∈[0,1] the
series representation
∞ 2X
−1 n
X
Bt (ω) = Z(ω)e(t) + Zn,k (ω)en,k (t), t ∈ [0, 1], (1.3.7)
n=0 k=0

holds w.r.t. uniform convergence on [0, 1] for P -almost every ω ∈ Ω, where
Z := B1 , and Zn,k := −2n/2 ∆n,k B (n ≥ 0, 0 ≤ k ≤ 2n − 1)
are independent random variables with standard normal distribution.
Proof. It only remains to verify that the coefficients Z and Zn,k are independent with
standard normal distribution. A vector given by finitely many of these random variables
has a multivariate normal distribution, since it is a linear transformation of increments
of the Brownian motion Bt . Hence it suffices to show that the random variables are
uncorrelated with variance 1. This is left as an exercise to the reader.
Lévy’s construction of Brownian motion

The series representation (1.3.7) can be used to construct Brownian motion starting
from independent standard normal random variables. The resulting construction does
not only prove existence of Brownian motion but it is also very useful for numerical
implementations:
Theorem 1.14 (P. Lévy 1948). Let Z and Zn,k (n ≥ 0, 0 ≤ k ≤ 2n − 1) be independent

standard normally distributed random variables on a probability space (Ω, A, P ). Then
the series in (1.3.7) converges uniformly on [0, 1] with probability 1. The limit process
(Bt )t∈[0,1] is a Brownian motion.
The convergence proof relies on a combination of the Borel-Cantelli Lemma and the
Weierstrass criterion for uniform convergence of series of functions. Moreover, we will
need the following result to identify the limit process as a Brownian motion:
Lemma 1.15 (Parseval relation for Schauder functions). For any s, t ∈ [0, 1],
∞ 2X
−1n
X
e(t)e(s) + en,k (t)en,k (s) = min(t, s).
n=0 k=0

Proof. Note that for g ∈ H and s ∈ [0, 1], we have

ˆ1
g(s) = g(s) − g(0) = g ′ · I(0,s) = (g, h(s) )H ,
0
´t
where h(s) (t) := I(0,s) = min(s, t). Hence the Parseval relation (1.3.4) applied to
0
the functions h(s) and h(t) yields
X
e(t)e(s) + en,k (t)en,k (s)
n,k
X
= (e, h(t) )(e, h(s) ) + (en,k , h(t) )(en,k , h(s) )
n,k
ˆ1
(t) (s)
= (h , h ) = I(0,t) I(0,s) = min(t, s).
0
Proof of Theorem 1.14. We proceed in 4 steps:

(1). Uniform convergence for P -a.e. ω: By the Weierstrass criterion, a series of func-
tions converges uniformly if the sum of the supremum norms of the summands is
finite. To apply the criterion, we note that for any fixed t ∈ [0, 1] and n ∈ N, only
one of the functions en,k , k = 0, 1, . . . , 2n − 1, does not vanish at t. Moreover,
|en,k (t)| ≤ 2−n/2 . Hence
2n −1
X

sup Zn,k (ω)en,k (t) ≤ 2−n/2 · Mn (ω), (1.3.8)
t∈[0,1] k=0

where
Mn := max |Zn,k |.
0≤k<2n
We now apply the Borel-Cantelli Lemma to show that with probability 1, Mn
grows at most linearly. Let Z denote a standard normal random variable. Then
we have
n 2n
P [Mn > n] ≤ 2 · P [|Z| > n] ≤ · E[|Z| ; |Z| > n]
n
ˆ∞ r
2 · 2n −x2 /2 2 2n −n2 /2
= √ xe dx = ·e
n · 2π π n
n

for any n ∈ N. Since the sequence on the right hand side is summable, Mn ≤ n
holds eventually with probability one. Therefore, the sequence on the right hand
side of (1.3.8) is also summable for P -almost every ω. Hence, by (1.3.8) and the
Weierstrass criterion, the partial sums
m 2X
−1 n
(m)
X
Bt (ω) = Z(ω)e(t) + Zn,k (ω)en,k (t), m ∈ N,
n=0 k=0
converge almost surely uniformly on [0, 1]. Let

(m)
Bt = lim Bt
m→∞
denote the almost surely defined limit.
(2). L2 convergence for fixed t: We now want to prove that the limit process (Bt )
is a Brownian motion, i.e., a continuous Gaussian process with E[Bt ] = 0 and
Cov[Bt , Bs ] = min(t, s) for any t, s ∈ [0, 1]. To compute the covariances we first
(m)
show that for a given t ∈ [0, 1] the series approximation Bt of Bt converges
also in L . Let l, m ∈ N with l < m. Since the Zn,k are independent (and hence
2
uncorrelated) with variance 1, we have

 !2 
m 2X n −1 m X
(m) (l) 2
X X
E[(Bt − Bt ) ] = E  Zn,k en,k (t)  = en,k (t)2 .
n=l+1 k=0 n=l+1 k
P
The right hand side converges to 0 as l, m → ∞ since en,k (t)2 < ∞ by Lemma
n,k
(m)
1.15. Hence Bt , m ∈ N, is a Cauchy sequence in L2 (Ω, A, P ). Since Bt =
(m)
lim Bt almost surely, we obtain
m→∞
(m) m→∞
Bt −→ Bt in L2 (Ω, A, P ).
(3). Expectations and Covariances: By the L2 convergence we obtain for any s, t ∈

[0, 1]:
(m)
E[Bt ] = lim E[Bt ] = 0, and
m→∞
(m)
Cov[Bt , Bs ] = E[Bt Bs ] = lim E[Bt Bs(m) ]
m→∞
n −1
m 2X
X
= e(t)e(s) + lim en,k (t)en,k (s).
m→∞
n=0 k=0

Here we have used again that the random variables Z and Zn,k are independent
with variance 1. By Parseval’s relation (Lemma 1.15), we conclude
Cov[Bt , Bs ] = min(t, s).
Since the process (Bt )t∈[0,1] has the right expectations and covariances, and, by
construction, almost surely continuous paths, it only remains to show that (Bt ) is
a Gaussian process in oder to complete the proof:
(4). (Bt )t∈[0,1] is a Gaussian process: We have to show that (Bt1 , . . . , Btl ) has a mul-
tivariate normal distribution for any 0 ≤ t1 < . . . < tl ≤ 1. By Theorem 1.5,
it suffices to verify that any linear combination of the components is normally
distributed. This holds by the next Lemma since
l
X l
X (m)
pj Btj = lim pj Btj P -a.s.
m→∞
j=1 j=1
is an almost sure limit of normally distributed random variables for any

p1 , . . . , pl ∈ R.
Combining Steps 3, 4 and the continuity of sample paths, we conclude that (Bt )t∈[0,1] is
indeed a Brownian motion.
Lemma 1.16. Suppose that (Xn )n∈N is a sequence of normally distributed random vari-
ables defined on a joint probability space (Ω, A, P ), and Xn converges almost surely to
a random variable X. Then X is also normally distributed.
Proof. Suppose Xn ∼ N(mn , σn2 ) with mn ∈ R and σn ∈ (0, ∞). By the Dominated
Convergence Theorem,
1 2 2
E[eipX ] = lim E[eipXn ] = lim eipmn e− 2 σn p .
n→∞ n→∞
The limit on the right hand side only exists for all p, if either σn → ∞, or the sequences
σn and mn both converge to finite limits σ ∈ [0, ∞) and m ∈ R. In the first case,
the limit would equal 0 for p 6= 0 and 1 for p = 0. This is a contradiction, since

characteristic functions are always continuous. Hence the second case occurs, and,
therefore
1 2 p2
E[eipX ] = eipm− 2 σ for any p ∈ R,
i.e., X ∼ N(m, σ 2 ).
So far, we have constructed Brownian motion only for t ∈ [0, 1]. Brownian motion on
any finite time interval can easily be obtained from this process by rescaling. Brownian
motion defined for all t ∈ R+ can be obtained by joining infinitely many Brownian
motions on time intervals of length 1:
B (2)
B (3)
1 2 3
B (1)
(1) (2)
Theorem 1.17. Suppose that Bt , Bt , . . . are independent Brownian motions starting
at 0 defined for t ∈ [0, 1]. Then the process
⌊t⌋
(⌊t⌋+1)
X (i)
Bt := Bt−⌊t⌋ + B1 , t ≥ 0,
i=1
is a Brownian motion defined for t ∈ [0, ∞).

The proof is left as an exercise.
1.4 The Brownian Sample Paths

In this section we study some properties of Brownian sample paths in dimension one.
We show that a typical Brownian path is nowhere differentiable, and Hölder-continuous
with parameter α if and only if α < 1/2. Furthermore, the set Λa = {t ≥ 0 : Bt = a}
of all passage times of a given point a ∈ R is a fractal. We will show that almost surely,
Λa has Lebesgue measure zero but any point in Λa is an accumulation point of Λa .
We consider a one-dimensional Brownian motion (Bt )t≥0 with B0 = 0 defined on a
probability space (Ω, A, P ). Then:
Typical Brownian sample paths are nowhere differentiable

Bt+h −Bt
For any t ≥ 0 and h > 0, the difference quotient h
is normally distributed with
mean 0 and standard deviation
√
σ[(Bt+h − Bt )/h] = σ[Bt+h − Bt ]/h = 1/ h.
This suggests that the derivative

d Bt+h − Bt
Bt = lim
dt hց0 h
does not exist. Indeed, we have the following stronger statement.
Theorem 1.18 (Paley, Wiener, Zygmund 1933). Almost surely, the Brownian sample
path t 7→ Bt is nowhere differentiable, and

Bs − Bt

lim sup = ∞ for any t ≥ 0.
sցt s − t

1.4. THE BROWNIAN SAMPLE PATHS 53
Note that, since there are uncountably many t ≥ 0, the statement is stronger than claim-
ing only the almost sure non-differentiability for any given t ≥ 0.
Proof. It suffices to show that the set

1
N=
ω ∈ Ω ∃ t ∈ [0, T ], k, L ∈ N ∀ s ∈ (t, t + ) : |Bs (ω) − Bt (ω)| ≤ L|s − t|
k
is a null set for any T ∈ N. Hence fix T ∈ N, and consider ω ∈ N. Then there exist
k, L ∈ N and t ∈ [0, T ] such that
1
|Bs (ω) − Bt (ω)| ≤ L · |s − t| holds for s ∈ (t, t + ). (1.4.1)
k
To make use of the independence of the increments over disjoint intervals, we note that
for any n > 4k, we can find an i ∈ {1, 2, . . . , nT } such that the intervals ( ni , i+1
n
),
( i+1
n
, i+2
n
), and ( i+2
n
, i+3
n
) are all contained in (t, t + k1 ):
i−1 i i+1 i+2 i+3

n n n n n
t t+ 1
k
1/k > 4/n
Hence by (1.4.1), the bound

B j+1 (ω) − B j (ω) ≤ B j+1 (ω) − Bt (ω) + Bt (ω) − B j (ω)
n n n n
j+1 j 8L
≤ L·( − t) + L · ( − t) ≤
n n n
holds for j = i, i + 1, i + 2. Thus we have shown that N is contained in the set
[ \ [ nT
e 8L
N := B j+1 − B j ≤ for j = i, i + 1, i + 2 .
k,L∈N n>4k i=1
n n n

e = 0. By independence and stationarity of the increments we have

We now prove P [N]

8L
P B j+1 − B j ≤ for j = i, i + 1, i + 2
n n n
3 3
8L 8L
= P B 1 ≤ = P |B1 | ≤ √ (1.4.2)
n n n
3
1 16L 163 L3
≤ √ √ = √ 3 · 3/2
2π n 2π n
for any i and n. Here we have used that the standard normal density is bounded from
√
above by 1/ 2π. By (1.4.2) we obtain
" nT #
\ [ 8L

P B j+1 − B j ≤ for j = i, i + 1, i + 2
n>4k i=1
n n n
163
≤ √ 3 · inf nT L3 /n3/2 = 0.
2π n>4k
e = 0, and therefore N is a null set.
Hence, P [N]
Hölder continuity
The statement of Theorem 1.18 says that a typical Brownian path is not Lipschitz contin-
uous on any non-empty open interval. On the other hand, the Wiener-Lévy construction
shows that the sample paths are continuous. We can almost close the gap between these
two statements by arguing in both cases slightly more carefully:
Theorem 1.19. The following statements hold almost surely:
(1). For any α > 1/2,

|Bs − Bt |
lim sup = ∞ for all t ≥ 0.
sցt |s − t|α
(2). For any α < 1/2,

|Bs − Bt |
sup α
< ∞ for all T > 0.
s,t∈[0,T ] |s − t|
s6=t

Hence a typical Brownian path is nowhere Hölder continuous with parameter α > 1/2,
but it is Hölder continuous with parameter α < 1/2 on any finite interval. The critical
case α = 1/2 is more delicate, and will be briefly discussed below.
Proof of Theorem 1.19. The first statement can be shown by a similar argument as in
the proof of Theorem 1.18. The details are left to the reader.
To prove the second statement for T = 1, we use the Wiener-Lévy representation
∞ 2X
−1 n
X
Bt = Z ·t+ Zn,k en,k (t) for any t ∈ [0, 1]
n=0 k=0
with independent standard normal random variables Z, Zn,k . For t, s ∈ [0, 1] we obtain
X X
|Bt − Bs | ≤ |Z| · |t − s| + Mn |en,k (t) − en,k (s)|,
n k
where Mn := max |Zn,k | as in the proof of Theorem 1.14. We have shown above that
k
by the Borel-Cantelli Lemma, Mn ≤ n eventually with probability one, and hence
Mn (ω) ≤ C(ω) · n
for some almost surely finite constant C(ω). Moreover, note that for each s, t and n, at
P
most two summands in k |en,k (t) − en,k (s)| do not vanish. Since |en,k (t)| ≤ 21 · 2−n/2
and |e′n,k (t)| ≤ 2n/2 , we obtain the estimates
|en,k (t) − en,k (s)| ≤ 2−n/2 , and (1.4.3)

|en,k (t) − en,k (s)| ≤ 2n/2 · |t − s|. (1.4.4)
For given s, t ∈ [0, 1], we now choose N ∈ N such that
2−N ≤ |t − s| < 21−N . (1.4.5)
By applying (1.4.3) for n > N and (1.4.4) for n ≤ N, we obtain

N ∞
!
X X
|Bt − Bs | ≤ |Z| · |t − s| + 2C · n2n/2 · |t − s| + n2−n/2 .
n=1 n=N +1
By (1.4.5) the sums on the right hand side can both be bounded by a constant multiple of
|t − s|α for any α < 1/2. This proves that (Bt )t∈[0,1] is almost surely Hölder-continuous
of order α.

Law of the iterated logarithm

Khintchine’s version of the law of the iterated logarithm is a much more precise state-
ment on the local regularity of a typical Brownian path at a fixed time s ≥ 0. It implies
in particular that almost every Brownian path is not Hölder continuous with parameter
α = 1/2. We state the result without proof:
Theorem 1.20 (Khintchine 1924). For s ≥ 0, the following statements hold almost
surely:
Bs+t − Bs Bs+t − Bs
lim sup p = +1, and lim inf p = −1.
tց0 2t log log(1/t) tց0 2t log log(1/t)
For the proof cf. e.g. Breiman, Probability, Section 12.9.

By a time inversion, the Theorem translates into a statement on the global asymptotics
of Brownian paths:
Corollary 1.21. The following statements hold almost surely:

Bt Bt
lim sup √ = +1, and lim inf √ = −1.
t→∞ 2t log log t t→∞ 2t log log t
bt =
Proof. This follows by applying the Theorem above to the Brownian motion B
t · B1/t . For example, substituting h = 1/t, we have
Bt h · B1/h
lim sup p = lim sup p = +1
t→∞ 2t log log(t) hց0 2h log log 1/h
almost surely.
The corollary is a continuous time analogue of Kolmogorov’s law of the iterated log-
Pn
arithm for Random Walks stating that for Sn = ηi , ηi i.i.d. with E[ηi ] = 0 and
i=1
Var[ηi ] = 1, one has
Sn Sn
lim sup √ = +1 and lim inf √ = −1
n→∞ 2n log log n n→∞ 2n log log n

almost surely. In fact, one way to prove Kolmogorov’s LIL is to embed the Random
Walk into a Brownian motion, cf. e.g. Rogers and Williams, Vol. I, Ch. 7 or Section 3.3
Passage times
We now study the set of passage times to a given level a for a one-dimensional Brownian
motion (Bt )t≥0 . This set has interesting properties – in particular it is a random fractal.
Fix a ∈ R, and let
Λa (ω) = {t ≥ 0 : Bt (ω) = a} ⊆ [0, ∞).
Assuming that every path is continuous, the random set Λa (ω) is closed for every ω.
Moreover, scale invariance of Brownian motion implies a statistical self similarity prop-
erty for the sets of passage times: Since the rescaled process (c−1/2 Bct )t≥0 has the same
distribution as (Bt )t≥0 for any c > 0, we can conclude that the set valued random vari-
able c · Λa/√c has the same distribution as Λa . In particular, Λ0 is a fractal in the sense
that
Λ0 ∼ c · Λ0 for any c > 0.
bbbb b b bb b b b bbb b b b b b b b bb b bbbb b b
Figure 1: Brownian motion with corresponding level set Λ0 .
Figure 1.4: Brownian motion with corresponding level set Λ0 .

Moreover, by Fubini’s Theorem one easily verifies that Λa has almost surely Lebesgue
measure zero. In fact, continuity of t 7→ Bt (ω) for any ω implies that (t, ω) 7→ Bt (ω) is
product measurable (Exercise). Hence {(t, ω) : Bt (ω) = a} is contained in the product
σ-algebra, and
∞ 
ˆ ˆ∞
E[λ(Λa )] = E  I{a} (Bt ) dt = P [Bt = a] dt = 0.
0 0
Theorem 1.22 (Unbounded oscillations, recurrence).

P sup Bt = +∞ = P inf Bt = −∞ = 1.
t≥0 t≥0
In particular, for any a ∈ R, the random set Λa is almost surely unbounded, i.e. Brow-
nian motion is recurrent.
Proof. By scale invariance,
sup Bt ∼ c−1/2 sup Bct = c−1/2 sup Bt for any c > 0.

t≥0 t≥0 t≥0
Hence,
√
P sup Bt ≥ a = P sup Bt ≥ a · c
t≥0 t≥0
for any c > 0, and therefore sup Bt ∈ {0, ∞} almost surely. The first part of the asser-
tion now follows since sup Bt is almost surely strictly positive. By reflection symmetry,
we also obtain inf Bt = −∞ with probability one.
The last Theorem makes a statement on the global structure of the set Λa . By invariance
w.r.t. time inversion this again translates into a local regularity result:
Theorem 1.23 (Fine structure of Λa ). The set Λa is almost surely a perfect set, i.e., any
t ∈ Λa is an accumulation point of Λa .

Proof. We prove the statement for a = 0, the general case being left as an exercise. We
proceed in three steps:
S TEP 1: 0 is almost surely an accumulation point of Λ0 : This holds by time-reversal.

Setting Bbt = t · B1/t , we see that 0 is an accumulation point of Λ0 if and only of
bt = 0, i.e., if and only if the zero set
for any n ∈ N there exists t > n such that B
bt is unbounded. By Theorem 1.22, this holds almost surely.
of B
S TEP 2: For any s ≥ 0, Ts := min(Λa ∩ [s, ∞)) = min{t ≥ s : Bt = a} is almost

surely an accumulation point of Λa : For the proof we need the strong Markov
property of Brownian motion which will be proved in the next section. By The-
orem 1.22, the random variable Ts is almost surely finite. Hence, by continuity,
BTs = a almost surely. The strong Markov property says that the process
et := BTs +t − BTs ,
B t ≥ 0,
is again a Brownian motion starting at 0. Therefore, almost surely, 0 is an accu-

et by Step 1. The claim follows since almost
mulation point of the zero set of B
surely
et = 0} = {t ≥ 0 : BTs +t = BTs } = {t ≥ Ts : Bt = a} ⊆ Λa .
{t ≥ 0 : B
S TEP 3: To complete the proof note that we have shown that the following properties
hold with probability one:
(1). Λa is closed.
(2). min(Λa ∩ [s, ∞)) is an accumulation point of Λa for any s ∈ Q+ .
Since Q+ is a dense subset of R+ , (1) and (2) imply that any t ∈ Λa is an accu-
mulation point of Λa . In fact, for any s ∈ [0, t] ∩ Q, there exists an accumulation
point of Λa in (s, t] by (2), and hence t is itself an accumulation point.
Remark. It can be shown that the set Λa has Hausdorff dimension 1/2.

1.5 Strong Markov property and reflection principle

In this section we prove a strong Markov property for Brownian motion. Before, we give
another motivation for our interest in an extension of the Markov property to random
times.
Maximum of Brownian motion
Suppose that (Bt )t≥0 is a one-dimensional continuous Brownian motion starting at 0

defined on a probability space (Ω, A, P ). We would like to compute the distribution of
the maximal value
Ms = max Bt
t∈[0,s]
attained before a given time s ∈ R+ . The idea is to proceed similarly as for Random
Walks, and to reflect the Brownian path after the first passage time
Ta = min{t ≥ 0 : Bt = a}
to a given level a > 0:
bt
B
Ta
Bt

1.5. STRONG MARKOV PROPERTY AND REFLECTION PRINCIPLE 61
It seems plausible (e.g. by the heuristic path integral representation of Wiener measure,
or by a Random Walk approximation) that the reflected process (B bt )t≥0 defined by

Bt for t ≤ Ta
Bbt :=
a − (B − a) for t > T
t a
is again a Brownian motion. At the end of this section, we will prove this reflection
principle rigorously by the strong Markov property. Assuming the reflection principle
is true, we can compute the distribution of Ms in the following way:
P [Ms ≥ a] = P [Ms ≥ a, Bs ≤ a] + P [Ms ≥ a, Bs > a]

bs ≥ a] + P [Bs > a]
= P [B
= 2 · P [Bs ≥ a]
= P [|Bs | ≥ a].
Thus Ms has the same distribution as |Bs |.

cs = max{B
Furthermore, since Ms ≥ a if and only if M bt : t ∈ [0, s]} ≥ a, we obtain
the stronger statement
cs ≥ a, B
P [Ms ≥ a, Bs ≤ c] = P [M bs ≥ 2a − c] = bs ≥ 2a − c]
P [B
ˆ∞
1
= √ exp(−x2 /2s) dx
2πs
2a−c
for any a ≥ 0 and c ≤ a. As a consequence, we have:
Theorem 1.24 (Maxima of Brownian paths).
(1). For any s ≥ 0, the distribution of Ms is absolutely continuous with density

2
fMs (x) = √ exp(−x2 /2s) · I(0,∞) (x).
2πs
(2). The joint distribution of Ms and Bs is absolutely continuous with density

2x − y (2x − y)2
fMs ,Bs (x, y) = 2 √ exp − I(0,∞) (x)I(−∞,x) (y).
2πs3 2s

Proof. (1) holds since Ms ∼ |Bs |. For the proof of (2) we assume w.l.o.g. s = 1. The
general case can be reduced to this case by the scale invariance of Brownian motion
(Exercise). For a ≥ 0 and c ≤ a let
G(a, c) := P [M1 ≥ a, B1 ≤ c].
By the reflection principle,
G(a, c) = P [B1 ≥ 2a − c] = 1 − Φ(2a − c),
where Φ denotes the standard normal distribution function. Since lim G(a, c) = 0 and
a→∞
lim G(a, c) = 0, we obtain
c→−∞
ˆ∞ ˆc
∂2G
P [M1 ≥ a, B1 ≤ c] = G(a, c) = − (x, y) dydx
∂x∂y
x=a y=−∞
ˆ∞ ˆc
2x − y (2x − y)2
= 2· √ · exp − dydx.
2π 2
x=a y=−∞
This implies the claim for s = 1, since M1 ≥ 0 and B1 ≤ M1 by definition of M1 .
The Theorem enables us to compute the distributions of the first passage times Ta . In
fact, for a > 0 and s ∈ [0, ∞) we obtain
√
P [Ta ≤ s] = P [Ms ≥ a] = 2 · P [Bs ≥ a] = 2 · P [B1 ≥ a/ s]
r ˆ∞
2 2
= e−x /2 dx. (1.5.1)
π √
a/ s
Corollary 1.25 (Distribution of Ta ). For any a ∈ R \ {0}, the distribution of Ta is

absolutely continuous with density
|a| 2 /2s
fTa (s) = √ · e−a .
2πs3

Proof. For a > 0, we obtain

a 2 /2s
fTa (s) = FT′ a (s) = √ e−a
2πs3
by (1.5.1). For a < 0 the assertion holds since Ta ∼ T−a by reflection symmetry of
Brownian motion.
Next, we prove a strong Markov property for Brownian motion. Below we will then
complete the proof of the reflection principle and the statements above by applying the
strong Markov property to the passage time Ta .
Strong Markov property for Brownian motion

Suppose again that (Bt )t≥0 is a d-dimensional continuous Brownian motion starting at
0 on a probability space (Ω, A, P ), and let
FtB = σ(Bs : 0 ≤ s ≤ t), t ≥ 0,
denote the σ-algebras generated by the process up to time t.
Definition. A random variable T : Ω → [0, ∞] is called an (FtB )-stopping time if and

only if
{T ≤ t} ∈ FtB for any t ≥ 0.
Example. Clearly, for any a ∈ R, the first passage time
Ta = min{t ≥ 0 : Bt = a}
to a level a is an (FtB )-stopping time.

The σ-algebra FTB describing the information about the process up to a stopping time T
is defined by
FTB = {A ∈ A : A ∩ {T ≤ t} ∈ FtB for any t ≥ 0}.

Note that for (FtB ) stopping times S and T with S ≤ T we have FSB ⊆ FTB , since for
t≥0
A ∩ {S ≤ t} ∈ FtB =⇒ A ∩ {T ≤ t} = A ∩ {S ≤ t} ∩ {T ≤ t} ∈ FtB .
For any constant s ∈ R+ , the process (Bs+t − Bs )t≥0 is a Brownian motion independent
of FsB .
A corresponding statement holds for stopping times:
Theorem 1.26 (Strong Markov property). Suppose that T is an almost surely finite
et )t≥0 defined by
(FtB ) stopping time. Then the process (B
et = BT +t − BT
B if T < ∞, 0 otherwise,
is a Brownian motion independent of FTB .
Proof. We first assume that T takes values only in C ∪ {∞} where C is a countable
subset of [0, ∞). Then for A ∈ FTB and s ∈ C, we have A ∩ {T = s} ∈ FsB and
Bet = Bt+s −Bs on A∩{T = s}. By the Markov property, (Bt+s −Bs )t≥0 is a Brownian
motion independent of FsB . Hence for any measurable subset Γ of C([0, ∞], Rd), we
have
X
et )t≥0 ∈ Γ} ∩ A] =
P [{(B P [{(Bt+s − Bs )t≥0 ∈ Γ} ∩ A ∩ {T = s}]
s∈C
X
= µ0 [Γ] · P [A ∩ {T = s}] = µ0 [Γ] · P [A]
s∈C
where µ0 denotes the distribution of Brownian motion starting at 0. This proves the
assertion for discrete stopping times.
For an arbitrary (FtB ) stopping time T that is almost surely finite and n ∈ N, we set
Tn = n1 ⌈nT ⌉, i.e.,

k k−1 k
Tn = on <T ≤ for any k ∈ N.
n n n

Since the event {Tn = k/n} is Fk/n

B
-measurable for any k ∈ N, Tn is a discrete (FtB )
stopping time. Therefore, (BTn +t − BTn )t≥0 is a Brownian motion that is independent
of FTBn , and hence of the smaller σ-algebra FTB . As n → ∞, Tn → T , and thus, by
continuity,
et = BT +t − BT = lim (BTn +t − BTn ).
B
n→∞
et )t≥0 is again a Brownian motion that is independent of

Now it is easy to verify that (B
FTB .
A rigorous reflection principle
We now apply the strong Markov property to prove a reflection principle for Brownian
motion. Consider a one-dimensional continuous Brownian motion (Bt )t≥0 starting at 0.
For a ∈ R let
Ta = min{t ≥ 0 : Bt = a} (first passage time),

BtTa = Bmin{t,Ta } (process stopped at Ta ), and
Bet = BTa +t − BTa (process after Ta ).
Theorem 1.27 (Reflection principle). The joint distributions of the following random
variables with values in R+ × C([0, ∞)) × C([0, ∞)) agree:
et )t≥0 )
(Ta , (BtTa )t≥0 , (B ∼ et )t≥0 )
(Ta , (BtTa )t≥0 , (−B
e is a Brownian motion starting at

Proof. By the strong Markov property, the process B
0 independent of FTa , and hence of Ta and B Ta = (BtTa )t≥0 . Therefore,
e −1 = P ◦ (Ta , B Ta )−1 ⊗ µ0 = P ◦ (Ta , B Ta , −B)

P ◦ (Ta , B Ta , B) e −1 .

Bt
a
bt
B
Ta
As a consequence of the theorem, we can complete the argument given at the beginning
of this section: The "shadow path" Bbt of a Brownian path Bt with reflection when
reaching the level a is given by

B Ta for t ≤ Ta
bt = t
B ,
a − B
et−Ta for t > Ta
whereas 
B Ta for t ≤ Ta
t
Bt = .
a + B
e for t > Ta
t−Ta
bt )t≥0 has the same distribution as (Bt )t≥0 . Therefore, and since
By the Theorem 1.27, (B
max Bt ≥ a if and only if max B bt ≥ a, we obtain for a ≥ c:
t∈[0,s] t∈[0,s]

P max Bt ≥ a, Bs ≤ c = P max B bt ≥ a, B
bs ≥ 2a − c
t∈[0,s] t∈[0,s]
h i
= P B bs ≥ 2a − c
ˆ∞
1 2 /2s
= √ e−x dx.
2πs
2a−c

Part II
Introduction to Stochastic Analysis
67
Chapter 2
Martingales in discrete time
Classical analysis starts with studying convergence of sequences of real numbers. Sim-
ilarly, stochastic analysis relies on basic statements about sequences of real-valued ran-
dom variables. Any such sequence can be decomposed uniquely into a martingale, i.e.,
a real.valued stochastic process that is “constant on average”, and a predictable part.
Therefore, estimates and convergence theorems for martingales are crucial in stochastic
analysis.
2.1 Definitions and examples

We fix a probability space (Ω, A, P ). Moreover, we assume that we are given an in-
creasing sequence Fn (n = 0, 1, 2, . . .) of sub-σ-algebras of A. Intuitively, we often
think of Fn as describing the information available to us at time n. Formally, we define:
Definition (Filtration, adapted process). (1). A filtration on (Ω, A) is an increasing

sequence
F0 ⊆ F1 ⊆ F2 ⊆ . . .
of σ-algebras Fn ⊆ A.
(2). A stochastic process (Xn )n≥0 is adapted to a filtration (Fn )n≥0 iff each Xn is
Fn -measurable.
68
2.1. DEFINITIONS AND EXAMPLES 69
Example. (1). The canonical filtration (FnX ) generated by a stochastic process (Xn )
is given by
FnX = σ(X0 , X1 , . . . , Xn ).
If the filtration is not specified explicitly, we will usually consider the canonical
filtration.
(2). Alternatively, filtrations containing additional information are of interest, for ex-
ample the filtration
Fn = σ(Z, X0 , X1 , . . . , Xn )
generated by the process (Xn ) and an additional random variable Z, or the filtra-
tion
Fn = σ(X0 , Y0, X1 , Y1 , . . . , Xn , Yn )
generated by the process (Xn ) and a further process (Yn ).
Clearly, the process (Xn ) is adapted to any of these filtrations. In general, (Xn ) is
adapted to a filtration (Fn ) if and only if FnX ⊆ Fn for any n ≥ 0.
Martingales and supermartingales

We can now formalize the notion of a real-valued stochastic process that is constant
(respectively decreasing or increasing) on average:
Definition (Martingale, supermartingale, submartingale). (1). A sequence of real-

valued random variables Mn : Ω → R (n = 0, 1, . . .) on the probability space
(Ω, A, P ) is called a martingale w.r.t. the filtration (Fn ) if and only if
(a) (Mn ) is adapted w.r.t. (Fn ),

(b) Mn is integrable for any n ≥ 0, and
(c) E[Mn | Fn−1 ] = Mn−1 for any n ∈ N.

70 CHAPTER 2. MARTINGALES IN DISCRETE TIME
(2). Similarly, (Mn ) is called a supermartingale (resp. a submartingale) w.r.t. (Fn )

if and only if (a) holds, the positive part Mn+ (resp. the negative part Mn− ) is inte-
grable for any n ≥ 0, and (c) holds with “=” replaced by “≤”, “≥” respectively.
Condition (c) in the martingale definition can equivalently be written as
(c’) E[Mn+1 − Mn | Fn ] = 0 for any n ∈ Z+ ,
and correspondingly with “=” replaced by “≤” or “≥” for super- or submartingales.
Intuitively, a martingale is a ”fair game´´, i.e., Mn−1 is the best prediction (w.r.t. the
mean square error) for the next value Mn given the information up to time n − 1. A su-
permartingale is “decreasing on average”, a submartingale is “increasing on average”,
and a martingale is both “decreasing” and “increasing”, i.e., “constant on average”. In
particular, by induction on n, a martingale satisfies
E[Mn ] = E[M0 ] for any n ≥ 0.
Similarly, for a supermartingale, the expectation values E[Mn ] are decreasing. More
generally, we have:
Lemma 2.1. If (Mn ) is a martingale (respectively a supermartingale) w.r.t. a filtration

(Fn ) then
(≤)
E[Mn+k | Fn ] = Mn P -almost surely for any n, k ≥ 0.
Proof. By induction on k: The assertion holds for k = 0, since Mn is Fn -measurable.

Moreover, the assertion for k − 1 implies

E[Mn+k | Fn ] = E E[Mn+k | Fn+k−1] Fn
= E[Mn+k−1 | Fn ] = Mn P -a.s.
by the tower property for conditional expectations.

Remark (Supermartingale Convergence Theorem). A key fact in analysis is that

any lower bounded decreasing sequence of real numbers converges to its infimum. The
counterpart of this result in stochastic analysis is the Supermartingale Convergence The-
orem: Any lower bounded supermartingale converges almost surely, cf. Theorem 4.5
below.
Some fundamental examples

a) Sums of independent random variables
A Random Walk n
X
Sn = ηi , n = 0, 1, 2, . . . ,
i=1
with independent increments ηi ∈ L1 (Ω, A, P ) is a martingale w.r.t. to the filtration
Fn = σ(η1 , . . . , ηn ) = σ(S0 , S1 , . . . , Sn )
if and only if the increments ηi are centered random variables. In fact, for any n ∈ N,
E[Sn − Sn−1 | Fn−1 ] = E[ηn | Fn−1 ] = E[ηn ]
by independence of the increments. Correspondingly, (Sn ) is an (Fn ) supermartingale

if and only if E[ηi ] ≤ 0 for any i ∈ N.
b) Products of independent non-negative random variables
A stochastic process
n
Y
Mn = Yi , n = 0, 1, 2, . . . ,
i=1
with independent non-negative factors Yi ∈ L1 (Ω, A, P ) is a martingale respectively a

supermartingale w.r.t. the filtration
Fn = σ(Y1 , . . . , Yn )
if and only if E[Yi ] = 1 for any i ∈ N, or E[Yi ] ≤ 1 for any i ∈ N respectively. In fact,
as Mn is Fn -measurable and Yn+1 is independent of Fn , we have
E[Mn+1 | Fn ] = E[Mn · Yn+1 | Fn ] = Mn · E[Yn+1] for any n ≥ 0.

Martingales and supermartingales of this type occur naturally in stochastic growth mod-
els.
Pn
Example (Exponential martingales). Consider a Random Walk Sn = i=1 ηi with
i.i.d. increments ηi , and let
Z(λ) = E[exp(ληi )], λ ∈ R,
denote the moment generating function of the increments. Then for any λ ∈ R with
Z(λ) < ∞, the process
n
Y
Mnλ := eλSn /Z(λ)n = eληi /Z(λ)
i=1
is a martingale. This martingale can be used to prove exponential bounds for Ran-
dom Walks, cf. e.g. Chernov’s theorem [“Einführung in die Wahrscheinlichkeitstheo-
rie”, Theorem 8.3].
Example (CRR model of stock market). In the Cox-Ross-Rubinstein binomial model

of mathematical finance, the price of an asset is changing during each period either by
a factor 1 + a or by a factor 1 + b with a, b ∈ (−1, ∞) such that a < b. We can model
the price evolution in a fixed number N of periods by a stochastic process
n
Y
Sn = S0 · Xi , n = 0, 1, 2, . . . , N,
i=1
defined on Ω = {1 + a, 1 + b}N , where the initial price S0 is a given constant, and

Xi (ω) = ωi . Taking into account a constant interest rate r > 0, the discounted stock
price after n periods is
Yn
Xi
Sen = Sn /(1 + r)n = S0 · .
i=1
1+r
A probability measure P on Ω is called a martingale measure if the discounted stock

price is a martingale w.r.t. P and the filtration Fn = σ(X1 , . . . , Xn ). Martingale mea-
sures are important for option pricing under no arbitrage assumptions, cf. Section 2.3
below. For 1 ≤ n ≤ N,

e
E[Sn | Fn−1] = E Sen−1 ·
Xn Fn−1 = Sen−1 · E[Xn | Fn−1 ] .
1+r 1+r

Hence (Sen ) is an (Fn ) martingale w.r.t. P if and only if
E[Xn | Fn−1 ] = 1+r for any 1 ≤ n ≤ N. (2.1.1)
On the other hand, since in the CRR model Xn only takes the values 1 + a and 1 + b,
we have
E[Xn | Fn−1 ] = (1 + a) · P [Xn = 1 + a | Fn−1] + (1 + b) · P [Xn = 1 + b | Fn−1 ]

= 1 + a + (b − a) · P [Xn = 1 + b | Fn−1].
Therefore, by (2.1.1), (Sen ) is a martingale if and only if

r−a
P [Xn = 1 + b | Fn−1 ] = for any n = 1, . . . , N,
b−a
i.e., if and only if the growth factors X1 , . . . , XN are independent with
r−a b−r
P [Xn = 1 + b] = and P [Xn = 1 + a] = . (2.1.2)
b−a b−a
Hence for r 6∈ [a, b], a martingale measure does not exist, and for r ∈ [a, b], the product
measure P on Ω satisfying (2.1.2) is the unique martingale measure. Intuitively this
is plausible: If r < a or r > b respectively, then the stock price is always growing
more or less than the discount factor (1 + r)n , so the discounted stock price can not be
a martingale. If, on the other hand, a < r < b, then (Sen ) is a martingale provided the
growth factors are independent with
P [Xn = 1 + b] (1 + r) − (1 + a)
= .
P [Xn = 1 + a] (1 + b) − (1 + r)
We remark, however, that uniqueness of the martingale measure only follows from
(2.1.1) since we have assumed that each Xn takes only two possible values (binomial
model). In a corresponding trinomial model there are infinitely many martingale mea-
sures!
c) Successive prediction values
Let F be an integrable random variable, and let (Fn ) be a filtration on a probability

space (Ω, A, P ). Then the process
Mn := E[F | Fn ], n = 0, 1, 2, . . . ,

of successive prediction values for F based on the information up to time n is a martin-

gale. Indeed, by the tower property for conditional expectations, we have

E[Mn | Fn−1 ] = E E[F | Fn ] Fn−1 = E F Fn−1 = Mn−1
almost surely for any n ∈ N.

Remark (Representing martingales as successive prediction values). The class of
martingales that have a representation as successive prediction values almost contains
general martingales. In fact, for an arbitrary (Fn ) martingale (Mn ) and any finite integer
m ≥ 0, the representation
Mn = E[Mm | Fn ]
holds for any n = 0, 1, . . . , m. Moreover, the L1 Martingale Convergence Theorem
implies that under a uniform integrability assumption, the limit M∞ = lim Mm exists
m→∞
in L1 , and the representation
Mn = E[M∞ | Fn ]
holds for any n ≥ 0, see Section 4.3 below .
d) Functions of martingales
By Jensen’s inequality for conditional expectations, convex functions of martingales are

submartingales, and concave functions of martingales are supermartingales:
Theorem 2.2 (Convex functions of martingales). Suppose that (Mn )n≥0 is an (Fn )
martingale, and u : R → R is a convex function that is bounded from below. Then
(u(Mn )) is an (Fn ) submartingale.
Proof. Since u is lower bounded, u(Mn )− is integrable for any n. Jensen’s inequality
for conditional expectations now implies

E[u(Mn+1 ) | Fn ] ≥ u E[Mn+1 | Fn ] = u(Mn )
almost surely for any n ≥ 0.

Example. If (Mn ) is a martingale then (|Mn |p ) is a submartingale for any p ≥ 1.

e) Functions of Markov chains
Let p(x, dy) be a transition kernel on a measurable space (S, B).
Definition (Markov chain, superharmonic function). (1). A discrete time stochas-

tic process (Xn )n≥0 with state space (S, B) defined on the probability space
(Ω, A, P ) is called a (time-homogeneous) Markov chain with transition kernel
p w.r.t. the filtration (Fn ), if and only if
(a) (Xn ) is (Fn ) adapted, and

(b) P [Xn+1 ∈ B | Fn ] = p(Xn , B) P -almost surely for any B ∈ B and n ≥ 0.
(2). A measurable function h : S → R is called superharmonic (resp. subharmonic)

w.r.t. p if and only if the integrals
ˆ
(ph)(x) := p(x, dy)h(y), x ∈ S,
exist, and
(ph)(x) ≤ h(x) (respectively (ph)(x) ≥ h(x))
holds for any x ∈ S.

The function h is called harmonic iff it is both super- and subharmonic, i.e., iff
(ph)(x) = h(x) for any x ∈ S.
By the tower property for conditional expectations, any (Fn ) Markov chain is also a
Markov chain w.r.t. the canonical filtration generated by the process.
Example (Classical Random Walk on Zd ). The standard Random Walk (Xn )n≥0 on
Zd is a Markov chain w.r.t. the filtration FnX = σ(X0 , . . . , Xn ) with transition prob-
abilities p(x, x + e) = 1/2d for any unit vector e ∈ Zd . The coordinate processes
(Xni )n≥0 , i = 1, . . . , d, are Markov chains w.r.t. the same filtration with transition prob-
abilities
1 2d − 2
p(x, x + 1) = p(x, x − 1) = , p(x, x) = .
2d 2d
A function h : Zd → R is superharmonic w.r.t. p if and only if

d
X
∆Zd h(x) = = h(x + ei ) − 2h(x) + h(x − ei ) = 2d ((ph)(x) − h(x)) ≤ 0
i=1
for any x ∈ Zd .
A function h : Z → R is harmonic w.r.t. p if and only if h(x) = ax + b with a, b ∈ R,
and h is superharmonic if and only if it is concave.
It is easy to verify that (super-)harmonic functions of Markov chains are (super-)mar-

tingales:
Theorem 2.3 (Superharmonic functions of Markov chains are supermartingales).

Suppose that (Xn ) is an (Fn ) Markov chain. Then the real-valued process
Mn := h(Xn ), n = 0, 1, 2, . . . ,
is a martingale (resp. a supermartingale) w.r.t. (Fn ) for every harmonic (resp. super-
harmonic) function h : S → R such that h(Xn ) (resp. h(Xn )+ ) is integrable for all
n.
Proof. Clearly, (Mn ) is again (Fn ) adapted. Moreover,
E[Mn+1 | Fn ] = E[h(Xn+1 ) | Fn ] = (ph)(Xn ) P -a.s.
The assertion now follows immediately from the definitions.
Below, we will show how to construct more general martingales from Markov chains,
cf. Theorem 2.5. At first, however, we consider a simple example that demonstrates the
usefulness of martingale methods in analyzing Markov chains:
Example (Wright model for evolution). In the Wright model for a population of N
individuals (replicas) with a finite number of possible types, each individual in genera-
tion n + 1 inherits a type from a randomly chosen predecessor in the n th generation.

The number Xn of individuals of a given type in generation n is a Markov chain with

state space S = {0, 1, . . . , N} and transition kernel
p(k, •) = Bin(N, k/N).

p(k, •)
k N
Figure 2.1: Transition function of (Xn ).
Moreover, as the average of this binomial distribution is k, the function h(x) = x is

harmonic, and the expected number of individuals in generation n+ 1 given X0 , . . . , Xn
is
E[Xn+1 | X0 , . . . , Xn ] = Xn .
Hence, the process (Xn ) is a bounded martingale. The Martingale Convergence The-
orem now implies that the limit X∞ = lim Xn exists almost surely, cf. Section 4.2
below. Since Xn takes discrete values, we can conclude that Xn = X∞ eventually with
probability one. In particular, X∞ is almost surely an absorbing state. Hence

P Xn = 0 or Xn = N eventually = 1. (2.1.3)
In order to compute the probabilities of the events “Xn = 0 eventually” and “Xn = N
eventually” we can apply the Optional Stopping Theorem for martingales, cf. Section
2.3 below. Let
T := min{n ≥ 0 : Xn = 0 or Xn = N}, min ∅ := ∞,
denote the first hitting time of the absorbing states. If the initial number X0 of individ-
uals of the given type is k, then by the Optional Stopping Theorem,
E[XT ] = E[X0 ] = k.

Hence by (2.1.3) we obtain

1 k
P Xn = N eventually = P [XT = N] = E[XT ] = , and
N N
k N −k
P Xn = 0 eventually = 1 − = .
N N
Hence eventually all individuals have the same type, and a given type occurs eventually
with probability determined by its initial relative frequency in the population.
2.2 Doob Decomposition and Martingale Problem

We will show now that any adapted sequence of real-valued random variables can be
decomposed into a martingale and a predictable process. In particular, the variance
process of a martingale (Mn ) is the predictable part in the corresponding Doob decom-
position of the process (Mn2 ). The Doob decomposition for functions of Markov chains
implies the martingale problem characterization of Markov chains.
Doob Decomposition
Let (Ω, A, P ) be a probability space and (Fn )n≥0 a filtration on (Ω, A).
Definition (Predictable process). A stochastic process (An )n≥0 is called predictable

w.r.t. (Fn ) if and only if A0 is constant and An is measurable w.r.t. Fn−1 for any n ∈ N.
Intuitively, the value An (ω) of a predictable process can be predicted by the information
available at time n − 1.
Theorem 2.4 (Doob decomposition). Every (Fn ) adapted sequence of integrable ran-
dom variables Yn (n ≥ 0) has a unique decomposition (up to modification on null sets)
Yn = Mn + An (2.2.1)

2.2. DOOB DECOMPOSITION AND MARTINGALE PROBLEM 79
into an (Fn ) martingale (Mn ) and a predictable process (An ) such that A0 = 0. Ex-
plicitly, the decomposition is given by
n
X
An = E[Yk − Yk−1 | Fk−1], and Mn = Yn − An . (2.2.2)
k=1
Remark. (1). The increments E[Yk −Yk−1 |Fk−1] of the process (An ) are the predicted
increments of (Yn ) given the previous information.
(2). The process (Yn ) is a supermartingale (resp. a submartingale) if and only if the
predictable part (An ) is decreasing (resp. increasing).
Proof of Theorem 2.4. Uniqueness: For any decomposition as in (2.2.1) we have
Yk − Yk−1 = Mk − Mk−1 + Ak − Ak−1 for any k ∈ N.
If (Mn ) is a martingale and (An ) is predictable then
E[Yk − Yk−1 | Fk−1] = E[Ak − Ak−1 | Fk−1] = Ak − Ak−1 P -a.s.
This implies that (2.2.2) holds almost surely if A0 = 0.
Existence: Conversely, if (An ) and (Mn ) are defined by (2.2.2) then (An ) is predictable
with A0 = 0 and (Mn ) is a martingale, since
E[Mk − Mk−1 | Fk−1] = 0 P -a.s. for any k ∈ N.
Conditional Variance Process

Consider a martingale (Mn ) such that Mn is square integrable for any n ≥ 0. Then,
by Jensen’s inequality, (Mn2 ) is a submartingale and can again be decomposed into a
fn ) and a predictable process hMin such that hMi0 = 0:
martingale (M
Mn2 = fn + hMin
M for any n ≥ 0.

The increments of the predictable process are given by
hMik − hMik−1 = E[Mk2 − Mk−1

2
| Fk−1]

= E (Mk − Mk−1 )2 Fk−1 + 2 · E Mk−1 · (Mk − Mk−1 ) Fk−1

= Var Mk − Mk−1 Fk−1 for any k ∈ N.
Here we have used in the last step that E[Mk − Mk−1 | Fk−1] vanishes since (Mn ) is a
martingale.
Definition (Conditional variance process). The predictable process

n
X
hMin := Var [Mk − Mk−1 | Fk−1 ] , n ≥ 0,
k=1
is called the conditional variance process of the square integrable martingale (Mn ).
Pn
Example (Random Walks). If Mn = i=1 ηi is a sum of independent centered random
variables ηi and Fn = σ(η1 , . . . , ηn ) then the conditional variance process is given by
P
hMin = ni=1 Var[ηi ].
The conditional variance process is crucial for generalizations of classical limit theo-
rems such as the Law of Large Numbers or the Central Limit Theorem from sums of
independent random variables to martingales. A direct consequence of the fact that
Mn2 − hMin is a martingale is that
E[Mn2 ] = E[M02 ] + E[hMin ] for any n ≥ 0.
This can often be used to derive L2 -estimates for martingales.
Example (Discretizations of stochastic differential equations). Consider an ordinary

differential equation
dXt
= b(Xt ), t ≥ 0, (2.2.3)
dt
where b : Rd → Rd is a given vector field. In order to take into account unpredictable

effects on a system, one is frequently interested in studying random perturbations of the
dynamics (2.2.3) of type
dXt = b(Xt ) dt + “noise” (2.2.4)
with a random noise term. The solution (Xt )t≥0 of such a stochastic differential equa-
tion (SDE) is a stochastic process in continuous time defined on a probability space
(Ω, A, P ) where also the random variables describing the noise effects are defined. The
vector field b is called the (deterministic) “drift”. We will make sense of general SDE
later, but we can already consider time discretizations.
For simplicity let us assume d = 1. Let b, σ : R → R be continuous functions, and let

(ηi )i∈N be a sequence of i.i.d. random variables ηi ∈ L2 (Ω, A, P ) describing the noise
effects. We assume
E[ηi ] = 0 and Var[ηi ] = 1 for any i ∈ N.
Here, the values 0 and 1 are just a convenient normalization, but it is an important
assumption that the random variables are independent with finite variances. Given an
initial value x0 ∈ R and a fine discretization step size h > 0, we now define a stochastic
(h) (h)
process (Xn ) in discrete time by X0 = x0 , and
(h) (h) (h) (h)
√
Xk+1 − Xk = b(Xk ) · h + σ(Xk ) h ηk+1 , for k = 0, 1, 2, . . . (2.2.5)
(h)
One should think of Xk as an approximation for the value of the process (Xt ) at time
t = k · h. The equation (2.2.5) can be rewritten as
n−1
X n−1
X
(h) (h)
√
Xn(h) = x0 + b(Xk ) ·h+ σ(Xk ) · h · ηk+1 . (2.2.6)
k=0 k=0
√
To understand the scaling factors h and h we note first that if σ ≡ 0 then (2.2.5) re-
spectively (2.2.6) is the Euler discretization of the ordinary differential equation (2.2.3).
√
Furthermore, if b ≡ 0 and σ ≡ 1, then the diffusive scaling by a factor h in the second
(h)
term ensures that the continuous time process X⌊t/h⌋ , t ∈ [0, ∞), converges in distri-
bution as h ց 0. Indeed, the functional central limit theorem (Donsker’s invariance

principle) states that the limit process in this case is a Brownian motion (Bt )t∈[0,∞) . In
general, (2.2.6) is an Euler discretization of a stochastic differential equation of type
dXt = b(Xt ) dt + σ(Xt ) dBt
where (Bt )t≥0 is a Brownian motion. Let Fn = σ(η1 , . . . , ηn ) denote the filtration gen-
erated by the random variables ηi . The following exercise summarizes basic properties
of the process X (h) in the case of normally distributed increments.
Exercise. Suppose that the random variables ηi are standard normally distributed.
(1). Prove that the process X (h) is a time-homogeneous (Fn ) Markov chain with tran-
sition kernel
p(x, • ) = N(x + b(x)h, σ(x)2 h)[ • ].
(2). Show that the Doob decomposition X (h) = M (h) + A(h) is given by
n−1
X n−1
X
(h) (h)
√
An(h) = b(Xk ) · h, Mn(h) = x0 + σ(Xk ) h ηk+1 , (2.2.7)
k=0 k=0
and the conditional variance process of the martingale part is

n−1
X
(h) (h)
hM in = σ(Xk )2 · h. (2.2.8)
k=0
(3). Conclude that

n−1
X (h)
E[(Mn(h) − x0 )2 ] = E[σ(Xk )2 ] · h. (2.2.9)
k=0
The last equation can be used in combination with the maximal inequality for mar-
tingales to derive bounds for the processes (X (h) ) in an efficient way, cf. Section 2.4
below.
Remark (Quadratic variation). The quadratic variation of a square integrable martin-

gale (Mn ) is the process [M]n defined by
n
X
[M]n = (Mk − Mk−1 )2 , n ≥ 0.
k=1

It is easy to verify that Mn2 − [M]n is again a martingale. However, [M]n is not pre-
dictable. For continuous martingales in continuous time, the quadratic variation and the
conditional variance process coincide. In discrete time or for discontinuous martingales
they are usually different.
Martingale problem
For a Markov chain (Xn ) we obtain a Doob decomposition
f (Xn ) = Mn[f ] + A[f

n
]
(2.2.10)
for any function f on the state space such that f (Xn ) is integrable for each n. Compu-
tation of the predictable part leads to the following general result:
Theorem 2.5 (Martingale problem for time-homogeneuous Markov chains). Let p

be a stochastic kernel on a measurable space (S, B). Then for an (Fn ) adapted stochas-
tic process (Xn )n≥0 with state space (S, B) the following statements are equivalent:
(1). (Xn ) is a time homogeneous (Fn ) Markov chain with transition kernel p.
(2). (Xn ) is a solution of the martingale problem for the operator L = p − I, i.e.,
there is a decomposition
n−1
X
f (Xn ) = Mn[f ] + (L f )(Xk ), n ≥ 0,
k=0
[f ]
with an (Fn ) martingale (Mn ) for every function f : S → R such that f (Xn ) is
integrable for each n, or, equivalently, for every bounded function f : S → R.
In particular, we see once more that if f (Xn ) is integrable and f is harmonic (L f = 0)

then f (Xn ) is a martingale, and if f is superharmonic (L f ≤ 0), then f (Xn ) is a
supermartingale. The theorem hence extends Theorem 2.3 above.

Proof. The implication “(i)⇒(ii)” is just the Doob decomposition for f (Xn ). In fact,
by Theorem 2.4, the predictable part is given by
n−1
X
A[f
n
]
= E[f (Xk+1) − f (Xk ) | Fk ]
k=0
n−1
X n−1
X
= (pf (Xk ) − f (Xk )) = (L f )(Xk ),
k=0 k=0
[f ] [f ]
and Mn = f (Xn ) − An is a martingale.
[f ]
To prove the converse implication “(ii)⇒(i)” suppose that Mn is a martingale for any
bounded f : S → R. Then
[f ]
0 = E[Mn+1 − Mn[f ] | Fn ]
= E[f (Xn+1 ) − f (Xn ) | Fn ] − ((pf )(Xn ) − f (Xn ))
= E[f (Xn+1 ) | Fn ] − (pf )(Xn )
almost surely for any bounded function f . Hence (Xn ) is an (Fn ) Markov chain with
transition kernel p.
Example (One dimensional Markov chains). Suppose that under Px , the process (Xn )
is a time homogeneous Markov chain with state space S = R or S = Z, initial state
X0 = x, and transition kernel p. Assuming Xn ∈ L2 (Ω, A, P ) for each n, we define the
“drift” and the “fluctuations” of the process by
b(x) := Ex [X1 − X0 ]
a(x) = Varx [X1 − X0 ].
We now compute the Doob decomposition of (Xn ). Choosing f (x) = x we have

ˆ
(p − I)f (x) = y p(x, dy) − x = Ex [X1 − X0 ] = b(x).
Hence by Theorem 2.5,

n−1
X
Xn = Mn + b(Xk ) (2.2.11)
k=0

2.3. GAMBLING STRATEGIES AND STOPPING TIMES 85
with an (Fn ) martingale (Mn ). To obtain detailed information on Mn , we compute the

variance process: By (2.2.11) and the Markov property, we obtain
n−1
X n−1
X n−1
X
hMin = Var[Mk+1 − Mk | Fk ] = Var[Xk+1 − Xk | Fk ] = a(Xk ).
k=0 k=0 k=0
Therefore
n−1
X
Mn2 = fn +
M a(Xk ) (2.2.12)
k=0
fn ). The functions a(x) and b(x) can now be used in

with another (Fn ) martingale (M
connection with fundamental results for martingales as e.g. the maximal inequality (cf.
2.4 below) to derive bounds for Markov chains in an efficient way.
2.3 Gambling strategies and stopping times

Throughout this section, we fix a filtration (Fn )n≥0 on a probability space (Ω, A, P ).
Martingale transforms
Suppose that (Mn )n≥0 is a martingale w.r.t. (Fn ), and (Cn )n∈N is a predictable sequence
of real-valued random variables. For example, we may think of Cn as the stake in the
n-th round of a fair game, and of the martingale increment Mn − Mn−1 as the net gain
(resp. loss) per unit stake. In this case, the capital In of a player with gambling strategy
(Cn ) after n rounds is given recursively by
In = In−1 + Cn · (Mn − Mn−1 ) for any n ∈ N,
i.e.,
n
X
In = I0 + Ck · (Mk − Mk−1 ).
k=1

Definition (Martingale transform). The stochastic process C• M defined by

n
X
(C• M)n := Ck · (Mk − Mk−1 ) for any n ≥ 0,
k=1
is called the martingale transform of the martingale (Mn )n≥0 w.r.t. the predictable
sequence (Cn )n≥1 , or the discrete stochastic integral of C w.r.t. M.
We will see later that the process C• M is a time-discrete version of the stochastic inte-
gral Cs dMs of a predictable continuous-time process C w.r.t. a continuous-time mar-
´
´n
tingale M. To be precise, (C• M)n coincides with the Itô integral 0 C⌈t⌉ dM⌊t⌋ of the
left continuous jump process t 7→ C⌈t⌉ w.r.t. the right continuous martingale t 7→ M⌊t⌋ .
Example (Martingale strategy). One origin of the word “martingale” is the name of
a well-known gambling strategy: In a standard coin-tossing game, the stake is doubled
each time a loss occurs, and the player stops the game after the first time he wins. If the
net gain in n rounds with unit stake is given by a standard Random Walk
Mn = η1 + . . . + ηn , ηi i.i.d. with P [ηi = 1] = P [ηi = −1] = 1/2,
then the stake in the n-th round is
Cn = 2n−1 if η1 = . . . = ηn−1 = −1, and Cn = 0 otherwise.
Clearly, with probability one, the game terminates in finite time, and at that time the
player has always won one unit, i.e.,
P [(C• M)n = 1 eventually] = 1.

(C• M)n
2
n
−1
−2
−3
−4
−5
−6
−7
At first glance this looks like a safe winning strategy, but of course this would only be
the case, if the player had unlimited capital and time available.
Theorem 2.6 (You can’t beat the system!). (1). If (Mn )n≥0 is an (Fn ) martingale,
and (Cn )n≥1 is predictable with Cn · (Mn − Mn−1 ) ∈ L1 (Ω, A, P ) for any n ≥ 1,
then C• M is again an (Fn ) martingale.
(2). If (Mn ) is an (Fn ) supermartingale and (Cn )n≥1 is non-negative and predictable
with Cn · (Mn − Mn−1 ) ∈ L1 for any n, then C• M is again a supermartingale.
Proof. For n ≥ 1 we have
E[(C• M)n − (C• M)n−1 | Fn−1 ] = E[Cn · (Mn − Mn−1 ) | Fn−1 ]

= Cn · E[Mn − Mn−1 | Fn−1 ] = 0 P -a.s.

This proves the first part of the claim. The proof of the second part is similar.
The theorem shows that a fair game (a martingale) can not be transformed by choice of
a clever gambling strategy into an unfair (or “superfair”) game. In models of financial
markets this fact is crucial to exclude the existence of arbitrage possibilities (riskless
profit).
Example (Martingale strategy, cont.). For the classical martingale strategy, we obtain
E[(C• M)n ] = E[(C• M)0 ] = 0 for any n ≥ 0
by the martingale property, although
lim (C• M)n = 1 P -almost surely.

n→∞
This is a classical example showing that the assertion of the dominated convergence
theorem may not hold if the assumptions are violated.
Remark. The integrability assumption in Theorem 2.6 is always satisfied if the random
variables Cn are bounded, or if both Cn and Mn are square-integrable for any n.
Example (Financial market model with one risky asset). Suppose that during each
time interval (n − 1, n), an investor is holding Φn units of an asset with price Sn per
unit at time n. We assume that (Sn ) is an adapted and (Φn ) is a predictable stochastic
process w.r.t. a filtration (Fn ). If the investor always puts his remaining capital onto
a bank account with guaranteed interest rate r (“riskless asset”) then the change of his
capital Vn during the time interval (n − 1, n) is given by
Vn = Vn−1 + Φn · (Sn − Sn−1 ) + (Vn−1 − Φn · Sn−1 ) · r. (2.3.1)
Considering the discounted quantity Ven = Vn /(1 + r)n , we obtain the equivalent
recursion
Ven = Ven−1 + Φn · (Sen − Sen−1 ) for any n ≥ 1. (2.3.2)
In fact, (2.3.1) holds if and only if
Vn − (1 + r)Vn−1 = Φn · (Sn − (1 + r)Sn−1 ),

which is equivalent to (2.3.2). Therefore, the discounted capital at time n is given by
Ven = V0 + (Φ• S)
e n.
By Theorem 2.6, we can conclude that, if the discounted price process (Sen ) is an (Fn )
martingale w.r.t. a given probability measure, then (Ven ) is a martingale as well. In this
case, assuming that V0 is constant, we obtain in particular
E[Ven ] = V0 ,
or, equivalently,
E[Vn ] = (1 + r)n V0 for any n ≥ 0. (2.3.3)
This fact, together with the existence of a martingale measure, can now be used for
option pricing under a no-arbitrage assumption. To this end we assume that the payoff
of an option at time N is given by an (FN )-measurable random variable F . For example,
the payoff of a European call option with strike price K based on the asset with price
process (Sn ) is SN − K if the price Sn at maturity exceeds K, and 0 otherwise, i.e.,
F = (SN − K)+ .
Suppose further that the option can be replicated by a hedging strategy (Φn ), i.e., there
exists an F0 -measurable random variable V0 and a predictable sequence of random vari-
ables (Φn )1≤n≤N such that
F = VN
is the value at time N of a portfolio with initial value V0 w.r.t. the trading strategy (Φn ).
Then, assuming the non-existence of arbitrage possibilities, the option price at time
0 has to be V0 , since otherwise one could construct an arbitrage strategy by selling
the option and investing money in the stock market with strategy (Φn ), or conversely.
Therefore, if a martingale measure exists (i.e., an underlying probability measure such
that the discounted stock price (Sen ) is a martingale), then the no-arbitrage price of the
option at time 0 can be computed by (2.3.3) where the expectation is taken w.r.t. the
martingale measure.
The following exercise shows how this works out in the Cox-Ross-Rubinstein binomial
model:

Exercise (No-Arbitrage Pricing in the CRR model). Consider the CRR binomial
model, i.e., Ω = {1 + a, 1 + b}N with −1 < a < r < b < ∞, Xi (ω1 , . . . , ωN ) = ωi ,
Fn = σ(X1 , . . . , Xn ), and
n
Y
Sn = S0 · Xi , n = 0, 1, . . . , N,
i=1
where S0 is a constant.
(1). Completeness of the CRR model: Prove that for any function F : Ω → R there
exists a constant V0 and a predictable sequence (Φn )1≤n≤N such that F = VN
where (Vn )1≤n≤N is defined by (2.3.1), or, equivalently,
F
= VeN = V0 + (Φ• S)
e N.
(1 + r)N
Hence in the CRR model, any FN -measurable function F can be replicated by
a predictable trading strategy. Market models with this property are called com-
plete.
Hint: Prove inductively that for n = N, N − 1, . . . , 0, Fe = F/(1 + r)N can be

represented as
N
X
Fe = Ven + Φi · (Sei − Sei−1 )
i=n+1
with an Fn -measurable function Ven and a predictable sequence (Φi )n+1≤i≤N .
(2). Option pricing: Derive a general formula for the no-arbitrage price of an option
with payoff function F : Ω → R in the CRR model. Compute the no-arbitrage
price for a European call option with maturity N and strike K explicitly.
Stopped Martingales
One possible strategy for controlling a fair game is to terminate the game at a time
depending on the previous development. Recall that a random variable T : Ω →
{0, 1, 2, . . .} ∪ {∞} is called a stopping time w.r.t. the filtration (Fn ) if and only if
the event {T = n} is contained in Fn for any n ≥ 0, or equivalently, iff {T ≤ n} ∈ Fn
for any n ≥ 0.

Example (Hitting times). (1). The first hitting time
TB = min{n ≥ 0 : Xn ∈ B} (where min ∅ := ∞)
and the first passage or return time
SB = min{n ≥ 1 : Xn ∈ B}
to a measurable subset B of the state space by an (Fn ) adapted stochastic process

are (Fn ) stopping times. For example, for n ≥ 0,
{TB = n} = {X1 ∈ B C , . . . , Xn−1 ∈ B C , Xn ∈ B} ∈ Fn .
If one decides to sell an asset as soon as the price Sn exceeds a given level λ > 0
then the selling time equals T(λ,∞) and is hence a stopping time.
(2). On the other hand, the last visit time
LB := sup{n ≥ 0 : Xn ∈ B} (where sup ∅ := 0)
is not a stopping time in general. Intuitively, to decide whether LB = n, informa-

tion on the future development of the process is required.
We consider an (Fn )-adapted stochastic process (Mn )n≥0 , and an (Fn )-stopping time
T on the probability space (Ω, A, P ). The process stopped at time T is defined as
(MT ∧n )n≥0 where

Mn (ω) for n ≤ T (ω),
MT ∧n (ω) = MT (ω)∧n (ω) =
M for n ≥ T (ω).
T (ω) (ω)
For example, the process stopped at a hitting time TB gets stuck at the first time it enters
the set B.
Theorem 2.7 (Optional Stopping Theorem,Version 1). If (Mn )n≥0 is a martingale

(resp. a supermartingale) w.r.t. (Fn ), and T is an (Fn )-stopping time, then the stopped
process (MT ∧n )n≥0 is again an (Fn )-martingale (resp. supermartingale). In particular,
we have
(≤)
E[MT ∧n ] = E[M0 ] for any n ≥ 0.

Proof. Consider the following strategy:
Cn = I{T ≥n} = 1 − I{T ≤n−1} ,
i.e., we put a unit stake in each round before time T and quit playing at time T . Since
T is a stopping time, the sequence (Cn ) is predictable. Moreover,
MT ∧n − M0 = (C• M)n for any n ≥ 0. (2.3.4)
In fact, for the increments of the stopped process we have

( )
Mn − Mn−1 if T ≥ n
MT ∧n − MT ∧(n−1) = = Cn · (Mn − Mn−1 ),
0 if T ≤ n − 1
and (2.3.4) follows by summing over n. Since the sequence (Cn ) is predictable, bounded
and non-negative, the process C• M is a martingale, supermartingale respectively, pro-
vided the same holds for M.
Remark (IMPORTANT). (1). In general, it is NOT TRUE under the assumptions in

Theorem 2.7 that
E[MT ] = E[M0 ], E[MT ] ≤ E[M0 ] respectively. (2.3.5)
Suppose for example that (Mn ) is the classical Random Walk starting at 0 and
T = T{1} is the first hitting time of the point 1. Then, by recurrence of the
Random Walk, T < ∞ and MT = 1 hold almost surely although M0 = 0.
(2). If, on the other hand, T is a bounded stopping time, then there exists n ∈ N such
that T (ω) ≤ n for any ω. In this case, the optional stopping theorem implies
(≤)
E[MT ] = E[MT ∧n ] = E[M0 ].
More general sufficient conditions for (2.3.5) are given in Theorems 2.8, 2.9 and 2.10
below.
Example (Classical ruin problem). Let a, b, x ∈ Z with a < x < b. We consider the
classical Random Walk
n
X 1
Xn = x + ηi , ηi i.i.d. with P [ηi = ±1] = ,
i=1
2

with initial value X0 = x. We now show how to apply the Optional Stopping Theorem
to compute the distributions of the exit time
T (ω) = min{n ≥ 0 : Xn (ω) 6∈ (a, b)},
and the exit point XT . These distributions can also be computed by more traditional
methods (first step analysis, reflection principle), but martingales yield an elegant and
general approach.
(1). Ruin probability r(x) = P [XT = a].

The process (Xn ) is a martingale w.r.t. the filtration Fn = σ(η1 , . . . , ηn ), and
T < ∞ almost surely holds by elementary arguments. As the stopped process
XT ∧n is bounded (a ≤ XT ∧n ≤ b), we obtain
n→∞
x = E[X0 ] = E[XT ∧n ] → E[XT ] = a · r(x) + b · (1 − r(x))
by the Optional Stopping Theorem and the Dominated Convergence Theorem.

Hence
b−x
r(x) = . (2.3.6)
a−x
(2). Mean exit time from (a, b).

To compute the expectation E[T ], we apply the Optional Stopping Theorem to
the (Fn ) martingale
Mn := Xn2 − n.
By monotone and dominated convergence, we obtain
x2 = E[M0 ] = E[MT ∧n ] = E[XT2 ∧n ] − E[T ∧ n]

n→∞
−→ E[XT2 ] − E[T ].
Therefore, by (2.3.6),
E[T ] = E[XT2 ] − x2 = a2 · r(x) + b2 · (1 − r(x)) − x2

= (b − x) · (x − a). (2.3.7)

(3). Mean passage time of b.

The first passage time Tb = min{n ≥ 0 : Xn = b} is greater or equal than the
exit time from the interval (a, b) for any a < x. Thus by (2.3.7), we have
E[Tb ] ≥ lim (b − x) · (x − a) = ∞,
a→−∞
i.e., Tb is not integrable! These and some other related passage times are im-
portant examples of random variables with a heavy-tailed distribution and infinite
first moment.
(4). Distribution of passage times.

We now compute the distribution of the first passage time Tb explicitly in the case
x = 0 and b = 1. Hence let T = T1 . As shown above, the process
Mnλ := eλXn /(cosh λ)n , n ≥ 0,
is a martingale for each λ ∈ R. Now suppose λ > 0. By the Optional Stopping

Theorem,
1 = E[M0λ ] = E[MTλ∧ n ] = E[eλXT ∧n /(cosh λ)T ∧n ] (2.3.8)
for any n ∈ N. As n → ∞, the integrands on the right hand side converge

to eλ (cosh λ)−T · I{T <∞} . Moreover, they are uniformly bounded by eλ , since
XT ∧n ≤ 1 for any n. Hence by the Dominated Convergence Theorem, the expec-
tation on the right hand side of (2.3.8) converges to E[eλ /(cosh λ)T ; T < ∞],
and we obtain the identity
E[(cosh λ)−T ; T < ∞] = e−λ for any λ > 0. (2.3.9)
Taking the limit as λ ց 0, we see that P [T < ∞] = 1. Taking this into account,
and substituting s = 1/ cosh λ in (2.3.9), we can now compute the generating
function of T explicitly:
√
E[sT ] = e−λ = (1 − 1 − s2 )/s for any s ∈ (0, 1). (2.3.10)
Developing both sides into a power series finally yields

∞ ∞
!
X X 1/2
sn · P [T = n] = (−1)m+1 s2m−1 .
n=0 m=1 m

Therefore, the distribution of the first passage time of 1 is given by

m+1 1/2 m+1 1 1 1
P [T = 2m−1] = (−1) = (−1) · · − ··· − m + 1 /m!
m 2 2 2
and P [T = 2m] = 0 for any mN.
Optional Stopping Theorems

Stopping times occurring in applications are typically not bounded. Therefore, we need
more general conditions guaranteeing that (2.3.5) holds nevertheless. A first general
criterion is obtained by applying the Dominated Convergence Theorem:
Theorem 2.8 (Optional Stopping Theorem, Version 2). Suppose that (Mn ) is a mar-
tingale w.r.t. (Fn ), T is an (Fn )-stopping time with P [T < ∞] = 1, and there exists a
random variable Y ∈ L1 (Ω, A, P ) such that
|MT ∧n | ≤ Y P -almost surely for any n ∈ N.
Then
E[MT ] = E[M0 ].
Proof. Since P [T < ∞] = 1, we have
MT = lim MT ∧n P -almost surely.

n→∞
By Theorem 2.7, E[M0 ] = E[MT ∧n ], and by the Dominated Convergence Theorem,

E[MT ∧n ] −→ E[MT ] as n → ∞.
Remark (Weakening the assumptions). Instead of the existence of an integrable ran-

dom variable Y dominating the random variables MT ∧n , n ∈ N, it is enough to assume
that these random variables are uniformly integrable, i.e.,

sup E |MT ∧n | ; |MT ∧n | ≥ c → 0 as c → ∞.
n∈N
A corresponding generalization of the Dominated Convergence Theorem is proven in

Section 4.3 below.

For non-negative supermartingales, we can apply Fatou’s Lemma instead of the Domi-
nated Convergence Theorem to pass to the limit as n → ∞ in the Stopping Theorem.
The advantage is that no integrability assumption is required. Of course, the price to
pay is that we only obtain an inequality:
Theorem 2.9 (Optional Stopping Theorem, Version 3). If (Mn ) is a non-negative

supermartingale w.r.t. (Fn ), then
E[M0 ] ≥ E[MT ; T < ∞]
holds for any (Fn ) stopping time T .
Proof. Since MT = lim MT ∧n on {T < ∞}, and MT ≥ 0, Theorem 2.7 combined

n→∞
with Fatou’s Lemma implies
h i
E[M0 ] ≥ lim inf E[MT ∧n ] ≥ E lim inf MT ∧n ≥ E[MT ; T < ∞].
n→∞ n→∞
Example (Dirichlet problem for Markov chains). Suppose that w.r.t. the probability
measure Px , the process (Xn ) is a time-homogeneous Markov chain with measurable
state space (S, B), transition kernel p, and start in x. Let D ∈ B be a measurable
subset of the state space, and f : D C → R a measurable function (the given “boundary
values”), and let
T = min{n ≥ 0 : Xn ∈ D C }
denote the first exit time of the Markov chain from D. By conditioning on the first
step of the Markov chain, one can show that if f is non-negative or bounded, then the
function
h(x) = Ex [f (XT ) ; T < ∞], (x ∈ S),
is a solution of the Dirichlet problem
(ph)(x) = h(x) for x ∈ D,

h(x) = f (x) for x ∈ D C ,

see [XXXStochastic Processes].
DC
By considering the martingale h(XT ∧n ) for a function h that is harmonic on D, we

obtain a converse statement:
Exercise (Uniqueness of the Dirichlet problem). Suppose that Px [T < ∞] = 1 for

any x ∈ S.
(1). Prove that h(XT ∧n ) is a martingale w.r.t. Px for any bounded solution h of the
Dirichlet problem and any x ∈ S.
(2). Conclude that if f is bounded, then
h(x) = Ex [f (XT )] (2.3.11)
is the unique bounded solution of the Dirichlet problem.
(3). Similarly, show that for any non-negative f , the function h defined by (2.3.11) is
the minimal non-negative solution of the Dirichlet problem.
We finally state a version of the Optional Stopping Theorem that applies in particular to
martingales with bounded increments:

Corollary 2.10 (Optional Stopping for martingales with bounded increments). Sup-
pose that (Mn ) is an (Fn ) martingale, and there exists a finite constant K ∈ (0, ∞) such
that
E[|Mn+1 − Mn | | Fn ] ≤ K P -almost surely for any n ≥ 0. (2.3.12)
Then for any (Fn ) stopping time T with E[T ] < ∞, we have
E[MT ] = E[M0 ].
Proof. For any n ≥ 0,

∞
X
|MT ∧n | ≤ |M0 | + |Mi+1 − Mi | · I{T >i} .
i=0
Let Y denote the expression on the right hand side. We will show that Y is an integrable
random variable – this implies the assertion by Theorem 2.8. To verify integrability of
Y note that the event {T > i} is contained in Fi for any i ≥ 0 since T is a stopping
time. Therefore and by (2.3.12),
E[|Mi+1 − Mi | ; T > i] = E[E[|Mi+1 − Mi | | Fi ] ; T > i] ≤ k · P [T > i].
Summing over i, we obtain

∞
X
E[Y ] ≤ E[|M0 |] + k · P [T > i] = E[|M0 |] + k · E[T ] < ∞
i=0
by the assumptions.
Exercise (Integrability of stopping times). Prove that the expectation E[T ] of a stop-
ping time T is finite if there exist constants ε > 0 and k ∈ N such that
P [T ≤ n + k | Fn ] ≥ ε P -a.s. for any n ∈ N.

2.4. MAXIMAL INEQUALITIES 99
Wald’s identity for random sums

We finally apply the Optional Stopping Theorem to sums of independent random vari-
ables with a random number T of summands. The point is that we do not assume that
T is independent of the summands but only that it is a stopping time w.r.t. the filtration
generated by the summands.
Let Sn = η1 + . . . + ηn with i.i.d. random variables ηi ∈ L1 (Ω, A, P ). Denoting by m
the expectations of the increments ηi , the process
Mn = Sn − n · m
is a martingale w.r.t. Fn = σ(η1 , . . . , ηn ). By applying Corollary 2.10 to this martingale,

we obtain:
Theorem 2.11 (Wald’s identity). Suppose that T is an (Fn ) stopping time with E[T ] <
∞. Then
E[ST ] = m · E[T ].
Proof. For any n ≥ 0, we have
E[|Mn+1 − Mn | | Fn ] = E[|ηn+1 − m| |Fn ] = E[|ηn+1 − m|]
by the independence of the ηi . As the ηi are identically distributed and integrable, the
right hand side is a finite constant. Hence Corollary 2.10 applies, and we obtain
0 = E[M0 ] = E[MT ] = E[ST ] − m · E[T ].
2.4 Maximal inequalities

For a standard Random Walk Sn = η1 + . . . + ηn , ηi i.i.d. with P [ηi = ±1] = 1/2, the
reflection principle implies the identity
P [max(S0 , S1 , . . . , Sn ) ≥ c] = P [Sn ≥ c] + P [Sn < c; max(S0 , S1 , . . . , Sn ) ≥ c]

= P [|Sn | > c] + P [Sn > c]

for any c ∈ N. In combination with the Markov-Čebyšev inequality this can be used to
control the running maximum of the Random Walk in terms of the moments of the last
value Sn .
Maximal inequalities are corresponding estimates for max(M0 , M1 , . . . , Mn ) or sup Mk
k≥0
when (Mn ) is a sub- or supermartingale respectively. These estimates are an important
tool in stochastic analysis. They are a consequence of the Optional Stopping Theorem.
Doob’s inequality
We first prove the basic version of maximal inequalities for sub- and supermartingales:
Theorem 2.12 (Doob).
(1). Suppose that (Mn )n≥0 is a non-negative supermartingale. Then

1
P sup Mk ≥ c ≤ · E[M0 ] for any c > 0.
k≥0 c
(2). Suppose that (Mn )n≥0 is a non-negative submartingale. Then

1 1
P max Mk ≥ c ≤ ·E Mn ; max Mk ≥ c ≤ ·E[Mn ] for any c > 0.
0≤k≤n c 0≤k≤n c
Proof. (1). For c > 0 we consider the stopping time
Tc = min{k ≥ 0 : Mk ≥ c}, min ∅ = ∞.
Note that Tc < ∞ whenever sup Mk > c. Hence by the version of the Optional
Stopping Theorem for non-negative supermartingales, we obtain
1 1
P [sup Mk > c] ≤ P [Tc < ∞] ≤ E[MTc ; Tc < ∞] ≤ E[M0 ].
c c
Here we have used in the second and third step that (Mn ) is non-negative. Re-
placing c by c − ε and letting ε tend to zero we can conclude
1 1
P [sup Mk ≥ c] = lim P [sup Mk > c − ε] ≤ lim inf E[M0 ] = · E[M0 ].
εց0 εց0 c − ε c

(2). For a non-negative supermartingale, we obtain

1
P max Mk ≥ c = P [Tc ≤ n] ≤ E[MTc ; Tc ≤ n]
0≤k≤n c
n n
1X 1X
= E[Mk ; Tc = k] ≤ E[Mn ; Tc = k]
c k=0 c k=0
1
= · E[Mn ; Tc ≤ n].
c
Here we have used in the second last step that E[Mk ; Tc = k] ≤ E[Mn ; Tc = k]
since (Mn ) is a supermartingale and {Tc = k} is in Fk .
First consequences of Doob’s maximal inequality for submartingales are extensions of

the classical Markov- Čebyšev inequalities:
Corollary 2.13. (1). Suppose that (Mn )n≥0 is an arbitrary submartingale (not neces-
sarily non-negative!). Then

1 +
P max Mk ≥ c ≤ E Mn ; max Mk ≥ c for any c > 0, and
k≤n c k≤n

−λc λMn
P max Mk ≥ c ≤ e E e ; max Mk ≥ c for any λ, c > 0.
k≤n k≤n
(2). If (Mn ) is a martingale then, moreover, the estimates

1 p
P max |Mk | ≥ c ≤ p E |Mn | ; max |Mk | ≥ c
k≤n c k≤n
hold for any c > 0 and p ∈ [1, ∞).
Proof. The corollary follows by applying the maximal inequality to the non-negative
submartingales Mn+ , exp(λMn ), |Mn |p respectively. These processes are indeed sub-
martingales, as the functions x 7→ x+ and x 7→ exp(λx) are convex and non-decreasing
for any λ > 0, and the functions x 7→ |x|p are convex for any p ≥ 1.

Lp inequalities
The last estimate in Corollary 2.13 can be used to bound the Lp norm of the running
maximum of a martingale in terms of the Lp -norm of the last value. The resulting bound,
known as Doob’s Lp -inequality, is crucial for stochastic analysis. We first remark:
´y
Lemma 2.14. If Y : Ω → R+ is a non-negative random variable, and G(y) = g(x)dx
0
is the integral of a non-negative function g : R+ → R+ , then
ˆ∞
E[G(Y )] = g(c) · P [Y ≥ c] dc.
0
Proof. By Fubini’s theorem we have

 Y  ∞ 
ˆ ˆ
E[G(Y )] = E  g(c) dc = E  I[0,Y ] (c)g(c) dc
0 0
ˆ∞
= g(c) · P [Y ≥ c] dc.
0
Theorem 2.15 (Doob’s Lp inequality). Suppose that (Mn )n≥0 is a martingale, and let
Mn∗ := max |Mk |, and M ∗ := sup |Mk |.

k≤n k
1 1
Then, for any p, q ∈ (1, ∞) such that p
+ q
= 1, we have
kMn∗ kLp ≤ q · kMn kLp , and kM ∗ kLp ≤ q · sup kMn kLp .

n
In particular, if (Mn ) is bounded in Lp then M ∗ is contained in Lp .

Proof. By Lemma 2.14, Corollary 2.13 applied to the martingales Mn and (−Mn ), and
Fubini’s theorem, we have
ˆ∞
2.14
E[(Mn∗ )p ] = pcp−1 · P [Mn∗ ≥ c] dc
0
ˆ∞
2.13
≤ pcp−2E[|Mn | ; Mn∗ ≥ c] dc
0
 ∗ 
ˆMn
Fub.
= E |Mn | · pcp−2 dp
0
p
= E[|Mn | · (Mn∗ )p−1]
p−1
p
for any n ≥ 0 and p ∈ (1, ∞). Setting q = p−1
and applying Hölder’s inequality to the
right hand side, we obtain
E[(Mn∗ )p ] ≤ q · kMn kLp · k(Mn∗ )p−1 kLq = q · kMn kLp · E[(Mn∗ )p ]1/q ,
i.e.,
kMn∗ kLp = E[(Mn∗ )p ]1−1/q ≤ q · kMn kLp . (2.4.1)
This proves the first inequality. The second inequality follows as n → ∞, since

kM ∗ kLp = lim Mn∗ = lim inf kMn∗ kLp ≤ q · sup kMn kLp
n→∞ Lp n→∞ n∈N
by Fatou’s Lemma.
Hoeffding’s inequality
For a standard Random Walk (Sn ) starting at 0, the reflection principle combined with
Bernstein’s inequality implies the upper bound
P [max(S0 , . . . , Sn ) ≥ c] ≤ 2 · P [Sn ≥ c] ≤ 2 · exp(−2c2 /n)
for any n ∈ N and c ∈ (0, ∞). A similar inequality holds for arbitrary martingales with
bounded increments:

Theorem 2.16 (Azuma, Hoeffding). Suppose that (Mn ) is a martingale such that
|Mn − Mn−1 | ≤ an P -almost surely
for a sequence (an ) of non-negative constants. Then

, n !
1 X
P max(Mk − M0 ) ≥ c ≤ exp − c2 a2i (2.4.2)
k≤n 2 i=1
for any n ∈ N and c ∈ (0, ∞).
Proof. W.l.o.g. we may assume M0 = 0. Let Yn = Mn − Mn−1 denote the martingale

increments. We will apply the exponential form of the maximal inequality. For λ > 0
and n ∈ N, we have,
" n #
Y
E[eλMn ] = E eλYi = E eλMn−1 · E eλYn | Fn−1 . (2.4.3)
i=1
To bound the conditional expectation, note that
1 an − Yn −λan 1 an + Yn λan
eλYn ≤ e + e
2 an 2 an
holds almost surely, since x 7→ exp(λx) is a convex function, and −an ≤ Yn ≤

an . Indeed, the right hand side is the value at Yn of the secant connecting the points
(−an , exp(−λan )) and (an , exp(λan )). Since (Mn ) is a martingale, we have
E[Yn |Fn−1] = 0,
and therefore
2
E[eλYn | Fn−1 ] ≤ e−λan + eλan /2 = cosh(λan ) ≤ e(λan ) /2
almost surely. Now, by (2.4.3), we obtain

2 /2
E[eλYn ] ≤ E[eλMn−1 ] · e(λan ) .

Hence, by induction on n,
n
!
1 2X 2
E[eλMn ] ≤ exp λ a for any n ∈ N, (2.4.4)
2 i=1 i
and, by the exponential maximal inequality (cf. Corollary 2.13),

n
!
1 X 2
P [max Mk ≥ c] ≤ exp −λc + λ2 a (2.4.5)
k≤n 2 i=1 i
holds for any n ∈ N and c, λ > 0.

For a given c and n, the expression on the right hand side of (2.4.5) is minimal for λ =
P
c/ ni=1 a2i . Choosing λ correspondingly, we finally obtain the upper bound (2.4.2).
Hoeffding’s concentration inequality has numerous applications, for example in the

analysis of algorithms, cf. [Mitzenmacher, Upful: Probability and Computing]. Here,
we just consider one simple example to illustrate the way it typically is applied:
Example (Pattern Matching). Suppose that X1 , X2 , . . . , Xn is a sequence of indepen-

dent, uniformly distributed random variables (“letters”) taking value sin a finite set S
(the underlying “alphabet”), and let
n−l
X
N = I{Xi+1 =a1 ,Xi+2 =ax ,...,Xi+l =al } (2.4.6)
i=0
denote the number of occurrences of a given “word” a1 a2 · · · al with l letters in the

random text. In applications, the “word” could for example be a DNA sequence. We
easily obtain
n−l
X
E[N] = P [Xi+k = ak for k = 1, . . . , l] = (n − l + 1)/|S|l. (2.4.7)
i=0
To estimate the fluctuations of the random variable N around its mean value, we con-
sider the martingale
Mi = E[N | σ(X1 , . . . , Xi )], (i = 0, 1, . . . , n)

with initial value M0 = E[N] and terminal value Mn = N. Since at most l of the
summands in (2.4.6) are not independent of i, and each summand takes values 0 and 1
only, we have
|Mi − Mi−1 | ≤ l for each i = 0, 1, . . . , n.
Therefore, by Hoeffding’s inequality, applied in both directions, we obtain
P [|N − E[N]| ≥ c] = P [|Mn − M0 | ≥ c] ≤ 2 exp(−c2 /(2nl2 ))
for any c > 0, or equivalently,

√
P [|N − E[N]| ≥ ε · l n] ≤ 2 · exp(−ε2 /2) for any ε > 0. (2.4.8)
The equation (2.4.7) and the bound (2.4.8) show that N is highly concentrated around
√
its mean if l is small compared to n.

Chapter 3
Martingales in continuous time
The notion of a martingale, sub- and supermartingale in continuous time can be defined
similarly as in the discrete parameter case. Fundamental results such as the optional
stopping theorem or the maximal inequality carry over from discrete parameter to con-
tinuous time martingales under additional regularity conditions as, for example, conti-
nuity of the sample paths. Similarly as for Markov chains in discrete time, martingale
methods can be applied to derive explicit expressions and bounds for probabilities and
expectations of Brownian motion in a clear and efficient way.
We start with the definition of martingales in continuous time. Let (Ω, A, P ) denote a
probability space.
Definition. (1). A continuous-time filtration on (Ω, A) is a family (Ft )t∈[0,∞) of σ-

algebras Ft ⊆ A such that Fs ⊆ Ft for any 0 ≤ s ≤ t.
(2). A real-valued stochastic process (Mt )t∈[0,∞) on (Ω, A, P ) is called a martingale

(or super-, submartingale) w.r.t. a filtration (Ft ) if and only if
(a) (Mt ) is adapted w.r.t. (Ft ), i.e., Mt is Ft measurable for any t ≥ 0.

(b) For any t ≥ 0, the random variable Mt (resp. Mt+ , Mt− ) is integrable.
(≤,≥)
(c) E[Mt | Fs ] = Ms P -almost surely for any 0 ≤ s ≤ t.
107
108 CHAPTER 3. MARTINGALES IN CONTINUOUS TIME
3.1 Some fundamental martingales of Brownian Motion

In this section, we identify some important martingales that are functions of Brownian
motion. Let (Bt )t≥0 denote a d-dimensional Brownian motion defined on (Ω, A, P ).
Filtrations generated by Brownian motion

Any stochastic process (Xt )t≥0 in continuous time generates a filtration
FtX = σ(Xs : 0 ≤ s ≤ t), t ≥ 0.
However, not every hitting time that we are interested in is a stopping time w.r.t. this
filtration. For example, for one-dimensional Brownian motion (Bt ), the first hitting
time T = inf{t ≥ 0 : Bt > c} of the open interval (c, ∞) is not an (FtB ) stopping
time. An intuitive explanation for this fact is that for t ≥ 0, the event {T ≤ t} is not
contained in FtB , since for a path with Bs ≤ c on [0, t] and Bt = c, we can not decide
at time t, if the path will enter the interval (c, ∞) in the next instant. For this and other
reasons, we also consider the right-continuous filtration
\
B
Ft := Ft+ε , t ≥ 0,
ε>0
that takes into account “infinitesimal information on the future development.”
Exercise (Hitting times as stopping times). Prove that the first hitting time TA =
inf{t ≥ 0 : Bt ∈ A} of a set A ⊆ Rd is an (FtB ) stopping time if A is closed, whereas
TA is an (Ft ) stopping time, but not necessarily an (FtB ) stopping time if A is open.
It is easy to verify that a d-dimensional Brownian motion (Bt ) is also a Brownian motion
w.r.t. the right-continuous filtration (Ft ):
Lemma 3.1. For any 0 ≤ s < t, the increment Bt − Bs is independent of Fs with

distribution N(0, (t − s) · Id ).
Proof. Since t 7→ Bt is almost surely continuous, we have
Bt − Bs = lim (Bt − Bs+ε ) P -a.s. (3.1.1)

εց0
ε∈Q

3.1. SOME FUNDAMENTAL MARTINGALES OF BROWNIAN MOTION 109
For small ε > 0 the increment Bt −Bs+ε is independent of Fs+ε

B
, and hence independent
of Fs . Therefore, by (3.1.1), Bt − Bs is independent of Fs as well.
Another filtration of interest is the completed filtration (FtP ). A σ-algebra F is called

complete w.r.t. a probability measure P iff it contains all subsets of P -measure zero
sets. The completion of a σ-algebra A w.r.t. a probability measure P on (Ω, A) is the
complete σ-algebra
AP = {A ⊆ Ω : ∃A1 , A2 ∈ A : A1 ⊆ A ⊆ A2 , P [A2 \ A1 ] = 0}
generated by all sets in A and all subsets of P -measure zero sets in A.
It can be shown that the completion (FtP ) of the right-continuous filtration (Ft ) is again
right-continuous. The assertion of Lemma 3.1 obviously carries over to the completed
filtration.
Remark (The “usual conditions”). Some textbooks on stochastic analysis consider

only complete right-continuous filtrations. A filtration with these properties is said to
satisfy the usual conditions. A disadvantage of completing the filtration, however, is
that (FtP ) depends on the underlying probability measure P (or, more precisely, on its
null sets). This can cause problems when considering several non-equivalent probability
measures at the same time.
Brownian Martingales
We now identify some basic martingales of Brownian motion:
Theorem 3.2 (Elementary martingales of Brownian motion). For a d-dimensional

Brownian motion (Bt ) the following processes are martingales w.r.t. each of the filtra-
tions (FtB ), (Ft ) and (FtP ):
(i)
(1). The coordinate processes Bt , 1 ≤ i ≤ d.
(i) (j)
(2). Bt Bt − t · δij for any 1 ≤ i, j ≤ d.

(3). exp(α · Bt − 21 |α|2 t) for any α ∈ Rd .
The processes Mtα = exp(α · Bt − 12 |α|2t) are called exponential martingales.
Proof. We only prove the second assertion for d = 1 and the right-continuous filtration
(Ft ). The verification of the remaining statements is left as an exercise.
For d = 1, since Bt is normally distributed, the Ft -measurable random variable Bt2 − t
is integrable for any t. Moreover, by Lemma 3.1,
E[Bt2 − Bs2 | Fs ] = E[(Bt − Bs )2 | Fs ] + 2Bs · E[Bt − Bs | Fs ]

= E[(Bt − Bs )2 ] + 2Bs · E[Bt − Bs ] = t − s
almost surely. Hence
E[Bt2 − t | Fs ] = Bs2 − s P -a.s. for any 0 ≤ s ≤ t,
i.e., Bt2 − t is an (Ft ) martingale.
Remark (Doob decomposition, variance process of Brownian motion). For a one-

dimensional Brownian motion (Bt ), the theorem yields the Doob decomposition
Bt2 = Mt + t
of the submartingale (Bt2 ) into a martingale (Mt ) and the continuous increasing adapted
process hBit = t.
A Doob decomposition of the process f (Bt ) for general functions f ∈ C 2 (R) will be
obtained below as a consequence of Itô’s celebrated formula. It states that
ˆt ˆt
′ 1
f (Bt ) − f (B0 ) = f (Bs ) dBs + f ′′ (Bs ) ds (3.1.2)
2
0 0
where the first integral is an Itô stochastic integral, cf. Section 6.3. If, for example, f ′ is
bounded, then the Itô integral is a martingale as a function of t. If f is convex then f (Bt )
is a submartingale and the second integral is a continuous increasing adapted process in

3.1. SOME FUNDAMENTAL MARTINGALES OF BROWNIAN MOTION 111
t. It is a consequence of (3.1.2) that Brownian motion solves the martingale problem for
the operator L f = f ′′ /2 with domain Dom(L ) = {f ∈ C 2 (R) : f ′ bounded}.
Itô’s formula (3.1.2) can also be extended to the multi-dimensional case, see Section
P 2
6.4 below. The second derivative is then replaced by the Laplacian ∆f = di=1 ∂∂xf2 .
i
The multi-dimensional Itô formula implies that a sub- or superharmonic function of d-
dimensional Brownian motion is a sub- or supermartingale respectively, if appropriate
integrability conditions hold. We now give a direct proof of this fact by the mean value
property:
Lemma 3.3 (Mean value property for harmonic function in Rd ). Suppose that h ∈
C 2 (Rd ) is a (super-)harmonic function, i.e.,
(≤)
∆h(x) = 0 for any x ∈ Rd .
Then for any x ∈ Rd and any rotationally invariant probability measure µ on Rd ,

ˆ
(≤)
h(x + y) µ(dy) = h(x). (3.1.3)
Proof. By the classical mean value property, h(x) is equal to (resp. greater or equal
than) the average value h of h on any sphere ∂Br (x) with center at x and radius
ffl
∂Br (x)
r > 0, cf. e.g. [XXXKönigsberger: Analysis II]. Moreover, if µ is a rotationally invari-
ant probability measure then the integral in (3.1.3) is an average of average values over
spheres: ˆ ˆ
(≤)
h(x + y) µ(dy) = h µR (dr) = h(x),
∂Br (x)
where µR is the distribution of R(x) = |x| under µ.
Theorem 3.4 (Superharmonic functions of Brownian motion are supermartin-

gales). If h ∈ C 2 (Rd ) is a (super-) harmonic function then (h(Bt )) is a (super-) mar-
tingale w.r.t. (Ft ) provided h(Bt ) (resp. h(Bt )+ ) is integrable for any t ≥ 0.

Proof. By Lemma 3.1 and the mean value property, we obtain
E[h(Bt ) | Fs ](ω) = E[h(Bs + Bt − Bs ) | Fs ](ω)

= E[h(Bs (ω) + Bt − Bs )]
ˆ
= h(Bs (ω) + y) N(0, (t − s) I)(dy)
(≤)
= h(Bs (ω))
for any 0 ≤ s ≤ t and P -almost every ω.
3.2 Optional Sampling and Optional Stopping
The Optional Sampling Theorem

The optional stopping theorem can be easily extended to continuous time martingales
with continuous sample paths. We directly prove a generalization:
Theorem 3.5 (Optional Sampling Theorem). Suppose that (Mt )t∈[0,∞] is a martingale
w.r.t. an arbitrary filtration (Ft ) such that t 7→ Mt (ω) is continuous for P -almost every
ω. Then
E[MT | FS ] = MS P -almost surely (3.2.1)
for any bounded (Ft ) stopping times S and T with S ≤ T .
We point out that an additional assumption on the filtration (e.g. right-continuity) is not
required in the theorem. Stopping times and the σ-algebra FS are defined for arbitrary
filtrations in complete analogy to the definitions for the filtration (FtB ) in Section 1.5.
Remark (Optional Stopping). By taking expectations in the Optional Sampling The-

orem, we obtain
E[MT ] = E[E[MT | F0 ]] = E[M0 ]
for any bounded stopping time T . For unbounded stopping times,
E[MT ] = E[M0 ]

3.2. OPTIONAL SAMPLING AND OPTIONAL STOPPING 113
holds by dominated convergence provided T < ∞ almost surely, and the random vari-
ables MT ∧n , n ∈ N, are uniformly integrable.
Proof of Theorem 3.5. We verify the defining properties of the conditional expectation
in (3.4) by approximating the stopping times by discrete random variables:
(1). MS has an FS -measurable modification: For n ∈ N let Sen = 2−n ⌊2n S⌋, i.e.,
Sen = k · 2−n on {k · 2−n ≤ S < (k + 1)2−n } for any k = 0, 1, 2, . . . .
We point out that in general, Sen is not a stopping time w.r.t. (Ft ). Clearly, the
sequence (Sen )n∈N is increasing with S = lim Sn . By almost sure continuity
MS = lim MSen P -almost surely. (3.2.2)

n→∞
On the other hand, each of the random variables MSen is FS -measurable. In fact,
X
MSen · I{S≤t} = Mk·2−n · I{k2−n ≤S<(k+1)2−n and S≤t}
k:k·2−n ≤t
is Ft -measurable for any t ≥ 0 since S is an (Ft ) stopping time. Therefore, by

fS := lim sup M e is an FS -measurable modifica-
(3.2.2), the random variable M Sn
n→∞
tion of MS .
(2). E[MT ; A] = E[MS ; A] for any A ∈ FS : For n ∈ N, the discrete random vari-
ables Tn = 2−n · ⌈2n T ⌉ and Sn = 2−n · ⌈2n S⌉ are (Ft ) stopping times satisfying
Tn ≥ Sn ≥ S, cf. the proof of Theorem 1.26. In particular, FS ⊆ FSn ⊆ FTn .
Furthermore, (Tn ) and (Sn ) are decreasing sequences with T = lim Tn and
S = lim Sn . As T and S are bounded random variables by assumption, the
sequences (Tn ) and (Sn ) are uniformly bounded by a finite constant c ∈ (0, ∞).
Therefore, we obtain

S(ω) Sn (ω)
Sen (ω)
Figure 3.1: Two ways to approximate a continuous stopping time.
X
E[MTn ; A] = E[Mk·2−n ; A ∩ {Tn = k · 2−n }]
k:k·2−n ≤c
X
= E[Mc ; A ∩ {Tn = k · 2−n }] (3.2.3)
k:k·2−n ≤c
= E[Mc ; A] for any A ∈ FTn ,
and similarly
E[MSn ; A] = E[Mc ; A] for any A ∈ FSn . (3.2.4)
In (3.2.3) we have used that (Mt ) is an (Ft ) martingale, and A ∩ {Tn = k · 2−n } ∈
Fk·2−n . A set A ∈ FS is contained both in FTn and FSn . Thus by (3.2.3) and
(3.2.4),
E[MTn ; A] = E[MSn ; A] for any n ∈ N and any A ∈ FS . (3.2.5)
As n → ∞, MTn → MT and MSn → MS almost surely by continuity. It remains

to show that the expectations in (3.2.5) converge as well. To this end note that by
(3.2.3) and (3.2.4),
MTn = E[Mc | FTn ] and MSn = E[Mc | FSn ] P -almost surely.

We will prove in Section 4.3 that any family of conditional expectations of a

given random variable w.r.t. different σ-algebras is uniformly integrable, and that
for uniformly integrable random variables a generalized Dominated Convergence
Theorem holds, cf. Theorem 4.13. Therefore, we finally obtain
E[MT ; A] = E[lim MTn ; A] = lim E[MTn ; A]

= lim E[MSn ; A] = E[lim MSn ; A] = E[MS ; A],
completing the proof of the theorem.
Remark (Measurability and completion). In general, the random variable MS is not

necessarily FS -measurable. However, we have shown in the proof that MS always has
fS . If the filtration contains all measure zero sets, then
an FS -measurable modification M
this implies that MS itself is FS -measurable and hence a version of E[MT | FS ].
Ruin probabilities and passage times revisited

Similarly as for random walks, the Optional Sampling Theorem can be applied to com-
pute distributions of passage times and hitting probabilities for Brownian motion. For a
one-dimensional Brownian motion (Bt ) starting at 0, and a, b > 0, let
T = inf{t ≥ 0 : Bt 6∈ (−b, a)} and Ta = inf{t ≥ 0 : Bt = a}
denote the first exit time from the interval (−b, a) and the first passage time to the point
a, respectively. In Section 1.5 we have computed the distribution of Ta by the reflection
principle. This and other results can be recovered by applying optional stopping to the
basic martingales of Brownian motion. The advantage of this approach is that it carries
over to other diffusion processes.
Exercise (Exit and passage times of Brownian motion). Prove by optional stopping:
(1). Law of the exit point: P [BT = a] = b/(a + b), P [BT = −b] = a/(a + b),
(2). Mean exit time: E[T ] = a · b and E[Ta ] = ∞,

(3). Laplace transform of passage times: For any s > 0,

√
E[exp(−sTa )] = exp(−a 2s).
Conclude that the distribution of Ta on (0, ∞) is absolutely continuous with density
fTa (t) = a · (2πt3 )−1/2 · exp(−a2 /2t).
Exit laws and Dirichlet problem

Applying optional stopping to harmonic functions of a multidimensional Brownian mo-
tion yields a generalization of the mean value property and a stochastic representation
for solutions of the Dirichlet problem. This will be exploited in full generality in Chap-
ter 7. Here, we only sketch the basic idea.
Suppose that h ∈ C 2 (Rd ) is a harmonic function and that (Bt )t≥0 is a d-dimensional
Brownian motion starting at x w.r.t. the probability measure Px . Assuming that
Ex [h(Bt )] < ∞ for any t ≥ 0,
the mean value property for harmonic functions implies that h(Bt ) is a martingale under
Px , cf. Theorem 3.4. The first hitting time T = inf{t ≥ 0 : Bt ∈ Rd \ D} of the com-
plement of an open set D ⊆ Rd is a stopping time w.r.t. the filtration (FtB ). Therefore,
by Theorem 3.5 and the remark below, we obtain
Ex [h(BT ∧n )] = Ex [h(B0 )] = h(x) for any n ∈ N. (3.2.6)
Now let us assume in addition that the set D is bounded. Then T is almost surely
finite, and the sequence of random variables h(BT ∧n ) (n ∈ N) is uniformly bounded
because BT ∧n takes values in the closure D for any n ∈ N. Applying the Dominated
Convergence Theorem to (3.2.6), we obtain the integral representation
ˆ
h(x) = Ex [h(BT )] = h(y)µx (dy) (3.2.7)
∂D
where µx = Px ◦ BT−1 denotes the exit law from D for Brownian motion starting at x.
In Chapter 7, we show that the representation (3.2.7) still holds true if h is a continuous

function defined on D that is C 2 and harmonic on D. The proof requires localization

techniques that will be developed below in the context of stochastic calculus. For the
moment we note that the representation (3.2.7) has several important aspects and appli-
cations:
Generalized mean value property for harmonic functions. For any bounded do-
main D ⊆ Rd and any x ∈ D, h(x) is the average of the boundary values of h on ∂D
w.r.t. the measure µx .
Stochastic representation for solutions of the Dirichlet problem. A solution h ∈

C (D) ∩ C(D) of the Dirichlet problem
2
∆h(x) = 0 for x ∈ D, (3.2.8)

h(x) = f (x) for x ∈ ∂D,
has a stochastic representation
h(x) = Ex [f (BT )] for any x ∈ D. (3.2.9)
Monte Carlo solution of the Dirichlet problem. The stochastic representation (3.2.9)
can be used as the basis of a Monte Carlo method for computing the harmonic function
h(x) approximately by simulating a large number n of sample paths of Brownian motion
starting at x, and estimating the expectation by the corresponding empirical average. Al-
though in many cases classical numerical methods are more efficient, the Monte Carlo
method is useful in high dimensional cases. Furthermore, it carries over to far more
general situations.
Computation of exit law. Conversely, if the Dirichlet problem (3.2.8) has a unique
solution h, then computation of h (for example by standard numerical methods) enables
us to obtain the expectations in (3.2.8). In particular, the probability h(x) = Px [BT ∈ A]
for Brownian motion exiting the domain on a subset A ⊆ ∂D is informally given as the
solution of the Dirichlet problem
∆h = 0 on D, h = IA on ∂D.

This can be made rigorous under regularity assumptions. The full exit law is the har-
monic measure, i.e., the probability measure µx such that the representation (3.2.7) holds
for any function h ∈ C 2 (D) ∩ C(D) with ∆h = 0 on D. For simple domains such as
half-spaces, balls and cylinders, this harmonic measure can be computed explicitly.
Example (Exit laws from balls). For d ≥ 2, the exit law from the unit ball D = {y ∈
Rd : |y| < 1} for Brownian motion starting at a point x ∈ Rd with |x| < 1 is given by
1 − |x|2
µx (dy) = ν(dy)
|y − x|d
where ν denotes the normalized surface measure on the unit sphere S d−1 = {y ∈ Rd :
|y| = 1}. Indeed, the classical Poisson integral formula states that for any f ∈ C(S d−1 ),
the function ˆ
h(x) = f (y) µx (dy)
solves the Dirichlet problem on D with boundary values lim h(x) = f (z) for any z ∈
x→z
S d−1 , cf. e.g. [XXX Karatzas/Shreve, Ch. 4]. Hence by (3.2.9),
1 − |x|2
ˆ
Ex [f (BT )] = f (y) ν(dy)
|y − x|d
holds for any f ∈ C(S d−1 ), and thus by a standard approximation argument, for any
indicator function of a measurable subset of S d−1 .
3.3 Maximal inequalities and the Law of the Iterated

Logarithm
The extension of Doob’s maximal inequality to the continuous time case is straight-
forward. As a first application, we give a proof for the upper bound in the law of the
iterated logarithm.
Maximal inequalities in continuous time

3.3. MAXIMAL INEQUALITIES AND THE LIL 119
Theorem 3.6 (Doob’s Lp inequality in continuous time). Suppose that (Mt )t∈[0,∞) is
a martingale with almost surely right continuous sample paths t 7→ Mt (ω). Then the
1 1
following estimates hold for any a ∈ [0, ∞), p ∈ [1, ∞), q ∈ (1, ∞] with p
+ q
= 1,
and c > 0:
" #
(1). P sup |Mt | ≥ c ≤ c−p · E[|Ma |p ],
t∈[0,a]

(2). sup |Mt | ≤ q · kMa kLp .
t∈[0,a]
Lp
Remark. The same estimates hold for non-negative submartingales.
Proof. Let (πn ) denote an increasing sequence of partitions of the interval [0, a] such
that the mesh size of πn goes to 0 as n → ∞. By Corollary 2.13 applied to the discrete
time martingale (Mt )t∈πn , we obtain

P max |Mt | ≥ c ≤ E[|Ma |p ]/cp for any n ∈ N.
t∈πn
Moreover, as n → ∞,
max |Mt | ր sup |Mt | almost surely

t∈πn t∈[0,a]
by right continuity of the sample paths. Hence

" # " #
[
P sup |Mt | > c = P max |Mt | > c
t∈[0,a] t∈τn
n

= lim P max |Mt | > c ≤ E[|Ma |p ]/cp .
n→∞ t∈τn
The first assertion now follows by replacing c by c − ε and letting ε tend to 0. The
second assertion follows similarly from Theorem 2.15.
As a first application of the maximal inequality to Brownian motion, we derive an upper

bound for the probability that the graph of one-dimensional Brownian motion passes a
line in R2 :

Lemma 3.7 (Passage probabilities for lines). For a one-dimensional Brownian motion
(Bt ) starting at 0 we have
P [Bt ≥ β + αt/2 for some t ≥ 0] ≤ exp(−αβ) for any α, β > 0.
Proof. Applying the maximal inequality to the exponential martingale
Mtα = exp(αBt − α2 t/2)
yields
" #
P [Bt ≥ β + αt/2 for some t ∈ [0, a]] = P sup (Bt − αt/2) ≥ β
t∈[0,a]
" #
= P sup Mtα ≥ exp(αβ) ≤ exp(−αβ) · E[Maα ] = exp(−αβ)
t∈[0,a]
for any a > 0. The assertion follows in the limit as a → ∞.
With slightly more effort, it is possible to compute the passage probability and the dis-
tribution of the first passage time of a line explicitly, cf. ?? below.
Application to LIL
A remarkable consequence of Lemma 3.7 is a simplified proof for the upper bound half
of the Law of the Iterated Logarithm:

Theorem 3.8 (LIL, upper bound). For a one-dimensional Brownian motion (Bt ) start-
ing at 0,
Bt
lim sup p ≤ +1 P -almost surely. (3.3.1)
tց0 2t log log t−1
Proof. Let δ > 0. We would like to show that almost surely,
Bt ≤ (1 + δ)h(t) for sufficiently small t > 0,
p
where h(t) := 2t log log t−1 . Fix θ ∈ (0, 1). The idea is to approximate the function
h(t) by affine functions
ln (t) = βn + αn t/2
on each of the intervals [θn , θn−1 ], and to apply the upper bounds for the passage prob-
abilities from the lemma.

θ5 θ4 θ3 θ2 θ 1
We choose αn and βn in a such way that ln (θn ) = h(θn ) and ln (0) = h(θn )/2, i.e.,
βn = h(θn )/2 and αn = h(θn )/θn .
For this choice we have ln (θn ) ≥ θ · ln (θn−1 ), and hence
ln (θn )
ln (t) ≤ ln (θn−1 ) ≤ (3.3.2)
θ
h(θn ) h(t)
= ≤ for any t ∈ [θn , θn−1 ].
θ θ
ln (t)
h(t)
h(θn )
h(θn )/2
θn θn−1
We now want to apply the Borel-Cantelli lemma to show that with probability one,
Bt ≤ (1 + δ)ln (t) for large n. By Lemma 3.7,
P [Bt ≥ (1 + δ)ln (t) for some t ≥ 0] ≤ exp(−αn βn · (1 + δ)2 )

h(θn )2 2
= exp − · (1 + δ) .
2θn
p
Choosing h(t) = 2t log log t−1 , the right hand side is equal to a constant multiple of
2
n−(1+δ) , which is a summable sequence. Note that we do not have to know the precise
form of h(t) in advance to carry out the proof – we just choose h(t) in such a way that
the probabilities become summable!
Now, by Borel-Cantelli, for P -almost every ω there exists N(ω) ∈ N such that
Bt (ω) ≤ (1 + δ)ln (t) for any t ∈ [0, 1] and n ≥ N(ω). (3.3.3)
By (3.3.2), the right hand side of (3.3.3) is dominated by (1+δ)h(t)/θ for t ∈ [θn , θn−1 ].
Hence
1+δ [
Bt ≤ h(t) for any t ∈ [θn , θn−1 ],
θ n≥N
i.e., for any t ∈ (0, θ N −1

), and therefore,
Bt 1+δ
lim sup ≤ P -almost surely.
tց0 h(t) θ
The assertion then follows in the limit as θ ր 1 and δ ց 0.

Since (−Bt ) is again a Brownian motion starting at 0, the upper bound (3.3.1) also
implies
Bt
lim inf p ≥ −1 P -almost surely. (3.3.4)
tց0 2t log log t−1
The converse bounds are actually easier to prove since we can use the independence of
the increments and apply the second Borel-Cantelli Lemma. We only mention the key
steps and leave the details as an exercise:
Exercise (Complete proof of LIL). Prove the Law of the Iterated Logarithm:
Bt Bt
lim sup = +1 and lim inf = −1
tց0 h(t) tց0 h(t)
p
where h(t) = 2t log log t−1 . Proceed in the following way:
(1). Let θ ∈ (0, 1) and consider the increments Zn = Bθn − Bθn+1 , n ∈ N. Show that
if ε > 0, then
P [Zn > (1 − ε)h(θn ) infinitely often] = 1.

´∞
(Hint: x
exp(−z 2 /2)dz ≤ x−1 exp(−x2 /2))
(2). Conclude that by (3.3.4),

Bt
lim sup ≥ 1−ε P -almost surely for any ε > 0,
tց0 h(t)
and complete the proof of the LIL by deriving the lower bounds
Bt Bt
lim sup ≥ 1 and lim inf ≤ −1 P -almost surely. (3.3.5)
tց0 h(t) tց0 h(t)

Chapter 4
Martingale Convergence Theorems
The strength of martingale theory is partially due to powerful general convergence the-
orems that hold for martingales, sub- and supermartingales. In this chapter, we study
convergence theorems with different types of convergence including almost sure, L2
and L1 convergence, and consider first applications.
At first, we will again focus on discrete-parameter martingales – the results can then be
easily extended to continuous martingales.
4.1 Convergence in L2
Already when proving the Law of Large Numbers, L2 convergence is much easier to
show than, for example, almost sure convergence. The situation is similar for mar-
tingales: A necessary and sufficient condition for convergence in the Hilbert space
L2 (Ω, A, P ) can be obtained by elementary methods.
Martingales in L2
Consider a discrete-parameter martingale (Mn )n≥0 w.r.t. a filtration (Fn ) on a probabil-
ity space (Ω, A, P ). Throughout this section we assume:
Assumption (Square integrability). E[Mn2 ] < ∞ for any n ≥ 0.
We start with an important remark:
125
126 CHAPTER 4. MARTINGALE CONVERGENCE THEOREMS
Lemma 4.1. The increments Yn = Mn − Mn−1 of a square-integrable martingale are

centered and orthogonal in L2 (Ω, A, P ) (i.e. uncorrelated).
Proof. By definition of a martingale, E[Yn | Fn−1 ] = 0 for any n ≥ 0. Hence E[Yn ] = 0

and E[Ym Yn ] = E[Ym · E[Yn | Fn−1 ]] = 0 for 0 ≤ m < n.
Since the increments are also orthogonal to M0 by an analogue argument, a square

integrable martingale sequence consists of partial sums of a sequence of uncorrelated
random variables: n
X
Mn = M0 + Yk for any n ≥ 0.
k=1
The Convergence Theorem

The central result of this section shows that an L2 -bounded martingale (Mn ) can always
be extended to n ∈ {0, 1, 2, . . .} ∪ {∞}:
Theorem 4.2 (L2 Martingale Convergence Theorem). The martingale sequence

(Mn ) converges in L2 (Ω, A, P ) as n → ∞ if and only if it is bounded in L2 in the
sense that
sup E[Mn2 ] < ∞. (4.1.1)
n≥0
In this case, the representation
Mn = E[M∞ | Fn ]
holds almost surely for any n ≥ 0, where M∞ denotes the limit of Mn in L2 (Ω, A, P ).
We will prove in the next section that (Mn ) does also converge almost surely to M∞ .
An analogue result to Theorem 4.2 holds with L2 replaced by Lp for any p ∈ (1, ∞) but
not for p = 1, cf. Section 4.3 below.
Proof. (1). Let us first note that
E[(Mn − Mm )2 ] = E[Mn2 ] − E[Mm

2
] for 0 ≤ m ≤ n. (4.1.2)

4.1. CONVERGENCE IN L2 127
Indeed,
E[Mn2 ] − E[Mm
2
] = E[(Mn − Mm )(Mn + Mm )]
= E[(Mn − Mm )2 ] + 2E[Mm · (Mn − Mm )],
and the last term vanishes since the increment Mn − Mm is orthogonal to Mm in

L2 .
(2). To prove that (4.1.1) is sufficient for L2 convergence, note that the sequence
(E[Mn2 ])n≥0 is increasing by (4.1.2). If (4.1.1) holds then this sequence is bound-
ed, and hence a Cauchy sequence. Therefore, by (4.1.2), (Mn ) is a Cauchy se-
quence in L2 . Convergence now follows by completeness of L2 (Ω, A, P ).
(3). Conversely, if (Mn ) converges in L2 to a limit M∞ , then the L2 norms are bound-
ed. Moreover, by Jensen’s inequality, for each fixed k ≥ 0,
E[Mn | Fk ] −→ E[M∞ | Fk ] in L2 (Ω, A, P ) as n → ∞.
As (Mn ) is a martingale, we have E[Mn | Fk ] = Mk for n ≥ k, and hence
Mk = E[M∞ | Fk ] P -almost surely.
Remark (Functional analytic interpretation of L2 convergence theorem). The asser-

tion of the L2 martingale convergence theorem can be rephrased as a purely functional
analytic statement:
P
∞
An infinite sum Yk of orthogonal vectors Yk in the Hilbert space L2 (Ω, A, P ) is
k=1
P
n
convergent if and only if the sequence of partial sums Yk is bounded.
k=1
How can boundedness in L2 be verified for martingales? Writing the martingale (Mn )
as the sequence of partial sums of its increments Yn = Mn − Mn−1 , we have
n n
! n
X X X
2 2
E[Mn ] = M0 + Yk , M0 + Yk = E[M0 ] + E[Yk2 ]
k=1 k=1 L2 k=1

by orthogonality of the increments and M0 . Hence

∞
X
sup E[Mn2 ] = E[M02 ] + E[Yk2 ].
n≥0
k=1
Alternatively, we have E[Mn2 ] = E[M02 ] + E[hMin ]. Hence by monotone convergence
sup E[Mn2 ] = E[M02 ] + E[hMi∞ ]

n≥0
where hMi∞ = suphMin .
Summability of sequences with random signs

As a first application we study the convergence of series with coefficients with random
signs. In an introductory analysis course it is shown as an application of the integral and
Leibniz criterion for convergence of series that
P
∞
n−α converges ⇐⇒ α>1 , whereas
n=1
P
∞
(−1)n n−α converges ⇐⇒ α > 0.
n=1
Therefore, it seems interesting to see what happens if the signs are chosen randomly.
The L2 martingale convergence theorem yields:
Corollary 4.3. Let (an ) be a real sequence. If (εn ) is a sequence of independent random
variables on (Ω, A, P ) with P [εn = +1] = P [εn = −1] = 1/2, then
∞
X ∞
X
2
εn an converges in L (Ω, A, P ) ⇐⇒ a2n < ∞.
n=1 n=1
P
n
Proof. The sequence Mn = εk ak of partial sums is a martingale with
k=1
∞
X ∞
X
sup E[Mn2 ] = E[ε2k a2k ] = a2k .
n≥0
k=1 k=1

4.1. CONVERGENCE IN L2 129
P
∞
Example. The series εn · n−α converges in L2 if and only if α > 12 .
n=1
Remark (Almost sure asymptotics). By the Supermartingale Convergence Theorem

P P
(cf. Theorem 4.5 below), the series εn an also converges almost surely if a2n < ∞.
P 2
On the other hand, if an = ∞ then the series of partial sums has almost surely
unbounded oscillations:
P P
n
Exercise. Suppose that an = ∞, and let Mn = εk ak .
k=1
(1). Compute the conditional variance process hMin .
(2). For c > 0 let Tc = inf{n ≥ 0 : |Mn | ≥ c}. Apply the Optional Stopping
Theorem to the martingale in the Doob decomposition of (Mn2 ), and conclude
that P [Tc = ∞] = 0.
(3). Prove that (Mn ) has almost surely unbounded oscillations.
L2 convergence in continuous time

The L2 convergence theorem directly extends to the continuous-parameter case.
Theorem 4.4 (L2 Martingale Convergence Theorem in continuous time). Let a ∈

(0, ∞]. If (Mt )t∈[0,a) is a martingale w.r.t. a filtration (Ft )t∈[0,a) such that
sup E[Mt2 ] < ∞

t∈[0,u)
then Mu = lim Mt exists in L2 (Ω, A, P ) and (Mt )t∈[0,u] is again a square-integrable

tրu
martingale.
Proof. Choose any increasing sequence tn ∈ [0, u) such that tn → u. Then (Mtn ) is an
L2 -bounded discrete-parameter martingale. Hence the limit Mu = lim Mtn exists in L2 ,
and
Mtn = E[Mu | Ftn ] for any n ∈ N. (4.1.3)

For an arbitrary t ∈ [0, u), there exists n ∈ N with tn ∈ (t, u). Hence
Mt = E[Mtn | Ft ] = E[Mu | Ft ]
by (4.1.3) and the tower property. In particular, (Mt )t∈[0,u] is a square-integrable mar-
tingale. By orthogonality of the increments,
E[(Mu − Mtn )2 ] = E[(Mu − Mt )2 ] + E[(Mt − Mtn )2 ] ≥ E[(Mu − Mt )2 ]
whenever tn ≤ t ≤ u. Since Mtn → Mu in L2 , we obtain
lim E[(Mu − Mt )2 ] = 0.
tրu
Remark. (1). Note that in the proof it is enough to consider a fixed sequence tn ր u.
(2). To obtain almost sure convergence, an additional regularity condition on the sam-
ple paths (e.g. right-continuity) is required, cf. below. This assumption is not
needed for L2 convergence.
4.2 Almost sure convergence of supermartingales

Let (Zn )n≥0 be a discrete-parameter supermartingale w.r.t. a filtration (Fn )n≥0 on a
probability space (Ω, A, P ). The following theorem yields a stochastic counterpart to
the fact that any lower bounded decreasing sequence of reals converges to a finite limit:
Theorem 4.5 (Supermartingale Convergence Theorem, Doob). If sup E[Zn− ] < ∞

n≥0
then (Zn ) converges almost surely to a random variable Z∞ ∈ L1 (Ω, A, P ).
In particular, supermartingales that are uniformly bounded from below converge almost
surely to an integrable random variable.

4.2. ALMOST SURE CONVERGENCE OF SUPERMARTINGALES 131
Remark (L1 boundedness vs. L1 convergence). (1). The condition sup E[Zn− ] < ∞
holds if and only if (Zn ) is bounded in L1 . Indeed, as E[Zn+ ] < ∞ by our defini-
tion of a supermartingale, we have
E[ |Zn | ] = E[Zn ] + 2E[Zn− ] ≤ E[Z0 ] + 2E[Zn− ] for any n ≥ 0.
(2). Although (Zn ) is bounded in L1 and the limit is integrable, L1 convergence does
not hold in general, cf. the examples below.
For proving the Supermartingale Convergence Theorem, we introduce the number

U (a,b) (ω) of upcrossings of an interval (a, b) by the sequence Zn (ω), cf. below for the
exact definition.
1st upcrossing 2nd upcrossing

Note that if U (a,b)
(ω) is finite for every non-empty bounded interval [a, b] then
lim sup Zn (ω) and lim inf Zn (ω) coincide, i.e., the sequence (Zn (ω)) converges. There-
fore, to show almost sure convergence of (Zn ), we derive an upper bound for U (a,b) . We
first prove this key estimate and then complete the proof of the theorem.
Doob’s upcrossing inequality

(a,b)
For n ∈ N and a, b ∈ R with a < b, we define the number Un of upcrossings of the
interval [a, b] before time n by

Un(a,b) = max k ≥ 0 : ∃ 0 ≤ s1 < t1 < s2 < t2 . . . < sk < tk ≤ n :

Zsi (ω) ≤ a, Zti (ω) ≥ b .

Lemma 4.6 (Doob). If (Zn ) is a supermartingale then
(b − a) · E[Un(a,b) ] ≤ E[(Zn − a)− ] for any a < b and n ≥ 0.
Proof. We may assume E[Zn− ] < ∞ since otherwise there is nothing to prove. The key
idea is to set up a predictable gambling strategy that increases our capital by (b − a)
for each completed upcrossing. Since the net gain with this strategy should again be a
supermartingale this yields an upper bound for the average number of upcrossings. Here
is the strategy:
• Wait until Zk ≤ a.
repeat
• Then play unit stakes until Zk ≥ b.
The stake Ck in round k is


1 if Z0 ≤ a,
C1 =
0 otherwise,
and for k ≥ 2,

1 if (Ck−1 = 1 and Zk−1 ≤ b) or (Ck−1 = 0 and Zk−1 ≤ a),
Ck = .
0 otherwise
Clearly, (Ck ) is a predictable, bounded and non-negative sequence of random variables.

Moreover, Ck · (Zk − Zk−1 ) is integrable for any k ≤ n, because Ck is bounded and

E |Zk | = 2E[Zk+ ] − E[Zk ] ≤ 2E[Zk+ ] − E[Zn ] ≤ 2E[Zk+ ] − E[Zn− ]
for k ≤ n. Therefore, by Theorem 2.6 and the remark below, the process
k
X
(C• Z)k = Ci · (Zi − Zi−1 ), 0 ≤ k ≤ n,
i=1
is again a supermartingale.
Clearly, the value of the process C• Z increases by at least (b − a) units during each

completed upcrossing. Between upcrossing periods, the value of (C• Z)k is constant.
Finally, if the final time n is contained in an upcrossing period, then the process can
decrease by at most (Zn − a)− units during that last period (since Zk might decrease
before the next upcrossing is completed). Therefore, we have
(C• Z)n ≥ (b − a) · Un(a,b) − (Zn − a)− , i.e.,
(b − a) · Un(a,b) ≤ (C• Z)n + (Zn − a)− .
Zn
Gain ≥ b − a Gain ≥ b − a Loss ≤ (Zn − a)−

Since C• Z is a supermartingale with initial value 0, we obtain the upper bound
(b − a)E[Un(a,b) ] ≤ E[(C• Z)n ] + E[(Zn − a)− ] ≤ E[(Zn − a)− ].
Proof of Doob’s Convergence Theorem

We can now complete the proof of Theorem 4.5.
Proof. Let
U (a,b) = sup Un(a,b)
n∈N
denote the total number of upcrossings of the supermartingale (Zn ) over an interval
(a, b) with −∞ < a < b < ∞. By the upcrossing inequality and monotone conver-
gence,
1
E[U (a,b) ] = lim E[Un(a,b) ] ≤ · sup E[(Zn − a)− ]. (4.2.1)
n→∞ b − a n∈N

Assuming sup E[Zn− ] < ∞, the right hand side of (4.2.1) is finite since (Zn − a)− ≤
|a| + Zn− . Therefore,
U (a,b) < ∞ P -almost surely,
and hence the event

[
{lim inf Zn 6= lim sup Zn } = {U (a,b) = ∞}
a,b∈Q
a<b
has probability zero. This proves almost sure convergence.
It remains to show that the almost sure limit Z∞ = lim Zn is an integrable random
variable (in particular, it is finite almost surely). This holds true as, by the remark below
Theorem 4.5, sup E[Zn− ] < ∞ implies that (Zn ) is bounded in L1 , and therefore
E[ |Z∞ | ] = E[lim inf |Zn | ] ≤ lim inf E[ |Zn | ] < ∞
by Fatou’s lemma.
Examples and first applications

We now consider a few prototypic applications of the almost sure convergence theorem:
Example (1. Sums of i.i.d. random variables). Consider a Random Walk

n
X
Sn = ηi
i=1
on R with centered and bounded increments
ηi i.i.d. with |ηi | ≤ c and E[ηi ] = 0, c ∈ R.
Suppose that P [ηi 6= 0] > 0. Then there exists ε > 0 such that P [|ηi | ≥ ε] > 0. As the
increments are i.i.d., the event {|ηi | ≥ ε} occurs infinitely often with probability one.
Therefore, almost surely, the martingale (Sn ) does not converge as n → ∞.
Now let a ∈ R. We consider the first hitting time
Ta = inf{t ≥ 0 : Sn ≥ a}

of the interval [a, ∞). By the Optional Stopping Theorem, the stopped Random Walk
(STa ∧n )n≥0 is again a martingale. Moreover, as Sk < a for any k < Ta and the incre-
ments are bounded by c, we obtain the upper bound
STa ∧n < a + c for any n ∈ N.
Therefore, the stopped Random Walk converges almost surely by the Supermartingale
Convergence Theorem. As (Sn ) does not converge, we can conclude that
P [Ta < ∞] = 1 for any a > 0, i.e., lim sup Sn = ∞ almost surely.
Since (Sn ) is also a submartingale, we obtain
lim inf Sn = −∞ almost surely
by an analogue argument. A generalization of this result is given in Theorem 4.7 below.
Remark (Almost sure vs. Lp convergence). In the last example, the stopped process
does not converge in Lp for any p ∈ [1, ∞). In fact,
lim E[STa ∧n ] = E[STa ] ≥ a whereas E[STa ∧n ] = E[S0 ] = 0 for all n.

n→∞
Example (2. Products of non-negative i.i.d. random variables). Consider a growth

process
n
Y
Zn = Yi
i=1
with i.i.d. factors Yi ≥ 0 with finite expectation α ∈ (0, ∞). Then
Mn = Zn /αn
is a martingale. By the almost sure convergence theorem, a finite limit M∞ exists al-
most surely, because Mn ≥ 0 for all n. For the almost sure asymptotics of (Zn ), we
distinguish three different cases:
(1). α < 1: In this case,

Zn = Mn · αn
converges to 0 exponentially fast with probability one.

(2). α = 1: Here (Zn ) is a martingale and converges almost surely to a finite limit. If
P [Yi 6= 1] > 0 then there exists ε > 0 such that Yi ≥ 1 + ε infinitely often with
probability one. This is consistent with convergence of (Zn ) only if the limit is
zero. Hence, if (Zn ) is not almost surely constant, then also in the critical case,
Zn → 0 almost surely.
(3). α > 1 (supercritical): In this case, on the set {M∞ > 0},
Zn = Mn · αn ∼ M∞ · αn ,
i.e., (Zn ) grows exponentially fast. The asymptotics on the set {M∞ = 0} is not
evident and requires separate considerations depending on the model.
Although most of the conclusions in the last example could have been obtained without
martingale methods (e.g. by taking logarithms), the martingale approach has the advan-
tage of carrying over to far more general model classes. These include for example
branching processes or exponentials of continuous time processes.
Example (3. Boundary behaviors of harmonic functions). Let D ⊆ Rd be a bounded

open domain, and let h : D → R be a harmonic function on D that is bounded from
below:
∆h(x) = 0 for any x ∈ D, inf h(x) > −∞. (4.2.2)
x∈D
To study the asymptotic behavior of h(x) as x approaches the boundary ∂D, we con-
struct a Markov chain (Xn ) such that h(Xn ) is a martingale: Let r : D → (0, ∞) be a
continuous function such that
0 < r(x) < dist(x, ∂D) for any x ∈ D, (4.2.3)
and let (Xn ) w.r.t Px denote the canonical time-homogeneous Markov chain with state
space D, initial value x, and transition probabilities
p(x, dy) = Uniform distribution on {y ∈ Rd : |y − x| = r(x)}.

D
x
r(x)
By (4.2.3), the function h is integrable w.r.t. p(x, dy), and, by the mean value property,
(ph)(x) = h(x) for any x ∈ D.
Therefore, the process h(Xn ) is a martingale w.r.t. Px for each x ∈ D. As h(Xn ) is

lower bounded by (4.2.2), the limit as n → ∞ exists Px -almost surely by the Super-
martingale Convergence Theorem. In particular, since the coordinate functions x 7→ xi
are also harmonic and lower bounded on D, the limit X∞ = lim Xn exists Px -almost
n→∞
surely. Moreover, X∞ is in ∂D, because r is bounded from below by a strictly positive
constant on any compact subset of D.
Summarizing we have shown:
(1). Boundary regularity: If h is harmonic and bounded from below on D then the
limit lim h(Xn ) exists along almost every trajectory Xn to the boundary ∂D.
n→∞
(2). Representation of h in terms of boundary values: If h is continuous on D, then

h(Xn ) → h(X∞ ) Px -almost surely and hence
h(x) = lim Ex [h(Xn )] = E[h(X∞ )],

n→∞
i.e., the law of X∞ w.r.t. Px is the harmonic measure on ∂D.
Note that, in contrast to classical results from analysis, the first statement holds without
any smoothness condition on the boundary ∂D. Thus, although boundary values of h
may not exist in the classical sense, they do exist along almost every trajectory of the
Markov chain!

Generalized Borel-Cantelli Lemma

Another application of the almost sure convergence theorem is a generalization of the
Borel-Cantelli lemmas. We first prove a dichotomy for the asymptotic behavior of mar-
tingales with L1 -bounded increments:
Theorem 4.7 (Asymptotics of martingales with L1 bounded increments). Suppose

that (Mn ) is a martingale, and there exists an integrable random variable Y such that
|Mn − Mn−1 | ≤ Y for any n ∈ N.
Then for P -almost every ω, the following dichotomy holds:
Either: The limit lim Mn (ω) exists in R,

n→∞
or: lim sup Mn (ω) = +∞ and lim inf Mn (ω) = −∞.

n→∞ n→∞
The theorem and its proof are a generalization of Example 1 above.
Proof. For a ∈ (−∞, 0) let Ta = min{n ≥ 0 : Mn ≥ a}. By the Optional Stopping

Theorem, (MTa ∧n ) is a martingale. Moreover,
MTa ∧n ≥ min(M0 , a − Y ) for any n ≥ 0,
and the right hand side is an integrable random variable. Therefore, (Mn ) converges
almost surely on {Ta = ∞}. Since this holds for every a < 0, we obtain almost sure
convergence on the set
[
{lim inf Mn > −∞} = {Ta = ∞}.
a<0
a∈Q
Similarly, almost sure convergence follows on the set {lim sup Mn < ∞}.
Now let (Fn )n≥0 be an arbitrary filtration. As a consequence of Theorem 4.7 we obtain:

Corollary 4.8 (Generalized Borel-Cantelli Lemma). If (An ) is a sequence of events

with An ∈ Fn for any n, then the equivalence
∞
X
ω ∈ An infinitely often ⇐⇒ P [An | Fn−1 ](ω) = ∞
n=1
holds for almost every ω ∈ Ω.
P
n P
n
Proof. Let Sn = IAk and Tn = E[IAk | Fk−1]. Then Sn and Tn are almost surely
k=1 k=1
increasing sequences. Let S∞ = sup Sn and T∞ = sup Tn denote the limits on [0, ∞].
The claim is that almost surely,
S∞ = ∞ ⇐⇒ T∞ = ∞. (4.2.4)
To prove (4.2.4) we note that Sn − Tn is a martingale with bounded increments. There-

fore, almost surely, Sn − Tn converges to a finite limit, or (lim sup(Sn − Tn ) = ∞ and
lim inf(Sn − Tn ) = −∞). In the first case, (4.2.4) holds. In the second case, S∞ = ∞
and T∞ = ∞, so (4.2.4) holds, too.
The assertion of Corollary 4.8 generalizes both classical Borel-Cantelli Lemmas: If

(An ) is an arbitrary sequence of events in a probability space (Ω, A, P ) then we can
consider the filtration Fn = σ(A1 , . . . , An ). By Corollary 4.8 we obtain:
P P
1st Borel-Cantelli Lemma: If P [An ] < ∞ then P [An | Fn−1] < ∞ almost surely,
and therefore
P [An infinitely often] = 0.
P
2nd Borel-Cantelli Lemma: If P [An ] = ∞ and the An are independent then
P P
P [An | Fn−1 ] = P [An ] = ∞ almost surely, and therefore
P [An infinitely often] = 1.

Upcrossing inequality and convergence theorem in continuous time

The upcrossing inequality and the supermartingale convergence theorem carry over im-
mediately to the continuous time case if we assume right continuity (or left continuity)
of the sample paths. Let u ∈ (0, ∞], and let (Zs )s∈[0,u) be a supermartingale in contin-
uous time w.r.t. a filtration (Fs ). We define the number of upcrossings of (Zs ) over an
interval (a, b) before time t as the supremum of the number of upcrossings over all time
discretizations (Zs )s∈π where π is a partition of the interval [0, t]:
(a,b)
Ut [Z] := sup U (a,b) [(Zs )s∈π ].
π⊂[0,t]
finite
Note that if (Zs ) has right-continuous sample paths and (πn ) is a sequence of partitions
of [0, t] such that 0, t ∈ π0 , πn ⊂ πn+1 and mesh(πn ) → 0 then
(a,b)
Ut [Z] = lim U (a,b) [(Zs )s∈πn ].
n→∞
Theorem 4.9 (Supermatingale Convergence Theorem in continuous time). Suppose

that (Zs )s∈[0,u) is a right continuous supermartingale.
(1). Upcrossing inequality: For any t ∈ [0, u) and a < b,
(a,b) 1
E[Ut ] ≤ E[(Zt − a)− ].
b−a
(2). Convergence Theorem: If sup E[Zs− ] < ∞, then the limit Zu− = lim Zs exists
s∈[0,u) sրu
almost surely, and Zu− is an integrable random variable.
Proof. (1). By the upcrossing inequality in discrete time,
E[U (a,b) [(Zs )s∈πn ]] ≤ E[(Zt − a)− ] for any n ∈ N,
where (πn ) is a sequence of partitions as above. The assertion now follows by the
Monotone Convergence Theorem.

4.3. UNIFORM INTEGRABILITY AND L1 CONVERGENCE 141
(2). The almost sure convergence can now be proven in the same way as in the discrete
time case.
More generally than stated above, the upcrossing inequality also implies that for a right-
continuous supermartingale (Zs )s∈[0,u) all the left limits lim Zs , t ∈ [0, u), exist simul-
sրt
taneously with probability one. Thus almost every sample path is càdlàg (continue à
droite, limites a gauche, i.e., right continuous with left limits). By similar arguments,
the existence of a modification with right continuous (and hence càdlàg) sample paths
can be proven for any supermartingale (Zs ) provided the filtration is right continuous
and complete, and s 7→ E[Zs ] is right continuous, cf. e.g. [XXXRevuz/Yor, Ch.II,§2].
4.3 Uniform integrability and L1 convergence
The Supermartingale Convergence Theorem shows that every supermartingale (Zn ) that
is bounded in L1 converges almost surely to an integrable limit Z∞ . However, L1 con-
vergence does not necessarily hold:
Qn
Example. (1). Suppose that Zn = i=1 Yi where the Yi are i.i.d. with E[Yi ] = 1,
P [Yi 6= 1] > 0. Then, Zn → 0 almost surely, cf. Example 2 in Section 4.2. On
the other hand, L1 convergence does not hold as E[Zn ] = 1 for any n.
(2). Similarly, the exponential martingale Mt = exp(Bt − t/2) of a Brownian motion

converges to 0 almost surely, but E[Mt ] = 1 for any t.
L1 convergence of martingales is of interest because it implies that a martingale se-

quence (Mn ) can be extended to n = ∞, and the random variables Mn are given as
conditional expectations of the limit M∞ . Therefore, we now prove a generalization of
the Dominated Convergence Theorem that leads to a necessary and sufficient condition
for L1 convergence.

Uniform integrability
Let (Ω, A, P ) be a probability space. The key condition required to deduce L1 conver-
gence from convergence in probability is uniform integrability. To motivate the defini-
tion we first recall two characterizations of integrable random variables:
Lemma 4.10. If X : Ω → R is an integrable random variable on (Ω, A, P ), then
(1). lim E[|X| ; |X| ≥ c] = 0, and

c→∞
(2). for any ε > 0 there exists δ > 0 such that
E[|X| ; A] < ε for any A ∈ A with P [A] < δ.
The second statement says that the positive measure
Q(A) = E[|X| ; A], A ∈ A,
with relative density |X| w.r.t. P is absolutely continuous w.r.t. P in the following
sense: For any ε > 0 there exists δ > 0 such that
P [A] < δ ⇒ Q(A) < ε.
Proof. (1). For an integrable random variable X the first assertion holds by the Mono-
tone Convergence Theorem, since |X| · I{|X|≥c} ց 0 as c ր ∞.
(2). Let ε > 0. By (1),
E[|X| ; A] = E[|X| ; A ∩ {|X| ≥ c}] + E[|X| ; A ∩ {|X| < c}]

≤ E[|X| ; |X| ≥ c] + c · P [A]
ε ε
< + = ε
2 2
provided c ∈ (0, ∞) is chosen appropriately and P [A] < ε/2c.

Uniform integrability means that properties (1) and (2) hold uniformly for a family of
random variables:
Definition (Uniform integrability). A family {Xi : i ∈ I} of random variables on

(Ω, A, P ) is called uniformly integrable if and only if
sup E[|Xi | ; |Xi | ≥ c] −→ 0 as c → ∞.

i∈I
Exercise (Equivalent characterization of uniform integrability). Prove that {Xi :

i ∈ I} is uniformly integrable if and only if sup E[|Xi | ; A] < ∞, and the measures
Qi (A) = E[|Xi | ; A] are uniformly absolutely continuous, i.e., for any ε > 0 there
exists δ > 0 such that
P [A] < δ ⇒ sup E[|Xi | ; A] < ε.

i∈I
We will prove below that convergence in probability plus uniform integrability is equiv-
alent to L1 convergence. Before, we state two lemmas giving sufficient conditions for
uniform integrability (and hence for L1 convergence) that can often be verified in appli-
cations:
Lemma 4.11 (Sufficient conditions for uniform integrability). A family {Xi : i ∈ I}

of random variables is uniformly integrable if one of the following conditions holds:
(1). There exists an integrable random variable Y such that
|Xi | ≤ Y for any i ∈ I.
(2). There exists a measurable function g : R+ → R+ such that

g(x)
lim = ∞ and sup E[g(|Xi |)] < ∞.
x→∞ x i∈I
Proof. (1). If |Xi | ≤ Y then
sup E[|Xi | ; |Xi | ≥ c] ≤ E[Y ; Y ≥ c].

i∈I
The right hand side converges to 0 as c → ∞ if Y is integrable.

(2). The second condition implies uniform integrability, because

y
sup E[|Xi | ; |Xi | ≥ c] ≤ sup · sup E[g(|Xi |)].
i∈I y≥c g(y) i∈I
The first condition in Lemma 4.11 is the classical assumption in the Dominated Con-
vergence Theorem. The second condition holds in particular if
sup E[|Xi |p ] < ∞ for some p > 1 (Lp boundedness),

i∈I
or, if
sup E[|Xi |(log |Xi |)+ ] < ∞ (Entropy condition)
i∈I
is satisfied. Boundedness in L1 , however, does not imply uniform integrability, cf. the
examples at the beginning of this section.
The next observation is crucial for the application of uniform integrability to martin-
gales:
Lemma 4.12 (Conditional expectations are uniformly integrable). If X is an inte-

grable random variable on (Ω, A, P ) then the family
{E[X | F ] : F ⊆ A σ-algebra}
of all conditional expectations of X given sub-σ-algebras of A is uniformly integrable.
Proof. By Lemma 4.10, for any ε > 0 there exists δ > 0 such that
E[|E[X | F ]| ; |E[X | F ]| ≥ c] ≤ E[E[|X| | F ] ; |E[X | F ]| ≥ c] (4.3.1)

= E[|X| ; |E[X | F ]| ≥ c] < ε
holds for c > 0 with P [|E[X | F ]| ≥ c] < δ. Since
1 1
P [|E[X | F ]| ≥ c] ≤ E[|E[X | F ]|] ≤ E[ |X| ],
c c
(4.3.1) holds simultaneously for all σ-algebras F ⊆ A if c is sufficiently large.

Definitive version of Lebesgue’s Dominated Convergence Theorem
Theorem 4.13. Suppose that (Xn )n∈N is a sequence of integrable random variables.
Then (Xn ) converges to a random variable X w.r.t. the L1 norm if and only if Xn
converges to X in probability and the family {Xn : n ∈ N} is uniformly integrable.
Proof. (1). We first prove the “if” part of the assertion under the additional assumption
that the random variables |Xn | are uniformly bounded by a finite constant c: For
ε > 0,
E[ |Xn − X| ] = E[ |Xn − X| ; |Xn − X| > ε] + E[ |Xn − X| ; |Xn − X| ≤ ε]

≤ 2c · P [ |Xn − X| > ε] + ε. (4.3.2)
Here we have used that |Xn | ≤ c and hence |X| ≤ c with probability one, because
a subsequence of (Xn ) converges almost surely to X. For sufficiently large n, the
right hand side of (4.3.2) is smaller than 2ε. Therefore, E[ |Xn − X| ] → 0 as
n → ∞.
(2). To prove the “if” part under the uniform integrability condition, we consider the
cut-off-functions
φc (x) = (x ∧ c) ∨ (−c)
φc
−c c
−c

For c ∈ (0, ∞), the function φc : R → R is a contraction. Therefore,
|φc (Xn ) − φc (X)| ≤ |Xn − X| for any n ∈ N.
If Xn → X in probability then φc (Xn ) → φc (X) in probability. Hence by (1),
E[ |φc (Xn ) − φc (X)| ] −→ 0 for any c > 0. (4.3.3)
We would like to conclude that E[ |Xn − X| ] → 0 as well. Since (Xn ) is

uniformly integrable, and a subsequence converges to X almost surely, we have
E[ |X| ] ≤ lim inf E[ |Xn | ] < ∞ by Fatou’s Lemma. We now estimate
E[ |Xn − X| ]
≤ E[ |Xn − φc (Xn )| ] + E[ |φc (Xn ) − φc (X)| ] + E[ |φc (X) − X| ]
≤ E[ |Xn | ; |Xn | ≥ c] + E[ |φc (Xn ) − φc (X)| ] + E[ |X| ; |X| ≥ c].
Let ε > 0 be given. Choosing c large enough, the first and the last summand on
the right hand side are smaller than ε/3 for all n by uniform integrability of {Xn :
n ∈ N} and integrability of X. Moreover, by (4.3.3), there exists n0 (c) such that
the middle term is smaller than ε/3 for n ≥ n0 (c). Hence E[ |Xn − X| ] < ε for
n ≥ n0 , and thus Xn → X in L1 .
(3). Now suppose conversely that Xn → X in L1 . Then Xn → X in probability by

Markov’s inequality. To prove uniform integrability, we observe that
E[ |Xn | ; A] ≤ E[ |X| ; A] + E[ |X − Xn | ] for any n ∈ N and A ∈ A.
For ε > 0, there exist n0 ∈ N and δ > 0 such that
E[ |X − Xn | ] < ε/2 for any n > n0 , and

E[ |X| ; A] < ε/2 whenever P [A] < δ,
cf. Lemma 4.10. Hence, if P [A] < δ then supn≥n0 E[ |Xn | ; A] < ε.
Moreover, again by Lemma 4.10, there exist δ1 , . . . , δn0 > 0 such that for n ≤ n0 ,
E[ |Xn | ; A] < ε if P [A] < δn .

Choosing δe = min(δ, δ1 , δ2 , . . . , δn0 ), we obtain
sup E[ |Xn | ; A] < ε whenever e

P [A] < δ.
n∈N
Therefore, {Xn : n ∈ N} is uniformly integrable by the exercise below the defi-

nition of uniform integrability on page 143.
L1 convergence of martingales
If X is an integrable random variable and (Fn ) is a filtration then Mn = E[X | Fn ]
is a martingale w.r.t. (Fn ). The next result shows that an arbitrary martingale can be
represented in this way if and only if it is uniformly integrable:
Theorem 4.14 (L1 Martingale Convergence Theorem). Suppose that (Mn )n≥0 is a
martingale w.r.t. a filtration (Fn ). Then the following statements are equivalent:
(1). {Mn : n ≥ 0} is uniformly integrable.
(2). The sequence (Mn ) converges w.r.t. the L1 norm.
(3). There exists an integrable random variable X such that
Mn = E[X | Fn ] for any n ≥ 0.
Proof.
(3) ⇒ (1) holds by Lemma 4.12.
(1) ⇒ (2): If the sequence (Mn ) is uniformly integrable then it is bounded in L1

because
sup E[ |Mn | ] ≤ sup E[ |Mn | ; |Mn | ≥ c] + c ≤ 1 + c

n n

for c ∈ (0, ∞) sufficiently large. Therefore, the limit M∞ = lim Mn exists al-
most surely and in probability by the almost sure convergence theorem. Uniform
integrability then implies
Mn → M∞ in L1
by Theorem 4.13.
(2) ⇒ (3): If Mn converges to a limit M∞ in L1 then
Mn = E[M∞ | Fn ] for any n ≥ 0.
Indeed, Mn is a version of the conditional expectation since it is Fn -measurable

and
E[M∞ ; A] = lim E[Mk ; A] = E[Mn ; A] for any A ∈ Fn (4.3.4)

k→∞
by the martingale property.
A first consequence of the L1 convergence theorem is a limit theorem for conditional

expectations:
Corollary 4.15. If X is an integrable random variable and (Fn ) is a filtration then
E[X | Fn ] → E[X | F∞ ] almost surely and in L1 ,

S
where F∞ := σ( Fn ).
Proof. Let Mn := E[X | Fn ]. By the almost sure and the L1 martingale convergence
theorem, the limit M∞ = lim Mn exists almost surely and in L1 . To obtain a measurable
function that is defined everywhere, we set M∞ := lim sup Mn . It remains to verify, that
M∞ is a version of the conditional expectation E[X | F∞ ]. Clearly, M∞ is measurable
w.r.t. F∞ . Moreover, for n ≥ 0 and A ∈ Fn ,
E[M∞ ; A] = E[Mn ; A] = E[X ; A]

S
by (4.3.4). Since Fn is stable under finite intersections,
E[M∞ ; A] = E[X ; A]
S
holds for all A ∈ σ( Fn ) as well.
Example (Existence of conditional expectations). The common existence proof for

conditional expectations relies either on the Radon-Nikodym Theorem or on the exis-
tence of orthogonal projections onto closed subspaces of the Hilbert space L2 . Martin-
gale convergence can be used to give an alternative existence proof. Suppose that X is
an integrable random variable on a probability space (Ω, A, P ) and F is a separable
sub-σ-algebra of A, i.e., there exists a countable collection (Ai )i∈N of events Ai ∈ A
such that F = σ(Ai : i ∈ N). Let
Fn = σ(A1 , . . . , An ), n ≥ 0.
Note that for each n ≥ 0, there exist finitely many atoms B1 , . . . , Bk ∈ A (i.e. disjoint
S
events with Bi = Ω) such that Fn = σ(B1 , . . . , Bk ). Therefore, the conditional
expectation given Fn can be defined in an elementary way:
X
E[X | Fn ] := E[X | Bi ] · IBi .
i : P [Bi ]6=0
Moreover, by Corollary 4.15, the limit M∞ = lim E[X | Fn ] exists almost surely and in
L1 , and M∞ is a version of the conditional expectation E[X | F ].
You might (and should) object that the proofs of the martingale convergence theorems
require the existence of conditional expectations. Although this is true, it is possible
to state the necessary results by using only elementary conditional expectations, and
thus to obtain a more constructive proof for existence of conditional expectations given
separable σ-algebras.
Another immediate consequence of Corollary 4.15 is an extension of Kolmogorov’s 0-1

law:

Corollary 4.16 (0-1 Law of P.Lévy). If (Fn ) is a filtration on (Ω, A, P ) then for any
S
event A ∈ σ( Fn ),
P [A | Fn ] −→ IA P -almost surely. (4.3.5)
Example (Kolmogorov’s 0-1 Law). Suppose that Fn = σ(A1 , . . . , An ) with indepen-

dent σ-algebras Ai ⊆ A. If A is a tail event, i.e., A is in σ(An+1 , An+2 , . . .) for every
n ∈ N, then A is independent of Fn for any n. Therefore, the corollary implies that
P [A] = IA P -almost surely, i.e.,
P [A] ∈ {0, 1} for any tail event A.
The L1 Martingale Convergence Theorem also implies that any martingale that is Lp
bounded for some p ∈ (1, ∞) converges in Lp :
Exercise (Lp Martingale Convergence Theorem). Let (Mn ) be an (Fn ) martingale

with sup E[ |Mn |p ] < ∞ for some p ∈ (1, ∞).
(1). Prove that (Mn ) converges almost surely and in L1 , and Mn = E[M∞ | Fn ] for
any n ≥ 0.
(2). Conclude that |Mn − M∞ |p is uniformly integrable, and Mn → M∞ in Lp .
Note that uniform integrability of |Mn |p holds automatically and has not to be assumed !
Backward Martingale Convergence

We finally remark that Doob’s upcrossing inequality can also be used to prove that the
conditional expectations E[X | Fn ] of an integrable random variable given a decreasing
T
sequence (Fn ) of σ-algebras converge almost surely to E[X | Fn ]. For the proof one
considers the martingale M−n = E[X | Fn ] indexed by the negative integers:
Exercise (Backward Martingale Convergence Theorem and LLN). Let (Fn )n≥0 be
a decreasing sequence of sub-σ-algebras on a probability space (Ω, A, P ).

(1). Prove that for any random variable X ∈ L1 (Ω, A, P ), the limit M−∞ of the
sequence M−n := E[X | Fn ] as n → −∞ exists almost surely and in L1 , and
\
M−∞ = E[X | Fn ] almost surely.
(2). Now let (Xn ) be a sequence of i.i.d. random variables in L1 (Ω, A, P ), and let
Fn = σ(Sn , Sn+1 , . . .) where Sn = X1 + . . . + Xn . Prove that
Sn
E[X1 | Fn ] = ,
n
and conclude that the strong Law of Large Numbers holds:
Sn
−→ E[X1 ] almost surely.
n

Chapter 5
Stochastic Integration w.r.t.

Continuous Martingales
Suppose that we are interested in a continuous-time scaling limit of a stochastic dynam-

(h)
ics of type X0 = x0 ,
(h) (h) (h)
√
Xk+1 − Xk = σ(Xk ) · h · ηk+1 , k = 0, 1, 2, . . . , (5.0.1)
with i.i.d. random variables ηi ∈ L2 such that E[ηi ] = 0 and Var[ηi ] = 1, a continuous
function σ : R → R, and a scale factor h > 0. Equivalently,
n−1
X
(h)
√ (h)
Xn(h) = X0 + h· σ(Xk ) · ηk+1 , n = 0, 1, 2, . . . . (5.0.2)
k=0
(h)
If σ is constant then as h ց 0, the rescaled process (X⌊t/h⌋ )t≥0 converges in distribution
to (σ · Bt ) where (Bt ) is a Brownian motion. We are interested in the scaling limit for
general σ. One can prove that the rescaled process again converges in distribution, and
the limit process is a solution of a stochastic integral equation
ˆt
Xt = X0 + σ(Xs ) dBs , t ≥ 0. (5.0.3)
0
Here the integral is an Itô stochastic integral w.r.t. a Brownian motion (Bt ). Usually the
equation (5.0.3) is written briefly as
dXt = σ(Xt ) dBt , (5.0.4)
152
153
and interpreted as a stochastic differential equation. Stochastic differential equations

occur more generally when considering scaling limits of appropriately rescaled Markov
chains on Rd with finite second moments. The goal of this section is to give a meaning
to the stochastic integral, and hence to the equations (5.0.3), (5.0.4) respectively.
Example (Stock prices, geometric Brownian motion). A simple discrete time model
for stock prices is given by
Xk+1 − Xk = Xk · ηk+1 , ηi i.i.d.
To set up a corresponding continuous time model we consider the rescaled equation

(5.0.1) as h ց 0. The limit in distribution is a solution of a stochastic differential
equation
dXt = Xt dBt (5.0.5)
w.r.t. a Brownian motion (Bt ). Although with probability one, the sample paths of
Brownian motion are nowhere differentiable, we can give a meaning to this equation by
rewriting it in the form (5.0.3) with an Itô stochastic integral.
A naive guess would be that the solution of (5.0.5) with initial condition X0 = 1 is
Xt = exp Bt . However, more careful considerations show that this can not be true! In
fact, the discrete time approximations satisfy
(h)
√ (h)
Xk+1 = (1 + hηk+1 ) · Xk for k ≥ 0.
(h)
Hence (Xk ) is a product martingale:
n
Y √
Xn(h) = (1 + hηk ) for any n ≥ 0.
k=1
(h)
In particular, E[Xn ] = 1. We would expect similar properties for the scaling limit
(Xt ), but exp Bt is not a martingale and E[exp(Bt )] = exp(t/2).
It turns out that in fact, the unique solution of (5.0.5) with X0 = 1 is not exp(Bt ) but
the exponential martingale
Xt = exp(Bt − t/2),
which is also called a geometric Brownian motion. The reason is that the irregularity of
Brownian paths enforces a correction term in the chain rule for stochastic differentials
leading to Itô’s famous formula, which is the fundament of stochastic calculus.

CHAPTER 5. STOCHASTIC INTEGRATION W.R.T. CONTINUOUS
154 MARTINGALES
5.1 Defining stochastic integrals: A first attempt

Let us first fix some notation that will be used constantly below: By a partition π of
R+ we mean an increasing sequence 0 = t0 < t1 < t2 < . . . such that sup tn = ∞. The
mesh size of the partition is
mesh(π) = sup{|ti − ti−1 | : i ∈ N}.
We are interested in defining integrals of type
ˆt
It = Hs dXs , t ≥ 0, (5.1.1)
0
for continuous functions and, respectively, continuous adapted processes (Hs ) and (Xs ).
For a given t ≥ 0 and a given partition π of R+ , we define the increments of (Xs ) up to

time t by
δXs := Xs′ ∧t − Xs∧t for any s ∈ π,
where s′ := min{u ∈ π : u > s} denotes the next partition point after s. Note that
the increments δXs vanish for s ≥ t. In particular, only finitely many of the increments
are not equal to zero. A nearby approach for defining the integral It in (5.1.1) would be
Riemann sum approximations:
Riemann sum approximations

There are various possibilities to define approximating Riemann sums w.r.t. a given
sequence (πn ) of partitions with mesh(πn ) → 0, for example:
P
Variant 1 (non-anticipative): Itn = Hs δXs ,
s∈πn
P
Variant 2 (anticipative): Iˆtn = Hs′ δXs ,
s∈πn
◦ P 1
Variant 3 (anticipative): Itn = 2
(Hs + Hs′ )δXs .
s∈πn

5.1. DEFINING STOCHASTIC INTEGRALS: A FIRST ATTEMPT 155
Note that for finite t, in each of the sums, only finitely many summands do not vanish.
For example,
X X
Itn = Hs δXs = Hs · (Xs′ ∧t − Xs ).
s∈πn s∈πn
s<t s<t
Now let us consider at first the case where Hs = Xs and t = 1, i.e., we would like to
´1
define the integral I = Xs dXs . Suppose first that X : [0, 1] → R is a continuous
0
function of finite variation, i.e.,
( )
X
V (1) (X) = sup |δXs | : π partition of R+ < ∞.
s∈π
Then for H = X and t = 1 all the approximations above converge to the same limit as
n → ∞. For example,
X
kIˆ1n − I1n k = (δXs )2 ≤ V (1) (X) · sup |δXs |,
s∈πn
s∈πn
and the right-hand side converges to 0 by uniform continuity of X on [0, 1]. In this case
the limit of the Riemann sums is a Riemann-Stieltjes integral
ˆ1
lim I1n = lim Iˆ1n = Xs dXs ,
n→∞ n→∞
0
which is well-defined whenever the integrand is continuous and the integrator is of finite
variation or conversely. The sample paths of Brownian motion, however, are almost
surely not of finite variation. Therefore, the reasoning above does not apply, and in fact
if Xt = Bt is a one-dimensional Brownian motion and Ht = Xt then
X X
E[ |Iˆ1n − I1n | ] = E[(δBs )2 ] = δs = 1,
s∈πn s∈πn
i.e., the L1 -limits of the random sequence (I1n ) and (Iˆ1n ) are different if they exist. Below
◦
we will see that indeed the limits of the sequences (I1n ), (Iˆ1n ) and (I1n ) do exist in L2 ,
and all the limits are different. The limit of the non-anticipative Riemann sums I1n
´1
is the Itô stochastic integral 0 Bs dBs , the limit of (Iˆ1n ) is the backward Itô integral
´1
ˆ s , and the limit of I ◦ is the Stratonovich integral 1 Bs ◦ dBs . All three notions
´
0
Bs dB n 0

156 MARTINGALES
of stochastic integrals are relevant. The most important one is the Itô integral because
´t
the non-anticipating Riemann sum approximations imply that the Itô integral 0 Hs dBs
is a continuous time martingale transform of Brownian motion if the process (Hs ) is
adapted.
Itô integrals for continuous bounded integrands

We now give a first existence proof for Itô integrals w.r.t. Brownian motion. We start
with a provisional definition that will be made more precise later:
Preliminary Definition. For continuous functions or continuous stochastic processes

(Hs ) and (Xs ) and a given sequence (πn ) of partitions with mesh(πn ) → 0, the Itô
integral of H w.r.t. X is defined by
ˆt X
Hs dXs = lim Hs δXs
n→∞
0 s∈πn
whenever the limit exists in a sense to be specified.
Note that the definition is vague since the mode of convergence is not specified. More-
over, the Itô integral might depend on the sequence (πn ). In the following sections we
will see which kind of convergence holds in different circumstances, and in which sense
the limit is independent of (πn ).
To get started let us consider the convergence of Riemann sum approximations for the
´t
Itô integrals Hs dBs of a bounded continuous (Fs ) adapted process (Hs )s≥0 w.r.t. an
0
(Fs ) Brownian motion (Bs ). Let (πn ) be a fixed sequence of partitions with πn ⊆ πn+1
and mesh(πn ) → 0. Then for the Riemann-Itô sums
X X
Itn = Hs δBs = Hs (Bs′ ∧t − Bs )
s∈πn s∈πn
s<t
we have
X
Itn − Itm = (Hs − H⌊s⌋m ) δBs for any m ≤ n,
s∈πn
s<t

where ⌊s⌋m = max{r ∈ πm : r ≤ s} denotes the next partition point in πm below

s. Since Brownian motion is a martingale, we have E[δBs | Fs ] = 0 for any s ∈ πn .
Moreover, E[(δBs )2 | Fs ] = δs. Therefore, we obtain by conditioning on Fs , Fr
respectively:
XX
E[(Itn − Itm )2 ] = E[(Hs − H⌊s⌋m )(Hr − H⌊r⌋m )δBs δBr ]
s∈πn r∈πn
s<t r<t
X X
= E[(Hs − H⌊s⌋m )2 δs] ≤ E[Vm ] · δs = E[Vm ] · t,
s∈πn s∈πn
s<t s<t
where
Vm := sup (Hs − Hr )2 −→ 0 as m → ∞
|s−r|≤mesh(πm )
by uniform continuity of (Hs ) on [0, t]. Since H is bounded, E[Vm ] → 0 as m → ∞,

and hence (Itn ) is a Cauchy sequence in L2 (Ω, A, P ) for any given t ≥ 0. Thus we
obtain:
Theorem 5.1 (Itô integrals for bounded continuous integrands, Variant 1). Suppose
that (Hs )s≥0 is a bounded continuous (Fs ) adapted process, and (Bs )s≥0 is an (Fs )
Brownian motion. Then for any fixed t ≥ 0, the Itô integral
ˆt
Hs dBs = lim Itn (5.1.2)
n→∞
0
exists as a limit in L2 (Ω, A, P ). Moreover, the limit does not depend on the choice of a
sequence of partitions (πn ) with mesh (πn ) → 0.
Proof. An analogue argument as above shows that for any partitions π and π
e such that
e, the L2 distance of the corresponding Riemann sum approximations Itπ and Itπe is
π⊇π
bounded by a constant C(mesh(e
π )) that only depends on the maximal mesh size of the
two partitions. Moreover, the constant goes to 0 as the mesh sizes go to 0. By choosing
a joint refinement and applying the triangle inequality, we see that
kItπ − Itπe kL2 (P ) ≤ 2 C(∆)

158 MARTINGALES
holds for arbitrary partitions π, π

e such that max(mesh(π)), mesh(e
π )) ≤ ∆. The asser-
tion now follows by completeness of L2 (P ).
The definition of the Itô integral suggested by Theorem 5.1 has two obvious drawbacks:
´t
Drawback 1: The integral 0
Hs dBs is only defined as an equivalence class in L2 (Ω, A, P ),
i.e., uniquely up to modification on P -measure zero sets. In particular, we do not have a
´t
pathwise definition of 0 Hs (ω) dBs (ω) for a given Brownian sample path s 7→ Bs (ω).
Drawback 2: Even worse, the construction above works only for a fixed integra-
tion interval [0, t]. The exceptional sets may depend on t and therefore, the process
´t
t 7→ 0 Hs dBs does not have a meaning yet. In particular, we do not know yet if there
exists a version of this process that is almost surely continuous.
The first drawback is essential: In certain cases it is indeed possible to define stochastic
integrals pathwise, cf. Chapter 6 below. In general, however, pathwise stochastic inte-
grals cannot be defined. The extra impact needed is the Lévy area process, cf. the rough
paths theory developed by T. Lyons and others [XXXLyons, Friz and Victoir, Friz and
Hairer].
Fortunately, the second drawback can be overcome easily. By extending the Itô isom-
etry to an isometry into the space Mc2 of continuous L2 bounded martingales, we can
´t
construct the complete process t 7→ 0 Hs dBs simultaneously as a continuous martin-
gale. The key observation is that by the maximal inequality, continuous L2 bounded
martingales can be controlled uniformly in t by the L2 norm of their final value.
The Hilbert space M2c

Fix u ∈ (0, ∞] and suppose that for t ∈ [0, u], (Itn ) is a sequence of Riemann sum
´t
approximations for 0 Hs dBs as considered above. It is not difficult to check that for
each fixed n ∈ N, the stochastic process t 7→ Itn is a continuous martingale. Our aim is
to prove convergence of these continuous martingales to a further continuous martingale
´t
It = 0 Hs dBs . Since the convergence holds only almost surely, the limit process

will not necessarily be (Ft ) adapted. To ensure adaptedness, we have to consider the
completed filtration
FtP = {A ∈ A : P [A △ B] = 0 for some B ∈ Ft }, t ≥ 0,
where A △ B = (A \ B) ∪ (B \ A) is the symmetric difference of the sets A and B.

Note that the conditional expectations given Ft and FtP agree P -almost surely. Hence,
if (Bt ) is a Brownian motion resp. a martingale w.r.t. the filtration (Ft ) then it is also a
Brownian motion or a martingale w.r.t. (FtP ).
Let M2 ([0, u]) denote the space of all L2 -bounded (FtP ) martingales (Mt )0≤t≤u on
(Ω, A, P ). By M2c ([0, u]) and M2d ([0, u]) we denote the subspaces consisting of all
continuous (respectively right continuous) martingales M ∈ M2 ([0, u]). Recall that
by the L2 martingale convergence theorem, any (right) continuous L2 -bounded martin-
gale (Mt ) defined for t ∈ [0, u) can be extended to a (right) continuous martingale in
M2 ([0, u]).
f ∈ M2 ([0, u]) are called modifications of each other if
Two martingales M, M
ft ] = 1
P [Mt = M for any t ∈ [0, u].
If the martingales are right-continuous then two modifications agree almost surely, i.e.,
ft ∀t ∈ [0, u]] = 1.
P [Mt = M
In order to obtain norms and not just semi-norms, we consider the spaces
M 2 ([0, u]) := M2 ([0, u])/ ∼ and Mc2 ([0, u]) := M2c ([0, u])/ ∼
of equivalence classes of martingales that are modifications of each other. We will

frequently identify equivalence classes and their representatives.
We endow the space M 2 ([0, u]) with the inner product
(M, N)M 2 ([0,u]) = (Mu , Nu )L2 = E[Mu Nu ].
As the process (Mt2 ) is a submartingale for any M ∈ M 2 ([0, u]), the norm correspond-
ing to the inner product is given by
kMk2M 2 ([0,u]) = E[Mu2 ] = sup E[Mt2 ].

0≤t≤u

160 MARTINGALES
Moreover, if (Mt ) is right continuous then by Doob’s L2 -maximal inequality,

sup |Mt | ≤ 2 · sup kMt kL2 (Ω,A,P ) = 2kMkM 2 ([0,u]) . (5.1.3)
0≤t≤u
0≤t≤u
L2 (Ω,A,P )
This crucial estimate shows that on the subspaces Mc2 and Md2 , the M 2 norm is equiva-
lent to the L2 norm of the supremum of the martingale. Therefore, the M 2 norm can be
used to control (right) continuous martingales uniformly in t!
Lemma 5.2 (Completeness). (1). The space M 2 ([0, u]) is a Hilbert space, and the
linear map M 7→ Mu from M 2 ([0, u]) to L2 (Ω, Fu , P ) is onto and isometric.
(2). The spaces Mc2 ([0, u]) and Md2 ([0, u]) are closed subspaces of M 2 ([0, u]), i.e.,
if (M n ) is a Cauchy sequence in Mc2 ([0, u]), or in Md2 ([0, u]) respectively, then
there exists a (right) continuous martingale M ∈ M 2 ([0, u]) such that
sup |Mtn − Mt | −→ 0 in L2 (Ω, A, P ).

t∈[0,u]
Proof. (1). By definition of the inner product on M 2 ([0, u]), the map M 7→ Mu is
an isometry. Moreover, for X ∈ L2 (Ω, Fu , P ), the process Mt = E[X | Fu ]
is in M 2 ([0, u]) with Mu = X. Hence, the range of the isometry is the whole
space L2 (Ω, Fu , P ). Since L2 (Ω, Fu , P ) is complete w.r.t. the L2 norm, the space
M 2 ([0, u]) is complete w.r.t. the M 2 norm.
(2). If (M n ) is a Cauchy sequence in Mc2 ([0, u]) or in Md2 ([0, u]) respectively, then by
(5.1.3),
kM n − M m ksup := sup |Mtn − Mtm | −→ 0 in L2 (Ω, A, P ).

0≤t≤u
In particular, we can choose a subsequence (M nk ) such that
P [ kM nk+1 − M nk ksup ≥ 2−k ] ≤ 2−k for all k ∈ N.
Hence, by the Borel-Cantelli Lemma,
P [ kM nk+1 − M nk ksup < 2−k eventually ] = 1,

and therefore Mtnk converges almost surely uniformly in t as k → ∞. The limit

of the sequence (M n ) in M 2 ([0, u]) exists by (1), and the process M defined by

lim M nk if (M nk ) converges uniformly,
t
Mt := (5.1.4)
0 otherwise,
is a continuous (respectively right continuous) representative of the limit. Indeed,
by Fatou’s Lemma,
kM nk − Mk2M 2 ([0,u]) ≤ E[ kM nk − Mk2sup ] = E[ lim kM nk − M nl k2sup ]

l→∞
≤ lim inf E[ kM nk − M nl k2sup ],
l→∞
and the right hand side converges to 0 as k → ∞. Finally, M is a martingale w.r.t.

(FtP ), and hence an element in Mc2 ([0, u]) or in Md2 ([0, u]) respectively.
Remark. We point out that the (right) continuous representative (Mt ) defined by (5.1.4)
is a martingale w.r.t. the complete filtration (FtP ), but it is not necessarily adapted w.r.t.
(Ft ).
Definition of Itô integral in M2c

Let u ∈ R+ . For any bounded continuous (Ft ) adapted process (Ht ) and any sequence
(πn ) of partitions of R+ , the processes
X
Itn = Hs (Bs′ ∧t − Bs∧t ), t ∈ [0, u],
s∈πn
are continuous L bounded martingales on [0, u]. We can therefore restate Theorem 5.1
2
in the following way:
Corollary 5.3 (Itô integrals for bounded continuous integrands, Variant 2). Suppose
that (Hs )s≥0 is a bounded continuous (Fs ) adapted process. Then for any fixed u ≥ 0,
the Itô integral
ˆ•
Hs dBs = lim (Itn )t∈[0,u] (5.1.5)
n→∞
0

162 MARTINGALES
exists as a limit in Mc2 ([0, u]). Moreover, the limit does not depend on the choice of a
sequence of partitions (πn ) with mesh (πn ) → 0.
Proof. The assertion is an immediate consequence of the definition of the M 2 norm,

Theorem 5.1 and Lemma 5.2.
Similar arguments as above apply if Brownian motion is replaced by a bounded mar-

tingale with continuous sample paths. In the rest of this chapter we will work out the
construction of the Itô integral w.r.t. Brownian motion and more general continuous
martingales more systematically and for a broader class of integrands.
5.2 Itô’s isometry

Let (Mt )t≥0 be a continuous (or, more generally, right continuous) martingale w.r.t. a
filtration (Ft ) on a probability space (Ω, A, P ). We now develop a more systematic
´t
approach for defining stochastic integrals 0 Hs dMs of adapted processes (Ht ) w.r.t.
(Mt ).
Predictable step functions
In a first step, we define the integrals for predictable step functions (Ht ) of type
n−1
X
Ht (ω) = Ai (ω)I(ti ,ti+1 ] (t)
i=0
with n ∈ N, 0 ≤ t0 < t1 < t2 < . . . < tn , and bounded Fti -measurable random vari-
ables Ai , i = 0, 1, . . . , n − 1. Let E denote the vector space consisting of all stochastic
processes of this form.

5.2. ITÔ’S ISOMETRY 163
Definition (Itô integral for predictable step functions). For stochastic processes H ∈
E and t ≥ 0 we define
ˆt n−1
X X
Hs dMs := Ai · (Mti+1 ∧t − Mti ∧t ) = Ai · (Mti+1 ∧t − Mti ).
0 i=0 i : ti <t
The stochastic process H• M given by

ˆt
(H• M)t := Hs dMs for t ∈ [0, ∞]
0
is called the Itô integral of H w.r.t. M.
Note that the map (H, M) 7→ H• M is bilinear. The process H• M is a continuous time
martingale transform of M w.r.t. H. It models for example the net gain up to time t
if we hold Ai units of an asset with price process (Mt ) during each of the time intervals
(ti , ti+1 ].
Lemma 5.4. For any H ∈ E , the process H• M is a continuous (Ft ) martingale up to

time t = ∞.
Similarly to the discrete time case, the fact that Ai is predictable, i.e., Fti -measurable,
is essential for the martingale property:
Proof. By definition, H• M is continuous and (Ft ) adapted. It remains to verify that
E[(H• M)t | Fs ] = (H• M)s for any 0 ≤ s ≤ t. (5.2.1)
We do this in three steps:
(1). At first we note that (5.2.1) holds for s, t ∈ {t0 , t1 , . . . , tn }. Indeed, since Ai is
Fti -measurable, the process
j−1
X
(H• M)tj = Ai · (Mti+1 − Mti ), j = 0, 1, . . . , n,
i=0

164 MARTINGALES
is a martingale transform of the discrete time martingale (Mti ), and hence again
a martingale.
(2). Secondly, suppose s, t ∈ [tj , tj+1 ] for some j ∈ {0, 1, 2, . . . , n − 1}. Then
E[(H• M)t −(H• M)s | Fs ] = E[Aj · (Mt −Ms ) | Fs ] = Aj · E[Mt −Ms | Fs ] = 0
because Aj is Ftj -measurable and hence Fs -measurable, and (Mt ) is a martingale.
(3). Finally, suppose that s ∈ [tj , tj+1 ] and t ∈ [tk , tk+1] with j < k.
tj s tj+1 tk t tk+1
Then by the tower property for conditional expectations and by (1) and (2),
E[(H• M)t | Fs ] = E[E[E[(H• M)t | Ftk ] | Ftj+1 ] | Fs ]

(2) (1)
= E[E[(H• M)tk | Ftj+1 ] | Fs ] = E[(H• M)tj+1 | Fs ]
(2)
= (H• M)s .
Remark (Riemann sum approximations). Non-anticipative Riemann sum approxi-

mations of stochastic integrals are Itô integrals of predictable step functions: If (Ht ) is
an adapted stochastic process and π = {t0 , t1 , . . . , tn } is a partition then
n−1
X ˆt
Hti · (Mti+1 ∧t − Mti ∧t ) = Hsπ dMs (5.2.2)
i=0 0
P
n−1
where H π := Hti · I(ti ,ti+1 ] is a process in E .
i=0

Itô’s isometry for Brownian motion

Recall that our goal is to prove that non-anticipative Riemann sum approximations
for a stochastic integral converge. Let (πn ) be a sequence of partitions of [0, t] with
mesh(πn ) → 0. By the remark above, the corresponding Riemann-Itô sums I πn defined
by (5.2.2) are integrals of predictable step functions H πn . Hence in order to prove that
the sequence (I πn ) converges in the Hilbert space Mc2 it suffices to show that
(1). (H πn ) is a Cauchy sequence w.r.t. an appropriate norm on the vector space E ,

and
(2). the “Itô map” J : E → Mc2 defined by
ˆ•
J (H) = H• M = Hs dMs
0
is continuous w.r.t. this norm.
It turns out that we can even identify explicitly a simple norm on E such that the Itô
map is an isometry. We first consider the case where (Mt ) is a Brownian motion:
Theorem 5.5 (Itô’s isometry for Brownian motion). If (Bt ) is an (Ft ) Brownian mo-
tion on (Ω, A, P ) then for any u ∈ [0, ∞], and for any process H ∈ E ,
 u 2   u 
ˆ ˆ
kH• Bk2M 2 ([0,u]) = E  Hs dBs   = E  Hs2 ds = kHk2L2 (P ⊗λ(0,u) ) (5.2.3)
0 0
Pn−1
Proof. Suppose that H = i=0 Ai · I(ti ,ti+1 ] with n ∈ N, 0 ≤ t0 < t1 < . . . < tn and
Ai bounded and Fti -measurable. With the notation δi B := Bti+1 ∧u − Bti ∧u , we obtain
"ˆ  !2 
u 2 # n−1
X n−1
X
E Hs dBs = E Ai δi B  = E [Ai Ak · δi Bδk B] . (5.2.4)
0 i=0 i,k=0

166 MARTINGALES
By the martingale property, the summands on the right hand side vanish for i 6= k.
Indeed, if, for instance, i < k then
E[Ai Ak δi Bδk B] = E[Ai Ak δi B · E[δk B | Ftk ]] = 0.
Here we have used in an essential way, that Ak is Ftk -measurable. Similarly,
E[A2i · (δi B)2 ] = E[A2i E[(δi B)2 | Fti ]] = E[A2i · δi t]
by the independence of the increments of Brownian motion. Therefore, by (5.2.4) we

obtain
"ˆ 2 # n−1 ˆ
u X u
E Hs dBs = E[A2i · (ti+1 ∧ u − ti ∧ u)] = E Hs2 ds .
0 i=0 0
The assertion now follows by definition of the M 2 norm.
Theorem 5.5 shows that the linear map

ˆ r
J :E → M2c ([0, u]), J (H) = Hs dBs ,
0 r∈[0,u]
is an isometry if the space E of simple predictable processes (s, ω) 7→ Hs (ω) is en-

dowed with the L2 norm
ˆ u 1/2
kHkL2 (P ⊗λ(0,u) ) = E Hs2 ds
0
on the product space Ω × (0, u). In particular, J respects P ⊗ λ classes, i.e., if Hs (ω) =
e s (ω) for P ⊗ λ-almost every (ω, s) then • H dB = • H e dB P -almost surely. Hence
´ ´
H 0 0
J also induces a linear map between the corresponding spaces of equivalence classes.
As usual, we do not always differentiate between equivalence classes and functions, and
so we denote the linear map on equivalence classes again by J :
J : E ⊂ L2 (P ⊗ λ(0,u) ) → Mc2 ([0, u]),

kJ (H)kM 2([0,u]) = kHkL2 (P ⊗λ(0,u) ) . (5.2.5)

Itô’s isometry for martingales

An Itô isometry also holds if Brownian motion is replaced by a continuous square-
integrable martingale (Mt ). More generally, suppose that (Mt )t≥0 is a right continuous
square integrable (Ft ) martingale satisfying the following assumption:
Assumption A. There exists a non-decreasing adapted continuous process t 7→ hMit

such that hMi0 = 0 and Mt2 − hMit is a martingale.
For continuous square integrable martingales, the assumption is always satisfied. In-
deed, assuming continuity, the “angle bracket process” hMit coincides almost surely
with the quadratic variation process [M]t of M, cf. Section 6.3 below. For Brownian
motion, Assumption A holds with
hBit = t.
Note that for any 0 ≤ s ≤ t, Assumption A implies

E (Mt − Ms )2 | Fs = E Mt2 − Ms2 | Fs = E [hMit − hMis | Fs] . (5.2.6)
Since t 7→ hMit (ω) is continuous and non-decreasing for a given ω, it is the distribution
function of a unique positive measure hMi(ω, dt) on R+ .
Theorem 5.6 (Itô’s isometry for martingales). Suppose that (Mt )t≥0 is a right con-
tinuous (Ft ) martingale with angle bracket process hMi satisfying Assumption A. Then
for any u ∈ [0, ∞], and for any process H ∈ E ,
"ˆ 2 # ˆ
u u
2
kH• MkM 2 ([0,u]) = E Hs dMs = E Hs2 dhMis (5.2.7)
0 0
where dhMi denotes integration w.r.t. the positive measure with distribution function
F (t) = hMit .
For Brownian motion hBit = t, so (5.2.7) reduces to (5.2.3).

168 MARTINGALES
Proof. The proof is similar to the proof of Theorem 5.5 above. Suppose again that
Pn−1
H = i=0 Ai · I(ti ,ti+1 ] with n ∈ N, 0 ≤ t0 < t1 < . . . < tn and Ai bounded and Fti -
measurable. With the same notation as in the proof above, we obtain by the martingale
properties of M and M 2 − hMi,
E[Ai Ak δi M δk M] = 0 for i 6= k, and
E[A2i · (δi M)2 ] = E[A2i E[(δi M)2 | Fti ]] = E[A2i E[δi hMi | Fti ]] = E[A2i · δi hMi].
cf. (5.2.6). Therefore,

"ˆ  !2 
u 2 # n−1
X n−1
X
E Hs dMs = E Ai δi M  = E [Ai Ak δi M δk M]
0 i=0 i,k=0
n−1
X ˆ u
= E[A2i δi hMi] = E Hs2 dhMis .
i=0 0
For a continuous square integrable martingale, Theorem 5.6 implies that the linear map
ˆ r
J :E → M2c ([0, u]), J (H) = Hs dMs ,
0 r∈[0,u]
is an isometry if the space E of simple predictable processes (s, ω) 7→ Hs (ω) is en-

dowed with the L2 norm
ˆ u 1/2
kHkL2(Ω×(0,u),PhM i ) = E Hs2 dhMis
0
on the product space Ω × (0, u) endowed with the positive measure
PhM i (dω dt) = P (dω) hMi(ω, dt). (5.2.8)
Again, we denote the corresponding linear map induced on equivalence classes by the
same letter J .

Definition of Itô integrals for square-integrable integrands

From now on we assume that (Mt ) is a continuous square integrable (Ft ) martingale
with angle bracket process hMit . We fix u ∈ [0, ∞] and consider the isometry
J : E ⊂ L2 (Ω × (0, u), PhM i) → Mc2 ([0, u]), . (5.2.9)

H 7→ H• M
mapping an elementary predictable process H to the continuous martingale

ˆ t
(H• M)t = Hs dMs .
0
More precisely, we consider the induced map on equivalence classes.
Let Eu denote the closure of the space E in L2 (Ω × (0, u), PhM i ). Since J is linear with
kJ (H)kM 2 ([0,u]) = kHkL2 (Ω×(0,u),PhM i ) for any H ∈ E,
there is a unique extension to a continuous (and even isometric) linear map
J : Eu ⊆ L2 (Ω × (0, u), PhM i ) → Mc2 ([0, u]).
This can be used to define the Itô integral for any process in Eu , i.e., for any process that
can be approximated by predictable step functions w.r.t. the L2 (PhM i ) norm:
ˆ t
H• B := J (H), Hs dBs := (H• B)t .
0
Explicitly, we obtain the following definition of stochastic integrals for integrands in Eu :
´t
Definition (Itô integral). For H ∈ Eu the process H• M = ( 0
Hs dMs )t∈[0,u] is the up
to modifications unique continuous martingale on [0, u] satisfying
(H• M)t = lim (H•n M)t in L2 (P ) for any t ∈ [0, u]

n→∞
whenever (H n ) is a sequence of elementary predictable processes such that H n → H

in L2 (Ω × (0, u), PhM i).

170 MARTINGALES
Remark. (1). By construction, the map H 7→ H• M is an isometry from Eu endowed

with the L2 (PhM i ) norm to Mc2 ([0, u]). If t 7→ hMit is absolutely continuous, then
the closure Eu of the elementary processes actually contains any (FtP ) adapted
process (ω, t) 7→ Ht (ω) that is square-integrable w.r.t. PhM i , see XXX below.
(2). The definition above is consistent in the following sense: If H• M is the stochastic
integral defined on the time interval [0, v] and u ≤ v, then the restriction of H• M
to [0, u] coincides with the stochastic integral on [0, u].
For 0 ≤ s ≤ t we define
ˆ t
Hr dMr := (H• M)t − (H• M)s .
s
Exercise. Verify that for any H ∈ Et ,

ˆ t ˆ t ˆ t ˆ t
Hr dMr = Hr dBr − I(0,s) (r)Hr dMr = I(s,t) (r)Hr dMr .
s 0 0 0
Having defined the Itô integral, we now show that bounded adapted processes with
left-continuous sample paths are contained in the closure of the simple predictable pro-
cesses, and the corresponding stochastic integrals are limits of predictable Riemann sum
approximations. As above, we consider a sequence (πn ) of partitions of R+ such that
mesh(πn ) → 0.
Theorem 5.7 (Approximation by Riemann-Itô sums). Let u ∈ (0, ∞), and suppose
that (Ht )t∈[0,u) is an (FtP ) adapted stochastic process on (Ω, A, P ) such that (t, ω) 7→
Ht (ω) is product-measurable and bounded. If t 7→ Ht is P -almost surely left continuous
then H is in Eu , and
ˆ t X
Hs dMs = lim Hs (Ms′ ∧t − Ms∧t ), t ∈ [0, u], (5.2.10)
0 n→∞
s∈πn
w.r.t. convergence uniformly in t in the L2 (P ) sense.

Remark. (1). In particular, a subsequence of the predictable Riemann sum approxi-

mations converges uniformly in t with probability one.
(2). The assertion also holds if H is unbounded with sups≤u |Hs | ∈ L2 (P ).
Proof. For any t ∈ [0, u], the Riemann sums on the right hand side of (5.2.10) are the
´t
stochastic integrals 0 Hsn dMs of the predictable step functions
X
Htn := Hs · I(s,s′] (t), n ∈ N.
s∈πn ,s<u
By left-continuity, Htn → Ht as n → ∞ for any t ∈ [0, u], P -almost surely. Therefore,

H n → H PhM i -almost surely, and, by dominated convergence,
Hn → H in L2 (PhM i ).
Here we have used that the sequence (H n ) is uniformly bounded since H is bounded by
assumption. Now, by Itô’s isometry,
ˆ • ˆ •
Hs dMs = lim Hsn dMs in Mc2 ([0, u]).
0 n→∞ 0
Identification of admissible integrands

Let u ∈ (0, ∞]. We have already shown that if u < ∞ then any product-measurable
adapted bounded process with left-continuous sample paths is in Eu . More generally,
we will prove now that if Mt = Bt is a Brownian motion then any adapted process
in L2 (P ⊗ λ[0,u)) is contained in Eu , and hence “integrable” w.r.t. (Bt ). Let L2a (0, u)
denote the linear space of all product-measurable, (FtP ) adapted stochastic processes
(ω, t) 7→ Ht (ω) defined on Ω × (0, u) such that
ˆ u
2
E Ht dt < ∞.
0
The corresponding space of equivalence classes of P ⊗λ versions is denoted by L2a (0, u).
Lemma 5.8. L2a (0, u) is a closed linear subspace of L2 (P ⊗ λ(0,u) ).

172 MARTINGALES
Proof. It only remains to show that an L2 (P ⊗ λ) limit of (FtP ) adapted processes again
has an (FtP ) adapted P ⊗ λ-version. Hence consider a sequence H n ∈ L2a (0, u) with
H n → H in L2 (P ⊗ λ). Then there exists a subsequence (H nk ) such that Htnk (ω) →
e defined by H
Ht (ω) for P ⊗ λ-almost every (ω, t) ∈ Ω × (0, u). The process H e t (ω) :=
e t (ω) := 0 otherwise, is an (F P ) adapted version of
lim Htnk (ω) if the limit exists, H t
H.
We can now identify the class of integrands H for which the stochastic integral H• B is
well-defined as a limit of integrals of predictable step functions in Mc2 ([0, u]):
Theorem 5.9 (Admissible integrands for Brownian motion). For any u ∈ (0, ∞],
Eu = L2a (0, u).
Proof. Since E ⊆ L2a (0, u) it only remains to show the inclusion “⊇”. Hence fix a
process H ∈ L2a (0, u). We will prove in several steps that H can be approximated by
simple predictable processes w.r.t. the L2 (P ⊗ λ(0,u) ) norm:
(1). Suppose first that H is bounded and has almost surely continuous trajectories.
Then for u < ∞, H is in Eu by Theorem 5.7. For u = ∞, H is still in Eu
provided there exists t0 ∈ (0, ∞) such that Ht vanishes for t ≥ t0 .
(2). Now suppose that (Ht ) is bounded and, if u = ∞, vanishes for t ≥ t0 . To

prove H ∈ Eu we approximate H by continuous adapted processes. To this end
let ψn : R → [0, ∞), n ∈ N, be continuous functions such that ψ(s) = 0 for
´∞
s∈/ (0, 1/n) and −∞ ψn (s) ds = 1. Let H n := H ∗ ψn , i.e.,
ˆ 1/n
Htn (ω) = Ht−ε (ω)ψn (ε) dε, (5.2.11)
0
where we set Ht := 0 for t ≤ 0. We prove that
(a) H n → H in L2 (P ⊗ λ(0,u) ), and

(b) H n ∈ Eu for any n ∈ N.

Combining (a) and (b), we see that H is in Eu as well.
(a) Since H is in L2 (P ⊗ λ(0,u) ), we have

ˆ u
Ht (ω)2 dt < ∞ (5.2.12)
0
for P -almost every ω. It is a standard fact from analysis that (5.2.12) implies
ˆ u
|Htn (ω) − Ht (ω)|2 dt −→ 0 as n → ∞.
0
By dominated convergence, we obtain

ˆ u
n 2
E |Ht − Ht | dt −→ 0 as n → ∞ (5.2.13)
0
because H is bounded, the sequence (Hn ) is uniformly bounded, and H and

H n vanish for t ≥ t0 + 1.
(b) This is essentially a consequence of part (1) of the proof. We sketch how to
verify that H n satisfies the assumptions made there:
• The sample paths t 7→ Htn (ω) are continuous for all ω.
• |Htn | is bounded by sup |H|.
• The map (ω, t) 7→ Htn (ω) is product measurable by (5.2.11) and Fu-
bini’s Theorem, because the map (ω, t, ε) 7→ Ht−ε (ω)ψε (ω) is product
measurable.
• If the process (Ht ) is progressively measurable, i.e., if the map (s, ω) 7→
Hs (ω) (s ∈ (0, t), ω ∈ Ω) is measurable w.r.t. the product σ-algebra
B(0, t) ⊗ FtP for any t ≥ 0, then (Htn ) is (FtP ) adapted by (5.2.11)
and Fubini’s Theorem. This is for example the case if (Ht ) is right
continuous or left continuous.
• In general, one can prove that (Ht ) has a progressively measurable mod-
ification, whence (Htn ) has an (FtP ) adapted modification. We omit the
details.
(3). We finally prove that general H ∈ L2a (0, u) are contained in Eu . This is a conse-
quence of (2), because we can approximate H by the processes
Htn := ((Ht ∧ n) ∨ (−n)) · I(0,n) (t), n ∈ N.

174 MARTINGALES
These processes are bounded, they vanish for t ≥ n, and H n → H in L2 (P ⊗

λ(0,u) ). By (2), H n is contained in Eu for any n, so H is in Eu as well.
Remark (Riemann sum approximations). For discontinuous integrands, the predict-

able Riemann sum approximations considered above do not converge to the stochastic
integral in general. However, one can prove that for u < ∞ any process H ∈ L2a (0, u)
is the limit of the simple predictable processes
n −1
2X ˆ i2−n u
Htn = 2 n
Hs ds · I(i2−n u,(i+1)2−n u] (t)
i=1 (i−1)2−n u
w.r.t. the L2 (P ⊗ λ[0,u)) norm, cf. [XXXSteele: “Stochastic calculus and financial ap-
´t
plications”, Sect 6.6]. Therefore, the stochastic integral 0 H dB can be approximated
for t ≤ u by the correspondingly modified Riemann sums.
For continuous martingales, a similar statement as in Theorem 5.9 holds provided the
angle bracket process is absolutely continuous. Let L2a (0, u; M) denote the linear space
of all product-measurable, (FtP ) adapted stochastic processes (ω, t) 7→ Ht (ω) such that
ˆ u
2
E Ht dhMit < ∞.
0
The corresponding space of equivalence classes w.r.t. PhM i is denoted by L2a (0, u; M).
Exercise (Admissible integrands w.r.t. martingales). Suppose that (Mt ) is a contin-

uous square integrable (Ft ) martingale. Show that if almost surely, t 7→ hMit is abso-
lutely continuous, then the closure Eu of the elementary processes w.r.t. the L2 (PhM i )
norm is given by
Eu = L2a (0, u; M).
5.3 Localization
Square-integrability of the integrand is still an assumption that we would like to avoid,
since it is not always easy to verify or may even fail to hold. The key to extending

5.3. LOCALIZATION 175
the class of admissible integrands further is localization, which enables us to define a

stochastic integral w.r.t. a continuous martingale for any continuous adapted process.
The price we have to pay is that for integrands that are not square integrable, the Itô
integral is in general not a martingale, but only a local martingale.
Throughout this section we assume that Mt is a continuous square integrable martingale
with absolutely continuous angle bracket process hMit .
Local dependence on integrand and integrator

The approximations considered in the last section imply that the stochastic integral de-
pends locally both on the integrand and on the integrator in the following sense:
Corollary 5.10. Suppose that T : Ω → [0, ∞] is a random variable, M, M f are square

f
integrable martingales with absolutely continuous angle bracket processes hMi, hMi,
and H, H e are processes in L2 (0, ∞; M), L2 (0, ∞; M) f respectively, such that almost
a a
surely, Ht = He t for any t ∈ [0, T ) and Mt = Mft for any t ∈ [0, T ]. Then almost surely,
ˆ t ˆ t
Hs dMs = e s dM
H fs for any t ∈ [0, T ]. (5.3.1)
0 0
Proof. We go through the same approximations as in the proof of Theorem 5.9 above:
e t are almost surely continuous and bounded, and there
(1). Suppose first that Ht and H
exists t0 ∈ R+ such that Ht = H e t = 0 for t ≥ t0 . Let (πn ) be a sequence of
partitions with mesh(πn ) → 0. Then by Theorem 5.7,
ˆ t X
H dM = lim Hs · (Ms′ ∧t − Ms ), and
0 n→∞
s∈πn
s<t
ˆ t X
e dM
H f = lim e s · (M
H fs′ ∧t − M
fs )
0 n→∞
s∈πn
s<t
with P -almost sure uniform convergence on finite time-intervals along a common

subsequence. For t ≤ T the right-hand sides coincide, and thus (5.3.1) holds true.

176 MARTINGALES
e are bounded and Ht = H

(2). Now suppose that H and H e t = 0 for t ≥ t0 . Then the
approximations
ˆ 1/n ˆ 1/n
Htn = Ht−ε ψn (ε) dε, en =
H e t−ε ψn (ε)
H
t
0 0
e t := 0 for t < 0)
(with ψn defined as in the proof of Theorem 5.9 and Ht := H
coincide for t ≤ T . Hence by (1), on {t ≤ T },
ˆ t ˆ t ˆ t ˆ t
H dM = lim n
H dM = = lim en f =
H dM e dM
H f,
0 0 0 0
where the convergence holds again almost surely uniformly in t along a subse-
quence.
e by
(3). Finally, in the general case the assertion follows by approximating H and H
the bounded processes
e n = ((H
Htn = ((Ht ∧ n) ∨ (−n)) · I[0,n] (t), H e t ∧ n) ∨ (−n)) · I[0,n] (t).
t
Itô integrals for locally square-integrable integrands

Let M be a continuous square integrable martingale with absolutely continuous angle
bracket process hMi, and let T : Ω → [0, ∞] be an (FtP ) stopping time. We will also
be interested in the case where T = ∞. Let L2a,loc (0, T ; M) denote the linear space
consisting of all stochastic processes (t, ω) 7→ Ht (ω) defined for t ∈ [0, T (ω)) such that
the trivially extended process

 Ht for t < T,
e t :=
H
0 for t ≥ T,
is product measurable in (t, ω), adapted w.r.t. the filtration (FtP ), and
t 7→ Ht (ω) is in L2loc ([0, T (ω)), dhMi(ω)) for P -a.e. ω. (5.3.2)

Here for u ∈ (0, ∞], the space L2loc ([0, u), dhMi(ω)) consists of all measurable func-
´s
tions f : [0, u) → [−∞, ∞] such that 0 f (t)2 dhMit (ω) < ∞ for any s ∈ (0, u). In
particular, it contains all continuous functions.
e t )0≤t<∞ of
From now on, we use the notation Ht · I{t<T } for the trivial extension (H
a process (Ht )0≤t<T beyond the stopping time T . Locally square integrable adapted
processes allow for a localization by stopping times:
Lemma 5.11 (Localization by stopping). If (Ht )0≤t<T is a process in L2a,loc (0, T ; M)

then there exists an increasing sequence (Tn )n∈N of (FtP ) stopping times such that T =
sup Tn almost surely, and
Ht · I{t<Tn } ∈ L2a (0, ∞; M) for any n ∈ N.
Proof. One easily verifies that the random variables Tn defined by

ˆ t
2
Tn := inf 0 ≤ t < T : Hs dhMis ≥ n ∧ T, n ∈ N, (5.3.3)
0
are (FtP ) stopping times. Moreover, for almost every ω, the function t 7→ Ht (ω) is
´t
in L2loc ([0, T ), dhMi(ω)). Hence the function t 7→ 0 Hs (ω)2 dhMis is increasing and
finite on [0, T (ω)), and therefore Tn (ω) ր T (ω) as n → ∞. Since Tn is an (FtP )
stopping time, the process Ht · I{t<Tn } is (FtP )-adapted, and by (5.3.3),
ˆ ∞ ˆ Tn
2 2
E (Hs · I{s<Tn } ) dhMis = E Hs dhMis ≤ n for any n.
0 0
A sequence of stopping times as in the lemma will also be called a localizing sequence.
We can now extend the definition of the Itô integral to locally square-integrable adapted
integrands:
Definition (Itô integral with locally square integrable integrand). For a process H ∈
L2a,loc (0, T ; M), the Itô stochastic integral w.r.t. the martingale M is defined for t ∈
[0, T ) by
ˆ t ˆ t
Hs dMs := Hs · I{s<T̂ } dMs for any t ∈ [0, T̂ ] (5.3.4)
0 0

178 MARTINGALES
whenever T̂ is an (FtP ) stopping time such that Ht · I{t<T̂ } ∈ L2a (0, ∞; M).
´t
Theorem 5.12. For H ∈ L2a,loc (0, T ; M) the Itô integral t 7→ 0
Hs dMs is almost surely
well defined by (5.3.4) as a continuous process on [0, T ).
Proof. We have to verify that the definition does not depend on the choice of the local-
izing stopping times. This is a direct consequence of Corollary 5.10: Suppose that T̂
and Te are stopping times such that Ht · I and Ht · I e are both in L2 (0, ∞; M).
{t<T̂ } {t<T } a
Since the two trivially extended processes agree on [0, T̂ ∧ Te), Corollary 5.10 implies
that almost surely,
ˆ t ˆ t
Hs · I{s<T̂ } dMs = Hs · I{s<Te} dMs for any t ∈ [0, T̂ ∧ Te).
0 0
Hence, by Lemma 5.11, the stochastic integral is well defined on [0, T ).
Stochastic integrals as local martingales

Itô integrals w.r.t. square integrable martingales are not necessarily martingales if the
integrands are not square integrable. However, they are still local martingales in the
sense of the definition stated below.
Definition (Predictable stopping time). An (FtP ) stopping time T is called predictable

iff there exists an increasing sequence (Tk )k∈N consisting of (FtP ) stopping times such
that Tk < T on {T 6= 0} for any k, and T = sup Tk .
Example (Hitting time of a closed set). The hitting time TA of a closed set A by a
continuous adapted process is predictable, as it can be approximated from below by the
hitting times TAk of the neighbourhoods Ak = {x : dist(x, A) < 1/k}. On the other
hand, the hitting time of an open set is not predictable in general.

Definition (Local martingale). Suppose that T : Ω → [0, ∞] is a predictable stopping

time. A stochastic process Mt (ω) defined for 0 ≤ t < T (ω) is called a local martingale
up to time T , if and only if there exists an increasing sequence (Tk ) of stopping times
with T = sup Tk such that for any k ∈ N, Tk < T on {T > 0}, and the stopped process
(Mt∧Tk ) is a martingale for t ∈ [0, ∞).
Recall that by the Optional Stopping Theorem, a continuous martingale stopped at a

stopping time is again a martingale. Therefore, any continuous martingale (Mt )t≥0 is a
local martingale up to T = ∞. Even if (Mt ) is assumed to be uniformly integrable, the
converse implication fails to hold:
Exercise (A uniformly integrable local martingale that is not a martingale). Let

x ∈ R3 with x 6= 0, and suppose that (Bt ) is a three-dimensional Brownian motion with
initial value B0 = x. Prove that the process Mt = 1/|Bt | is a uniformly integrable local
martingale up to T = ∞, but (Mt ) is not a martingale.
On the other hand, note that if (Mt ) is a continuous local martingale up to T = ∞, and
the family {Mt∧Tk : k ∈ N} is uniformly integrable for each fixed t ≥ 0, then (Mt ) is a
martingale, because for 0 ≤ s ≤ t
E[Mt | Fs ] = lim E[Mt∧Tk | Fs ] = lim Ms∧Tk = Ms

k→∞ k→∞
with convergence in L1 .
As a consequence of the definition of the Itô integral by localization, we immediately

obtain:
Theorem 5.13 (Itô integrals as local martingales). Suppose that T is a predictable

stopping time w.r.t. (FtP ). Then for any H ∈ L2a,loc (0, T ; M), the Itô integral process
´t
t 7→ 0 Hs dMs is a continuous local martingale up to time T .

180 MARTINGALES
Proof. We can choose an increasing sequence (Tk ) of stopping times such that Tk < T
on {T > 0} and Ht · I{t<Tk } ∈ L2a (0, ∞; M) for any k. Then, by definition of the Itô
integral,
ˆ t∧Tk ˆ t∧Tk
Hs dMs = Hs · I{s<Tk } dMs almost surely for any k ∈ N,
0 0
and the right-hand side is a continuous martingale in Mc2 ([0, ∞)).
The theorem shows that for a predictable (FtP ) stopping time T , the Itô map H 7→
´•
0
H dM extends to a linear map
J : L2loc (0, T ; M) −→ Mc,loc ([0, T )),
where L2loc (0, T ; M) is the space of equivalence classes of processes in L2loc (0, T ; M)
that coincide for PhM i -a.e. (ω, t), and Mc,loc ([0, T )) denotes the space of equivalence
classes of continuous local (FtP ) martingales up to time T w.r.t. P -almost sure coin-
cidence. Note that different notions of equivalence are used for the integrands and the
integrals.
We finally observe that continuous local martingales (and hence stochastic integrals
w.r.t. continuous martingales) can always be localized by a sequence of bounded mar-
tingales in Mc2 ([0, ∞):
Exercise (Localization by bounded martingales). Suppose that (Mt ) is a continuous

local martingale up to time T , and (Tk ) is a localizing sequence of stopping times.
(1). Show that

Tek = Tk ∧ inf{t ≥ 0 : |Mt | ≥ k} ∧ k

is another localizing sequence, and for all k, the stopped processes Mt∧Tek
t∈[0,∞)
are bounded martingales in Mc2 ([0, ∞)).
(2). Show that if T = ∞ then T̂k := inf{t ≥ 0 : |Mt | ≥ k} is also a localizing

sequence for M.

Approximation by Riemann-Itô sums
If the integrand (Ht ) of a stochastic integral H dB has continuous sample paths then
´
local square integrability always holds, and the stochastic integral is a limit of Riemann-
Itô sums: Let (πn ) be a sequence of partition of R+ with mesh(πn ) → 0.
Theorem 5.14. Suppose that T is a predictable stopping time, and (Ht )0≤t<T is a
stochastic process defined for t < T . If the sample paths t 7→ Ht (ω) are continuous on
[0, T (ω)) for any ω, and the trivially extended process Ht · I{t<T } is (FtP ) adapted, then
H is in L2a,loc (0, T ; M), and for any t ≥ 0,
ˆ t X
Hs dMs = lim Hs · (Ms′ ∧t − Ms ) on {t < T } (5.3.5)
0 n→∞
s∈πn
s<t
with convergence in probability.
Proof. Let ⌊t⌋n = max{s ∈ πn : s ≤ t} denote the next partition point below t. By
continuity,
Ht · I{t<T } = lim H⌊t⌋n · I{t<T } .

n→∞
Hence (Ht · I{t<T } ) is (FtP ) adapted. It is also product-measurable, because
X
H⌊t⌋n · I{t<T } = Hs · I{s<T } · I[s,s′) (t) · I(0,∞) (T − t).
s∈πn
By continuity, t 7→ Ht (ω) is locally bounded for every ω, and thus H is in L2a,loc (0, T ; M).
Moreover, suppose that (Tk ) is a sequence of stopping times approaching T from below
in the sense of the definition of a predictable stopping time given above. Then
Tek := Tk ∧ inf{t ≥ 0 : |Ht | ≥ k}, k ∈ N,

182 MARTINGALES
is a localizing sequence of stopping times with Ht · I{t<Tk } in L2a (0, T ; M) for any k,
and Tek ր T . Therefore, by definition of the Itô integral and by Theorem 5.7,
ˆ t ˆ t ˆ t
Hs dMs = Hs · I{s<Tek } dMs = Hs · I{s≤Tek } dMs
0 0 0
X
= lim Hs · (Ms′ ∧t − Ms ) on {t ≤ Tek }
n→∞
s∈πn
s<t
w.r.t. convergence in probability. Since

" #
[
P {t < T } \ {t ≤ Tek } = 0,
k
we obtain (5.3.5).

Chapter 6
Itô’s formula and pathwise integrals
Our approach to Itô’s formula in this chapter follows that of [Föllmer: Stochastic Anal-
ysis, Vorlesungsskript Uni Bonn WS91/92]. We start with a heuristic derivation of the
formula that will be the central topic of this chapter.
Suppose that s 7→ Xs is a function from [0, t] to R, and F is a smooth function on R. If
(πn ) is a sequence of partitions of the interval [0, t] with mesh(πn ) → 0 then by Taylor’s
theorem
1
F (Xs′ ) − F (Xs ) = F ′ (Xs ) · (Xs′ − Xs ) + F ′′ (Xs ) · (Xs′ − Xs )2 + higher order terms.
2
Summing over s ∈ πn we obtain
X 1
F (Xt ) − F (X0 ) = F ′ (Xs ) · (Xs′ − Xs ) + F ′′ (Xs ) · (Xs′ − Xs )2 + . . . (6.0.1)
s∈π
2
n
We are interested in the limit of this formula as n → ∞.
(a) Classical case, e.g. X continuously differentiable For X ∈ C 1 we have

dXs ′
Xs ′ − Xs = (s − s) + O(|s − s′ |2 ), and
ds
(Xs′ − Xs )2 = O(|s − s′ |2 ).
Therefore, the second order terms can be neglected in the limit of (6.0.1) as mesh(πn ) →
0. Similarly, the higher order terms can be neglected, and we obtain the limit equation
ˆt
F (Xt ) − F (X0 ) = F ′ (Xs ) dXs , (6.0.2)
0
183
184 CHAPTER 6. ITÔ’S FORMULA AND PATHWISE INTEGRALS
or, in differential notation,
dF (Xt ) = F ′ (Xt ) dXt , (6.0.3)
Of course, (6.0.3) is just the chain rule of classical analysis, and (6.0.2) is the equivalent
chain rule for Stieltjes integrals, cf. Section 6.1 below.
(b) Xt Brownian motion If (Xt ) is a Brownian motion then
E[(Xs′ − Xs )2 ] = s′ − s.
Summing these expectations over s ∈ πn , we obtain the value t independently of n. This

shows that the sum of the second order terms in (6.0.1) can not be neglected anymore.
Indeed, as n → ∞, a law of large numbers type result implies that we can almost surely
replace the squared increments (Xs′ −Xs )2 in (6.0.1) asymptotically by their expectation
values. The higher order terms are on average O(|s′ − s|3/2 ) whence their sum can be
neglected. Therefore, in the limit of (6.0.1) as n → ∞ we obtain the modified chain
rule
ˆt ˆt
′ 1
F (Xt ) − F (X0 ) = F (Xs ) dXs + F ′′ (Xs ) ds (6.0.4)
2
0 0
with probability one. The equation (6.0.4) is the basic version of Itô’s celebrated for-
mula.
In Section 6.1, we will first introduce Stieltjes integrals and the chain rule from Stieltjes
calculus systematically. In Section 6.2 we prove a general version of Itô’s formula for
continuous functions with finite quadratic variation in dimension one. Here the setup
and the proof are still purely deterministic. As an aside we obtain a pathwise definition
for stochastic integrals involving only a single one-dimensional process due to Föllmer.
After computing the quadratic variation of Brownian motion in Section 6.3, we consider
first consequences of Itô’s formula for Brownian motions and continuous martingales.
Section 6.4 contains extensions to the multivariate and time-dependent case, as well as
further applications.

6.1. STIELTJES INTEGRALS AND CHAIN RULE 185
6.1 Stieltjes integrals and chain rule

In this section, we define Lebesgue-Stieltjes integrals w.r.t. deterministic functions of
finite variation, and prove a corresponding chain rule. The resulting calculus can then
be applied path by path to stochastic processes with sample paths of finite variation.
Lebesgue-Stieltjes integrals
Fix u ∈ (0, ∞], and suppose that t 7→ At is a right-continuous and non-decreasing
function on [0, u). Then At − A0 is the distribution function of the positive measure µ
on (0, u) determined uniquely by
µA [(s, t]] = At − As for any 0 ≤ s ≤ t < u.
´t
Therefore, we can define integrals of type Hs dAs as Lebesgue integrals w.r.t. the
s
measure µA . We extend µ trivially to the interval [0, u), so L1loc ([0, u), µA) is the space
of all functions H : [0, u) → R that are integrable w.r.t. µA on any interval (0, t) with
t < u. Then for any u ∈ [0, ∞] and any function H ∈ L1loc ([0, u), µA), the Lebesgue-
Stieltjes integral of H w.r.t. A is defined by
ˆt ˆ
Hr dAr := Hr · I(s,t] (r)µA (dr) for 0 ≤ s ≤ t < u.
s
It is easy to verify that the definition is consistent, i.e., varying u does not change the
´t
definition of the integrals, and that t 7→ Hr dAr is again a right-continuous function.
0
For an arbitrary right-continuous function A : [0, u) → R, the (first order) variation of

A on an interval [0, t) is defined by
(1)
X
Vt (A) := sup |As′ ∧t − As∧t | for 0 ≤ t < u,
π
s∈π
where the supremum is over all partitions π of R+ . The function t 7→ At is said to be

(1)
(locally) of finite variation on the interval [0, u) iff Vt (A) < ∞ for any t ∈ [0, u).

Any right-continuous function of finite variation can be written as the difference of two
non-decreasing right-continuous functions. In fact, we have
At = Aր ց
t − At (6.1.1)
with
X 1 (1)
Aր
t = sup (As′ ∧t − As∧t )+ = (Vt (A) + At ), (6.1.2)
π
s∈π
2
X 1 (1)
Aց
t = sup (As′ ∧t − As∧t )− = (Vt (A) − At ). (6.1.3)
π
s∈π
2
Exercise. Prove that if At is right-continuous and is locally of finite variation on [0, u)

(1)
then the functions Vt (A), Aր ց
t and At are all right-continuous and non-decreasing for
t < u.
Remark (Hahn-Jordan decomposition). The functions Aր ր ց ց

t − A0 and At − A0 are
again distribution functions of positive measures µ+ −

A and µA on (0, u). Correspondingly,
At − A0 is the distribution function of the signed measure
µA [B] := µ+ −
A [B] − µA [B], B ∈ B(0, u), (6.1.4)
(1)
and Vt is the distribution of the measure |µA | = µ+ −
A −µA . It is a consequence of (6.1.5)
and (6.1.6) that the measures µ+ −
A and µA are singular, i.e., the mass is concentrated on
disjoint sets S + and S − . The decomposition (6.1.7) is hence a particular case of the
Hahn-Jordan decomposition of a signed measure µ of finite variation into a positive and
a negative part, and the measure |µ| is the total variation measure of µ, cf. e.g. [Alt:
Lineare Funktionalanalysis].
We can now apply (6.1.1) to define Lebesgue-Stieltjes integrals w.r.t. functions of finite
variation. A function is integrable w.r.t. a signed measure µ if and only if it is integrable
w.r.t. both the positive part µ+ and the negative part µ− . The Lebesgue integral w.r.t. µ
is then defined as the difference of the Lebesgue integrals w.r.t. µ+ and µ− . Correspond-
ingly, we define the Lebesgue-Stieltjes integral w.r.t. a function At of finite variation as
the Lebesgue integral w.r.t. the associated signed measure µA :

Definition. Suppose that t 7→ At is right-continuous and locally of finite variation on

[0, u). Then the Lebesgue-Stieltjes integral w.r.t. A is defined by
ˆt ˆ ˆ
Hr dAr := Hr · I(s,t] (r) dAր
r − Hr · I(s,t] (r) dAց
r , 0 ≤ s ≤ t < u,
s
for any function H ∈ L1loc ((0, u), |dA|) where
L1loc ((0, u), |dA|) := L1loc ((0, u), dAր) ∩ L1loc ((0, u), dAց)
is the intersection of the local L1 spaces w.r.t. the positive measures dAր = µ+
A and
dAց = µ− 1
A on [0, u), or, equivalently, the local L space w.r.t. the total variation mea-
sure |dA| = |µA |.
P
n−1
Remark. (1). Simple integrands: If Ht = ci · I(ti ,ti+1 ] is a step function with
i=0
0 ≤ t0 < t1 < . . . < tn < u and c0 , c1 , . . . , cn−1 ∈ R then
ˆt n−1
X
Hs dAs = ci · (Ati+1 ∧t − Ati ∧t ).
0 i=0
(2). Continuous integrands; Riemann-Stieltjes integral: If H : [0, u) → R is a contin-

uous function then the Stieltjes integral can be approximated by Riemann sums:
ˆt X
Hs dAs = lim Hs · (As′ ∧t − As ), t ∈ [0, u),
n→∞
0 s∈πn
s<t
for any sequence (πn ) of partitions of R+ such that mesh(πn ) → 0. For the proof
note that the step functions
X
Hrn = Hs · I(s,s′ ] (r), r ∈ [0, u),
s∈πn
s<t

converge to Hr pointwise on (0, u) by continuity. Moreover, again by continuity,

Hr is locally bounded on [0, u),and hence the sequence Hrn is locally uniformly
bounded. Therefore,
ˆ ˆ
Hrn I(0,t] (r) dAr −→ Hr I(0,t] (r) dAr
for any t < n by the dominated convergence theorem.
(3). Absolutely continuous integrators: If At is an absolutely continuous function on

[0, u) then At has locally finite variation
ˆt
(1)
Vt (A) = |A′s | ds < ∞ for t ∈ [0, u).
0
The signed measure µA with distribution function At − A0 is then absolutely

continuous w.r.t. Lebesgue measure with Radon-Nikodym density
dµA
(t) = A′t for almost every t ∈ [0, u).
dt
Therefore,
L1loc ([0, u), |dA|) = L1loc ([0, u), |A′|dt),
and the Lebesgue-Stieltjes integral of a locally integrable function H is given by

ˆt ˆt
Hs dAs = Hs A′s ds for t ∈ [0, u).
0 0
In the applications that we are interested in, the integrand will mostly be continuous,
and the integrator absolutely continuous. Hence Remarks (2) and (3) above apply.
The chain rule in Stieltjes calculus

We are now able to prove Itô’s formula in the special situation where the integrator has
finite variation. In this case, the second order correction disappears, and Itô’s formula
reduces to the classical chain rule from Stieltjes calculus:

Theorem 6.1 (Fundamental Theorem of Stieltjes Calculus). Suppose that A :

[0, u) → R is a continuous function of locally finite variation. Then for any F ∈ C 2 (R),
ˆt
F (At ) − F (A0 ) = F ′ (As ) dAs ∀t ∈ [0, u). (6.1.5)
0
Proof. Let t ∈ [0, u) be given. Choose a sequence of partitions (πn ) of R+ with

mesh(πn ) → 0, and let
δAs := As′ ∧t − As∧t for s ∈ πn ,
where, as usual, s′ denotes the next partition point. By Taylor’s formula, for s ∈ πn
with s < t we have
1
F (As′ ∧t ) − F (As ) = F ′ (As )δAs + F ′′ (Zs ) · (δAs )2 ,
2
where Zs is an intermediate value between As and As′ ∧t . Summing over s ∈ πn , we

obtain
X 1 X ′′
F (At ) − F (A0 ) = F ′ (As )δAs + F (Zs )(δAs )2 . (6.1.6)
s∈πn
2 s∈π
n
s<t s<t
´t
As n → ∞, the first (Riemann) sum converges to the Stieltjes integral F ′ (As ) dAs by
0
continuity of F ′ (As ), cf. Remark (2) above.
To see that the second sum converges to zero, note that the range of the continuous
function A restricted to [0, t] is a bounded interval. Since F ′′ is continuous by assump-
tion, F ′′ is bounded on this range by a finite constant c. As Zs is an intermediate value
between As and As′ ∧t , we obtain

X ′′ X
F (Z )(δA ) 2 ≤ c· (1)
(δAs )2 ≤ c · Vt (A) · sup |δAs |.
s s
s∈πn s∈πn
s∈πn
s<t s<t
s<t

(1)
Since Vt (A) < ∞, and A is a uniformly continuous function on [0, t], the right hand
side converges to 0 as n → ∞. Hence we obtain (6.1.5) in the limit of (6.1.6) as
n → ∞.
To see that (6.1.5) can be interpreted as a chain rule, we write the equation in differential
form:
dF (A) = F ′ (A)dA. (6.1.7)
In general, the equation (6.1.7) is to be understood mathematically only as an abbrevia-

tion for the integral equation (6.1.5). For intuitive arguments, the differential notation is
obviously much more attractive than the integral form of the equation. However, for the
differential form to be useful at all, we should be able to multiply the equation (6.1.7)
by another function, and still obtain a valid equation. This is indeed possible due to the
next result, which states briefly that if dI = H dA then also G dI = GH dA:
´s
Theorem 6.2 (Stieltjes integrals w.r.t. Stieltjes integrals). Suppose that Is = Hr dAr
0
where A : [0, u) → R is a function of locally finite variation, and H ∈ L1loc ([0, u), |dA|).
Then the function s 7→ Is is again right continuous with locally finite variation
(1) ´t
Vt (I) ≤ |H| |dA| < ∞, and, for any function G ∈ L1loc ([0, u), |dI|),
0
ˆt ˆt
Gs dIs = Gs Hs dAs for t ∈ [0, u). (6.1.8)
0 0
Proof. Right continuity of It and the upper bound for the variation are left as an exercise.
We now use Riemann sum approximations to prove (6.1.8) if G is continuous. For a

partition 0 = t0 < t1 < . . . < tk = t, we have
n−1
X n−1
X ˆti+1 ˆt
Gti (Iti+1 − Iti ) = Gti · Hs dAs = G⌊s⌋ Hs dAs
i=0 i=0 ti 0

6.2. QUADRATIC VARIATION AND ITÔ’S FORMULA 191
where ⌊s⌋ denotes the largest partition point ≤ s. Choosing a sequence (πn ) of parti-
tions with mesh(πn ) → 0, the integral on the right hand side converges to the Lebesgue-
´t
Stieltjes integral Gs Hs dAs by continuity of G and the dominated convergence the-
0
´t
orem, whereas the Riemann sum on the left hand side converges to Gs dIs . Hence
0
(6.1.8) holds for continuous G. The equation for general G ∈ L1loc ([0, u), |dI|) follows
then by standard arguments.
6.2 Quadratic variation and Itô’s formula

Our next goal is to derive a generalization of the chain rule from Stieltjes calculus to
continuous functions that are not of finite variation. Examples of such functions are
typical sample paths of Brownian motion. As pointed out above, an additional term will
appear in the chain rule in this case.
Quadratic variation
Consider once more the approximation (6.1.6) that we have used to prove the funda-
mental theorem of Stieltjes calculus. We would like to identify the limit of the last sum
P ′′
F (Zs )(δAs )2 when A has unfinite variation on finite intervals. For F ′′ = 1 this
s∈πn
limit is called the quadratic variation of A if it exists:
Definition. Let u ∈ (0, ∞] and let (πn ) be a sequence of partitions of R+ with

mesh(πn ) → 0. The quadratic variation [X]t of a continuous function X : [0, u) → R
w.r.t. the sequence (πn ) is defined by
X
[X]t = lim (Xs′∧t − Xs∧t )2 for t ∈ [0, u)
n→∞
s∈πn
whenever the limit exists.
WARNINGS (Dependence on partition, classical 2-variation)

(1). The quadratic variation should not be confused with the classical 2-variation de-
fined by
(2)
X
Vt (X) := sup |Xs′ ∧t − Xs∧t |2
π
s∈π
(2)
where the supremum is over all partitions π. The classical 2-variation Vt (X)
is strictly positive for any function X that is not constant on [0, t] whereas [X]t
vanishes in many cases, cf. Example (1) below.
(2). In general, the quadratic variation may depend on the sequence of partitions con-
sidered. See however Examples (1) and (3) below.
Example. (1). Functions of finite variation: For any continuous function A : [0, u) →
R of locally finite variation, the quadratic variation along (πn ) vanishes:
[A]t = 0 for any t ∈ [0, u).
In fact, for δAs = As′ ∧t − As∧t we have

X (1)
|δAs |2 ≤ Vt (A) · sup |δAs | → 0 as n → ∞
s∈πn
s∈πn
s<t
(1)
by uniform continuity and since Vt (A) < ∞.
(2). Perturbations by functions of finite variation: If the quadratic variation [X]t of X

w.r.t. (πn ) exists, then [X + A]t also exists, and
[X + A]t = [X]t .
This holds since

X X X X
|δ(X + A)|2 = (δX)2 + 2 δXδA + (δA)2 ,
and the last two sums converge to 0 as mesh(πn ) → 0 by Example (1) and the
Cauchy-Schwarz inequality.
(3). Brownian motion: If (Bt )t≥0 is a one-dimensional Brownian motion then P -

almost surely,
[B]t = t for all t ≥ 0

w.r.t. any fixed sequence (πn ) of partitions such that mesh(πn ) → 0, cf. Theorem
6.6 below.
´t
(4). Itô processes: If It = Hs dBs is the stochastic integral of a process H ∈
0
L2a,loc (0, ∞) w.r.t. Brownian motion then almost surely, the quadratic variation
w.r.t. a fixed sequence of partitions is
ˆt
[I]t = Hs2 ds for all t ≥ 0.
0
(5). Continuous martingales: [M] exists and is almost surely finite, see below.
Note that in Examples (3), (4) and (5), the exceptional sets may depend on the sequence
(πn ). If it exists, the quadratic variation [X]t is a non-decreasing function in t. In
particular, Stieltjes integrals w.r.t. [X] are well-defined provided [X] is right continuous.
Lemma 6.3. Suppose that X : [0, u) → R is a continuous function. If the quadratic

variation [X]t along (πn ) exists for t ∈ [0, u), and t 7→ [X]t is continuous then
X ˆt
Hs · (Xs′ ∧t − Xs )2 −→ Hs d[X]s as n → ∞ (6.2.1)
s∈πn 0
s<t
for any continuous function H : [0, u) → R and any t ≥ 0.
Remark. Heuristically, the assertion of the lemma says that

ˆ ˆ
“ H d[X] = H (dX)2 ”,
i.e., the infinitesimal increments of the quadratic variation are something like squared
infinitesimal increments of X. This observation is crucial for controlling the second
order terms in the Taylor expansion used for proving the Itô-Doeblin formula.
Proof. The sum on the left-hand side of (6.2.1) is the integral of H w.r.t. the finite
positive measure
X
µn := (Xs′ ∧t − Xs )2 · δs
s∈πn
s<t

on the interval [0, t). The distribution function of µn is

X
Fn (u) = : (Xs′ ∧t − Xs )2 , u ∈ [0, t].
s∈πn
s≤t
As n → ∞, Fn (u) → [X]u for any u ∈ [0, t] by continuity of X. Since [X]u is a contin-

uous function of u, convergence of the distribution functions implies weak convergence
of the measures µn to the measure d[X] on [0, t) with distribution function [X]. Hence,
ˆ ˆ
Hs µn (ds) −→ Hs d[X]s as n → ∞
for any continuous function H : [0, t] → R.
Itô’s formula and pathwise integrals in R1

We are now able to complete the proof of the following purely deterministic (pathwise)
version of the one-dimensional Itô formula going back to [Föllmer: Calcul d’Itô sans
probabilités, Sém. Prob XV, LNM850XXX]:
Theorem 6.4 (Itô’s formula without probability). Suppose that X : [0, u) → R is a

continuous function with continuous quadratic variation [X] w.r.t. (πn ). Then for any
function F that is C 2 in a neighbourhood of X([0, u)), and for any t ∈ [0, u), the Itô
integral
ˆt X
F ′ (Xs ) dXs = lim F ′ (Xs ) · (Xs′ ∧t − Xs ) (6.2.2)
n→∞
0 s∈πn
s<t
exists, and Itô’s formula

ˆt ˆt
′ 1
F (Xt ) − F (X0 ) = F (Xs ) dXs + F ′′ (Xs ) d[X]s (6.2.3)
2
0 0
holds. In particular, if the quadratic variation [X] does not depend on (πn ) then the Itô
integral (6.2.2) does not depend on (πn ) either.

´t
Note that the theorem implies the existence of f (Xs ) dXs for any function f ∈
0
C 1 (R)! Hence this type of Itô integrals can be defined in a purely deterministic way
without relying on the Itô isometry. Unfortunately, the situation is more complicated in
higher dimensions, cf. ?? below.
Proof. Fix t ∈ [0, u) and n ∈ N. As before, for s ∈ πn we set δXs = Xs′ ∧t − Xs∧t
where s′ denotes the next partition point. Then as above we have
X 1 X ′′ (n)
F (Xt ) − F (X0 ) = F ′ (Xs )δXs + F (Zs )(δXs )2
s∈πn
2 s∈π
n
s<t s<t
(6.2.4)
X 1 X ′′ X
= F ′ (Xs )δXs + F (Xs )(δXs )2 + Rs(n) ,
s∈πn
2 s∈π s∈π
n n
s<t s<t s<t
(6.2.5)
(n) (n) (n)

where Zs is an intermediate point between Xs and Xs′ ∧t , and Rs := 12 (F ′′ (Zs ) −
F ′′ (Xs )) · (δXs )2 . As n → ∞, the second sum on the right hand side of (6.2.4) con-
´t (n)
verges to F ′′ (Xs ) d[X]s by Lemma 6.3. We claim that the sum of the remainders Rs
0
(n)
converges to 0. To see this note that Zs = Xr for some r ∈ [s, s′ ∧ t], whence
1
|Rs(n) | = |F ′′ (Zs(n) ) − F ′′ (Xs )| · (δXs )2 ≤ εn (δXs )2 ,
2
where
εn := sup |F ′′ (Xa ) − F ′′ (Xb )|.
a,b∈[0,t]
|a−b|≤mesh(πn )
As n → ∞, εn converges to 0 by uniform continuity of F ′′ ◦ X on the interval [0, t].

Thus
X 1 X
|Rs(n) | ≤ εn (δXs )2 → 0 as well,
2 s∈π
n
s<t
because the sum of the squared increments converges to the finite quadratic variation
[X]t .
We have shown that all the terms on the right hand side of (6.2.4) except the first

´t
Riemann-Itô sum converge as n → ∞. Hence, by (6.2.4), the limit F ′ (Xs ) dXs
0
of the Riemann Itô sums also exists, and the limit equation (6.2.2) holds.
Remark. (1). In differential notation, we obtain the Itô chain rule

1
dF (X) = F ′ (X) dX + F ′′ (X) d[X]
2
which includes a second order correction term due to the quadratic variation. A
justification for the differential notation is given in Section ??.
(2). For functions X with [X] = 0 we recover the classical chain rule dF (X) =
F ′ (X) dX from Stieltjes calculus as a particular case of Itô’s formula.
Example. (1). Exponentials: Choosing F (x) = ex in Itô’s formula, we obtain

ˆt ˆt
Xt X0 Xs 1
e −e = e dXs + eXs d[X]s ,
2
0 0
or, in differential notation,

1
deX = eX dX + eX d[X].
2
Thus eX does not solve the Itô differential equation
dZ = Z dX (6.2.6)
if [X] 6= 0. An appropriate renormalization is required instead. We will see below

that the correct solution of (6.2.6) is given by
Zt = exp (Xt − [X]/2) ,
cf. the first example below Theorem 6.18.
(2). Polynomials: Similarly, choosing F (x) = xn for some n ∈ N, we obtain

n(n − 1) n−2
dX n = nX n−1 dX + X [X].
2
Again, X n does not solve the equation dX n = nX n−1 dX. Here, the appropriate
renormalization leads to the Hermite polynomials : X :n , cf. the second example
below Theorem 6.18.

The chain rule for anticipative integrals

The form of the second order correction term appearing in Itô’s formula depends cru-
cially on choosing non-anticipative Riemann sum approximations. For limits of antic-
ipative Riemann sums, we obtain different correction terms, and hence also different
notions of integrals.
Theorem 6.5. Suppose that X : [0, u) → R is continuous with continuous quadratic

variation [X] along (πn ). Then for any function F that is C 2 in a neighbourhood of
X([0, u)) and for any t ≥ 0, the backward Itô integral
ˆt X
ˆ s := lim
F ′ (Xs ) dX F ′ (Xs′ ∧t ) · (Xs′ ∧t − Xs ),
n→∞
0 s∈πn
s<t
and the Stratonovich integral

ˆt X1
F ′ (Xs ) ◦ dXs := lim (F ′ (Xs ) + F ′ (Xs′ ∧t )) · (Xs′ ∧t − Xs )
n→∞
s∈πn
2
0
s<t
exist, and
ˆt ˆt
F (Xt ) − F (X0 ) = ˆ s−1
F ′ (Xs ) dX F ′′ (Xs ) d[X]s (6.2.7)
2
0 0
ˆt
= F ′ (Xs ) ◦ dXs . (6.2.8)
0
Proof. The proof of the backward Itô formula (6.2.7) is completely analogous to that of
Itô’s formula. The Stratonovich formula (6.2.8) follows by averaging the Riemann sum
approximations to the forward and backward Itô rule.
Note that Stratonovich integrals satisfy the classical chain rule
◦dF (X) = F ′ (X) ◦ dX.

This makes them very attractive for various applications. For example, in stochastic dif-
ferential geometry, the chain rule is of fundamental importance to construct stochastic
processes that stay on a given manifold. Therefore, it is common to use Stratonovich
instead of Itô calculus in this context, cf. the corresponding example in the next sec-
tion. On the other hand, Stratonovich calculus has a significant disadvantage against Itô
calculus: The Stratonovich integrals
ˆt X1
Hs ◦ dBs = lim (Hs + Hs′ ∧t )(Bs′ ∧t − Bs )
n→∞ 2
0
w.r.t. Brownian motion typically are not martingales, because the coefficients 12 (Hs +
Hs′ ∧t ) are not predictable.
6.3 Itô’s formula for Brownian motion and martingales

Our next aim is to compute the quadratic variation and to state Itô’s formula for typical
sample paths of Brownian motion. More generally, we will show that the quadratic
variation exists almost surely for continuous local martingales.
Let (πn ) be a sequence of partitions of R+ with mesh(πn ) → 0. We note first that for
any function t 7→ Xt the identity
X
Xt2 − X02 = (Xs2′ ∧t − Xs2 ) = Vtn + 2Itn (6.3.1)
s∈πn
s<t
with
X X
Vtn = (Xs′ ∧t − Xs )2 and Itn = Xs · (Xs′ ∧t − Xs )
s∈πn s∈πn
s<t s<t
holds. The equation (6.3.1) is a discrete approximation of Itô’s formula for the function
F (x) = x2 . The remainder terms in the approximation vanish in this particular case.
Note that by (6.3.1), the quadratic variation [X]t = limn→∞ Vtn exists if and only if the
´t
Riemann sum approximations Itn to the Itô integral 0 Xs dXs converge:
ˆ t
n
∃ [X]t = lim Vt ⇐⇒ ∃ Xs dXs = lim Itn .
n→∞ 0 n→∞

6.3. ITÔ’S FORMULA FOR BROWNIAN MOTION AND MARTINGALES 199
Now suppose that (Xt ) is a continuous martingale with E[Xt2 ] < ∞ for any t ≥ 0.
Then the Riemann sum approximations (Itn ) are continuous martingales for any n ∈ N.
Therefore, by the maximal inequality, for a given u > 0, the processes (Itn ) and (Vtn )
converge uniformly for t ∈ [0, u] in L2 (P ) if and only if the random variables Iun or Vun
respectively converge in L2 (P ).
Quadratic variation of Brownian motion
For the sample paths of a Brownian motion B, the quadratic variation [B] exists almost
surely along any fixed sequence of partitions (πn ) with mesh(πn ) → 0, and [B]t = t
a.s. In particular, [B] is a deterministic function that does not depend on (πn ). The
reason is a law of large numbers type effect when taking the limit of the sum of squared
increments as n → ∞.
Theorem 6.6 (P. Lévy). If (Bt ) is a one-dimensional Brownian motion on (Ω, A, P )

then as n → ∞

X

sup (Bs′ ∧t − Bs ) − t
2
−→ 0 P -a.s. and in L2 (Ω, A, P ) (6.3.2)
t∈[0,u] s∈π
s<tn
for any u ∈ (0, ∞), and for each sequence (πn ) of partitions of R+ with mesh(πn ) → 0.
Warning. (1). Although the almost sure limit in (6.3.2) does not depend on the se-
quence (πn ), the exceptional set may depend on the chosen sequence!
(2) P
(2). The classical quadratic variation Vt (B) = supπ s∈π (δBs )
2
is almost surely
infinite for any t ≥ 0. The classical p-variation is almost surely finite if and only
if p > 2.

Proof. (1). L2 -convergence for fixed t: As usual, the proof of L2 convergence is com-
P
paratively simple. For Vtn = (δBs )2 with δBs = Bs′ ∧t − Bs∧t , we have
s∈πn
X X
E[Vtn ] = E[(δBs )2 ] = δs = t, and
s∈πn s∈πn
X X
Var[Vtn ] = 2
Var[(δBs ) ] = E[((δBs )2 − δs)2 ]
s∈πn s∈πn
X
2 2
= E[(Z − 1) ] · (δs)2 ≤ const. · t · mesh(πn )
s∈πn
where Z is a standard normal random variable. Hence, as n → ∞,
Vtn − t = Vtn − E[Vtn ] → 0 in L2 (Ω, A, P ).
Moreover, by (6.3.1), Vtn − Vtm = Itn − Itm is a continuous martingale for any
n, m ∈ N. Therefore, the maximal inequality yields uniform convergence of Vtn
to t for t in a finite interval in the L2 (P ) sense.
P
(2). Almost sure convergence if mesh(πn ) < ∞: Similarly, by applying the max-
imal inequality to the process Vtn − Vtm and taking the limit as m → ∞, we
obtain
" #
2
P sup |Vtn − t| > ε ≤ E[(Vtn − t)2 ] ≤ const. · t · mesh(πn )
t∈[0,u] ε2
P
for any given ε > 0 and u ∈ (0, ∞). If mesh(πn ) < ∞ then the sum of
the probabilities is finite, and hence sup |Vtn − t| → 0 almost surely by the
t∈[0,u]
Borel-Cantelli Lemma.
P
(3). Almost sure convergence if mesh(πn ) = ∞: In this case, almost sure conver-
gence can be shown by the backward martingale convergence theorem. We refer
to Proposition 2.12 in [Revuz, YorXXX], because for our purposes almost sure
convergence w.r.t arbitrary sequences of partitions is not essential.

Itô’s formula for Brownian motion

By Theorem 6.6, we can apply Theorem 6.7 to almost every sample path of a one-
dimensional Brownian motion (Bt ):
Theorem 6.7 (Itô’s formula for Brownian motion). Suppose that F ∈ C 2 (I) where
I ⊆ R be an open interval. Then almost surely,
ˆt ˆt
1
F (Bt ) − F (B0 ) = F ′ (Bs ) dBs + F ′′ (Bs ) ds for all t < T, (6.3.3)
2
0 0
where T = inf{t ≥ 0 : Bt 6∈ I} is the first exit time from I.
Proof. For almost every ω, the quadratic variation of t 7→ Bt (ω) along a fixed sequence
of partitions is t. Moreover, for any r < T (ω), the function F is C 2 on a neighbourhood
of {Bt (ω) : t ∈ [0, r]}. The assertion now follows from Theorem 6.7 by noting that
the pathwise integral and the Itô integral as defined in Section 5 coincide almost surely
since both are limits of Riemann-Itô sums w.r.t. uniform convergence for t in a finite
interval, almost surely along a common (sub)sequence of partitions.
Consequences
´t
(1). Doob decomposition in continuous time: The Itô integral MtF = F ′ (Bs ) dBs
0
is a local martingale up to T , and MtF is a square integrable martingale if I = R
and F ′ is bounded. Therefore, (6.3.3) can be interpreted as a continuous time
Doob decomposition of the process (F (Bt )) into the (local) martingale part M F
and an adapted process of finite variation. This process takes over the role of the
predictable part in discrete time.
In particular, we obtain:

Corollary 6.8 (Martingale problem for Brownian motion). Brownian motion is a so-
1 d2
lution of the martingale problem for the operator L = with domain Dom(L ) =
dF
2 dx2
2
{F ∈ C (R) : dx is bounded}, i.e., the process
ˆt
MtF = F (Bt ) − F (B0 ) − (L f )(Bs ) ds
0
is a martingale for any F ∈ Dom(L ).
The corollary demonstrates how Itô’s formula can be applied to obtain solutions of
martingale problems, cf./ below for generalizations.
(2). Kolmogorov’s forward equation: Taking expectation values in (6.3.3), we recover

Kolmogorov’s equation
ˆ t
E[F (Bt )] = E[F (B0 )] + E[(L F )(Bs )] ds ∀t≥ 0
0
for any F ∈ Cb2 (R). In differential form,

d 1
E[F (Bt )] = E[(F ′′ )(Bt )].
dt 2
(3). Computation of expected values: The Itô formula can be applied in many ways to
compute expectation values:
Example. (a) For any n ∈ N, the process

ˆt ˆt
n(n − 1)
Btn − Bsn−2 ds = n · Bsn−1 dBs
2
0 0
is a martingale. By taking expectation values for t = 1 we obtain the recur-

sion
ˆ1 ˆ1
n(n − 1) n(n − 1)
E[B1n ] = E[Bsn−2 ] ds = sn−2/2 ds · E[B1n−2 ]
2 2
0 0
= (n − 1) · E[B1n−2 ]

for the moments of the standard normally distributed random variable B1 .

Of course this identity can be obtained directly by integration by parts in the
2
Gaussian integral xn · e−x /2 dx.
´
(b) For α ∈ R, the process

t t
α2
ˆ ˆ
exp(αBt ) − exp(αBs ) ds = α exp(αBs ) dBs
2 0 0
´t
is a martingale because E[ 0
exp(2αBs ) ds] < ∞. Denoting by Tb =
min{t ≥ 0 : Bt = b} the first passage time to a level b > 0, we obtain the
identity
ˆ Tb
2 αb
E exp(αBs ) ds = (e − 1) for any α > 0
0 α2
by optional stopping and dominated convergence.
Itô’s formula is also the key tool to derive or solve stochastic differential equations
for various stochastic processes of interest:
Example (Brownian motion on S1 ). Brownian motion on the unit circle S 1 =

{z ∈ C : |z| = 1} is the process given by
Zt = exp(iBt ) = cos Bt + i · sin Bt
where (Bt ) is a standard Brownian motion on R1 . Itô’s formula yields the stochas-
tic differential equation
1
dZt = t(Zt ) dBt − n(Zt ) dt, (6.3.4)
2
iz

where t(z) = iz is the unit tangent vector to S 1 at the point z, and n(z) = z is the
outer normal vector. If we would omit the correction term − 21 n(Zt ) dt in (6.3.4),
the solution to the s.d.e. would not stay on the circle. This is contrary to classical
o.d.e. where the correction term is not required. For Stratonovich integrals, we
obtain the modified equation
◦dZt = t(Zt ) ◦ dBt ,
which does not involve a correction term!
Quadratic variation of continuous martingales

Next, we will show that the sample paths of continuous local martingales almost surely
have finite quadratic variation. Let (Mt ) be a continuous local martingale, and fix a
sequence (πn ) of partitions of R+ with mesh(πn ) → 0. Let
X
Vtn = (Ms′ ∧t − Ms∧t )2
s∈πn
denote the quadratic variation of M along πn . Recall the crucial identity

X
Mt2 − M02 = Ms2′ ∧t − Ms∧t
2
= Vtn + 2Itn (6.3.5)
s∈πn
P
where = s∈πn Ms (Ms′ ∧t − Ms∧t ) are the Riemann sum approximations to the Itô
Itn
´t
integral 0 M dM. The identity shows that Vtn converges (uniformly) as n → ∞ if and
only if the same holds for Itn . Moreover, in this case, we obtain the limit equation
ˆ t
2 2
Mt − M0 = [M]t + 2 Ms dMs (6.3.6)
0
which is exactly Itô’s equation for F (x) = x2 .
Theorem 6.9 (Existence of quadratic variation). Suppose that (Mt ) is a continuous

local martingale on (Ω, A, P ). Then there exist a continuous non-decreasing process
´t
t 7→ [M]t and a continuous local martingale t 7→ 0 M dM such that as n → ∞,

n
ˆ s
n
sup |Vs − [M]s | → 0 and
sup Is − M dM → 0
s∈[0,t] s∈[0,t] 0

in probability for any t ≥ 0, and in L2 (P ) respectively if M is bounded. Moreover, the

identity (6.3.6) holds.
Notice that in the theorem, we do not assume the existence of an angle bracket process
hMi. Indeed, the theorem proves that for continuous local martingales, the angle bracket
process always exists and it coincides almost surely with the quadratic variation process
[M] ! We point out that for discontinuous martingales, hMi and [M] do not coincide.
Proof. We first assume that M is a bounded martingale: |Mt | ≤ C for some finite
constant C. We then show that (In ) is a Cauchy sequence in the Hilbert space Mc2 ([0, t])
for any given t ∈ R+ . To this end let n, m ∈ N. We assume without loss of generality
that πm ⊆ πn , otherwise we compare to a common refinement of both partitions. For
s ∈ πn , we denote the next partition point in πn by s′ , and the previous partition point
in πm by ⌊s⌋m . Fix t ≥ 0. Then
X
Itn − Itm = (Ms − M⌊s⌋m ) (Ms′ ∧t − Ms ), and hence
s∈πn
s<t

kI n − I m k2M 2 ([0,t]) = E (Itn − Itm )2
X
= E (Ms − M⌊s⌋m )2 (Ms′ ∧t − Ms )2
s∈πn
s<t
X 2 1/2
2 1/2 2
≤ E δm E (δMs ) , (6.3.7)
where δm := sup{|Ms − Mr |2 : |s − r| ≤ mesh(πm )}. Here we have used that the

non-diagonal summands cancel because M is a martingale.
Since M is bounded and continuous, dominated convergence shows that E[δm
2
] → 0 as
m → ∞. Furthermore,
 !2  " # " #
X X X
E (δMs )2  = E (δMs )4 + E (δMr )2 (δMs )2
s s r,s:r<s
" # " " ##
X X X
≤ 4C 2 E (δMs )2 + 2E (δMr )2 E (δMs )2 |Fr
s r s>r
2
≤ 6C E[Mt2 − M02 ] ≤ 6 C < ∞.4

Here we have used that by the martingale property,

" #
X
E (δMs )2 = E[Mt2 − M02 ] ≤ C 2 , and
s
" #
X
E (δMs )2 |Fr = E Mt2 − Mr2 |Fr ≤ C 2 .
s>r
By (6.3.7), kI n − I m k2M 2 ([0,t]) → 0 as n, m → ∞. Hence (Isn )s∈[0,t] converges uni-

formly as n → ∞ in the L2 (P ) sense. By (6.3.5), (Vsn )s∈[0,t] converges uniformly as
´•
n → ∞ in the L2 (P ) sense as well. Hence the limits 0 M dM and [M] exist, the
stochastic integral is in Mc2 ([0, t]), and the identity (6.3.6) holds.
It remains to extend the result from bounded martingales to local martingales. If M is a
continuous local martingale then there exists a sequence of stopping times Tk ↑ ∞ such
that the stopped processes (MTk ∧t )t≥0 are continuous bounded martingales. Hence the
corresponding quadratic variations [MTk ∧• ] converge uniformly in the L2 (P ) sense for
any finite t and k. Therefore, the approximations Vtn for the quadratic variation of M
converge uniformly in the L2 (P ) sense on each of the random intervals [0, Tk ∧ t], and
thus for any ε, δ > 0,

n n
P sup |Vs − [M]s | > ε ≤ P [t > Tk ] + P sup |Vs − [M]s | > ε ≤ δ
s≤t s≤Tk
for k, n sufficiently large.
Having shown the existence of the quadratic variation [M] for continuous local martin-
gales, we observe next that [M] is always non-trivial if M is not constant:
Theorem 6.10 (Non-constant continuous martingales have non-trivial quadratic

variation). Suppose that (Mt ) is a continuous local martingale. If [M]t = 0 almost
surely for some t ≥ 0, then M is almost surely constant on the interval [0, t].
Proof. Again, we assume at first that M is a bounded martingale. Then the Itô integral
´•
0
M dM is a martingale as well. Therefore, by (6.3.6),
kM − M0 k2M 2 ([0,t]) = E[(Mt − M0 )2 ] = E[Mt2 − M02 ] = E[ [M]t ] = 0,

i.e., Ms = M0 for any s ∈ [0, t]. In the general case, the assertion follows once more by
localization.
The theorem shows in particular that every local martingale with continuous finite vari-
ation paths is almost surely constant, i.e., the Doob type decomposition of a continu-
ous stochastic process into a local martingale and a continuous finite variation pro-
cess starting at 0 is unique up to equivalence. As a consequence we observe that the
quadratic variation is the unique angle bracket process of M. In particular, up to mod-
ification on measure zero sets, [M] does not depend on the chosen partition sequence
(πn ):
Corollary 6.11 (Quadratic variation as unique angle bracket process). Suppose that
(Mt ) is a continuous local martingale. Then [M] is the up to equivalence unique contin-
uous process of finite variation such that [M]0 = 0 and Mt2 −[M]t is a local martingale.
Proof. By (6.3.6), Mt2 − [M]t is a continuous local martingale. To prove uniqueness,

et ) are continuous finite variation processes with A0 = A
suppose that (At ) and (A e0 = 0
et are local martingales. Then At − A
such that both Mt2 −At and Mt2 − A et is a continuous
local martingale as well. Since the paths have finite variation, the quadratic variation of
A−A e vanishes. Hence almost surely, At − Aet = A0 − A e0 = 0 for all t.
From continuous martingales to Brownian motion
A remarkable consequence of Itô’s formula for martingales is that any continuous local
martingale (Mt ) (up to T = ∞) with quadratic variation given by [M]t = t for any
t ≥ 0 is a Brownian motion ! In fact, for 0 ≤ s ≤ t and p ∈ R, Itô’s formula yields
ˆt ˆt
ipMt ipMs ipMr p2
e −e = ip e dMr − eipMr dr
2
s s

where the stochastic integral can be identified as a local martingale. From this identity
it is not difficult to conclude that the increment Mt − Ms is conditionally independent
of FsM with characteristic function
2 (t−s)/2
E[eip(Mt −Ms ) ] = e−p for any p ∈ R,
i.e., (Mt ) has independent increments with distribution Mt − Ms ∼ N(0, t − s).
Theorem 6.12 (P. Lévy 1948). A continuous local martingale (Mt )t∈[0,∞) is a Brownian
motion if and only if almost surely,
[M]t = t for any t ≥ 0.
Exercise (Lévy’s characterization of Brownian motion). Extend the sketch above to

a proof of Theorem 6.12.
Lévy’s Theorem is the basis for many important developments in stochastic analysis
including transformations and weak solutions for stochastic differential equations. An
extension to the multi-dimensional case with a detailled proof, as well as several appli-
cations, are contained in Section 11.1 below.
One remarkable consequence of Lévy’s characterization of Brownian motion is that ev-

ery continuous local martingale can be represented as a time-changed Brownian motion
(in general possibly on an extended probability space):
Exercise (Continuous local martingales as time-changed Brownian motions). Let

(Mt )t∈[0,∞) be a continuous local martingale, and assume for simplicity that t 7→ [M]t
is almost surely strictly increasing with limt→∞ [M]t = ∞. Prove that there exists a
Brownian motion (Bt )t∈[0,∞) such that
Mt = B[M ]t for t ∈ [0, ∞). (6.3.8)
Hint: Set Ba = MTa where Ta = [M]−1 (a) = inf{t ≥ 0 : [M]t = a}, and verify by
Lévy’s characterization that B is a Brownian motion.

6.4. MULTIVARIATE AND TIME-DEPENDENT ITÔ FORMULA 209
In a more general form, the representation of continuous local martingales as time-

changed Brownian motions is due to Dambis and Dubins-Schwarz (1965), cf. [37] or
Section 11.2 below for details. Remarkably, even before Itô, Wolfgang Doeblin, the
son of Alfred Doeblin, had developed an alternative approach to stochastic calculus
where stochastic integrals are defined as time changes of Brownian motion. Doeblin
died when fighting as a French soldier at the German front in World War II, and his
results that were hidden in a closed envelope at the Académie de Sciences have become
known and been published only recently, more than fifty years after their discovery, cf.
[Doeblin, Sur l’équation de Kolmogoroff, 1940/2000], [Yor: Présentation du pli cacheté,
C.R.Acad.Sci. Paris 2000].
6.4 Multivariate and time-dependent Itô formula

We now extend Itô’s formula to Rd -valued functions and stochastic processes. Let
(1) (d)
u ∈ (0, ∞] and suppose that X : [0, u) → D, Xt = (Xt , . . . , Xt ), is a continu-
ous function taking values in an open set D ⊆ Rd . As before, we fix a sequence (πn ) of
partitions of R+ with mesh(πn ) → 0. For a function F ∈ C 2 (D), we have similarly as
in the one-dimensional case:
F (Xs′∧t ) − F (Xs ) = ∇F (Xs ) · (Xs′ ∧t − Xs ) + (6.4.1)

d
1 X ∂2F (i) (j)
(Xs )(Xs′ ∧t − Xs(i) )(Xs′ ∧t − Xs(j) ) + Rs(n)
2 i,j=1 ∂xi ∂xj
(n)
for any s ∈ πn with s < t where the dot denotes the Euclidean inner product Rs is the
remainder term in Taylor’s formula. We would like to obtain a multivariate Itô formula
by summing over s ∈ πn with s < t and taking the limit as n → ∞. A first problem
that arises in this context is the identification of the limit of the sums
X
g(Xs )δXs(i) δXs(j)
s∈πn
s<t
for a continuous function g : D → R as n → ∞.

Covariation
Suppose that X, Y : [0, u) → R are continuous functions with continuous quadratic
variations [X]t and [Y ]t w.r.t. (πn ).
Definition. The function

X
[X, Y ]t = lim (Xs′ ∧t − Xs∧t )(Ys′ ∧t − Ys∧t ), t ∈ [0, u),
n→∞
s∈πn
is called the covariation of X and Y w.r.t. (πn ) if the limit exists.
The covariation [X, Y ]t is the bilinear form corresponding to the quadratic form [X]t .
In particular, [X, X] = [X]. Furthermore:
Lemma 6.13 (Polarization identity). The covariation [X, Y ]t exists and is a continu-
ous function in t if and only if the quadratic variation [X + Y ]t exists and is continuous
respectively. In this case,
1
[X, Y ]t = ([X + Y ]t − [X]t − [Y ]t ).
2
Proof. For n ∈ N we have
X X X X
2 δXs δYs = (δXs + δYs )2 − (δXs )2 − (δYs )2 .
s∈πn s∈πn s∈πn s∈πn
The assertion follows as n → ∞ because the limits [X]t and [Y ]t of the last two terms
are continuous functions by assumption.
Remark. Note that by the polarization identity, the covariation [X, Y ]t is the difference
of two increasing functions, i.e., t 7→ [X, Y ]t has finite variation.
Example. (1). Functions and processes of finite variation: If Y has finite variation
then [X, Y ]t = 0 for any t ≥ 0. Indeed,

X X

δXs δYs ≤ sup |δXs | · |δYs |
s∈πn
s∈πn s∈πn

and the right hand side converges to 0 by uniform continuity of X on [0, t]. In
particular, we obtain again
[X + Y ] = [X] + [Y ] + 2[X, Y ] = [X].
et ) are independent Brownian mo-

(2). Independent Brownian motions: If (Bt ) and (B
tions on a probability space (Ω, A, P ) then for any given sequence (πn ),
X
e t = lim
[B, B] es = 0
δBs δ B for any t ≥ 0
n→∞
s∈πn
√
et )/ 2 is again a Brownian
P -almost surely. For the proof note that (Bt + B
motion, whence
√ 1 1 e t t
e t = [(B + B)/
[B, B] e 2]t − [B]t − [B] t = t− − = 0 almost surely.
2 2 2 2
´t ´t
(3). Itô processes: If It = 0
Gs dBs and Ft = Hs dBs with continuous adapted
0
et ) then
processes (Gt ) and (Ht ) and Brownian motions (Bt ) and (B
[I, J]t = 0 e are independent, and

if B and B (6.4.2)
ˆt
[I, J]t = Gs Hs ds e
if B = B, (6.4.3)
0
cf. Theorem ?? below.
More generally, under appropriate assumptions on G, H, X and Y , the identity

ˆt
[I, J]t = Gs Hs d[X, Y ]s
0
´t ´t
holds for Itô integrals It = Gs dXs and Jt = Hs dYs , cf. e.g. Corollary ??.
0 0
Itô to Stratonovich conversion

The covariation also occurs as the correction term in Itô compared to Stratonovich inte-
grals:

Theorem 6.14. If the Itô integral

ˆt X
Xs Ys = lim Xs δYs
n→∞
0 s∈πn
s<t
and the covariation [X, Y ]t exists along a sequence (πn ) of partitions with mesh(πn ) →
´t
0 then the corresponding backward Itô integral Xs dY ˆ s and the Stratonovich integral
0
´t
Xs ◦ dYs also exist, and
0
ˆt ˆt
ˆ s =
Xs dY Xs Ys + [X, Y ]t , and
0 0
ˆt ˆt
1
Xs ◦ dYs = Xs Ys + [X, Y ]t .
2
0 0
Proof. This follows from the identities

X X X
XS ′ ∧t δYs = Xs δYs + δXs δYs , and
X1 X 1 X
(Xs + Xs′ ∧t )δYs = Xs δYs + δXs δYs .
2 2
Itô’s formula in Rd
By the polarization identity, if [X]t , [Y ]t and [X + Y ]t exist and are continuous then
[X, Y ]t is a continuous function of finite variation.
Lemma 6.15. Suppose that X, Y and X + Y are continuous function on [0, u) with
continuous quadratic variations w.r.t. (πn ). Then
X ˆt
Hs (Xs′ ∧t − Xs )(Ys′ ∧t − Ys ) −→ Hs d[X, Y ]s as n → ∞
s∈πn 0
s<t

for any continuous function H : [0, u) → R and any t ≥ 0.
Proof. The assertion follows from Lemma 6.3 by polarization.
By Lemma 6.15, we can take the limit as mesh(πn ) → 0 in the equation derived by
summing (6.4.2) over all s ∈ πn with s < t. In analogy to the one-dimensional case,
this yields the following multivariate version of the pathwise Itô formula:
Theorem 6.16 (Multivariate Itô formula without probability). Suppose that

X : [0, u) → D ⊆ Rd is a continuous function with continuous covariations
[X (i) , X (j) ]t , 1 ≤ i, j ≤ d, w.r.t. (πn ). Then for any F ∈ C 2 (D) and t ∈ [0, u),
ˆt d ˆ t
1X ∂2F
F (Xt ) = F (X0 ) + ∇F (Xs ) · dXs + (Xs ) d[X (i) , X (j)]s ,
2 i,j=1 ∂xi ∂xj
0 0
where the Itô integral is the limit of Riemann sums along (πn ):
ˆt X
∇F (Xs ) · dXs = lim ∇F (Xs ) · (Xs′ ∧t − Xs ). (6.4.4)
n→∞
0 s∈πn
s<t
The details of the proof are similar to the one-dimensional case and left as an exercise
to the reader. Note that the theorem shows in particular that the Itô integral in (6.4.4) is
independent of the sequence (πn ) if the same holds for the covariations [X (i) , X (j) ].
Remark (Existence of pathwise Itô integrals). The theorem implies the existence of
´t
the Itô integral b(Xs ) · dXs if b = ∇F is the gradient of a C 2 function F : D ⊆
0
Rd → R. In contrast to the one-dimensional case, not every C 1 vector field b : D → Rd
´t
is a gradient. Therefore, for d ≥ 2 we do not obtain existence of 0 b(Xs ) · dXs for
any b ∈ C 1 (D, Rd ) from Itô’s formula. In particular, we do not know in general if the
´ t ∂F (i)
integrals 0 ∂x i
(Xs ) dXs , 1 ≤ i ≤ d, exists and if
ˆt Xd ˆ t
∂F
∇F (Xs ) · dXs = (Xs ) dXs(i) .
i=1
∂xi
0 0

If (Xt ) is a Brownian motion this is almost surely the case by the existence proof for Itô
integrals w.r.t. Brownian motion from Section 5.
(1) (d)
Example (Itô’s formula for Brownian motion in Rd ). Suppose that Bt = (Bt , . . . , Bt )
is a d-dimensional Brownian motion defined on a probability space (Ω, A, P ). Then the
(1) (d)
component processes Bt , . . . , Bt are independent one-dimensional Brownian mo-
tions. Hence for a given sequence of partitions (πn ) with mesh(πn ) → 0, the covari-
ations [B (i) , B (j) ], 1 ≤ i, j ≤ d, exists almost surely by Theorem 6.6 and the example
above, and
[B (i) , B (j) ] = t · δij ∀t ≥ 0
P -almost surely. Therefore, we can apply Itô’s formula to almost every trajectory. For
an open subset D ⊆ Rd and a function F ∈ C 2 (D) we obtain:
ˆt ˆt
1
F (Bt ) = F (B0 )+ ∇F (Bs )·dBs + ∆F (Bs )ds ∀t < TDC P -a.s. (6.4.5)
2
0 0
where TDC := inf{t ≥ 0 : Bt 6∈ D} denotes the first exit time from D. As in

the one-dimensional case, (6.4.5) yields a decomposition of the process F (Bt ) into a
continuous local martingale and a continuous process of finite variation, cf. Section ??
for applications.
Product rule, integration by parts

As a special case of the multivariate Itô formula, we obtain the following integration by
parts identity for Itô integrals:
Corollary 6.17. Suppose that X, Y : [0, u) → R are continuous functions with contin-
uous quadratic variations [X] and [Y ], and continuous covariation [X, Y ]. Then
ˆt !
Ys
Xt Y t − X0 Y 0 = · d Xs Ys + [X, Y ]t for any t ∈ [0, u). (6.4.6)
Xs
0

´t ´t
If one, or, equivalently, both of the Itô integrals Ys dXs and Xs dYs exist then (6.4.6)
0 0
yields
ˆt ˆt
Xt Y t − X0 Y 0 = Ys dXs + Xs dYs + [X, Y ]t . (6.4.7)
0 0
Proof. The identity (6.4.6) follows by applying Itô’s formula in R2 to the process (Xt , Yt )
´t ´t
and the function F (x, y) = xy. If one of the integrals 0 Y dX or 0 X dY exists, then
the other exists as well, and
ˆt ! ! ˆt ˆt
Ys Xs
·d = Ys dXs + Xs dYs .
Xs Ys
0 0 0
As it stands, (6.4.7) is an integration by parts formula for Itô integrals which involves the
correction term [X, Y ]t . In differential notation, it is a product rule for Itô differentials:
d(XY ) = X dY + Y dX + [X, Y ].
Again, in Stratonovich calculus a corresponding product rule holds without the correc-
tion term [X, Y ]:
◦d(XY ) = X ◦ dY + Y ◦ dX.
Remark / Warning (Existence of X dY, Lévy area). Under the conditions of the
´
´t ´t
theorem, the Itô integrals X dY and Y dX do not necessarily exist! The following
0 0
statements are equivalent:
´t
(1). The Itô integral Ys dXs exists (along (πn )).
0
´t
(2). The Itô integral Xs dYs exists.
0

(3). The Lévy area At (X, Y ) defined by

ˆt X
At (X, Y ) = (Y dX − X dY ) = lim (Ys ∆Xs − Xs ∆Ys )
n→∞
0 s∈πn
s<t
exists.
Hence, if the Lévy area At (X, Y ) is given, the stochastic integrals X dY and
´ ´
Y dX
can be constructed pathwise. Pushing these ideas further leads to the rough paths theory
developed by T. Lyons and others, cf. [Lyons, St. Flour], [Friz: Rough paths theory].
Example (Integrating finite variation processes w.r.t. Brownian motion). If (Ht ) is

an adapted process with continuous sample paths of finite variation and (Bt ) is a one-
dimensional Brownian motion then [H, B] = 0, and hence
ˆt ˆt
Ht Bt − H0 B0 = Hs dBs + Bs dHs .
0 0
This integration by parts identity can be used as an alternative definition of the stochastic
´t
integral H dB for integrands of finite variation, which can then again be extended to
0
general integrands in L2a (0, t) by the Itô isometry.
Time-dependent Itô formula

The multi-dimensional Itô formula can be applied to functions that depend explicitly
on the time variable t or on the quadratic variation [X]t . For this purpose we simply
add t or [X]t respectively as an additional component to the function, i.e., we apply the
multi-dimensional Itô formula to Yt = (t, Xt ) or Yt = (t, [X]t ) respectively.
Theorem 6.18. Suppose that X : [0, u) → Rd is a continuous function with continuous

covariations [X (i) , X (j) ]t , along (πn ), and let F ∈ C 2 (A([0, u))×Rd ). If A : [0, u) → R
is a continuous function of finite variation then the integral
ˆt X
∇x F (As , Xs ) · dXs = lim ∇x F (As , Xs ) · (Xs′ ∧t − Xs )
n→∞
0 s∈πn
s<t

exists, and the Itô formula

ˆt ˆt
∂F
F (At , Xt ) = F (0, X0) + ∇x F (As , Xs ) · dXs + (6.4.8)
(As , Xs ) dAs +
∂a
0 0
d ˆt
1X ∂2F
(As , Xs ) d[X (i) , X (j) ]s (6.4.9)
2 i,j=1 ∂xi ∂xj
0
holds for any t ≥ 0. Here ∂F/∂a denotes the derivative of F (a, x) w.r.t. the first com-
ponent, and ∇x F and ∂ 2 F/∂xi ∂xj are the gradient and the second partial derivatives
w.r.t. the other components. The most important application of the theorem is for At = t.
Here we obtain the time-dependent Itô formula
d
∂F 1 X ∂2F
dF (t, Xt ) = ∇x F (t, Xt ) · dXt + (t, Xt ) dt + (t, Xt ) d[X (i) , X (j) ]t .
∂t 2 i,j=1 ∂xi ∂xj
(6.4.10)
Similarly, if d = 1 and At = [X]t then we obtain

∂F ∂F 1 ∂2F
dF ([X]t , Xt ) = ([X]t , Xt ) dt + + ([X]t , Xt ) d[X]t . (6.4.11)
∂t ∂a 2 ∂x2
If (X)t is a Brownian motion and d = 1 then both formulas coincide.
(0) (1) (d)

Proof. Let Yt = (Yt , Yt , . . . , Yt ) := (At , Xt ). Then [Y (0) , Y (i) ]t = 0 for any t ≥ 0
(0)
and 0 ≤ i ≤ d because Yt = At has finite variation. Therefore, by Itô’s formula in
R d+1
,
d
1 X ∂2F
F (At , Xt ) = F (A0 , X0 ) + It + (As , Xs ) d[X (i) , X (j) ]s
2 i,j=1 ∂xi ∂xj
where
!
X As′ ∧t − As
Rd+1
It = lim ∇ F (As , Xs ) ·
n→∞
s∈πn Xs′ ∧t − Xs
s<t
X X
∂F
= lim (As , Xs )(As′ ∧t − As ) + ∇x F (As , Xs ) · (Xs′ ∧t − Xs ) .
n→∞ ∂a

´t ∂F
The first sum on the right hand side converges to the Stieltjes integral 0 ∂a
(As , Xs )dAs
as n → ∞. Hence, the second sum also converges, and we obtain (6.4.7) in the limit as
n → ∞.
Note that if h(t, x) is a solution of the dual heat equation
∂h 1 ∂ 2 h
+ = 0 for t ≥ 0, x ∈ R, (6.4.12)
∂t 2 ∂x2
then by (6.4.11),
ˆt
∂h
h([X]t , Xt ) = h(0, X0 ) + ([X]s , Xs ) dXs .
∂x
0
In particular, if (Xt ) is a Brownian motion, or more generally a local martingale, then

h([X]t , Xt ) is also a local martingale. The next example considers two situations where
this is particular interesting:
Example. (1). Itô exponentials: For any α ∈ R, the function
h(t, x) = exp(αx − α2 t/2)
satisfies (6.4.12) and ∂h/∂x = αh. Hence the function

(α) 1 2
Zt := exp αXt − α [X]t
2
is a solution of the Itô differential equation
(α) (α)
dZt = αZt dXt
(α) (α)
with initial condition Z0 = 1. This shows that in Itô calculus, the functions Zt
are the correct replacements for the exponential functions. The additional factor
exp(−α2 [X]t /2) should be thought of as an appropriate renormalization in the
continuous time limit.
For a Brownian motion (Xt ), we obtain the exponential martingales as general-
ized exponentials.

(2). Hermite polynomials: For n = 0, 1, 2, . . ., the Hermite polynomials
∂n 1
hn (t, x) = n
exp(αx − α2 t)|α=0
∂α 2
also satisfy (6.4.12). The first Hermite polynomials are 1, x, x2 − t, x3 − 3tx, . . ..

Note also that
∞
X
2 αn
exp(αx − α t/2) = hn (t, x)
n=0
n!
by Taylor’s theorem. Moreover, the following properties can be easily verified:
2 /2 dn −x2 /2
hn (1, x) = ex (−1)n
n
e for any x ∈ R, (6.4.13)
dx
√
hn (t, x) = tn/2 hn (1, x/ t) for any t ≥ 0, x ∈ R, (6.4.14)
∂hn ∂hn 1 ∂ 2 hn
= nhn−1 , + = 0. (6.4.15)
∂x ∂t 2 ∂x2
For example, (6.4.13) holds since
exp(αx − α2 /2) = exp(−(x − a)2 /2) exp(x2 /2)
yields
dn
2
hN (1, x) = exp(x /2)(−1) exp(−β 2
n
/2) ,
dβ n
β=x
and (6.4.14) follows from

√ √ √
exp(αx − α2 t/2) = exp(α t · (x/ t) − (α t)2 /2)
X∞
αn n/2 √
= t hn (1, x/ t).
n=0
n!
By (6.4.13) and (6.4.14), hn is a polynomial of degree n. For any n ≥ 0, the

function
(n)
Ht := hn ([X]t , Xt )
is a solution of the Itô equation
(n) (n−1)
dHt = nHt dXt . (6.4.16)

Therefore, the Hermite polynomials are appropriate replacements for the ordinary
(n)
monomials xn in Itô calculus. If X0 = 0 then H0 = 0 for n ≥ 1, and we obtain
inductively
ˆt ˆ ˆ t ˆs
(0) (1) (2)
Ht = 1, Ht = dXs , Ht = Hs(1) dXs = dXr dXs ,
0 0 0
and so on.
Corollary 6.19 (Itô 1951). If X : [0, u) → R is continuous with X0 = 0 and continuous

quadratic variation then for t ∈ [0, u),
ˆ t ˆsn ˆs2
1
··· dXs1 · · · dXsn−1 dXsn = hn ([X]t , Xt ).
n!
0 0 0
Proof. The equation follows from (6.4.16) by induction on n.
Iterated Itô integrals occur naturally in Taylor expansions of Itô calculus. Therefore, the
explicit expression from the corollary is valuable for numerical methods for stochastic
differential equations, cf. Section ?? below.

Chapter 7
Brownian Motion and Partial

Differential Equations
The stationary and time-dependent Itô formula enable us to work out the connection of
Brownian motion to several partial differential equations involving the Laplace operator
in detail. One of the many consequences is the evaluation of probabilities and expec-
tation values for Brownian motion by p.d.e. methods. More generally, Itô’s formula
establishes a link between stochastic processes and analysis that is extremely fruitful in
both directions.
Suppose that (Bt ) is a d-dimensional Brownian motion defined on a probability space
(Ω, A, P ) such that every sample path t 7→ Bt (ω) is continuous. We first note that Itô’s
formula shows that Brownian motion solves the martingale problem for the operator
1
L = ∆ in the following sense:
2
Corollary 7.1 (Time-dependent martingale problem). The process

ˆt
F ∂F 1
Mt = F (t, Bt ) − F (0, B0) − + ∆F (s, Bs ) ds
∂s 2
0
is a continuous (FtB ) martingale for any C 2 function F : [0, ∞) × Rd → R with

bounded first derivatives. Moreover, M F is a continuous local martingale up to TDC =
inf{t ≥ 0 : Bt 6∈ D} for any F ∈ C 2 ([0, ∞) × D), D ⊆ Rd open.
221
222 CHAPTER 7. BROWNIAN MOTION AND PDE
Proof. By the continuity assumptions one easily verifies that M F is (FtB ) adapted.
Moreover, by the time-dependent Itô formula (6.4.10),
ˆt
MtF = ∇x F (s, Bs ) · dBs for t < TDC ,
0
which implies the claim.
Choosing a function F that does not explicitly depend on t, we obtain in particular that
ˆt
1
MtF = F (Bt ) − F (B0 ) − ∆F (Bs ) ds
2
0
is a martingale for any f ∈ Cb2 (Rd ), and a local martingale up to TDC for any F ∈
C 2 (D).
7.1 Dirichlet problem, recurrence and transience

As a first consequence of Corollary 7.1 we can now complete the proof of the stochastic
representation for solutions of the Dirichlet problem that has been already mentioned in
Section 3.2 above. By solving the Dirichlet problem for balls explicitly, we will then
study recurrence, transience and polar sets for multi-dimensional Brownian motion.
The Dirichlet problem revisited

Suppose that h ∈ C 2 (D) ∩ C(D) is a solution of the Dirichlet problem
∆h = 0 on D, h=f on ∂D, (7.1.1)
for a bounded open set D ⊂ Rd and a continuous function f : ∂D → R. If (Bt ) is under

Px a continuous Brownian motion with B0 = x Px -almost surely, then by Corollary 7.1,
the process h(Bt ) is a local (FtB ) martingale up to TDC . By applying the optional

7.1. DIRICHLET PROBLEM, RECURRENCE AND TRANSIENCE 223
stopping theorem with a localizing sequence of bounded stopping times Sn ր TDC , we

obtain
h(x) = Ex [h(B0 )] = Ex [h(BSn )] for any n ∈ N.
Since Px [TDC < ∞] = 1 and h is bounded on D, dominated convergence then yields
the stochastic representation
h(x) = Ex [h(BTDC )] = Ex [f (BTDC )] for any x ∈ Rd .
We thus have shown:
Theorem 7.2 (Stochastic representation for solutions of the Dirichlet problem).

Suppose that D is a bounded open subset of Rd , f is a continuous function on the
boundary ∂D, and h ∈ C 2 (D) ∩ C(D) is a solution of the Dirichlet problem (7.1.1).
Then
h(x) = Ex [f (BT )] for any x ∈ D.
We will generalize this result substantially in Theorem 7.5 below. Before, we apply the
Dirichlet problem to study recurrence and transience of Brownian motions:
Recurrence and transience of Brownian motion in Rd

Let (Bt ) be a d-dimensional Brownian motion on (Ω, A, P ) with initial value B0 =
x0 , x0 6= 0. For r ≥ 0 let
Tr = inf{t > 0 : |Bt | = r}.
We now compute the probabilities P [Ta < Tb ] for a < |x0 | < b. Note that this is a
multi-dimensional analogue of the classical ruin problem. To compute the probability
for given a, b we consider the domain
D = {x ∈ Rd : a < |x| < b}.
For b < ∞, the first exit time TDC is almost surely finite,
TDC = min(Ta , Tb ), and P [Ta < Tb ] = P [|BTDC | = a].

x0
a
Suppose that h ∈ C(U) ∩ C 2 (U) is a solution of the Dirichlet problem


1 if |x| = a,
∆h(x) = 0 for all x ∈ D, h(x) = (7.1.2)
0 if |x| = b.
Then h(Bt ) is a bounded local martingale up to TDC and optional stopping yields
P [Ta < Tb ] = E[h(BTDC )] = h(x0 ). (7.1.3)
By rotational symmetry, the solution of the Dirichlet problem (7.1.2) can be computed
explicitly. The Ansatz h(x) = f (|x|) leads us to the boundary value problem
d2 f d − 1 df
2
(|x|) + (|x|) = 0, f (a) = 1, f (b) = 0,
dr |x| dr
for a second order ordinary differential equation. Solutions of the o.d.e. are linear
combinations of the constant function 1 and the function


 s for d = 1,


φ(s) := log s for d = 2,



s2−d for d ≥ 3.

7.1. DIRICHLET PROBLEM, RECURRENCE AND TRANSIENCE 225
φ(s)
Figure 7.1: The function φ(s) for different values of d: red (d = 1), blue (d = 2) and
purple (d = 3)
Hence, the unique solution f with boundary conditions f (a) = 1 and f (b) = 0 is
φ(b) − φ(r)
f (r) = .
φ(b) − φ(a)
Summarizing, we have shown:
Theorem 7.3 (Ruin problem in Rd ). For a, b > 0 with a < |x0 | < b,
φ(b) − φ(|x0 |)
P [Ta < Tb ] = , and
φ(b) − φ(a)

1 for d ≤ 2
P [Tb < ∞] =
(a/|x |)d−2 for d > 2.
0
Proof. The first equation follows by 6.4.12. Moreover,


1 for d ≤ 2
P [Ta < ∞] = lim P [Ta < Tb ] =
b→∞ φ(|x |)/φ(a) for d ≥ 3.
0

Corollary 7.4. For a Brownian motion in Rd the following statements hold for any
initial value x0 ∈ Rd :
(1). If d ≤ 2 then every non-empty ball D ⊆ Rd is recurrent, i.e., the last visit time of
D is almost surely infinite:
Ld = sup{t ≥ 0 : Bt ∈ D} = ∞ P -a.s.
(2). If d ≥ 3 then every ball D is transient, i.e.,
Ld < ∞ P -a.s.
(3). If d ≥ 2 then every point x ∈ Rd is polar, i.e.,
P [ ∃ t > 0 : Bt = x] = 0.
We point out that the last statement holds even if the starting point x0 coincides with x.
The first statement implies that a typical Brownian sample path is dense in R2 , whereas
by the second statement, lim |Bt | = ∞ almost surely for d ≥ 3.
t→∞
Proof.
(1),(2) The first two statements follow from Theorem 7.3 and the Markov property.

7.2. BOUNDARY VALUE PROBLEMS, EXIT AND OCCUPATION TIMES 227
(3). For the third statement we assume w.l.o.g. x = 0. If x0 6= 0 then
P [T0 < ∞] = lim P [T0 < Tb ]

b→∞
for any a > 0. By Theorem 7.3,
P [T0 < Tb ] ≤ inf P [Ta < Tb ] = 0 for d ≥ 2,

a>0
whence T0 = ∞ almost surely. If x0 = 0 then by the Markov property,
P [ ∃ t > ε : Bt = 0] = E[PBε [T0 < ∞]] = 0
for any ε > 0. thus we again obtain
P [T0 < ∞] = lim P [ ∃ t > ε : Bt = 0] = 0.

εց0
Remark (Polarity of linear subspaces). For d ≥ 2, any (d − 2) dimensional subspace

V ⊆ Rd is polar for Brownian motion. For the proof note that the orthogonal projection
of a one-dimensional Brownian motion onto the orthogonal complement V ⊥ is a 2-
dimensional Brownian motion.
7.2 Boundary value problems, exit and occupation times

The connection of Brownian motion to boundary value problems for partial differential
equations involving the Laplace operator can be extended substantially:
The stationary Feynman-Kac-Poisson formula

Suppose that f : ∂D → R, V : D → R and g : D → [0, ∞) are continuous functions
defined on an open bounded domain D ⊂ Rd , or on its boundary respectively. We
assume that under Px , (Bt ) is Brownian motion with Px [B0 = x] = 1, and that
 
ˆT
Ex exp V − (Bs ) ds < ∞ for any x ∈ D, (7.2.1)
0

where T = TDC is the first exit time from D.

Note that (7.2.1) always holds if V is non-negative.
Theorem 7.5. If u ∈ C 2 (D) ∩ C(D) is a solution of the boundary problem

1
∆u(x) = V (x)u(x) − g(x) for x ∈ D
2
(7.2.2)
u(x) = f (x) for x ∈ ∂D, (7.2.3)
and (7.2.1) holds then

   
ˆT
u(x) = Ex exp − V (Bs ) ds · f (BT ) + (7.2.4)
0
   
ˆT ˆt
Ex  exp − V (Bs ) ds · g(Bt ) dt
0 0
for any x ∈ D.
Remark. Note that we assume the existence of a smooth solution of the boundary value
problem (7.2.2). Proving that the function u defined by (7.2.4) is a solution of the b.v.p.
without assuming existence is much more demanding.
Proof. By continuity of V and (Bs ), the sample paths of the process
ˆt
At = V (Bs ) ds
0
are C 1 and hence of finite variation for t < T . Let
Xt = e−At u(Bt ), t < T.

Applying Itô’s formula with F (a, b) = e−a u(b) yields the decomposition
1
dXt = e−At ∇u(Bt ) · dBt − e−At u(Bt ) dAt + e−At ∆u(Bt ) dt
2
1
= e−At ∇u(Bt ) · dBt + e−At ∆u − V · u (Bt ) dt
2
of Xt into a local martingale up to time T and an absolutely continuous part. Since u
1
is a solution of (7.2.2), we have ∆u − V u = −g on D. By applying the optional
2
stopping theorem with a localizing sequence Tn ր T of stopping times, we obtain the
representation
 
ˆTn
u(x) = Ex [X0 ] = Ex [XTn ] + Ex  e−At g(Bt ) dt
0
 
ˆTn
= Ex [e−ATn u(BTn )] + Ex  e−At g(Bt ) dt
0
for x ∈ D. The assertion (7.2.4) now follows provided we can interchange the limit
as n → ∞ and the expectation values. For the second expectation on the right hand
side this is possible by the monotone convergence theorem, because g ≥ 0. For the first
expectation value, we can apply the dominated convergence theorem, because
 T 
−A
ˆ
e Tn u(BTn ) ≤ exp  V − (Bs ) ds · sup |u(y)| ∀n ∈ N,
y∈D
0
and the majorant is integrable w.r.t. each Px by Assumption 7.2.1.
Remark (Extension to diffusion processes). A corresponding result holds under ap-

propriate assumptions if the Brownian motion (Bt ) is replaced by a diffusion process
(Xt ) solving a stochastic differential equation of the type dXt = σ(Xt ) dBt + b(Xt ) dt,
and the operator 21 ∆ in (7.2.2) is replaced by the generator
d
1X ∂2
L = ai,j (x) + b(x) · ∇, a(x) = σ(x)σ(x)⊤ ,
2 i,j=1 ∂xi ∂xj
of the diffusion process, cf. ??. The theorem hence establishes a general connection
between Itô diffusions and boundary value problems for linear second order elliptic
partial differential equations.

By Theorem 7.5 we can compute many interesting expectation values for Brownian mo-
tion by solving appropriate p.d.e. We now consider various corresponding applications.
Let us first recall the Dirichlet problem where V ≡ 0 and g ≡ 0. In this case,
u(x) = Ex [f (Bt )]. We have already pointed out in the last section that this can be
used to compute exit distributions and to study recurrence, transience and polarity of
linear subspaces for Brownian motion in Rd . A second interesting case of Theorem 7.5
is the stochastic representation for solutions of the Poisson equation:
Poisson problem and mean exit time

If V and f vanish in Theorem 7.3, the boundary value problem (7.2.2) reduces to the
boundary value problem
1
∆u = −g on D, u = 0 on D,
2
for the Poisson equation. The solution has the stochastic representation
 T 
ˆ
u(x) = Ex  g(Bt ) dt , x ∈ D, (7.2.5)
0
which can be interpreted as an average cost accumulated by the Brownian path before
exit from the domain D. In particular, choosing g ≡ 1, we can compute the mean exit
time
u(x) = Ex [T ]
from D for Brownian motion starting at x by solving the corresponding Poisson prob-
lem.
Example. If D = {x ∈ Rd : |x| < r} is a ball around 0 of radius r > 0, then the

solution u(x) of the Poisson problem

1 −1 for |x| < r
∆u(x) =
2 0 for |x| = r
can be computed explicitly. We obtain
r 2 − |x|2
Ex [T ] = u(x) = for any x ∈ D.
d

Occupation time density and Green function

If (Bt ) is a Brownian motion in Rd then the corresponding Brownian motion with ab-
sorption at the first exit time from the domain D is the Markov process (Xt ) with state
space D ∪ {∆} defined by

Bt for t < T
Xt = ,
∆ for t ≥ T
where ∆ is an extra state added to the state space. By setting g(∆) = 0, the stochastic
representation (7.2.5) for a solution of the Poisson problem can be written in the form
∞ 
ˆ ˆ∞
u(x) = Ex  g(Xt ) dt = (pD
t g)(x) dt, (7.2.6)
0 0
where
pD
t (x, A) = Px [Xt ∈ A], A ⊆ Rd measurable,
is the transition function for the absorbed process (Xt ). Note that for A ⊂ Rd ,
pD
t (x, A) = Px [Bt ∈ A and t < T ] ≤ pt (x, A) (7.2.7)
where pt is the transition function of Brownian motion on Rd . For t > 0 and x ∈ Rd ,

the transition function pt (x, •) of Brownian motion is absolutely continuous. There-
fore, by (7.2.7), the sub-probability measure pD
t (x, •) restricted to R is also absolutely
d
continuous with non-negative density

|x − y|2
pD
t (x, y) ≤ pt (x, y) = (2πt) −d/2
exp − .
2t
The function pD
t is called the heat kernel on the domain D w.r.t. absorption on the
boundary. Note that
ˆ∞
D
G (x, y) = pD
t (x, y) dt
0
is an occupation time density, i.e., it measures the average time time a Brownian mo-
tion started in x spends in a small neighbourhood of y before it exits from the Domain

D. By (7.2.6), a solution u of the Poisson problem 21 ∆u = −g on D, u = 0 on ∂D, can

be represented as
ˆ
u(x) = GD (x, y)g(y) dy for x ∈ D.
D
This shows that the occupation time density GD (x, y) is the Green function (i.e.,
1
the fundamental solution of the Poisson equation) for the operator 2
with Dirichlet
boundary conditions on the domain D.
Note that although for domains with irregular boundary, the Green’s function might not
exist in the classical sense, the function GD (x, y) is always well-defined!
Stationary Feynman-Kac formula and exit time distributions

Next, we consider the case where g vanishes and f ≡ 1 in Theorem 7.5. Then the
boundary value problem (7.2.4) takes the form
1
∆u = V u on D, u = 1 on ∂D. (7.2.8)
2
1
The p.d.e. 2
∆u = V u is a stationary Schrödinger equation. We will comment on the
relation between the Feynman-Kac formula and Feynman’s path integral formulation of
quantum mechanics below. For the moment, we only note that for the solution of (??),
the stochastic representation
  
ˆT
u(x) = Ex exp − V (Bt ) dt
0
holds for x ∈ D.
As an application, we can, at least in principle, compute the full distribution of the exit
time T . In fact, choosing V ≡ α for some constant α > 0, the corresponding solution
uα of (7.2.8) yields the Laplace transform
ˆ∞
uα (x) = Ex [e−αT ] = e−αt µx (dt) (7.2.9)
0
of µx = Px ◦ T −1 .

Example (Exit times in R1 ). Suppose d = 1 and D = (−1, 1). Then (7.2.8) with
V = α reads
1 ′′
u (x) = αuα (x) for x ∈ (−1, 1), uα (1) = uα (−1) = 1.
2 α
This boundary value problem has the unique solution

√
cosh(x · 2α)
uα (x) = √ for x ∈ [−1, 1].
cosh( 2α)
By inverting the Laplace transform (7.2.9), one can now compute the distribution µx
of the first exit time T from (−1, 1). It turns out that µx is absolutely continuous with
density
∞
X
1 − (4n+1+x)
2
− (4n+1−x)
2
fT (t) = √ (4n + 1 + x)e 2t + (4n + 1 − x)e 2t , t ≥ 0.
2πt3 n=−∞
fT (t)
Figure 7.2: The density of the first exit time T depending on the starting point x ∈
[−1, 1] and the time t ∈ (0, 2].

Boundary value problems in Rd and total occupation time

Suppose we would like to compute the distribution of the total occupation time
ˆ∞
IA (Bs ) ds
0
of a bounded domain A ⊂ Rd for Brownian motion. This only makes sense for d ≥ 3,
since for d ≤ 2, the total occupation time of any non-empty open set is almost surely
infinite by recurrence of Brownian motion in R1 and R2 . The total occupation time is of
´∞
the form V (Bs ) ds with V = IA . Therefore, we should in principle be able to apply
0
Theorem 7.3, but we have to replace the exit time T by +∞ and hence the underlying
bounded domain D by Rd .
Corollary 7.6. Suppose d ≥ 3 and let V : Rd → [0, ∞) be continuous. If u ∈ C 2 (Rd )

is a solution of the boundary value problem
1
∆u = V u on Rd , lim u(x) = 1 (7.2.10)
2 |x|→∞
then   
ˆ∞
u(x) = Ex exp − V (Bt ) dt for any x ∈ Rd .
0
Proof. Applying the stationary Feynman-Kac formula on an open bounded subset D ⊂

Rd , we obtain the representation
  
TD C
ˆ
  
u(x) = Ex u(BTDC ) exp − V (Bt ) dt (7.2.11)
0
by Theorem 7.3. Now let Dn = {x ∈ Rd : |x| < n}. Then TDnC ր ∞ as n → ∞. Since
d ≥ 3, Brownian motion is transient, i.e., lim |Bt | = ∞, and therefore by (7.2.10)
t→∞
lim u(BTDC ) = 1 Px -almost surely for any x.

n→∞ n

Since u is bounded and V is non-negative, we can apply dominated convergence in

(7.2.11) to conclude
  
ˆ∞
u(x) = Ex exp − V (Bt ) dt .
0
Let us now return to the computation of occupation time distributions. consider a

bounded subset A ⊂ Rd , d ≥ 3, and let
  
ˆ∞
vα (x) = Ex exp −α IA (Bs ) ds , α > 0,
0
denote the Laplace transform of the total occupation time of A. Although V = αIA is
not a continuous function, a representation of vα as a solution of a boundary problem
holds:
Exercise. Prove that if A ⊂ Rd is a bounded domain with smooth boundary ∂A and

uα ∈ C 1 (Rd ) ∩ C 2 (Rd \ ∂A) satisfies
1
∆uα = αIA uα on Rd \ ∂A, lim uα (x) = 1, (7.2.12)
2 |x|→∞
then vα = uα .
Remark. The condition uα ∈ C 1 (Rd ) guarantees that uα is a weak solution of the p.d.e.
(7.2.10) on all of Rd including the boundary ∂U.
Example (Total occupation time of the unit ball in R3 ). Suppose A = {x ∈ R3 :

|x| < 1}. In this case the boundary value problem (7.2.10) is rotationally symmetric.
The ansatz uα (x) = fα (|x|), yields a Bessel equation for fα on each of the intervals
(0, 1) and (1, ∞):
1 ′′ 1 ′′
f (r) + r −1 fα′ (r) = αfα (r) for r < 1, f (r) + r −1 fα (r) = 0 for r > 1.
2 α 2 α
Taking into account the boundary condition and the condition uα ∈ C 1 (Rd ), one obtains
the rotationally symmetric solution
 √ !

 tanh( 2α)
1 +
 √ − 1 · r −1 for r ∈ [1, ∞),

 2α

 √
uα (x) = √sinh( 2αr) √ · r −1 for r ∈ (0, 1) .



 2α cosh 2α

 1

 √ for r = 0
cosh( 2α)
of (7.2.10), and hence an explicit formula for vα . In particular, for x = 0 we obtain the
simple formula
  
ˆ∞
1
E0 exp −α IA (Bt ) dt = uα (0) = √ .
cosh( 2α)
0
The right hand side has already appeared in the example above as the Laplace transform
of the exit time distribution of a one-dimensional Brownian motion starting at 0 from the
interval (−1, 1). Since the distribution is uniquely determined by its Laplace transform,
we have proven the remarkable fact that the total occupation time of the unit ball for a
standard Brownian motion in R3 has the same distribution as the first exit time from the
unit ball for a standard one-dimensional Brownian motion:
ˆ∞
3
I{|BR3 |<1} dt ∼ inf{t > 0 : |BtR | > 1}.
t
0
This is a particular case of a theorem of Ciesielski and Taylor who proved a correspond-
ing relation between Brownian motion in Rd+2 and Rd for arbitrary d.
7.3 Heat Equation and Time-Dependent Feynman-Kac

Formula
Itô’s formula also yields a connection between Brownian motion (or, more generally, so-
lutions of stochastic differential equations) and parabolic partial differential equations.
The parabolic p.d.e. are Kolmogorov forward or backward equations for the correspond-
ing Markov processes. In particular, the time-dependent Feynman-Kac formula shows

7.3. HEAT EQUATION AND TIME-DEPENDENT FK FORMULA 237
that the backward equation for Brownian motion with absorption is a heat equation with
dissipation.
Brownian Motion with Absorption

Suppose we would like to describe the evolution of a Brownian motion that is absorbed
during the evolution of a Brownian motion that is absorbed during an infinitesimal time
interval [t, t + dt] with probability V (t, x)dt where x is the current position of the pro-
cess. We assume that the absorption rate V (t, x) is given by a measurable locally-
bounded function
V : [0, ∞) × Rd → [0, ∞).
Then the accumulated absorption rate up to time t is given by the increasing process
ˆt
At = V (s, Bs ) ds, t ≥ 0.
0
We can think of the process At as an internal clock for the Brownian motion determining
the absorption time. More precisely, we define:
Definition. Suppose that (Bt )t≥0 is a d-dimensional Brownian motion and T is a with
parameter 1 exponentially distributed random variable independent of (Bt ). Let ∆ be
a separate state added to the state space Rd . Then the process (Xt ) defined by

Bt for At < T,
Xt :=
∆ for A ≥ T,
t
is called a Brownian motion with absorption rate V (t, x), and the random variable
ζ := inf{t ≥ 0 : At ≥ T }
is called the absorption time.

A justification for the construction is given by the following informal computation: For
an infinitesimal time interval [t, t + dt] and almost every ω,
P [ζ ≤ t + dt | (Bs )s≥0 , ζ > t](ω) = P [At+dt (ω) ≥ T | At (ω) < T ]

= P [At+dt (ω) − At (ω) ≥ T ]
= P [V (t, Bt (ω))dt ≥ T ]
= V (t, Bt (ω))dt
by the memoryless property of the exponential distribution, i.e., V (t, x) is indeed the
infinitesimal absorption rate.
Rigorously, it is not difficult to verify that (Xt ) is a Markov process with state space
Rd ∪ {∆} where ∆ is an absorbing state. The Markov process is time-homogeneous if
V (t, x) is independent of t.
For a measurable subset D ⊆ Rd and t ≥ 0 the distribution µt of Xt is given by
µt [D] = P [Xt ∈ D] = P [Bt ∈ D and At < T ]

= E[P [At < T | (Bt )] ; Bt ∈ D] (7.3.1)
   
ˆt
= E exp − V (s, Bs ) ds ; Bt ∈ D  .
0
Itô’s formula can be used to prove a Kolmogorov type forward equation:
Theorem 7.7 (Forward equation for Brownian motion with absorption). The sub-
probability measures µt on Rd solve the heat equation
∂µt 1
= ∆µt − V (t, •)µt (7.3.2)
∂t 2
in the following distributional sense:
ˆt ˆ
1
ˆ ˆ
f (x)µt (dx) − f (x)µ0 (dx) = ( ∆f (x) − V (s, x)f (x))µs (dx) ds
2
0
for any function f ∈ C02 (Rd ).

Here C02 (Rd ) denotes the space of C 2 -functions with compact support. Under additional
regularity assumptions it can be shown that µt has a smooth density that solves (7.3.1)
in the classical sense. The equation (7.3.1) describes heat flow with cooling when the
heat at x at time t dissipates with rate V (t, x).
Proof. By (7.3.1), ˆ
f dµt = E[exp(−At ) ; f (Bt )] (7.3.3)
for any bounded measurable function f : Rd → R. For f ∈ C02 (Rd ), an application of

Itô’s formula yields
ˆt ˆt
1
e−At f (Bt ) = f (B0 ) + Mt + e−As f (Bs )V (s, Bs ) ds + e−As ∆f (Bs ) ds,
2
0 0
for t ≥ 0, where (Mt ) is a local martingale. Taking expectation values for a localizing
sequence of stopping times and applying the dominated convergence theorem subse-
quently, we obtain
ˆt
−At 1
E[e f (Bt )] = E[f (B0 )] + E[e−As ( ∆f − V (s, •)f )(Bs )] ds.
2
0
Here we have used that 12 ∆f (x)−V (s, x)f (x) is uniformly bounded for (s, x) ∈ [0, t]×
Rd , because f has compact support and V is locally bounded. The assertion now follows
by (7.3.3).
Exercise (Heat kernel and Green’s function). The transition kernel for Brownian mo-
tion with time-homogeneous absorption rate V (x) restricted to Rd is given by
   
ˆt
pVt (x, D) = Ex exp − V (Bs ) ds ; Bt ∈ D  .
0
(1). Prove that for any t > 0 and x ∈ Rd , the sub-probability measure pVt (x, •) is
absolutely continuous on Rd with density satisfying
0 ≤ pVt (x, y) ≤ (2πt)−d/2 exp(−|x − y|2 /(2t)).

(2). Identify the occupation time density

ˆ∞
GV (x, y) = pVt (x, y) dt
0
as a fundamental solution of an appropriate boundary value problem. Adequate

regularity may be assumed.
Time-dependent Feynman-Kac formula

In Theorem 7.7 we have applied Itô’s formula to prove a Kolmogorov type forward
equation for Brownian motion with absorption. To obtain a corresponding backward
equation, we have to reverse time:
Theorem 7.8 (Feynman-Kac). Fix t > 0, and let f : Rd → R and V, g : [0, t] × Rd →

R be continuous functions. Suppose that f is bounded, g is non-negative, and V satisfies
 
ˆt
Ex exp V (t − s, Bs )− ds < ∞ for all x ∈ Rd . (7.3.4)
0
If u ∈ C 1,2 ((0, t] × Rd ) ∩ C([0, t] × Rd ) is a bounded solution of the heat equation
∂u 1
(s, x) = ∆u(s, x) − V (s, x)u(s, x) + g(s, x) for s ∈ (0, t], x ∈ Rd ,
∂s 2
(7.3.5)
u(0, x) = f (x),
then u has the stochastic representation

  
ˆt
u(t, x) = Ex f (Bt ) exp − V (t − s, Bs ) ds +
0
 t  r  
ˆ ˆ
Ex  g(t − r, Br ) exp − V (t − s, Bs ) ds dr  .
0 0

Remark. The equation (7.3.5) describes heat flow with sinks and dissipation.
Proof. We first reverse time on the interval [0, t]. The function
û(s, x) = u(t − s, x)
solves the p.d.e.

∂ û ∂u 1
(s, x) = − (t − s, x) = − ∆u − V u + g (t − s, x)
∂s ∂t 2

1
= − ∆û − V̂ û + ĝ (s, x)
2
on [0, t] with terminal condition û(t, x) = f (x). Now let Xr = exp(−Ar )û(r, Br ) for
r ∈ [0, t], where
ˆr ˆr
Ar := V̂ (s, Bs ) ds = V (t − s, Bs ) ds.
0 0
By Itô’s formula, we obtain for τ ∈ [0, t],

ˆτ ˆτ
−Ar −Ar ∂ û 1
Xτ − X0 = Mτ − e û(r, Br ) dAr + e + ∆û (r, Br ) dr
∂s 2
0 0
ˆτ
−Ar ∂ û 1
= Mτ + e + ∆û − V̂ û (r, Br ) dr
∂s 2
0
ˆτ
= Mτ − e−Ar ĝ(r, Br ) dr
0
with a local martingale (Mτ )τ ∈[0,t] vanishing at 0. Choosing a corresponding localizing

sequence of stopping times Tn with Tn ր t, we obtain by the optional stopping theorem
and by dominated convergence,
u(t, x) = û(0, x) = Ex [X0 ]

 t 
ˆ
= Ex [Xt ] + Ex  e−Ar ĝ(r, Br ) dr 
0
 
ˆt
= Ex [e−At u(0, Bt)] + Ex  e−Ar g(t − r, Br ) dr  .
0

Remark (Extension to diffusion processes). Again a similar result holds under a ap-
propriate regularity assumptions for Brownian motion replaced by a solution of a s.d.e.
dXt = σ(Xt )dBt + b(Xt )dt and 21 ∆ replaced by the corresponding generator, cf. ??.
Occupation times and arc-sine law

The Feynman-Kac formula can be used to study the distribution of occupation times
of Brownian motion. We consider an example where the distribution can be computed
explicitly: The proportion of time during the interval [0, t] spent by a one-dimensional
standard Brownian motion (Bt ) in the interval (0, ∞). Let
ˆt
At = λ({s ∈ [0, t] : Bs > 0}) = I(0,∞) (Bs ) ds.
0
Theorem 7.9 (Arc-sine law of P.Lévy). For any t > 0 and θ ∈ [0, 1],
√ ˆθ
2 1 ds
P0 [At /t ≤ θ] = arcsin θ = p .
π π s(1 − s)
0
2
π
0.5 1.0
Figure 7.3: Density of At /t.

Note that the theorem shows in particular that a law of large numbers does not hold!
Indeed, for each ε > 0,
 t 
ˆ
1 1
P0  I(0,∞) (Bs ) ds − > ε 6→ 0 as t → ∞.
t 2
0
Even for large times, values of At /t close to 0 or 1 are the most probable. By the func-
tional central limit theorem, the proportion of time that one player is ahead in a long
coin tossing game or a counting of election results is also close to the arcsine law. In
particular, it is more then 20 times more likely that one player is ahead for more than
98% of the time than it is that each player is ahead between 49% and 51% of the time
[Steele].
Before proving the arc-sine law, we give an informal derivation based on the time-
dependent Feynman-Kac formula.
The idea for determining the distribution of At is again to consider the Laplace trans-
forms
u(t, x) = Ex [exp(−βAt )], β > 0.
By the Feynman-Kac formula, we could expect that u solves the equation
∂u 1 ∂2u
= (7.3.6)
∂t 2 ∂x2
with initial condition u(0, x) = 1. To solve the parabolic p.d.e. (7.3.6), we consider
another Laplace transform: The Laplace transform
∞ 
ˆ∞ ˆ
vα (x) = e−αt u(t, x) dt = Ex  e−αt−βAt dt , α > 0,
0 0
of a solution u(t, x) of (7.3.6) w.r.t. t. An informal computation shows that vα should

satisfy the o.d.e.
ˆ∞
1 ′′ −αt 1 ∂2u
v − βI(0,∞) vα = e − βI(0,∞) u (t, •) dt
2 α 2 ∂x2
0
ˆ∞ ˆ∞
∂u
= e−αt (t, •) dt = e−αt u(t, •)|∞
0 −α e−αt u(t, •) dt
∂t
0 0
= 1 − αvα ,

i.e., vα should be a bounded solution of

1
αvα − vα′′ + βI0,∞ vα = g (7.3.7)
2
where g(x) = 1 for all x. The solution of (7.3.7) can then be computed explicitly, and
yield the arc-sine law by Laplace inversion.
Remark. The method of transforming a parabolic p.d.e. by the Laplace transform into
an elliptic equation is standard and used frequently. In particular, the Laplace trans-
form of a transition semigroup (pt )t≥0 is the corresponding resolvent (gα )α≥0 , gα =
´ ∞ −αt
0
e pt dt, which is crucial for potential theory.
Instead of trying to make the informal argument above rigorous, one can directly prove
the arc-sine law by applying the stationary Feynman-Kac formula:
Exercise. Prove Lévy’s arc-sine law by proceeding in the following way:
(1). Let g ∈ Cb (R). Show that if vα is a bounded solution of (7.3.7) on R \ {0} with
vα ∈ C 1 (R) ∩ C 2 (R \ {0}) then
∞ 
ˆ
vα (x) = Ex  g(Bt )e−αt−βAt dt for any x ∈ R.
0
(2). Compute a corresponding solution vα for g ≡ 1, and conclude that

ˆ∞
1
e−αt E0 [e−βAt ] dt = p .
α(α + β)
0
(3). Now use the uniqueness of the Laplace inversion to show that the distribution µt
of At /t under P• is absolutely continuous with density
1
fAt /t (s) = p .
π s · (1 − s)

Chapter 8
Stochastic Differential Equations:

Explicit Computations
Suppose that (Bt )t≥0 is a given Brownian motion defined on a probability space (Ω, A, P ).
We will now study solutions of stochastic differential equations (SDE) of type
dXt = b(t, Xt ) dt + σ(t, Xt ) dBt (8.0.1)
where b and σ are continuous functions defined on R+ × Rd or an appropriate subset.
Recall that FtB,P denotes the completion of the filtration FtB = σ(Bs | 0 ≤ s ≤ t)
generated by the Brownian motion. Let T be an (FtB,P ) stopping
time.
We call a
process (t, ω) 7→ Xt (ω) defined for t < T (ω) adapted w.r.t. FtB,P , if the trivially
extended process Xet = Xt · I{t<T } defined by

 Xt for t < T
et :=
X ,
0 for t ≥ T
is (FtB,P )-adapted.
245
246 CHAPTER 8. SDE: EXPLICIT COMPUTATIONS
Definition. An almost surely continuous stochastic process (t, ω) 7→ Xt (ω) defined for
t ∈ [0, T (ω)) is called a strong solution of the stochastic differential equation (8.0.1) if
it is (FtB,P )-adapted, and the equation
ˆt ˆt
Xt = X0 + b(s, Xs ) ds + σ(s, Xs ) dBs for t ∈ [0, T ) (8.0.2)
0 0
holds P -almost surely.
The terminology “strong” solution will be explained later when we introduce “weak”
solutions. The point is that a strong solution is adapted w.r.t. the filtration (FtB,P ) gener-
ated by the Brownian motion. Therefore, a strong solution is essentially (up to modifi-
cation on measure zero sets) a measurable function of the given Brownian motion! The
concept of strong and weak solutions of SDE is not related to the analytic definition of
strong and weak solutions for partial differential equations.
In this section we study properties of solutions and we compute explicit solutions for
one-dimensional SDE. We start with an example:
Example (Asset price model in continuous time). A nearby model for an asset price
process (Sn )n=0,1,2,... in discrete time is to define Sn recursively by
Sn+1 − Sn = αn (S0 , . . . , Sn )Sn + σn (S0 , . . . , Sn )Sn ηn+1
with i.i.d. random variables ηi , i ∈ N, and measurable functions αn , σn : Rn → R.

Trying to set up a corresponding model in continuous time, we arrive at the stochastic
differential equation
dSt = αt St dt + σt St dBt (8.0.3)
with an (Ft )-Brownian motion (Bt ) and (FtP ) adapted continuous stochastic processes
(αt )t≥0 and (σt )t≥0 , where (Ft ) is a given filtration on a probability space (Ω, A, P ).
The processes αt and σt describe the instantaneous mean rate of return and the volatility.
Both are allowed to be time dependent and random.

247
In order to compute a solution of (8.0.3), we assume St > 0 for any t ≥ 0, and divide
the equation by St :
1
dSt = αt dt + σt dBt . (8.0.4)
St
We will prove in Section 8.1 that if an SDE holds then the SDE multiplied by a contin-
uous adapted process also holds, cf. Theorem 8.1. Hence (8.0.4) is equivalent to (8.0.3)
if St > 0. If (8.0.4) would be a classical ordinary differential equation then we could
1
use the identity d log St = St
dSt to solve the equation. In Itô calculus, however, the
classical chain rule is violated. Nevertheless, it is still useful to compute d log St by
Itô’s formula. The process (St ) has quadratic variation
 • 
ˆ ˆt
[S]t =  σr Sr dBr  = σr2 Sr2 dr for any t ≥ 0,
0 t 0
almost surely along an appropriate sequence (πn ) of partitions with mesh(πn ) → 0. The
´t
first equation holds by (8.0.3), since t 7→ αr Sr dr has finite variation, and the second
0
identity is proved in Theorem 8.1 below. Therefore, Itô’s formula implies:
1 1
d log St = dSt − 2 d[S]t
St 2St
1
= αt dt + σt dBt − σt2 dt
2
= µt dt + σt dBt ,
where µt := αt − σt2 /2, i.e.,

ˆt ˆt
log St − log S0 = µs ds + σs dBs ,
0 0
or, equivalently,  t 
ˆ ˆt
St = S0 · exp  µs ds + σs dBs  . (8.0.5)
0 0
Conversely, one can verify by Itô’s formula that (St ) defined by (8.0.5) is indeed a
solution of (8.0.3). Thus we have proven existence, uniqueness and an explicit repre-
sentation for a strong solution of (8.0.3). In the special case when αt ≡ α and σt ≡ σ
are constants in t and ω, the solution process

St = S0 exp σBt + (α − σ 2 /2)t

is called a geometric Brownian motion with parameters α and σ.
50
40
30
20
10
1 2 3 4 5 6 7 8 9 10
Figure 1: Three one dimensional geometric Brownian motions with α2 = 1 and

σ = 0.1 (blue), σ = 1.0 (red) and σ = 2.0 (magenta).
Figure 8.1: Three one dimensional geometric Brownian motions with α2 = 1 and σ =
0.1 (blue), σ = 1.0 (red) and σ = 2.0 (magenta).
8.1 Stochastic Calculus for Itô processes

By definition, any solution of an SDE of the form (8.0.1) is the sum of an absolutely
continuous adapted process and an Itô stochastic integral w.r.t. the underlying Brownian
motion, i.e.,
Xt = At + It for t < T, (8.1.1)
where
ˆt ˆt
At = Ks ds and It = Hs dBs (8.1.2)
0 0
with (Ht )t<T and (Kt )t<T almost surely continuous and (FtB,P )-adapted. A stochas-
tic process of type (8.1.1) is called an Itô process. In order to compute and analyze

8.1. STOCHASTIC CALCULUS FOR ITÔ PROCESSES 249
solutions of SDE we will apply Itô’s formula to Itô processes. Since the absolutely con-
tinuous process (At ) has finite variation, classical Stieltjes calculus applies to this part
of an Itô process. It remains to consider the stochastic integral part (It ):
Stochastic integrals w.r.t. Itô processes

Let (πn ) be a sequence of partitions of R+ with mesh(πn ) → 0. Recall that for t ≥ 0,
ˆt X
It = Hs dBs = lim Hs · (Bs′ ∧t − Bs )
n→∞
0 s∈πn
s<t
w.r.t. convergence in probability on {t < T }, cf. Theorem 5.14.
Theorem 8.1 (Composition rule and quadratic variation). Suppose that T is a pre-
dictable stopping time and (Ht )t<T is almost surely continuous and adapted.
(1). For any almost surely continuous, adapted process (Gt )0≤t<T , and for any t ≥ 0,
X ˆt
lim Gs (Is′ ∧t − Is ) = Gs Hs dBs (8.1.3)
n→∞
s∈πn 0
s<t
with convergence in probability on {t < T }. Moreover, if H is in L2a ([0, a]) and G

is bounded on [0, a] × Ω for some a > 0, then the convergence holds in Mc2 ([0, a])
and thus uniformly for t ∈ [0, a] in the L2 (P ) sense.
(2). For any t ≥ 0, the quadratic variation [I]t along (πn ) is given by
X ˆt
2
[I]t = lim (Is′ ∧t − Is ) = Hs2 ds (8.1.4)
n→∞
s∈πn 0
s<t
w.r.t. convergence in probability on {t < T }.
XXX gleich [I, J] berechnen, Beweis analog

Remark (Uniform convergence). Similarly to the proof of Theorem 5.14 one can show
that there is a sequence of bounded stopping times Tk ր T such that almost surely along
a subsequence, the convergence in (8.1.3) and (8.1.4) holds uniformly on [0, Tk ] for any
k ∈ N.
Proof. (1). We first fix a > 0 and assume that H is in L2a ([0, a)) and G is bounded,
′ ∧t
s´
left-continuous and adapted on [0, ∞) × Ω. Since Is′ ∧t − Is = Hr dBr , we
s
obtain
X ˆt
Gs (Is′ ∧t − Is ) = G⌊r⌋n Hr dBr
s∈πn 0
s<t
where ⌊r⌋n = max{s ∈ πn : s ≤ r} is the next partition point below r.

´t
As n → ∞, the right-hand side converges to Gr Hr dBr in Mc2 ([0, a]) because
0
G⌊r⌋n Hr → Gr Hr in L2 (P ⊗ λ[0,a) ) by continuity of G and dominated conver-
gence.
The assertion in the general case now follows by localization: Suppose (Sk ) and
(Tk ) are increasing sequences of stopping times with Tk ր T and Ht I{t≤Sk } ∈
L2a ([0, ∞)), and let
Tek = Sk ∧ Tk ∧ inf{t ≥ 0 : |Gt | > k} ∧ k.
Then Tek ր T , the process Ht

(k) (k)
:= Ht I{t≤Tk } is in L2a ([0, ∞)) the process Gt :=
Gt I{t≤Tk } is bounded, left-continuous and adapted, and
ˆs
Is = Hr(k) dBr , Gs = G(k)
s for any s ∈ [0, t]
0
holds almost surely on {t ≤ Tek }. Therefore as n → ∞,

X X
Gs (Is′ ∧t − Is ) = G(k)
s (Is′ ∧t − Is )
s∈πn s∈πn
s<t s<t
ˆt ˆt
→ G(k)
r Hr
(k)
dBr = Gr Hr dBr
0 0

uniformly for t ≤ Tek in L2 (P ). The claim follows, since

" #
[
P {t < T } \ {t ≤ Tek } = 0.
k
(2). We first assume that H is in L2a ([0, ∞)), continuous and bounded. Then for s ∈
πn ,
ˆs′ ∧t
δIs = Is′ ∧t − Is = Hr dBr = Hs δBs + Rs(n)
s
′ ∧t
s´
(n)
where Rs := (Hr − H⌊r⌋n ) dBr . Therefore,
s
X X X X
(δIs )2 = Hs2 (δBs )2 + 2 Rs(n) Hs δBs + (Rs(n) )2 .
s∈πn s∈πn s∈πn s∈πn
s<t s<t s<t s<t
Since [B]t = t almost surely, the first term on the right-hand side converges
´t
to Hs2 ds with probability one. It remains to show that the remainder terms
0
converge to 0 in probability as n → ∞. This is the case, since
s′ ∧t
hX i X Xˆ
E (Rs(n) )2 = E[(Rs(n) )2 ] = E[(Hr − H⌊r⌋n )2 ] dr
s
ˆt
= E[(Hr − H⌊r⌋n )2 ] dr −→ 0
0
P (n)
by the Itô isometry and continuity and boundedness of H, whence (Rs )2 → 0
P (n)
in L1 and in probability, and Rs Hs δBs → 0 in the same sense by the Schwarz
inequality.
For H defined up to a stopping time T , the assertion now follows by a localization
procedure similar to the one applied above.
The theorem and the corresponding composition rule for Stieltjes integrals suggest that
we may define stochastic integrals w.r.t. an Itô process
ˆt ˆt
Xt = X0 + Hs dBs + Ks ds, t < T,
0 0

in the following way:
Definition. Suppose that (Bt ) is a Brownian motion on (Ω, A, P ) w.r.t. a filtration (Ft ),
X0 is an (F0P )-measurable random variable, T is a predictable (FtP )-stopping time,
and (Gt ), (Ht ) and (Kt ) are almost surely continuous, (FtP ) adapted processes defined
for t < T . Then the stochastic integral of (Gt ) w.r.t. (Xt ) is the Itô process defined by
ˆt ˆt ˆt
Gs dXs = Gs Hs dBs + Gs Ks ds, t < T.
0 0 0
By Theorem 8.1, this definition is consistent with a definition by Riemann sum approx-
imations. Moreover, the definition shows that the class of Itô processes w.r.t. a given
Brownian motion is closed under taking stochastic integrals! In particular, strong solu-
tions of SDE w.r.t. Itô processes are again Itô processes.
Calculus for Itô processes

We summarize calculus rules for Itô processes that are immediate consequences of the
definition above and Theorem 8.1: Suppose that (Xt ) and (Yt ) are Itô processes, and
et ) and (Ht ) are adapted continuous process that are all defined up to a stopping
(Gt ), (G
time T . Then the following calculus rules hold for Itô stochastic differentials:
Linearity:
d(X + cY ) = dX + c dY for any c ∈ R,

(G + cH) dX = G dX + cH dX for any c ∈ R.
Composition rule:
dY = G dX ⇒ e dY = GG
G e dX,
Quadratic variation:
dY = G dX ⇒ d[Y ] = G2 d[X],

Itô rule: For any function F ∈ C 1,2 (R+ × R),
∂F ∂F 1 ∂2F
dF (t, X) = (t, X) dX + (t, X) dt + (t, X) d[X]
∂x ∂t 2 ∂x2
All equations are to be understood in the sense that the corresponding stochastic inte-
grals over any interval [0, t], t < T , coincide almost surely.
The proofs are straightforward. For example, if
ˆt
Yt = Y0 + Gs dXs
0
and
ˆt ˆt
Xt = X0 + Ks ds + Hs dBs
0 0
then, by the definition above, for t < T ,

ˆt ˆt
Yt = Y0 + Gs Ks ds + Gs Hs dBs ,
0 0
and hence
ˆt ˆt ˆt ˆt
es dYs =
G es Gs Ks ds +
G es Gs Hs dBs =
G es Gs dXs
G
0 0 0 0
and  
ˆ• ˆt ˆt
[Y ]t =  Gs Hs dBs  = G2s Hs2 ds = G2s d[X]s .
0 t 0 0
Moreover, Theorem 8.1 guarantees that the stochastic integrals in Itô’s formula (which
are limits of Riemann-Itô sums) coincide with the stochastic integrals w.r.t. Itô processes
defined above.
Example (Option Pricing in continuous time I). We again consider the continuous
time asset price model introduced in the beginning of Chapter 8. Suppose an agent is
holding φt units of a single asset with price process (St ) at time t, and he invests the
remainder Vt − φt St of his wealth Vt in the money market with interest rate Rt . We

assume that (φt ) and (Rt ) are continuous adapted processes. Then the change of wealth
in a small time unit should be described by the Itô equation
dVt = φt dSt + Rt (Vt − φt St ) dt.
Similarly to the discrete time case, we consider the discounted wealth process
 
ˆt
Vet := exp − Rs ds Vt .
0
´t
Since t 7→ Rs ds has finite variation, the Itô rule and the composition rule for stochas-
0
tic integrals imply:
   
ˆt ˆt
dVet = exp − Rs ds dVt − exp − Rs ds Rt Vt dt
0
  0 
ˆt ˆt
= exp − Rs ds φt dSt − exp − Rs ds Rt φt St dt
0 0
     
ˆt ˆt
= φt · exp − Rs ds dSt − exp − Rs ds Rt St dt
0 0
= φt dSet ,
where Set is the discounted asset price process. Therefore,

ˆt
Vet − Ve0 = φs dSes ∀t ≥ 0 P -almost surely.
0
As a consequence, we observe that if (Set ) is a (local) martingale under a probability

measure P∗ that is equivalent to P then the discounted wealth process (Vet ) is also a
local martingale under P∗ . A corresponding probability measure P∗ is called an equiv-
alent martingale measure or risk neutral measure, and can be identified by Girsanov’s
theorem, cf. Section 9.3 below. Once we have found P∗ , option prices can be computed
similarly as in discrete time under the additional assumption that the true measure P for
the asset price process is equivalent to P∗ , see Section 9.4.

The Itô-Doeblin formula in R1
We will now apply Itô’s formula to solutions of stochastic differential equations. Let
b, σ ∈ C(R+ × I) where I ⊆ R is an open interval. Suppose that (Bt ) is an (Ft )-
Brownian motion on (Ω, A, P ), and (Xt )0≤t<T is an (FtP )-adapted process with values
in I and defined up to an (FtP ) stopping time T such that the SDE
ˆt ˆt
Xt − X0 = b(s, Xs ) ds + σ(s, Xs ) dBs for any t < T (8.1.5)
0 0
holds almost surely.
Corollary 8.2 (Doeblin 1941, Itô 1944). Let F ∈ C 1,2 (R+ × I). Then almost surely,
ˆt
F (t, Xt ) − F (0, X0 ) = (σF ′ )(s, Xs ) dBs (8.1.6)
0
ˆt
∂F 1 2 ′′ ′
+ + σ F + bF (s, Xs ) ds for any t < T ,
∂t 2
0
where F ′ = ∂F/∂x denotes the partial derivative w.r.t. x.
Proof. Let (πn ) be a sequence of partitions with mesh(πn ) → 0. Since the process t 7→
´t
X0 + b(s, Xs ) ds has sample paths of locally finite variation, the quadratic variation
0
of (Xt ) is given by
 
ˆ• ˆt
[X]t =  σ(s, Xs ) dBs  = σ(s, Xs )2 ds ∀t < T
0 t 0

w.r.t. almost sure convergence along a subsequence of (πn ). Hence Itô’s formula can be
applied to almost every sample path of (Xt ), and we obtain
ˆt ˆt ˆt
∂F 1
F (t, Xt ) − F (0, X0) = F ′ (s, Xs ) dXs + (s, Xs ) ds + F ′′ (s, Xs ) d[X]s
∂t 2
0 0 0
ˆt ˆt ˆt ˆt
∂F 1
= (σF ′ )(s, Xs ) dBs + (bF ′ )(s, Xs ) ds + (s, Xs ) ds + (σ 2 F ′′ )(s, Xs ) ds
∂t 2
0 0 0 0
for all t < T , P -almost surely. Here we have used (8.1.5) and the fact that the Itô
integral w.r.t. X is an almost sure limit of Riemann-Itô sums after passing once more to
an appropriate subsequence of (πn ).
Exercise (Black Scholes partial differential equation). A stock price is modeled by a

geometric Brownian Motion (St ) with parameters α, σ > 0. We assume that the interest
rate is equal to a real constant r for all times. Let c(t, x) be the value of an option at
time t if the stock price at that time is St = x. Suppose that c(t, St ) is replicated by a
hedging portfolio, i.e., there is a trading strategy holding φt shares of stock at time t and
putting the remaining portfolio value Vt − φt St in the money market account with fixed
interest rate r so that the total portfolio value Vt at each time t agrees with c(t, St ).
“Derive” the Black-Scholes partial differential equation
∂c ∂c 1 ∂2c
(t, x) + rx (t, x) + σ 2 x2 2 (t, x) = rc(t, x) (8.1.7)
∂t ∂x 2 ∂x
and the delta-hedging rule
∂c
φt = (t, St ) (=: Delta ). (8.1.8)
∂x
Hint: Consider the discounted portfolio value Vet = e−rt Vt and, correspondingly, the
discounted option value e−rt c(t, St ). Compute the Ito differentials, and conclude that
both processes coincide if c is a solution to (8.1.7) and φt is given by (8.1.8).

8.2. STOCHASTIC GROWTH 257
Martingale problem for solutions of SDE

The Itô-Doeblin formula shows that if (Xt ) is a solution of (8.1.5) then
ˆt
F
Mt = F (t, Xt ) − F (0, X0 ) − (L F )(s, Xs ) ds
0
is a local martingale up to T for any F ∈ C (R+ × I) and

1,2
1
(L F )(t, x) = σ(t, x)2 F ′′ (t, x) + b(t, x)F ′ (t, x).
2
In particular, in the time-homogeneous case and for T = ∞, any solution of (8.1.5)
solves the martingale problem for the operator L F = 21 σ 2 F ′′ +bF ′ with domain C02 (I).
Similarly as for Brownian motion, the martingales identified by the Itô-Doeblin formula
can be used to compute various expectation values for the Itô diffusion (Xt ). In the next
section we will look at first examples.
Remark (Uniqueness and Markov property of strong solutions). If the coefficients
are, for example, Lipschitz continuous, then the strong solution of the SDE (8.1.5) is
unique, and it has the strong Markov property, i.e., it is a diffusion process in the
classical sense (a strong Markov process with continuous sample paths). By the Itô-
Doeblin formula, the generator of this Markov process is an extension of the operator
(L , C02 (I)).
Although in general, uniqueness and the Markov property may not hold for solutions of
the SDE (8.1.5), we call any solution of this equation an Itô diffusion.
8.2 Stochastic growth

In this section we consider time-homogeneous Itô diffusions taking values in the inter-
val I = (0, ∞). They provide natural models for stochastic growth processes, e.g. in
mathematical biology, financial mathematics and many other application fields. Ana-
logue results also hold if I is replaced by an arbitrary non-empty open interval.
Suppose that (Xt )0≤t<T is a strong solution of the SDE
dXt = b(Xt ) dt + σ(Xt ) dBt for t ∈ [0, T ),

X0 = x0 ,

with a given Brownian motion (Bt ), x0 ∈ (0, ∞), and continuous time-homogeneous
coefficients b, σ : (0, ∞) → R. We assume that the solution is defined up to the explo-
sion time
T = sup Tε,r , Tε,r = inf{t ≥ 0 | Xt 6∈ (ε, r)}.
ε,r>0
The corresponding generator is

1
L F = bF ′ + σ 2 F ′′ .
2
Before studying some concrete models, we show in the general case how harmonic func-
tions can be used to compute exit distributions (e.g. ruin probabilities) and to analyze
the asymptotic behaviour of Xt as t ր T .
Scale functions and exit distributions

To determine the exit distribution from a finite subinterval (ε, r) ⊂ (0, ∞) we compute
the harmonic functions of L . For h ∈ C 2 (0, ∞) with h′ > 0 we obtain:
2b ′ 2b
Lh = 0 ⇐⇒ h′′ = − h ⇐⇒ (log h′ )′ = − .
σ2 σ2
Therefore, the two-dimensional vector space of harmonic functions is spanned by the
constant function 1 and by the function
 
ˆx ˆz
2b(y) 
s(x) = exp − dy dz.
σ(y)2
x0 x0
s(x) is called a scale function of the process (Xt ). It is strictly increasing and harmonic
on (0, ∞). Hence we can think of s : (0, ∞) → (s(0), s(∞)) as a coordinate transfor-
mation, and the transformed process s(Xt ) is a local martingale up to the explosion time
T . Applying the martingale convergence theorem and the optional stopping theorem to
s(Xt ) one obtains:
Theorem 8.3. For any ε, r ∈ (0, ∞) with ε < x0 < r we have:

(1). The exit time Tε,r = inf{t ∈ [0, T ) : Xt 6∈ (ε, r)} is almost surely less than T .
s(r) − s(x)
(2). P [Tε < Tr ] = P [XTε,r = ε] = .
s(r) − s(ε)
The proof of Theorem 8.3 is left as an exercise.
Remark. (1). Note that any affine transformation e

s(x) = cs(x) + d with constants
c > 0 and d ∈ R is also harmonic and strictly increasing, and hence a scale
function. The ratio (s(r) − s(x))/(s(r) − s(ε)) is invariant under non-degenerate
affine transformations of s.
(2). The scale function and the ruin probabilities depend only on the ratio b(x)/σ(x)2 .
Recurrence and asymptotics

We now apply the formula for the exit distributions in order to study the asymptotics of
one-dimensional non-degenerate Itô diffusions as t ր T . For ε ∈ (0, x0 ) we obtain
P [Tε < T ] = P [ Tε < Tr for some r ∈ (x0 , ∞)]

s(r) − s(x0 )
= lim P [Tε < Tr ] = lim .
r→∞ r→∞ s(r) − s(ε)
In particular, we have
P [Xt = ε for some t ∈ [0, T )] = P [Tε < T ] = 1
if and only if s(∞) = lim s(r) = ∞.

rր∞
Similarly, one obtains for r ∈ (x0 , ∞):
P [Xt = r for some t ∈ [0, T )] = P [Tr < T ] = 1
if and only if s(0) = lim s(ε) = −∞.

εց0
Moreover,
" #
[ \ s(x0 ) − s(ε)
P [Xt → ∞ as t ր T ] = P {Tr < Tε } = lim lim ,
εց0 rր∞ s(r) − s(ε)
ε>0 r<∞

and
" #
[ \ s(x0 ) − s(ε)
P [Xt → 0 as t ր T ] = P {Tε < Tr } = lim lim .
rր∞ εց0 s(r) − s(ε)
r<∞ ε>0
Corollary 8.4 (Asymptotics of one-dimensional Itô diffusions).
(1). If s(0) = −∞ and s(∞) = ∞, then the process (Xt ) is recurrent, i.e.,
P [Xt = y for some t ∈ [0, T )] = 1 for any x0 , y ∈ (0, ∞).
(2). If s(0) > −∞ and s(∞) = ∞ then lim Xt = 0 almost surely.

tրT
(3). If s(0) = −∞ and s(∞) < ∞ then lim Xt = ∞ almost surely.

tրT
(4). If s(0) > −∞ and s(∞) < ∞ then

s(∞) − s(x0 )
P lim Xt = 0 =
tրT s(∞) − s(0)
and
s(x0 ) − s(0)
P lim Xt = ∞ =
tրT s(∞) − s(0)
Intuitively, if s(0) = −∞, in the natural scale the boundary is transformed to −∞,
which is not a possible limit for the local martingale s(Xt ), whereas otherwise s(0) is
finite and approached by s(Xt ) with strictly positive probability.
Example. Suppose that b(x)/σ(x)2 ≈ γx−1 as x ր ∞ and b(x)/σ(x)2 ≈ δx−1 as

x ց 0 holds for γ, δ ∈ R in the sense that b(x)/σ(x)2 − γx−1 is integrable at ∞ and
b(x)/σ(x)2 − δx−1 is integrable at 0. Then s′ (x) is of order x−2γ as x ր ∞ and of
order x−2δ as x ց 0. Hence
1 1
s(∞) = ∞ ⇐⇒ γ≤ , s(0) = −∞ ⇐⇒ δ≥ .
2 2
1
In particular, recurrence holds if and only if γ ≤ 2
and δ ≥ 21 .

More concrete examples will be studied below.
Remark (Explosion in finite time, Feller’s test). Corollary 8.4 does not tell us whether
the explosion time T is infinite with probability one. It can be shown that this is always
the case if (Xt ) is recurrent. In general, Feller’s test for explosions provides a necessary
and sufficient condition for the absence of explosion in finite time. The idea is to com-
pute a function g ∈ C(0, ∞) such that e−t g(Xt ) is a local martingale and to apply the
optional stopping theorem. The details are more involved than in the proof of corollary
above, cf. e.g. Section 6.2 in [Durrett: Stochastic calculus].
Geometric Brownian motion

A geometric Brownian motion with parameters α ∈ R and σ > 0 is a solution of the
s.d.e.
dSt = αSt dt + σSt dBt . (8.2.1)
We have already shown in the beginning of Section ?? that for B0 = 0, the unique
strong solution of (8.2.1) with initial condition S0 = x0 is

St = x0 · exp σBt + (α − σ 2 /2)t .
The distribution of St at time t is a lognormal distribution, i.e., the distribution of c · eY

where c is a constant and Y is normally distributed. Moreover, one easily verifies that
(St ) is a time-homogeneous Markov process with log-normal transition densities

1 (log(y/x) − µt)2
pt (x, y) = √ exp − , t, x, y > 0,
2πtσ 2 2tσ 2
where µ = α − σ 2 /2. By the Law of Large Numbers for Brownian motion,

+∞ if µ > 0
lim St = .
t→∞ 0 if µ < 0
If µ = 0 then (St ) is recurrent since the same holds for (Bt ).

We now convince ourselves that we obtain the same results via the scale function:
The ratio of the drift and diffusion coefficient is
b(x) αx α
2
= 2
= 2 ,
σ(x) (σx) σ x

and hence
 
ˆx
2α  = const. · x−2α/σ2 .
s′ (x) = const. · exp − dy
σ2 y
x0
Therefore,
s(∞) = ∞ ⇐⇒ 2α/σ 2 ≤ 1, s(0) = ∞ ⇐⇒ 2α/σ 2 ≥ 1,
which again shows that St → ∞ for α > σ 2 /2, St → 0 for α < σ 2 /2, and St is
recurrent for α = σ 2 /2.
Feller’s branching diffusion

Our second growth model is described by the stochastic differential equation
p
dXt = βXt dt + σ Xt dBt , X0 = x0 , (8.2.2)
with given constants β ∈ R, σ > 0, and values in R+ . Note that in contrast to the
√
equation of geometric Brownian motion, the multiplicative factor Xt in the noise term
is not a linear function of Xt . As a consequence, there is no explicit formula for a
solution of (8.2.2). Nevertheless, a general existence result guarantees the existence of
a strong solution defined up to the explosion time
T = sup TR\(ε,r) ,
ε,r>0
cf. ??. SDEs similar to (8.2.2) appear in various applications.
Example (Diffusion limits of branching processes). We consider a Galton-Watson

branching process Zth with time steps t = 0, h, 2h, 3h, . . . of size h > 0, i.e., Z0h is a
given initial population size, and
h
Zt
X
h
Zt+h = Ni , t/h for t = k · h, k = 0, 1, 2, . . . ,
i=1
with independent identically distributed random variables Ni,k , i ≥ 1, k ≥ 0. The

random variable Zkh
h
describes the size of a population in the k-th generation when Ni,l

is the number of offspring of the i-th individual in the l-th generation. We assume that
the mean and the variance of the offspring distribution are given by
E[Ni,l ] = 1 + βh and Var[Ni,l ] = σ 2
for finite constants β, σ ∈ R.
We are interested in a scaling limit of the model as the size h of time steps goes to 0. To
establish convergence to a limit process as h ց 0 we rescale the population size by h,
i.e., we consider the process
Xth := h · Z⌊t⌋
h
, t ∈ [0, ∞).
The mean growth (“drift”) of this process in one time step is
h
E[Xt+h − Xth | Fth ] = h · E[Zt+h
h
− Zth | Fth ] = hηhZth = hβXth ,
and the corresponding condition variance is
h
Var[Xt+h − Xth | Fth ] = h2 · Var[Zt+h
h
− Zth | Fth ] = h2 σ 2 Zth = hσ 2 Xth ,
where Fth = σ(Ni,l | i ≥ 1, 0 ≤ l ≤ k) for t = k · h. Since both quantities are of order

O(h), we can expect a limit process (Xt ) as h ց 0 with drift coefficient β · Xt and
√
diffusion coefficient σ 2 Xt , i.e., the scaling limit should be a diffusion process solving
a s.d.e. of type (8.2.2). A rigorous derivation of this diffusion limit can be found e.g. in
Section 8 of [Durrett: Stochastic Calculus].
We now analyze the asymptotics of solutions of (8.2.2). The ratio of drift and diffusion
√
coefficient is βx/(σ x)2 = β/σ, and hence the derivative of a scale function is
s′ (x) = const. · exp(−2βx/σ).
Thus s(0) is always finite, and s(∞) = ∞ if and only if β ≤ 1. Therefore, by Corollary
8.4, in the subcritical and critical case β ≤ 1, we obtain
lim Xt = 0 almost surely,

tրT

whereas in the supercritical case β > 1,

P lim Xt = 0 > 0 and P lim Xt = ∞ > 0.
tրT tրT
This corresponds to the behaviour of Galton-Watson processes in discrete time. It can

be shown by Feller’s boundary classification for one-dimensional diffusion processes
that if Xt → 0 then the process actually dies out almost surely in finite time, cf. e.g.
Section 6.5 in [Durrett: Stochastic Calculus]. On the other hand, for trajectories with
Xt → ∞, the explosion time T is almost surely infinite and Xt grows exponentially as
t → ∞.
Cox-Ingersoll-Ross model
The CIR model is a model for the stochastic evolution of interest rates or volatilities.
The equation is
p
dRt = (α − βRt ) dt + σ Rt dBt R0 = x0 , (8.2.3)
with a one-dimensional Brownian motion (Bt ) and positive constants α, β, σ > 0. Al-
though the s.d.e. looks similar to the equation for Feller’s branching diffusion, the
behaviour of the drift coefficient near 0 is completely different. In fact, the idea is that
the positive drift α pushes the process away from 0 so that a recurrent process on (0, ∞)
is obtained. We will see that this intuition is true for α ≥ σ 2 /2 but not for α < σ 2 /2.
Again, there is no explicit solution for the s.d.e. (8.13), but existence of a strong solution
holds. The ratio of the drift and diffusion coefficient is (α − βx)/σ 2 x, which yields
2 2
s′ (x) = const. · x−2α/σ e2βx/σ .
Hence s(∞) = ∞ for any β > 0, and s(0) = ∞ if and only if 2α ≥ σ 2 . Therefore, the
CIR process is recurrent if and only if α ≥ σ 2 /2, whereas Xt → 0 as t ր T almost
surely otherwise.
By applying Itô’s formula one can now prove that Xt has finite moments, and compute
the expectation and variance explicitly. Indeed, taking expectation values in the s.d.e.
ˆt ˆt p
Rt = x0 + (α − βRs ) ds + σ Rs dBs ,
0 0

8.3. LINEAR SDE WITH ADDITIVE NOISE 265
we obtain informally
d
E[Rt ] = α − βE[Rt ],
dt
and hence by variation of constants,
α
E[Rt ] = x0 · e−βt + (1 − e−βt ).
β
´t √
To make this argument rigorous requires proving that the local martingale t 7→ σ Rs dBs
0
is indeed a martingale:
Exercise. Consider a strong solution (Rt )t≥0 of (8.13) for α ≥ σ 2 /2.
(1). Show by applying Itô’s formula to x 7→ |x|p that E[|Rt |p ] < ∞ for any t ≥ 0 and
p ≥ 1.
(2). Compute the expectation of Rt , e.g. by applying Itô’s formula to eβt x.
(3). Proceed in a similar way to compute the variance of Rt . Find its asymptotic value
lim Var[Rt ].
t→∞
8.3 Linear SDE with additive noise
We now consider stochastic differential equations of the form
dXt = βt Ct dt + σt dBt , X0 = x, (8.3.1)
where (Bt ) is a Brownian motion, and the coefficients are deterministic continuous
functions β, σ : [0, ∞) → R. Hence the drift term βt Xt is linear in Xt , and the diffusion
coefficient does not depend on Xt , i.e., the noise increment σt dBt is proportional to
white noise dBt with a proportionality factor that does not depend on Xt .

Variation of constants
An explicit strong solution of the SDE (8.3.1) can be computed by a “variation of con-
stants” Ansatz. We first note that the general solution in the deterministic case σt ≡ 0 is
given by  t 
ˆ
Xt = const. · exp  βs ds .
0
To solve the SDE in general we try the ansatz

 t 
ˆ
Xt = Ct · exp  βs ds
0
with a continuous Itô process (Ct ) driven by the Brownian motion (Bt ). By the Itô
product rule,  t 
ˆ
dXt = βt Xt dt + exp  βs ds dCt .
0
Hence (Xt ) solves (8.3.1) if and only if

 
ˆt
dCt = exp − βs ds σt dBt ,
0
i.e.,  
ˆt ˆr
Ct = C0 + exp − βs ds σr dBr .
0 0
We thus obtain:
Theorem 8.5. The almost surely unique strong solution of the SDE (8.3.1)with initial
value x is given by
   t 
ˆt ˆt ˆ
Xtx = x · exp − βs ds + exp  βs ds σr dBr .
0 0 r

Note that the theorem not only yields an explicit solution but it also shows that the
solution depends smoothly on the initial value x. The effect of the noise on the solution
is additive and given by a Wiener-Itô integral, i.e., an Itô integral with deterministic
integrand. The average value
 t 
ˆ
E[Xtx ] = x · exp  Bs ds , (8.3.2)
0
coincides with the solution in the absence of noise, and the mean-square deviation from
this solution due to random perturbation of the equation is
 t  t    t 
ˆ ˆ ˆt ˆ
Var[Xtx ] = Var  exp  βs ds σr dBr  = exp 2 βs ds σr2 dr
0 r 0 r
by the Itô isometry.
Solutions as Gaussian processes

We now prove that the solution (Xt ) of a linear s.d.e. with additive noise is a Gaussian
process. We first observe that Xt is normally distributed for any t ≥ 0.
Lemma 8.6. For any deterministic function h ∈ L2 (0, t), the Wiener-Itô integral It =
´t ´t
hs dBs is normally distributed with mean 0 and variance h2s ds.
0 0
P
n−1
Proof. Suppose first that h = ci · I(ti ,ti+1 ] is a step function with n ∈ N, c1 , . . . , cn ∈
i=0
P
n−1
R, and 0 ≤ t0 < t1 < . . . < tn . Then It = ci · (Bti+1 − Bti ) is normally distributed
i=0
with mean zero and variance
n−1
X ˆt
Var[It ] = c2i (ti+1 − ti ) = h2s ds.
i=0 0
In general, there exists a sequence (h(n) )n∈N of step functions such that h(n) → h in
L2 (0, t), and
ˆt ˆt
It = h dB = lim h(n) dB in L2 (Ω, A, P ).
n→∞
0 0

Hence It is again normally distributed with mean zero and

 t 
ˆ ˆt
Var[It ] = lim Var  h(n) dB  = h2 ds.
n→∞
0 0
Theorem 8.7 (Wiener-Itô integrals are Gaussian processes). Suppose that h ∈

´t
L2loc ([0, ∞), R). Then It = hs dBs is a continuous Gaussian process with
0
ˆt∧s
E[It ] = 0 and Cov[It , Is ] = h2r ds for any t, s ≥ 0.
0
Proof. Let 0 ≤ t1 < . . . < tn . To show that (It1 , . . . , Itn ) has a normal distribution it
suffices to prove that any linear combination of the random variables It1 , . . . , Itn is nor-
mally distributed. This holds true since any linear combination is again an Itô integral
with deterministic integrand:
n
X ˆtn X
n
λi Iti = λi · I(0,ti ) (s)hs dBs
i=1 0 i=1
for any n ∈ N and λ1 , . . . , λn ∈ R. Hence (It ) is a Gaussian process with E[It ] = 0

and
Cov[It , Is ] = E[It Is ]
∞ 
ˆ ˆ∞
= E  hr · I(0,t) (r) dBr hr · I(0,s) (r) dBr 
0 0
= (h · I(0,t) , h · I(0,s) )L2 (0,∞)
ˆs∧t
= h2r dr.
0

Example (Brownian motion). If h ≡ 1 then It = Bt . The Brownian motion (Bt ) is a

centered Gaussian process with Cov[Bt , Bs ] = t ∧ s.
More generally, by Theorem 8.7 and Theorem 8.5, any solution (Xt ) of a linear SDE
with additive noise and deterministic (or Gaussian) initial value is a continuous Gaussian
process. In fact by (8.3.1), the marginals of (Xt ) are affine functions of the correspond-
ing marginals of a Wiener-Itô integral:
   
ˆt ˆr
1 
Xtx = · x + hr σr dBr  with hr = exp − βu du .
ht
0 0
Hence all finite dimensional marginals of (Xtx ) are normally distributed with
ˆt∧s
1
E[Xtx ] = x/Ht and Cov[Xtx , Xsx ] = · h2r σr2 dr.
ht hs
0
The Ornstein-Uhlenbeck process

In 1905, Einstein introduced a model for the movement of a “big” particle in a fluid.
Suppose that Vtabs is the absolute velocity of the particle, V t is the mean velocity of the
fluid molecules and Vt = Vtabs − V t is the velocity of the particle relative to the fluid.
Then the velocity approximatively can be described as a solution to an s.d.e.
dVt = −γVt dt + σdBt . (8.3.3)
Here (Bt ) is a Brownian motion in Rd , d = 3, and γ, σ are strictly positive constants

that describe the damping by the viscosity of the fluid and the magnitude of the random
collisions. A solution to the s.d.e. (8.3.3) is called an Ornstein-Uhlenbeck process.
Although it has first been introduced as a model for the velocity of physical Brown-
ian motion, the Ornstein-Uhlenbeck process is a fundamental stochastic process that is
almost as important as Brownian motion for mathematical theory and stochastic model-
ing. In particular, it is a continuous-time analogue of an AR(1) autoregressive process.
Note that (8.3.3) is a system of d decoupled one-dimensional stochastic differential
(i) (i) (i)
equations dVt = −γVt + σdBt . Therefore, we will assume w.l.o.g. d = 1. By the
considerations above, the one-dimensional Ornstein-Uhlenbeck process is a continuous

Gaussian process. The unique strong solution of the s.d.e. (8.3.3) with initial condition
x is given explicitly by
 
ˆt
Vtx = e−γt x + σ eγs dBs  . (8.3.4)
0
In particular,
E[Vtx ] = e−γt x,
and
ˆt∧s
Cov[Vtx , Vsx ] = e−γ(t+s) σ 2 e2γr dr
0
2
σ −γ|t−s|
= (e − e−γ(t+s) ) for any t, s ≥ 0.
2γ
Note that as t → ∞, the effect of the initial condition decays exponentially fast with rate
γ. Similarly, the correlations between Vtx and Vsx decay exponentially as |t − s| → ∞.
The distribution at time t is

σ2
Vtx ∼ −γt
N e x, (1 − e−2γt
) . (8.3.5)
2γ
In particular, as t → ∞
D σ2
Vtx −→ N 0, .
2γ
One easily verifies that N(0, σ 2 /2γ) is an equilibrium for the process: If V0 ∼ N(0, σ 2 /2γ)
and (Bt ) is independent of V0 then
ˆt
Vt = e−γt V0 + σ eγ(s−t) dBs
0
 
2 ˆt
σ −2γt
∼ N 0, e + σ2 e2γ(s−t) ds = N(0, σ 2 /2γ)
2γ
0
for any t ≥ 0.

Theorem 8.8. The Ornstein-Uhlenbeck process (Vtx ) is a time-homogeneous Markov

process w.r.t. the filtration (FtB,P ) with stationary distribution N(0, σ 2 /2γ) and transi-
tion probabilities

−γt σ √
pt (x, A) = P e x + √ 1−e−2γt Z∈A , Z ∼ N(0, 1).
2γ
Proof. We first note that by (8.3.5),

σ √
Vtx ∼ e−γt x + √ 1 − e−2γt Z for any t ≥ 0
2γ
with Z ∼ N(0, 1). Hence,
E[f (Vtx )] = (pt f )(x)
for any non-negative measurable function f : R → R. We now prove a pathwise coun-
terpart to the Markov property: For t, r ≥ 0, by (8.3.4)
 
ˆt ˆt+r
−γ(t+r) 
x
Vt+r = e x + σ e dBs  + σ
γs
eγ(s−t−r) dBs
0 0
ˆr
= e−γr Vtx + σ eγ(u−r) dB u , (8.3.6)
0
where B u := Bt+u − Bt is a Brownian motion that is independent of FtB,P . Hence, the

´r
random variable σ · 0 eγ(u−r) dB u is also independent of FtB,P and, by (8.3.4), it has
the same distribution as the Ornstein-Uhlenbeck process with initial condition 0:
ˆr
σ · eγ(u−r) dB u ∼ Vr0 .
0
Therefore, by (8.3.6), the conditional distribution of Vt+r

x
given FtB,P coincides with the
distribution of the process with initial Vtx at time r:
x
E[f (Vt+r ) | FtB,P ] = E[f (e−γr Vtx (ω) + Vr0 )]
x (ω)
= E[f (VrVt )] = (pr f )(Vtx (ω)) for P -a.e. ω.
This proves that (Vtx ) is a Markov process with transition kernels pr , r ≥ 0.

Remark. The pathwise counterpart of the Markov property used in the proof above is
called cocycle property of the stochastic flow x 7→ Vtx .
The Itô-Doeblin formula can now be used to identify the generator of the Ornstein-
Uhlenbeck process: Taking expectation values, we obtain the forward equation
ˆt
E[F (Vtx )] = F (x) + E[(L F )(Vsx )] ds
0
for any function F ∈ C02 (R) and t ≥ 0, where

1 2 ′′
(L F )(x) = σ f (x) − γxf ′ (x).
2
For the transition function this yields
ˆt
(pt F )(x) = F (x) + (ps L F )(x) for any x ∈ R,
0
whence
ˆt
(pt f )(x) − f (x) 1
lim = lim E[(L f )(Vsx )] ds = (L f )(x)
tց0 t tց0 t
0
by continuity and dominated convergence. This shows that the infinitesimal generator
of the Ornstein-Uhlenbeck process is an extension of the operator (L , C02 (R)).
Change of time-scale
We will now prove that Wiener-Itô integrals can also be represented as Brownian motion
with a coordinate transformation on the time axis. Hence solutions of one-dimensional
linear SDE with additive noise are affine functions of time changed Brownian motions.
´t
We first note that a Wiener-Itô integral It = 0 hr dBr with h ∈ L2loc (0, ∞) is a contin-
uous centered Gaussian process with covariance
ˆt∧s
Cov[It , Is ] = h2r dr = τ (t) ∧ τ (s)
0

where
ˆt
τ (t) := h2r dr = Var[It ]
0
is the corresponding variance process. The variance process should be thought of as an

“internal clock” for the process (It ). Indeed, suppose h > 0 almost everywhere. Then
τ is strictly increasing and continuous, and
τ : [0, ∞) → [0, τ (∞)) is a homeomorphism.
Transforming the time-coordinate by τ , we have
Cov[Iτ −1 (t) , Iτ −1 (s) ] = t ∧ s for any t, s ∈ [0, τ (∞)].
These are exactly the covariance of a Brownian motion. Since a continuous Gaussian
process is uniquely determined by its expectations and covariances, we can conclude:
Theorem 8.9 (Wiener-Itô integrals as time changed Brownian motions). The pro-
es := Iτ −1 (s) , 0 ≤ s < τ (∞), is a Brownian motion, and
cess B
eτ (t)
It = B for any t ≥ 0, P -almost surely.
Proof. Since (B es )0≤s<τ (∞) has the same marginal distributions as the Wiener-Itô in-
es ) is again a continuous centered Gaussian
tegral (It )t≥0 (but at different times), (B
process. Moreover, Cov[Bt , Bs ] = t ∧ s, so that (Bs ) is indeed a Brownian motion.
Note that the argument above is different from previous considerations in the sense that
the Brownian motion (B es ) is constructed from the process (It ) and not vice versa.
This means that we can not represent (It ) as a time-change of a given Brownian motion
(e.g. (Bt )) but we can only show that there exists a Brownian motion (B es ) such that I
e This way of representing stochastic processes w.r.t. Brownian
is a time-change of B.
motions that are constructed from the process corresponds to the concept of weak solu-
tions of stochastic differential equations, where driving Brownian motion is not given a

priori. We return to these ideas in Section 9, where we will also prove that continuous
local martingales can be represented as time-changed Brownian motions.
Theorem 8.9 enables us to represent solution of linear SDE with additive noise by time-
changed Brownian motions. We demonstrate this with an example: By the explicit
formula (8.3.4) for the solution of the Ornstein-Uhlenbeck SDE, we obtain:
Corollary 8.10 (Mehler formula). A one-dimensional Ornstein-Uhlenbeck process Vtx

with initial condition x can be represented as
e 1 2γt )
Vtx = e−γt (x + σ B (e −1)
2γ
et )t≥0 such that B

with a Brownian motion (B e0 = 0.
Proof. The corresponding time change for the Wiener-Itô integral is given by
ˆt
τ (t) = exp(2γs) ds = (exp(2γt) − 1)/2γ.
0
8.4 Brownian bridge

In many circumstances one is interested in conditioning diffusion process on taking a
given value at specified times. A basic example is the Brownian bridge which is Brow-
nian motion conditioned to end at a given point x after time t0 . We now present several
ways to describe and characterize Brownian bridges. The first is based on the Wiener-
Lévy construction and specific to Brownian motion, the second extends to Gaussian
processes, whereas the final characterization of the bridge process as the solution of a
time-homogeneous SDE can be generalized to other diffusion processes. From now on,
we consider a one-dimensional Brownian motion (Bt )0≤t≤1 with B0 = 0 that we would
like to condition on taking a given value y at time 1

8.4. BROWNIAN BRIDGE 275
Wiener-Lévy construction
Recall that the Brownian motion (Bt ) has the Wiener-Lévy representation
∞ X
X
Bt (ω) = Y (ω)t + 2n − 1Yn,k (ω)en,k (t) for t ∈ [0, 1] (8.4.1)
n=0 k=0
where en,k are the Schauder functions, and Y and Yn,k (n ≥ 0, k = 0, 1, 2, . . . , 2n −

1) are independent and standard normally distributed. The series in (8.4.1) converges
almost surely uniformly on [0, 1], and the approximating partial sums are piecewise
linear approximations of Bt . The random variables Y = B1 and
∞ 2X
−1 n
X
Xt := Yn,k en,k (t) = Bt − tB1
n=0 k=0
are independent. This suggests that we can construct the bridge by replacing Y (ω) by
the constant value y. Let
Xty := yt + Xt = Bt + (y − B1 ) · t,
and let µy denote the distribution of the process (Xty )0≤t≤1 on C([0, 1]). The next theo-
rem shows that Xty is indeed a Brownian motion conditioned to end at y at time 1:
Theorem 8.11. The map y 7→ µy is a regular version of the conditional distribution of

(Bt )0≤t≤1 given B1 , i.e.,
(1). µy is a probability measure on C([0, 1]) for any y ∈ R,
(2). P [(Bt )0≤t≤1 ∈ A | B1 ] = µB1 [A] holds P -almost surely for any given Borel
subset A ⊆ C([0, 1]).
(3). If F : C([0, 1]) → R is a bounded and continuous function (w.r.t. the supremum
´
norm on C([0, 1])) then the map y 7→ F dµy is continuous.
The last statement says that <7→ µy is a continuous function w.r.t. the topology of weak
convergence.

Proof. By definition, µy is a probability measure for any y ∈ R. Moreover, for any

Borel set A ⊆ C([0, 1]),
P [(Bt )0≤t≤1 ∈ A | B1 ](ω) = P [(Xt + tB1 ) ∈ A | B1 ](ω)

Bt (ω)
= P [(Xt + tB1 (ω)) ∈ A] = P [(Xt ) ∈ A] = µB1 (ω) [A]
for P -almost every ω by independence of (XT ) and B1 . Finally, if F : C([0, 1]) → R is

continuous and bounded then
ˆ
F dµy = E[F ((yt + Xt )0≤t≤1 )]
is continuous in y by dominated convergence.
Finite-dimensional distributions
We now compute the marginals of the Brownian bridge Xty :
Corollary 8.12. For any n ∈ N and 0 < t1 < . . . < tn < 1, the distribution of
(Xty1 , . . . , Xtyn ) on Rn is absolutely continuous with density
pt1 (0, x1 )pt2 −t1 (x1 , x2 ) · · · ptn −tn−1 (xn−1 , xn )p1−tn (xn , y)
fy (x1 , . . . , xn ) = . (8.4.2)
p1 (0, y)
Proof. The distribution of (Bt1 , . . . , Btn , B1 ) is absolutely continuous with density
fBt1 ,...,Btn ,B1 (x1 , . . . , xn , y) = pt1 (0, x0 )pt2 −t1 (x1 , x2 ) · · · ptn −tn−1 (xn−1 , xn )p1−tn (xn , y).
Since the distribution of (Xty1 , . . . , Xtyn ) is a regular version of the conditional distribu-
tion of (Bt1 , . . . , Btn ) given B1 , it is absolutely continuous with the conditional density
fBt1 ,...,Btn ,B1 (x1 , . . . , xn , y)

fBt1 ,...,Btn |B1 (x1 , . . . , xn |y) = ´ ´
· · · fBt1 ,...,Btn ,B1 (x1 , . . . , xn , y) dx1 · · · dxn
= fy (x1 , . . . , xn ).

In general, any almost surely continuous process on [0, 1] with marginals given by
(8.4.2) is called a Brownian bridge from 0 to y in time 1. A Brownian bridge from x
to y in time t is defined correspondingly for any x, y ∈ R and any t > 0. In fact, this
definition of the bridge process in terms of the marginal distributions carries over from
Brownian motion to arbitrary Markov processes with strictly positive transition densi-
ties. In the case of the Brownian bridge, the marginals are again normally distributed:
Theorem 8.13 (Brownian bridge as a Gaussian process). The Brownian bridge from
0 to y in time 1 is the (in distribution unique) continuous Gaussian process (Xty )t∈[0,1]
with
E[Xty ] = ty and Cov[Xty , Xsy ] = t ∧ s − ts for any s, t ∈ [0, 1]. (8.4.3)
Proof. A continuous Gaussian process is determined uniquely in distribution by its

means and covariances. Therefore, it suffices to show that the bridge Xty = Bt + (y −
B1 )t defined above is a continuous Gaussian process such that (8.4.3) holds. This holds
true: By (8.4.2), the marginals are normally distributed, and by definition, t 7→ Xty is
almost surely continuous. Moreover,
E[Xty ] = E[Bt ] + E[y − B1 ] · t = yt, and

Cov[Xty , Xsy ] = Cov[Bt , Bs ] − t · Cov[B1 , Bs ] − s · Cov[Bt , B1 ] + ts Var[B1 ]
= t ∧ s − ts − st + ts = t ∧ s − ts.
Remark (Covariance as Green function, Cameron-Martin space). The covariances

of the Brownian bridge are given by

t · (1 − s) for t ≤ s,
c(t, s) = Cov[Xty , Xsy ] =
(1 − t) · s for t ≥ s.
The function c(t, s) is the Green function of the operator d2 /dt2 with Dirichlet boundary
conditions on the interval [0, 1]. This is related to the fact that the distribution of the

Brownian bridge from 0 to 0 can be viewed as a standard normal distribution on the

space of continuous paths ω : [0, 1] → R with ω(0) = ω(1) = 0 w.r.t. the Cameron-
Martin inner product
ˆ1
(g, h)H = g ′ (s)h′ (s) ds.
0
The second derivative d2 /dt2 is the linear operator associated with this quadratic from.
SDE for the Brownian bridge

Our construction of the Brownian bridge by an affine transformation of Brownian mo-
tion has two disadvantages:
• It can not be carried over to more general diffusion processes with possibly non-
linear drift and diffusion coefficients.
• The bridge Xty = Bt + t(y − B1 ) does not depend on (Bt ) in an adapted way,
because the terminal value B1 is required to define Xty for any t > 0.
We will now show how to construct a Brownian bridge from a Brownian motion in an
adapted way. The idea is to consider an SDE w.r.t. the given Brownian motion with a
drift term that forces the solution to end at a given point at time 1. The size of the drift
term will be large if the process is still far away from the given terminal point at a time
close to 1. For simplicity we consider a bridge (Xt ) from 0 to 0 in time 1. Brownian
bridges with other end points can be constructed similarly. Since the Brownian bridge
is a Gaussian process, we may hope that there is a linear stochastic differential equation
with additive noise that has a Brownian bridge as a solution. We therefore try the Ansatz
dXt = −βt Xt dt + dBt , X0 = 0 (8.4.4)
with a given continuous deterministic function βt , 0 ≤ t < 1. By variation of constants,

the solution of (8.4.4) is the Gaussian process Xt , 0 ≤ t < 1, given by
 t 
ˆt
1
ˆ
Xt = hr dBr where ht = exp  βs ds .
ht
0 0

The process (Xt ) is centered and has covariances

ˆt∧s
1
Cov[Xt , Xs ] = h2r dr.
ht hs
0
Therefore, (Xt ) is a Brownian bridge if and only if
Cov[Xt , Xs ] = t · (1 − s) for any t ≤ s,
i.e., if and only if

ˆt
1
h2r dr = hs · (1 − s) for any 0 < t ≤ s. (8.4.5)
tht
0
The equation (8.4.5) holds if and only if ht is a constant multiple of 1/1 − t, and in this
case
d h′ 1
log ht = t =
βt = for t ∈ [0, 1].
dt ht 1−t
Theorem 8.14. If (Bt ) is a Brownian motion then the process (Xt ) defined by
ˆt
1−t
Xt = dBr for t ∈ [0, 1], X1 = 0,
1−r
0
is a Brownian bridge from 0 to 0 in time 1. It is the unique continuous process solving

the SDE
Xt
dXt = − dt + dBt for t ∈ [0, 1). (8.4.6)
1−t
Proof. As shown above, (Xt )t∈[0,1) is a continuous centered Gaussian process with the
covariances of the Brownian bridge. Hence its distribution on C([0, 1)) coincides with
that of the Brownian bridge from 0 to 0. In particular, this implies lim Xt = 0 almost
tր1
surely, so the trivial extension from [0, 1) to [0, 1] defined by X1 = 0 is a Brownian
bridge.

If the Brownian bridge is replaced by a more general conditioned diffusion process,

the Gaussian characterization does not apply. Nevertheless, it can still be shown by
different means (the keyword is “h-transform”) that the bridge process solves an SDE
generalizing (8.4.6), cf. ?? below.
8.5 Stochastic differential equations in Rn

We now explain how to generalize our considerations to systems of stochastic differen-
tial equations, or, equivalently, SDE in several dimensions. For the moment, we will not
initiate a systematic study but rather consider some examples. The setup is the follow-
ing: We are given a d-dimensional Brownian motion Bt = (Bt1 , . . . , Btd ). The compo-
nent processes Btk , 1 ≤ k ≤ d, are independent one-dimensional Brownian motions that
drive the stochastic dynamics. We are looking for a stochastic process Xt : Ω → Rn
solving an SDE of the form
d
X
dXt = b(t, Xt ) dt + σk (t, Xt ) dBtk . (8.5.1)
k=1
Here n and d may be different, and b, σ1 , . . . , σd : R+ × Rn → Rn are time-dependent

continuous vector fields on Rn . In matrix notation,
dXt = b(t, Xt ) dt + σ(t, Xt ) dBt (8.5.2)
where σ(t, x) = (σ1 (t, x)σ2 (t, x) · · · σd (t, x)) is an n × d-matrix.
Existence, uniqueness and stability
Assuming Lipschitz continuity of the coefficients, existence, uniqueness and stability of

strong solutions of the SDE (8.5.2) can be shown by similar arguments as for ordinary
differential equations.

8.5. STOCHASTIC DIFFERENTIAL EQUATIONS IN RN 281
Theorem 8.15 (Existence, uniqueness and stability under global Lipschitz condi-
tions). Suppose that b and σ satisfy a global Lipschitz condition of the following form:
For any t0 ∈ R, there exists a constant L ∈ R+ such that
|b(t, x)−b(t, x
e)|+||σ(t, x)−σ(t, x
e)|| ≤ L·|x−e e ∈ Rn . (8.5.3)
x| ∀ t ∈ [0, t0 ], x, x
Then for any initial value x ∈ Rn , the SDE (8.5.2) has a unique (up to equivalence)
strong solution (Xt )t∈[0,∞) such that X0 = x P -almost surely.
et ) are two strong solutions with arbitrary initial conditions,
Furthermore, if (Xt ) and (X
then for any t ∈ R+ , there exists a finite constant C(t) such that
" #
h i
E sup |Xs − X es | ≤ C(t) · E |X0 − X e 0 |2 .
s∈[0,t]
The proof of Theorem 8.15 is outlined in the exercises below. In Section 12.1, we will
prove more general results that contain the assertion of the theorem as a special case. In
particular, we will see that existence up to an explosion time and uniqueness of strong
solutions still hold true if one assumes only a local Lipschitz condition.
The key step for proving stability and uniqueness is to control the deviation

εt := E sup |Xs − X es | 2
s≤t
between two solutions up to time t. Existence of strong solutions can then be shown by
a Picard-Lindelöf approximation based on a corresponding norm:
et ) are strong
Exercise (Proof of stability and uniqueness). Suppose that (Xt ) and (X
solutions of (8.5.2), and let t0 ∈ R+ . Apply Itô’s isometry and Gronwall’s inequality to
show that if (8.5.3) holds, then there exists a finite constant C ∈ R+ such that for any
t ≤ t0 ,
ˆ t
εt ≤ C · ε0 + εs ds , and (8.5.4)
0
εt ≤ C · eCt ε0 . (8.5.5)

Hence conclude that two strong solutions with the same initial value coincide almost
surely.
Exercise (Existence of strong solutions). Define approximate solutions of (8.5.2) with

initial value x ∈ Rn inductively by setting Xt0 := x for all t, and
ˆ t ˆ t
n+1 n
Xt := x + b(s, Xs ) ds + σ(s, Xsn ) dBs .
0 0
Let ∆nt := E[sups≤t |Xsn+1 − Xsn |2 ]. Show that if (8.5.3) holds, then for any t0 ∈ R+ ,
there exists a finite constant C(t0 ) such that
ˆ t
n+1
∆t ≤ C(t0 ) ∆ns ds for any n ≥ 0 and t ≤ t0 , and
0
tn 0
∆nt ≤ ∆ C(t0 )n
for any n ∈ N and t ≤ t0 .
n! t
Hence conclude that the limit Xs = limn→∞ Xsn exists uniformly for s ∈ [0, t0 ] with
probability one, and X is a strong solution of (8.5.2) with X0 = x.
Itô processes driven by several Brownian motions

Any solution to the SDE (8.5.1) is an Itô process pf type
ˆt d ˆ
X
t
Xt = Gs ds + Hsk dBsk (8.5.6)

0 k=1 0
with continuous (FtB,P ) adapted stochastic processes Gs , Hs1, Hs2 , . . . , Hsd. We now
extend the stochastic calculus rules to such Itô processes that are driven by several in-
dependent Brownian motions. Let Hs and H e s be continuous (FtB,P ) adapted processes.
Lemma 8.16. If (πn ) is a sequence of partitions of R+ with mesh(πn ) → 0 then for

´t
any 1 ≤ k, l ≤ d and a ∈ R+ , the covariation of the Itô integrals t 7→ Hs dBsk and
0
´t
t 7→ e s dBsl exists almost surely uniformly for t ∈ [0, a] along a subsequence of (πn ),
H
0
and  
ˆ• ˆ• ˆt ˆt
 H dB k , e dB l  =
H e d[B k , B l ] = δkl
HH e s ds.
Hs H
0 0 t 0 0

The proof is an extension of the proof of Theorem 8.1(ii), where the assertion has been
derived for k = l and H = H.e The details are left as an exercise.
Similarly to the one-dimensional case, the lemma can be used to compute the covariation
of Itô integrals w.r.t. arbitrary Itô processes. If Xs and Ys are Itô processes as in (8.5.1),
and Ks and Ls are adapted and continuous then we obtain
ˆ • ˆ • ˆ t
K dX, L dY = Ks Ls d[X, Y ]s
0 0 t 0
almost surely uniformly for t ∈ [0, u], along an appropriate subsequence of (πn ).
Multivariate Itô-Doeblin formula
We now assume again that (Xt )t≥0 is a solution of a stochastic differential equation of
the form (8.5.1). By Lemma 8.16, we can apply Itô’s formula to almost every sample
path t 7→ Xt (ω):
Theorem 8.17 (Itô-Doeblin). Let F ∈ C 1,2 (R+ × Rn ). Then almost surely,
ˆt
F (t, Xt ) = F (0, X0) + (σ ⊤ ∇x F )(s, Xs ) · dBs
0
ˆt
∂F
+ +LF (s, Xs ) ds for all t ≥ 0,
∂t
0
where ∇x denotes the gradient in the space variable, and

n n
1X ∂2F X ∂F
(L F )(t, x) := ai,j (t, x) (t, x) + bi (t, x) (t, x)
2 i,j=1 ∂xi ∂xj i=1
∂xi
with a(t, x) := σ(t, x)σ(t, x)⊤ ∈ Rn×n .

Proof. If X is a solution to the SDE then

Xhˆ ˆ i
i j
[X , X ]t = σk (s, X) dB , σlj (s, X) dB l
i k
t
k,l
Xˆ t ˆ t
= (σki σlj )(s, X) k
d[B , B ] l
= aij (s, Xs ) ds
k,l 0 0
P
where aij = k σki σkj , i.e.,
a(s, x) = σ(s, x)σ(s, x)T ∈ Rn×n .
Therefore, Itô’s formula applied to the process (t, Xt ) yields

d
∂F 1 X ∂2F
dF (t, X) = (t, X) dt + ∇x F (t, X) · dX + i j
(t, X) d[X i , X j ]
∂t 2 i,j=1 ∂x ∂x
∂F
= (σ T ∇x F )(t, X) · dB + + L F (t, X) dt,
∂t
for any F ∈ C 1,2 (R+ × Rn ).
The Itô-Doeblin formula shows that for any F ∈ C 2 (R+ × Rn ), the process
ˆs
∂F
MsF = F (s, Xs ) − F (0, X0 ) − +LF (t, Xt ) dt
∂t
0
is a local martingale. If σ ⊤ ∇x F is bounded then M F is a global martingale.
Exercise (Drift and diffusion coefficients). Show that the processes

ˆs
Msi = Xsi − X0i − bi (s, Xs ) ds, 1 ≤ i ≤ n,
0
are local martingales with covariations
[M i , M j ]s = ai,j (s, Xs ) for any s ≥ 0, P -almost surely.
The vector field b(s, x) is called the drift vector field of the SDE, and the coefficients
ai,j (s, x) are called diffusion coefficients.

General Ornstein-Uhlenbeck processes

XXX to be included
Example (Stochastic oscillator).
Examples
Example (Physical Brownian motion with external force).
Example (Kalman-Bucy filter).
Example (Heston model for stochastic volatility).

Chapter 9
Change of measure
9.1 Local and global densities of probability measures
A thorough understanding of absolute continuity and relative densities of probability

measures is crucial at many places in stochastic analysis. Martingale convergence yields
an elegant approach to these issues including a proof of the Radon-Nikodym and the
Lebesgue Decomposition Theorem. We first recall the definition of absolute continuity.
Absolute Continuity
Suppose that P and Q are probability measures on a measurable space (Ω, A), and F is
a sub-σ-algebra of A.
Definition. (1). The measure Q is called absolutely continuous w.r.t. P on the σ-

algebra F if and only if Q[A] = 0 for any A ∈ F with P [A] = 0.
(2). The measures Q and P are called singular on F if and only if there exists A ∈ F
such that P [A] = 0 and Q[AC ] = 0.
286
9.1. LOCAL AND GLOBAL DENSITIES OF PROBABILITY MEASURES 287
We use the notations Q ≪ P for absolute continuity of Q w.r.t. P , Q ≈ P for mutual

absolute continuity, and Q P for singularity of Q and P . The definitions above extend

to signed measures.
Example. The Dirac measure δ1/2 is obviously singular w.r.t. Lebesgue measure λ(0,1]
on the Borel σ-algebra B((0, 1]). However, δ1/2 is absolutely continuous w.r.t. λ(0,1]
on each of the σ-algebras Fn = σ(Dn ) generated by the dyadic partitions Dn = {(k ·
S
2−n , (k + 1)2−n ] : 0 ≤ k < 2n }, and B([0, 1)) = σ( Dn ).
The next lemma clarifies the term “absolute continuity.”
Lemma 9.1. The probability measure Q is absolutely continuous w.r.t. P on the σ-

algebra F if and only if for any ε > 0 there exists δ > 0 such that for A ∈ F ,
P [A] < δ ⇒ Q[A] < ε. (9.1.1)
Proof. The “if” part is obvious. If P [A] = 0 and (9.1.1) holds for each ε > 0 with δ
depending on ε then Q[A] < ε for any ε > 0, and hence Q[A] = 0.
To prove the “only if” part, we suppose that there exists ε > 0 such that (9.1.1) does not
hold for any δ > 0. Then there exists a sequence (An ) of events in F such that
Q[An ] ≥ ε and P [An ] ≤ 2−n .
Hence, by the Borel-Cantelli-Lemma,
P [An infinitely often] = 0,
whereas
" # " #
\ [ [
Q[An infinitely often] = Q Am = lim Q Am ≥ ε.
n→∞
n m≥n m≥n
Therefore Q is not absolutely continuous w.r.t. P .
Example (Absolute continuity on R). A probability measure µ on a real interval is

absolutely continuous w.r.t. Lebesgue measure if and only if the distribution function
F (t) = µ[(−∞, t]] satisfies:

288 CHAPTER 9. CHANGE OF MEASURE
For any ε > 0 there exists δ > 0 such that for any n ∈ N and a1 , . . . , an , b1 , . . . , bn ∈ R,
n
X n
X
|bi − ai | < ε ⇒ |F (bi ) − F (ai )| < δ, (9.1.2)
i=1 i=1
cf. e.g. [Billingsley: Probability and Measures].
Definition (Absolutely continuous functions). A function F : (a, b) ⊂ R → R is

called absolutely continuous iff (9.1.2) holds.
The Radon-Nikodym Theorem states that absolute continuity is equivalent to the exis-
tence of a relative density.
Theorem 9.2 (Radon-Nikodym). The probability measure Q is absolutely continuous

w.r.t. P on the σ-algebra F if and only if there exists a non-negative random variable
Z ∈ L1 (Ω, F , P ) such that
ˆ
Q[A] = Z dP for any A ∈ F . (9.1.3)
A
The relative density Z of Q w.r.t. P on F is determined by (9.1.3) uniquely up to modi-

fication on P -measure zero sets. It is also called the Radon-Nikodym derivative or the
likelihood ratio of Q w.r.t. P on F . We use the notation

dQ
Z = ,
dP F
and omit the F when the choice of the σ-algebra is clear.
Example (Finitely generated σ-algebra). Suppose that the σ-algebra F is generated

S
by finitely many disjoint atoms B1 , . . . , Bk with Ω = Bi . Then Q is absolutely
continuous w.r.t. P if and only if for any i,
P [Bi ] = 0 =⇒ Q[Bi ] = 0.

In this case, the relative density is given by

X
dQ Q[Bi ]
= · IBi .
dP F P [Bi ]
i : P [Bi ]>0
From local to global densities

Let (Fn ) be a given filtration on (Ω, A).
Definition (Local absolutely continuity). The measure Q is called locally absolutely

continuous w.r.t. P and the filtration (Fn ) if and only if Q is absolutely continuous w.r.t.
P on the σ-algebra Fn for each n.
Example (Dyadic partitions). Any probability measure on the unit interval [0, 1] is
locally absolutely continuous w.r.t. Lebesgue measure on the filtration Fn = σ(Dn )
generated by the dyadic partitions of the unit interval. The Radon-Nikodym derivative
on Fn is the dyadic difference quotient defined by

dµ µ[((k − 1) · 2−n , k · 2−n )] F (k · 2−n ) − F ((k − 1) · 2−n )
(x) = = (9.1.4)
dλ Fn λ[((k − 1) · 2−n , k · 2−n )] 2−n
for x ∈ ((k − 1)2−n , k2−n ].
N
∞ N
∞
Example (Product measures). If Q = ν and P = µ are infinite products of
i=1 i=1
probability measures ν and µ, and ν is absolutely continuous w.r.t. µ with density ̺,
then Q is locally absolutely continuous w.r.t. P on the filtration
Fn = σ(X1 , . . . , Xn )
generated by the coordinate maps Xi (ω) = ωi . The local relative density is

n
dQ Y
= ̺(Xi )
dP Fn i=1
However, if ν 6= µ, then Q is not absolutely continuous w.r.t. P on F∞ = σ(X1 , X2 , . . .),

Pn
since by the LLN, n−1 IA (Xi ) converges Q almost surely to ν[A] and P -almost
i=1
surely to µ[A].

Now suppose that Q is locally absolutely continuous w.r.t. P on a filtration (Fn ) with
relative densities
dQ
Zn = .
dP Fn
The L1 martingale convergence theorem can be applied to study the existence of a global
density on the σ-algebra
[
F∞ = σ( Fn ).
Let Z∞ := lim sup Zn .
Theorem 9.3 (Convergence of local densities, Lebesgue decomposition).
(1). The sequence (Zn ) of successive relative densities is an (Fn )-martingale w.r.t. P .
In particular, (Zn ) converges P -almost surely to Z∞ , and Z∞ is integrable w.r.t.
P.
(2). The following statements are equivalent:
(a) (Zn ) is uniformly integrable w.r.t. P .

(b) Q is absolutely continuous w.r.t. P on F∞ .
(c) Q[A] = Z∞ dP for any P on F∞ .
´
A
(3). In general, the decomposition Q = Qa + Qs holds with

ˆ
Qa [A] = Z∞ dP, Qs [A] = Q[A ∩ {Z∞ = ∞}]. (9.1.5)
A
Qa and Qs are positive measure with Qa ≪ P and Qs P .

The decomposition Q = Qa + Qs into an absolutely continuous and a singular part is

called the Lebesgue decomposition of the measure Q w.r.t. P on the σ-algebra F∞ .
Proof. (1). For n ≥ 0, the density Zn is in L1 (Ω, Fn , P ), and
EP [Zn ; A] = Q[A] = EP [Zn+1 ; A] for any A ∈ Fn .

Hence Zn = EP [Zn+1 | Fn ], i.e., (Zn ) is a martingale w.r.t. P . Since Zn ≥ 0, the

martingale converges P -almost surely, and the limit is integrable.
(2). (a) ⇒ (c): If (Zn ) is uniformly integrable w.r.t. P , then
Zn = EP [Z∞ | Fn ] P -almost surely for any n,
by the L1 convergence theorem. Hence for A ∈ Fn ,
Q[A] = EP [Zn ; A] = EP [Z∞ ; A].

S
This shows that Q[A] = EP [Z∞ ; A] holds for any A ∈ Fn , and thus for any
S
A ∈ F∞ = σ( Fn ).
(c) ⇒ (b) is evident.
(b) ⇒ (a): If Q ≪ P on F∞ then Zn converges also Q-almost surely to a finite

limit Z∞ . Hence for n0 ∈ N and c > 1,
sup EP [ |Zn | ; |Zn | ≥ c] = sup EP [Zn ; Zn ≥ c] = sup Q[Zn ≥ c]

n n n
≤ max Q[Zn ≥ c] + sup Q[Zn ≥ c]
n<n0 n≥n0
≤ max Q[Zn ≥ c] + Q[Z∞ ≥ c − 1] + sup Q[|Zn − Z∞ | ≥ 1].
n<n0 n≥n0
Given ε > 0, the last summand is smaller than ε/3 for n0 sufficiently large, and
the other two summands on the right hand side are smaller than ε/3 if c is chosen
sufficiently large depending on n0 . Hence (Zn ) is uniformly integrable w.r.t. P .
(3). In general, Qa [A] = EP [Z∞ ; A] is a positive measure on F∞ with Qa ≤ Q,

since for n ≥ 0 and A ∈ Fn ,
Qa [A] = EP [lim inf Zk ; A] ≤ lim inf EP [Zk ; A] = EP [Zn ; A] = Q[A]

k→∞ k→∞
by Fatou’s Lemma and the martingale property. It remains to show that
Qa [A] = Q[A ∩ {Z∞ < ∞}] for any A ∈ F∞ . (9.1.6)
If (9.1.6) holds, then Q = Qa + Qs with Qs defined by (9.1.5). In particular, Qs

is then singular w.r.t. P , since P [Z∞ = ∞] = 0 and Qs [Z∞ = ∞] = 0, whereas

Qa is absolutely continuous w.r.t. P by definition.

Since Qa ≤ Q, it suffices to verify (9.1.6) for A = Ω. Then
(Q − Qa )[A ∩ {Z∞ < ∞}] = (Q − Qa )[Z∞ < ∞] = 0,
and therefore
Q[A ∩ {Z∞ < ∞}] = Qa [A ∩ {Z∞ < ∞}] = Qa [A]
for any A ∈ F∞ .
To prove (9.1.6) for A = Ω we observe that for c ∈ (0, ∞),

Q lim sup Zn < c ≤ lim sup Q[Zn < c] = lim sup EP [Zn ; Zn < c]
n→∞ n→∞ n→∞

≤ EP lim sup Zn · I{Zn <c} ≤ EP [Z∞ ] = Qa [Ω]
n→∞
by Fatou’s Lemma. As c → ∞, we obtain
Q[Z∞ < ∞] ≤ Qa [Ω] = Qa [Z∞ < ∞] ≤ Q[Z∞ < ∞]
and hence (9.1.6) with A = Ω. This completes the proof
As a first consequence of Theorem 9.3, we prove the Radon-Nikodym Theorem on a

separable σ-algebra A. Let P and Q be probability measures on (Ω, A) with Q ≪ P .
Proof of the Radon-Nikodym Theorem for separable σ-algebras. We fix a filtration

S
(Fn ) consisting of finitely generated σ-algebras Fn ⊆ A with A = σ( Fn ). Since
Q is absolutely continuous w.r.t. P , the local densities Zn = dQ/dP |Fn on the finitely
generated σ-algebras Fn exist, cf. the example above. Hence by Theorem 9.3,
ˆ
Q[A] = Z∞ dP for any A ∈ A.
A
The approach above can be generalized to probability measures that are not absolutely
continuous:

Exercise (Lebesgue decomposition, Lebesgue densities). Let P and Q be arbitrary

(not necessarily absolutely continuous) probability measures on (Ω, A). A Lebesgue
density of Q w.r.t. P is a random variable Z : Ω → [0, ∞] such that Q = Qa + Qs with
ˆ
Qa [A] = Z dP, Qs [A] = Q[A ∩ {Z = ∞}] for any A ∈ A.
A
The goal of the exercise is to prove that a Lebesgue density exists if the σ-algebra A is
separable.
(1). Show that if Z is a Lebesgue density of Q w.r.t. P then 1/Z is a Lebesgue density
of P w.r.t. Q. Here 1/∞ := 0 and 1/0 := ∞.
S
From now on suppose that the σ-algebra is separable with A = σ( Fn ) where (Fn ) is
a filtration consisting of σ-algebras generated by finitely many atoms.
(1). Write down Lebesgue densities Zn of Q w.r.t. P on each Fn . Show that
Q[Zn = ∞ and Zn+1 < ∞] = 0 for any n,
and conclude that (Zn ) is a non-negative supermartingale under P , and (1/Zn) is

a non-negative supermartingale under Q.
(2). Prove that the limit Z∞ = lim Zn exists both P -almost surely and Q-almost
surely, and P [Z∞ < ∞] = 1 and Q[Z∞ > 0] = 1.
(3). Conclude that Z∞ is a Lebesgue density of P w.r.t. Q on A, and 1/Z∞ is a

Lebesgue density of Q w.r.t. P on A.
Derivatives of monotone functions

Suppose that F : [0, 1] → R is a monotone and right-continuous function. After an
appropriate linear transformation we may assume that F is non decreasing with F (0) =
0 and F (1) = 1. Let µ denote the probability measure with distribution function F .
By the example above, the Radon-Nikodym derivative of µ w.r.t. Lebesgue measure on
the σ-algebra Fn = σ(Dn ) generated by the n-th dyadic partition of the unit interval
is given by the dyadic difference quotients (9.1.4) of F . By Theorem 9.3, we obtain a
version of Lebesgue’s Theorem on derivatives of monotone functions:

Corollary 9.4 (Lebesgue’s Theorem). Suppose that F : [0, 1] → R is monotone (and

right continuous). Then the dyadic derivative

′ dµ
F (t) = lim (t)
n→∞ dλ
Fn
exists for almost every t and F ′ is an integrable function on (0, 1). Furthermore, if F is
absolutely continuous then
ˆs
F (s) = F ′ (t) dt for all s ∈ [0, 1]. (9.1.7)
0
Remark. Right continuity is only a normalization and can be dropped from the assump-
tions. Moreover, the assertion extends to function of finite variation since these can be
represented as the difference of two monotone functions, cf. ?? below. Similarly, (9.1.7)
also holds for absolutely continuous functions that are not monotone. See e.g. [Elstrodt:
Maß- und Integrationstheorie] for details.
Absolute continuity of infinite product measures

∞
Suppose that Ω = × Si , and
i=1
∞
O ∞
O
Q = νi and P = µi
i=1 i=1
are products of probability measures νi and µi defined on measurable spaces (Si , Si ).

We assume that νi and µi are mutually absolutely continuous for every i ∈ N. Denot-
ing by Xk : Ω → Sk the evaluation of the k-th coordinate, the product measures are
mutually absolutely continuous on each of the σ-algebras
Fn = σ(X1 , . . . , Xn ), n ∈ N,
with relative densities

dQ dP
= Zn and = 1/Zn ,
dP Fn dQ Fn

where n
Y dνi
Zn = (Xi ) ∈ (0, ∞) P -almost surely.
i=1
dµ i
In particular, (Zn ) is a martingale under P , and (1/Zn ) is a martingale under Q. Let

F∞ = σ(X1 , X2 , . . .) denote the product σ-algebra.
Theorem 9.5 (Kakutani’s dichotomy). The infinite product measures Q and P are
either singular or mutually absolutely continuous with relative density Z∞ . More pre-
cisely, the following statements are equivalent:
(1). Q ≪ P on F∞ .
(2). Q ≈ P on F∞ .
r
Q∞ ´ dνi
(3). dµi > 0.
i=1 dµi
P
∞
(4). d2H (νi , µi ) < ∞.
i=1
Here the squared Hellinger distance d2H (νi , µi ) of mutually absolutely continuous prob-
ability measures ν and µ is defined by
s !2
ˆ r !2
1 dν 1 dµ
ˆ
d2H = −1
dµ = −1 dν
2 dµ 2 dν
ˆ s ˆ r
dν dµ
= 1− dµ = 1 − dν.
dµ dν
Remark. (1). If mutual absolutely continuity holds then the relative densities on F∞
are
dQ dP 1
= lim Zn P -almost surely, and = lim Q-almost surely.
dP n→∞ dQ n→∞ Zn
(2). If ν and µ are absolutely continuous w.r.t. a measure dx then

1 p p 2
ˆ ˆ p
2
dH (ν, µ) = f (x) − g(x) dx = 1 − f (x)g(x) dx.
2

dνi
Proof. (1) ⇐⇒ (3): For i ∈ N let Yi := dµi
(Xi ). Then the random variables Yi are
independent under both P and Q with EP [Yi ] = 1, and
Zn = Y 1 · Y 2 · · · Y n .
By Theorem 9.3, the measure Q is absolutely continuous w.r.t. P if and only if the mar-
tingale (Zn ) is uniformly integrable. To obtain a sharp criterion for uniform integrability
we switch from L1 to L2 , and consider the non-negative martingale
√ √ √
p ˆ s
Y1 Y2 Yn dνi
Mn = · ··· with βi = EP [ Yi ] = dµi
β1 β2 βn dµi
under the probability measure P . Note that for n ∈ N,
n
, n !2
Y Y
E[Mn2 ] = E[Yi ]/βi2 = 1 βi .
i=1 i=1
If (3) holds then (Mn ) is bounded in L (Ω, A, P ). Therefore, by Doob’s L2 inequality,

2
the supremum of Mn is in L2 (Ω, A, P ), i.e.,
E[sup |Zn | ] = E[sup Mn2 ] < ∞.
Thus (Zn ) is uniformly integrable and Q ≪ P on F∞ .

Conversely, if (3) does not hold then
n
Y
Zn = MN2 · βi −→ 0 P -almost surely,
i=1
since Mn converges to a finite limit by the martingale convergence theorem. Therefore,

the absolute continuous part Qa vanishes by Theorem 9.3 (3), i.e., Q is singular w.r.t.
P.
Q
∞ P
∞
(3) ⇐⇒ (4): For reals βi ∈ (0, 1), the condition βi > 0 is equivalent to (1−βi ) <
i=1 i=1
∞. For βi as above, we have
ˆ s
dνi
1 − βi = 1 − dµi = d2H (νi , µi ).
dµi
(2) ⇒ (1) is obvious.
(4) ⇒ (2): Condition (4) is symmetric in νi and µi . Hence, if (4) holds then both Q ≪ P
and P ≪ Q.

N
∞ N
∞
Example (Gaussian products). Let P = N(0, 1) and Q = N(ai , 1) where
i=1 i=1
(ai )i∈N is a sequence of reals. The relative density of the normal distributions νi :=
N(ai , 1) and µ := N(0, 1) is
dνi exp(−(x − ai )2 )/2
(x) = = exp(ai x − a2i /2),
dµ exp(−x2 /2)
and
ˆ s ˆ∞
dνi 1 1 2
dµ = √ exp − (x − ai x + ai /2) dx = exp(−a2i /8).
2
dµ 2π 2
−∞
Therefore, by condition (3) in Theorem 9.5,

∞
X
Q ≪ P ⇐⇒ Q ≈ P ⇐⇒ a2i < ∞.
i=1
Hence mutual absolute continuity holds for the infinite products if and only if the se-
quence (ai ) is contained in ℓ2 , and otherwise Q and P are singular.
Remark (Relative entropy). (1). In the singular case, the exponential rate of degen-
eration of the relative densities on the σ-algebras Fn is related to the relative
entropies
dνi dνi dνi
ˆ ˆ
H(νi | µi ) = log dµi = log dνi .
dµi dµi dµi
For example in the i.i.d. case µi ≡ µ and νi ≡ ν, we have
n
1 1X dν
log Zn = log (Xi ) −→ H(ν | µ) Q-a.s., and
n n i=1 dµ
1 1
− log Zn = log Z −1 −→ H(µ | ν) P -a.s.
n n
as n → ∞ by the Law of Large Numbers.
P
n P
n
In general, log Zn − H(νi |µi ) is a martingale w.r.t. Q, and log Zn + H(νi |µi )
i=1 i=1
is a martingale w.r.t. P .
(2). The relative entropy is related to the squared Hellinger distance by the inequality
1
H(ν | µ) ≥ d2H (ν | µ),
2
which follows from the elementary inequality
1 √ √
log x−1 = − log x ≥ 1 − x for x > 0.
2

9.2 Translations of Wiener measure

We now return to stochastic processes in continuous time. We endow the continuous
path space C([0, ∞), Rd) with the σ-algebra generated by the evolution maps Xt (ω) =
ω(t), and with the filtration
FtX = σ(Xs | s ∈ [0, t]), t ≥ 0.
Note that FtX consists of all sets of type

ω ∈ C([0, ∞), Rd ) : ω|[0,t] ∈ Γ with Γ ∈ B(C([0, t], Rd )).
In many situations one is interested in the distribution on path space of a process
Bth = Bt + h(t)
h(t)
Bt + h(t)
Bt
obtained by translating a Brownian motion (Bt ) by a deterministic function h : [0, ∞) →

Rd . In particular, it is important to know if the distribution of (Bth ) has a density w.r.t.
the Wiener measure on the σ-algebras FtX , and how to compute the densities if they
exist.
Example. (1). Suppose we would like to evaluate the probability that sup |Bs −
s∈[0,t]
g(s)| < ε for a given t > 0 and a given function g ∈ C([0, ∞), Rd) asymptotically
as ε ց 0. One approach is to study the distribution of the translated process
Bt − g(t) near 0.

9.2. TRANSLATIONS OF WIENER MEASURE 299
(2). Similarly, computing the passage probability P [Bs ≥ a+bs for some s ∈ [0, t]]
to a line s 7→ a + bs for a one-dimensional Brownian motion is equivalent to
computing the passage probability to the point a for the translated process Bt −bt.
(3). A solution to a stochastic differential equation
dYt = dBt + b(t, Yt )dt
is a translation of the Brownian motion Bt − B0 by the stochastic process Ht =

´t
Y0 + b(s, Ys ) ds. Again, in the simplest case (when b(t, y) only depends on t),
0
Ht is a deterministic function.
The Cameron-Martin Theorem

Let (Bt ) denote a continuous Brownian motion with B0 = 0, and let h ∈ C([0, ∞), Rd).
The distribution
µh := P ◦ (B + h)−1
of the translated process Bth = Bt + h(t) is the image of Wiener measure µ0 under the
translation map
τh : C([0, ∞), Rd) −→ C([0, ∞), Rd ), τh (x) = x + h.
Recall that Wiener measure is a Gaussian measure on the infinite dimensional space
C([0, ∞), Rd). The next exercise discusses translations of Gaussian measures in Rn :
Exercise (Translations of normal distributions). Let C ∈ Rn×n be a symmetric non-

negative definite matrix, and let h ∈ Rn . the image of the normal distribution N(0, C)
under the translation map x 7→ x + h on Rn is the normal distribution N(h, C).
(1). Show that if C is non-degenerate then N(h, C) ≈ N(0, C) with relative density
dN(h, C) 1
(x) = e(h,x)− 2 (h,h) for x ∈ Rn , (9.2.1)
dN(0, C)
where (g, h) = (g, C −1, h) for g, h ∈ Rn .

(2). Prove that in general, N(h, C) is absolutely continuous w.r.t. N(0, C) if and only
if h is orthogonal to the kernel of C w.r.t. the Euclidean inner product.
On C([0, ∞), Rd), we can usually not expect the existence of a global density of the
translated measures µh w.r.t. µ0 . The Cameron-Martin Theorem states that for t ≥ 0, a
relative density on FtX exists if and only if h is contained in the corresponding Cameron-
Martin space:
Theorem 9.6 (Cameron, Martin). For h ∈ C([0, ∞), Rd) and t ∈ R+ the translated
measure µh = µ ◦ τh−1 is absolutely continuous w.r.t. Wiener measure µ0 on FtX if and
´t
only if h is an absolutely continuous function on [0, t] with h(0) = 0 and 0 |h′ (s)|2 ds <
∞. In this case, the relative density is given by
ˆ t
dµh 1 t ′
ˆ
′ 2
= exp h (s) dXs − |h (s)| ds . (9.2.2)
dµ0 F X 0 2 0
t
´t
where 0
h′ (s) dXs is the Itô integral w.r.t. the canonical Brownian motion (X, µ0 ).
Before giving a rigorous proof let us explain heuristically why the result should be true.
Clearly, absolute continuity does not hold if h(0) 6= 0, since then the translated paths do
not start at 0. Now suppose h(0) = 0, and fix t ∈ (0, ∞). Absolute continuity on FtX
means that the distribution µth of (Bsh )0≤s≤t on C([0, t], Rd ) is absolutely continuous
w.r.t. Wiener measure µt0 on this space. The measure µt0 , however, is a kind of infinite
dimensional standard normal distribution w.r.t. the inner product
ˆ t
(x, y)H = x′ (s) · y ′ (s) ds
0
on functions x, y : [0, t] → Rd vanishing at 0, and the translated measure µth is a Gaus-

sian measure with mean h and the same covariances.
Choosing an orthonormal basis (ei )i∈N w.r.t. the H-inner product (e.g. Schauder func-
N
∞ N
∞
tions), we can identify µt0 and µth with the product measures N(0, 1) and N(αi , 1)
i=1 i=1

respectively where αi = (h, ei )H is the i-th coefficient of h in the basis expansion.

Therefore, µth should be absolutely continuous w.r.t. µt0 if and only if
∞
X
(h, h)H = αi2 < ∞,
i=1
i.e., if and only if h is absolutely continuous with h′ ∈ L2 (0, t).
Moreover, in analogy to the finite-dimensional case (9.2.1), we would expect informally

a relative density of the form
ˆ t t
dµt 1
ˆ
1
“ ht (x) = e(h,x)H − 2 (h,h)H = exp ′ ′
h (s) · x (s) ds − ′ 2
|h (s)| ds ”
dµ0 0 2 0
Since µt0 -almost every path x ∈ C([0, ∞), Rd) is not absolutely continuous, this ex-
pression does not make sense. Nevertheless, using finite dimensional approximations,
we can derive the rigorous expression (9.2.2) for the relative density where the integral
´t ′ ′ ´t
0
h x ds is replaced by the almost surely well-defined stochastic integral 0 h′ dx :
Proof of Theorem 9.6. We assume t = 1. The proof for other values of t is similar.
Moreover, as explained above, it is enough to consider the case h(0) = 0.
(1). Local densities: We first compute the relative densities when the paths are only
evaluated at dyadic time points. Fix n ∈ N, let ti = i · 2−n , and let
δi x = xti+1 − xti
denote the i-th dyadic increment. Then the increments δi B h (i = 0, 1, . . . , 2n − 1)

of the translated Brownian motion are independent random variables with distri-
butions
δi B h = δi B + δi h ∼ N(δi h, (δt) · Id ), δt = 2−n .
Consequently, the marginal distribution of (Bth1 , Bth2 , . . . , Bth2n ) is a normal distri-

bution with density w.r.t. Lebesgue measure proportional to
n −1
2X
!
|δi x − δi h|2 n
exp − , x = (xt1 , xt2 , . . . , xt2n ) ∈ R2 d .
i=0
2δt

Since the normalization constant does not depend on h, the joint distribution
of (Bth1 , Bth2 , . . . , Bth2n ) is absolutely continuous w.r.t. that of (Bt1 , Bt2 , . . . , Bt2n )
with relative density
2 !
X δi h 1 X δi h
exp · δi x − δt δt . (9.2.3)
δt 2
Consequently, µh is always absolutely continuous w.r.t. µ0 on each of the σ-

algebras
Fn = σ(Xi·2−n : i = 0, 1, . . . , 2n − 1), n ∈ N,
with relative densities

n −1
2X 2n −1 2 !
δi h 1 X δi h
Zn = exp · δi X − δt . (9.2.4)
i=0
δt 2 i=0 δt
(2). Limit of local densities: Suppose that h is absolutely continuous with

ˆ 1
|h′ (t)|2 dt < ∞.
0
We now identify the limit of the relative densities Zn as n → ∞.

First, we note that
n −1
2X
δi h 2 1
ˆ
δt −→ |h′ (t)|2 dt as n → ∞.
δt
i=0 0
In fact, the sum on the right hand side coincides with the squared L2 norm
ˆ 1 2

dh/dt|σ(Dn ) dt
0
of the dyadic derivative

n −1
2X
dh δi h
= · I((i−1)·2−n ,i·2−n ]
dt σ(Dn ) i=0
δt
on the σ-algebra generated by the intervals ((i − 1) · 2 , i · 2 ]. If h is abso-

−n −n
dh
lutely continuous with h′ ∈ L2 (0, 1) then → h′ (t) in L2 (0, 1) by the L2
dt σ(Dn )
martingale convergence theorem.

Furthermore, by Itô’s isometry,

n −1
2X 1
δi h
ˆ
· δi X → h′ (s) dXs in L2 (µ0 ) as n → ∞. (9.2.5)
i=0
δt 0
Indeed,
the sum on the right-hand side is the Itô integral of the step function
dh
w.r.t. X, and as remarked above, these step functions converge to h′ in
dt σ(Dn )
L2 (0, 1). Along a subsequence, the convergence in (9.2.5) holds µ0 -almost surely,
and hence by (9.2.4),
 1 
ˆ1
1
ˆ
lim Zn = exp  h′ (s) dXs − |h′ (s)|2 ds µ0 -a.s. (9.2.6)
n→∞ 2
0 0
(3). Absolute continuity on F1X : We still assume h′ ∈ L2 (0, 1). Note that F1X =
S
σ( Fn ). Hence for proving that µh is absolutely continuous w.r.t. µ0 on F1X with
density given by (9.2.6), it suffices to show that lim sup Zn < ∞ µh -almost surely
(i.e., the singular part in the Lebesgue decomposition of µh w.r.t. µ0 vanishes).
Since µh = µ0 ◦ τh−1 , the process
Wt = Xt − h(t) is a Brownian motion w.r.t. µh ,
and by (9.2.3) and (9.2.4),

n −1
2X 2n −1 2 !
δi h 1 X δi h
Zn = exp · δi W + δt .
i=0
δt 2 i=0 δt
Note that the minus sign in front of the second sum has turned into a plus by the
translation! Arguing similarly as above, we see that along a subsequence, (Zn )
converges µh -almost surely to a finite limit:
 1 
ˆ1
1
ˆ
lim Zn = exp  h′ (s) dWs + |h′ (s)|2 ds µh -a.s.
2
0 0
Hence µh ≪ µ0 with density lim Zn .

(4). Singularity on F1X : Conversely, let us suppose now that h is not absolutely con-
tinuous or h′ is not in L2 (0, 1). Then
n −1
2X ˆ1 2
δi h 2 dh
δt = dt −→ ∞ as n → ∞.
δi t dt
i=0 σ(Dn )
0
Since
2n −1 n −1 2 !1/2
X δ h 2X
δi h
i
· δ X
i = δt ,
δt δt
i=0 L2 (µ0 ) i=0
we can conclude by (9.2.4) that
lim Zn = 0 µ0 -almost surely,
i.e., µh is singular w.r.t. µ0 .
In Section 11.5, we will give an alternative proof of the Cameron-Martin Theorem.
Passage times for Brownian motion with constant drift

We now consider a one-dimensional Brownian motion with constant drift β, i.e., a pro-
cess
Yt = Bt + βt, t ≥ 0,
where Bt is a Brownian motion starting at 0 and β ∈ R. We will apply the Cameron-

Martin Theorem to compute the distributions of the first passage times
TaY = min{t ≥ 0 : Yt = a}, a > 0.
Note that TaY is also the first passage time to the line t 7→ a − βt for the Brownian
motion (Bt ).

9.3. GIRSANOV TRANSFORM 305
Theorem 9.7. For a > 0 and β ∈ R, the restriction of the distribution of TaY to (0, ∞)
is absolutely continuous with density

a(a − βt)2
fa,β (t) = √ exp − .
2πt3 2t
In particular,
ˆ∞
P [TaY < ∞] = fa,β (s) ds.
0
Proof. Let h(t) = βt. By the Cameron-Martin Theorem, the distribution µh of (Yt ) is
absolutely continuous w.r.t. Wiener measure on FtX with density
Zt = exp(β · Xt − β 2 t/2).
Therefore, denoting by Ta = inf{t ≥ 0 : Xt = a} the passage time of the canonical

process, we obtain
P [TaY ≤ t] = µh [Ta ≤ t] = Eµ0 [Zt ; Ta ≤ t]

1
= Eµ0 [ZTa ; Ta ≤ t] = Eµ0 [exp(βa − β 2 Ta ) ; Ta ≤ t]
ˆ t 2
= exp(βa − β 2 s/2) fTa (s) ds
0
by the optional sampling theorem. The claim follows by inserting the explicit expression
for fTa derived in Corollary 1.25.
9.3 Girsanov transform

We will now extend the results in the previous section 9.2 considerably. To this end, we
will consider locally absolutely continuous changes of measure with local densities of
type ˆ
t t
1
ˆ
2
Zt = exp Gs · dXs − |Gs | ds ,
0 2 0

where (Gs ) is an adapted process. Recall that the densities in the Cameron-Martin-
Theorem took this form with the deterministic function Gs = h′ (s). We start with a
general discussion about changing measure on filtered probability spaces that will be
useful in other contexts as well.
Change of measure on filtered probability spaces
Let (Ft ) be a filtration on a measurable space (Ω, A), and fix t0 ∈ (0, ∞). We consider
two probability measures P and Q on (Ω, A) that are mutually absolutely continuous
on the σ-algebra Ft0 with relative density
dP
Zt0 = > 0 Q-almost surely.
dQ Ft0
Then P and Q are also mutually absolutely continuous on each of the σ-algebras Ft ,
t ≤ t0 , with Q- and P -almost surely strictly positive relative densities
dP dQ 1
Zt = = EQ Zt0 Ft and = .
dQ Ft dP Ft Zt
The process (Zt )t≤t0 is a martingale w.r.t. Q, and, correspondingly, (1/Zt )t≤t0 is a mar-
tingale w.r.t. P . From now on, we always choose a right continuous version of these
martingales.
Lemma 9.8. 1) For any 0 ≤ s ≤ t ≤ t0 , and for any Ft -measurable random vari-
able X : Ω → [0, ∞],
EQ [XZt |Fs ] EQ [XZt |Fs ]

EP [X|Fs ] = = P -a.s. and Q-a.s. (9.3.1)
EQ [Zt |Fs ] Zs
2) Suppose that (Mt )t≤t0 is an (Ft ) adapted right continuous stochastic process.
Then
(i) M is a martingale w.r.t. P ⇔ M · Z is a martingale w.r.t. Q,
(ii) M is a local martingale w.r.t. P ⇔ M · Z is a local martingale w.r.t. Q.

Proof. 1) The right hand side of (9.3.1) is Fs -measurable. Moreover, for any A ∈ Fs ,
EP [EQ [XZt |Fs ]/Zs ; A] = EQ [EQ [XZt |Fs ] ; A]

= EQ [XZt ; A] = EQ [X ; A].
2) (i) is a direct consequence of 1). Moreover, by symmetry, it is enough to prove

the implication "⇐" in (ii). Hence suppose that M · Z is a local Q-martingale with
localizing sequence (Tn ). We show that M Tn is a P -martingale, i.e.,
EP [Mt∧Tn ; A] = EP [Ms∧Tn ; A] for any A ∈ Fs , 0 ≤ s ≤ t ≤ t0 . (9.3.2)
To verify (9.3.2), we first note that
EP [Mt∧Tn ; A ∩ {Tn ≤ s}] = EP [Ms∧Tn ; A ∩ {Tn ≤ s}] (9.3.3)
since t ∧ Tn = Tn = s ∧ Tn on {Tn ≤ s}. Moreover, one verifies from the definition of

the σ-algebra Fs∧Tn that for any A ∈ Fs , the event A ∩ {Tn > s} is contained in Fs∧Tn ,
and hence in Ft∧Tn . Therefore,
EP [Mt∧Tn ; A ∩ {Tn > s}] = EQ [Mt∧Tn Zt∧Tn ; A ∩ {Tn > s}] (9.3.4)
= EQ [Ms∧Tn Zs∧Tn ; A ∩ {Tn > s}]] = EP [Ms∧Tn ; A ∩ {Tn > s}]
by the martingale property for (MZ)Tn , the optional sampling theorem, and the fact
that P ≪ Q on Ft∧Tn with relative density Zt∧Tn . (9.3.2) follows from (9.3.3) and
(9.3.4).
Girsanov’s Theorem
We now return to our original problem of identifying the change of measure induced
by a random translation of the paths of a Brownian motion. Suppose that (Xt ) is a
Brownian motion in Rd with X0 = 0 w.r.t. the probability measure Q and the filtration
(Ft ), and fix t0 ∈ [0, ∞). Let
ˆ t
Lt = Gs · dXs , t ≥ 0,
0

´t
with G ∈ L2a,loc R+ , Rd . Then [L]t = 0 |Gs |2 ds, and hence
ˆ t
1
ˆ t
2
Zt = exp Gs · dXs − |Gs | ds (9.3.5)
0 2 0
is the exponential of L. In particular, since L is a local martingale w.r.t. Q, Z is a non-

negative local martingale, and hence a supermartingale w.r.t. Q. It is a Q-martingale for
t ≤ t0 if and only if EQ [Zt0 ] = 1:
Exercise (Martingale property for exponentials). Let (Zt )t∈[0,t0 ] on (Ω, A, Q) be a

non-negative local martingale satisfying Z0 = 1.
a) Show that Z is a supermartingale.
b) Prove that Z is a martingale if and only if EQ [Zt0 ] = 1.
In order to use Zt0 for changing the underlying probability measure on Ft0 we have to
assume the martingale property:
Assumption. (Zt )t≤t0 is a martingale w.r.t. Q.
Theorem 9.10 below implies that the assumption is satisfied if

ˆ t
1 2
E exp |Gs | ds < ∞.
2 0
If the assumption holds then we can consider a probability measure P on A with
dP
= Zt0 Q-a.s. (9.3.6)
dQ Ft0
Note that P and Q are mutually absolutely continuous on Ft for any t ≤ t0 with
dP dQ 1
= Zt and =
dQ Ft dP Ft Zt
both P - and Q-almost surely. We are now ready to prove one of the most important
results of stochastic analysis:

Theorem 9.9 (Maruyama 1954, Girsanov 1960). Suppose that X is a d-dimensional

Brownian motion w.r.t. Q and (Zt )t≤t0 is defined by (9.3.5) with G ∈ L2a,loc (R+ , Rd ). If
EQ [Zt0 ] = 1 then the process
ˆ t
Bt := Xt − Gs ds, t ≤ t0 ,
0
is a Brownian motion w.r.t. any probability measure P on A satisfying (9.3.6).
Proof. By the extension of Lévy’s characterization of Brownian motion to the multi-

dimensional case, it suffices to show that (Bt )t≤t0 is an Rd -valued P -martingale with
[B i , B j ]t = δij t P -almost surely for any i, j ∈ {1, . . . , d}, cf. Theorem 11.2 below.
Furthermore, by Lemma 9.8, and since P and Q are mutually absolutely continuous
on Ft0 , this holds true provided (Bt Zt )t≤t0 is an Rd valued local martingale under Q,
and [B i , B j ] = δij t Q-almost surely. The identity for the covariations holds since (Bt )
differs from the Q-Brownian motion (Xt ) only by a continuous finite variation process.
To show that B · Z is a local Q-martingale, we apply Itô’s formula: For 1 ≤ i ≤ d,
d(B i Z) = B i dZ + Z dB i + d[B i , Z] (9.3.7)

= B i ZG · dX + Z dX i − Z Gi dt + ZGi dt,
where we have used that
d[B i , Z] = ZG · d[B i , X] = ZGi dt Q-almost surely.
The right-hand side of (9.3.7) is a stochastic integral w.r.t. the Q-Brownian motion X,
and hence a local Q-martingale.
The theorem shows that if X is a Brownian motion w.r.t. Q, and Z defined by (9.3.5) is
a Q-martingale, then X satisfies
dXt = Gt dt + dBt .
with a P -Brownian motion B. This can be used to construct weak solutions of stochastic
differential equations by changing the underlying probability measure, see Section 11.3

below. For instance, if we choose Gt = b(Xt ) then the Q-Brownian motion (Xt ) is a
solution to the SDE
dXt = b(Xt ) dt + dBt ,
where B is a Brownian motion under the modified probability measure P .
Furthermore, Girsanov’s Theorem generalizes the Cameron-Martin Theorem to non-

deterministic adapted translations
ˆ t
Xt (ω) −→ Xt (ω) − Ht (ω), Ht = Gs ds,
0
of a Brownian motion X.
Remark (Assumptions in Girsanov’s Theorem).

´t
1) Absolute continuity and adaptedness of the “translation process” Ht = 0
Gs ds are
essential for the assertion of Theorem 9.9.
2) The assumption EQ [Zt0 ] = 1 ensuring that (Zt )t≤t0 is a Q-martingale is not always
satisfied − a sufficient condition is given in Theorem 9.10 below. If (Zt ) is not a martin-
gale w.r.t. Q it can still be used to define a positive measure Pt with density Zt w.r.t. Q
on each σ-algebra Ft . However, in this case, Pt [Ω] < 1. The sub-probability measures
Pt correspond to a transformed process with finite life-time.
Novikov’s condition
To verify the assumption in Girsanov’s theorem, we now derive a sufficient condition
for ensuring that the exponential

Zt = exp Lt − 1/2 [L]t
of a continuous local (Ft ) martingale (Lt ) is a martingale. Recall that Z is always a

non-negative local martingale, and hence a supermartingale w.r.t. (Ft ).

Theorem 9.10 (Novikov 1971). Let t0 ∈ R+ . If E[exp [L]t0 /2 ] < ∞ then (Zt )t≤t0 is
an (Ft ) martingale.

We only prove the theorem under the slightly more restrictive condition
E [exp(p[L]t /2)] < ∞ for some p > 1. (9.3.8)
This simplifies the proof considerably, and the condition is sufficient for many applica-
tions. For a proof in the general case and under even weaker assumptions see e.g. [37].
Proof. Let (Tn )n∈N be a localizing sequence for the martingale Z. Then (Zt∧Tn )t≥0 is a
martingale for any n. To carry over the martingale property to the process (Zt )t∈[0,t0 ] , it
is enough to show that the random variables Zt∧Tn , n ∈ N, are uniformly integrable for
each fixed t ≤ t0 . However, for c > 0 and p, q ∈ (1, ∞) with p−1 + q −1 = 1, we have
E[Zt∧Tn ; Zt∧Tn ≥ c]
p p−1
= E exp Lt∧Tn − [L]t∧Tn exp [L]t∧Tn ; Zt∧Tn ≥ c (9.3.9)
2 2
p2 1/p p−1 1/q
≤ E exp pLt∧Tn − [L]t∧Tn · E exp q · [L]t∧Tn ; Zt∧Tn ≥ c
2 2
p 1/q
≤ E exp [L]t ; Zt∧Tn ≥ c
2
for any n ∈ N. Here we have used Hölder’s inequality and the fact that exp pLt∧Tn −
p2

2
[L]t∧Tn is an exponential supermartingale. If exp 2p [L]t is integrable then the right
hand side of (9.3.9) converges to 0 uniformly in n as c → ∞, because
P [Zt∧Tn ≥ 0] ≤ c−1 E[Zt∧Tn ] ≤ c−1 −→ 0
uniformly in n as c → ∞. Hence {Zt∧Tn : n ∈ N} is indeed uniformly integrable, and

thus (Zt )t∈[0,t0 ] is a martingale.
´t
Example (Bounded drifts). If Lt = 0
Gs · dXs with a Brownian motion (Xt ) and
an adapted process (Gt ) that is uniformly bounded on [0, t] for any finite t then the
´t
quadratic variation [L]t = 0 |Gs |2 ds is also bounded for finite t. Hence exp(L − 12 [L])
is an (Ft ) martingale for t ∈ [0, ∞).
Example (Option pricing in continuous time II: Risk-neutral measure). We con-

sider the asset price model in continuous time introduced in the beginning of Chapter 8.
The stock price is modelled by an SDE
dSt = αt St dt + σt St dXt , (9.3.10)

and the interest rate is given by (Rt ). We assume that (Xt ) is a Brownian motion and
(αt ), (Rt ), (σt ) and (1/σt ) are adapted bounded continuous processes, all defined on a
filtered probability space (Ω, A, Q, (Ft )). Then the discounted asset price
ˆ t
Set := exp − Rs ds St
0
satisfies
dSet = (αt − Rt )Set dt + σt Set dXt = σt Set dBt , (9.3.11)
where t
αs − Rs
ˆ
Bt := Xt + ds.
0 σs
We can apply Girsanov’s Theorem and the Novikov condition to conclude that the pro-
cess (Bt ) is a Brownian motion under a probability measure P on (Ω, A) with local
densities w.r.t. Q on Ft given by
ˆ t 1 t
ˆ
2
Zt = exp Gs · dXs − |Gs | ds where Gt = (Rt − αt )/σt .
0 2 0
Therefore, by (9.3.11) and by the assumptions on the coefficients, the process (Set ) is a
martingale under Q. The measure Q can now be used to compute option prices under
a no-arbitrage assumption similarly to the discrete time case considered in Section 2.3
above, see Section 9.4.
9.4 Itô’s Representation Theorem and Option Pricing

We now prove two basic representation theorems for functionals and martingales that
are adapted w.r.t. the filtration generated by a Brownian motion. Besides their intrin-
sic interest, such representation theorems are relevant e.g. for the theory of financial
markets, and for stochastic filtering. Throughout this section, (Bt ) denotes a Brownian
motion starting at 0 on a probability space (Ω, A, P ), and
Ft = σ(Bs : s ∈ [0, t])P , t ≥ 0,
is the completed filtration generated by (Bt ). It is crucial that the filtration does not con-
tain additional information. By the factorization lemma, this implies that Ft measurable

9.4. ITÔ’S REPRESENTATION THEOREM AND OPTION PRICING 313
random variables F : Ω → R are almost surely functions of the Brownian path (Bs )s≤t .
Indeed, we will show that such functions can be represented as stochastic integrals.
Representation theorems for functions and martingales

The first version of Itô’s Representation Theorem states that random variables that are
measurable w.r.t. the σ-algebra F1 = F1B,P can be represented as stochastic integrals:
Theorem 9.11 (Itô). For any function F ∈ L2 (Ω, F1 , P ) there exists a unique process
G ∈ L2a (0, 1) such that
ˆ 1
F = E[F ] + Gs · dBs P -almost surely. (9.4.1)
0
An immediate consequence of Theorem 9.11 is a corresponding representation for mar-

tingales w.r.t. the Brownian filtration Ft = FtB,P :
Corollary 9.12 (Itô representation for martingales). For any right-continuous L2 -

bounded (Ft ) martingale (Mt )t∈[0,1] there exists a unique process G ∈ L2a (0, 1) such
that ˆ t
Mt = M0 + Gs · dBs for any t ∈ [0, 1], P -a.s.
0
The corollary is of fundamental importance in financial mathematics where it is related

to completeness of financial markets. It also proves the remarkable fact that every
martingale w.r.t. the Brownian filtration has a continuous modification! Of course,
this result can not be true w.r.t. a general filtration.
We first show that the corollary follows from Theorem 9.11, and then we prove the
theorem:

Proof of Corollary 9.12. If (Mt )t∈[0,1] is an L2 bounded (Ft ) martingale then M1 ∈

L2 (Ω, F1 , P ), and
Mt = E[M1 |Ft ] a.s. for any t ∈ [0, 1].
Hence, by Theorem 9.11, there exists a unique process G ∈ L2a (0, 1) such that
ˆ 1 ˆ 1
M1 = E[M1 ] + G · dB = M0 + G · dB a.s.,
0 0
and thus
ˆ t
Mt = E[M1 |Ft ] = M0 + G · dB a.s. for any t ∈ [0, 1].
0
Since both sides in the last equation are almost surely right continuous, the identity
actually holds simultaneously for all t ∈ [0, 1] with probability 1.
e∈
Proof of Theorem 9.11. Uniqueness. Suppose that (9.4.1) holds for two processes G, G
L2a (0, 1). Then ˆ 1 ˆ 1
G · dB = e · dB,
G
0 0
and hence, by Itô’s isometry,
ˆ
e L2 (P ⊗λ) e
||G − G|| = (G − G) · dB = 0.
L2 (P )
et (ω) for almost every (t, ω).

Hence Gt (ω) = G
Existence. We prove the existence of a representation as in (9.4.1) in several steps −

starting with “simple” functions F .
1. Suppose that F = exp(ip · (Bt − Bs )) for some p ∈ Rd and 0 ≤ s ≤ t ≤ 1. By Itô’s
formula,
t
1 1 1
ˆ
exp(ip · Bt + |p|2 t) = exp(ip · Bs + |p|2 s) + exp ip · Br + |p|2 r ip · dBr .
2 2 s 2
Rearranging terms, we obtain an Itô representation for F with a bounded adapted inte-
grand G.
Q
n
2. Now suppose that F = Fk where Fk = exp ipk · (Btk − Btk−1 ) for some n ∈ N,
k=1

p1 , . . . , pn ∈ Rd , and 0 ≤ t0 ≤ t1 ≤ · · · ≤ tn ≤ 1. Denoting by Gk the bounded

adapted process in the Itô representation for Fk , we have
Yn ˆ tk+1
F = E[Fk ] + Gk · dB .
k=1 tk
Qn
We show that the right hand side can be written as the sum of k=1 E[Fk ] and a stochas-
tic integral w.r.t. B. For this purpose, it suffices to verify that the product of two stochas-
´t ´t
tic integrals Xt = 0 G · dB and Yt = 0 H · dB with bounded adapted processes G and
´1
H is the stochastic integral of a process in L2a (0, 1) provided 0 Gt · Ht dt = 0. This
holds true, since by the product rule,
ˆ 1 ˆ 1 ˆ 1
X1 Y 1 = Xt Ht · dBt + Yt Gt · dBt + Gt · Ht dt,
0 0 0
and XH + Y G is square-integrable by Itô’s isometry.
3. Clearly, an Itô representation also holds for any linear combination of functions as in
Step 2.
4. To prove an Itô representation for arbitrary functions in L2 (Ω, F1 , P ), we first note

that the linear combinations of the functions in Step 2 form a dense subspace of the
Hilbert space L2 (Ω, F1 , P ). Indeed, if φ is an element in L2 (Ω, F1 , P ) that is orthogonal
to this subspace then
h Yn i
E φ exp(ipk · (Btk − Btk−1 )) = 0
k=1
for any n ∈ N, p1 , . . . , pn ∈ Rd and 0 ≤ t0 ≤ t1 ≤ · · · ≤ tn ≤ 1. By Fourier inversion,

this implies
E[φ | σ(Btk − Btk−1 : 1 ≤ k ≤ n)] = 0 a.s.
for any n ∈ N and 0 ≤ t0 ≤ · · · ≤ tn ≤ 1, and hence φ = 0 a.s. by the Martingale

Convergence Theorem.
Now fix an arbitrary function F ∈ L2 (Ω, F1 , P ). Then by Step 3, there exists a sequence
(Fn ) of functions in L2 (Ω, F1 , P ) converging to F in L2 that have a representation of
the form ˆ 1
Fn − E[Fn ] = G(n) · dB (9.4.2)
0

with processes G(n) ∈ L2a (0, 1). As n → ∞,
Fn − E[Fn ] −→ F − E[F ] in L2 (P ).
Hence, by (9.4.2) and Itô’s isometry, (G(n) ) is a Cauchy sequence in L2 (P ⊗ λ(0,1) ).

Denoting by G the limit process, we obtain the representation
ˆ 1
F − E[F ] = G · dB
0
by taking the L2 limit on both sides of (9.4.2).
Application to option pricing

We return to the asset price model considered at the end of Section 9.3. For simplicity,
we now assume that the coefficients in (9.3.10) and (9.3.11) are constant:
αt ≡ α ∈ R, σt ≡ σ ∈ (0, ∞), Rt ≡ r ∈ R.
Then the change of measure is given by the local densities

2 !
r−α 1 r−α
Zt = exp Xt − t , (9.4.3)
σ 2 σ
and by (9.3.11), the discounted stock price is proportional to the Itô exponential of σB
α−r
where Bt = Xt + σ
t is a Brownian motion under the risk-neutral measure Q:
Set = S0 · exp(σBt − σ 2 t/2) (9.4.4)
Now suppose that we want to compute the no-arbitrage price of an option. For example,
let us consider a European call option where the payoff at the final time t0 is given by
Vt0 = (St0 − K)+
for a positive constant K. By (9.4.4), the discounted payoff

+
Vet0 = Set0 − e−rt0 K (9.4.5)

is an FtB,P
0
measurable random variable. Therefore, by Itô’s Representation Theorem
and (9.4.4), there exists a process G ∈ L2a (0, t0 ) such that
h i ˆ t0 h i ˆ t0
Vet0 = EP Vet0 + Gr dBr = EP Vet0 + Φr dSer ,
0 0
where Φr := Gr /(σ Ser ). Hence (Φr ) is a replicating strategy for the option, i.e., in-
vesting Φr units in the stock and putting the remaining money on the bank account
yields
h exactly
i the payoff for the option at time t0 provided our initial capital is given by
EP Vet0 . Since otherwise there would be an arbitrage opportunity by selling the option
and investing the gain by the strategy Φ, or conversely, we can conclude that under a
no-arbitrage assumption, the only possible option price at time 0 is given by
h i +
e σBt0 −σ2 t0 /2 −rt0
EP Vt0 = EP S0 e −e K
Noting that Bt0 ∼ N(0, t0 ) under P , we obtain the Black-Scholes formula for the no-
arbitrage price of a European call option. Notice in particular that the price does not
depend on the usually unknown model parameter α (the mean rate of return).
Application to stochastic filtering

XXX to be included

Part III
Stochastic Analysis
318
Appendix
548
Appendix A
Conditional expectations
A.1 Conditioning on discrete random variables

We first consider conditioning on the value of a random variable Y : Ω → S where S is
countable. In this case, we can define the conditional probability measure
P [A ∩ {Y = z}]
P [A | Y = z] = , A ∈ A,
P [Y = z]
and the conditional expectations
E[X; Y = z]
E[X | Y = z] = , X ∈ L1 (Ω, A, P ),
P [Y = z]
for any z ∈ S with P [Y = z] > 0 in an elementary way. Note that for z ∈ S with
P [Y = z] = 0, the conditional probabilities are not defined.
Conditional expectations as random variables
It will turn out to be convenient to consider the conditional probabilities and expecta-
tions not as functions of the outcome z, but as functions of the random variable Y . In
this way, the conditional expectations become random variables:
549
550 APPENDIX A. CONDITIONAL EXPECTATIONS
Definition (Conditional expectation given a discrete random variable). Let X :

Ω → R be a random variable such that E[X − ] < ∞, and let Y : Ω → S be a dis-
crete random variable. The random variable E[X | Y ] that is P -almost surely uniquely
defined by
X
E[X | Y ] := g(Y ) = g(z) · I{Y =z}
z∈S
with 
E[X | Y = z] if P [Y = z] > 0
g(z) :=
arbitrary if P [Y = z] = 0
is called (a version of the) conditional expectation of X given Y . For an event A ∈ A,
the random variable
P [A | Y ] := E[IA | Y ]
is called (a version of the) conditional probability of A given Y .
The conditional expectation E[X | Y ] and the conditional probability P [A | Y ] are again
random variables.They take the values E[X | Y = z] and P [A | Y = z], respectively,
on the sets {Y = z}, z ∈ S with P [Y = z] > 0. On each of the null sets {Y =
z}, z ∈ S with P [Y = z] = 0, an arbitrary constant value is assigned to the conditional
expectation. Hence the definition is only almost surely unique.
Characteristic properties of conditional expectations

Let X : Ω → R be a non-negative or integrable random variable on a probability space
(Ω, A, P ). The following alternative characterisation of the conditional expectation of
X given Y can be verified in an elementary way:
Theorem A.1. A real random variable X ≥ 0 (or X ∈ L1 ) on (Ω, A, P ) is a version

of the conditional expectation E[X | Y ] if and only if
(I) X = g(Y ) for a function g : S → R, and

A.2. GENERAL CONDITIONAL EXPECTATIONS 551

(II) E X · f (Y ) = E[X · f (Y )] for all non-negative or bounded functions f :
S → R, respectively.
A.2 General conditional expectations

If Y is a real-valued random variable on a probability space (Ω, A, P ) with continuous
distribution function, then P [Y = z] = 0 for any z ∈ R. Therefore, conditional
probabilities given Y = z can not be defined in the same way as above. Alternatively,
one could try to define conditional probabilities given Y as limits:
P [A | Y = z] = lim P [A | z − h ≤ Y ≤ z + h]. (A.2.1)

hց0
In certain cases this is possible but in general, the existence of the limit is not guaran-
teed.
Instead, the characterization in Theorem A.1 is used to provide a definition of condi-

tional expectations given general random variables Y . The conditional probability of a
fixed event A given Y can then be defined almost surely as a special case of a conditional
expectation:
P [A | Y ] := E[IA | Y ]. (A.2.2)
Note, however, that in general, the exceptional set will depend on the event A !
The factorization lemma

We first prove an important measure theoretic statement.
Theorem A.2 (Factorization lemma). Suppose that (S, S) is a measurable space and
Y : Ω → S is a map. Then a map X : Ω → R is measurable w.r.t. σ(Y ) if and only if
X = f (Y ) = f ◦ Y

for a S-measurable function f : S → R.
Y
(Ω, σ(Y )) (S, S) (R, B(R))
Proof. (1). If X = f ◦ Y for a measurable function f , then
X −1 (B) = Y −1 (f −1 (B)) ∈ σ(Y ) holds for all B ∈ B(R),
as f −1 (B) ∈ S. Therefore, X is σ(Y )-measurable.
(2). Coversely, we have to show that σ(Y )-measurability of X implies that X is a

measurable function of Y . This is done in several steps:
(a) If X = IA is an indicator function of an set A ∈ σ(Y ), then A = Y −1 (B)

with B ∈ S, and thus
X(ω) = IY −1 (B) (ω) = IB (Y (ω)) for all ω ∈ Ω.

Pn
(b) For X = i=1 ci IAi with Ai ∈ σ(Y ) and ci ∈ R we have correspondingly
n
X
X= ci IBi (Y ),
i=1
where wobei Bi are sets in S such that Ai = Y −1 (Bi ).

(c) For an arbitrary non-negative, σ(Y )-measurable map X : Ω → R, there
exists a sequence of σ(Y )-measurable elementary functions such that Xn ր
X. By (b), Xn = fn (Y ) with S-measurable functions fn . Hence
X = sup Xn = sup fn (Y ) = f (Y ),
where f = sup fn is again S-measurable.

(d) For a general σ(Y )-measurable map X : Ω → R, both X + and X − are
measurable functions of Y , hence X is a measurable function of Y as well.

The factorization lemma can be used to rephrase the characterizing properties (I) und
(II) of conditional expectations in Theorem A.1 in the following way:
X is a version of E[X | Y ] if and only if
(i) X ist σ(Y )-messbar,
(ii) E[X ; A] = E[X ; A] für alle A ∈ σ(Y ).
The equivalence of (I) und (i) is a consequence of the factorization lemma, and the
equivalence of (II) and (ii) follows by monotone classes, since (ii) states that
E[X · IB (Y )] = E[X · IB (Y )] holds for all B ∈ S.
Conditional expecations given σ-algebras

A remarkable consequence of the characterization of conditional expectations by Con-
ditions (i) and (ii) is that the conditional expectation E[X | Y ] depends on the random
variable Y only via the σ-algebra σ(Y ) generated by Y ! If two random variables Y
and Z are functions of each other then σ(Y ) = σ(Z), and hence the conditional expec-
tations E[X | Y ] and E[X | Z] coincide (with probability 1). Therefore it is plausible
to define directly the conditional expectation given a σ-Algebra. The σ-algebra (e.g.
σ(Y ), or σ(Y1 , . . . , Yn )) then describes the available “information” on which we are
conditioning.
The characterization of conditional expectations by (i) and (ii) can be extended immedi-
ately to the case of general conditional expectations given a σ-algebra or given arbitrary
random variables. To this end let X : Ω → R be a non-negative (or integrable) random
variable on a probability space (Ω, A, P ).
Definition (Conditional expectation, general). (1). Let F ⊆ A be a σ-algebra. A

non-negative (or integrable) random variable X : Ω → R is called a version of
the conditional expectation E[X | F ] iff:

(a) X is F -measurable, and

(b) E[X ; A] = E[X ; A] for any A ∈ F .
(2). For arbitrary random variables Y, Y1 , Y2 , . . . , Yn on (Ω, A, P ) we define
E[X | Y ] := E[X | σ(Y )],

E[X | Y1 , . . . Yn ] := E[X | (Y1 , . . . , Yn )] = E[X | σ(Y1 , . . . , Yn )].
(3). For an event A ∈ A we define
P [A | F ] := E[IA | F ], and correspondingly P [A | Y ] = E[IA | Y ].
Remark. By monotone classes it can be shown that Condition (b) is equivalent to:
(b’) E[X · Z] = E[X · Z] for any non-negative (resp. bounded) F -measurable

Z : Ω → R.
Theorem A.3 (Existence and uniqueness of conditional expectations). Let X ≥ 0 or

X ∈ L1 , and let F ⊆ A be a σ-algebra. Then:
(1). There exists a version of the conditional expectation E[X | F ].
(2). Any two versions coincide P -almost surely.
Proof. Existence can be shown as a consequence of the Radon-Nikodym theorem. In

Theorem A.8 below, we give a different proof of existence that only uses elementary
methods.
e be two versions of E[X | F ]. Then both X and X
For proving uniqueness let X and X e
are F -measurable, and
E[X ; A] = e ; A]
E[X for any A ∈ F .
e P -almost surely.
Therefore, X = X

Properties of conditional expectations

Starting form the definition, we now derive several basic properties of conditional ex-
pectations that are used frequently:
Theorem A.4. Let X, Y and Xn (n ∈ N) be non-negative or integrable random vari-

ables on (Ω, A, P ), and let F , G ⊆ A be σ-algebras.
The following assertions hold:
(1). Linearity: E[λX + µY | F ] = λ E[X | F ] + µ E[Y | F ] P -almost surely for
any λ, µ ∈ R.
(2). Monotonicity: If X ≥ 0 P -almost surely, then E[X | F ] ≥ 0 P -almost surely.
(3). If X = Y P -almost surely then E[X | F ] = E[Y | F ] P -almost surely.
(4). Monotone Convergence: If (Xn ) is increasing with X1 ≥ 0, then
E[sup Xn | F ] = sup E[Xn | F ] P -almost surely.
(5). Tower Property: If G ⊆ F then
E[E[X | F ] | G] = E[X | G] P -almost surely.
In particular,
E[E[X | Y, Z] | Y ] = E[X|Y ] P -almost surely.
(6). Taking out what is known: Let Y be F -measurable such that Y · X ∈ L1 or ≥ 0.

Then
E[Y · X | F ] = Y · E[X | F ] P -almost surely.
(7). Independence: If X is independent of F then E[X | F ] = E[X] P -almost surely.
(8). Let (S, S) and (T, T ) be measurable spaces. If Y : Ω → S is F -measurable

and X : Ω → T is independent of F , then for any product-measurable function
f : S × T → [0, ∞) we have
E[f (X, Y ) | F ](ω) = E[f (X, Y (ω))] für P -fast alle ω.

Proof. (1). Aus der Linearität des Erwartungswertes folgt, dass λE[X |F ]+µE[Y |F ]
eine Version der bedingten Erwartung E[λX + µY | F ] ist.
(2). Sei X eine Version von E[X | F ]. Aus X ≥ 0 P -fast sicher folgt wegen {X <
0} ∈ F :
E[X ; X < 0] = E[X ; X < 0] ≥ 0,
und damit X ≥ 0 P -fast sicher.
(3). Dies folgt unmittelbar aus (1) und (2).
(4). Ist Xn ≥ 0 und monoton wachsend, dann ist sup E[Xn | F ] eine nichtnegative
F -messbare Zufallsvariable (mit Werten in [0, ∞]), und nach dem "‘klassischen
"’ Satz von der monotonen Konvergenz gilt:
E[sup E[Xn | F ] · Z] = sup E[E[Xn | F ] · Z] = sup E[Xn · Z] = E[sup Xn · Z]

für jede nichtnegative F -messbare Zufallsvariable Z. Also ist sup E[Xn | F ] eine
Version der bedingten Erwartung von sup Xn gegeben F .
(5). Wir zeigen, dass jede Version von E[X | G] auch eine Version von E[E[X | F ] | G]
ist, also die Eigenschaften (i) und (ii) aus der Definition der bedingten Erwartung
erfüllt:
(i) E[X | G] ist nach Definition G-messbar.

(ii) Für A ∈ G gilt auch A ∈ F , und somit E[E[X | G] ; A] = E[X ; A] =
E[E[X | F ] ; A].
(6) und (7). Auf ähnliche Weise verifiziert man, dass die Zufallsvariablen, die auf der rechten
Seite der Gleichungen in (6) und (7) stehen, die definierenden Eigenschaften der
bedingten Erwartungen auf der linken Seite erfüllen (Übung).
(8). Dies folgt aus (6) und (7) in drei Schritten:
(a) Gilt f (x, y) = g(x) · h(y) mit messbaren Funktionen g, h ≥ 0, dann folgt
nach (6) und (7) P -fast sicher:
E[f (X, Y ) | F ] = E[g(X) · h(Y ) | F ] = h(Y ) · E[g(X)|F ]
= h(Y ) · E[g(X)],

A.3. CONDITIONAL EXPECTATION AS BEST L2 -APPROXIMATION 557
und somit
E[f (X, Y )|F ](ω) = E[g(X)·h(Y (ω))] = E[f (X, Y (ω))] für P -fast alle ω.
(b) Um die Behauptung für Indikatorfunktionen f (x, y) = IB (x, y) von pro-

duktmessbaren Mengen B zu zeigen, betrachten wir das Mengensystem
D = {B ∈ S ⊗ T | Behauptung gilt für f = IB }.
D ist ein Dynkinsystem, das nach (a) alle Produkte B = B1 ×B2 mit B1 ∈ S
und B2 ∈ T enthält. Also gilt auch
D ⊇ σ({B1 × B2 | B1 ∈ S, B2 ∈ T }) = S ⊗T.
(c) Für beliebige produktmessbare Funktionen f : S × T → R+ folgt die Be-

hauptung nun durch maßtheoretische Induktion.
Remark (Convergence theorems for conditional expectations). The Monotone Con-

vergence Theorem (Property (4)) implies versions of Fatou’s Lemma and of the Domi-
nated Convergence Theorem for conditional expectations. The proofs are similar to the
unconditioned case.
The last property in Theorem A.4 is often very useful. For independent random variables
X and Y it implies
E[f (X, Y ) | Y ](ω) = E[f (X, Y (ω))] für P -fast alle ω, (A.2.3)
We stress that independence of X and Y ist essential for (A.2.3) to hold true. The
application of (A.2.3) without independence is a common mistake in computations with
conditional expectations.
A.3 Conditional expectation as best L2-approximation

In this section we show that the conditional expectation of a square integrable random
variable X given a σ-algebra F can be characterized alternatively as the best approxi-
mation of X in the subspace of F -measurable, square integrable random variables, or,

equivalently, as the orthogonal projection of X onto this subspace. Besides obvious

applications to non-linear predictions, this point of view is also the basis for a simple
existence proof of conditional expectations
Jensen’s inequality
Jensen’s inequality is valid for conditional expectations as well. Let (Ω, A, P ) be a
probability space, X ∈ L1 (Ω, A, P ) an integrable random variable, and F ⊆ A a σ-
algebra.
Theorem A.5 (Jensen). If u : R → R is a convex function with u(X) ∈ L1 or u ≥ 0,

then
E[u(X) | F ] ≥ u(E[X | F ]) P -almost surely.
Proof. Jede konvexe Funktion u lässt sich als Supremum von abzählbar vielen affinen
Funktionen darstellen, d.h. es gibt an , bn ∈ R mit
u(x) = sup (an x + bn ) für alle x ∈ R.

n∈N
Zum Beweis betrachtet man die Stützgeraden an allen Stellen einer abzählbaren dichten
Teilmenge von R, siehe z.B. [Williams: Probability with martingales, 6.6]. Wegen der
Monotonie und Linearität der bedingten Erwartung folgt
E[u(X) | F ] ≥ E[an X + bn | F ] = an · E[X | F ] + bn
P -fast sicher für alle n ∈ N, also auch
E[u(X) | F ] ≥ sup (an · E[X | F ] + bn ) P -fast sicher.

n∈N

Corollary A.6 (Lp -contractivity). The map X 7→ E[X | F ] is a contraction on

Lp (Ω, A, P ) for every p ≥ 1, i.e.,
E [|E[X | F ]|p ] ≤ E[|X|p ] for any X ∈ L1 (Ω, A, P ).
Proof. Nach der Jensenschen Ungleichung gilt:
|E[X | F ]|p ≤ E[|X|p | F ] P -fast sicher.
Die Behauptung folgt durch Bilden des Erwartungswertes.
The proof of the corollary shows in particular that for a random variable X ∈ Lp , the
conditional expectation E[X | F ] is contained in Lp as well. We now restrict ourselves
to the case p = 2.
Conditional expectation as best L2-prediction value

The space L2 (Ω, A, P ) = L2 (Ω, A, P )/ ∼ of equivalence classes of square integrable
random variables is a Hilbert space with inner product (X, Y )L2 = E[XY ]. If F ⊆ A is
a sub-σ-algebra then L2 (Ω, F , P ) is a closed subspace of L2 (Ω, A, P ), because limits
of F -measurable random variables are F -measurable as well. For X ∈ L2 (Ω, A, P ),
each version of the conditional expectation E[X | F ] is contained in the subspace
L2 (Ω, F , P ) by Jensen’s inequality. Furthermore, the conditional expectation respects
equivalence classes, see Theorem A.3. Therefore, X 7→ E[X | F ] induces a lin-
ear map from the Hilbert space L2 (Ω, A, P ) of equivalence classes onto the subspace
L2 (Ω, F , P ).
Theorem A.7 (Characterization of the conditional expectation as best L2 approxi-

mation and as orthogonal projection). For Y ∈ L2 (Ω, F , P ) the following statements
are all equivalent:

(1). Y is a version of the conditional expectation E[X | F ].
(2). Y is a “best approximation” of X in the subspace L2 (Ω, F , P ), i.e.,
E[(X − Y )2 ] ≤ E[(X − Z)2 ] for any Z ∈ L2 (Ω, F , P ).
(3). Y is a version of the orthogonal projection of X onto the subspace

L2 (Ω, F , P ) ⊆ L2 (Ω, A, P ), i.e.,
E[(X − Y ) · Z] = 0 for any Z ∈ L2 (Ω, F , P ).
X
L2 (Ω, A, P )
E[X | F ]
0
L2 (Ω, F , P )
Figure A.1: X 7→ E[X | F ] as orthogonal projection onto the subspace L2 (Ω, F , P ).
Proof. (1) ⇐⇒ (3): Für Y ∈ L2 (Ω, F , P ) gilt:
Y ist eine Version von E[X | F ]

⇐⇒ E[Y · IA ] = E[X · IA ] für alle A ∈ F
⇐⇒ E[Y · Z] = E[X · Z] für alle Z ∈ L2 (Ω, F , P )
⇐⇒ E[(X − Y ) · Z] = 0 für alle Z ∈ L2 (Ω, F , P )

Hierbei zeigt man die zweite Äquivalenz mit den üblichen Fortsetzungsverfahren
(maßtheoretische Induktion).
(3) ⇒ (2): Sei Y eine Version der orthogonalen Projektion von X auf L2 (Ω, F , P ).
Dann gilt für alle Z ∈ L2 (Ω, F , P ):
E[(X − Z)2 ] = E[((X − Y ) + (Y − Z))2 ]

= E[(X − Y )2 ] + E[(Y − Z)2 ] + 2E[(X − Y ) (Y − Z) ]
| {z }
∈L2 (Ω,F ,P )
≥ E[(X − Y )2 ]
Hierbei haben wir im letzten Schritt verwendet, dass Y −Z im Unterraum L2 (Ω, F , P )

enthalten, also orthogonal zu X − Y ist.
(2) ⇒ (3): Ist umgekehrt Y eine beste Approximation von X in L2 (Ω, F , P ) und Z ∈
L2 (Ω, F , P ), dann gilt
E[(X − Y )2 ] ≤ E[(X − Y + tZ)2 ]

= E[(X − Y )2 ] + 2tE[(X − Y )Z] + t2 E[Z 2 ]
für alle t ∈ R, also E[(X − Y ) · Z] = 0.
The equivalence of (2) and (3) is a well-known functional analytic statement: the best
approximation of a vector in a closed subspace of a Hilbert space is the orthogonal
projection of the vector onto this subspace. The geometric intuition behind this fact is
indicated in Figure A.1.
Theorem A.7 is a justification for the interpretation of the conditional expectation as

a predicion value. For example, by the factorization lemma, E[X | Y ] is the best L2 -
prediction for X among all functions of type g(Y ), g : R → R measurable.
Existence of conditional expectations

By the characterization of the conditional expectation as the best L2 -approximation,
the existence of conditional expectations of square integrable random variables is an

immediate consequence of the existence of the best approximation of a vector in a closed

subspace of a Hilbert space. By monotone approximation, the existence of conditional
expectations of general non-negative random variables then follows easily.
Theorem A.8 (Existence of conditional expectations). For every random variable

X ≥ 0 or X ∈ L1 (Ω, A, P ), and every σ-algebra F ⊆ A, there exists a version of the
conditional expectation E[X | F ].
Proof. (1). Wir betrachten zunächst den Fall X ∈ L2 (Ω, A, P ). Wie eben bemerkt, ist
der Raum L2 (Ω, F , P ) ein abgeschlossener Unterraum des Hilbertraums L2 (Ω, A, P ).
Sei d = inf{kZ − XkL2 | Z ∈ L2 (Ω, F , P )} der Abstand von X zu diesem Un-
terraum. Um zu zeigen, dass eine beste Approximation von X in L2 (Ω, F , P ) ex-
istiert, wählen wir eine Folge (Xn ) aus diesem Unterraum mit kXn − XkL2 → d.
Mithilfe der Parallelogramm-Identität folgt für n, m ∈ N:
kXn − Xm k2L2 = k(Xn − X) − (Xm − X)k2L2

= 2 · kXn − Xk2L2 + 2 · kXm − Xk2L2 − k(Xn − X) + (Xm − X)k2L2
2
X n + X m
= 2 · kXn − Xk2L2 +2 · kXm − Xk2L2 −4
− X ,
2
| {z } | {z } 2 L
→d2 →d2 | {z }
≤d2
und damit
lim sup kXn − Xm k2L2 ≤ 0.
n,m→∞
Also ist die Minimalfolge (Xn ) eine CauchyLfolge in dem vollständigen Raum
L2 (Ω, F , P ), d.h. es existiert ein Y ∈ L2 (Ω, F , P ) mit
kXn − Y kL2 → 0.
Für Y gilt
kY − XkL2 = k lim Xn − XkL2 ≤ lim inf kXn − XkL2 ≤ d,

n→∞ n→∞
d.h. Y ist die gesuchte Bestapproximation, und damit eine Version der bedingten
Erwartung E[X | F ].

(2). Für eine beliebige nichtnegative Zufallsvariable X auf (Ω, A, P ) existiert eine
monoton wachsende Folge (Xn ) nichtnegativer quadratintegrierbarer Zufallsvari-
ablen mit X = sup Xn . Man verifiziert leicht, dass sup E[Xn | F ] eine Version
n
von E[X | F ] ist.
(3). Entsprechend verifiziert man, dass für allgemeine X ∈ L1 (Ω, A, P ) durch E[X|F ] =
E[X + | F ] − E[X − | F ] eine Version der bedingten Erwartung gegeben ist.

Index
absolute continuity first hitting time, 82

- of probability measures, 139
Gaussian
absolutely continuous
- measure, 29
- function, 140
Gaussian process, 24
locally -, 141
Green function, 231
adapted, 59
harmonic, 66
Brownian motion sub-, 66
canonical model for -, 20 super-, 66
definition, 13 heat kernel, 231
geometric -, 246
i.i.d., 9
Lévy process, 14
integrable
Markov property, 15
uniformly -, 84
canonical model for Brownian motion, 20 Integral
colored noise, 9 Stochastic -
completion of a σ-algebra, 99 Discrete -, 77
Conditional Variance Process, 71 Itô diffusion, 256
Itô integral
Dirichlet problem, 88 - for predictable step functions, 165
Doob backward -, 201
- maximal inequality, 172 Itô map, 167, 175
Itô’s formula
Filtration, 59, 97
- without probabilities, 198
completed -, 99, 171
right-continuous -, 98 Lebesgue decomposition, 143
564
INDEX 565
Lebesgue density, 145 partition, 161

likelihood predictable step functions, 165
- ratio, 141 probability measure
local martingale, 184 absolute continuity of -, 139
localizing sequence, 183 singular -, 139
lognormal distribution, 260 Process
adapted -, 59
Markov
Conditional Variance -, 71
- chain, 66
predictable -, 69
Martingale, 60
process
- transform, 77
Gaussian -, 24
Definition, 97
Sub-, 60 Quadratic Variation, 196
Super-, 60
Radon-Nikodym
Martingale Problem
- derivative, 141
Solution of-, 74
martingale transform, 165 scale function, 257
measure Schauder functions, 36
Gaussian -, 29 singularity
Wiener -, 20 - of probability measures, 139
Mehler formula, 272 Stochastic integral
mesh size - Definition, 175
- of a partition, 161 Discrete -, 77
modification stopping time, 55
- of a martingale, 172 Stratonovich integral, 201
subharmonic, 66
noise
superharmonic, 66
colored -, 9
white -, 9 uniformly integrable, 84
usual conditions, 99
occupation time density, 231
Ornstein-Uhlenbeck process, 268 Variance Process
orthonormal basis, 36 Conditional -, 71

566 INDEX
white noise, 9, 25
Wiener
- measure, 20

Bibliography
[1]
[2]
[3] Sur l’équation de Kolmogoroff, par W. Doeblin. Éditions Elsevier, Paris, 2000. C.
R. Acad. Sci. Paris Sér. I Math. 331 (2000), Special Issue.
[4] Robert J. Adler and Jonathan E. Taylor. Random fields and geometry. Springer
Monographs in Mathematics. Springer, New York, 2007.
[5] David Applebaum. Lévy Processes and Stochastic Calculus. Cambridge UP, Cam-
bridge, 2004.
[6] Ole E. Barndorff-Nielsen. Processes of normal inverse Gaussian type. Finance

Stoch., 2(1):41–68, 1998.
[7] Richard F. Bass. Stochastic Processes. Cambridge University Press, Cambridge,

2011.
[8] Jean Bertoin. Lévy Processes. Cambridge UP, Cambridge, 1998.
[9] Jean-Michel Bismut. Large deviations and the Malliavin calculus. Birkhaeuser,
Boston, 1994.
[10] Amir Dembo and Ofur Zeitouni. Large Deviations Techniques and Applications.
Springer, 1998, Berlin.
[11] Frank den Hollander. Large Deviations. American Mathematical Society, Provi-
dence, 2008.
567
568 BIBLIOGRAPHY
[12] Andreas Eberle. Markov processes.

http://www.uni-bonn.de/~eberle/MarkovProcesses/MPSkript.pdf,
2009.
[13] Andreas Eberle. Wahrscheinlichkeitstheorie und Stochastische Prozesse.

http://www.uni-bonn.de/~eberle/StoPr/WthSkriptneu.pdf,
2010.
[14] Andreas Eberle. Introduction to Stochastic Analysis.

http://www.uni-bonn.de/~eberle/StoAn1011/StoAnSkriptneu.pdf,
2011.
[15] David Elworthy. On the Geometry of Diffusion Operators and Stochastic Flows.
Springer, 1999, Berlin.
[16] Stewart N. Ethier and Thomas G. Kurtz. Markov Processes. Wiley-Interscience,

Hoboken, NJ, 2005.
[17] William Feller. Introduction to Probability Theory and Its Applications. Wiley,
New York, 1957.
[18] Peter K. Friz and Martin Hairer. A course on rough paths. Universitext. Springer,
Cham, 2014. With an introduction to regularity structures.
[19] Peter K. Friz and Nicolas Victoir. Multidimensional Stochastic Processes as Rough
Paths. Cambridge UP, Cambridge, 2010.
[20] W. Hackenbroch and A. Thalmaier. Stochastic Analysis. Teubner, 1991.
[21] Martin Hairer. On malliavin’s proof of hörmander’s theorem. Bull. Sci. Math.,
135(6-7):650–666, 2011.
[22] Elton Hsu. Stochastic Analysis on Manifolds. Oxford University Press, Oxford,
2002.
[23] Nobuyuki Ikeda and Takehiko Ed. Watanabe. Stochastic Differential Equations
and Diffusion Processes. North-Holland, Amsterdam, 1981.

BIBLIOGRAPHY 569
[24] Jean Jacod and Albert Shiryaev. Limit Theorems for Stochastic Processes.
Springer, Berlin, 2003.
[25] Ioannis Karatzas and Steven E. Shreve. Brownian motion and stochastic calculus.
Springer, New York, 2010.
[26] Peter E. Kloeden and Eckhard Platen. Numerical solution of stochastic differen-
tial equations, volume 23 of Applications of Mathematics (New York). Springer-
Verlag, Berlin, 1992.
[27] Hiroshi Kunita. Stochastic Flows and Stochastic Differential Equations. Cam-
bridge University Press, Cambridge, 1997.
[28] Andreas Kyprianou. Introductory Lectures on Fluctuations of Lévy Processes with

Applications. Springer, Berlin, 2006.
[29] Terry J. Lyons. Differential Equations Driven by Rough Paths, Ecole d’Eté de
Probabilités de Saint-Flour XXXIV. Springer, 2004.
[30] Zhi-Ming Ma and Michael Röckner. Introduction to the Theory of (Non-

Symmetric) Dirichlet Forms. Springer, Berlin, 1992.
[31] Dilip B. Madan and Eugene Seneta. The variance gamma model for share market
returns. Journal of Business, 63(4):511–524, 1990.
[32] Paul Malliavin. Stochastic Analysis. Springer, Berlin, 2002.
[33] G. N. Milstein and M. V. Tretyakov. Stochastic numerics for mathematical physics.

Scientific Computation. Springer-Verlag, Berlin, 2004.
[34] David Nualart. The Malliavin Calculus and Related Topics. Springer, Berlin,
2006.
[35] Giuseppe Da Prato. Introduction to Stochastic Analysis and Malliavin Calculus.

Edizioni della Normale, Pisa, 2009.
[36] Philip E. Protter. Stochastic Integration and Differential Equations. Springer,

Berlin, 2005.

570 BIBLIOGRAPHY
[37] Daniel Revuz and Marc Yor. Continuous martingales and Brownian motion.
[38] L. C. G. Rogers and David Williams. Diffusions, Markov processes and martin-
gales, Vol. 2: Ito calculus. Cambridge UP, Camebridge, 2008.
[39] Steven E. Shreve. Stochastic calculus for finance. II. Springer Finance. Springer-
Verlag, New York, 2004. Continuous-time models.
[40] Daniel W. Stroock and S.R. S. Varadhan. Multidimensional Diffusion Processes.

[41] Shinzo Watanabe. Lectures on stochastic differential equations and Malliavin cal-
culus. Springer, Berlin, 1984.

Eberle Introduction To Stochastic Analysis

Uploaded by

Copyright:

Available Formats

Eberle Introduction To Stochastic Analysis

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Eberle Introduction To Stochastic Analysis

Uploaded by

Copyright:

Available Formats

Introduction to Stochastic Analysis

February 23, 2016

Maximum of Brownian motion . . . . . . . . . . . . . . . . . . . . . . 60

II Introduction to Stochastic Analysis 67

2 Martingales in discrete time 68

3 Martingales in continuous time 107

University of Bonn 2015/2016

3.3 Maximal inequalities and the LIL . . . . . . . . . . . . . . . . . . . . . 118

4 Martingale Convergence Theorems 125

5 Stochastic Integration w.r.t. Continuous Martingales 152

Stochastic Analysis Andreas Eberle

Definition of Itô integrals for square-integrable integrands . . . . . . . . 169

6 Itô’s formula and pathwise integrals 183

7 Brownian Motion and PDE 221

University of Bonn 2015/2016

7.2 Boundary value problems, exit and occupation times . . . . . . . . . . 227

8 SDE: Explicit Computations 245

Stochastic Analysis Andreas Eberle

8.5 Stochastic differential equations in Rn . . . . . . . . . . . . . . . . . . 280

9 Change of measure 286

III Stochastic Analysis 318

10 Lévy processes and Poisson point processes 313

University of Bonn 2015/2016

Examples with infinite jump intensity . . . . . . . . . . . . . . . . . . 321

11 Transformations of SDE 360

Stochastic Analysis Andreas Eberle

Doob’s h-transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388

12 Extensions of Itô calculus 406

University of Bonn 2015/2016

Continuous modifications of random fields . . . . . . . . . . . . . . . . 449

13 Variations of parameters in SDE 457

14 Stochastic calculus for semimartingales with jumps 478

Stochastic Analysis Andreas Eberle

Covariation and integration by parts . . . . . . . . . . . . . . . . . . . 509

A Conditional expectations 549

University of Bonn 2015/2016

Stochastic Analysis Andreas Eberle

This introduction to stochastic analysis starts with an introduction to Brownian motion.

random variables, Brownian motion shows up as a universal scaling limit of Random

1.1 From Random Walks to Brownian Motion

Xn+1 = Xn + Φn+1 (Xn , ηn+1 ), n = 0, 1, 2, . . . . (1.1.1)

Sn+1 = Sn + ηn+1 for n = 0, 1, 2, . . . .

Xn+1 − Xn = Φn+1 (Xn , Sn+1 − Sn ), n = 0, 1, 2, . . . . (1.1.2)

University of Bonn 2015/2016

Central Limit Theorem

converge in distribution to a multivariate normal distribution N(0, C) with covariance

on a probability space (Ω, A, P ) with independent increments ηi ∈ L2 (Ω, A, P ) nor-

E[ηi ] = 0 and Var[ηi ] = 1. (1.1.3)

Stochastic Analysis Andreas Eberle

500 1000 5000 10000

for some constants cm > 0. If t is a multiple of 1/m, then

University of Bonn 2015/2016

Stochastic Analysis Andreas Eberle

of the rescaled process is a rescaled sum of m · (ti+1 − ti ) i.i.d. random variables

Our considerations motivate the following definition:

Definition (Brownian Motion).

(1). Let a ∈ R. A continuous-time stochastic process Bt : Ω → R, t ≥ 0, defined on

(a) B0 (ω) = a for each ω ∈ Ω.