Stochastic Analysis

Andreas Eberle

January 28, 2013 Contents

Contents 2

1 Lévy processes and Poisson point processes 6 1.1 Lévyprocesses ...... 7 Basicexamples ...... 8 Characteristicexponents ...... 13 MartingalesofLévyprocesses ...... 15 CompoundPoissonprocesses...... 16 Superpositions and subdivisions of Poisson processes ...... 20 1.2 Poissonpointprocesses...... 21 Deﬁnition and construction of Poisson point processes ...... 23 Construction of compound Poisson processes from PPP ...... 26 Stochastic integrals w.r.t. Poisson point processes ...... 28 Lebesgueintegrals ...... 31 Itô integrals w.r.t. compensated Poisson point processes ...... 32 1.3 Lévyprocesseswithinﬁnitejumpintensity ...... 34 ConstructionfromPoissonpointprocesses ...... 34 TheLévy-Itôdecomposition ...... 38 Subordinators ...... 41 Stableprocesses...... 44

2 Stochastic integrals and Itô calculus for semimartingales 47 Semimartingalesindiscretetime ...... 48

2 CONTENTS 3

Semimartingalesincontinuoustime ...... 49 2.1 Finitevariationcalculus...... 52 Lebesgue-Stieltjesintegralsrevisited ...... 53 Productrule...... 54 Chainrule...... 57 Exponentialsofﬁnitevariationfunctions ...... 58 2.2 Stochasticintegrationforsemimartingales ...... 66 Integralswithrespecttoboundedmartingales ...... 66 Localization...... 72 Integrationw.r.t.semimartingales ...... 74 2.3 Quadraticvariationandcovariation...... 76 Covariationandintegrationbyparts ...... 77 Quadratic variation and covariation of local martingales ...... 78 Covariationofstochasticintegrals ...... 84 The Itô isometry for stochastic integrals w.r.t.martingales...... 86 2.4 Itôcalculusforsemimartingales ...... 87 Integrationw.r.t.stochasticintegrals ...... 87 Itô’sformula...... 88 ApplicationtoLévyprocesses ...... 92 ApplicationstoItôdiffusions ...... 94 Exponentialsofsemimartingales ...... 97 2.5 Generalpredictableintegrandsandlocaltimes ...... 100 Deﬁnition of stochastic integrals w.r.t. semimartingales ...... 102 Localization...... 103 Propertiesofthestochasticintegral...... 105 Localtime...... 108 Itô-Tanakaformula ...... 112

3 SDE I: Strong solutions and approximation schemes 115 3.1 Existenceanduniquenessofstrongsolutions ...... 118 Lp Stability ...... 119 Existenceofstrongsolutions ...... 122

UniversityofBonn WinterSemester2012/2013 4 CONTENTS

Non-explosioncriteria ...... 124 3.2 StochasticﬂowandMarkovproperty...... 126 Existenceofacontinuousﬂow ...... 127 Markovproperty ...... 129 Continuityoflocaltime...... 130 3.3 Stratonovichdifferentialequations ...... 132 Itô-Stratonovichformula ...... 133 StratonovichSDE...... 134 Brownianmotiononhypersurfaces ...... 135 Doss-Sussmannmethod...... 138 WongZakaiapproximationsofSDE ...... 140 3.4 Stochastic Taylor expansions and numerical methods ...... 142 Itô-Taylorexpansions ...... 142 NumericalschemesforSDE ...... 147 Strongconvergenceorder...... 148

4 SDEII:Transformationsandweaksolutions 152 4.1 Lévycharacterizations ...... 154 Lévy’scharacterizationofBrownianmotion ...... 155 LévycharacterizationofLévyprocesses ...... 159 Lévycharacterizationofweaksolutions ...... 161 4.2 Randomtimechange ...... 163 Continuous local martingales as time-changed Brownian motions. . . . 164 Time-change representations of stochastic integrals ...... 166 Time substitution in stochastic differential equations ...... 166 One-dimensionalSDE ...... 169 4.3 Girsanovtransformation ...... 172 Changeofmeasureonﬁlteredprobabilityspaces ...... 174 TranslatedBrownianmotions...... 176 FirstapplicationstoSDE ...... 178 Novikov’scondition...... 180 4.4 DrifttransformationsforItôdiffusions ...... 181

Stochastic Analysis Andreas Eberle CONTENTS 5

TheMarkovproperty ...... 182 Bridgesandheatkernels ...... 183 DrifttransformationsforgeneralSDE ...... 186 Doob’s h-transformanddiffusionbridges ...... 187 4.5 Largedeviationsonpathspaces ...... 189 TranslationsofWienermeasure ...... 189 Schilder’sTheorem ...... 191 Randomperturbationsofdynamicalsystems ...... 195 4.6 Changeofmeasureforjumpprocesses ...... 197 Poissonpointprocesses...... 197 ApplicationtoLévyprocesses ...... 199 Ageneraltheorem...... 200

5 Variations of SDE 202 5.1 VariationsofparametersinSDE ...... 203 Differentationofsolutionsw.r.t.aparameter ...... 204 DerivativeﬂowandstabilityofSDE ...... 205 Consequencesforthetransitionsemigroup ...... 209 5.2 Malliavin gradient and Bismut integration by parts formula ...... 210 Gradient and integration by parts for smooth functions ...... 212 Skorokhodintegral ...... 216 DeﬁnitionofMalliavingradientII ...... 216 Productandchainrule ...... 219 5.3 DigressiononRepresentationTheorems ...... 219 ItôsRepresentationTheorem ...... 220 Clark-Oconeformula ...... 223 5.4 FirstapplicationstoSDE ...... 223 5.5 Existenceandsmoothnessofdensities ...... 223

Index 224

Bibliography 226

UniversityofBonn WinterSemester2012/2013 Chapter 1

Lévy processes and Poisson point processes

A widely used class of possible discontinuous driving processes in stochastic differential equations are Lévy processes. They include Brownian motion, Poisson and compound Poisson processes as special cases. In this chapter, we outline basics from the theory of Lévy processes, focusing on prototypical examples of Lévy processes and their construction. For more details we refer to the monographs of Applebaum [3] and Bertoin [6]. Apart from simple transformations of Brownian motion, Lévy processes do not have continuous paths. Instead, we will assume that the paths are càdlàg (continue à droite, limites à gauche), i.e., right continuous with left limits. This can always be assured by choosing an appropriate modiﬁcation. We now summarize a few notations and facts about càdlàg functions that are frequently used below. If x : I R is a càdlàg function → deﬁned on a real interval I, and s is a point in I except the left boundary point, then we denote by

xs = lim xs ε − ε 0 − ↓ the left limit of x at s, and by

∆xs = xs xs − − 6 1.1. LÉVY PROCESSES 7

the size of the jump at s. Note that the function s xs is left continuous with right → − limits. Moreover, x is continuous if and only if ∆x =0 for all s. Let (I) denote the s D linear space of all càdlàg functions x : I R. → Exercise (Càdlàg functions). Prove the following statements: 1) If I is a compact interval, then for any function x (I), the set ∈D s I : ∆x >ε { ∈ | s| } is ﬁnite for any ε > 0. Conclude that any function x ([0, )) has at most ∈ D ∞ countably many jumps. 2) A càdlàg function deﬁned on a compact interval is bounded. 3) A uniform limit of a sequence of càdlàg functions is again càdlàg .

1.1 Lévy processes

Lévy processes are Rd-valued stochastic processes with stationary and independent increments. More generally, let ( t)t 0 be a filtration on a probability space (Ω, ,P ). F ≥ A Definition. An ( ) Lévy process is an ( ) adapted càdlàg stochastic process Ft Ft X : Ω Rd such that w.r.t. P , t → (a) X X is independent of for any s, t 0, and s+t − s Fs ≥ (b) X X X X for any s, t 0. s+t − s ∼ t − 0 ≥ Any Lévy process (X ) is also a Lévy process w.r.t.the filtration ( X ) generated by the t Ft process. Often continuity in probability is assumed instead of càdlàg sample paths. It can then be proven that a càdlàg modification exists, cf. [32, Ch.I Thm.30]. Remark (Lévy processes in discrete time are Random Walks). A discrete-time process (Xn)n=0,1,2,... with stationary and independent increments is a Random Walk: n Xn = X0 + ηj with i.i.d. increments ηj = Xj Xj 1. j=1 − − Remark (Lévy processes and infinite divisibility). The increments X X of a s+t − s Lévy process are infinitely divisible random variables, i.e., for any n N there ex- ∈ ist i.i.d.random variables Y ,...,Y such that X X has the same distribution as 1 n s+t − s UniversityofBonn WinterSemester2012/2013 8 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES

n Yi. Indeed, we can simply choose Yi = Xs+it/n Xs+i(t 1)/n. The Lévy-Khinchin i=1 − − formula gives a characterization of all distributions of inﬁnitely divisible random variables, cf.e.g. [3]. The simplest examples of inﬁnitely divisible distributions are normal and Poisson distributions.

Exercise (Strong Markov property for Lévy processes). Let (X ) be an ( ) Lévy t Ft process, and let T be a finite stopping time. Show that Y = X X is a process t T +t − T that is independent of , and X and Y have the same law. FT n n Hint: Consider the sequence of stopping times defined by T = (k + 1)2− if k2− n ≤ n T < (k + 1)2− . Notice that T T as n . In a first step show that for any m N n ↓ →∞ ∈ and t < t <...

E [f(X X ,...,X X )I ]= E [f(X ,...,X )] P [A]. Tn+t1 − Tn Tn+tm − Tn A t1 tm

Basic examples

We now consider ﬁrst examples of continuous and discontinuous Lévy processes.

Example (Brownian motion and Gaussian Lévy processes). A d-dimensional Brow- nian motion (Bt) is by deﬁnition a continuous Lévy process with

B B N(0, (t s)I ) for any 0 s < t. t − s ∼ − d ≤

Moreover, Xt = σBt + bt is a Lévy process with normally distributed marginals for d d d any σ R × and b R . Note that these Lévy processes are precisely the driving ∈ ∈ processes in SDE considered so far.

First examples of discontinuous Lévy processes are Poisson and, more generally, compound Poisson processes.

Example (Poisson processes). The most elementary example of a pure jump Lévy process in continuous time is the Poisson process. It takes values in 0, 1, 2,... and { } Stochastic Analysis Andreas Eberle 1.1. LÉVY PROCESSES 9 jumps up one unit each time after an exponentially distributed waiting time. Explicitly, a Poisson process (Nt)t 0 with intensity λ> 0 is given by ≥ ∞ N Nt = I Sn t = ♯ n : Sn t (1.1) { ≤ } { ∈ ≤ } n=1 where S = T + T + + T with independent random variables T Exp(λ). The n 1 2 n i ∼ increments N N of a Poisson process over disjoint time intervals are independent t − s and Poisson distributed with parameter λ(t s), cf. [12, Satz 10.12]. Note that by (1.1), − the sample paths t N (ω) are càdlàg. In general, any Lévy process with → t X X Poisson (λ(t s)) for any 0 s t t − s ∼ − ≤ ≤ is called a Poisson process with intensity λ, and can be represented as above. The paths of a Poisson process are increasing and hence of ﬁnite variation. The compensated Poisson process M := N E[N ] = N λt t t − t t − is an ( N ) martingale, yielding the semimartingale decomposition Ft

Nt = Mt + λt with the continuous ﬁnite variation part λt. On the other hand, there is the alternative trivial semimartingale decomposition Nt =0+ Nt with vanishing martingale part. This demonstrates that without an additional regularity condition, the semimartingale decomposition of discontinuous processes is not unique. A compensated Poisson process is a Lévy process which has both a continuous and a pure jump part.

Exercise (Martingales of Poisson processes). Prove that the compensated Poisson process M = N λt and the process M 2 λt are ( N ) martingales. t t − t − Ft

Exercise (A characterization of Poisson processes). Let (Xt)t 0 be a Lévy process ≥ with X0 = 0 a.s. Suppose that the paths of X are piecewise constant, increasing, all jumps of X are of size 1, and X is not identically 0. Prove that X is a Poisson process.

Hint: Apply the Strong Markov property to the jump times (Ti)i=1,2,... of X to conclude that the random variables Ui := Ti Ti 1 are i.i.d. (with T0 := 0). Then, it remains to − − show that U1 is an exponential random variable with some parameter λ> 0.

UniversityofBonn WinterSemester2012/2013 10 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES

For a different point of view on Poisson processes let

+(S)= δ (y ) finite or countable sequence in S Mc yi i denote the set of all counting measures on a set S. A Poisson process (Nt)t 0 can be ≥ viewed as the distribution function of a random counting measure, i.e., of a random variable N : Ω +([0, )). →Mc ∞ Definition. Let ν be a σ-finite measure on a measurable space (S, ). A collection of S random variables N(B), B , on a probability space (Ω, ,P ) is called a Poisson ∈ S A random measure (or spatial Poisson process) of intensity ν if and only if

(i) B N(B)(ω) is a counting measure for any ω Ω, → ∈ (ii) if B ,...,B are disjoint then the random variables N(B ),...,N(B ) are 1 n ∈ S 1 n independent,

(iii) N(B) is Poisson distributed with parameter ν(B) for any B with ν(B) < . ∈ S ∞ A Poisson random measure N with ﬁnite intensity ν can be constructed as the empirical measure of a Poisson distributed number of independent samples from the normalized measure ν/ν(S):.

K N = δ with X ν/ν(s) i.i.d., K Poisson(ν(S)) independent. Xj j ∼ ∼ j=1 If the intensity measure ν does not have atoms then almost surely, N( x ) 0, 1 for { } ∈{ } any x S, and N = x A δx for a random subset A of S. For this reason, a Poisson ∈ ∈ random measure is often called a Poisson point process, but we will use this terminology differently below.

A real-valued process (Nt)t 0 is a Poisson process of intensity λ > 0 if and only if ≥ t N (ω) is the distribution function of a Poisson random measure N(dt)(ω) on → t ([0, )) with intensity measure ν(dt) = λ dt. The Poisson random measure N(dt) B ∞ can be interpreted as the derivative of the Poisson process:

N(dt)= δs(dt).

s: ∆Ns=0 Stochastic Analysis Andreas Eberle 1.1. LÉVY PROCESSES 11

In a stochastic differential equation of type dYt = σ(Yt ) dNt, N(dt) is the driving − Poisson noise.

Any linear combination of independent Lévy processes is again a Lévy process:

Example (Superpositions of Lévy processes). If (Xt) and (Xt′) are independent Lévy Rd Rd′ Rn processes with values in and then αXt +βXt′ is a Lévy process with values in n d n d′ for any constant matrices α R × and β R × . For example, linear combinations ∈ ∈ of independent Brownian motions and Poisson processes are again Lévy processes.

Example (Inverse Gaussian subordinators). Let (Bt)t 0 be a one-dimensional Brow- ≥ nian motion with B =0 w.r.t. a right continuous filtration ( ), and let 0 Ft T = inf t 0 : B = s s { ≥ t } denote the first passage time to a level s R. Then (Ts)s 0 is an increasing stochastic ∈ ≥ process that is adapted w.r.t.the filtration ( Ts )s 0. For any ω, s Ts(ω) is the gener- F ≥ → alized right inverse of the Brownian path t B (ω). Moreover, by the strong Markov → t property, the process

B(s) := B B , t 0, t Ts+t − t ≥ is a Brownian motion independent of for any s 0, and FTs ≥ T = T + T (s) for s,u 0, (1.2) s+u s u ≥

(s) (s) (s) where Tu = inf t 0 : B = u is the ﬁrst passage time to u for the process B . ≥ t (s) Bt Bt

s + k s

Ts Ts+k

UniversityofBonn WinterSemester2012/2013 12 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES

By (1.2), the increment T T is independent of , and, by the reﬂection principle, s+u − s FTs 2 u 3/2 u Ts+u Ts Tu x− exp I(0, )(x) dx. − ∼ ∼ √ π −2x ∞ 2 Hence (Ts) is an increasing process with stationary and independent increments. The process (Ts) is left-continuous, but it is not difﬁcult to verify that

(s) Ts+ = lim Ts+ε = inf t 0 : Bt >u ε 0 ≥ ↓ is a càdlàg modiﬁcation, and hence a Lévy process. (Ts+) is called “The Lévy subordinator”, where “subordinator” stands for an increasing Lévy process. We will see below that subordinators are used for random time transformations (“subordination”) of other Lévy processes.

More generally, if Xt = σBt + bt is a Gaussian Lévy process with coefﬁcients σ > 0, b R, then the right inverse ∈ T X = inf t 0 : X = s , s 0, s { ≥ t } ≥ is called an Inverse Gaussian subordinator.

Exercise (Sample paths of Inverse Gaussian processes). Prove that the process (Ts)s 0 ≥ is increasing and purely discontinuous, i.e., with probability one, (Ts) is not continuous on any non-empty open interval (a, b) [0, ). ⊂ ∞ Example (Stable processes). Stable processes are Lévy processes that appear as scaling n Rd limits of Random Walks. Suppose that Sn = j=1 ηj is a Random Walk in with i.i.d. increments ηj. If the random variables ηj are square-integrable with mean zero then Donsker’s invariance principle (the “functional central limit theorem”) states that the 1/2 diffusively rescaled process (k− S kt )t 0 converges in distribution to (σBt)t 0 where ⌊ ⌋ ≥ ≥ (B ) is a Brownian motion in Rd and σ is a non-negative deﬁnite symmetric d d t × matrix. However, the functional central limit theorem does not apply if the increments

ηj are not square integrable (“heavy tails”). In this case, one considers limits of rescaled (k) 1/α Random Walks of the form Xt = k− S kt where α (0, 2] is a ﬁxed constant. It is ⌊ ⌋ (k) ∈ not difﬁcult to verify that if (Xt ) converges in distribution to a limit process (Xt) then

(Xt) is a Lévy process that is invariant under the rescaling, i.e.,

1/α k− X X for any k (0, ) and t 0. (1.3) kt ∼ t ∈ ∞ ≥

Stochastic Analysis Andreas Eberle 1.1. LÉVY PROCESSES 13

Deﬁnition. Let α (0, 2]. A Lévy process (X ) satisfying (1.3) is called (strictly) ∈ t α-stable.

The reason for the restriction to α (0, 2] is that for α > 2, an α-stable process ∈ does not exist. This will become clear by the proof of Theorem 1.2 below. There is a broader class of Lévy processes that is called α-stable in the literature, cf. e.g. [25]. Throughout these notes, by an α-stable process we always mean a strictly α-stable process as deﬁned above.

For b R, the deterministic process X = bt is a 1-stable Lévy process. Moreover, ∈ t 1 a Lévy process X in R is 2-stable if and only if Xt = σBt for a Brownian motion (B ) and a constant σ [0, ). Other examples of stable processes will be considered t ∈ ∞ below.

Characteristic exponents

From now on we restrict ourselves w.l.o.g. to Lévy processes with X0 = 0. The distribution of the sample paths is then uniquely determined by the distributions of the increments X X = X for t 0. Moreover, by stationarity and independence of t − 0 t ≥ the increments we obtain the following representation for the characteristic functions ϕ (p)= E[exp(ip X )]: t t

Theorem 1.1 (Characteristic exponent). If (Xt)t 0 is a Lévy process with X0 = 0 ≥ then there exists a continuous function ψ : Rd C with ψ(0) = 0 such that →

ip Xt tψ(p) d E[e ] = e− for any t 0 and p R . (1.4) ≥ ∈ Moreover, if (X ) has ﬁnite ﬁrst or second moments, then ψ is 1, 2 respectively, and t C C 2 k l ∂ ψ E[Xt] = it ψ(0) , Cov[Xt ,Xt ] = t (0) (1.5) ∇ ∂pk∂pl for any k,l =1,...,d and t 0. ≥ Proof. Stationarity and independence of the increments implies the identity

ϕ (p)= E[exp(ip X )] = E[exp(ip X )] E[exp(ip (X X ))] t+s t+s s t+s − s = ϕ (p) ϕ (p) (1.6) t s

UniversityofBonn WinterSemester2012/2013 14 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES for any p Rd and s, t 0. For a given p Rd, right continuity of the paths and ∈ ≥ ∈ dominated convergence imply that t ϕ (p) is right-continuous. Since → t

ϕt ε(p) = E[exp(ip (Xt Xε))], − − the function t ϕ (p) is also left continuous, and hence continuous. By (1.6) and since → t ϕ (p)=1, we can now conclude that for each p Rd, there exists ψ(p) C such that 0 ∈ ∈ (1.4) holds. Arguing by contradiction we then see that ψ(0) = 0 and ψ is continuous, since otherwise ϕt would not be continuous for all t. Moreover, if X is (square) integrable then ϕ is 1 (resp. 2), and hence ψ is also t t C C 1 (resp. 2). The formulae in (1.5) for the ﬁrst and second moment now follow by C C computing the derivatives w.r.t. p at p =0 in (1.4).

The function ψ is called the characteristic exponent of the Lévy process.

Examples. 1) For the Gaussian Lévy processes considered above, 1 1 ψ(p) = σT p 2 ib p = p ap ib p with a = σσT . 2| | − 2 − 2) The characteristic exponent of a Poisson process with intensity λ is

ψ(p) = λ(1 eip). − 3) In the superposition example above,

T T ψαX+βX′ (p) = ψX (α p)+ ψY (β p).

Characteristic exponents can be applied to classify all α-stable processes:

Theorem 1.2 (Characterization of stable processes). For α (0, 2] and a Lévy pro- ∈ 1 cess (Xt) in R with X0 =0 the following statements are equivalent:

(i) (Xt) is strictly α-stable.

(ii) ψ(cp)= cαψ(p) for any c 0 and p R. ≥ ∈ (iii) There exists constants σ 0 and R such that ≥ ∈ ψ(p) = σα p α(1 + i sgn(p)). | |

Stochastic Analysis Andreas Eberle 1.1. LÉVY PROCESSES 15

Proof. (i) (ii). The process (X ) is strictly α-stable if and only if X α cX for ⇔ t c t ∼ t any c, t 0, i.e., if and only if ≥ α tψ(cp) ipcXt ipX α c tψ(p) e− = E e = E e c t = e− for any c, t 0 and p R. ≥ ∈ (ii) (iii). Clearly, Condition (ii) holds if and only if there exist complex numbers ⇔ z+ and z such that − α z+ p for p 0, ψ(p) = | | ≥ z p α for p 0.  −| | ≤ Moreover, since ϕ (p) = exp( tψ(p)) is a characteristic function of a probability t −  measure for any t 0, the characteristic exponent ψ satisﬁes ψ( p) = ψ(p) and ≥ − (ψ(p)) 0. Therefore, z = z+ and (z+) 0. ℜ ≥ − ℜ ≥ Example (Symmetric α-stable processes). A Lévy process in Rd with characteristic exponent ψ(p) = σα p α | | for some σ 0 and a (0, 2] is called a symmetric α-stable process. It can be shown ≥ ∈ by Fourier transformation that a symmetric α-stable process is a Markov process with generator σα( ∆)α/2. In particular, Brownian motion is a symmetric 2-stable process. − −

Martingales of Lévy processes

The notion of a martingale immediately extends to complex or vector valued processes by a componentwise interpretation. As a consequence of Theorem 1.1 we obtain:

Corollary 1.3. If (Xt) is a Lévy process with X0 = 0 and characteristic exponent ψ, then the following processes are martingales:

(i) exp(ip X + tψ(p)) for any p Rd, t ∈ (ii) M = X bt with b = i ψ(0), provided X 1 t 0. t t − ∇ t ∈L ∀ ≥ j k jk jk ∂2ψ 2 (iii) Mt M a t with a = (0) (j, k = 1,...,d), provided Xt t − ∂pj ∂pk ∈ L t 0. ∀ ≥ UniversityofBonn WinterSemester2012/2013 16 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES

Proof. We only prove (ii) and (iii) for d = 1 and leave the remaining assertions as an exercise to the reader. If d =1 and (X ) is integrable then for 0 s t, t ≤ ≤

E[X X ] = E[X X ] = i(t s)ψ′(0) t − s |Fs t − s − by independence and stationarity of the increments and by (1.5). Hence M = X t t − itψ′(0) is a martingale. Furthermore,

M 2 M 2 =(M + M )(M M )=2M (M M )+(M M )2. t − s t s t − s s t − s t − s

If (Xt) is square integrable then the same holds for (Mt), and we obtain

E[M 2 M 2 ]= E[(M M )2 ] = Var[M M ] t − s |Fs t − s |Fs t − s |Fs = Var[Xt Xs s] = Var[Xt Xs] = Var[Xt s]=(t s)ψ′′(0) − |F − − −

2 Hence M tψ′′(0) is a martingale. t − Note that Corollary 1.3 (ii) shows that an integrable Lévy process is a semimartingale with martingale part Mt and continuous ﬁnite variation part bt. The identity (1.4) can be used to classify all Lévy processes, c.f. e.g. [3]. In particular, we will prove below that by Corollary 1.3, any continuous Lévy process with X0 =0 is of the type Xt = σBt +bt d d d with a d-dimensional Brownian motion (B ) and constants σ R × and b R . t ∈ ∈ From now on, we will focus on discontinuous Lévy processes.

Compound Poisson processes

Compound Poisson processes are pure jump Lévy processes, i.e., the paths are constant apart from a finite number of jumps in finite time. A compound Poisson process is a continuous time Random Walk defined by

Nt X = η , t 0, t j ≥ j=1 with a Poisson process (Nt) of intensity λ > 0 and with independent identically distributed random variables η : Ω Rd (j N) that are independent of the Poisson j → ∈ Stochastic Analysis Andreas Eberle 1.1. LÉVY PROCESSES 17 process as well.

The process (Xt) is again a pure jump process with jump times that do not accumulate. A compound Poisson process has jumps of size y with intensity

ν(dy) = λ π(dy),

where π denotes the joint distribution of the random variables ηj.

Lemma 1.4. A compound Poisson process is a Lévy process with characteristic exponent ip y ψ(p)= (1 e ) ν(dy). (1.7) − Proof. Let 0= t < t < < t . Then the increments 0 1 n

Ntk X X = η , k =1, 2, . . . , n , (1.8) tk − tk−1 j j=N +1 tk−1 are conditionally independent given the σ-algebra generated by the Poisson process d (Nt)t 0. Therefore, for p1,...,pn R , ≥ ∈ n n E exp i p (X X (N ) = E[exp(ip (X X ) (N )] k tk − tk−1 t k tk − tk−1 | t k=1 k=1 n Nt Nt − = ϕ(pk) k − k 1 , k=1 where ϕ denotes the characteristic function of the jump sizes ηj. By taking the expectation value on both sides, we see that the increments in (1.8) are independent and stationary, since the same holds for the Poisson process (Nt). Moreover, by a similar computation,

E[exp(ip X )] = E[E[exp(ip X ) (N )]] = E[ϕ(p)Nt ] t t | s ∞ k λt (λt) k λt(ϕ(p) 1) = e− ϕ(p) = e − k! k=0 for any p Rd, which proves (1.7). ∈ UniversityofBonn WinterSemester2012/2013 18 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES

The paths of a compound Poisson process are again of finite variation and càdlàg. One can show that every pure jump Lévy process with finitely many jumps in finite time is a compound Poisson process , cf.Theorem 1.13 below.

Exercise (Martingales of Compound Poisson processes). Show that the following processes are martingales:

(a) M = X bt where b = y ν(dy) provided η 1, t t − 1 ∈L (b) M 2 at where a = y 2 ν(dy) provided η 2. | t| − | | 1 ∈L We have shown that a compound Poisson process with jump intensity measure ν(dy) is a Lévy process with characteristic exponent

ip y d ψ (p)= (1 e )ν(dy) , p R . (1.9) ν − ∈ Since the distribution of a Lévy process on the space ([0, ), Rd) of càdlàg paths is D ∞ uniquely determined by its characteristic exponent, we can prove conversely:

Lemma 1.5. Suppose that ν is a ﬁnite positive measure on Rd 0 with total mass B \{ } λ = ν(Rd 0 ), and (X ) is a Lévy process with X = 0 and characteristic exponent \{ } t 0 ψ , deﬁned on a complete probability space (Ω, ,P ). Then there exists a sequence ν A 1 (ηj)j N of i.i.d.random variables with distribution λ− ν and an independent Poisson ∈ Process (N ) with intensity λ on (Ω, ,P ) such that almost surely, t A Nt

Xt = ηj . (1.10) j=1 Proof. Let (ηj) be an arbitrary sequence of i.i.d.random variables with distribution 1 d λ− ν, and let (Nt) be an independent Poisson process of intensity ν(R 0 ), all \ { } defined on a probability space (Ω, , P ). Then the compound Poisson process Xt = N A ft j=1 ηj is also a Lévy process with X0 =0 and characteristic exponent ψν . Therefore, the finite dimensional marginals of (Xt) and (Xt), and hence the distributions of (Xt) d and (Xt) on ([0, ), R ) coincide. In particular, almost every path t Xt(ω) has D ∞ → only finitely many jumps in a finite time interval, and is constant inbetween. Now set S0 =0 and let

Sj = inf s>Sj 1 : ∆Xs =0 for j N { − } ∈ Stochastic Analysis Andreas Eberle 1.1. LÉVY PROCESSES 19

denote the successive jump-times of (Xt). Then (Sj) is a sequence of non-negative random variables on (Ω, ,P ) that is almost surely finite and strictly increasing with A lim S = . Defining η := ∆X if S < , η =0 otherwise, and j ∞ j Sj j ∞ j N := s (0, t] : ∆X =0 = j N : S t , t { ∈ s } { ∈ j ≤ } as the successive jump sizes and the number of jumps up to time t, we conclude that almost surely, (Nt) is finite, and the representation (1.10) holds. Moreover, for any j N and t 0 , ηj and Nt are measurable functions of the process (Xt)t 0. Hence the ∈ ≥ ≥ joint distribution of all these random variables coincides with the joint distribution of the random variables η (j N) and N (t 0), which are the corresponding measurable j ∈ t ≥ functions of the process (Xt). We can therefore conclude that (ηj)j N is a sequence ∈ 1 of i.i.d.random variables with distributions λ− ν and (Nt) is an independent Poisson process with intensity λ.

The lemma motivates the following formal deﬁnition of a compound Poisson process:

Definition. Let ν be a finite positive measure on Rd, and let ψ : Rd C be the ν → function defined by (1.9).

1) The unique probability measure π on (Rd) with characteristic function ν B ip y d e π (dy) = exp( ψ (p)) p R ν − ν ∀ ∈ is called the compound Poisson distribution with intensity measure ν.

2) A Lévy process (X ) on Rd with X X π for any s, t 0 is called a t s+t − s ∼ tν ≥ compound Poisson process with jump intensity measure (Lévy measure) ν.

K The compound Poisson distribution πν is the distribution of j=1 ηj where K is a Pois- Rd son random variable with parameter λ = ν( ) and (ηj) is a sequence of i.i.d.random 1 variables with distribution λ− ν. By conditioning on the value of K , we obtain the explicit series representation

∞ k λ λ k π = e− ν∗ , ν k! k=0 k where ν∗ denotes the k-fold convolution of ν.

UniversityofBonn WinterSemester2012/2013 20 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES

Superpositions and subdivisions of Poisson processes

The following assertion about Poisson processes is intuitively clear from the interpretation of a Poisson process as the distribution function of a Poisson random measure. Compound Poisson processes enable us to give a simple proof of the second part of the lemma:

Theorem 1.6. Let K be a countable set.

(k) 1) Suppose that (Nt )t 0, k K, are independent Poisson processes with intensi- ≥ ∈ ties λk. Then N = N (k) , t 0, t t ≥ k K ∈ is a Poisson process with intensity λ = λ provided λ< . k ∞ 2) Conversely, if (Nt)t 0 is a Poisson process with intensity λ > 0, and (Cn)n N is ≥ ∈ a sequence of i.i.d.random variables C : Ω K that is also independent of n → (Nt), then the processes

Nt (k) Nt = I Cj =k , t 0, { } ≥ j=1 are independent Poisson processes of intensities qkλ, where qk = P [C1 = k].

The subdivision in the second assertion can be thought of as colouring the points in the support of the corresponding Poisson random measure N(dt) independently with (k) random colours Cj, and decomposing the measure into parts N (dt) of equal colour.

Proof. The first part is rather straightforward, and left as an exercise. For the second part, we may assume w.l.o.g. that K is finite. Then the process N : Ω RK defined t → by Nt (k) Nt := Nt = ηj with ηj = I k (Cj) k K { } k K ∈ j=1 ∈ is a compound Poisson process on RK, and hence a Lévy process. Moreover, by the proof of Lemma 1.4, the characteristic function of N for t 0 is given by t ≥ E exp ip N = exp (λt(ϕ(p) 1)) , p RK , t − ∈ Stochastic Analysis Andreas Eberle 1.2. POISSON POINT PROCESSES 21 where

ipk ϕ(p)= E [exp(ip η1)] = E exp i pkI k (C1) = qke . { } k K k K ∈ ∈

Noting that qk =1, we obtain E[exp(ip N )] = exp(λtq (eipk 1)) for any p RK and t 0. t k − ∈ ≥ k K ∈ The assertion follows, because the right hand side is the characteristic function of a Lévy process in RK whose components are independent Poisson processes with intensities qkλ.

1.2 Poisson point processes

A compensated Poisson process has only finitely many jumps in a finite time interval. General Lévy jump processes may have a countably infinite number of (small) jumps in finite time. In the next section, we will construct such processes from their jumps. As a preparation we will now study Poisson point processes that encode the jumps of Lévy processes. The jump part of a Lévy process can be recovered from these counting measure valued processes by integration, i.e., summation of the jump sizes.

Note ﬁrst that a Lévy process (Xt) has only countably many jumps, because the paths are càdlàg. The jumps can be encoded in the counting measure-valued stochastic process N : Ω +(Rd 0 ), t →Mc \{ }

N (dy) = δ (dy), t 0, t ∆Xs ≥ s t ∆X≤s=0 or, equivalently, in the random counting measure N : Ω + R (Rd 0 ) → Mc + × \{ } deﬁned by

N(dt dy) = δ(s,∆Xs)(dt dy). s t ∆X≤s=0

UniversityofBonn WinterSemester2012/2013 22 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES

d R Xt

∆Xt

The process (Nt)t 0 is increasing and adds a Dirac mass at y each time the Lévy pro- ≥ cess has a jump of size y. Since (Xt) is a Lévy process, (Nt) also has stationary and independent increments:

N (B) N (B) N (B) for any s, t 0 and B (Rd 0 ). s+t − s ∼ t ≥ ∈ B \{ } Hence for any set B with N (B) < a.s. for all t, the integer valued stochastic process t ∞ (Nt(B))t 0 is a Lévy process with jumps of size +1. By an exercise in Section 1.1, we ≥ can conclude that (N (B)) is a Poisson process. In particular, t E[N (B)] is a linear t → t function.

Stochastic Analysis Andreas Eberle 1.2. POISSON POINT PROCESSES 23

Definition. The jump intensity measure of a Lévy process (Xt) is the σ-finite measure ν on the Borel σ-algebra (Rd 0 ) determined by B \{ } E[N (B)] = t ν(B) t 0, B (Rd 0 ). (1.11) t ∀ ≥ ∈ B \{ } It is elementary to verify that for any Lévy process, there is a unique measure ν satisfying (1.11). Moreover, since the paths of a Lévy process are càdlàg, the measures Nt and ν are finite on y Rd : y ε for any ε> 0. { ∈ | | ≥ } Example (Jump intensity of stable processes). The jump intensity measure of strictly α-stable processes in R1 can be easily found by an informal argument. Suppose we rescale in space and time by y cy and t cαt. If the jump intensity is ν(dy) = → → f(y) dy, then after rescaling we would expect the jump intensity cαf(cy)c dy. If scale 1 α invariance holds then both measures should agree, i.e., f(y) y − − both for y > 0 ∝ | | and for y < 0 respectively. Therefore, the jump intensity measure of a strictly α-stable process on R1 should be given by

1 α ν(dy) = c+I(0, )(y)+ c I( ,0)(y) y − − dy (1.12) ∞ − −∞ | | with constants c+,c [0, ). − ∈ ∞

If (Xt) is a pure jump process with finite jump intensity measure (i.e., finitely many jumps in a finite time interval) then it can be recovered from (Nt) by adding up the jump sizes: X X = ∆X = y N (dy). t − 0 s t s t ≤ In the next section, we are conversely going to construct more general Lévy jump processes from the measure-valued processes encoding the jumps. As a first step, we are going to define formally the counting-measure valued processes that we are interested in.

Deﬁnition and construction of Poisson point processes

Let (S, , ν) be a σ-ﬁnite measure space. S UniversityofBonn WinterSemester2012/2013 24 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES

Deﬁnition. A collection N (B), t 0, B , of random variables on a probability t ≥ ∈ S space (Ω, ,P ) is called a Poisson point process of intensity ν if and only if A (i) B N (B)(ω) is a counting measure on S for any t 0 and ω Ω, → t ≥ ∈

(ii) if B1,...,Bn are disjoint then (Nt(B1))t 0,..., (Nt(Bn))t 0 are indepen- ∈ S ≥ ≥ dent stochastic processes and

(iii) (Nt(B))t 0 is a Poisson process of intensity ν(B) for any B with ν(B) < . ≥ ∈ S ∞ A Poisson point process adds random points with intensity ν(dt) dy in each time instant dt. It is the distribution function of a Poisson random measure N(dt dy) on R+ × S with intensity measure dt ν(dy), i.e.

N (B) = N((0, t] B) for any t 0 and B . t × ≥ ∈ S

b y b

b b b b b b b b b B Nt(B) b b b b b b b

The distribution of a Poisson point process is uniquely determined by its intensity measure: If (Nt) and (Nt) are Poisson point processes with intensity ν then

(Nt(B1),...,Nt(Bn))t 0 (Nt(B1),..., (Nt(Bn))t 0 ≥ ∼ ≥ for any ﬁnite collection of disjoint sets B ,...,B , and, hence, for any ﬁnite 1 n ∈ S collection of measurable arbitrary sets B ,...,B . 1 n ∈ S Applyinga measurable map to the points of a Poisson point process yields a new Poisson point process:

Stochastic Analysis Andreas Eberle 1.2. POISSON POINT PROCESSES 25

Exercise (Mapping theorem). Let (S, ) and (T, ) be measurable spaces and let S T f : S T be a measurable function. Prove that if (N ) is a Poisson point process with → t 1 intensity measure ν then the image measures N f − , t 0, form a Poisson point t ◦ ≥ 1 process on T with intensity measure ν f − . ◦

An advantage of Poisson point processes over Lévy processes is that the passage from ﬁnite to inﬁnite intensity (of points or jumps respectively) is not a problem on the level of Poisson point processes because the resulting sums trivially exist by positivity:

Theorem 1.7 (Construction of Poisson point processes).

1) Suppose that ν is a ﬁnite measure with total mass λ = ν(S). Then

Nt = δηj j=1

is a Poisson point process of intensity ν provided the random variables ηj are 1 independent with distribution λ− ν, and (Kt) is an independent Poisson process of intensity λ.

2) If (N (k)), k N, are independent Poisson point processes on (S, ) with intensity t ∈ S measures νk then

∞ (k) Nt = Nt k=1

is a Poisson point process with intensity measure ν = νk. The statements of the theorem are consequences of the subdivision and superposition properties of Poisson processes. The proof is left as an exercise.

Conversely, one can show that any Poisson point process with ﬁnite intensity measure ν can be almost surely represented as in the ﬁrst part of Theorem 1.7, where Kt = Nt(S). The proof uses uniqueness in law of the Poisson point process, and is similar to the proof of Lemma 1.5.

UniversityofBonn WinterSemester2012/2013 26 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES

Construction of compound Poisson processes from PPP

We are going to construct Lévy jump processes from Poisson point processes. Suppose first that (N ) is a Poisson point process on Rd 0 with finite intensity measure ν. t \{ } Then the support of N is almost surely finite for any t 0. Therefore, we can define t ≥

Xt = y Nt(dy) = y Nt( y ), Rd 0 { } \{ } y supp(Nt) ∈ d Theorem 1.8. If ν(R 0 ) < then (Xt)t 0 is a compound Poisson process with \{ } ∞ ≥ jump intensity ν. More generally, for any Poisson point process with ﬁnite intensity measure ν on a measurable space (S, ) and for any measurable function f : S Rn, S → n N, the process ∈ N f := f(y)N (dy) , t 0, t t ≥ 1 is a compound Poisson process with intensity measure ν f − . ◦ Proof. By Theorem 1.7 and by the uniqueness in law of a Poisson point process with

Kt given intensity measure, we can represent (Nt) almost surely as Nt = j=1 δηj with i.i.d. random variables η ν/ν(S) and an independent Poisson process (K ) of inten- j ∼ t sity ν(S). Thus,

Kt f Nt = f(y)Nt(dy) = f(ηj) almost surely. j=1 Since the random variables f(η ), j N, are i.i.d. and independent of (K ) with distri- j ∈ t 1 f bution ν f − , (N ) is a compound Poisson process with this intensity measure. ◦ t As a direct consequence of the theorem and the properties of compound Poisson processes derived above, we obtain:

Corollary 1.9 (Martingales of Poisson point processes). Suppose that (Nt) is a Pois- son point process with ﬁnite intensity measure ν. Then the following processes are martingales w.r.t. the ﬁltration N = σ(N (B) 0 s t, B ): Ft s | ≤ ≤ ∈ S f (i) N = N f t fdν for any f 1(ν), t t − ∈L f g 2 (ii) Nt Nt t fg dν for any f,g (ν), − ∈L Stochastic Analysis Andreas Eberle 1.2. POISSON POINT PROCESSES 27

(iii) exp (ipN f + t (1 eipf ) dν) for any measurable f : S R and p R. t − → ∈ Proof. If f is in p(ν) for p =1, 2 respectively, then L p 1 p x ν f − (dx)= f(y) ν(dy) < , | | ◦ | | ∞ 1 1 x ν f − (dx) = f dν, and xy ν (fg)− (dxdy)= fg dν . ◦ ◦ Therefore (i) and (ii) (and similarly also (iii)) follow from the corresponding statements for compound Poisson processes.

With a different proof and an additional integrability assumption, the assertion of Corol- lary 1.9 extends to σ-ﬁnite intensity measures:

Exercise (Expectation values and martingales for Poisson point processes with in-

finite intensity). Let (Nt) be a Poisson point process with σ-finite intensity ν. a) By considering first elementary functions, prove that for t 0, the identity ≥

E f(y)Nt(dy) = t f(y)ν(dy) holds for any measurable function f : S [0, ]. Conclude that for f 1(ν), f → ∞ ∈L the integral Nt = f(y)Nt(dy) exists almost surely and deﬁnes a random variable in L1(Ω, ,P ). A b) Proceeding similarly as in a), prove the identities

E[N f ] = t f dν for any f 1(ν), t ∈L Cov[N f , N g] = t fg dν for any f,g 1(ν) 2(ν), t t ∈L ∩L E[exp(ipN f )] = exp(t (eipf 1) dν) for any f 1(ν). t − ∈L c) Show that the processes considered in Corollary 1.9 are again martingales provided f 1(ν), f,g 1(ν) 2(ν) respectively. ∈L ∈L ∩L

If (Nt) is a Poisson point process with intensity measure ν then the signed measure valued stochastic process

N (dy) := N (dy) t ν(dy) , t 0, t t − ≥ UniversityofBonn WinterSemester2012/2013 28 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES is called a compensated Poisson point process . Note that by Corollary 1.9 and the exercise, f Nt = f(y)Nt(dy) is a martingale for any f 1(ν), i.e., (N ) is a measure-valued martingale. ∈L t Stochastic integrals w.r.t. Poisson point processes

Let (S, , ν) be a σ-finite measure space, and let ( ) be a filtration on a probability S Ft d space (Ω, ,P ). Our main interest is the case S = R . Suppose that (Nt(dy))t 0 is an A ≥ ( ) Poisson point process on (S, ) with intensity measure ν. As usual, we denote by Ft S N = N tν the compensated Poisson point process, and by N(dt dy) and N(dt dy) the t t− corresponding uncompensated and compensated Poisson random measure on R+ S. × Recall that for A, B with ν(A) < and ν(B) < , the processes N (A), N (B), ∈ S ∞ ∞ t t and Nt(A)Nt(B) tν(A B) are martingales. Our goal is to define stochastic integrals − ∩ of type

(G N)t = Gs(y) N(ds dy), (1.13) • (0,t] S ×

(G N)t = Gs(y) N(ds dy) (1.14) • (0,t] S × respectively for predictable processes (ω,s,y) G (y)(ω) defined on Ω R S. In → s × + × particular, choosing Gs(y)(ω)= y, we will obtain Lévy processes with possibly infinite jump intensity from Poisson point processes. If the measure ν is finite and has no atoms, the process G.N is defined in an elementary way as

(G.N)t = Gs(y). (s,y) supp(N), s t ∈ ≤ Definition. The predictable σ-algebra on Ω R S is the σ-algebra generated × + × P by all sets of the form A (s, t] B with 0 s t, A and B . A stochastic × × ≤ ≤ ∈Fs ∈ S process defined on Ω R S is called ( ) predictable iff it is measurable w.r.t. . × + × Ft P It is not difficult to verify that any adapted left-continuous process is predictable:

Stochastic Analysis Andreas Eberle 1.2. POISSON POINT PROCESSES 29

Exercise. Prove that is the σ-algebra generated by all processes (ω,t,y) G (y)(ω) P → t such that G is measurable for any t 0 and t G (y)(ω) is left-continuous t Ft × S ≥ → t for any y S and ω Ω. ∈ ∈

Example. If (Nt) is an ( t) Poisson process then the left limit process Gt(y)= Nt is F − predictable, since it is left-continuous. However, Gt(y) = Nt is not predictable. This is intuitively convincing since the jumps of a Poisson process can not be “predicted in advance”. A rigorous proof of the non-predictability, however, is surprisingly difﬁcult and seems to require some background from the general theory of stochastic processes, cf. e.g. [5].

We denote by the vector space consisting of all elementary predictable processes G E of the form n 1 m − Gt(y)(ω) = Zi,k(ω) I(ti,ti+1](t) IBk (y) (1.15) i=0 k=1 with m, n N, 0 t < t < < t , B ,...,B disjoint with ν(B ) < , and ∈ ≤ 0 1 n 1 m ∈ S k ∞ R Zi,k : Ω bounded and ti -measurable. For G , the stochastic integral G N is → F ∈ E • a well-deﬁned Lebesgue integral given by

n 1 m − (G N)t = Zi,k Nti t(Bk) Nti t(Bk) , (1.16) • +1∧ − ∧ i=0 k=1

Notice that the summands vanish for ti t and that G N is an ( t) adapted process ≥ • F with càdlàg paths.

Stochastic integrals w.r.t. Poisson point processes have properties reminiscent of those known from Itô integrals based on Brownian motion:

Lemma 1.10 (Elementary properties of stochastic integrals w.r.t. PPP). Let G . ∈E Then the following assertions hold:

1) For any t 0, ≥

E [(G N)t] = E Gs(y) ds ν(dy) . • (0,t] S ×

UniversityofBonn WinterSemester2012/2013 30 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES

2) The process G N deﬁned by •

(G N)t = Gs(y) N(ds dy) Gs(y) ds ν(dy) • (0,t] S − (0,t] S × × is a square integrable ( t) martingale with (G N)0 =0. F • 3) For any t 0, G N satisﬁes the Itô isometry ≥ • 2 2 E (G N)t = E Gs(y) ds ν(dy) . • (0,t] S × 2 2 4) The process (G N)t (0,t] S Gs(y) ds ν(dy) is an ( t) martingale. • − × F Proof. 1) Since the processes (Nt(Bk)) are Poisson processes with intensities ν(Bk), we obtain by conditioning on : Fti

E [(G N)t] = E Zi,k Nti t(Bk) Nti (Bk) • +1∧ − i,k:ti

= E Gs(y) ds ν(dy) . (0,t] S × 2) The process G N is bounded and hence square integrable. Moreover, it is a martin- • gale, since by 1), for any 0 s t and A s, ≤ ≤ ∈F

E [(G N)t (G N)s; A] = E IA Gr(y) I(s,t](r) N(dr dy) • − • (0,t] S ×

= E IA Gr(y) I(s,t](r) dr ν(dy) (0,t] S ×

= E Gr(y) dr ν(dy) Gr(y) dr ν(dy); A (0,t] S − (0,s] S × ×

= E Gs(y) ds ν(dy) . (0,t] S ×

3) We have (G N)t = Zi,k ∆iN(Bk), where • i,k ∆iN(Bk) := Nti t(Bk) Nti t(Bk) +1∧ − ∧ Stochastic Analysis Andreas Eberle 1.2. POISSON POINT PROCESSES 31 are increments of independent compensated Poisson point processes. Noticing that the summands vanish if t t, we obtain i ≥ 2 E (G N)t = E Zi,kZj,l∆iN(Bk)∆jN(Bl) • i,j,k,l = 2 E Z Z ∆ N(B )E[∆ N(B ) ] i,k j,l i k j l |Ftj icovariance E[∆ N(B )∆ N(B )] = δ ν(B )∆ t. i k Fti i k i l kl k i 4) now follows similarly as 2), and is left as an exercise to the reader.

Lebesgue integrals

If the integrand Gt(y) is non-negative, then the integrals (1.13) and (1.14) are well- deﬁned Lebesgue integrals for every ω. By Lemma 1.10 and monotone convergence, the identity

E [(G N)t] = E Gs(y) ds ν(dy) (1.17) • (0,t] S × holds for any predictable G 0. ≥ Now let u (0, ], and suppose that G : Ω (0,u) S R is predictable and ∈ ∞ × × → integrable w.r.t. the product measure P λ ν. Then by (1.17), ⊗ (0,u) ⊗

E Gs(y) N(dsdy) = E Gs(y) ds ν(dy) < . (0,u] S | | (0,u] S | | ∞ × × + Hence the processes G N and G−N are almost surely finite on [0,u], and, correspond- • • + ingly G N = G N G−N is almost surely well-defined as a Lebesgue integral, and it • • − • satisfies the identity (1.17).

Theorem 1.11. Suppose that G 1(P λ ν) is predictable. Then the following ∈L ⊗ (0,u) ⊗ assertions hold:

UniversityofBonn WinterSemester2012/2013 32 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES

P 1) G N is an ( t ) adapted stochastic process satisfying (1.17). • F P 2) The compensated process G N is an ( t ) martingale. • F

3) The sample paths t (G N)t are càdlàg with almost surely ﬁnite variation → •

(1) Vt (G N) Gs(y) N(ds dy). • ≤ (0,t] S | | × Proof. 1) extends by a monotone class argument from elementary predictable G to general non-negative predictable G, and hence also to integrable predictable G.

2) can be veriﬁed similarly as in the proof of Lemma 1.10.

+ 3) We may assume w.l.o.g. G 0, otherwise we consider G N and G−N separately. ≥ • • Then, by the Monotone Convergence Theorem,

(G N)t+ε (G N)t = Gs(y) N(ds dy) 0, and • − • (t,t+ε] S → ×

(G N)t (G N)t ε Gs(y) N(ds dy) • − • − → t S { }× as ε 0. This shows that the paths are càdlàg. Moreover, for any partition π of [0,u], ↓

(G N)r′ (G N)r = Gs(y) N(ds dy) | • − • | (r,r′] S r π r π × ∈ ∈ Gs(y) N(ds dy) < a.s. ≤ (0,u] S | | ∞ ×

Remark (Watanabe characterization). It can be shown that a counting measure valued process (N ) is an ( ) Poisson point process if and only if (1.17) holds for any t Ft non-negative predictable process G.

Itô integrals w.r.t. compensated Poisson point processes

Suppose that (ω,s,y) G (y)(ω) is a predictable process in 2(P λ ν) for → s L ⊗ (0,u) ⊗ some u (0, ]. If G is not integrable w.r.t. the product measure, then the integral G N ∈ ∞ • Stochastic Analysis Andreas Eberle 1.2. POISSON POINT PROCESSES 33 does not exist in general. Nevertheless, under the square integrability assumption, the integral G N w.r.t. the compensated Poisson point process exists as a square integrable • martingale. Note that square integrability does not imply integrability if the intensity measure ν is not finite. To define the stochastic integral G N for square integrable integrands G we use the Itô • isometry. Let 2([0,u]) = M 2([0,u]) t M (ω) càdlàg for any ω Ω Md ∈M | → t ∈ denote the space of all square-integrable càdlàg martingales w.r.t. the completed filtration ( P ). Recall that the L2 maximal inequality Ft 2 2 2 2 E sup Mt E[ Mu ] t [0,u] | | ≤ 2 1 | | ∈ − holds for any right-continuous martingale in 2([0,u]). Since a uniform limit of càdlàg M 2 functions is again càdlàg, this implies that the space Md ([0,u]) of equivalence classes of indistinguishable martingales in 2([0,u]) is a closed subspace of the Hilbert space Md M 2([0,u]) w.r.t. the norm

2 1/2 M 2 = E[ M ] . || ||M ([0,u]) | u| Lemma 1.10, 3), shows that for elementary predictable processes G,

G N M 2([0,u]) = G L2(P λ ν). (1.18) || • || || || ⊗ (0,u)⊗ On the other hand, it can be shown that any predictable process G L2(P λ ν) ∈ ⊗ (0,u) ⊗ is a limit w.r.t. the L2(P λ ν) norm of a sequence (G(k)) consisting of elementary ⊗ (0,u) ⊗ predictable processes. Hence isometric extension of the linear map G G N can be → • 2 2 used to deﬁne G N Md (0,u) for any predictable G L (P λ(0,u) ν) in such a • ∈ ∈ ⊗ ⊗ way that G(k)N G N in M 2 whenever G(k) G in L2. • −→ • → Theorem 1.12 (Itô isometry and stochastic integrals w.r.t. compensated PPP). Suppose that u (0, ]. Then there is a unique linear isometry G G N from ∈ ∞ → • 2 2 L (Ω (0,u) S, ,P λ ν) to Md ([0,u]) such that G N is given by (1.16) for × × P ⊗ ⊗ • any elementary predictable process G of the form (1.15). UniversityofBonn WinterSemester2012/2013 34 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES

Proof. As pointed out above, by (1.18), the stochastic integral extends isometrically to the closure ¯ of the subspace of elementary predictable processes in the Hilbert space E L2(Ω (0,u) S, ,P λ ν). It only remains to show that any square integrable × × P ⊗ ⊗ predictable process G is contained in ¯, i.e., G is an L2 limit of elementary predictable E processes. This holds by dominated convergence for bounded left-continuous processes, and by a monotone class argument or a direct approximation for general bounded predictable processes, and hence also for predictable processes in L2. The details are left to the reader.

The deﬁnition of stochastic integrals w.r.t. compensated Poisson point processes can be extended to locally square integrable predictable processes G by localization we refer − to [3] for details.

Example (Deterministic integrands). If H (y)(ω) = h(y) for some function h s ∈ 2(S, , ν) then L S h (H N)t = h(y) Nt(dy) = Nt , • 1 i.e., H N is a Lévy martingale with jump intensity measure ν h− . • ◦ 1.3 Lévy processes with inﬁnite jump intensity

In this section, we are going to construct general Lévy processes from Poisson point processes and Brownian motion. Afterwards, we will consider several important classes of Lévy jump processes with inﬁnite jump intensity.

Construction from Poisson point processes

Let ν(dy) be a positive measure on Rd 0 such that (1 y 2) ν(dy) < , i.e., \{ } ∧| | ∞ ν( y >ε) < for any ε> 0, and (1.19) | | ∞ y 2 ν(dy) < . (1.20) y 1 | | ∞ | |≤ Note that the condition (1.19) is necessary for the existence of a Lévy process with jump intensity ν. Indeed, if (1.19) would be violated for some ε > 0 then a corresponding

Stochastic Analysis Andreas Eberle 1.3. LÉVYPROCESSESWITHINFINITEJUMPINTENSITY 35

Lévy process should have inﬁnitely many jumps of size greater than ε in ﬁnite time. This contradicts the càdlàg property of the paths. The square integrability condition (1.20) controls the intensity of small jumps. It is crucial for the construction of a Lévy process with jump intensity ν given below, and actually it turns out to be also necessary for the existence of a corresponding Lévy process.

In order to construct the Lévy process, let N (dy), t 0, be a Poisson point process t ≥ with intensity measure ν deﬁned on a probability space (Ω, ,P ), and let N (dy) := A t Nt(dy) t ν(dy) denote the compensated process. For a measure and a measurable − set A, we denote by A(B) = (B A) ∩ A the part of the measure on the set A, i.e., (dy)= IA(y)(dy). The following decomposition property is immediate from the deﬁnition of a Poisson point process:

Remark (Decomposition of Poisson point processes). If A, B (Rd 0 ) are ∈ B \{ } A B disjoint sets then (Nt )t 0 and (Nt )t 0 are independent Poisson point processes with ≥ ≥ intensity measures νA, νB respectively.

If A B (y) = for some ε > 0 then the measure νA has ﬁnite total mass νA(Rd) = ∩ e ∅ ν(A) by (1.19). Therefore,

A A Xt := y Nt(dy) = y Nt (dy) A is a compound Poisson process with intensity measure νA, and characteristic exponent

ψ A (p) = (1 exp(ip y)) ν(dy). X − A On the other hand, if y 2 ν(dy) < then A | | ∞ A A Mt = y Nt(dy) = y Nt (dy) A is a square integrable martingale. If both conditions are satisﬁed simultaneously then

M A = XA t y ν(dy). t t − A UniversityofBonn WinterSemester2012/2013 36 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES

In particular, in this case M A is a Lévy process with characteristic exponent

ψ A (p) = (1 exp(ip y)+ ip y) ν(dy). M − A By (1.19) and (1.20), we are able to construct a Lévy process with jump intensity measure ν that is given by

r Xt = y Nt(dy) + y Nt(dy). (1.21) y >r y r | | | |≤ for any r (0, ). Indeed, let ∈ ∞

r Xt := y Nt(dy) = yI y >r N(dsdy), and (1.22) y >r (0,t] Rd {| | } | | × ε,r Mt := y Nt(dy). (1.23) ε< y r | |≤ for ε,r [0, ) with ε

Theorem 1.13 (Existence of Lévy processes with inﬁnite jump intensity). Let ν be a positive measure on Rd 0 satisfying (1 y 2) ν(dy) < . \{ } ∧| | ∞ r 1) For any r > 0, (Xt ) is a compound Poisson process with intensity measure r ν (dy)= I y >r ν(dy). {| | } 0,r 2) The process (Mt ) is a Lévy martingale with characteristic exponent

ip y d ψr(p) = (1 e + ip y) ν(dy) p R . (1.24) y r − ∀ ∈ | |≤ Moreover, for any u (0, ), ∈ ∞

ε,r 0,r 2 E sup Mt Mt 0 as ε 0. (1.25) 0 t u | − | → ↓ ≤ ≤

0,r r r r 0,r 3) The Lévy processes (Mt ) and (Xt ) are independent, and Xt := Xt + Mt is a Lévy process with characteristic exponent ip y d ψr(p) = 1 e + ip yI y r ν(dy) p R . (1.26) − {| |≤ } ∀ ∈ Stochastic Analysis Andreas Eberle 1.3. LÉVYPROCESSESWITHINFINITEJUMPINTENSITY 37

Proof. 1) is a consequence of Theorem 1.8. 0,r 2) By (1.20), the stochastic integral (Mt ) is a square integrable martingale on [0,u] for any u (0, ). Moreover, by the Itô isometry, ∈ ∞ u 0,r ε,r 2 0,ε 2 2 M M M 2([0,u]) = M M 2([0,u]) = y I y ε ν(dy) dt 0 − | | {| |≤ } → 0 as ε 0. By Theorem 1.8, (M ε,r) is a compensated compound Poisson process with ↓ t intensity I ε< y r ν(dy) and characteristic exponent { | |≤ }

ip y ψε,r(p) = (1 e + ip y) ν(dy). ε< y r − | |≤ ip y 2 As ε 0, ψε,r(p) converges to ψr(p) since 1 e + ip y = ( y ). Hence the limit ↓ 0,r 1/n,r − O | | martingale Mt = lim Mt also has independent and stationary increments, and n characteristic function→∞

0,r 1/n,1 E[exp(ip Mt )] = lim E[exp(ip Mt )] = exp( tψr(p)). n →∞ −

3) Since I y r Nt(dy) and I y >r Nt(dy) are independent Poisson point processes, {| |≤ } {| | } 0,r r r 0,r r the Lévy processes (Mt ) and (Xt ) are also independent. Hence Xt = Mt + Xt is a Lévy process with characteristic exponent ip y ψr(p)= ψr(p)+ (1 e ) ν(dy). y >r − | |

Remark. All the partially compensated processes (Xr), r (0, ), are Lévy pro- t ∈ ∞ cesses with jump intensity ν. Actually, these processes differ only by a ﬁnite drift term, since for any 0 <ε

ε r Xt = Xt + bt, where b = y ν(dy). ε< y r | |≤ A totally uncompensated Lévy process

Xt = lim y Nt(dy) n y 1/n →∞ | |≥ does exist only under additional assumptions on the jump intensity measure:

UniversityofBonn WinterSemester2012/2013 38 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES

Corollary 1.14 (Existence of uncompensated Lévy jump processes). Suppose that (1 y ) ν(dy) < , or that ν is symmetric (i.e., ν(B) = ν( B) for any B ∧ | | ∞ − ∈ (Rd 0 )) and (1 y 2) ν(dy) < . Then there exists a Lévy process (X ) with B \{ } ∧| | ∞ t characteristic exponent

ip y d ψ(p) = lim 1 e ν(dy) p R (1.27) ε 0 y >ε − ∀ ∈ ↓ | | such that ε 2 E sup Xt Xt 0 as ε 0. (1.28) 0 t u | − | → ↓ ≤ ≤ Proof. For 0 <ε

ε r ε,r Xt = Xt + Mt + t y ν(dy). ε< y r | |≤ As ε 0, M ε,r converges to M 0,r in M 2([0,u]) for any ﬁnite u. Moreover, under the ↓ assumption imposed on ν, the integral on the right hand side converges to bt where

b = lim y ν(dy). ε 0 ε< y r ↓ | |≤ Therefore, (Xε) converges to a process (X ) in the sense of (1.28) as ε 0. The t t ↓ limit process is again a Lévy process, and, by dominated convergence, the characteristic exponent is given by (1.27).

Remark (Lévy processes with finite variation paths). If (1 y ) ν(dy) < then ∧| | ∞ the process Xt = y Nt(dy) is defined as a Lebesgue integral. As remarked above, in that case the paths of (Xt) are almost surely of finite variation:

V (1)(X) y N (dy) < a.s. t ≤ | | t ∞ The Lévy-Itô decomposition

We have constructed Lévy processes corresponding to a given jump intensity measure ν under adequate integrability conditions as limits of compound Poisson processes or partially compensated compound Poisson processes, respectively. Remarkably, it turns out that by taking linear combinations of these Lévy jump processes and Gaussian Lévy

Stochastic Analysis Andreas Eberle 1.3. LÉVYPROCESSESWITHINFINITEJUMPINTENSITY 39 processes, we obtain all Lévy processes. This is the content of the Lévy-Itô decomposition theorem that we will now state before considering in more detail some important classes of Lévy processes. Already the classical Lévy-Khinchin formula for infinity divisible random variables (see Corollary 1.16 below) shows that any Lévy process on Rd can be characterized by three d d d quantities: a non-negative definite symmetric matrix a R × , a vector b R , and a ∈ ∈ σ-finite measure ν on (Rd 0 ) such that B \{ } (1 y 2) ν(dy) < . (1.29) ∧| | ∞ Note that (1.29) holds if and only if ν is finite on complements of balls around 0, and 2 y 1 y ν(dy) < . The Lévy-Itô decomposition gives an explicit representation of | |≤ | | ∞ a Lévy process with characteristics (a, b, ν). d d T Let σ R × with a = σσ , let (B ) be a d-dimensional Brownian motion, and let (N ) ∈ t t be an independent Poisson point process with intensity measure ν. We define a Lévy process (Xt) by setting

Xt = σBt + bt + y Nt(dy)+ y (Nt(dy) tν(dy)) . (1.30) y >1 y 1 − | | | |≤ The ﬁrst two summands are the diffusion part and the drift of a Gaussian Lévy process, the third summand is a pure jump process with jumps of size greater than 1, and the last summand represents small jumps compensated by drift. As a sum of independent Lévy processes, the process (Xt) is a Lévy process with characteristic exponent

1 ip y ψ(p) = p ap ib p + (1 e + ip yI y 1 ) ν(dy). (1.31) 2 − Rd 0 − {| |≤ } \{ } We have thus proved the ﬁrst part of the following theorem:

Theorem 1.15 (Lévy-Itô decomposition). 1) The expression (1.30) deﬁnes a Lévy process with characteristic exponent ψ.

2) Conversely, any Lévy process (Xt) can be decomposed as in (1.30) with the Pois- son point process

N = ∆X , t 0, (1.32) t s ≥ s t ∆X≤s=0 UniversityofBonn WinterSemester2012/2013 40 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES

d d d an independent Brownian motion (B ), a matrix σ R × , a vector b R , and t ∈ ∈ a σ-finite measure ν on Rd 0 satisfying (1.29). \{ } We will not prove the second part of the theorem here. The principal way to proceed is to define (Nt) via (1.29), and to consider the difference of (Xt) and the integrals w.r.t. (Nt) on the right hand side of (1.30). One can show that the difference is a continuous Lévy process which can then be identified as a Gaussian Lévy process by the Lévy characterization, cf. Section 4.1 below. Carrying out the details of this argument, however, is still a lot of work. A detailed proof is given in [3].

As a byproduct of the Lévy-Itô decomposition, one recovers the classical Lévy-Khinchin formula for the characteristic functions of inﬁnitely divisible random variables, which can also be derived directly by an analytic argument.

Corollary 1.16 (Lévy-Khinchin formula). For a function ψ : Rd C the following → statements are all equivalent:

(i) ψ is the characteristic exponent of a Lévy process.

(ii) exp( ψ) is the characteristic function of an infinitely divisible random variable. − d d (iii) ψ satisfies (1.31) with a non-negative definite symmetric matrix a R × , a ∈ vector b Rd, and a measure ν on (Rd 0 ) such that (1 y 2) ν(dy) < . ∈ B \{ } ∧| | ∞ Proof. (iii) (i) holds by the first part of Theorem 1.15. ⇒ (i) (ii): If (X ) is a Lévy process with characteristic exponent ψ then X X is an ⇒ t 1 − 0 infinitely divisible random variable with characteristic function exp( ψ). − (ii) (iii) is the content of the classical Lévy-Khinchin theorem, see e.g. [16]. ⇒ We are now going to consider several important subclasses of Lévy processes. The class of Gaussian Lévy processes of type

Xt = σBt + bt

d d d with σ R × , b R , and a d-dimensional Brownian motion (B ) has already been ∈ ∈ t introduced before. The Lévy-Itô decomposition states in particular that these are the only Lévy processes with continuous paths!

Stochastic Analysis Andreas Eberle 1.3. LÉVYPROCESSESWITHINFINITEJUMPINTENSITY 41

Subordinators

A subordinator is by deﬁnition a non-decreasing real-valued Lévy process. The name comes from the fact that subordinators are used to change the time-parametrization of a Lévy process, cf. below. Of course, the deterministic processes X = bt with b 0 are t ≥ subordinators. Furthermore, any compound Poisson process with non-negative jumps is a subordinator. To obtain more interesting examples, we assume that ν is a positive measure on (0, ) with ∞ (1 y) ν(dy) < . (0, ) ∧ ∞ ∞

Then a Poisson point process (Nt) with intensity measure ν satisﬁes almost surely

supp(N ) [0, ) for any t 0. t ⊂ ∞ ≥ Hence the integrals X = y N (dy) , t 0, t t ≥ deﬁne a non-negative Lévy process with X0 =0. By stationarity, all increments of (Xt) are almost surely non-negative, i.e., (Xt) is increasing. In particular, the sample paths are (almost surely) of ﬁnite variation.

Example (Gamma process). The Gamma distributions form a convolution semigroup of probability measures on (0, ), i.e., ∞ Γ(r, λ) Γ(s,λ) = Γ(r + s,λ) for any r,s,λ > 0. ∗

Therefore, for any a,λ > 0 there exists an increasing Lévy process (Γt)t 0 with incre- ≥ ment distributions

Γ Γ Γ(at, λ) for any s, t 0. t+s − s ∼ ≥ Computation of the Laplace transform yields

at u − ∞ uxy 1 λy E[exp( uΓt)] = 1+ = exp t (1 e− )ay− e− dy (1.33) − λ − 0 − UniversityofBonn WinterSemester2012/2013 42 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES for every u 0, cf. e.g. [25, Lemma 1.7]. Since Γ 0, both sides in (1.33) have a ≥ t ≥ unique analytic extension to u C : (u) 0 . Therefore, we can replace u by ip { ∈ ℜ ≥ } − in (1.33) to conclude that the characteristic exponent of (Γt) is

∞ ipy 1 λy ψ(p) = (1 e ) ν(dy), where ν(dy)= ay− e− dy. − 0 Hence the Gamma process is a non-decreasing pure jump process with jump intensity measure ν.

Example (Inverse Gaussian processes). An explicit computation of the characteristic function shows that the Lévy subordinator (Ts) is a pure jump Lévy process with Lévy measure 1/2 3/2 ν(dy) = (2π)− y− I(0, )(y) dx. ∞

More generally, if Xt = σBt + bt is a Gaussian Lévy process with coefﬁcients σ > 0, b R, then the right inverse ∈ T X = inf t 0 : X = s , s 0, s { ≥ t } ≥ is a Lévy jump process with jump intensity

1/2 3/2 2 ν(dy) = (2π)− y− exp( b y/2)I(0, )(y) dy. − ∞ Remark (Finite variation Lévy jump processes on R1).

Suppose that (Nt) is a Poisson point process on R 0 with jump intensity measure ν \{ } (0, ) ( ,0) satisfying (1 y ) ν(dy) < . Then the decomposition N = N ∞ + N −∞ into ∧| | ∞ t t t the independent restrictions of (Nt) to R+, R respectively induces a corresponding − decomposition

(0, ) ( ,0) Xt = Xtր + Xtց , Xtր = y Nt ∞ (dy) , Xtց = y Nt −∞ (dy), of the associated Lévy jump process Xt = y Nt(dy) into a subordinator Xtր and a decreasing Lévy process Xtց. In particular, we see once more that (Xt) has almost surely paths of ﬁnite variation.

An important property of subordinators is that they can be used for random time transformations of Lévy processes:

Stochastic Analysis Andreas Eberle 1.3. LÉVYPROCESSESWITHINFINITEJUMPINTENSITY 43

Exercise (Time change by subordinators). Suppose that (Xt) is a Lévy process with Laplace exponent η : R R, i.e., X + → E[exp( αX )] = exp( tη (α)) for any α 0. − t − X ≥

Prove that if (Ts) is an independent subordinator with Laplace exponent ηT then the time-changed process X := X , s 0, s Ts ≥ is again a Lévy process with Laplace exponent

η(p) = ηT (ηX (p)).

The characteristic exponent can be obtained from this identity by analytic continuation.

Example (Subordinated Lévy processes). Let (Bt) be a Brownian motion.

1) If (Nt) is an independent Poisson process with parameter λ > 0 then (BNt ) is a compensated Poisson process with Lévy measure

1/2 2 ν(dy) = λ(2π)− exp( y /2) dy. − 2) If (Γ ) is an independent Gamma process then for σ, b R the process t ∈

Xt = σBΓt + bΓt

is called a Variance Gamma process. It is a Lévy process with characteristic exponent ψ(p)= (1 eipy) ν(dy), where −

1 λy y ν(dy) = c y − e− I(0, )(y)+ e− | |I( ,0)(y) dy | | ∞ −∞ with constants c,λ, > 0. In particular, a Variance Gamma process satisﬁes X =Γ(1) Γ(2) with two independent Gamma processes. Thus the increments of t t − t (Xt) have exponential tails. Variance Gamma processes have been introduced and applied to option pricing by Madan and Seneta [28] as an alternative to Brownian motion taking into account longer tails and allowing for a wider modeling of skewness and kurtosis.

UniversityofBonn WinterSemester2012/2013 44 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES

3) Normal Inverse Gaussian (NIG) processes are time changes of Brownian motions with drift by inverse Gaussian subordinators [4]. Their increments over unit time intervals have a normal inverse Gaussian distribution, which has slower de- caying tails than a normal distribution. NIG processes are applied in statistical modelling in ﬁnance and turbulence.

Stable processes

We have noted in (1.12) that the jump intensity measure of a strictly α-stable process in R1 is given by

1 α ν(dy) = c+I(0, )(y)+ c I( ,0)(y) y − − dy (1.34) ∞ − −∞ | | with constants c+,c [0, ). Note that for any α (0, 2), the measure ν is ﬁnite on − ∈ ∞ ∈ R 2 ( 1, 1), and [ 1,1] y ν(dy) < . \ − − | | ∞ We will prove now that if α (0, 1) (1, 2) then for each choice of the constants c ∈ ∪ + and c , there is a strictly α-stable process with Lévy measure (1.34). For α = 1 this − is only true if c+ = c , whereas a non-symmetric 1-stable process is given by Xt = bt − with b R 0 . To deﬁne the corresponding α-stable processes, let ∈ \{ }

ε Xt = y Nt(dy) R [ ε,ε] \ − where (N ) is a Poisson point process with intensity measure ν. Setting X = t || ||u 2 1/2 E[supt a Xt ] , an application of Theorem 1.13 yields: ≤ | | Corollary 1.17 (Construction of α-stable processes). Let ν be the probability measure on R 0 deﬁned by (1.34) with c+,c [0, ). \{ } − ∈ ∞

1) If c+ = c then there exists a symmetric α-stable process X with characteristic − exponent ψ(p)= γ p α, γ = (1 cos y) ν(dy) R, such that X1/n X | | − ∈ || − ||u → 0 for any u (0, ). ∈ ∞ 2) If α (0, 1) then (1 y ) ν(dy) < , and X = y N (dy) is an α-stable ∈ ∧| | ∞ t t process with characteristic exponent ψ(p)= z p α, z = 1 eiy ν(dy) C. | | − ∈ Stochastic Analysis Andreas Eberle 1.3. LÉVYPROCESSESWITHINFINITEJUMPINTENSITY 45

3) For α =1 and b R, the deterministic process X = bt is α-stable with charac- ∈ t teristic exponent ψ(p)= ibp. −

4) Finally, if α (1, 2) then ( y y 2) ν(dy) < , and the compensated process ∈ | |∧| | ∞ Xt = y Nt(dy) is an α-stable martingale with characteristic exponent ψ(p) = α iy z p , z = (1 e + iy) ν(dy). | | − Proof.By Theorem 1.13 it is sufﬁcient to prove convergence of the characteristic exponents

ψ (p) = 1 eipy ν(dy) = p α 1 eix ν(dx), ε − | | − R [ε,ε] R [ εp,εp] \ − \ − ψ (p) = 1 eipy + ipy ν(dy) = p α 1 eix + ix ν(dx) ε − | | − R R [ ε,ε] [ εp,εp] \ − \ − to ψ(p), ψ(p) respectively as ε 0. This is easily veriﬁed in cases 1), 2) and 4) by ↓ ix ix 2 ix noting that 1 e + 1 e− = 2(1 cos x) = (x ), 1 e = ( x ), and − − − O − O | | 1 eix + ix = ( x 2). − O | |

Notice that although the characteristic exponents in the non-symmetric cases 2), 3) and 4) above take a similar form (but with different constants), the processes are actually very different. In particular, for α > 1, a strictly α-stable process is always a limit of compensated compound Poisson processes and hence a martingale!

Example (α-stable subordinators vs. α-stable martingales). For c = 0 and α − ∈ (0, 1), the α-stable process with jump intensity ν is increasing, i.e., it is an α-stable subordinator. For c = 0 and α (1, 2) this is not the case since the jumps are − ∈ “compensated by an inﬁnite drift”. The graphics below show simulations of samples from α-stable processes for c =0 and α =3/2, α =1/2 respectively. For α (0, 2), − ∈ a symmetric α-stable process has the same law as (√2BTs ) where (Bt) is a Brownian motion and (Ts) is an independent α/2-stable subordinator.

UniversityofBonn WinterSemester2012/2013 46 CHAPTER 1. LÉVY PROCESSES AND POISSON POINT PROCESSES

100

500 1000 1500 2000 2500 3000

-100

-200

5 ´ 106

4 ´ 106

3 ´ 106

2 ´ 106

1 ´ 106

500 1000 1500 2000 2500 3000

Stochastic Analysis Andreas Eberle Chapter 2

Stochastic integrals and Itô calculus for semimartingales

[Stochastic calculus for semimartingales] Our aim in this chapter is to develop a stochas- (1) (2) (d) tic calculus for functions of ﬁnitely many real-valued stochastic processes Xt ,Xt ,...,Xt . In particular, we will make sense of stochastic differential equations of type

d (k) dYt = σk(t, Yt ) dXt − k=1

n n with continuous time-dependent vector fields σ1,...,σd : R+ R R . The sample (k) × → paths of the driving processes (Xt ) and of the solution (Yt) may be discontinuous, but we will always assume that they are càdlàg, i.e., right-continuous with left limits. In most relevant cases this can be assured by choosing an appropriate modification. For example, a martingale or a Lévy process w.r.t.a right-continuous complete filtration always has a càdlàg modification, cf. [33, Ch.II, §2] and [32, Ch.I Thm.30].

An adequate class of stochastic processes for which a stochastic calculus can be developed are semimartingales, i.e., sums of local martingales and adapted finite variation processes with càdlàg trajectories. To understand why this is a reasonable class of processes to consider, we first briefly review the discrete time case.

47 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 48 SEMIMARTINGALES

Semimartingales in discrete time

If ( ) is a discrete-time ﬁltration on a probability space (Ω, ,P ) then any Fn n=0,1,2,... A ( ) adapted integrable stochastic process (X ) has a unique Doob decomposition Fn n

X = X + M + Aր Aց (2.1) n 0 n n − n into an ( ) martingale (M ) and non-decreasing predictable processes (Aր) and (Aց) Fn n n n such that M0 = A0ր = A0ց = 0, cf. [13, Thm. 2.4]. The decomposition is determined by choosing

Mn Mn 1 = Xn Xn 1 E[Xn Xn 1 n 1], − − − − − − − |F −

+ Anր Anր 1 = E[Xn Xn 1 n 1] , and Anց Anց 1 = E[Xn Xn 1 n 1]−. − − − − |F − − − − − |F −

In particular, (Xn) is a sub- or supermartingale if and only if Anց = 0 for any n, or

Anր =0 for any n, respectively. The discrete stochastic integral

(G X)n = Gk (Xk Xk 1) • − − k=1 of a bounded predictable process (Gn) w.r.t. (Xn) is again a martingale if (Xn) is a martingale, and an increasing (decreasing) process if G 0 for any n, and (X ) n ≥ n is increasing (respectively decreasing). For a bounded adapted process (Hn), we can deﬁne correspondingly the integral

(H X)n = Hk 1 (Xk Xk 1) − • − − − k=1 of the predictable process H =(Hk 1)k N w.r.t. X. − − ∈ The Taylor expansion of a function F C2(R) yields a primitive version of the Itô ∈ formula in discrete time. Indeed, notice that for k N, ∈ 1 F (Xk) F (Xk 1) = F ′(Xk 1 + s∆Xk) ds ∆Xk − − 0 − 1 s 2 = F ′(Xk 1) ∆Xk + F ′′(Xk 1 + r∆Xk) drds (∆Xk) . − − 0 0 Stochastic Analysis Andreas Eberle 49

where ∆Xk := Xk Xk 1. By summing over k, we obtain − − n 1 s 2 F (Xn) = F (X0)+(F ′(X) X)n + F ′′(Xk 1 + r∆Xk) drds (∆Xk) . − • 0 0 − k=1 Itô’s formula for a semimartingale (Xt) in continuous time will be derived in Theorem 2.22 below. It can be rephrased in a way similartothe formula above, where the last term on the right-hand side is replaced by an integral w.r.t. the quadratic variation process

[X]t of X, cf. (XXX).

Semimartingales in continuous time

In continuous time, it is no longer true that any adapted process can be decomposed into a local martingale and an adapted process of ﬁnite variation (i.e., the sum of an increasing and a decreasing process). A counterexample is given by fractional Brownian motion, cf.Section 2.3 below. On the other hand, a large class of relevant processes has a corresponding decomposition.

Deﬁnition. Let ( t)t 0 be a ﬁltration. A real-valued ( t)-adapted stochastic process F ≥ F (Xt)t 0 on a probability space (Ω, ,P ) is called an (Ft) semimartingale if and only ≥ A if it has a decomposition

X = X + M + A , t 0, (2.2) t 0 t t ≥ into a strict local ( )-martingale (M ) with càdlàg paths, and an ( )-adapted process Ft t Ft (At) with càdlàg ﬁnite-variation paths such that M0 = A0 =0.

Here a strict local martingale is a process that can be localized by martingales with uniformly bounded jumps, see Section 2.2 for the precise definition. Any continuous local martingale is strict. In general, it can be shown that if the filtration is right continuous and complete then any local martingale can be decomposed into a strict local martingale and an adapted finite variation process (“Fundamental Theorem of Local Martingales”, cf. [32]). Therefore, the notion of a semimartingale defined above is not changed if the word “strict” is dropped in the definition. Since the non-trivial proof of the Fundamental Theorem of Local Martingales is not included in these notes, we nevertheless stick to the definition above.

UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 50 SEMIMARTINGALES

Remark. (Assumptions on path regularity). Requiring (At) to be càdlàg is just a standard convention ensuring in particular that t A (ω) is the distribution function of → t a signed measure. The existence of right and left limits holds for any monotone function, and, therefore, for any function of finite variation. Similarly, every local martingale w.r.t.a right-continuous complete filtration has a càdlàg modification.

Without additional conditions on (At), the semimartingale decomposition in (2.2) is not unique, see the example below. Uniqueness holds if, in addition, (At) is assumed to be predictable, cf. [5,32]. Under the extra assumption that (At) is continuous, uniqueness is a consequence of Corollary 2.15 below.

Example (Semimartingale decompositions of a Poisson process). An ( ) Poisson Ft process (Nt) with intensity λ has the semimartingale decompositions

Nt = Nt + λt = 0+ Nt into a martingale and an adapted finite variation process. Only in the first decomposition, the finite variation process is predictable and continuous respectively.

The following examples show that semimartingales form a sufﬁciently rich class of stochastic processes.

Example (Stochastic integrals). Let (B ) and (N ) be a d-dimensional ( ) Brownian t t Ft motion and an ( ) Poisson point process on a σ-ﬁnite measure space (S, , ν) respec- Ft S tively. Then any process of the form

t t Xt = Hs dBs+ Gs(y)N(ds dy)+ Ks ds+ Ls(y)N(ds dy) (2.3) 0 (0,t] S 0 (0,t] S × × is a semimartingale provided the integrands H,G,K,L are predictable, H and G are (locally) square integrable w.r.t. P λ, P λ ν respectively, and K and L are ⊗ ⊗ ⊗ (locally) integrable w.r.t. these measures. In particular, by the Lévy-Itô decomposition, every Lévy process is a semimartingale. Similarly, the components of solutions of SDE driven by Brownian motions and Poisson point processes are semimartingales. More generally, Itô’s formula yields an explicit semimartingale decomposition of f(t, Xt) for an arbitrary function f 2 (R Rn) and (X ) as above, cf. Section 2.4 below. ∈ C + × t

Stochastic Analysis Andreas Eberle 51

Example (Functions of Markov processes). If (X ) is a time-homogeneous ( ) t Ft Markov process on a probability space (Ω, ,P ), and f is a function in the domain A of the generator , then f(X ) is a semimartingale with decomposition L t t f(X )= local martingale + ( f)(X ) ds, (2.4) t L s 0 cf.e.g. [11] or [15]. Indeed, it is possible to deﬁne the generator of a Markov process L through a solution to a martingale problem as in (2.4).

Many results for continuous martingales carry over to the càdlàg case. However, there are some important differences and pitfalls to be noted: Exercise (Càdlàg processes).

1) A stopping time is called predictable iff there exists an increasing sequence (Tn) of stopping times such that T < T on T > 0 and T = sup T . Show that for n { } N a càdlàg stochastic process (Xt)t 0, the ﬁrst hitting time ≥ T = inf t 0 : X A A { ≥ t ∈ } of a closed set A R is not predictable in general. ⊂ 2) Prove that for a right continuous ( t) martingale (Mt)t 0 and an ( t) stopping F ≥ F time T , the stopped process (Mt T )t 0 is again an ( t) martingale. ∧ ≥ F 3) Prove that a càdlàg local martingale (Mt) can be localized by a sequence (Mt Tn ) ∧ of bounded martingales provided the jumps of (Mt) are uniformly bounded, i.e.,

sup ∆M (ω) : t 0, ω Ω < . {| t | ≥ ∈ } ∞ 4) Give an example of a càdlàg local martingale that can not be localized by bounded martingales.

Our next goal is to deﬁne the stochastic integral G X w.r.t. a semimartingale X for • the left limit process G = (Ht ) of an adapted càdlàg process H, and to build up a − corresponding stochastic calculus. Before studying integration w.r.t. càdlàg martingales in Section 2.2, we will consider integrals and calculus w.r.t. ﬁnite variation processes in Section 2.1.

UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 52 SEMIMARTINGALES

2.1 Finite variation calculus

In this section we extend Stieltjes calculus to càdlàg paths of ﬁnite variation. The results are completely deterministic. They will be applied later to the sample paths of the ﬁnite variation part of a semimartingale.

Fix u (0, ], and let A : [0,u) R be a right-continuous function of ﬁnite variation. ∈ ∞ → In particular, A is càdlàg. We recall that there is a σ-ﬁnite measure A on (0,u) with distribution function A, i.e.,

((s, t]) = A A for any 0 s t < u. (2.5) A t − s ≤ ≤

The function A has the decomposition

c d At = At + At (2.6) into the pure jump function

d At := ∆As (2.7) s t ≤ and the continuous function Ac = A Ad. Indeed, the series in (2.7) converges abso- t t − t lutely since

∆A V (1)(A) < for any t [0,u). | s| ≤ t ∞ ∈ s t ≤

The measure A can be decomposed correspondingly into

A = Ac + Ad where

d = ∆A δ A s s s (0,u) ∈ ∆As=0 is the atomic part, and Ac does not contain atoms. Note that Ac is not necessarily absolutely continuous!

Stochastic Analysis Andreas Eberle 2.1. FINITE VARIATION CALCULUS 53

Lebesgue-Stieltjes integrals revisited

Let 1 ([0,u), ) := 1 ([0,u), ) where denotes the positive measure with Lloc A Lloc | A| | A| distribution function V (1)(A). For G 1 ([0,u), ), the Lebesgue-Stieltjes integral t ∈ Lloc A of H w.r.t. A is deﬁned as u G dA = G I (r) (dr) for 0 s t < u. r r r (s,t] A ≤ ≤ s A crucial observation is that the function t I := G dA = G (dr) , t [0,u), t r r r A ∈ 0 (0,t] is the distribution function of the measure

I (dr) = Gr A(dr) with density G w.r.t. A. This has several important consequences: 1) The function I is again càdlàg and of ﬁnite variation with

t t V (1)(I) = G (dr) = G dV (1)(A). t | r|| A| | r| r 0 0 2) I decomposes into the continuous and pure jump parts

t t c c d d It = Gr dAr , It = Gr dAr = Gs ∆As. 0 0 s t ≤ 3) For any G 1 ( ), ∈Lloc I t t Gr dIr = GrGr dAr, 0 0 i.e., if “dI = G dA” then also “G dI = GG dA ”.

Theorem 2.1 (Riemann sum approximations for Lebesgue-Stieltjes integrals). Sup- pose that H : [0,u) R is a càdlàg function. Then for any a [0,u) and for any → ∈ sequence (π ) of partitions with mesh(π ) 0, n n → t lim Hs(As′ t As) = Hs dAs uniformly for t [0, a]. n ∧ − − ∈ →∞ s πn 0 s

UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 54 SEMIMARTINGALES

Remark. If (At) is continuous then

t t Hs dAs = Hs dAs, − 0 0 t because 0 ∆HsdAs = s t ∆Hs∆As = 0 for any càdlàg function H. In general, ≤ however, the limit of the Riemann sums in Theorem 2.1 takes the modiﬁed form t t c Hs dAs = Hs dAs + Hs ∆As. − − 0 0 s t ≤ Proof. For n N and t 0, ∈ ≥

′ Hs(As t As) = Hs dAr = H r n dAr ∧ − ′ ⌊ ⌋ s πn s πn (s,s t] (0,t] s

t t

sup H r n dAr Hr dAr H r n Hr A (dr) 0 t a 0 ⌊ ⌋ − 0 − ≤ (0,a] ⌊ ⌋ − − | | → ≤ as n by dominated convergence. →∞

Product rule

The covariation [H, A] of two functions H, A : [0,u) R w.r.t. a sequence (π ) of → n partitions with mesh(π ) 0 is deﬁned by n →

[H, A]t = lim (Hs′ t Hs)(As′ t As), (2.8) n ∧ − ∧ − →∞ s πn s

Lemma 2.2. If H and A are càdlàg and A has ﬁnite variation then the covariation exists and is independently of (πn) given by

[H, A]t = ∆Hs∆As 0

~~Stochastic Analysis Andreas Eberle 2.1. FINITE VARIATION CALCULUS 55~~

~~Proof. We again represent the sums as integrals:~~

~~t ′ ′ (Hs t Hs)(As t As) = (H r n t H r n ) dAr ∧ − ∧ − ⌈ ⌉ ∧ − ⌊ ⌋ s πn 0 s~~

~~Remark. 1) If H or A is continuous then [H, A]=0. 2) In general, the proof above shows that~~

~~t t Hs dAs = Hs dAs + [H, A]t, − 0 0 i.e., [H, A] is the difference between limits of right and left Riemann sums.~~

~~Theorem 2.3 (Integration by parts, product rule). Suppose that H, A : [0,u) R → are right continuous functions of ﬁnite variation. Then~~

~~t t HtAt H0A0 = Hr dAr + Ar dHr + [H, A]t for any t [0,u). (2.9) − − − ∈ 0 0 In particular, the covariation [H, A] is a càdlàg function of ﬁnite variation, and for a~~

~~d(HA)r = Hr dAr + Ar dHr + d[H, A]r. − − As special cases we note that if H and A are continuous then HA is continuous with~~

~~d(HA)r = Hr dAr + Ar dHr, and if H and A are pure jump functions (i.e. Hc = Ac = 0) then HA is a pure jump function with~~

~~∆(HA)r = Hr ∆Ar + Ar ∆Hr + ∆Ar∆Hr . − − UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 56 SEMIMARTINGALES~~

~~In the latter case, (2.9) implies~~

~~H A H A = ∆(HA) . t t − 0 0 r r t ≤ Note that this statement is not completely trivial, as it holds even when the jump times of HA form a countable dense subset of [0, t]!~~

~~Since the product rule is crucial but easy to prove, we give two proofs of Theorem 2.3:~~

~~Proof 1. For (π ) with mesh(π ) 0, we have n n →~~

~~HtAt H0A0 = (Hs′ tAs′ t HsAs) − ∧ ∧ − s πn s~~

= Hs(As′ t As)+ As(Hs′ t Hs)+ (As′ t As)(Hs′ t Hs). ∧ − ∧ − ∧ − ∧ − As n , (2.9) follows by Theorem 2.1 above. Moreover, the convergence of the → ∞ covariation is uniform for t [0, a], a

~~s>r~~

(Ht H0)(At A0) = H (dr) A(ds) − − (0,t] (0,t] × is the area of (0, t] (0, t] w.r.t. the product measure . By dividing the square × H ⊗ A (0, t] (0, t] into the parts (s,r) sr and the diagonal (s,r) s = r × { | } { | } { | } we see that this area is given by

~~t t + + = (Ar A0) dHr + (Hs H0) dAs + ∆Hs∆As, − − − − sr s=r 0 0 s t ≤ The assertion follows by rearranging terms in the resulting equation.~~

~~Stochastic Analysis Andreas Eberle 2.1. FINITE VARIATION CALCULUS 57~~

~~Chain rule~~

~~The chain rule can be deduced from the product rule by iteration and approximation of 1 functions by polynomials: C~~

Theorem 2.4 (Change of variables, chain rule, Itô formula for ﬁnite variation functions). Suppose that A : [0,u) R is right continuous with ﬁnite variation, and let → F 1(R). Then for any t [0,u), ∈ C ∈

~~t F (At) F (A0) = F ′(As ) dAs + (F (As) F (As ) F ′(As )∆As) , − − − − − − 0 s t ≤ (2.10) or, equivalently,~~

~~t c F (At) F (A0) = F ′(As ) dAs + (F (As) F (As )) . (2.11) − − − − 0 s t ≤~~

If A is continuous then F (A) is also continuous, and (2.10) reduces to the standard chain rule t F (A ) F (A ) = F ′(A ) dA . t − 0 s s 0 If A is a pure jump function then the theorem shows that F (A) is also a pure jump function (this is again not completely obvious!) with

F (At) F (A0) = (F (As) F (As )) . − − − s t ≤ Remark. Note that by Taylor’s theorem, the sum in (2.10) converges absolutely when- 2 ever s t(∆As) < . This observation will be crucial for the extension to Itô’s ≤ ∞ formula for processes with ﬁnite quadratic variation, cf. Theorem 2.22 below.

Proof of Theorem 2.4. Let denote the linear space consisting of all functions F A ∈ 1(R) satisfying (2.10). Clearly the constant function 1 and the identity F (t)= t are in C UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 58 SEMIMARTINGALES

~~. We now prove that is an algebra: Let F,G . Then by the integration by parts A A ∈A identity and by (2.11),~~

(FG)(A ) (FG)(A ) t − 0 t t = F (As ) dG(A)s + G(As ) dF (A)s + ∆F (A)s∆G(A)s − − 0 0 s t ≤ t c = (F (As )G′(As )+ G(As )F ′(As )) dAs − − − − 0 + (F (As )∆G(A)s + G(As )∆F (A)s + ∆F (A)s∆G(A)s) − − s t ≤ t c = (FG)′(As ) dAs + ((FG)(As) (FG)(As )) − − − 0 s t ≤ for any t [0,u), i.e., FG is in . ∈ A Since is an algebra containing 1 and t, it contains all polynomials. Moreover, if F A is an arbitrary 1 function then there exists a sequence (p ) of polynomials such that C n p F and p′ F ′ uniformly on the bounded set A s t . Since (2.11) holds n → n → { s | ≤ } for the polynomials pn, it also holds for F .

~~Exponentials of ﬁnite variation functions~~

~~Let A : [0, ) R be a right continuous finite variation function. The exponen- ∞ → tial of A is defined as the right-continuous finite variation function (Zt)t 0 solving the ≥ equation~~

dZt = Zt dAt , Z0 =1 , i.e., − t Zt = 1+ Zs dAs for any t 0. (2.12) − ≥ 0 If A is continuous then Zt = exp(At) solves (2.12) by the chain rule. On the other hand, if A is piecewise constant with ﬁnitely many jumps then Zt = s t(1+∆As) solves ≤ (2.12), since

~~Zt = Z0 + ∆Zs = 1+ Zs ∆As = 1+ Zs dAs. − − s t s t (0,t] ≤ ≤ In general, we obtain:~~

~~Stochastic Analysis Andreas Eberle 2.1. FINITE VARIATION CALCULUS 59~~

~~Theorem 2.5. The unique càdlàg function solving (2.12) is~~

~~Z = exp(Ac) (1+∆A ), (2.13) t t s s t ≤ where the product converges for any t 0. ≥ Proof. 1) We ﬁrst show convergence of the product~~

~~Pt = (1+∆As). s t ≤ Recall that since A is càdlàg, there are only ﬁnitely many jumps with ∆A > 1/2. | s| Therefore, we can decompose~~

Pt = exp  log(1 + ∆As) (1+∆As) (2.14) s t s t ≤ ≤  ∆As 1/2  ∆As >1/2 | |≤  | |   in the sense that the product Pt converges if and only if the series converges. The series converges indeed absolutely for A with finite variation, since log(1+x) can be bounded by a constant times x for x 1/2. The limit S of the series defines a pure jump | | | | ≤ t function with variation V (1)(S) const. V (1)(A) for any t 0. t ≤ t ≥ 2) Equation for P : The chain and product rule now imply by (2.14) that t P is also t → t a finite variation pure jump function. Therefore,

t d Pt = P0 + ∆Ps = 1+ Ps ∆As = 1+ Ps dAs, t 0, − − ∀ ≥ s t s t 0 ≤ ≤ (2.15) d i.e., P is the exponential of the pure jump part At = s t ∆As. ≤ c c 3) Equation for Zt: Since Zt = exp(At )Pt and exp(A) is continuous, the product rule and (2.15) imply

~~t t c c As As c Zt 1 = e dPs + Ps e dAs − 0 0 − t t c As d c = e Ps d(A + A )s = Zs dAs. − − 0 0~~

~~UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 60 SEMIMARTINGALES~~

~~4) Uniqueness: Suppose that Z is another càdlàg solution of (2.12), and let Xt :=~~

~~Zt Zt. Then X solves the equation − t Xt = Xs dAs t 0 − ∀ ≥ 0 with zero initial condition. Therefore,~~

t Mt n Mt n+1 N Xt Vs dVs Vt t 0, n . (2.16) | | ≤ n! − ≤ (n + 1)! ∀ ≥ ∈ 0 Note that the correction terms in the chain rule are non-negative since V 0 and t ≥ [V ] 0 for all t. As n , the right hand side in (2.16) converges to 0 since M and t ≥ →∞ t V are ﬁnite. Hence X =0 for each t 0. t t ≥ From now on we will denote the unique exponential of (A ) by ( A). t Et Remark (Taylor expansion). By iterating the equation (2.12) for the exponential, we obtain the convergent Taylor series expansion

~~n A (n) t = 1+ dAsk dAsk−1 dAs1 + Rt , E (0,t] (0,s1) (0,sn−1) k=1 where the remainder term can be estimated by~~

~~R(n) M V n+1/(n + 1)!. | t | ≤ t t If A is continuous then the iterated integrals can be evaluated explicitly:~~

~~dA dA dA = (A A )k/k!. sk sk−1 s1 t − 0 (0,t] (0,s1) (0,sk−1) Stochastic Analysis Andreas Eberle 2.1. FINITE VARIATION CALCULUS 61~~

~~If A is increasing but not necessarily continuous then the right hand side still is an upper bound for the iterated integral.~~

~~We now derive a formula for A B where A and B are right-continuous ﬁnite variation Et Et functions. By the product rule and the exponential equation,~~

t t A B A B B A A B t t 1 = s d s + s d s + ∆ s ∆ s E E − E − E E − E E E 0 0 s t ≤ t A B A B = s s d(A + B)s + s s ∆As∆Bs E −E − E −E − 0 s t ≤ t A B = s s d(A + B + [A, B])s E −E − 0 for any t 0. This shows that in general, A B = A+B. ≥ E E E Theorem 2.6. If A, B : [0, ) R are right continuous with ﬁnite variation then ∞ → A B = A+B+[A,B]. E E E Proof. The left hand side solves the deﬁning equation for the exponential on the right hand side.

In particular, choosing B = A, we obtain: − 1 A+[A] = − A E E Example (Geometric Poisson process). A geometric Poisson process with parameters λ> 0 and σ, α R is deﬁned as a solution of a stochastic differential equation of type ∈

dSt = σSt dNt + αSt dt (2.17) − w.r.t.a Poisson process (Nt) with intensity λ. Geometric Poisson processes are relevant for ﬁnancial models, cf.e.g. [35]. The equation (2.17) can be interpreted pathwise as the Stieltjes integral equation

~~t t St = S0 + σ Sr dNr + α Srdr , t 0. − ≥ 0 0 UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 62 SEMIMARTINGALES~~

~~Deﬁning At = σNt + αt, (2.17) can be rewritten as the exponential equation~~

~~dSt = St dAt , − which has the unique solution~~

S = S A = S eαt (1 + σ∆N ) = S eαt(1 + σ)Nt . t 0 Et 0 s 0 s t ≤ Note that for σ > 1, a solution (S ) with positive initial value S is positive for all t, − t 0 whereas in general the solution may also take negative values. If α = λσ then (A ) − t is a martingale. We will show below that this implies that (St) is a local martingale.

~~Indeed, it is a true martingale which for S0 =1 takes the form~~

~~Nt λσt St = (1+ σ) e− .~~

~~Corresponding exponential martingales occur as “likelihood ratio” when the intensity of a Poisson process is modiﬁed, cf.Chapter 4 below.~~

Example (Exponential martingales for compound Poisson processes). For compound Poisson processes, we could proceed as in the last example. To obtain a different point of view, we go in the converse direction: Let

Xt = ηj j=1 be a compound Poisson process on Rd with jump intensity measure ν = λ where λ ∈ (0, ) and is a probability measure on Rd 0 . Hence the η are i.i.d. , and (K ) is ∞ \{ } j ∼ t an independent Poisson process with intensity λ. Suppose that we would like to change the jump intensity measure to an absolutely continuous measure ν¯(dy) = ̺(y)ν(dy) with relative density ̺ 1(ν), and let λ¯ =ν ¯(Rd 0 ). Intuitively, we could expect ∈ L \{ } that the change of the jump intensity is achieved by changing the underlying probability measure P on X with relative density (“likelihood ratio”) Ft

~~Kt (λ λ¯)t (λ λ¯)t Zt = e − ̺(ηj) = e − ̺(∆Xs). j=1 s t ∆X≤s=0 Stochastic Analysis Andreas Eberle 2.1. FINITE VARIATION CALCULUS 63~~

In Chapter 4, as an application of Girsanov’s Theorem, we will prove rigorously that this heuristics is indeed correct. For the moment, we identify (Zt) as an exponential martingale. Indeed, Z = A with t Et A = (λ λ¯)t + (̺(∆X ) 1) t − s − s t ∆X≤s=0 = (λ¯ λ)t + (̺(y) 1) N (dy). (2.18) − − − t Kt Here Nt = j=1 δηj denotes the corresponding Poisson point process with intensity measure ν. Note that (At) is a martingale, since it is a compensated compound Poisson process A = (̺(y) 1) N (dy) , where N := N tν. t − t t t − By the results in the next section, we can then conclude that the exponential (Zt) is a local martingale. We can write down the SDE

~~t Zt = 1+ Zs dAs (2.19) − 0 in the equivalent form~~

t Zt = 1+ Zs (̺(y) 1) N(ds dy) (2.20) (0,t] Rd − − × where N(ds dy) := N(ds dy) ds ν(dy) is the random measure on R+ Rd with − × d N((0, t] B)= Nt(B) for any t 0 and B (R ). In differential notation, (2.20) is × ≥ ∈ B an SDE driven by the compensated Poisson point process (Nt):

dZt = Zt (̺(y) 1) N(dt dy). y Rd − − ∈ Example (Stochastic calculus for finite Markov chains ). Functions of continuous time Markov chains on finite sets are semimartingales with finite variation paths. There- fore, we can apply the tools of finite variation calculus. Our treatment follows Rogers & Williams [34] where more details and applications can be found. Suppose that (X ) on (Ω, ,P ) is a continuous-time, time-homogeneous Markov pro- t A cess with values in a finite set S and càdlàg paths. We denote the transition matrices by

~~UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 64 SEMIMARTINGALES~~

1 pt and the generator (Q-matrix) by = ( (a, b))a,b S. Thus = limt 0 t− (pt I), L L ∈ L ↓ − i.e., for a = b, (a, b) is the jump rate from a to b, and (a, a)= b S,b=a (a, b) is L L − ∈ L the total (negative) intensity for jumping away from a. In particular,

( f)(a) := (a, b)f(b) = (a, b)(f(b) f(a)) L L L − b S b S,b=a ∈ ∈ for any real-valued function f = (f(a))a S on S. It is a standard fact that ((Xt),P ) ∈ solves the martingale problem for , i.e., the process L t M [f] = f(X ) ( f)(X ) ds , t 0, (2.21) t t − L s ≥ 0 is an ( X) martingale for any f : S R. Indeed, this is a direct consequence of the Ft → Markov property and the Kolmogorov forward equation, which imply

t [f] [f] X E[Mt Ms s ] = E[f(Xt) f(Xs) ( f)(Xr) dr s] − |F − − s L |F t = (pt sf)(Xs) f(Xs) (pr s f)(Xs) ds =0 − − − − L s for any 0 s t. In particular, choosing f = I b for b S, we see that ≤ ≤ { } ∈ t b Mt = I b (Xt) (Xs, b) ds (2.22) { } − L 0 is a martingale, and, in differential notation,

~~b dI b (Xt) = (Xt, b) dt + dMt . (2.23) { } L Next, we note that by the results in the next section, the stochastic integrals~~

t a,b b Nt = I a (Xs ) dMs , t 0, { } − ≥ 0 are martingales for any a, b S. Explicitly, for any a = b, ∈ a,b Nt = I a (Xs ) IS b (Xs )I b (Xs) I b (Xs )IS b (Xs) { } − \{ } − { } − { } − \{ } s t ≤ t I a (Xs) (Xs, b) ds , i.e., − { } L 0

~~Stochastic Analysis Andreas Eberle 2.1. FINITE VARIATION CALCULUS 65~~

N a,b = J a,b (a, b) La (2.24) t t −L t a,b where Jt = s t : Xs = a, Xs = b is the number of jumps from a to b until |{ ≤ − }| time t, and t a Lt = Ia(Xs) ds 0 is the amount of time spent at a before time t (“local time at a”). In the form of an SDE, dJ a,b = (a, b) dLa + dN a,b for any a = b. (2.25) t L t t More generally, for any function g : S S R, the process × →

~~[g] a,b Nt = g(a, b)Nt a,b S ∈ is a martingale. If g(a, b)=0 for a = b then by (2.24),~~

t [g] T Nt = g(Xs ,Xs) ( g )(Xs,Xs) ds (2.26) − − L s t 0 ≤ Finally, the exponentials of these martingales are again local martingales. For example, we ﬁnd that αN a,b a,b a = (1+ α)Jt exp( α (a, b)L ) Et − L t is an exponential martingale for any α R and a, b S. These exponential martingales ∈ ∈ appear again as likelihood ratios when changing the jump rates of the Markov chains.

Exercise (Change of measure for ﬁnite Markov chains). Let (X ) on (Ω, ,P, ( )) t A Ft be a continuous time Markov chain with ﬁnite state space S and generator (Q-matrix) , i.e., L t M [f] := f(X ) f(X ) ( f)(X ) ds t t − 0 − L s 0 is a martingale w.r.t. P for each function f : S R. We assume (a, b) > 0 for a = b. → L Let g(a, b) := (a, b)/ (a, b) 1 for a = b, g(a, a) := 0, L L − where is another Q-matrix. L UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 66 SEMIMARTINGALES

1) Let λ(a)= b=a (a, b)= (a, a) and λ(a)= (a, a) denote the total jump L −L −L intensities ata. We deﬁne a “likelihood quotient” for the trajectories of Markov chains with generators and by Z = ζ /ζ where L L t t t t ζt = exp λ(Xs) ds (Xs ,Xs), − 0 L − s t:Xs−=Xs ≤ [g] and ζt is deﬁned correspondingly. Prove that (Zt) is the exponential of (Nt ), and

conclude that (Zt) is a martingale with E[Zt]=1 for any t. 2) Let P denote a probability measure on that is absolutely continuous w.r.t. P on A t with relative density Zt for every t 0. Show that for any f : S R, F ≥ → t M [f] := f(X ) f(X ) ( f)(X ) ds t t − 0 − L s 0

is a martingale w.r.t.P . Hence under the new probability measure P , (Xt) is a Markov chain with generator . L [f] Hint: You may assume without proof that (Mt ) is a local martingale w.r.t. P if [f] and only if (ZtMt ) is a local martingale w.r.t. P . A proof of this fact is given in Section 3.3.

~~2.2 Stochastic integration for semimartingales~~

Throughout this section we fix a probability space (Ω, ,P ) with filtration ( t)t 0. We A F ≥ now define the stochastic integral of the left limit of an adapted càdlàg process w.r.t. a semimartingale in several steps. The key step is the first, where we prove the existence for the integral Hs dMs of a bounded adapted càdlàg process H w.r.t. a bounded − martingale M.

~~Integrals with respect to bounded martingales~~

~~P Suppose that M = (Mt)t 0 is a uniformly bounded càdlàg ( t ) martingale, and H = ≥ F P (Ht)t 0 is a uniformly bounded càdlàg ( t ) adapted process. In particular, the left limit ≥ F process~~

H := (Ht )t 0 − − ≥ Stochastic Analysis Andreas Eberle 2.2. STOCHASTICINTEGRATIONFORSEMIMARTINGALES 67 is left continuous with right limits and ( P ) adapted. For a partition π of R we con- Ft + sider the elementary processes

π ′ π ′ Ht := Hs I[s,s )(t), and Ht = Hs I(s,s ](t). − s π s π ∈ ∈ The process Hπ is again càdlàg and adapted, and the left limit Hπ is left continuous and − (hence) predictable . We consider the Riemann sum approximations

~~π ′ It := Hs(Ms t Ms) ∧ s π − s~~

t to the integral Hs dMs to be defined. Note that if we define the stochastic integral 0 − of an elementary process in the obvious way then t π π It = Hs dMs . − 0 We remark that a straightforward pathwise approach for the existence of the limit of Iπ(ω) as mesh(π) 0 is doomed to fail, if the sample paths are not of finite variation: → Exercise. Let ω Ω and t (0, ), and suppose that (π ) is a sequence of partitions ∈ ∈ ∞ n of R+ with mesh(πn) 0. Prove that if s π hs(Ms′ t(ω) Ms(ω)) converges for → s

The assertion of the exercise is just a restatement of the standard fact that the dual space of ([0, t]) consists of measures with ﬁnite total variation. There are approaches to ex- C tend the pathwise approach by restricting the class of integrands further or by assuming extra information on the relation of the paths of the integrand and the integrator (Young integrals, rough paths theory, cf. [26], [17]). Here, following the standard development of stochastic calculus, we also restrict the class of integrands further (to predictable processes), but at the same time, we give up the pathwise approach. Instead, we consider stochastic modes of convergence.

~~For H and M as above, the process Iπ is again a bounded càdlàg ( P ) martingale as Ft is easily veriﬁed. Therefore, it seems natural to study convergence of the Riemann sum~~

~~UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 68 SEMIMARTINGALES~~

2 2 approximations in the space Md ([0, a]) of equivalence classes of càdlàg L -bounded ( P ) martingales deﬁned up to a ﬁnite time a. The following fundamental theorem Ft settles this question completely:

Theorem 2.7 (Convergence of Riemann sum approximations to stochastic integrals). Let a (0, ) and let M and H be as deﬁned above. Then for every γ > 0 ∈ ∞ there exists a constant ∆ > 0 such that

~~π π 2 I I 2 < γ (2.27) || − e||M ([0,a]) holds for any partitions π and π of R+ with mesh(π) < ∆ and mesh(π) < ∆.~~

The constant ∆ in the theorem depends on M,H and a. The proof of the theorem for discontinuous processes is not easy, but it is worth the effort. For continuous processes, the proof simpliﬁes considerably. The theorem can be avoided if one assumes existence of the quadratic variation of M. However, proving the existence of the quadratic variation requires the same kind of arguments as in the proof below (cf. [15]), or, alternatively, a lengthy discussion of general semimartingale theory (cf. [34]).

Proof of Theorem 2.7. Let C (0, ) be a common uniform upper bound for the pro- ∈ ∞ cesses (Ht) and (Mt). To prove the estimate in (2.27), we assume w.l.o.g.that both partitions π and π contain the end point a, and π is a refinement of π. If this is not the case, we may first consider a common refinement and then estimate by the triangle inequality. Under the additional assumption, we have

π π ′ Ia Iae = (Hs H s )(Ms Ms) (2.28) − − ⌊ ⌋ − s π ∈ where from now on, we only sum over partition points less than a, s′ denotes the suc- cessor of s in the ﬁne partition π, and

s := max t π : t s ⌊ ⌋ { ∈ ≤ } Stochastic Analysis Andreas Eberle 2.2. STOCHASTICINTEGRATIONFORSEMIMARTINGALES 69 is the next partition point of the rough partition π below s. Now ﬁx ε > 0. By (2.28), the martingale property for M, and the adaptedness of H, we obtain π π 2 π π 2 I I 2 = E (I I ) || − e||M ([0,a]) a − ae 2 2 = E (Hs H s ) (Ms′ Ms) (2.29) − ⌊ ⌋ − s π ∈ 2 2 2 2 ε E (M ′ M ) + (2C) E (M ′ M ) ≤ s − s s − s s π t π s π ∈ ∈e τt(ε) ∈s< t ≤ ⌈ ⌉ where t := min u π : u > t is the next partition point of the rough partition, and ⌈ ⌉ { ∈ } τ (ε) := min s π,s>t : H H >ε t . t { ∈ | s − t| } ∧ ⌈ ⌉ is the ﬁrst time after t where H deviates substantially from Hs. Note that τt is a random variable.

~~The summands on the right hand side of (2.29) are now estimated separately. Since M is a bounded martingale, we can easily control the ﬁrst summand:~~

~~2 2 2 2 2 2 E (M ′ M ) = E M ′ M = E M M C . (2.30) s − s s − s a − 0 ≤ The second summand is more difﬁcult to handle. Noting that~~

~~2 2 2 E (M ′ M ) = E M ′ M on τ s , s − s |Fτt s − s |Fτt { t ≤ } we can rewrite the expectation value as~~

~~2 E E (M ′ M ) (2.31) s − s |Fτt t π τt s< t ∈e ≤⌈ ⌉ 2 2 2 = E E M t Mτt τt = E (M t Mτt ) =: B ⌈ ⌉ − |F ⌈ ⌉ − t π t π ∈e ∈e~~

~~Note that M t Mτt =0 only if τt < t , i.e., if H oscillates more than ε in the interval ⌈ ⌉ − ⌈ ⌉ [t, τt]. We can therefore use the càdlàg property of H and M to control (2.31). Let~~

Dε/2 := r [0, a]: Hr Hr > ε/2 { ∈ | − −| } denote the (random) set of “large” jumps of H. Since H is càdlàg, Dε/2 contains only ﬁnitely many elements. Moreover, for given ε, ε¯ > 0 there exists a random variable δ(ω) > 0 such that for u, v [0, a], ∈ UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 70 SEMIMARTINGALES

(i) u v δ H H ε or (u, v] D = , | − | ≤ ⇒ | u − v| ≤ ∩ ε/2 ∅ (ii) r D , u,v [r, r + δ] M M ε.¯ ∈ ε/2 ∈ ⇒ | u − v| ≤ Here we have used that H is càdlàg, Dε/2 is ﬁnite, and M is right continuous. Let ∆ > 0. By (i) and (ii), the following implication holds on ∆ δ : { ≤ }

τt < t Hτt Ht >ε [t, τt] Dε/2 = M t Mτt ε,¯ ⌈ ⌉ ⇒ | − | ⇒ ∩ ∅ ⇒ | ⌈ ⌉ − | ≤ i.e., if τ < t and ∆ δ then the increment of M between τ and t is small. t ⌈ ⌉ ≤ t ⌈ ⌉ Now ﬁx k N and ε>¯ 0. Then we can decompose B = B + B where ∈ 1 2 2 2 B1 = E (M t Mτt ) ; ∆ δ, Dε/2 k kε¯ , (2.32) ⌈ ⌉ − ≤ | | ≤ ≤ t π ∈e 2 B2 = E (M t Mτt ) ; ∆ > δ or Dε/2 >k ⌈ ⌉ − | | t π ∈e 2 2 1/2 1/2 E ( (M t Mτt ) ) P ∆ > δ or Dε/2 >k (2.33) ≤ ⌈ ⌉ − | | t π ∈e √6 C2 P ∆ > δ + P D >k 1/2 . ≤ | ε/2| In the last step we have used the following upper bound for the martingale increments

ηt := M t Mτt : ⌈ ⌉ − 2 2 4 2 2 E ηt = E ηt +2E ηt ηu t π t t u>t ∈e 4C2E η2 +2E η2E η2 ≤ t t u |Ft t t u>t 6C2E η2 6C2E M 2 M 2 6C4. ≤ t ≤ a − 0 ≤ t This estimate holds by the Optional Sampling Theorem, and since E[ η2 ] u>t u |Ft ≤ E[M 2 M 2 ] C2 by the orthogonality of martingale increments M M u − t |Ft ≤ Ti+1 − Ti over disjoint time intervals (Ti, Ti+1] bounded by stopping times. We now summarize what we have shown. By (2.29), (2.30) and (2.31),

π π 2 2 2 2 I I 2 ε C +4C (B + B ) (2.34) || − e||M ([0,a]) ≤ 1 2 where B1 and B2 are estimated in (2.32) and (2.33). Let γ > 0 be given. To bound the right hand side of (2.34) by γ we choose the constants in the following way:

~~Stochastic Analysis Andreas Eberle 2.2. STOCHASTICINTEGRATIONFORSEMIMARTINGALES 71~~

1. Choose ε> 0 such that C2ε2 < γ/4. 2. Choose k N such that 4√6 C4P D >k 1/2 < γ/4, ∈ | ε/2| 2 2 3. Choose ε¯ > 0 such that 4C kε¯ < γ/4, then choose the random variable δ depending on ε and ε¯ such that (i) and (ii) hold. 4. Choose ∆ > 0 such that 4√6 C4P ∆ > δ 1/2 < γ/4. Then for this choice of ∆ we finally obtain π π 2 γ I I 2 < 4 = γ || − e||M ([0,a]) 4 whenever mesh(π) ∆ and π is a refinement of π. ≤ The theorem proves that the stochastic integral H M is well-defined as an M 2 limit of −• the Riemann sum approximations:

Definition (Stochastic integral for left limits of bounded adapted càdlàg processes w.r.t. bounded martingales). For H and M as above, the stochastic integral H M is −• the unique equivalence class of càdlàg ( P ) martingales on [0, ) such that Ft ∞ πn 2 H M = lim H M in Md ([0, a]) −• [0,a] n −• [0,a] →∞ for any a (0, ) and for any sequence (π ) of partitions of R with mesh(π ) 0. ∈ ∞ n + n → Note that the stochastic integral is defined uniquely only up to càdlàg modifications. We will often denote versions of H M by • Hs dMs, but we will not always distinguish −• 0 − between equivalence classes and their representatives carefully. Many basic properties of stochastic integrals with left continuous integrands can be derived directly from the Riemann sum approximations:

~~Lemma 2.8 (Elementary properties of stochastic integrals). For H and M as above, the following statements hold:~~

1) If t Mt has almost surely ﬁnite variation then H M coincides almost surely → −• with the pathwise deﬁned Lebesgue-Stieltjes integral • Hs dMs. 0 − 2) ∆(H M)= H ∆M almost surely. −• − 3) If T : Ω [0, ] is a random variable, and H, H, M, M are processes as → ∞ above such that Ht = Ht for any t < T and Mt = Mt for any t T then, ≤ almost surely, H M = H M on [0, T ]. −• −•

~~UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 72 SEMIMARTINGALES~~

~~Proof. The statements follow easily by Riemann sum approximation. Indeed, let (πn) be a sequence of partitions of R such that mesh(π ) 0. Then almost surely along a + n → subsequence (πn),~~

(H M)t = lim Hs(Ms′ t Ms) −• n ∧ − →∞ s t s≤πn ∈e w.r.t. uniform convergence on compact intervals. This proves that H M coincides −• almost surely with the Stieltjes integral if M has ﬁnite variation. Moreover, for t> 0 it implies

∆(H M)t = lim H t n (Mt Mt ) = Ht ∆Mt (2.35) −• n ⌊ ⌋ − − →∞ − almost surely, where t denotes the next partition point of (π ) below t. Since both ⌊ ⌋n n H M and M are càdlàg, (2.35) holds almost surely simultaneously for all t> 0. The −• third statement can be proven similarly.

~~Localization~~

We now extend the stochastic integral to local martingales. It turns out that unbounded jumps can cause substantial difﬁculties for the localization. Therefore, we restrict ourselves to local martingales that can be localized by martingales with bounded jumps. Remark 2 below shows that this is not a substantial restriction.

Suppose that (Mt)t 0 is a càdlàg ( t) adapted process, where ( t) is an arbitrary filtra- ≥ F F tion. For an ( ) stopping time T , the stopped process M T is defined by Ft T Mt := Mt T for any t 0. ∧ ≥ Definition (Local martingale, Strict local martingale). A localizing sequence for M is a non-decreasing sequence (Tn)n N of ( t) stopping times such that sup TN = , ∈ F ∞ and the stopped process M Tn is an ( ) martingale for each n. The process M is called Ft a local (Ft) martingale iff there exists a localizing sequence. Moreover, M is called a

~~Tn strict local (Ft) martingale iff there exists a localizing sequence (Tn) such that M has uniformly bounded jumps for each n, i.e.,~~

~~sup ∆M (ω) : 0 t T (ω) , ω Ω < n N. {| t | ≤ ≤ n ∈ } ∞ ∀ ∈~~

~~Stochastic Analysis Andreas Eberle 2.2. STOCHASTICINTEGRATIONFORSEMIMARTINGALES 73~~

Remark. 1) Any continuous local martingale is a strict local martingale. 2) In general, any local martingale is the sum of a strict local martingale and a local martingale of ﬁnite variation. This is the content of the “Fundamental Theorem of Local Martingales”, cf. [32]. The proof of this theorem, however, is not trivial and is omitted here. The next example indicates how (local) martingales can be decomposed into strict (local) martingales and ﬁnite variation processes:

Example (Lévy martingales). Suppose that X = y (N (dy) tν(dy)) is a compen- t t − sated Lévy jump process on R1 with intensity measure ν satisfying ( y y 2) ν(dy) < | |∧| | . Then (X ) is a martingale but, in general, not a strict local martingale. However, ∞ t we can easily decompose Xt = Mt + At where At = y I y >1 (Nt(dy) t ν(dy)) {| | } − is a ﬁnite variation process, and Mt = yI y 1 (Nt(dy) tν(dy)) is a strict (local) {| |≤ } − martingale.

~~Strict local martingales can be localized by bounded martingales:~~

~~Lemma 2.9. M is a strict local martingale if and only if there exists a localizing se-~~

~~Tn quence (Tn) such that M is a bounded martingale for each n.~~

Proof. If M Tn is a bounded martingale then also the jumps of M Tn are uniformly bounded. To prove the converse implication, suppose that (Tn) is a localizing sequence such that ∆M Tn is uniformly bounded for each n. Then

~~S := T inf t 0 : M n , n N, n n ∧ { ≥ | t| ≥ } ∈ is a non-decreasing sequence of stopping times with sup S = , and the stopped n ∞ processes M Sn are uniformly bounded, since~~

~~Tn Mt Sn n + ∆MSn = n + ∆MS for any t 0. | ∧ | ≤ | | | n | ≥~~

Deﬁnition (Stochastic integrals of left limits of adapted càdlàg processes w.r.t. strict P local martingales). Suppose that (Mt)t 0 is a strict local ( t ) martingale, and ≥ F UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 74 SEMIMARTINGALES

P (Ht)t 0 is càdlàg and ( t ) adapted. Then the stochastic integral H M is the unique ≥ F −• equivalence class of local ( P ) martingales satisfying Ft H M = H M a.s., (2.36) −• [0,T ] −• [0,T ] whenever T is an ( P ) stopping time, H is a bounded càdlàg ( P ) adapted process Ft Ft P with H [0,T ) = H [0,T ) almost surely, and M is a bounded càdlàg ( t ) martingale with | | F M = M almost surely. [0,T ] [0,T] You should convince yourself that the integral H M is well deﬁned by (2.36) because −• of the local dependence of the stochastic integral w.r.t.bounded martingales on H and

~~M (Lemma 2.8, 3). Note that Ht and Ht only have to agree for t < T , so we may choose Ht = Ht I t~~

~~(n) Tn H M = H M on [0, Tn], −• −• and, by Lemma 2.8, ∆(H(n)M Tn )= H(n)∆M Tn is uniformly bounded for each n. −• −~~

~~Integration w.r.t. semimartingales~~

The stochastic integral w.r.t.a semimartingale can now easily be deﬁned via a semimartingale decomposition. Indeed, suppose that X is an ( P ) semimartingale with Ft decomposition X = X + M + A , t 0, t 0 t t ≥ into a strict local ( P ) martingale M and an ( P ) adapted process A with càdlàg ﬁnite- Ft Ft variation paths t A (ω). → t Stochastic Analysis Andreas Eberle 2.2. STOCHASTICINTEGRATIONFORSEMIMARTINGALES 75

Deﬁnition (Stochastic integrals of left limits of adapted càdlàg processes w.r.t. semi- P martingales ). For any ( t ) adapted process (Ht)t 0 with càdlàg paths, the stochastic F ≥ integral of H w.r.t. X is deﬁned by

H X = H M + H A, −• −• −• where M and A are the strict local martingale part and the finite variation part in a semimartingale decomposition as above, H M is the stochastic integral of H −• − t w.r.t. M, and (H A)t = Hs dAs is the pathwise defined Stieltjes integral of H −• 0 − − w.r.t. A. Note that the semimartingale decomposition of X is not unique. Nevertheless, the integral H X is uniquely defined up to modifications: −• Theorem 2.10. Suppose that (π ) is a sequence of partitionsof R with mesh(π ) 0. n + n → Then for any a [0, ), ∈ ∞

~~(H X)t = lim Hs(Xs′ t Xs) −• n ∧ − →∞ s πn s~~

continuous w.r.t. P then each version of the integral (H X)P deﬁned w.r.t. P is −• a version of the integral (H X)Q deﬁned w.r.t. Q. −• The proofs of this and the next theorem are left as exercises to the reader.

Theorem 2.11 (Elementary properties of stochastic integrals). P 1) Semimartingale decomposition: The integral H X is again an ( t ) semi- −• F martingale with decomposition H X = H M + H A into a strict local mar- −• −• −• tingale and an adapted ﬁnite variation process.

~~UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 76 SEMIMARTINGALES~~

~~2) Linearity: The map (H,X) H X is bilinear. → −• 3) Jumps: ∆(H X)= H ∆X almost surely. −• − 4) Localization: If T is an ( P ) stopping time then Ft~~

~~T T (H X) = H X = (H I[0,T )) X. −• −• −•~~

~~2.3 Quadratic variation and covariation~~

From now on we fix a probability space (Ω, ,P ) with a filtration ( ). The vector A Ft space of (equivalence classes of) strict local ( P ) martingales and of ( P ) adapted Ft Ft processes with càdlàg finite variation paths are denoted by Mloc and FV respectively. Moreover,

= M + FV S loc denotes the vector space of ( P ) semimartingales. If there is no ambiguity, we do not Ft distinguish carefully between equivalence classes of processes and their representatives. The stochastic integral induces a bilinear map , (H,X) H X on the S×S →S → −• equivalence classes that maps M to M and FV to FV. S × loc loc S × A suitable notion of convergence on (equivalence classes of) semimartingales is uniform convergence in probability on compact time intervals:

~~Deﬁnition (ucp-convergence). A sequence of semimartingales X converges to a n ∈ S limit X uniformly on compact intervals in probability iff ∈ S~~

~~n P R sup Xt Xt 0 as n for any a +. t a | − | −→ →∞ ∈ ≤~~

By Theorem (2.10), for H,X and any sequence of partitions with mesh(π ) 0, ∈ S n → the stochastic integral H dX is a ucp-limit of predictable Riemann sum approxima- − tions, i.e., of the integrals of the elementary predictable processes Hπn . − Stochastic Analysis Andreas Eberle 2.3. QUADRATIC VARIATION AND COVARIATION 77

~~Covariation and integration by parts~~

The covariation is a symmetric bilinear map FV. Instead of going once S×S → more through the Riemann sum approximations, we can use what we have shown for stochastic integrals and deﬁne the covariation by the integration by parts identity

~~t t XtYt X0Y0 = Xs dYs + Ys dXs + [X,Y ]t. − − − 0 0 The approximation by sums is then a direct consequence of Theorem 2.10.~~

~~Deﬁnition (Covariation of semimartingales). For X,Y , ∈ S~~

[X,Y ] := XY X0Y0 X dY Y dX. − − − − − Clearly, [X,Y ] is again an ( P ) adapted càdlàg process. Moreover, (X,Y ) [X,Y ] Ft → is symmetric and bilinear, and hence the polarization identity

~~1 [X,Y ] = ([X + Y ] [X] [Y ]) 2 − − holds for any X,Y where ∈ S [X] = [X,X] denotes the quadratic variation of X. The next corollary shows that [X,Y ] deserves the name “covariation”:~~

~~Corollary 2.12. For any sequence (π ) of partitions of R with mesh(π ) 0, n + n →~~

~~[X,Y ]t = ucp lim (Xs′ t Xs)(Ys′ t Ys). (2.37) − n ∧ − ∧ − →∞ s πn s~~

~~UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 78 SEMIMARTINGALES~~

~~Proof. (2.37) is a direct consequence of Theorem 2.10, and 1) follows from (2.37) and the polarization identity. 2) follows from Theorem 2.11, which yields~~

~~∆[X,Y ] = ∆(XY ) ∆(X Y ) ∆(Y X) − −• − −• = X ∆Y + Y ∆X + ∆X∆Y X ∆Y Y ∆X − − − − − − = ∆X∆Y.~~

~~3) follows similarly and is left as an exercise and 4) holds by (2.37) and the Cauchy- Schwarz formula for sums.~~

~~Statements 1) and 2) of the corollary show that [X,Y ] is a ﬁnite variation process with decomposition c [X,Y ]t = [X,Y ]t + ∆Xs∆Ys (2.38) s t ≤ into a continuous part and a pure jump part.~~

~~If Y has ﬁnite variation then by Lemma 2.2,~~

~~[X,Y ]t = ∆Xs∆Ys. s t ≤ Thus [X,Y ]c =0 and if, moreover, X or Y is continuous then [X,Y ]=0.~~

~~More generally, if X and Y are semimartingales with decompositions X = M + A, Y = N + B into M, N M and A, B FV then by bilinearity, ∈ loc ∈~~

~~[X,Y ]c = [M, N]c + [M, B]c + [A, N]c + [A, B]c = [M, N]c.~~

~~It remains to study the covariations of the local martingale parts which turn out to be the key for controlling stochastic integrals effectively.~~

~~Quadratic variation and covariation of local martingales~~

~~If M is a strict local martingale then by the integration by parts identity, M 2 [M] is a − strict local martingale as well. By localization and stopping we can conclude:~~

~~Stochastic Analysis Andreas Eberle 2.3. QUADRATIC VARIATION AND COVARIATION 79~~

Theorem 2.13. Let M and a [0, ). Then M 2([0, a]) if and only if ∈ Mloc ∈ ∞ ∈ Md M 2 and [M] 1. In this case, M 2 [M] (0 t a) is a martingale, and 0 ∈L a ∈L t − t ≤ ≤ 2 2 M 2 = E M + E [M] . (2.39) || ||M ([0,a]) 0 a Proof. We may assume M = 0; otherwise we consider M = M M . Let (T ) be a 0 − 0 n joint localizing sequence for the local martingales M and M 2 [M] such that M Tn is − bounded. Then by optional stopping,

~~2 N E Mt Tn = E [M]t Tn for any t 0 and any n . (2.40) ∧ ∧ ≥ ∈ Since M2 is a submartingale, we have~~

~~2 2 2 E[Mt ] lim inf E[Mt Tn ] E[Mt ] (2.41) n ∧ ≤ →∞ ≤ by Fatou’s lemma. Moreover, by the Monotone Convergence Theorem,~~

~~E [M]t = lim E [M]t Tn . n ∧ →∞ Hence by (2.41), we obtain~~

E[M 2] = E [M] for any t 0. t t ≥ For t a, the right-hand side is dominated from above by E [M] , Therefore, if [M] ≤ a a 2 2 is integrable then M is in Md ([0, a]) with M norm E [M]a . Moreover, in this case, the sequence M 2 [M] is uniformly integrable for each t [0, a], because, t Tn t Tn n N ∧ − ∧ ∈ ∈ 2 2 1 sup Mt [M]t sup Mt + [M]a , t a | − | ≤ t a | | ∈L ≤ ≤ 2 Therefore, the martingale property carries over from the stopped processes Mt Tn ∧ − 2 [M]t Tn to Mt [M]t. ∧ − Remark. The assertion of Theorem 2.13 also remains valid for a = in the sense that ∞ 2 1 if M0 is in and [M] = limt [M]t is in then M extends to a square integrable L ∞ →∞ L martingale (Mt)t [0, ] satisfying (2.40) with a = . The existence of the limit M = ∈ ∞ ∞ ∞ 2 limt Mt follows in this case from the L Martingale Convergence Theorem. →∞ The next corollary shows that the M 2 norms also control the covariations of square integrable martingales.

~~UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 80 SEMIMARTINGALES~~

~~Corollary 2.14. The map (M, N) [M, N] is symmetric, bilinear and continuous on → 2 Md ([0, a]) in the sense that~~

~~E[sup [M, N] t] M M 2([0,a]) N M 2([0,a]). t a | | ≤ || || || || ≤ Proof. By the Cauchy-Schwarz inequality for the covariation (Cor. 2.12,4),~~

~~[M, N] [M]1/2[N]1/2 [M]1/2[N]1/2 t a. | t| ≤ t t ≤ a a ∀ ≤ Applying the Cauchy-Schwarz inequality w.r.t.the L2-inner product yields~~

~~1/2 1/2 E[sup [M, N]t ] E [M]a E [N]a M M 2([0,a]) N M 2([0,a]) t a | | ≤ ≤ || || || || ≤ by Theorem 2.13.~~

Corollary 2.15. Let M and suppose that [M] = 0 almost surely for some ∈ Mloc a a [0, ]. Then almost surely, ∈ ∞ M = M for any t [0, a]. t 0 ∈ In particular, continuous local martingales of ﬁnite variation are almost surely constant.

Proof. By Theorem 2.13, M M 2 = E [M] =0. || − 0||M ([0,a]) a The assertion also extends to the case when a is replaced by a stopping time. Combined with the existence of the quadratic variation, we have now proven:

~~»Non-constant strict local martingales have non-trivial quadratic variation«~~

Example (Fractional Brownian motion is not a semimartingale). Fractional Brow- nian motion with Hurst index H (0, 1) is deﬁned as the unique continuous Gaussian ∈ H process (Bt )t 0 satisfying ≥ 1 E BH = 0 and Cov BH , BH = t2H + s2H t s 2H t s t 2 −| − | for any s, t 0. It has been introduced by Mandelbrot as an example of a self-similar ≥ process and is used in various applications, cf. [1]. Note that for H = 1/2, the covariance is equal to min(s, t), i.e., B1/2 is a standard Brownian motion. In general, one can

Stochastic Analysis Andreas Eberle 2.3. QUADRATIC VARIATION AND COVARIATION 81 prove that fractional Brownian motion exists for any H (0, 1), and the sample paths ∈ t BH (ω) are almost surely α-Hölder continuous if and only if α

V (1)(BH ) = for any t> 0 almost surely , and t ∞ 0 if H > 1/2 , H H H 2  [B ]t = lim Bs′ t Bs = t if H =1/2 , n ∧ −  →∞ s πn  s 1/2 and assume that there is a decomposition Bt = Mt + At into a continuous local martingale M and a continuous ﬁnite variation process A. Then

[M] = [BH ] = 0 almost surely , so by Corollary 2.15, M is almost surely constant, i.e., BH has ﬁnite variation paths. Since this is a contradiction, we see that also for H > 1/2, BH is not a continuous semimartingale, i.e., the sum of a continuous local martingale and a continuous adapted ﬁnite variation process. It is possible (but beyond the scope of these notes) to prove that any semimartingale that is continuous is a continuous semimartingale in the sense above (cf. [32]). Hence for H =1/2, fractional Brownian motion is not a semimartingale and classical Itô calculus is not applicable. Rough paths theory provides an alternative way to develop a calculus w.r.t. the paths of fractional Brownian motion, cf. [17].

~~The covariation [M, N] of local martingales can be characterized in an alternative way that is often useful for determining [M, N] explicitly.~~

Theorem 2.16 (Martingale characterization of covariation). For M, N M , the ∈ loc covariation [M, N] is the unique process A FV such that ∈ (i) MN A M , and − ∈ loc (ii) ∆A = ∆M ∆N , A0 =0 almost surely .

~~Proof. Since [M, N]= MN M0N0 M dN N dM, (i) and (ii) are satisﬁed − − − − − for A = [M, N]. Now suppose that A is another process in FV satisfying (i) and (ii).~~

~~UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 82 SEMIMARTINGALES~~

Then A A is both in M and in FV, and ∆(A A)=0 almost surely. Hence A A − loc − − is a continuous local martingale of ﬁnite variation, and thus A A = A0 A0 = 0 − − almost surely by Corollary 2.15. The covariation of two local martingales M and N yields a semimartingale decomposition for MN: MN = local martingale + [M, N]. However, such a decomposition is not unique. By Corollary 2.15 it is unique if we assume in addition that the ﬁnite variation part A is continuous with A0 = 0 (which is not the case for A = [M, N] in general).

Deﬁnition. Let M, N M . If there exists a continuous process A FV such that ∈ loc ∈ (i) MN A M , and − ∈ loc (ii) ∆A = 0 , A0 = 0 almost surely, then M, N = A is called the conditional covariance process of M and N. In general, a conditional covariance process as deﬁned above need not exist. General martingale theory (Doob-Meyer decomposition) yields the existence under an additional assumption if continuity is replaced by predictability, cf.e.g. [32]. For applications it is more important that in many situations the conditional covariance process can be easily determined explicitly, see the example below.

Corollary 2.17. Let M, N M . ∈ loc 1) If M is continuous then M, N = [M, N] almost surely. 2) In general, if M, N exists then it is unique up to modiﬁcations. 3) If M exists then the assertions of Theorem 2.13 hold true with [M] replaced by M . Proof. 1) If M is continuous then [M, N] is continuous. 2) Uniqueness follows as in the proof of 2.16. 3) If (T ) is a joint localizing sequence for M 2 [M] and M 2 M then, by monotone n − − convergence,

E M t = lim E M t Tn = lim E [M]t Tn = E [M]t n ∧ n ∧ →∞ →∞ for anyt 0. The assertions of Theorem 2.13 now follow similarly as above. ≥ Stochastic Analysis Andreas Eberle 2.3. QUADRATIC VARIATION AND COVARIATION 83

Examples (Covariations of Lévy processes). Rd k 1) Brownian motion: If (Bt) is a Brownian motion in then the components (Bt ) are independent one-dimensional Brownian motions. Therefore, the processes BkBl δ t t t − kl are martingales, and hence almost surely,

~~[Bk, Bl] = Bk, Bl = t δ for any t 0. t t kl ≥ 2) Lévy processes without diffusion part: Let~~

Xt = y Nt(dy) tI y 1 ν(dy) + bt Rd 0 − {| |≤ } \{ } with b Rd, a σ-finite measure ν on Rd 0 satisfying ( y 2 1) ν(dy) < , and a ∈ \{ } | | ∧ ∞ Poisson point process (N ) of intensity ν. Supposefirst that supp(ν) y Rd : y ε t ⊂ ∈ | | ≥ k for some ε> 0. Then the components X are finite variation processes, and hence

k l k l k l [X ,X ]t = ∆Xs ∆Xs = y y Nt(dy). (2.42) s t ≤ In general, (2.42) still holds true. Indeed, if X(ε) is the corresponding Lévy process with (ε) (ε),k k intensity measure ν (dy) = I y ε ν(dy) then X X M 2([0,a]) 0 as ε 0 {| |≥ } || − || → ↓ for any a R and k 1,...,d , and hence by Corollary 2.14, ∈ + ∈{ } k l (ε),k (ε),l k l X ,X = ucp - lim X ,X = ∆Xs ∆Xs. t ε 0 t ↓ s t ≤ On the other hand, we know that if X is square integrable then Mt = Xt it ψ(0) and 2 − ∇ M kM l t ∂ ψ (0) are martingales, and hence t t − ∂pk∂pl 2 k l k l ∂ ψ X ,X t = M , M t = t (0). ∂pk∂pl 3) Covariations of Brownian motion and Lévy jump processes: For B and X as above we have

Bk,Xl = [Bk,Xl] = 0 almost surely for any k and l. (2.43) Indeed, (2.43) holds true if Xl has finite variation paths. The general case then follows once more by approximating Xl by finite variation processes. Note that independence of B and X has not been assumed! We will see in Section 3.1 that (2.43) implies that a Brownian motion and a Lévy process without diffusion term defined on the same probability space are always independent.

~~UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 84 SEMIMARTINGALES~~

~~Covariation of stochastic integrals~~

We now compute the covariation of stochastic integrals. This is not only crucial for many computations, but it also yields an alternative characterization of stochastic integrals w.r.t. local martingales, cf. Corollary 2.19 below.

~~Theorem 2.18. Suppose that X and Y are ( P ) semimartingales, and H is ( P ) Ft Ft adapted and càdlàg. Then~~

H dX,Y = H d[X,Y ] almost surely. (2.44) − − Proof. 1. We first note that (2.44) holds if X or Y has finite variation paths. If, for example, X FV then also H dX FV, and hence ∈ − ∈ H dX,Y = ∆(H X)∆Y = H ∆X∆Y = H d[X,Y ] . − −• − − The same holds if Y FV. ∈ 2. Now we show that (2.44) holds if X and Y are bounded martingales, and H is bounded. For this purpose, we fix a partition π, and we approximate H by the elemen- − π ′ tary process H = s π Hs I(s,s ]. Let − ∈ π π ′ It = H dX = Hs(Xs t Xs). − ∧ − (0,t] s π ∈ We can easily verify that

[Iπ,Y ] = Hπ d[X,Y ] almost surely. (2.45) − Indeed, if (π ) is a sequence of partitions such that π π for any n and mesh(π ) 0 n ⊂ n n → then π π ′ ′ ′ (Ir′ t Ir )(Yr t Yr) = Hs (Xr t Xr)(Yr t Yr). ∧ − ∧ − ∧ − ∧ − r πn s π r πn r

~~Having veriﬁed (2.45) for any ﬁxed partition π, we choose again a sequence (πn) of partitions with mesh(π ) 0. Then n →~~

~~H dX = lim Iπn in M 2([0, a]) for any a (0, ), − n ∈ ∞ →∞ and hence, by Corollary 2.14 and (2.45),~~

H dX,Y = ucp - lim [Iπn ,Y ] = H d[X,Y ]. − n − →∞ 3. Now suppose that X and Y are strict local martingales. If T is a stopping time such T T that X and Y are bounded martingales, and HI[0,T ) is bounded as well, then by Step 2, Theorem 2.11 and Corollary 2.12,

T T T T T H dX,Y = H dX ,Y = (H I[0,T )) dX ,Y − − − T T T = H I[0,T ) d[X ,Y ] = H d[X,Y ] . − − Since this holds for all localizing stopping times as above, (2.45) is satisﬁed as well. 4. Finally, suppose that X and Y are arbitrary semimartingales. Then X = M + A and Y = N + B with M, N M and A, B FV. The assertion (2.44) now follows by ∈ loc ∈ Steps 1 and 3 and by the bilinearity of stochastic integral and covariation.

~~Perhaps the most remarkable consequences of Theorem 2.18 is:~~

~~Corollary 2.19 (Kunita-Watanabe characterization of stochastic integrals). P Let M Mloc and G = H with H ( t ) adapted and càdlàg. Then G M is the unique ∈ − F • element in Mloc satisfying~~

(i) (G M)0 = 0 , and • (ii) [G M, N] = G [M, N] for any N Mloc. • • ∈ Proof. By Theorem 2.18, G M satisﬁes (i) and (ii). It remains to prove uniqueness. Let • L M such that L =0 and ∈ loc 0

~~[L, N] = G [M, N] for any N Mloc. • ∈ UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 86 SEMIMARTINGALES~~

~~Then [L G M, N]=0 for any N Mloc. Choosing N = L G M, we conclude − • ∈ − • that [L G M]=0. Hence L G M is almost surely constant, i.e., − • − •~~

~~L G M L0 (G M)0 = 0. − • ≡ − •~~

~~Remark. Localization shows that it is sufﬁcient to verify Condition (ii) in the Kunita- Watanabe characterization for bounded martingales N.~~

The corollary tells us that in order to identify stochastic integrals w.r.t. local martingales it is enough to “test” with other (local) martingales via the covariation. This fact can be used to give an alternative deﬁnition of stochastic integrals that applies to general pre- P dictable integrands. Recall that a stochastic process (Gt)t 0 is called ( t ) predictable ≥ F iff the function (ω, t) G (ω) is measurable w.r.t. the σ-algebra P on Ω [0, ) → t × ∞ generated by all ( P ) adapted left-continuous processes. Ft Deﬁnition (Stochastic integrals with general predictable integrands). Let M M , and suppose that G is an ( P ) predictable process satisfying ∈ loc Ft t G2 d[M] < almost surely for any t 0. s s ∞ ≥ 0

~~If there exists a local martingale G M Mloc such that conditions (i) and (ii) in Corol- • ∈ lary 2.19 hold, then G M is called the stochastic integral of G w.r.t. M. •~~

~~Many properties of stochastic integrals can be deduced directly from this deﬁnition, see e.g. Theorem 2.21 below.~~

~~The Itô isometry for stochastic integrals w.r.t. martingales~~

Of course, Theorem 2.18 can also be used to compute the covariation of two stochastic integrals. In particular, if M is a semimartingale and G = H with H càdlàg and − adapted then [G M,G M] = G [M,G M] = G2 [M]. • • • • •

~~Stochastic Analysis Andreas Eberle 2.4.ITÔCALCULUSFORSEMIMARTINGALES 87~~

Corollary 2.20 (Itô isometry for martingales). Suppose that M M . Then also ∈ loc G dM 2 G2 d[M] M , and − ∈ loc 2 a 2 a G dM = E G dM = E G2 d[M] a 0, a.s . 2 M ([0,a]) 0 0 ∀ ≥ 2 Proof. If M Mloc then G M Mloc, and hence (G M) [G M] Mloc. Moreover, ∈ • ∈ • − • ∈ by Theorem 2.13,

~~2 2 G M M 2([0,a]) = E [G M]a = E (G [M])a . || • || • •~~

The Itô isometry for martingales states that the M 2([0, a]) norm of the stochastic integral G dM coincides with the L2 Ω (0, a],P norm of the integrand (ω, t) G (ω), × [M] → t where P is the measure on Ω R given by [M] × +

~~P[M](dω dt) = P (dω)[M](ω)(dt).~~

~~This can be used to prove the existence of the stochastic integral for general predictable integrands G L2(P ), cf. Section 2.5 below. ∈ [M]~~

~~2.4 Itô calculus for semimartingales~~

We are now ready to prove the two most important rules of Itô calculus for semimartingales: The so-called “Associative Law” which tells us how to integrate w.r.t. processes that are stochastic integrals themselves, and the change of variables formula.

~~Integration w.r.t. stochastic integrals~~

Suppose that X and Y are semimartingales satisfying dY = G dX for some predictable integrand G, i.e., Y Y0 = G dX. We would like to show that we are allowed to − multiply the differential equation formally by another predictable process G, i.e., we would like to prove that G dY = GG dX: dY = G dX = G dY = GG dX ⇒ UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 88 SEMIMARTINGALES

~~The covariation characterization of stochastic integrals w.r.t. local martingales can be used to prove this rule in a simple way.~~

Theorem 2.21 (“Associative Law”). Let X . Then ∈ S G (G X) = (GG) X (2.46) • • • holds for any processes G = H and G = H with H and H càdlàg and adapted. − − Remark. The assertion extends with a similar proof to more general predictable integrands.

Proof. We already know that (2.46) holds for X FV. Therefore, and by bilinearity of ∈ the stochastic integral, we may assume X M . By the Kunita-Watanabe characteri- ∈ loc zation it then sufﬁces to “test” the identity (2.46) with local martingales. For N M , ∈ loc Corollary 2.19 and the associative law for FV processes imply

~~[G (G X), N] = G [G X, N] = G (G [X, N]) • • • • • • = (GG) [X, N] = [(GG) X, N]. • • Thus (2.46) holds by Corollary 2.19.~~

~~Itô’s formula~~

We are now going to prove a change of variables formula for discontinuous semimartingales. To get an idea how the formula looks like we first briefly consider a semimartingale X with a finite number of jumps in finite time. Suppose that ∈ S 0 < T1 < T2 < . . . are the jump times, and let T0 = 0. Then on each of the intervals

~~[Tk 1, Tk), X is continuous. Therefore, by a similar argument as in the proof of Itô’s − formula for continuous paths (cf. [13, Thm.6.4]), we could expect that~~

~~F (Xt) F (X0) = F (Xt T ) F (Xt T − ) − ∧ k − ∧ k 1 k t Tk t Tk ∧ − 1 ∧ − = F ′(Xs ) dXs + F ′′(Xs ) d[X]s + F (XT ) F (XT ) − 2 − k − k− Tk−1~~

c where Xt = Xt s t ∆Xs denotes the continuous part of X. However, this formula − ≤ does not carry over to the case when the jumps accumulate and the paths are not of ﬁnite variation, since then the series may diverge and the continuous part Xc does not exist in general. This problem can be overcome by rewriting (2.47) in the equivalent form

~~t t 1 c F (Xt) F (X0) = F ′(Xs ) dXs + F ′′(Xs ) d[X]s (2.48) − − 2 − 0 0 + F (Xs) F (Xs ) F ′(Xs ) ∆Xs , − − − − s t ≤ which carries over to general semimartingales.~~

1 d Theorem 2.22 (Itô’s formula for semimartingales). Suppose that Xt =(Xt ,...,Xt ) with semimartingales X1,...,Xd . Then for every function F C2(Rd), ∈ S ∈ d d 2 ∂F i 1 ∂ F i j c F (Xt) F (X0)= (Xs ) dXs + (Xs ) d[X ,X ]s − ∂xi − 2 ∂xi∂xj − i=1 i,j=1 (0,t] (0,t] d ∂F i + F (Xs) F (Xs ) (Xs )∆Xs (2.49) − − − ∂xi − s (0,t] i=1 ∈ for any t 0, almost surely. ≥ i Remark. The existence of the quadratic variations [X ]t implies the almost sure absolute convergence of the series over s (0, t] on the right hand side of (2.49). Indeed, a ∈ Taylor expansion up to order two shows that

d ∂F i i 2 F (Xs) F (Xs ) (Xs )∆Xs Ct ∆Xs | − − − ∂xi − | ≤ | | s t i=1 s t i ≤ ≤ C [Xi] < , ≤ t t ∞ i where Ct = Ct(ω) is an almost surely ﬁnite random constant depending only on the maximum of the norm of the second derivativeof F on the convex hull of X : s [0, t] . { s ∈ } It is possible to prove this general version of Itô’s formula by a Riemann sum approximation, cf. [32]. Here, following [34], we instead derive the “chain rule” once more from the “product rule”:

~~UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 90 SEMIMARTINGALES~~

~~Proof. To keep the argument transparent, we restrict ourselves to the case d = 1. The generalization to higher dimensions is straightforward. We now proceed in three steps:~~

~~1. As in the ﬁnite variation case (Theorem 2.4), we ﬁrst prove that the set consisting A of all functions F C2(R) satisfying (2.48) is an algebra, i.e., ∈~~

~~F,G = FG . ∈A ⇒ ∈A~~

~~This is a consequence of the integration by parts formula~~

~~t t F (Xt)G(Xt) F (X0)G(X0)= F (X ) dG(X)+ G(X ) dF (X) − 0 − 0 − c + F (X),G(X) + ∆F (X)∆G(X), (2.50) (0,t] the associative law, which implies~~

1 c F (X ) dG(X) = F (X )G′(X ) dX + F (X )G′′(X ) d[X] − − − 2 − − + F (X ) (∆G(X) G′(X )∆X), (2.51) − − − the corresponding identity with F and G interchanged, and the formula

c c [F (X),G(X)] = F ′(X ) dX, G′(X ) dX (2.52) − − c c = F ′(X )G′(X ) d[X] = (F ′G′)(X ) d[X] − − − for the continuous part of the covariation. Both (2.51) and (2.52) follow from (2.49) and the corresponding identity for G. It is straightforward to verify that (2.50), (2.51) and (2.52) imply the change of variable formula (2.48) for FG, i.e., FG . Therefore, ∈ A by induction, the formula (2.48) holds for all polynomials F .

2. In the second step, we prove the formula for arbitrary F C2 assuming X = M +A ∈ with a bounded martingale M and a bounded process A FV. In this case, X is ∈ uniformly bounded by a ﬁnite constant C. Therefore, there exists a sequence (pn) of

~~Stochastic Analysis Andreas Eberle 2.4.ITÔCALCULUSFORSEMIMARTINGALES 91~~

~~polynomials such that p F , p′ F ′ and p′′ F ′′ uniformly on [ C,C]. For n → n → n → − t 0, we obtain ≥~~

F (Xt) F (X0) = lim pn(Xt) pn(X0) n − →∞ − t t Xs y 1 c = lim pn′ (Xs ) dXs + pn′′(Xs ) d[X]s + pn′′ (z) dz dy n − 2 − →∞ 0 0 s t Xs− Xs− ≤ t t Xs y 1 c = F ′(Xs ) dXs + F ′′(Xs ) d[X]s + F ′′(z) dz dy − 2 − 0 0 s t Xs− Xs− ≤ w.r.t. convergence in probability. Here we have used an expression of the jump terms in (2.48) by a Taylor expansion. The convergence in probability holds since X = M + A,

~~t t 2 E pn′ (Xs ) dMs F ′(Xs ) dMs 0 − − 0 − t 2 2 = E (pn′ F ′)(Xs ) d[M]s sup pn′ F ′ E [M]t 0 − − ≤ [ C,C] | − | − by Itô’s isometry, and~~

Xs y 1 2 (p′′ F ′′)(z) dz dy sup p′′ F ′′ (∆X ) . n − ≤ 2 | n − | s s t Xs− Xs− [ C,C] s t ≤ − ≤ 3. Finally, the change of variables formula for general semimartingales X = M + A with M M and A FV follows by localization. We can ﬁnd an increasing se- ∈ loc ∈ quence of stopping times (T ) such that sup T = a.s., M Tn is a bounded martingale, n n ∞ Tn and the process A − deﬁned by

~~At for t < Tn Tn A − := t  ATn for t Tn  − ≥~~

~~n Tn Tn is a bounded process in FV for any n. Itô’s formula then holds for X := M + A − for every n. Since Xn = X on [0, T ) and T a.s., this implies Itô’s formula for n n ր ∞ X.~~

Note that the second term on the right hand side of Itô’s formula (2.49) is a continuous ﬁnite variation process and the third term is a pure jump ﬁnite variation process. More- over, semimartingale decompositions of Xi, 1 i d, yield corresponding decomposi- ≤ ≤ tions of the stochastic integrals on the right hand side of (2.49). Therefore, Itô’s formula

~~UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 92 SEMIMARTINGALES~~

~~1 d can be applied to derive an explicit semimartingale decomposition of F (Xt ,...,Xt ) for any C2 function F . This will now be carried out in concrete examples.~~

~~Application to Lévy processes~~

~~We ﬁrst apply Itô’s formula to a one-dimensional Lévy process~~

Xt = x + σBt + bt + y Nt(dy) (2.53) with x, σ, b R, a Brownian motion (B ), and a compensated Poisson point process ∈ t N = N tν with intensity measure ν. We assume that ( y 2 y ) ν(dy) < . The t t − | | ∧| | ∞ only restriction to the general case is the assumed integrab ility of y at , which en- | | ∞ sures in particular that (Xt) is integrable. The process (Xt) is a semimartingale w.r.t. the ﬁltration ( B,N ) generated by the Brownian motion and the Poisson point process. Ft We now apply Itô’s formula to F (X ) where F C2(R). Setting C = y N (dy) we t ∈ t t ﬁrst note that almost surely, 2 2 2 [X]t = σ [B]t +2σ[B,C]t + [C]t = σ t + (∆Xs) . s t ≤ Therefore, by (2.54),

F (X ) F (X ) t − 0 t t 1 c = F ′(X ) dX + F ′′(X ) d[X] + F (X) F (X ) F ′(X )∆X − 2 − − − − − 0 0 s t ≤ t t 1 2 = (σF ′)(Xs ) dBs + (bF ′ + σ F ′′)(Xs) ds + F ′(Xs ) y N(ds dy) 0 − 0 2 − (0,t] R × + F (Xs + y) F (Xs ) F ′(Xs )y N(ds dy), (2.54) − − − − − (0,t] R × where N(ds dy) is the Poisson random measure on R R corresponding to the Pois- + × son point process, and N(ds dy)= N(ds dy) ds ν(dy). Here, we have used a rule for − evaluating a stochastic integral w.r.t.the process Ct = y Nt(dy) which is intuitively clear and can be veriﬁed by approximating the integrand by elementary processes. Note Stochastic Analysis Andreas Eberle 2.4.ITÔCALCULUSFORSEMIMARTINGALES 93

also that in the second integral on the right hand side we could replace Xs by Xs since − almost surely, ∆Xs =0 for almost all s. To obtain a semimartingale decomposition from (2.54), we note that the stochastic integrals w.r.t. (Bt) and w.r.t. (Nt) are local martingales. By splitting the last integral on the right hand side of (2.54) into an integral w.r.t. N(ds dy) (i.e., a local martingale) and an integral w.r.t.the compensator ds ν(dy), we have proven: Corollary 2.23 (Martingale problem for Lévy processes). For any F C2(R), the ∈ process t M [F ] = F (X ) F (X ) ( F )(X ) ds, t t − 0 − L s 0 1 ( F )(x) = (σF ′′)(x)+(bF ′)(x)+ F (x + y) F (x) F ′(x)y ν(dy), L 2 − − is a local martingale vanishing at 0. For F 2(R), M [F ] is a martingale, and ∈ Cb 1 ( F )(x) = lim E F (Xt) F (X0) . t 0 L ↓ t − [F ] Proof. M is a local martingale by the considerations above and since Xs(ω) = 2 Xs (ω) for almost all (s,ω). For F b , F is bounded since F (x + y) F (x) − ∈ C L − − 2 [F ] F ′(x)y = ( y y ). Hence M is a martingale in this case, and O | |∧| | 1 1 t E F (Xt) F (X0) = E ( F )(Xs) ds ( F )(x) t − t 0 L → L as t 0 by right continuity of( F )(X ). ↓ L s The corollary shows that is the inﬁnitesimal generator of the Lévy process. The L martingale problem can be used to extend results on the connection between Brownian motion and the Laplace operator to general Lévy processes and their generators. For example, exit distributions are related to boundary value problems (or rather complement value problems as is not a local operator), there is a potential theory for generators of L Lévy processes, the Feynman-Kac formula and its applications carry over, and so on.

Example (Fractional powers of the Laplacian). By Fourier transformation one veri- ﬁes that the generator of a symmetric α-stable process with characteristic exponent p α | | is = ( ∆)α/2. The behaviour of symmetric α-stable processes is therefore closely L − − linked to the potential theory of these well-studied pseudo-differential operators.

~~UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 94 SEMIMARTINGALES~~

~~Exercise (Exit distributions for compound Poisson processes). Let (Xt)t 0 be a com- ≥ pound Poisson process with X0 =0 and jump intensity measure ν = N(m, 1), m> 0.~~

~~i) Determine λ R such that exp(λX ) is a local martingale. ∈ t ii) Prove that for a< 0,~~

~~P [Ta < ] = lim P [Ta < Tb] exp(ma/2). b ∞ →∞ ≤~~

~~Why is it not as easy as for Brownian motion to compute P [Ta < Tb] exactly?~~

~~Applications to Itô diffusions~~

~~Next we consider a solution of a stochastic differential equation~~

dXt = b(t, Xt) dt + σ(t, Xt) dBt, X0 = x0, (2.55) defined on a filtered probability space (Ω, ,P, ( )). We assume that (B ) is an ( ) A Ft t Ft Brownian motion taking values in Rd, b, σ ,...,σ : R+ Rn Rn are continuous 1 d × → time-dependent vector fields in Rn, and σ(t, x)=(σ (t, x) σ (t, x)) is the n d 1 d × matrix with column vectors σ (t, x). A solution of (2.55) is a continuous ( P ) semi- i Ft martingale (Xt) satisfying

~~t d t k Xt = x0 + b(s,Xs) ds + σk(s,Xs) dBs t 0 a.s. (2.56) 0 0 ∀ ≥ k=1 If X is a solution then~~

~~i j i k j l [X ,X ]t = σk(s,X) dB , σl (s,X) dB t k,l t t i j k l ij = (σk σl )(s,X) d[B , B ] = a (s,Xs) ds 0 0 k,l ij i j where a = k σkσk, i.e.,~~

~~T n n a(s, x)= σ(s, x)σ(s, x) R × . ∈ Stochastic Analysis Andreas Eberle 2.4.ITÔCALCULUSFORSEMIMARTINGALES 95~~

~~Therefore, Itô’s formula applied to the process (t, Xt) yields~~

∂F 1 d ∂2F dF (t, X) = (t, X) dt + F (t, X) dX + (t, X) d[Xi,Xj] ∂t ∇x 2 ∂xi∂xj i,j=1 ∂F = (σT F )(t, X) dB + + F (t, X) dt, ∇x ∂t L for any F C2(R Rn), where ∈ + × 1 d ∂2F d ∂F ( F )(t, x) = aij(t, x) (t, x)+ bi(t, x) (t, x). L 2 ∂xi∂xj ∂xi i,j=1 i=1 We have thus derived the Itô-Doeblin formula

~~t t ∂F F (t, X ) F (0,X ) = (σT F )(s,X ) dB + + F (s,X ) ds t − 0 ∇ s s ∂t L s 0 0 (2.57)~~

Again, the formula provides a semimartingale decomposition for F (t, Xt). It establishes a connection between the stochastic differential equation (2.55) and partial differential equations involving the operator . L Example (Exit distributions and boundary value problems). Suppose that F ∈ C2(R Rn) is a classical solution of the p.d.e. + × ∂F (t, x)+( F )(t, x) = g(t, x) t 0, x U ∂t L − ∀ ≥ ∈ on an open subset U Rn with boundary values ⊂ F (t, x) = ϕ(t, x) t 0, x ∂U. ∀ ≥ ∈ Then by (2.57), the process

t Mt = F (t, Xt)+ g(s,Xs) ds 0 is a local martingale. If F and g are bounded on [0, t] U, then the process M T stopped × at the ﬁrst exit time T = inf t 0 : X / U is a martingale. Hence, if T is almost { ≥ t ∈ } surely ﬁnite then

~~T E[ϕ(T,XT )] + E g(s,Xs) ds = F (0, x0). 0 UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 96 SEMIMARTINGALES~~

This can be used, for example, to compute exit distributions (for g 0) and mean exit ≡ times (for ϕ 0, g 1) analytically or numerically. ≡ ≡ Similarly as in the example, the Feynman-Kac-formula and other connections between Brownian motion and the Laplace operator carry over to Itô diffusions and their generator in a straightforward way. Of course, the resulting partial differential equation L usually can not be solved analytically, but there is a wide range of well-established numerical methods for linear PDE available for explicit computations of expectation values.

Exercise (Feynman-Kac formula for Itô diffusions). Fix t (0, ), and suppose ∈ ∞ that ϕ : Rn R and V : [0, t] Rn [0, ) are continuous functions. Show that if → × → ∞ u C2((0, t] Rn) C([0, t] Rn) is a bounded solution of the heat equation ∈ × ∩ × ∂u (s, x) = ( u)(s, x) V (s, x)u(s, x) for s (0, t], x Rn, ∂s L − ∈ ∈ u(0, x) = ϕ(x), then u has the stochastic representation

t u(t, x) = E ϕ(X ) exp V (t s,X ) ds . x t − − s 0 Hint: Consider the time reversal uˆ(s, x) := u(t s, x) of u on [0, t]. Show first that − M := exp( A )û(r, X ) is a local martingale if A := r Vˆ (s,X ) ds. r − r r r 0 s Often, the solution of an SDE is only defined up to some explosion time ζ where it diverges or exits a given domain. By localization, we can apply the results above in this case as well. Indeed, suppose that U Rn is an open set, and let ⊆ U = x U : x 1/k , k N. k { ∈ | | } ∈

Then U = Uk. Let Tk denote the ﬁrst exit time of (Xt) from Uk. A solution (Xt) of the SDE (2.55) up to the explosion time ζ = sup Tk is a process (Xt)t [0,ζ) 0 such that ∈ ∪{ } for every k N, T <ζ almost surely on ζ (0, ) , and the stopped process XTk is ∈ k { ∈ ∞ } a semimartingale satisfying (2.56) for t T . By applying Itô’s formula to the stopped ≤ k processes, we obtain:

~~Stochastic Analysis Andreas Eberle 2.4.ITÔCALCULUSFORSEMIMARTINGALES 97~~

Corollary 2.24 (Martingale problem for Itô diffusions). If X : Ω U is a solution t → of (2.55) up to the explosion time ζ, then for any F C2(R U) and x U, the ∈ + × 0 ∈ process t ∂F Mt := F (t, Xt) + F (s,Xs) ds, t < ζ, − 0 ∂t L is a local martingale up to the explosion time ζ, and the stopped processes M Tk , k N, ∈ are localizing martingales.

Proof. We can choose functions F 2([0, a] U), k N, a 0, such that F (t, x)= k ∈ Cb × ∈ ≥ k F (t, x) for t [0, a] and x in a neighbourhood of U . Then for t a, ∈ k ≤ t Tk ∂Fk Mt = Mt Tk = Fk(t, Xt Tk ) + Fk (s,Xs Tk ) ds. ∧ ∧ − 0 ∂t L ∧ By (2.57), the right hand side is a bounded martingale.

~~Exponentials of semimartingales~~

~~If X is a continuous semimartingale then by Itô’s formula, 1 X = exp X [X] Et t − 2 t is the unique solution of the exponential equation~~

d X = X dX, X = 1. E E E0 In particular, X is a local martingale if X is a local martingale. Moreover, if E n ∂ 2 hn(t, x) = exp(αx α t/2) (2.58) ∂αn − α=0 denotes the Hermite polynomial of order n and X =0 then 0 n Ht = hn [X]t,Xt (2.59) solves the SDE n n 1 n dH = nH − dX, H0 =0, for any n N, cf. Section 6.4 in [13]. In particular, Hn is an iterated Itô integral: ∈ t sn s2 Hn = n! dX dX dX . t s1 s2 sn 0 0 0 The formula for the stochastic exponential can be generalized to the discontinuous case:

~~UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 98 SEMIMARTINGALES~~

~~Theorem 2.25 (Doléans-Dade). Let X . Then the unique solution of the exponen- ∈ S tial equation t Zt = 1+ Zs dXs, t 0, (2.60) − ≥ 0 is given by~~

~~1 Z = exp X [X]c (1+∆X ) exp( ∆X ). (2.61) t t − 2 t s − s s (0,t] ∈ Remarks. 1) In the ﬁnite variation case, (2.61) can be written as~~

1 Z = exp Xc [X]c (1+∆X ). t t − 2 t s s (0,t] ∈ In general, however, neither Xc nor (1+∆X) exist. 2) The analogues to the stochastic polynomials Hn in the discontinuous case do not have an equally simply expression as in (2.59) . This is not too surprising: Also for continuous two-dimensional semimartingales (Xt,Yt) there is no direct expression for the iterated integral t s dX dY = t(X X ) dY and for the Lévy area process 0 0 r s 0 s − 0 s t s t s A = dX dY dY dX t r s − r s 0 0 0 0 in terms of X,Y and their covariations. If X is a one-dimensional discontinuous semimartingale then X and X are different processes that have both to be taken into account − when computing iterated integrals of X.

~~Proof of Theorem 2.25. The proof is partially similar to the one given above for X ∈ FV, cf.Theorem 2.5. The key observation is that the product~~

~~P = (1+∆X ) exp( ∆X ) t s − s s (0,t] ∈ exists and deﬁnes a ﬁnite variation pure jump process. This follows from the estimate~~

~~log(1 + ∆X ) ∆X const. ∆X 2 const. [X] | s − s| ≤ | s| ≤ t 0~~

S = (log(1 + ∆X ) ∆X ), t 0, t s − s ≥ s t ∆Xs≤ 1/2 | |≤ defines almost surely a finite variation pure jump process. Therefore, (Pt) is also a finite variation pure jump process. Moreover, the process G = exp X 1 [X]c satisfies t t − 2 t G = 1+ G dX + (∆G G ∆X) (2.62) − − − by Itô’s formula. For Z = GP we obtain

~~∆X ∆X ∆Z = Z e (1+∆X)e− 1 = Z ∆X, − − − and hence, by integration by parts and (2.62),~~

Z 1 = P dG + G dP + [G,P ] − − − = P G dX + (P ∆G P G ∆X + G ∆P + ∆G ∆P ) − − − − − − − = Z dX + (∆Z Z ∆X) = Z dX. − − − − This proves that Z solves the SDE (2.60). Uniqueness of the solution follows from a general uniqueness result for SDE with Lipschitz continuous coefﬁcients, cf. Section 3.1.

Example (Geometric Lévy processes). Consider a Lévy martingale Xt = y Nt(dy) R where (Nt) is a Poisson point process on with intensity measure ν satisfying ( y | | ∧ y 2) ν(dy) < , and N = N tν. We derive an SDE for the semimartingale | | ∞ t t − Z = exp(σX + t), t 0, t t ≥ where σ and are real constants. Since [X]c 0, Itô’s formula yields ≡ σ∆X Zt 1 = σ Z dX + Z ds + Z e 1 σ∆X (2.63) − − − − − − (0,t] (0,t] (0,t] σy = σ Zs y N(ds dy)+ Zs ds + Zs e 1 σy N(ds dy). − − (0,t] R − − − (0,t] R (0,t] × × UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 100 SEMIMARTINGALES

~~If e2σy ν(dy) < then (2.63) leads to the semimartingale decomposition ∞~~

~~σ dZt = Zt dMt + αZt dt, Z0 =1, (2.64) − − where M σ = eσy 1 N (dy) t − t is a square-integrable martingale, and~~

~~α = + (eσy 1 σy) ν(dy). − −~~

~~In particular, we see that although (Zt) again solves an SDE driven by the compensated process (Nt), this SDE can not be written as an SDE driven by the Lévy process (Xt).~~

~~2.5 General predictable integrands and local times~~

So far, we have considered stochastic integrals w.r.t. general semimartingales only for integrands that are left limits of adapted càdlàg processes. This is indeed sufficient for many applications. For some results including in particular convergence theorems for stochastic integrals, martingale representation theorems and the existence of local time, stochastic integrals with more general integrands are important. In this section, we sketch the definition of stochastic integrals w.r.t. not necessarily continuous semimartingales for general predictable integrands. For details of the proofs, we refer to Chapter IV in Protter’s book [32]. At the end of the section, we apply the results to define the local time process (occupation time density) of a continuous semimartingale.

Throughout this section, we ﬁx a ﬁltered probability space (Ω, ,P, ( )). Recall that A Ft the predictable σ-algebra on Ω (0, ) is generated by all sets A (s, t] with P × ∞ × A and 0 s t, or, equivalently, by all left-continuous ( ) adapted processes ∈ Fs ≤ ≤ Ft (ω, t) G (ω). We denote by the vector space consisting of all elementary pre- → t E dictable processes G of the form

n 1 − Gt(ω) = Zi(ω)I(ti,ti+1](t) i=0 Stochastic Analysis Andreas Eberle 2.5. GENERALPREDICTABLEINTEGRANDSANDLOCALTIMES 101 with n N, 0 t < t < < t , and Z : Ω R bounded and -measurable. ∈ ≤ 0 1 n i → Fti For G and a semimartingale X , the stochastic integral G X defined by ∈E ∈ S • n 1 t − (G X)t = Gs dXs = Zi Xti t Xti t • +1∧ − ∧ 0 i=0 is again a semimartingale. Clearly, if A is a finite variation process then G A has finite • variation as well.

2 Now suppose that M Md (0, ) is a square-integrable martingale. Then G M ∈ ∞ • ∈ M 2(0, ), and the Itô isometry d ∞ 2 2 ∞ G M M 2(0, ) = E G dM || • || ∞ 0 ∞ 2 2 = E G d[M] = G dP[M] (2.65) 0 Ω R+ × holds, where

P[M](dω dt) = P (dω)[M](ω)(dt) is the Doléans measure of the martingale M on Ω R . The Itô isometry has been × + derived in a more general form in Corollary 2.20, but for elementary processes it can easily be veriﬁed directly (Excercise!). In many textbooks, the angle bracket process M is used instead of [M] to deﬁne stochastic integrals. The next remark shows that this is equivalent for predictable integrands: Remark ([M] vs. M). Let M M 2(0, ). If the angle-bracket process M exists ∈ d ∞ then the measures P[M] and P M coincide on predictable sets. Indeed, if C = A (s, t] × with A and 0 s t then ∈Fs ≤ ≤

~~P (C) = E [[M] [M] ; A] = E [E[[M] [M] ]; A] [M] t − s t − s|Fs = E [E[ M t M s s]; A] = P M (C). − |F~~

Since the collection of these sets C is an -stable generator for the predictable σ- ∩ algebra, the measures P[M] and P M coincide on . P UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 102 SEMIMARTINGALES

~~Example (Doléans measures of Lévy martingales). If M = X E[X ] with a square t t − t integrable Lévy process X : Ω R then t →~~

P[M] = P M = ψ′′(0) P λ(0, ) ⊗ ∞ where ψ is the characteristic exponent of X and λ(0, ) denotes Lebesgue measure on ∞ R+. Hence the Doléans measure of a general Lévy martingale coincides with the one for Brownian motion up to a multiplicative constant.

~~Deﬁnition of stochastic integrals w.r.t. semimartingales~~

We denote by 2 the vector space of all semimartingales vanishing at 0 of the form H X = M + A with M M 2(0, ) and A FV predictable with total variation ∈ d ∞ ∈ (1) 2 2 V (A) = ∞ dAs L (P ). In order to deﬁne a norm on the space , we as- ∞ 0 | | ∈ H sume without proof the following result, cf. e.g. Chapter III in Protter [32]: Fact. Any predictable local martingale with ﬁnite variation paths is almost surely constant.

~~The result implies that the Doob-Meyer semimartingale decomposition~~

X = M + A (2.66) is unique if we assume that M is local martingale and A is a predictable ﬁnite variation process vanishing at 0. Therefore, we obtain a well-deﬁned norm on 2 by setting H 2 2 2 (1) 2 ∞ X 2 = M M 2 + V (A) L2 = E [M] + dA . || ||H || || || ∞ || ∞ | | 0 Note that the M 2 norm is the restriction of the 2 norm to the subspace M 2(0, ) H ∞ ⊂ 2. As a consequence of (2.65), we obtain: H Corollary 2.26 (Itô isometry for semimartingales). Let X 2 with semimartingale ∈ H decomposition as above. Then

G X 2 = G X for any G , where || • ||H || || ∈E 2 2 2 ∞ G := G 2 + G dA . X L (P[M]) 2 || || || || 0 | || | L (P ) Stochastic Analysis Andreas Eberle 2.5. GENERALPREDICTABLEINTEGRANDSANDLOCALTIMES 103

2 Hence the stochastic integral : , X (G) = G X, has a unique isometric J E → H J • extension to the closure X of w.r.t. the norm in the space of all predictable E E || ||X 2 processes in L (P[M]).

Proof. The semimartingale decomposition X = M + A implies a corresponding decomposition G X = G M + G A for the stochastic integrals. One can verify that • • • 2 for G , G M is in Md (0, ) and G A is a predictable ﬁnite variation process. ∈ E • ∞ • Therefore, and by (2.65),

2 2 2 (1) 2 2 G X 2 = G M M 2 + V (G A) L2 = G L2(P ) + G dA . || • ||H || • || || ∞ • || || || [M] | || | L2(P ) X The Itô isometry yields a definition of the stochastic integral G X for G . For • ∈ E G = H with H càdlàg and adapted, this definition is consistent with the definition − given above since, by Corollary 2.20, the Itô isometry also holds for the integrals defined X above, and the isometric extension is unique. The class of admissible integrands is E already quite large:

X Lemma 2.27. contains all predictable processes G with G < . E || ||X ∞ Proof. We only mention the main steps of the proof, cf. [32] for details: 1) The approximation of bounded left-continuous processes by elementary predictable processes w.r.t. is straightforward by dominated convergence. ||||X 2) The approximabilityof bounded predictable processes by bounded left-continuous processes w.r.t. can be shown via the Monotone Class Theorem. ||||X n 3) For unbounded predictable G with G X < , the processes G := G I G n , || || ∞ { ≤ } n N, are predictable and bounded with Gn G 0. ∈ || − ||X →

~~Localization~~

2 Having deﬁned G X for X and predictable integrands G with G X < , the • ∈ H || || ∞ next step is again a localization. This localization is slightly different than before, because there might be unbounded jumps at the localizing stopping times. To overcome

UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 104 SEMIMARTINGALES this difﬁculty, the process is stopped just before the stopping time T , i.e., at T . How- − ever, stopping at T destroys the martingale property if T is not a predictable stopping − time. Therefore, it is essential that we localize semimartingales instead of martingales!

~~T For a semimartingale X and a stopping time T we deﬁne the stopped process X − by~~

Xt for t < T, T X − = X for t T > 0, t  T  − ≥ 0 for T =0.  The deﬁnition for T =0 is of course rather arbitrary. It will not be relevant below, since we are considering sequences (T ) of stopping times with T almost surely. We n n ↑ ∞ state the following result from Chapter IV in [32] without proof.

~~Fact. If X is a semimartingale with X0 = 0 then there exists an increasing sequence~~

Tn 2 (T ) of stopping times with sup T = such that X − for any n N. n n ∞ ∈ H ∈ Now we are ready to state the deﬁnition of stochastic integrals for general predictable integrands w.r.t. general semimartingales X. By setting G X = G (X X0) we may • • − assume X0 =0.

~~Deﬁnition. Let X be a semimartingale with X0 =0. A predictable process G is called integrable w.r.t. X iff there exists an increasing sequence (Tn) of stopping times such~~

Tn 2 that sup T = a.s., and for any n N, X − and G Tn− < . n ∞ ∈ ∈ H || ||X ∞ If G is integrable w.r.t. X then the stochastic integral G X is deﬁned by • t t Tn N (G X)t = Gs dXs = Gs dXs − for any t [0, Tn), n . • ∈ ∈ 0 0 Of course, one has to verify that G X is well-deﬁned. This requires in particular a • locality property for the stochastic integrals that are used in the localization. We do not carry out the details here, but refer once more to Chapter IV in [32].

Exercise (Sufﬁcient conditions for integrability of predictable processes). 1) Prove that if G is predictable and locally bounded in the sense that GTn is bounded for a sequence (T ) of stopping times with T , then G is integrable w.r.t. any n n ↑ ∞ semimartingale X . ∈ S Stochastic Analysis Andreas Eberle 2.5. GENERALPREDICTABLEINTEGRANDSANDLOCALTIMES 105

2) Suppose that X = M + A is a continuous semimartingale with M loc and ∈ Mc A FV . Prove that G is integrable w.r.t. X if G is predictable and ∈ c t t G2 d[M] + G dA < a.s. for any t 0. s s | s|| s| ∞ ≥ 0 0 Properties of the stochastic integral

Most of the properties of stochastic integrals can be extended easily to general predictable integrands by approximation with elementary processes and localization. The proof of Property (2) below, however, is not trivial. We refer to Chapter IV in [32] for detailed proofs of the following basic properties:

~~(1) The map (G,X) G X is bilinear. → • (2) ∆(G X)= G∆X almost surely. •~~

~~T T (3) (G X) =(GI[0,T ]) X = G X . • • •~~

T T (4) (G X) − = G X −. • • (5) G (G X)=(GG) X. • • • In all statements, X is a semimartingale, G is a process that is integrable w.r.t. X, T is a stopping time, and G is a process such that GG is also integrable w.r.t. X. We state the formula for the covariation of stochastic integrals separately below, because its proof is based on the Kunita-Watanabe inequality, which is of independent interest.

Exercise (Kunita-Watanabe inequality). Let X,Y , and let G,H be measurable ∈ S processes deﬁned on Ω (0, ) (predictability is not required). Prove that for any × ∞ a [0, ] and p, q [1, ] with 1 + 1 =1, the following inequalities hold: ∈ ∞ ∈ ∞ p q a a 1/2 a 1/2 G H d[X,Y ] G2 d[X] H2 d[Y ] , (2.67) 0 | || || | ≤ 0 0 a a 1/2 a 1/2 E G H d[X,Y ] G2 d[X] H2 d[Y ] . p q 0 | || || | ≤ 0 L 0 L (2.68) Hint: First consider elementary processes G,H.

~~UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 106 SEMIMARTINGALES~~

~~Theorem 2.28 (Covariation of stochastic integrals). For any X,Y and any pre- ∈ S dictable process G that is integrable w.r.t. X,~~

G dX,Y = Gd[X,Y ] almost surely. (2.69) Remark. If X and Y are local martingales, and the angle-bracket process X,Y exists, then also G dX,Y = Gd X,Y almost surely. Proof of Theorem 2.28. We only sketch the main steps brieﬂy, cf. [32] for details. Firstly, one veriﬁes directly that (2.69) holds for X,Y 2 and G . Secondly, for ∈ H ∈ E X,Y 2 and a predictable process G with G < there exists a sequence ∈ H || ||X ∞ (Gn) of elementary predictable processes such that Gn G 0, and || − ||X →

~~Gn dX,Y = Gn d[X,Y ] for any n N. ∈ As n , Gn dX G dX in 2 by the Itô isometry for semimartingales, and → ∞ → H hence Gn dX,Y G dX,Y u.c.p. −→ by Corollary 2.14. Moreover,~~

Gn d[X,Y ] Gd[X,Y ] u.c.p. −→ by the Kunita-Watanabe inequality. Hence (2.69) holds for G as well. Finally, by localization, the identity can be extended to general semimartingales X,Y and integrands G that are integrable w.r.t. X.

~~An important motivation for the extension of stochastic integrals to general predictable integrands is the validity of a Dominated Convergence Theorem:~~

Theorem 2.29 (Dominated Convergence Theorem for stochastic integrals). Suppose that X is a semimartingale with decomposition X = M + A as above, and let Gn, n N, and G be predictable processes. If ∈ Gn(ω) G (ω) for any t 0, almost surely, t −→ t ≥ Stochastic Analysis Andreas Eberle 2.5. GENERALPREDICTABLEINTEGRANDSANDLOCALTIMES 107 and if there exists a process H that is integrable w.r.t. X such that Gn H for any | | ≤ n N, then ∈ Gn X G X u.c.p. as n . • −→ • →∞ If, in addition to the assumptions above, X is in 2 and H < then even H || ||X ∞ n G X G X 2 0 as n . || • − • ||H −→ →∞ Proof. We may assume G = 0, otherwise we consider Gn G instead of Gn. Now − suppose ﬁrst that X is in 2 and H < . Then H || ||X ∞ 2 n 2 ∞ n 2 ∞ n G X = E G d[M]+ G dA 0 || || 0 | | 0 | || | −→ as n by the Dominated Convergence Theorem for Lebesgue integrals. Hence by → ∞ the Itô isometry, Gn X 0 in 2 as n . • −→ H →∞ The general case can now be reduced to this case by localization, where 2 convergence H is replaced by the weaker ucp-convergence.

We ﬁnally remark that basic properties of stochastic integrals carry over to integrals with respect to compensated Poisson point processes. We refer to the monographs by D.Applebaum [3] for basics, and to Jacod & Shiryaev [22] for a detailed study. We only state the following extension of the associative law, which has already been used in the last section:

Exercise (Integration w.r.t. stochastic integrals based on compensated PPP). Sup- pose that H : Ω R S R is predictable and square-integrable w.r.t. P λ ν, × + × → ⊗ ⊗ and G : Ω R R is a bounded predictable process. Show that if × + →

~~Xt = Hs(y) N(ds dy) (0,t] S × then t Gs dXs = GsHs(y) N(ds dy). 0 (0,t] S × Hint: Approximate G by elementary processes.~~

~~UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 108 SEMIMARTINGALES~~

~~Local time~~

The occupation time of a Borel set U R by a one-dimesional Brownian motion (B ) ⊆ t is given by t U Lt = IU (Bs) ds. 0 Brownian local time is an occupation time density for Brownian motion that is informally given by t a “ Lt = δa(Bs) ds ” 0 for any a R. It is a non-decreasing stochastic process satisfying ∈

U a Lt = Lt da. U We will now apply stochastic integration theory for general predictable integrands to a R deﬁne the local time process (Lt )t 0 for a rigorously for Brownian motion, and, ≥ ∈ more generally, for a continuous semimartingale (Xt). Note that by Itô’s formula,

t 1 t f(X ) f(X ) = f ′(X ) dX + f ′′(X ) d[X] . t − 0 s s 2 s s 0 0 Informally, if X is a Brownian motion then the last integral on the right hand side a should coincide with Lt if f ′′ = δa. A convex function with second derivative δa is + f(x)=(x a) . Noting that the left derivative of f is given by f ′ = I(a, ), this − − ∞ motivates the following deﬁnition:

~~Deﬁnition. For a continuous semimartingale X and a R, the process La deﬁned by ∈ t + + 1 a (Xt a) (X0 a) = I(a, )(Xs) dXs + Lt − − − ∞ 2 0 is called the local time of X at a.~~

Remark. 1) By approximating the indicator function by continuous functions it can be easily veriﬁed that the process I(a, )(Xs) is predictable and integrable w.r.t. X. ∞ 2) Alternatively, we could have deﬁned local time at a by the identity

~~t + + 1 ˆa (Xt a) (X0 a) = I[a, )(Xs) dXs + Lt − − − ∞ 2 0 Stochastic Analysis Andreas Eberle 2.5. GENERALPREDICTABLEINTEGRANDSANDLOCALTIMES 109~~

involving the right derivative I[a, ) instead of the left derivative I(a, ). Note that ∞ ∞ t a â Lt Lt = I a (Xs) dXs. − { } 0 This difference vanishes almost surely if X is a Brownian motion, or, more generally, a continuous local martingale. For semimartingales, however, the processes La and Lâ may disagree, cf. the example below Lemma 2.30. The choice of La in the definition of local time is then just a standard convention that is consistent with the convention of considering left derivatives of convex functions.

Lemma 2.30 (Properties of local time, Tanaka formulae). 1) Suppose that ϕ : R [0, ), n N, is a sequence of continuous functions with n → ∞ ∈ ϕ =1 and ϕ (x)=0 for x (a, a +1/n). Then n n ∈ t a Lt = ucp lim ϕn(Xs) d[X]s. − n →∞ 0 a In particular, the process (Lt )t 0 is non-decreasing and continuous. ≥ 2) The process La grows only when X = a, i.e.,

~~t a I Xs=a dLs = 0 for any t 0. { } ≥ 0 3) The following identities hold:~~

t + + 1 a (Xt a) (X0 a) = I(a, )(Xs) dXs + Lt , (2.70) − − − 0 ∞ 2 t 1 a (Xt a)− (X0 a)− = I( ,a](Xs) dXs + Lt , (2.71) − − − − 0 −∞ 2 t X a X a = sgn (X a) dX + La, (2.72) | t − |−| 0 − | s − s t 0 where sgn(x) := +1 for x> 0, and sgn(x) := 1 for x 0. − ≤ Remark. Note that we set sgn(0) := 1. This is related to our convention of using left − derivatives as sgn(x) is the left derivative of x . There are analogue Tanaka formulae | | for Lˆa with the intervals (a, ) and ( , a] replaced by [a, ) and ( , a), and the ∞ −∞ ∞ −∞ sign function deﬁned by sgnˆ (x) := +1 for x 0 and sgn(x) := 1 for x< 0. ≥ − UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 110 SEMIMARTINGALES

x y 2 Proof. 1) For n N let fn(x) := ϕn(z) dz dy. Then the function fn is C ∈ −∞ −∞ with fn′′ = ϕn. By Itô’s formula, t 1 t f (X ) f (X ) f ′ (X ) dX = ϕ (X ) d[X] . (2.73) n t − n 0 − n s s 2 n s s 0 0

~~As n , fn′ (Xs) converges pointwise to I(a, )(Xs). Hence →∞ ∞ t t~~

~~fn′ (Xs) dXs I(a, )(Xs) dXs → ∞ 0 0 in the ucp-sense by the Dominated Convergence Theorem 2.29. Moreover,~~

f (X ) f (X ) (X a)+ (X a)+. n t − n 0 → t − − 0 − The first assertion now follows from (2.73). R a 2) By 1), the measures ϕn(Xt) d[X]t on + converge weakly to the measure dLt with a distribution function L . Hence by the Portemanteau Theorem, and since ϕn(x)=0 for x (a, a +1/n), ∈ t t a I Xs a >ε dLs lim inf I Xs a >ε ϕn(Xs) d[X]s = 0 {| − | } ≤ n {| − | } 0 →∞ 0 for any ε > 0. The second assertion of the lemma now follows by the Monotone Convergence Theorem as ε 0. ↓ 3) The first Tanaka formula (2.70) holds by definition of La. Moreover, subtracting (2.71) from (2.70) yields

~~t (X a) (X a) = dX , t − − 0 − s 0 which is a valid equation. Therefore, the formulae (2.71) and (2.70) are equivalent. Finally, (2.72) follows by adding (2.70) and (2.71).~~

~~Remark. In the proof above it is essential that the Dirac sequence (ϕn) approximates~~

δa from the right. If X is a continuous martingale then the assertion 1) of the lemma also holds under the assumption that ϕn vanishes on the complement of the interval (a 1/n, a+1/n). For semimartingales however, approximating δ from the left would − a lead to an approximation of the process Lˆa, which in general may differ from La.

~~Stochastic Analysis Andreas Eberle 2.5. GENERALPREDICTABLEINTEGRANDSANDLOCALTIMES 111~~

Exercise (Brownian local time). Show that the local time of a Brownian motion B in a R is given by ∈ t a 1 Lt = ucp lim I(a ε,a+ε)(Bs) ds. − ε 0 2ε − → 0 Example (Reﬂected Brownian motion). Suppose that X = B where (B ) is a one- t | t| t dimensional Brownian motion starting at 0. By Tanaka’s formula (2.72), X is a semimartingale with decomposition

~~Xt = Wt + Lt (2.74)~~

t where Lt is the local time at 0 of the Brownian motion B and Wt := 0 sgn(Bs) dBs. By Lévy’s characterization, the martingale W is also a Brownian motion, cf. Theorem X 4.2. We now compute the local time Lt of X at 0. By (2.71) and Lemma 2.30, 2),

t 1 X Lt = Xt− X0− + I( ,0](Xs) dXs (2.75) 2 − 0 −∞ t t t = I 0 (Bs) dWs + I 0 (Bs) dLs = dLs = Lt a.s., { } { } 0 0 0 X t i.e., Lt =2Lt. Here we have used that I 0 (Bs) dWs vanishes almost surely by Itô’s 0 { } isometry, as both W and B are Brownian motions. Notice that on the other hand, t 1 ˆX Lt = Xt− X0− + I( ,0)(Xs) dXs = 0 a.s., 2 − −∞ 0 so the processes LX and LˆX do not coincide. By (2.74) and (2.75), the process X solves the singular SDE 1 dX = dW + dLX t t 2 t driven by the Brownian motion W . This justiﬁes thinking of X as Brownian motion reﬂected at 0.

~~The identity (2.74) can be used to compute the law of Brownian local time:~~

~~Exercise (The law of Brownian local time).~~

a) Prove Skorohod’s Lemma: If (yt)t 0 is a real-valued continuous function with ≥ y =0 then there exists a unique pair (x, k) of functions on [0, ) such that 0 ∞ UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 112 SEMIMARTINGALES

~~(i) x = y + k, (ii) x is non-negative, and~~

~~(iii) k is non-decreasing, continuous, vanishing at zero, and the measure dkt is carried by the set t : x =0 . { t }~~

~~The function k is given by kt = sups t( ys). ≤ −~~

~~b) Conclude that the local time process (Lt) at 0 of a one-dimensional Brownian~~

motion (Bt) starting at 0 and the maximum process St := sups t Bs have the ≤ same law. In particular, L B for any t 0. t ∼| t| ≥ c) More generally, show that the two-dimensional processes ( B , L) and (S B,S) | | − have the same law.

Notice that the maximum process (St)t 0 is the generalized inverse of the Lévy subor- ≥ dinator (Ta)a 0 introduced in Section 1.1. Thus we have identiﬁed Brownian local time ≥ at 0 as the inverse of a Lévy subordinator.

~~Itô-Tanaka formula~~

Local time can be used to extend Itô’s formula in dimension one from C2 to general convex functions. Recall that a function f : R R is convex iff → f(λx + (1 λ)y) λf(x)+(1 λ)f(y) λ [0, 1], x,y R. − ≤ − ∀ ∈ ∈ For a convex function f, the left derivatives f(x) f(x h) f ′ (x) = lim − − − h 0 h ↓ exist, the function f ′ is left-continuous and non-decreasing, and − b f(b) f(a) = f ′ (x) dx for any a, b R. − − ∈ a The second derivative of f in the distributional sense is the positive measure f ′′ given by

~~f ′′([a, b)) = f ′ (b) f ′ (a) for any a, b R. − − − ∈ Stochastic Analysis Andreas Eberle 2.5. GENERALPREDICTABLEINTEGRANDSANDLOCALTIMES 113~~

We will prove in Theorem 3.10 below that there is a version (t, a) La of the local → t time process of a continuous semimartingale X such that t La is continuous and → t a La is càdlàg. If X is a local martingale then La is even jointly continuous in t and → t t a. From now on, we ﬁx a corresponding version.

Theorem 2.31 (Itô-Tanaka formula, Meyer). Suppose that X is a continuous semimartingale, and f : R R is convex. Then → t 1 a f(Xt) f(X0) = f ′ (Xs) dXs + Lt f ′′(da). (2.76) − − 2 R 0 Proof. We proceed in several steps: 1) Equation (2.76) holds for linear functions f. 2) By localization, we may assume that X < C for a ﬁnite constant C. Then both | t| sides of (2.76) depend only on the values of f on ( C,C), so we may also assume − w.l.o.g. that f is linear on each of the intervals ( , C] and [C, ), i.e., −∞ − ∞

supp(f ′′) [ C,C]. ⊆ − Moreover, by subtracting a linear function and multiplying f by a constant, we may even assume that f vanishes on ( ,C], and f ′′ is a probability measure. Then −∞ x f ′ (y)= ( ,y) and f(x)= ( ,y) dy (2.77) − −∞ −∞ −∞ where := f ′′. + 3) Now suppose that = δa is a Dirac measure. Then f ′ = I(a, ) and f(x)=(x a) . − ∞ − Hence Equation (2.76) holds by definition of La. More generally, by linearity, (2.76) holds whenever has finite support, since then is a convex combination of Dirac measures. 4) Finally, if is a general probability measure then we approximate by measures with finite support. Suppose that Z is a random variable with distribution , and let n n n denote the law of Zn := 2− 2 Z . By 3), the Itô-Tanaka formula holds for the x ⌈ ⌉ functions fn(x) := n( ,y) dy, i.e., −∞ −∞ t 1 a fn(Xt) fn(X0) = fn′ (Xs) dXs + Lt n(da) (2.78) − − 2 R 0 UniversityofBonn WinterSemester2012/2013 CHAPTER 2. STOCHASTIC INTEGRALS AND ITÔ CALCULUS FOR 114 SEMIMARTINGALES for any n N. As n , ( ,X ) ( ,X ), and hence ∈ →∞ n −∞ s → −∞ s t t

fn′ (Xs) dXs f ′ (Xs) dXs − → − 0 0 in the ucp sense by dominated convergence. Similarly, f (X ) f (X ) f(X ) n t − n 0 → t − f(X ). Finally, the right continuity of a La implies that 0 → t

a a Lt n(da) Lt (da), R → R since Zn converges to Z from above. The Itô-Tanaka formula (2.76) for f now follows from (2.78) as n . →∞ Clearly, the Itô-Tanaka formula also holds for functions f that are the difference of two convex functions. If f is C2 then by comparing the Itô-Tanaka formula and Itô’s a formula, we can identify the integral Lt f ′′(da) over a as the stochastic time integral t 0 f ′′(Xs) d[X]s. The same remains true whenever the measure f ′′(da) is absolutely continuous with density denoted by f ′′(a):

Corollary 2.32. For any measurable function V : R [0, ), → ∞ t a Lt V (da) = V (Xs) d[X]s t 0. (2.79) R ∀ ≥ 0 Proof. The assertion assertion holds for any continuous function V : R [0, ) as → ∞ V can be represented as the second derivative of a C2 function f. The extension to measurable non-negative functions now follows by a monotone class argument.

~~Notice that for V = IB, the expression in (2.79) is the occupation time of the set B by~~

~~(Xt), measured w.r.t. the quadratic variation d[X]t.~~

~~Stochastic Analysis Andreas Eberle Chapter 3~~

~~SDE I: Strong solutions and approximation schemes~~

In this chapter we study strong solutions of stochastic differential equations and the corresponding stochastic ﬂows. We start with a crucial martingale inequality that is used frequently to derive Lp estimates for semimartingales. For real-valued càdlàg functions x =(xt)t 0 we set ≥

~~⋆ ⋆ xt := sup xs for t> 0, and x0 := x0 . s~~

~~115 116CHAPTER 3. SDE I: STRONG SOLUTIONS AND APPROXIMATION SCHEMES~~

~~Here, we only prove an easy special case of the Burkholder-Davis-Gundy inequality that will be sufﬁcient for our purposes. This estimate also holds for càdlàg local martingales:~~

~~Theorem 3.1 (Burkholder’s inequality). Let p [2, ). Then the estimate ∈ ∞~~

E[(M ⋆ )p]1/p γ E[[M]p/2]1/p (3.2) T ≤ p T holds for any strict local martingale M such that M =0, and for any stopping ∈Mloc 0 time T : Ω [0, ], where → ∞ (p 1)/2 1 − √ γp = 1+ p 1 p/ 2 e/2 p. − ≤ Remark. The estimate does not depend on the underlying ﬁltered probability space, the local martingale M, and the stopping time T . However, the constant γ goes to p ∞ as p . →∞

Proof. 1) We ﬁrst assume that T = and M is a bounded càdlàg martingale. Then, ∞ by the Martingale Convergence Theorem, M = lim Mt exists almost surely. Since ∞ t p 2 →∞ p 2 the function f(x) = x is C for p 2 with ϕ′′(x) = p(p 1) x − , Itô’s formula | | ≥ − | | implies

p ∞ 1 ∞ c M = ϕ′(Ms ) dMs + ϕ′′(Ms) d[M]s | ∞| − 2 0 0 + (ϕ(Ms) ϕ(Ms ) ϕ′(Ms )∆Ms, ) , (3.3) − − − − s where the ﬁrst term is a martingale since ϕ′ M is bounded, in the second term ◦

~~⋆ p 2 ϕ′′(Ms) p(p 1)(M ) − , ≤ − ∞ and the summand in the third term can be estimated by~~

~~1 2 ϕ(Ms) ϕ(Ms ) ϕ′(Ms )∆Ms sup(ϕ′′ M)(∆Ms) − − − − ≤ 2 ◦ 1 ⋆ p 2 2 p(p 1)(M ) − (∆Ms) . ≤ 2 − ∞~~

~~Stochastic Analysis Andreas Eberle 117~~

~~Hence by taking expectation values on both sides of (3.3), we obtain for q satisfying 1 1 p + q =1:~~

E[(M ⋆ )p] qp E[ M p] ∞ ≤ | ∞| p p(p 1) ⋆ p 2 c 2 q − E (M ) − [M] + (∆M) ≤ 2 ∞ ∞ p(p 1) p−2 p 2 qp − E[(M ⋆ )p] p E[[M] 2 ] p ≤ 2 ∞ ∞ by Doob’s inequality, Hölder’s inequality, and since [M]c + (∆M)2 = [M] . The ∞ ∞ p p 1 2 inequality (3.2) now follows by noting that q p(p 1) = q − p . − 2) For T = and a strict local martingale M M , there exists an increasing ∞ ∈ loc Tn sequence (Tn) of stopping times such that M is a bounded martingale for each n. Applying Burkholder’s inequality to M Tn yields

⋆ p Tn,⋆ p p Tn p/2 p p/2 E[(MTn ) ] = E[(M ) ] γp E[[M ] ] = γp E[[M]Tn ]. ∞ ≤ ∞ Burkholder’s inequality for M now follows as n . →∞ 3) Finally, the inequality for an arbitrary stopping time T can be derived from that for T = by considering the stopped process M T . ∞ For p 4, the converse estimate in (3.1) can be derived in a similar way: ≥ Exercise. Prove that for a given p [4, ), there exists a global constant c (1, ) ∈ ∞ p ∈ ∞ such that the inequalities

~~1 p/2 p p/2 cp− E [M] E [(M ∗ ) ] cp E [M] ∞ ≤ ∞ ≤ ∞ with Mt∗ = sups~~

~~Exercise. Let M be a continuous local martingale satisfying M0 =0. Show that x2 P sup Ms x ;[M]t c exp s t ≥ ≤ ≤ − 2c ≤ for any c, t, x [0, ). ∈ ∞~~

~~UniversityofBonn WinterSemester2012/2013 118CHAPTER 3. SDE I: STRONG SOLUTIONS AND APPROXIMATION SCHEMES~~

~~3.1 Existence and uniqueness of strong solutions~~

Let (S, , ν) be a σ-ﬁnite measure space, and let d, n N. Suppose that on a proba- S ∈ bility space (Ω, ,P ), we are given an Rd-valued Brownian motion (B ) and a Poisson A t random measure N(dt dy) over R+ S with intensity measure λ(0, ) ν. Let ( t) × ∞ ⊗ F denote a complete ﬁltration such that (B ) is an ( ) Brownian motion and N (B) = t Ft t N((0, t] B) is an ( ) Poisson point process, and let × Ft

N(dt dy) = N(dt dy) λ(0, )(dt) ν(dy). − ∞ If T is an ( ) stopping time then we call a predictable process (ω, t) G (ω) or Ft → t (ω,t,y) G (y)(ω) defined for finite t T (ω) and y S locally square integrable → t ≤ ∈ iff there exists an increasing sequence (T ) of ( ) stopping times with T = sup T n Ft n 2 such that for any n, the trivially extended process GtI t Tn is contained in (P { ≤ } L ⊗ λ), 2(P λ ν) respectively. For locally square integrable predictable integrands, L ⊗ ⊗ t the stochastic integrals 0 Gs dBs and (0,t] S Gs(y) N(ds dy) respectively are local × martingales defined for t [0, T ). ∈ In this section, we are going to study existence and pathwise uniqueness for solutions of stochastic differential equations of type

dXt = bt(X) dt + σt(X) dBt + ct (X,y) N(dt dy). (3.4) y S − ∈ n n n n d Here b : R (R , R ) R , σ : R (R , R ) R × , and c : R + × D + → + ×D + → + × (R , Rn) S Rn are càdlàg functions in the ﬁrst variable such that b , σ and c D + × → t t t are measurable w.r.t. the σ-algebras := σ(x x : s t), respectively for Bt → s ≤ Bt ⊗ S any t 0. We also assume local boundedness of the coefﬁcients, i.e., ≥

~~sup sup sup ( bs(x) + σs(x) + cs(x, y) ) < (3.5) s t. Corresponding~~

~~Stochastic Analysis Andreas Eberle 3.1. EXISTENCEANDUNIQUENESSOFSTRONGSOLUTIONS 119~~

statements hold for σt and ct. Condition (3.5) implies in particular that the jump sizes are locally bounded. Locally unbounded jumps could be taken into account by extending the SDE (3.4) by an additional term consisting of an integral w.r.t. an uncompensated Poisson point process.

~~Deﬁnition. Suppose that T is an ( ) stopping time. Ft 1) A solution of the stochastic differential equation (3.4) for t < T is a càdlàg ( ) Ft n adapted stochastic process (Xt)t~~

~~t t Xt = X0 + bs(X) ds + σs(X) dBs + cs (X,y) N(ds dy) (3.6) 0 0 (0,t] S − × holds for any t < T .~~

~~2) A solution (Xt)t~~

For a strong solution, Xt is almost surely a measurable function of the initial value X0 and the processes (Bs)s t and (Ns)s t driving the SDE up to time t. In Section 4.1, we ≤ ≤ will see an example of a solution to an SDE that does not possess this property.

Remark. The stochastic integrals in (3.6) are well-deﬁned strict local martingales. Indeed, the local boundedness of the coefﬁcients guarantees local square integrability of the integrands as well as local boundedness of the jumps for the integral w.r.t.

~~N. The process σs(X) is not necessarily predictable, but observing that σs(X(ω)) =~~

~~σs (X(ω)) for P λ almost every (ω,s), we may deﬁne − ⊗~~

~~σs(X) dBs := σs (X) dBs. −~~

~~Lp Stability~~

~~In addition to the assumptions above, we assume from now on that the coefﬁcients in the SDE (3.4) satisfy a local Lipschitz condition:~~

~~UniversityofBonn WinterSemester2012/2013 120CHAPTER 3. SDE I: STRONG SOLUTIONS AND APPROXIMATION SCHEMES~~

~~Assumption (A1). For any t R, and for any open bounded set U Rn, there 0 ∈ ⊂ exists a constant L R such that the following Lipschitz condition Lip(t , U) holds: ∈ + 0~~

2 bt(x) bt(x) + σt(x) σt(x) + ct(x, ) ct(x, ) L2(ν) L sup xs xs | − | || − || • − • ≤ s t | − | ≤ for any t [0, t ] and x, x (R , Rn) with x , x U for s t . ∈ 0 ∈D + s s ∈ ≤ 0 We now derive an a priori estimate for solutions of (3.4) that is crucial for studying existence, uniqueness, and dependence on the initial condition:

Theorem 3.2 (A priori estimate). Fix p [2, ) and an open set U Rn, and let ∈ ∞ ⊆ T be an ( ) stopping time. Suppose that (X ) and (X ) are solutions of (3.4) taking Ft t t values in U for t < T , and let

~~p ϕ(t) := E sup Xs Xs . s~~

Proof. We only prove the assertion for p = 2. For p > 2, the proof can be carried out essentially in a similar way by relying on Burkholder’s inequality instead of Itô’s isometry. Clearly, (3.8) follows from (3.7) by Gronwell’s lemma. To prove (3.7), note that

~~t t Xt = X0 + bs(X) ds+ σs(X) dBs + cs (X,y) N(ds dy) t < T, 0 0 (0,t] S − ∀ × and an analogue equation holds for X. Hence for t t , ≤ 0~~

~~⋆ (X X)t T I + II + III + IV, where (3.9) − ∧ ≤ Stochastic Analysis Andreas Eberle 3.1. EXISTENCEANDUNIQUENESSOFSTRONGSOLUTIONS 121~~

~~I = X X , | 0 − 0| t T ∧ II = bs(X) bs(X) ds, 0 | − | u III = sup (σs(X) σs(X)) dBs , and u~~

E[I2] = ϕ(0), and t T t 2 2 ∧ ⋆ 2 2 E[II ] L t E (X X)s ds L t ϕ(s) ds. ≤ 0 − ≤ 0 Denoting by Mu and Ku the stochastic integrals in III and IV respectively, Doob’s inequality and Itô’s isometry imply

2 ⋆ 2 2 E[III ] = E[Mt T ] 4E[Mt T ] ∧ ≤ ∧ t T t ∧ 2 2 = 4E σs(X) σs(X) ds 4L ϕ(s) ds, 0 || − || ≤ 0 2 ⋆ 2 2 E[IV ] = E[Kt T ] 4E[Kt T ] ∧ ≤ ∧ t T t ∧ 2 2 = 4E cs (X,y) cs (X,y) ν(dy) ds 4L ϕ(s) ds. 0 | − − − | ≤ 0 The assertion now follows since by (3.9),

⋆ 2 2 2 2 2 ϕ(t) = E (X X)t T ] 4 E[I + II + III + IV . − ∧ ≤ The a priori estimate shows in particular that under a global Lipschitz condition, solutions depend continuously on the initial condition in mean square. Moreover, it implies pathwise uniqueness under a local Lipschitz condition:

~~Corollary 3.3 (Pathwise uniqueness). Suppose that Assumption (A1) holds. If (Xt) and (Xt) are strong solutions of (3.1) with X0 = X0 almost surely then~~

~~P Xt = Xt for any t = 1. UniversityofBonn WinterSemester2012/2013 122CHAPTER 3. SDE I: STRONG SOLUTIONS AND APPROXIMATION SCHEMES~~

Proof. For any open bounded set U Rn and t R , the a priori estimate in Theorem ⊂ 0 ∈ + 3.2 implies that X and X coincide almost surely on [0, t T c ) where T c denotes the 0 ∧ U U ﬁrst exit time from U.

~~Existence of strong solutions~~

~~To prove existence of strong solutions, we need an additional assumption:~~

~~Assumption (A2). For any t R , 0 ∈ +~~

~~2 sup ct(0,y) ν(dy) < . t~~

~~2 2 2 ⋆ 2 sup bt(x) + σt(x) + ct(x, y) ν(dy) C(t0) (1 + xt0 ) . (3.10) t~~

Theorem 3.4 (Itô). Let ξ : Ω Rn be a random variable that is independent of the → Brownian motion B and the Poisson random measure N. 1) Suppose that the local Lipschitz condition (A1) and (A2) hold. Then (3.1) has a

~~strong solution (Xt)t<ζ with initial condition X0 = ξ that is deﬁned up to the explosion time~~

~~ζ = sup T , where T = inf t 0 : X k . k k { ≥ | t| ≥ } 2) If, moreover, the global Lipschitz condition Lip(t , Rn) holds for any t R , 0 0 ∈ + then ζ = almost surely. ∞~~

Proof of 3.4. We ﬁrst prove existence of a global strong solution (Xt)t [0, ) assuming ∈ ∞ (A2) and a global Lipschitz condition Lip(t , Rn) for any t R . The ﬁrst assertion 0 0 ∈ + will then follow by localization.

~~Stochastic Analysis Andreas Eberle 3.1. EXISTENCEANDUNIQUENESSOFSTRONGSOLUTIONS 123~~

For proving global existence we may assume w.l.o.g. that ξ is bounded and thus square integrable. We then construct a sequence (Xn) of approximate solutions to (3.1) by a Picard-Lindelöf iteration, i.e., for t 0 and n Z we deﬁne inductively ≥ ∈ +

0 Xt := ξ, (3.11) t t n+1 n n n Xt := ξ + bs(X ) ds + σs(X ) dBs + cs (X ,y) N(ds dy). 0 0 − (0,t] S × Fix t [0, ). We will show below that Assumption (A2) and the global Lipschitz 0 ∈ ∞ condition imply that

~~(i) for any n N, Xn is a square integrable ( 0) semimartingale on [0, t ] (i.e., ∈ Ft 0 the sum of a square integrable martingale and an adapted process with square integrable total variation), and~~

~~(ii) there exists a ﬁnite constant C(t0) such that the mean square deviations~~

~~∆n := E[(Xn+1 Xn)⋆ 2]. t − t of the approximations Xn and Xn+1 satisfy~~

~~t ∆n+1 C(t ) ∆n ds for any n 0 and t t . t ≤ 0 s ≥ ≤ 0 0 Then, by induction, tn ∆n C(t )n ∆0 for any n N and t t . t ≤ 0 n! t ∈ ≤ 0~~

n In particular, ∞ ∆ < . An application of the Borel-Cantelli Lemma now shows n=1 t0 ∞ n that the limit Xs = limn Xs exists uniformly for s [0, t0] with probability one. →∞ ∈ Moreover, X is a fixed point of the Picard-Lindelöf iteration, and hence a solution of the SDE (3.1). Since t0 has been chosen arbitrarily, the solution is defined almost surely on [0, ), and by construction it is adapted w.r.t. the filtration ( 0). ∞ Ft We now show by induction that Assertion (i) holds. If Xn is a square integrable ( 0) semimartingale on [0, t ] then, by the linear growth condition (3.10), the process Ft 0 b (Xn) 2 + σ (Xn) 2 + c (Xn,y) 2 ν(dy) is integrable w.r.t. the product measure | s | || s || | s | UniversityofBonn WinterSemester2012/2013 124CHAPTER 3. SDE I: STRONG SOLUTIONS AND APPROXIMATION SCHEMES

P λ . Therefore, by Itô’s isometry, the integrals on the right hand side of (3.11) ⊗ (0,t0) all deﬁne square integrable ( 0) semimartingales, and thus Xn+1 is a square integrable Ft ( 0) semimartingale, too. Ft Assertion (ii) is a consequence of the global Lipschitz condition. Indeed, by the Cauchy- n Schwarz inequality, Itô’s isometry and Lip(t0, R ), there exists a ﬁnite constant C(t0) such that

∆n+1 = E Xn+2 Xn+1 ⋆ 2 t − t t t 2 2 3t E b (Xn+1) b (Xn) ds + 3 E σ (Xn+1) σ (Xn) ds ≤ s − s s − s 0 0 t +3 E c (Xn+1,y) c (Xn,y) 2 ν(dy) ds s − s 0 t C(t ) ∆n ds for any n 0 and t t . ≤ 0 s ≥ ≤ 0 0 This completes the proof of global existence under a global Lipschitz condition.

Finally, suppose that the coefficients b, σ and c only satisfy the local Lipschitz condition (A1). Then for k N and t R , we can find functions bk, σk and ck that are globally ∈ 0 ∈ + Lipschitz continuous and that agree with b, σ and c on paths (xt) taking values in the ball B(0,k) for t t . The solution X(k) of the SDE with coefficients bk, σk, ck is then ≤ 0 a solution of (3.1) up to t T where T denotes the first exit time of X(k) from B(0,k). ∧ k k By pathwise uniqueness, the local solutions obtained in this way are consistent. Hence they can be combined to construct a solution of (3.1) that is defined up to the explosion time ζ = sup Tk.

~~Non-explosion criteria~~

Theorem 3.4 shows that under a global Lipschitz and linear growth condition on the coefficients, the solution to (3.1) is defined for all times with probability one. How- ever, this condition is rather restrictive, and there are much better criteria to prove that the explosion time ζ is almost surely infinite. Arguably the most generally applicable

~~Stochastic Analysis Andreas Eberle 3.1. EXISTENCEANDUNIQUENESSOFSTRONGSOLUTIONS 125 non-explosion criteria are those based on stochastic Lyapunov functions. Consider for example an SDE of type~~

~~dXt = b(Xt) dt + σ(Xt) dBt (3.12)~~

n n n n d where b : R R and σ : R R × are locally Lipschitz continuous, and let → → 1 n ∂2 = a (x) + b(x) , a(x) = σ(x)σ(x)T , L 2 ij ∂x ∂x ∇ i,j=1 i j denote the corresponding generator.

Theorem 3.5 (Lyapunov condition for non-explosion). Suppose that there exists a function ϕ C2(Rn) such that ∈ (i) ϕ(x) 0 for any x Rn, ≥ ∈ (ii) ϕ(x) as x , and →∞ | |→∞ (iii) ϕ λϕ for some λ R . L ≤ ∈ + Then the strong solution of (3.1) with initial value x Rn exists up to ζ = almost 0 ∈ ∞ surely.

Proof. We ﬁrst remark that by (iii), Z := exp( λt)ϕ(X ) is a supermartingale up to t − t the ﬁrst exit time T of the local solution X from a ball B(0,k) Rn. Indeed, by the k ⊂ product rule and the Itô-Doeblin formula,

λt λt λt dZ = λe− ϕ(X) dt + e− dϕ(X) = dM + e− ( ϕ λϕ)(X) dt − L − holds on [0, Tk] with a martingale M up to Tk. Now we ﬁx t 0. Then, by the Optional Stopping Theorem and by Condition (i), ≥

ϕ(x0) = E [ϕ(X0)] E [exp( λ(t Tk)) ϕ(Xt T )] ≥ − ∧ ∧ k E [exp( λt) ϕ(X ); T t] ≥ − Tk k ≤ exp( λt) inf ϕ(y) P [Tk t] ≥ − y =k ≤ | | for any k N. As k , inf y =k ϕ(y) by (ii). Therefore, ∈ →∞ | | →∞

P [sup Tk t] = lim P [Tk t]=0 k ≤ →∞ ≤ for any t 0, i.e., ζ = sup T = almost surely. ≥ k ∞ UniversityofBonn WinterSemester2012/2013 126CHAPTER 3. SDE I: STRONG SOLUTIONS AND APPROXIMATION SCHEMES

By applying the theorem with the function ϕ(x)=1+ x 2 we obtain: | | Corollary 3.6. If there exists λ R such that ∈ + 2x b(x)+tr(a(x)) λ (1 + x 2) for any x Rn ≤ | | ∈ then ζ = almost surely. ∞ Note that the condition in the corollary is satisﬁed if

x b(x) const. x and tr a(x) const. x 2 x ≤ | | ≤ | | | | for sufﬁciently large x Rn, i.e., if the outward component of the drift is growing at ∈ most linearly, and the trace of the diffusion matrix is growing at most quadratically.

~~3.2 Stochastic ﬂow and Markov property~~

~~d Let Ω = C0(R+, R ) endowed with Wiener measure 0 and the canonical Brownian motion Wt(ω)= ω(t). We consider the SDE~~

~~dXt = bt(X) dt + σt(X) dWt, X0 = a, (3.13)~~

~~n n n d with progressively measurable coefﬁcients b, σ : R C(R , R ) R , R × respec- + × + → tively satisfying the global Lipschitz condition~~

~~b (x) b (x) + σ (x) σ (x) L (x x)⋆ t, x, x (3.14) | t − t | || t − t || ≤ − t ∀ for some ﬁnite constant L R , as well as ∈ +~~

sup bs(0) + σs(0) < t. (3.15) s [0,t] | | || || ∞ ∀ ∈ Then by Itô’s existence and uniqueness theorem, there exists a unique global strong a Rn solution (Xt )t 0 of (3.13) for any initial condition a . Our next goal is to show ≥ ∈ that there is a continuous modiﬁcation (t, a) ξa of (Xa). The proof is based on the → t t multidimensional version of Kolmogorov’s continuity criterion for stochastic processes:

~~Stochastic Analysis Andreas Eberle 3.2. STOCHASTIC FLOW AND MARKOV PROPERTY 127~~

Theorem 3.7 (Kolmogorov, Centsovˇ ). Suppose that (E, ) is a Banach space, || || C = n I is a product of real intervals I ,...,I R, and X : Ω E, u C, is k=1 k 1 n ⊂ u → ∈ an E-valued stochastic process (a random ﬁeld) indexed by C. If there exists constants γ,c,ε R such that ∈ + E X X γ c u v n+ε for any u, v C, (3.16) || u − v|| ≤ | − | ∈ then there exists a modiﬁcation (ξu)u C of (Xu)u C such that ∈ ∈ ξ ξ γ u v for any (3.17) E sup || − α|| < α [0,ε/γ). u=v u v ∞ ∈ | − | In particular, u ξ is almost surely α-Hölder continuous for any α < ε/γ. → u For the proof cf. e.g. [33, Ch. I, (2.1)].

Example. Brownian motion satisﬁes (3.16) with d = 1 and ε = γ 1 for any γ 2 − ∈ (2, ). Letting γ tend to , we see that almost every Brownian path is α-Hölder ∞ ∞ continuous for any α< 1/2. This result is sharp in the sense that almost every Brownian 1 path is not 2 -Hölder-continuous, cf. [13, Thm. 1.20].

~~Existence of a continuous ﬂow~~

~~We now apply the Kolmogorov-Centsovˇ continuity criterion to the solution a (Xa) → s of the SDE (3.13) as a function of its starting point.~~

~~Theorem 3.8 (Flow of an SDE). Suppose that (3.14) and (3.15) hold.~~

1) There exists a function ξ : Rn Ω C(R , Rn), (a, ω) ξa(ω) such that × → + → a a Rn (i) ξ =(ξt )t 0 is a strong solution of (3.13) for any a , and ≥ ∈ (ii) the map a ξa(ω) is continuous w.r.t. uniform convergence on ﬁnite time → intervals for any ω Ω. ∈

2) If σ(t, x)= σ(xt) and b(t, x)= b(xt) with Lipschitz continuous functions n n d n n d σ : R R × and b : R R × then ξ satisﬁes the cocycle property → → a a ξt (ω) Rn ξt+s(ω) = ξs (Θt(ω)) s, t 0, a (3.18) ∀ ≥ ∈ UniversityofBonn WinterSemester2012/2013 128CHAPTER 3. SDE I: STRONG SOLUTIONS AND APPROXIMATION SCHEMES

~~for 0-almost every ω, where~~

~~Θ (ω) = ω( + t) C(R , Rd) t ∈ + denotes the shifted path, and the deﬁnition of ξ has been extended by~~

ξ(ω) := ξ(ω ω(0)) (3.19) − to paths ω C(R , Rd) with starting point ω(0) =0. ∈ + Proof. 1) We fix p>d. By the a priori estimate in Theorem 3.2 there exists a finite constant c R such that ∈ + E[(Xa Xa)⋆ p] c ect a a p for any t 0 and a, a Rn, (3.20) − e t ≤ | − | ≥ ∈ a where X denotes a version of the strong solution of (3.13) with initial condition a. Now fix t R . We apply the Kolmogorov-Centsovˇ Theorem with E = C([0, t], Rn) ∈ + endowed with the supremum norm X = X⋆. By (3.20), there exists a modification ξ || ||t t a a of (Xs )s t,a Rn such that a (ξs )s t is almost surely α-Hölder continuous w.r.t. t ≤ ∈ → ≤ p n a for any α < − . Clearly, for t t , the almost surely continuous map (s, a) ξ p 1 ≤ 2 → s constructed on [0, t ] Rn coincides almost surely with the restriction of the correspond- 1 × ing map on [0, t ] Rn. Hence we can almost surely extend the definition to R Rn 2 × + × in a consistent way.

~~2) Fix t 0 and a Rn. Then -almost surely, both sides of (3.18) solve the same ≥ ∈ 0 SDE as a function of s. Indeed,~~

t+s t+s a a a a ξt+s = ξt + b(ξu) du + σ(ξu) dWu t t s s a a a = ξ + b(ξ ) dr + σ(ξ ) d (Wr Θt) , t t+r t+r ◦ 0 0 s s a a a ξξt Θ = ξa + b ξξt Θ dr + σ(ξξt Θ ) d(W Θ ) s ◦ t t r ◦ t r ◦ t r ◦ t 0 0 hold 0-almost surely for any s 0 where r Wr Θt = Wr+t is again a Brownian a ≥a → ◦ ξt ξt (ω) motion, and ξr Θ (ω) := ξr (Θ (ω)). Pathwise uniqueness now implies ◦ t t a ξa = ξξt Θ for any s 0, almost surely. t+s s ◦ t ≥ Stochastic Analysis Andreas Eberle 3.2. STOCHASTIC FLOW AND MARKOV PROPERTY 129

~~Continuity of ξ then shows that the cocycle property (3.18) holds with probability one for all s, t and a simultaneously.~~

Remark (Extensions). 1) Joint Hölder continuity in t and a: Since the constant p in the proof above can be chosen arbitrarily large, the argument yields α-Hölder continuity of a ξa for any α< 1. By applying Kolmogorov’s criterion in dimension n+1, itisalso → possible to prove joint Hölder continuity in t and a. In Section 5.1 we will prove that under a stronger assumption on the coefficients b and σ, the flow is even continuously differentiable in a. 2) SDE with jumps: The first part of Theorem 3.8 extends to solutions of SDE of type (3.4) driven by a Brownian motion and a Poisson point process. In that case, under a global Lipschitz condition the same arguments go through if we replace C([0, t], Rn) by the Banach space ([0, t], Rn) when applying Kolmogorov’s criterion. Hence in spite D of the jumps, the solution depends continuously on the initial value a ! 3) Locally Lipschitz coefficients: By localization, the existence of a continuous flow can also be shown under local Lipschitz conditions, cf. e.g. [32]. Notice that in this case, the explosion time depends on the initial value.

~~Above we have shown the existence of a continuous ﬂow for the SDE (3.13) on the canonical setup. From this we can obtain strong solutions on other setups:~~

~~Exercise. Show that the unique strong solution of (3.13) w.r.t. an arbitrary driving a a Brownian motion B instead of W is given by Xt (ω)= ξt (B(ω)).~~

~~Markov property~~

~~In the time-homogeneous diffusion case, the Markov property for solutions of the SDE (3.13) is a direct consequence of the cocycle property:~~

Corollary 3.9. Suppose that σ(t, x)= σ(xt) and b(t, x)= b(xt) with Lipschitz contin- Rn Rn d Rn Rn a uous functions σ : × and b : . Then (ξt )t 0 is a time-homogeneous → → ≥ W,P ( t ) Markov process with transition function F p (a, B) = P [ξa B], t 0, a Rn. t t ∈ ≥ ∈

~~UniversityofBonn WinterSemester2012/2013 130CHAPTER 3. SDE I: STRONG SOLUTIONS AND APPROXIMATION SCHEMES~~

~~Proof. Let f : Rn R be a measurable function. Then for 0 s t, → ≤ ≤ Θ (ω) = ω(t)+ ω(t + ) ω(t) , t − and hence, by the cocycle property and by (3.19),~~

~~a f(ξa (ω)) = f ξξt (ω) ω(t + ) ω(t) s+t s − for a.e. ω. Since ω(t+ ) ω(t) is a Brownian motion starting at 0 independent of W,P , − Ft we obtain~~

~~W,P a E f(ξa ) (ω) = E f(ξξt (ω)) = (p f)(ξa(ω)) almost surely. s+t |Ft s s t~~

~~Remark. Without pathwise uniqueness, both the cocycle and the Markov property do not hold in general.~~

~~Continuity of local time~~

The Kolmogorov-Centsovˇ continuity criterion can also be applied to prove the existence of a jointly continuous version (a, t) La of the local time of a continuous local → t martingale. More generally, recall that the local time of a continuous semimartingale X = M + A is deﬁned by the Tanaka formula

t t 1 a + + Lt = (X0 a) (Xt a) I(a, )(Xs) dMs I(a, )(Xs) dAs (3.21) 2 − − − − ∞ − ∞ 0 0 almost surely for any a R. ∈ Theorem 3.10 (Yor). There exists a version (a, t) La of the local time process that → t is continuous in t and càdlàg in a with

~~t a a Lt Lt − = 2 I Xs=a dAs. (3.22) − { } 0 In particular, (a, t) La is jointly continuous if M is a continuous local martingale. → t~~

~~Stochastic Analysis Andreas Eberle 3.2. STOCHASTIC FLOW AND MARKOV PROPERTY 131~~

Proof. By localization, we may assume that M is a bounded martingale and A has (1) + bounded total variation V (A). The map (a, t) (Xt a) is jointly continuous in t ∞ → − and a. Moreover, by dominated convergence,

~~t a Zt := I(a, )(Xs) dAs ∞ 0 is continuous in t and càdlàg in a with~~

~~t a a Zt Zt − = I a (Xs) dAs. − − { } 0 Therefore it is sufﬁcient to prove that~~

t a Yt := I(a, )(Xs) dMs ∞ 0 a R Rn has a version such that the map a (Ys )s t from to C([0, t], ) is continuous for → ≤ any t [0, ). ∈ ∞ Hence ﬁx t 0 and p 4. By Burkholder’s inequality, ≥ ≥ s p a b ⋆p (3.23) E Y Y t = E sup I(a,b](X) dM − s

t t I(a,b](X) d[M] = I(a,b](X) d[X] 0 0 t = 2 f(Xt) f(X0) f ′(X) dX − − − 0 t 2 (1) (b a) + 2 f ′(X) dM + b a V (A) ≤ − | − | t 0 UniversityofBonn WinterSemester2012/2013 132CHAPTER 3. SDE I: STRONG SOLUTIONS AND APPROXIMATION SCHEMES

~~2 Here we have used in the last step that f ′ b a and f (b a) /2. Combining | |≤| − | | | ≤ − this estimate with 3.23 and applying Burkholder’s inequality another time, we obtain~~

t p/4 a b ⋆p p/2 2 E Y Y C (p, t) b a + E f ′(X) d[M] − t ≤ 2 | − | 0 C (p, t) b a p/2 (1 + [M]p/4) ≤ 2 | − | t with a finite constant C (p, t). The existence of a continuous modification of a 2 → a ˇ (Ys )s t now follows from the Kolmogorov-Centsov Theorem. ≤ a Remark. 1) The proof shows that for a continuous local martingale, a (Ls )s t is → ≤ α-Hölder continuous for any α< 1/2 and t R . ∈ + a â 2) For a continuous semimartingale, Lt − = Lt by (3.22).

~~3.3 Stratonovich differential equations~~

Replacing Itô by Statonovich integrals has the advantage that the calculus rules (product rule, chain rule) take the same form as in classical differential calculus. This is useful for explicit computations (Doss-Sussman method), for approximating solutions of SDE by solutions of ordinary differential equations, and in stochastic differential geometry. For simplicity, we only consider Stratonovich calculus for continuous semimartingales, cf. [32] for the discontinuous case. Let X and Y be continuous semimartingales on a filtered probability space (Ω, ,P, ( )). A Ft Definition (Fisk-Stratonovich integral). The Stratonovich integral X dY is the ◦ continuous semimartingale defined by t t 1 X dY := X dY + [X,Y ] for any t 0. s ◦ s s s 2 t ≥ 0 0 Note that a Stratonovich integral w.r.t. a martingale is not a local martingale in general. The Stratonovich integral is a limit of trapezoidal Riemann sum approximations:

~~Lemma 3.11. If (π ) is a sequence of partitions of R with mesh(π ) 0 then n + n → t Xs + Xs′ t Xs dYs = lim ∧ (Ys′ t Ys) in the ucp sense. ◦ n 2 ∧ − 0 →∞ s πn s~~

~~Stochastic Analysis Andreas Eberle 3.3. STRATONOVICH DIFFERENTIAL EQUATIONS 133~~

~~t Proof. This follows since X dY = ucp - lim Xs (Ys′ t Ys) and 0 s~~

For Stratonovich integrals w.r.t. continuous semimartingales, the classical chain rule holds: Theorem 3.12. Let X =(X1,...,Xd) with continuous semimartingales Xi. Then for any function F C2(Rd), ∈ d t ∂F F (X ) F (X ) = (X ) dXi t 0. (3.24) t − 0 ∂xi s ◦ s ∀ ≥ i=1 0 Proof. To simplify the proof we assume F C3. Under this condition, (3.24) is just a ∈ reformulation of the Itô rule d t ∂F 1 d t ∂2F F (X ) F (X ) = (X ) dXi + (X ) d[Xi,Xj] . t − 0 ∂xi s s 2 ∂xi∂xj s s i=1 0 i,j=1 0 (3.25) 2 ∂F Indeed, applying Itô’s rule to the C function ∂xi shows that ∂F ∂2F (X ) = A + (X ) dXj ∂xi t t ∂xi∂xj s s j for some continuous ﬁnite variation process A. Hence the difference between the Statonovich integral in (3.24) and the Itô integral in (3.25) is 2 1 ∂F i 1 ∂ F j i (X),X = (Xs) d[X ,X ]s. 2 ∂xi t 2 ∂xi∂xj j

Remark. For the extension of the proof to C2 functions F see e.g. [32], where also a generalization to càdlàg semimartingales is considered. The product rule for Stratonovich integrals is a special case of the chain rule: Corollary 3.13. For continuous semimartingales X,Y , t t X Y X Y = X dY + Y dX t 0. t t − 0 0 s ◦ s s ◦ s ∀ ≥ 0 0 Exercise (Associative law). Prove an associative law for Stratonovich integrals.

~~UniversityofBonn WinterSemester2012/2013 134CHAPTER 3. SDE I: STRONG SOLUTIONS AND APPROXIMATION SCHEMES~~

~~Stratonovich SDE~~

Since Stratonovich integrals differ from the corresponding Itô integrals only by the covariance term, equations involving Stratonovich integrals can be rewritten as Itô equations and vice versa, provided the coefﬁcients are sufﬁciently regular. We consider a Stratonovich SDE in Rd of the form

~~d dX = b(X ) dt + σ (X ) dBk, X = x , (3.26) ◦ t t k t ◦ t 0 0 k=1 with x Rn, continuous vector ﬁelds b, σ ,...,σ C(Rn, Rn), and an Rd-valued 0 ∈ 1 d ∈ Brownian motion (Bt).~~

~~Exercise (Stratonovich to Itô conversion). 1) Prove that for σ ,...,σ C1(Rn, Rn), 1 d ∈ the Stratonovich SDE (3.26) is equivalent to the Itô SDE~~

d k dXt = b(Xt) dt + σk(Xt) dBt , X0 = x0, (3.27) k=1 where 1 d b := b + σ σ . 2 k ∇ k k=1 2) Conclude that if b and σ1,...,σd are Lipschitz continuous, then there is a unique strong solution of (3.26). Theorem 3.14 (Martingale problem for Stratonovich SDE). Let b C(Rn, Rn) and ∈ 2 n n σ1,...,σd C (R , R ), and suppose that (Xt)t 0 is a solution of (3.26) on a given ∈ ≥ setup (Ω, ,P, ( ), (B )). Then for any function F C3(Rn), the process A Ft t ∈ t M F = F (X ) F (X ) ( F )(X ) ds, t t − 0 − L s 0 1 d F = σ (σ F )+ b F, L 2 k ∇ k ∇ ∇ k=1 is a local ( P ) martingale. Ft

~~Stochastic Analysis Andreas Eberle 3.3. STRATONOVICH DIFFERENTIAL EQUATIONS 135~~

~~Proof. By the Stratonovich chain rule and by (3.26),~~

t F (Xt) F (X0) = F (X) dX − 0 ∇ ◦ t t k = (b F )(X) dt + (σk F )(X) dB . (3.28) 0 ∇ 0 ∇ ◦ k By applying this formula to σ F , we see that k ∇ (σ F )(X ) = A + σ (σ F )(X) dBl k ∇ t t l ∇ k ∇ l with a continuous ﬁnite variation process (At). Hence

~~t t k k k (σk F )(X) dB = (σk F )(X) dB + [(σk F )(X), B ]t 0 ∇ ◦ 0 ∇ ∇ t = local martingale + σ (σ F )(X)dt. k ∇ k ∇ 0 (3.29)~~

~~The assertion now follows by (3.28) and (3.29).~~

The theorem shows that the generator of a diffusion process solving a Stratonovich SDE is in sum of squares form. In geometric notation, one brieﬂy writes b for the derivative b in the direction of the vector ﬁeld b. The generator then takes the form ∇ 1 = σ2 + b L 2 k k Brownian motion on hypersurfaces

One important application of Stratonovich calculus is stochastic differential geometry. Itô calculus can not be used directly for studying stochastic differential equations on manifolds, because the classical chain rule is essential for ensuring that solutions stay on the manifold if the driving vector ﬁelds are tangent vectors. Instead, one considers Stratonovich equations. These are converted to Itô form when computing expectation values. To avoid differential geometric terminology, we only consider Brownian motion on a hypersurface in Rn+1, cf. [34], [18] and [20] for stochastic calculus on more general Riemannian manifolds.

~~UniversityofBonn WinterSemester2012/2013 136CHAPTER 3. SDE I: STRONG SOLUTIONS AND APPROXIMATION SCHEMES~~

n+1 Let f C∞(R ) and suppose that c R is a regular value of f, i.e., f(x) =0 for ∈ ∈ ∇ 1 any x f − (c). Then by the implicit function theorem, the level set ∈ 1 n+1 M = f − (c) = x R : f(x)= c c ∈ is a smooth n-dimensional submanifold of Rn+1. For example, if f(x)= x 2 and c =1 | | n then Mc is the n-dimensional unit sphere S .

~~For x Mc, the vector ∈ f(x) n(x) = ∇ Sn f(x) ∈ |∇ | is the unit normal to Mc at x. The tangent space to Mc at x is the orthogonal complement~~

~~T M = span n(x) ⊥ . x c { } Let P (x): Rn+1 T M denote the orthogonal projection onto the tangent space w.r.t. → x c the Euclidean metric, i.e.,~~

~~P (x)v = v v n(x) n(x), v Rn+1. − ∈~~

~~η(x)~~

~~Mc′ TxMc~~

~~Stochastic Analysis Andreas Eberle 3.3. STRATONOVICH DIFFERENTIAL EQUATIONS 137~~

~~For k 1,...,n +1 , we set P (x)= P (x)e . ∈{ } k k~~

~~Deﬁnition. A Brownian motion on the hypersurface Mc with initial value x M is 0 ∈ c a solution (Xt) of the Stratonovich SDE~~

~~n+1 dX = P (X ) dB = P (X ) dBk, X = x , (3.30) ◦ t t ◦ t k t ◦ t 0 0 k=1 n+1 with respect to a Brownian motion (Bt) on R .~~

We now assume for simplicity that Mc is compact. Then, since c is a regular value of f, the vector fields Pk are smooth with bounded derivatives of all orders in a neigh- n+1 bourhood U of Mc in R . Therefore, there exists a unique strong solution of the SDE (3.30) in Rn+1 that is defined up to the first exit time from U. Indeed, this solution stays on the submanifold Mc for all times:

~~Theorem 3.15. If X is a solution of (3.30) with x M then almost surely, X M 0 ∈ c t ∈ c for any t 0. ≥ The proof is very simple, but it relies on the classical chain rule in an essential way:~~

~~Proof. We have to show that f(Xt) is constant. This is an immediate consequence of the Stratonovich formula:~~

~~t n+1 t k f(Xt) f(X0) = f(Xs) dXs = f(Xs) Pk(Xs) dBs = 0 − 0 ∇ ◦ 0 ∇ ◦ k=1 since P (x) is orthogonal to f(x) for any x. k ∇~~

Although we have deﬁned Brownian motion on the Riemannian manifold Mc in a non- intrinsic way, one can verify that it actually is an intrinsic object and does not depend on n+1 the embedding of Mc into R that we have used. We only convince ourselves that the corresponding generator is an intrinsic object. By Theorem 3.14, the Brownian motion

~~(Xt) constructed above is a solution of the martingale problem for the operator~~

1 n+1 1 n+1 = (P )P = P 2. L 2 k ∇ k ∇ 2 k k=1 k=1 1 From differential geometry it is well-known that this operator is 2 ∆Mc where ∆Mc denotes the (intrinsic) Laplace-Beltrami operator on Mc.

~~UniversityofBonn WinterSemester2012/2013 138CHAPTER 3. SDE I: STRONG SOLUTIONS AND APPROXIMATION SCHEMES~~

Exercise (Itô SDE for Brownian motion on Mc). Prove that the SDE (3.30) can be written in Itô form as 1 dX = P (X ) dB κ(X )n(X ) dt t t t − 2 t t 1 n where κ(x)= n div (x) is the mean curvature of Mc at x.

~~Doss-Sussmann method~~

~~Stratonovich calculus can also be used to obtain explicit solutions for stochastic differ- n ential equations in R that are driven by a one-dimensional Brownian motion (Bt). We consider the SDE~~

dX = b(X ) dt + σ(X ) dB , X = a, (3.31) ◦ t t t ◦ t 0 where a Rn, b : Rn Rn is Lipschitz continuous and σ : Rn Rn is C2 with ∈ → → bounded derivatives. Recall that (3.31) is equivalent to the Itô SDE 1 dX = b + σ σ (X ) dt + σ(X ) dB , X ¸ = ¸a. (3.32) t 2 ∇ t t t 0 We ﬁrst determine an explicit solution in the case b 0 by the ansatz X = F (B ) ≡ t t where F C2(R, Rn). By the Stratonovich rule, ∈

~~dX = F ′(B ) dB = σ(F (B )) dB ◦ t t ◦ t t ◦ t provided F is a solution of the ordinary differential equation~~

~~F ′(s) = σ(F (s)). (3.33)~~

~~Hence a solution of (3.31) with initial condition X0 = a is given by~~

Xt = F (Bt, a) where (s, x) F (s, x) is the flow of the vector field σ, i.e., F ( , a) is the unique → solution of (3.33) with initial condition a. Recall from the theory of ordinary differential equations that the flow of a vector field σ

Stochastic Analysis Andreas Eberle 3.3. STRATONOVICH DIFFERENTIAL EQUATIONS 139 as above deﬁnes a diffeomorphism a F (s, a) for any s R. To obtain a solution of → ∈ (3.31) in the general case, we try the “variation of constants” ansatz

Xt = F (Bt,Ct) (3.34) with a continuous semimartingale (Ct) satisfying C0 = a. In other words: we make a time-dependent coordinate transformation in the SDE that is determined by the ﬂow F and the driving Brownian path (Bt). By applying the chain rule to (3.34), we obtain ∂F ∂F dX = (B ,C ) dB + (B ,C ) dC ◦ t ∂s t t ◦ t ∂x t t ◦ t ∂F = σ(X ) dB + (B ,C ) dC t ◦ t ∂x t t ◦ t where ∂F (s, ) denotes the Jacobi matrix of the diffeomorphism F (s, ). Hence (X ) is ∂x t a solution of the SDE (3.31) provided (Ct) is almost surely absolutely continuous with derivative d ∂F 1 C = (B ,C )− b(F (B ,C )). (3.35) dt t ∂x t t t t For every given ω, the equation (3.35) is an ordinary differential equation for Ct(ω) which has a unique solution. Working out these arguments in detail yields the following result:

Theorem 3.16 (Doss 1977, Sussmann 1978). Suppose that b : Rn Rn is Lipschitz → continuous and σ : Rn Rn is C2 with bounded derivatives. Then the flow F of the → vector field σ is well-defined, F (s, ) is a C2 diffeomorphism for any s R, and the ∈ equation (3.35) has a unique pathwise solution (Ct)t 0 satisfying C0 = a. Moreover, ≥ the process Xt = F (Bt,Ct) is the unique strong solution of the equation (3.31), (3.32) respectively.

We refer to [23] for a detailed proof. Exercise (Computing explicit solutions). Solve the following Itô stochastic differential equations explicitly: 1 dX = X dt + 1+ X2 dB , X = 0, (3.36) t 2 t t t 0 2 2 dXt = Xt(1 + Xt) dt +(1+ Xt ) dBt, X0 = 1. (3.37) Do the solutions explode in ﬁnite time?

~~UniversityofBonn WinterSemester2012/2013 140CHAPTER 3. SDE I: STRONG SOLUTIONS AND APPROXIMATION SCHEMES~~

~~Exercise (Variation of constants). We consider nonlinear stochastic differential equations of the form~~

~~dXt = f(t, Xt) dt + c(t)Xt dBt, X0 = x, where f : R+ R R and c : R+ R are continuous (deterministic) functions. × → → Proceed as follows :~~

~~a) Find an explicit solution Z of the equation with f 0. t ≡ b) To solve the equation in the general case, use the Ansatz~~

~~X = C Z . t t t Show that the SDE gets the form~~

dC (ω) t = f(t, Z (ω) C (ω))/Z (ω); C = x. (3.38) dt t t t 0 Note that for each ω Ω, this is a deterministic differential equation for the ∈ function t C (ω). We can therefore solve (3.38) with ω as a parameter to ﬁnd → t Ct(ω).

~~c) Apply this method to solve the stochastic differential equation 1 dXt = dt + αXt dBt ; X0 = x> 0 , Xt where α is constant.~~

~~d) Apply the method to study the solution of the stochastic differential equation~~

~~γ dXt = Xt dt + αXt dBt ; X0 = x> 0 ,~~

~~where α and γ are constants. For which values of γ do we get explosion?~~

~~Wong Zakai approximations of SDE~~

~~A natural way to approximate the solution of an SDE driven by a Brownian motion is to replace the Brownian motion by a smooth approximation. The resulting equation can~~

Stochastic Analysis Andreas Eberle 3.3. STRATONOVICH DIFFERENTIAL EQUATIONS 141 then be solved pathwise as an ordinary differential equation. It turns out that the limit of this type of approximations as the driving smoothed processes converge to Brownian motion will usually solve the corresponding Stratonovich equation.

~~d Suppose that (Bt)t 0 is a Brownian motion in R with B0 = 0. For notational conve- ≥ nience we deﬁne Bt := 0 for t< 0. We approximate B by the smooth processes~~

2 (k) 1/2 t B := B⋆ϕ , ϕ (t) = (2πε)− exp . 1/k ε − 2ε Other smooth approximations could be used as well, cf. [23] and [21]. Let X(k) denote the unique solution to the ordinary differential equation

~~d d X(k) = b(X(k))+ σ(X(k)) B(k), X(k) = a (3.39) dt t t t dt t 0~~

~~n n n n d with coefﬁcients b : R R and σ : R R × . → →~~

~~Theorem 3.17 (Wong, Zakai 1965). Suppose that b is C1 with bounded derivatives and σ is C2 with bounded derivatives. Then almost surely as k , →∞~~

~~X(k) X uniformly on compact intervals, t −→ t where (Xt) is the unique solution of the Stratonovich equation~~

~~dX = b(X ) dt + σ(X ) dB , X = a. ◦ t t t ◦ t 0~~

If the driving Brownian motion is one-dimensional, there is a simple proof based on the Doss-Sussman representation of solutions. This shows that X(k) and X can be (k) (k) (k) represented in the form Xt = F (Bt ,Ct ) and Xt = F (Bt,Ct) with the flow F of the same vector field σ, and the processes C(k) and C solving (3.35) w.r.t. B(k), B respectively. Therefore, it is not difficult to verify that almost surely, X(k) X → uniformly on compact time intervals, cf. [23]. The proof in the more interesting general case is much more involved, cf. e.g. Ikeda & Watanabe [21, Ch. VI, Thm. 7.2].

~~UniversityofBonn WinterSemester2012/2013 142CHAPTER 3. SDE I: STRONG SOLUTIONS AND APPROXIMATION SCHEMES~~

~~3.4 Stochastic Taylor expansions and numerical methods~~

~~The goal of this section is to analyse the convergence order of numerical schemes for Itô stochastic differential equations of type~~

dXt = b(Xt) dt + σk(Xt) dBt (3.40) k=1 in RN , N N. We will assume throughout this section that the coefﬁcients b, σ ,...,σ ∈ 1 d N 1 d are C∞ vector ﬁelds on R , and B = (B ,...,B ) is a d-dimensional Brownian motion. Below, it will be convenient to set

~~0 Bt := t.~~

~~A solution of (3.40) satisﬁes~~

t+h d t+h k Xt+h = Xt + b(Xs) ds + σk(Xs) dBs (3.41) t t k=1 for any t, h 0. By approximating b(X ) and σ (X ) in (3.41) by b(X ) and σ (X ) re- ≥ s k s t k t spectively, we obtain an Euler approximation of the solution with step size h. Similarly, higher order numerical schemes can be obtained by approximating b(Xs) and σk(Xs) by stochastic Taylor approximations.

~~Itô-Taylor expansions~~

~~N Suppose that X is a solution of (3.40), and let f C∞(R ). Then the Itô-Doeblin ∈ formula for f(X) on the interval [t, t + h] can be written in the compact form~~

d t+h k f(Xt+h) = f(Xt) + ( kf)(Xs) dBs (3.42) t L k=0 for any t, h 0,where B0 = t, a = σσT , ≥ t 1 N ∂2f f = aij + b f, and (3.43) L0 2 ∂xi∂xj ∇ i,j=1 f = σ f, for k =1, . . . , d. (3.44) Lk k ∇ Stochastic Analysis Andreas Eberle 3.4. STOCHASTIC TAYLOR EXPANSIONS AND NUMERICAL METHODS 143

~~By iterating this formula, we obtain Itô-Taylor expansions for f(X). For example, a ﬁrst iteration yields~~

d t+h d t+h s k l k f(Xt+h) = f(Xt)+ ( kf)(Xt) dBs + ( l kf)(Xr) dBr dBs . L t t t L L k=0 k,l=0 The ﬁrst two terms on the right hand side constitute a ﬁrst order Taylor expansion for f(X) in terms of the processes Bk, k = 0, 1,...,d, and the iterated Itô integral in the third term is the corresponding remainder. Similarly, we obtain higher order expansions in terms of iterated Itô integrals where the remainders are given by higher order iterated integrals, cf. Theorem ?? below. The next lemma yields L2 bounds on the remainder terms:

~~Lemma 3.18. Suppose that G : Ω (t, t + h) R is an adapted process in 2(P × → L ⊗ λ(t,t+h)). Then~~

2 t+h s1 sn−1 n+m(k) kn k2 k1 h 2 E Gsn dBsn dBs2 dBs1 sup E Gs t t t ≤ n! s [t,t+h] ∈ for any n N and k =(k ,...,k ) 0, 1,...,d n, where ∈ 1 n ∈{ } m(k) := 1 i n : k =0 |{ ≤ ≤ i }| denotes the number of integrations w.r.t. dt.

~~Proof. By Itô’s isometry and the Cauchy-Schwarz inequality,~~

t+h 2 t+h E G dBk E G2 ds for any k =0, and s s ≤ s t t t+h 2 t+h E G ds h E G2 ds. s ≤ s t t By iteratively applying these estimates we see that the second moment of the iterated integral in the assertion is bounded from above by

~~t+h s1 sn−1 hm(k) E[G2 ] ds ds ds . sn n 2 1 t t t~~

~~UniversityofBonn WinterSemester2012/2013 144CHAPTER 3. SDE I: STRONG SOLUTIONS AND APPROXIMATION SCHEMES~~

The lemma can be applied to control the strong convergence order of stochastic Taylor expansions. For k N we denote by Ck(R) the space of all Ck functions with bounded ∈ b k derivatives up to order k. Notice that we do not assume that the functions in Cb are bounded.

~~Deﬁnition (Stochastic convergence order). Suppose that Ah, h > 0, and A are random variables, and let α> 0.~~

~~2 1) Ah converges to A with strong L order α iff~~

~~E A A 2 1/2 = O(hα). | h − | 2) Ah converges to A with weak order α iff~~

~~α 2(α+1) E [f(A )] E [f(A)] = O(h ) for any f C⌈ ⌉(R). h − ∈ b~~

Notice that convergence with strong order α requires that the random variables are de- ﬁned on a common probability space. For convergence with weak order α this is not necessary. If Ah converges to A with strong order α then we also write

~~A = A + (hα). h O~~

Examples. 1) If B is a Brownian motion then Bt+h converges to Bt almost surely as h 0. By the law of the iterated logarithm, the pathwise convergence order is ↓ 1/2 1 B B = O(h log log h− ) almost surely. t+h − t On the other hand, the strong L2 order is 1/2, and the weak order is 1 since by Kol- mogorov’s forward equation,

t+h 1 h E[f(B )] E[f(B )] = E[ ∆f(B )] ds sup ∆f t+h − t 2 s ≤ 2 t for any f C2. The exercise below shows that similar statements hold for more general ∈ b Itô diffusions. 2) The n-fold iterated Itô integrals w.r.t. Brownian motion considered in Lemma 3.18 have strong order (n + m)/2 where m is the number of time integrals.

~~Stochastic Analysis Andreas Eberle 3.4. STOCHASTIC TAYLOR EXPANSIONS AND NUMERICAL METHODS 145~~

Exercise (Order of Convergence for Itô diffusions). Let (Xt)t 0 be an N-dimensional ≥ stochastic process satisfying the SDE (3.40) where b, σ : RN RN ,k =1,...,d, are k → bounded continuous functions, and B is a d-dimensional Brownian motion. Prove that as h 0, ↓

~~2 1) Xt+h converges to Xt with strong L order 1/2.~~

~~2) Xt+h converges to Xt with weak order 1.~~

Corollary 3.19 (Itô-Taylor expansion with remainder of order α). Suppose that α = k/2 for some k N. If X is a solution of (3.40) with coefﬁcients b, σ1,...,σd 2α ∈ 2α+1 ∈ C⌊ ⌋(RN , RN ) then the following expansions hold for any f C⌊ ⌋(RN ): b ∈ b

f(X ) = − f (X ) (3.45) t+h Lkn Lkn 1 Lk1 t × n<2α k:n+m(k)<2α t+h s1 sn−1 G dBkn dBk2 dBk1 + (hα), × sn sn s2 s1 O t t t hn E [f(X )] = E [( nf)(X )] + O(hα). (3.46) t+h L0 t n! n<α Proof. Iteration of the Itô-Doeblin formula (3.42) shows that (3.45) holds with a remainder term that is a sum of iterated integrals of the form

t+h s1 sn−1 kn k2 k1 − f (X ) dB dB dB Lkn Lkn 1 Lk1 sn sn s2 s1 t t t with k = (k1,...,kn) satisfying n + m(k) > 2α and n 1+ m(k1,...,kn 1) < 2α. − − By Lemma 3.18, these iterated integrals are of strong L2 order (n + m(k))/2. Hence the full remainder term is of the order (hα). O Equation (3.46) follows easily by iterating the Kolmogorov forward equation

t+h E [f(X )] = E [f(X )] + E [( f)(X )] ds. t+h t L0 s t Alternatively, it can be derived from (3.45) by noting that all iterated integrals involving at least one integration w.r.t. a Brownian motion have mean zero.

~~UniversityofBonn WinterSemester2012/2013 146CHAPTER 3. SDE I: STRONG SOLUTIONS AND APPROXIMATION SCHEMES~~

~~Remark (Computation of iterated Itô integrals). Iterated Itô integrals involving only a single one dimensional Brownian motion B can be computed explicitly from the Brownian increments. Indeed,~~

t+h s1 sn−1 dB dB dB = h (h, B B )/n!, sn s2 s1 n t+h − t t t t where hn denotes the n-th Hermite polynomial, cf. (2.58). In the multi-dimensional case, however, the iterated Itô integrals can not be represented in closed form as functions of Brownian increments. Therefore, in higher order numerical schemes, these integrals have to be approximated separately. For example, the second iterated Itô integral h s kl k l k l Ih = dBr dBs = Bs dBs 0 0 kl lk k l of two components of a d dimensional Brownian motion satisﬁes Ih + Ih = BhBh. Hence the symmetric part can be computed easily. However, the antisymmetric part Ikl Ilk is the Lévy area process of the two dimensional Brownian motion (Bk, Bl). h − h The Lévy area can not be computed explicitly from the increments if k = l. Controlling the Lévy area is crucial for a pathwise stochastic integration theory, cf. [17, 26].

Exercise (Lévy Area). If c(t)=(x(t),y(t)) is a smooth curve in R2 with c(0) = 0, then t t t A(t) = (x(s)y′(s) y(s)x′(s)) ds = xdy y dx − − 0 0 0 describes the area that is covered by the secant from the origin to c(s) in the interval

~~[0, t]. Analogously, for a two-dimensional Brownian motion Bt =(Xt,Yt) with B0 =0, one deﬁnes the Lévy Area~~

~~t t A := X dY Y dX . t s s − s s 0 0 1) Let α(t), β(t) be C1-functions, p R, and ∈ α(t) V = ipA X2 + Y 2 + β(t) . t t − 2 t t~~

~~Vt 2 2 Show using Itô’s formula, that e is a local martingale provided α′(t)= α(t) p − and β′(t)= α(t).~~

~~Stochastic Analysis Andreas Eberle 3.4. STOCHASTIC TAYLOR EXPANSIONS AND NUMERICAL METHODS 147~~

~~2) Let t [0, ). The solutions of the ordinary differential equations for α and β 0 ∈ ∞ with α(t0)= β(t0)=0 are~~

α(t) = p tanh(p (t t)) , 0 − β(t) = log cosh(p (t t)) . − 0 − Conclude that ipA 1 E e t0 = p R . cosh(pt0) ∀ ∈ 3) Show that the distribution of At is absolutely continuous with density 1 fAt (x) = πx . 2t cosh( 2t )

~~Numerical schemes for SDE~~

Let X be a solution of the SDE (3.40), (3.41) respectively. By applying the Itô-Doeblin formula to σ (X ) and taking into account all terms up to strong order (h1), we obtain k s O the Itô-Taylor expansion

d X X = b(X ) h + σ (X )(Bk Bh) (3.47) t+h − t t k t t+h − t k=1 d t+h s + (σ σ )(X ) dBl dBk + h3/2 . l ∇ k t r s O k,l=1 t t Here the ﬁrst term on the right hand side has strong L2 order (h), the second term O (h1/2), and the third term (h). Taking into account only the ﬁrst two terms leads to O O the Euler-Maruyama scheme with step size h, whereas taking into account all terms up to order (h) yields the Milstein scheme: O Euler-Maruyama scheme with step size h • d Xh Xh = b(Xh) h + σ (Xh)(Bk Bh) (t =0, h, 2h, 3h, . . .) t+h − t t k t t+h − t k=1 Milstein scheme with step size h • d d t+h s h h h h k h h l k Xt+h Xt = b(Xt ) h + σk(Xt )(Bt+h Bt )+ (σl σk)(Xt ) dBr dBs − − ∇ t t k=1 k,l=1

~~UniversityofBonn WinterSemester2012/2013 148CHAPTER 3. SDE I: STRONG SOLUTIONS AND APPROXIMATION SCHEMES~~

The Euler and Milstein scheme provide approximations to the solution of the SDE (3.40) that are deﬁned for integer multiples t of the step size h. To analyse the approximation error it is convenient to extend the deﬁnition of the approximation schemes to all t 0 by considering the delay stochastic differential equations ≥

h h h k dXs = b(X s ) ds + σk(X s ) dBs , (3.48) ⌊ ⌋h ⌊ ⌋h k s h h h h l k dXs = b(X s h ) ds + σk(X s h )+(σl σk)(X s h ) dBr dBs (3.49) ⌊ ⌋ ⌊ ⌋ ∇ ⌊ ⌋ s h k,l ⌊ ⌋ respectively, where s := max t hZ : t s ⌊ ⌋h { ∈ ≤ } denotes the next discretization time below s. Notice that indeed, the Euler and Milstein scheme with step size h are obtained by evaluating the solutions of (3.48) and (3.49) respectively at t = kh with k Z. ∈

~~Strong convergence order~~

Fix a RN , let X be a solution of (3.40) with initial condition X = a, and let Xh be ∈ 0 a corresponding Euler or Milstein approximation satisfying (3.48, (3.49) respectively h with initial condition X0 = a.

~~Theorem 3.20 (Strong order for Euler and Milstein scheme). Let t [0, ). ∈ ∞~~

1) Suppose that the coefﬁcients b and σk are Lipschitz continuous. Then the Euler- Maruyama approximation on the time interval [0, t] has strong L2 order 1/2 in the following sense: h 1/2 sup Xs Xs = (h ). s t − O ≤ 3 2) If, moreover, the coefﬁcients b and σk are C with bounded derivatives then the Milstein approximation on the time interval [0, t] has strong L2 order 1, i.e.,

~~Xh X = (h). t − t O Stochastic Analysis Andreas Eberle 3.4. STOCHASTIC TAYLOR EXPANSIONS AND NUMERICAL METHODS 149~~

The assumptionson the coefﬁcients in the theorem are not optimal and can be weakened. However, it is well-known that even in the deterministic case a local Lipschitz condition is not sufﬁcient to guarantee convergence of the Euler approximations. The iterated integral in the Milstein scheme can be approximated by a Fourier expansion in such a way that the strong order (h) still holds, cf. Kloeden and Platen [?]XXX O Proof. For notational simplicity, we only prove the theorem in the one-dimensional case. The proof in higher dimensions is analogous.

~~h 1) By (3.48) and since X0 = X0, the difference of the Euler approximation and the solution of the SDE satisﬁes the equation~~

~~t t h h h Xt Xt = b(X s ) b(Xs) ds + σ(X s ) σ(Xs) dBs. − ⌊ ⌋h − ⌊ ⌋h − 0 0 This enables us to estimate the mean square error~~

~~h h 2 ε¯t := E sup Xs Xs . s t − ≤ By the Cauchy-Schwarz inequality and by Doob’s L2 inequality,~~

t t h h 2 h 2 ε¯t 2t E b(X s ) b(Xs) ds + 8 E σ(X s ) σ(Xs) ds ≤ ⌊ ⌋h − ⌊ ⌋h − 0 0 t 2 h 2 (2t + 8) L E X s Xs ds (3.50) ≤ ⌊ ⌋h − 0 t (4t + 16) L2 ε¯h ds + C h , ≤ s t 0 where t C is an increasing real-valued function, and L is a joint Lipschitz constant → t for b and σ. Here, we have used that by the triangle inequality,

h 2 h h 2 h 2 E X s Xs 2 E X s Xs + 2 E Xs Xs , ⌊ ⌋h − ≤ ⌊ ⌋h − − and the ﬁrst term representing the additional error by the time discretization on the interval [ s , s + h] is of order O(h) uniformly on ﬁnite time intervals by Corollary ⌊ ⌋h ⌊ ⌋h 3.19. By (3.50) and Gronwall’s inequality, we conclude that

ε¯h (4t + 16)L2C exp (4t + 16)L2t h, t ≤ t UniversityofBonn WinterSemester2012/2013 150CHAPTER 3. SDE I: STRONG SOLUTIONS AND APPROXIMATION SCHEMES and hence ε¯h = O(√h) for any t (0, ). This proves the assertion for the Euler t ∈ ∞ scheme. 2) To prove the assertion for the Milstein scheme we have to argue more carefully. By (3.49), the difference of the Milstein approximation and the solution of the SDE satisﬁes the equation

t h h Xt Xt = b(Xs) b(X s ) ds (3.51) − − ⌊ ⌋h 0 t h h + σ(Xs) σ(X s ) (σσ′)(X s )(Bs B s h ) dBs − ⌊ ⌋h − ⌊ ⌋h − ⌊ ⌋ 0 = It + IIt + IIIt + IVt where t I = b(X ) b(Xh) ds , t s − s 0 t II = σ(X ) σ(Xh) dB , t s − s s 0 t h h IIIt = b(Xs ) b(X s ) ds , − ⌊ ⌋h 0 t h h h IVt = σ(Xs ) σ(X s ) (σσ′)(X s )(Bs B s h ) dBs . − ⌊ ⌋h − ⌊ ⌋h − ⌊ ⌋ 0 Here terms I and II describe the continuation of the error made in previous time, whereas III and IV correspond to an additional error by the time discretization on the interval [ s ,s]. We are now going to bound the second moments of I, II, III ⌊ ⌋h and IV separately. Similarly as in the Euler case, we obtain by the Cauchy-Schwarz inequality and by Doob’s L2 inequality:

t t 2 h 2 2 h E It t E b(Xs) b(Xs ) ds L t εs ds, and (3.52) ≤ 0 − ≤ 0 t t 2 E II2 4t E σ(X ) σ(Xh) ds 4L2t εh ds, (3.53) t ≤ s − s ≤ s 0 0 where 2 εh := E Xh X . s s − s We are now going to prove that III andIV are of strong L2 order (h) uniformly on O ﬁnite time intervals. The proof can then be completed similarly as in the Euler case.

~~Stochastic Analysis Andreas Eberle 3.4. STOCHASTIC TAYLOR EXPANSIONS AND NUMERICAL METHODS 151~~

~~To bound IV we note that by (3.49) and Itô’s formula,~~

s h h h h σ(Xs ) σ(X s h ) σ(X s h )σ′(Xr ) dBr − ⌊ ⌋ − s ⌊ ⌋ ⌊ ⌋h s h h 1 h h 2 h = b(X s h )σ′(Xr )+ σ(X s h )+(σσ′)(X s h )(Br B s h ) σ′′(Xr ) dr. s ⌊ ⌋ 2 ⌊ ⌋ ⌊ ⌋ − ⌊ ⌋ ⌊ ⌋h Since all the coefﬁcients and their derivatives up to order 3 are assumed to be bounded, and, in particular, σ′ is Lipschitz continuous, we can conclude that

h h h σ(Xs ) σ(X s ) (σσ′)(X s )(Bs B s h ) h h ⌊ ⌋ − ⌊ ⌋ − ⌊ ⌋ s− 2 2 C1 (1 + Bs B s h + C2 1+ Br B r h dr ≤ − ⌊ ⌋ s − ⌊ ⌋ ⌊ ⌋h with constants C1 and C2 that do not depend on s. Hence

~~E IV 2 t ≤ LETZTER SCHRITT STIMMT NOCH NICHT GANZ !~~

~~UniversityofBonn WinterSemester2012/2013 Chapter 4~~

~~SDE II: Transformations and weak solutions~~

~~Let U Rn be an open set. We consider a stochastic differential equation of the form ⊆~~

dXt = b(t, Xt) dt + σ(t, Xt) dBt (4.1) with a d-dimensional Brownian motion (B ) and measurable coefficients b : [0, ) t ∞ × n n d U R and σ : [0, ) U R × . In applications one is often not interested in the → ∞ × → random variables X : Ω R themselves but only in their joint distribution. In that t → case, it is usually irrelevant w.r.t. which Brownian motion (Bt) the SDE (4.1) is satisfied. Therefore, we can “solve” the SDE in a very different way: Instead of constructing the solution from a given Brownian motion, we first construct a stochastic process (Xt,P ) by different types of transformations or approximations, and then we verify that the process satisfies (4.1) w.r.t. some Brownian motion (Bt) that is usually defined through (4.1).

~~Deﬁnition. A weak solution of the stochastic differential equation (4.1) is given by~~

~~(i) a “setup” consisting of a probability space (Ω, ,P ), a ﬁltration ( t)t 0 on A F ≥ (Ω, ) and an ( ) Brownian motion B : Ω Rd w.r.t. P , A Ft t → 152 153~~

~~(ii) a continuous ( ) adapted stochastic process (X ) where S is an (F ) stopping Ft t t~~

It is called a strong solution w.r.t. the given setup if and only if (Xt) is adapted w.r.t. the ﬁltration B,P generated by the Brownian motion and the initial condition. σ t ,X0 t 0 F ≥ Note that the concept of a weak solution of an SDE is not related to the analytic concept of a weak solution of a PDE!

Remark. A process (Xt)t 0 is a strong solution up to S < w.r.t. a given setup if and ≥ ∞ only if there exists a measurable map F : R Rn C R , Rd Rn, (t, x ,y) + × × + → 0 → n Ft(x0,y), such that the process (Ft)t 0 is adapted w.r.t. the ﬁltration (R ) t, t = ≥ B ⊗ B B σ(y y(s):0 s t), and → ≤ ≤ X = F (X , B) for any t 0 t t 0 ≥ holds almost surely. Hence strong solutions are (almost surely) functions of the given Brownian motion and the initial value!

There are SDE that haveweak but no strong solutions. An example is given in Section4.1. The deﬁnition of weak and strong solutions can be generalized to other types of SDE including in particular functional equations of the form

~~dXt = bt(X) dt + σt(X) dBt~~

~~n where (bt) and (σt) are (Bt) adapted stochastic processes deﬁned on C(R+, R ), as well as SDE driven by Poisson point processes, cf Chapter 3.~~

~~Different types of transformations of a stochastic process (Xt,P ) are useful for constructing weak solutions. These include:~~

Random time changes: (Xt)t 0 (XTa )a 0 where (Ta)a 0 is an increasing stochas- • ≥ → ≥ ≥ tic process on R such that T is a stopping time for any a 0. + a ≥ Transformations of the paths in space: These include for example coordinate changes • UniversityofBonn WinterSemester2012/2013 154 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS

(X ) (ϕ(X )), random translations (X ) (X +H ) where (H ) is another adapted t → t t → t t t process, and, more generally, a transformation that maps (Xt) to the strong solution (Yt) of an SDE driven by (Xt).

Change of measure: Here the random variables X are kept ﬁxed but the underlying • t probability measure P is replaced by a new measure P such that both measures are mutually absolutely continuous on each of the σ-algebras t, t R+ (but usually not on F ∈ ). F∞ In this chapter we study these transformations as well as relations between them. For identifying the transformed processes, the Lévy characterizations in Section 4.1 play a crucial rôle.

~~4.1 Lévy characterizations~~

~~Let (Ω, ,P, ( )) be a given ﬁltered probability space. We ﬁrst note that Lévy pro- A Ft cesses can be characterized by their exponential martingales:~~

Lemma 4.1. Let ψ : Rd C be a given function. An ( ) adapted càdlàg process → Ft X : Ω Rd is an ( ) Lévy process with characteristic exponent ψ if and only if the t → Ft complex-valued processes

Zp := exp ip X + tψ(p) , t 0, t t ≥ are ( ) martingales, or, equivalently, local ( ) martingales for any p Rd. Ft Ft ∈ Proof. By Corollary 1.3, the processes Zp are martingales if X is a Lévy process with characteristic exponent ψ. Conversely, suppose that Zp is a local martingale for any p Rd. Then, since these processes are uniformly bounded on ﬁnite time intervals, ∈ they are martingales. Hence for 0 s t and p Rd, ≤ ≤ ∈ E exp ip (X X ) = exp( (t s)ψ(p)), t − s Fs − − which implies that X X is independent of with characteristic function equal to t − s Fs exp( (t s)ψ). − −

~~Stochastic Analysis Andreas Eberle 4.1. LÉVY CHARACTERIZATIONS 155~~

Exercise (Lévy characterization of Poisson point processes). Let (S, , ν) be a σ- S ﬁnite measure space. Suppose that (Nt)t 0 on (Ω, ,P ) is an ( t) adapted process ≥ A F + taking values in the space Mc (S) consisting of all counting measures on S. Prove that the following statements are equivalent:

~~(i) (Nt) is a Poisson point processes with intensity measure ν. (ii) For any function f 1(S, , ν), the real valued process ∈L S~~

~~N f = f(y) N (dy), t 0, t t ≥ 1 is a compound Poisson process with jump intensity measure f − . ◦ (iii) For any function f 1(S, , ν), the complex valued process ∈L S~~

M f = exp(iN f + tψ(f)), t 0, ψ(f) = 1 eif dν, t t ≥ − is a local ( ) martingale. Ft Show that the statements are also equivalent if only elementary functions f L1(S, , ν) ∈ S are considered.

~~Lévy’s characterization of Brownian motion~~

d By Lemma 4.1, an R -valued process (Xt) is a Brownian motion if and only if the processes exp ip X + t p 2/2 are local martingales for all p Rd. This can be t | | ∈ Rd applied to prove the remarkable fact that any continuous valued martingale with the right covariations is a Brownian motion:

Theorem 4.2 (P. Lévy 1948). Suppose that M 1,...,M d are continuous local ( ) mar- Ft tingales with [M k, M l] = δ t P -a.s. for any t 0. t kl ≥ Then M =(M 1,...,M d) is a d-dimensional Brownian motion.

~~The following proof is due to Kunita and Watanabe (1967):~~

~~UniversityofBonn WinterSemester2012/2013 156 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS~~

Proof. Fix p Rd and let Φ := exp(ip M ). By Itô’s formula, ∈ t t 1 d dΦ = ip Φ dM Φ p p d[M k, M l] t t t − 2 t k l t k,l=1 1 = ip Φ dM Φ p 2 dt. t t − 2 t | | Since the ﬁrst term on the right hand side is a local martingale increment, the product rule shows that the process Φ exp( p 2 t/2) is a local martingale. Hence by Lemma t | | 4.1, M is a Brownian motion.

~~Lévy’s characterization of Brownian motion has a lot of remarkable direct consequences.~~

~~Example (Random orthogonal transformations). Suppose that X : Ω Rn is a t → solution of an SDE~~

dXt = Ot dBt, X0 = x0, (4.2) w.r.t. a d-dimensional Brownian motion (Bt), a product-measurable adapted process n d n (t, ω) O (ω) taking values in R × , and an initial vale x R . We verify that X is → t 0 ∈ an n-dimensional Brownian motion provided

~~O (ω) O (ω)T = I for any t 0, almost surely. (4.3) t t n ≥ Indeed, by (4.2) and (4.3), the components~~

~~d t i i ik k Xt = x0 + Os dBs 0 k=1 are continuous local martingales with covariations~~

~~i j ik jl k l ik jk [X ,X ] = O O d[B , B ] = O O dt = δij dt. k,l k Applications include inﬁnitesimal random rotations (n = d) and random orthogonal projections (n~~

Example (Bessel process). We derive an SDE for the radial component R = B of t | t| Brownian motion in Rd. The function r(x)= x is smooth on Rd 0 with r(x)= | | \{ } ∇ Stochastic Analysis Andreas Eberle 4.1. LÉVY CHARACTERIZATIONS 157

1 e (x), and ∆r(x)=(d 1) x − where e (x) = x/ x . Applying Itô’s formula to r − | | r | | d functions r C∞(R ), ε> 0, with r (x)= r(x) for x ε yields ε ∈ ε | | ≥ d 1 dRt = er(Bt) dBt + − dt for any t < T0 2Rt where T0 is the first hitting time of 0 for (Bt). By the last example, the process t W := e (B ) dB , t 0, t r s s ≥ 0 is a one-dimensional Brownian motion defined for all times (the value of er at 0 being irrelevant for the stochastic integral). Hence (Bt) is a weak solution of the SDE d 1 dRt = dWt + − dt (4.4) 2Rt up to the first hitting time of 0. The equation (4.4) makes sense for any particular d R ∈ and is called the Bessel equation. Much more on Bessel processes can be found in Revuz and Yor [33] and other works by M. Yor.

Exercise. a) Let (Xt)0 t<ζ be a solution of the Bessel equation ≤ d 1 dXt = − dt + dBt, X0 = x0, − 2Xt where (Bt)t 0 is a standard Brownian motion and d is a real constant. ≥ i) Find a non-constant function u : R R such that u(X ) is a local martingale up → t to the ﬁrst hitting time of 0.

ii) Compute the ruin probabilities P [T < T ] for a, b R with x [a, b] . a b ∈ + 0 ∈ iii) Proceeding similarly, determine the mean exit time E[T ], where T = min T , T . { a b} b) Now let (Xt)t 0 be a compound Poisson process with X0 = 0 and jump intensity ≥ measure ν = N(m, 1), m> 0.

~~i) Determine λ R such that exp(λX ) is a local martingale up to T . ∈ t 0 ii) Prove that for a< 0,~~

~~P [Ta < ] = lim P [Ta < Tb] exp(ma/2). b ∞ →∞ ≤ Why is it not as easy as aboveto computethe ruin probability P [Ta < Tb] exactly ?~~

~~UniversityofBonn WinterSemester2012/2013 158 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS~~

~~The next application of Lévy’s characterization of Brownian motion shows that there are SDE that have weak but no strong solutions.~~

~~Example (Tanaka’s example. Weak vs. strong solutions). Consider the one dimensional SDE~~

~~dXt = sgn(Xt) dBt (4.5)~~

~~+1 for x 0, where (Bt) is a Brownian motion and sgn(x) := ≥ . Note the unusual  1 for x< 0 − convention sgn(0) = 1 that is used below. We prove the following statements: ~~

~~1) X is a weak solution of (4.5) on (Ω, ,P, ( )) if and only if X is an ( ) Brow- A Ft Ft nian motion.~~

~~2) If X is a weak solution w.r.t. a setup (Ω, ,P, ( ), (B )) then for any t 0, the A Ft t ≥ X ,P process (Bs)s t is measurable w.r.t. the σ-algebra generated by 0 and | | . ≤ F F~~

~~3) There is no strong solution to (4.5).~~

The proof of 1) is again a consequence of the ﬁrst example above: If X is a weak solution then X is a Brownian motion by Lévy’s characterization. Conversely, if X is an ( ) Brownian motion then the process Ft

~~t Bt := sgn(Xs) dXs 0 is a Brownian motion as well, and, by the “associative law”,~~

~~t t sgn(X ) dB = sgn(X )2 dX = X X , s s s s t − 0 0 0 i.e., X is a weak solution to (4.5).~~

~~For proving 2) , we approximate r(x) = x by symmetric and concave functions r | | ε ∈ Stochastic Analysis Andreas Eberle 4.1. LÉVY CHARACTERIZATIONS 159~~

~~C∞(R) satisfying r (x) = x for x ε. Then the associative law, the Itô isometry, ε | | | | ≥ and Itô’s formula imply t t~~

Bt B0 = sgn(Xs) dXs = lim ϕε′′(Xs) dXs − ε 0 0 ↓ 0 1 t = lim ϕε(Xt) ϕε(X0) ϕε′′(Xs) ds ε 0 − − 2 ↓ 0 1 t = lim ϕε( Xt ) ϕε( X0 ) ϕε′′( Xs ) ds ε 0 | | − | | − 2 | | ↓ 0 with almost sure convergence along a subsequence ε 0. n ↓ Finally by 2), if X would be a strong solution w.r.t. a Brownian motion B then Xt X ,P would also be measurable w.r.t. the σ-algebra generated by and | | . This leads F0 Ft to a contradiction as one can verify that the event X 0 is not measurable w.r.t. this { t ≥ } σ-algebra for a Brownian motion (Xt).

~~Lévy characterization of Lévy processes~~

~~Lévy’s characterization has a natural extension to discontinuous martingales.~~

d d d Theorem 4.3. Let a R × , b R, and let ν be a σ-ﬁnite measure on R 0 ∈ ∈ \ { } satisfying ( y y 2) ν(dy) < . If X1,...,Xd : Ω R are càdlàg stochastic | |∧| | ∞ t t → processes such that (i) M k := Xk bkt is a local ( ) martingale for any k 1,...,d , t t − Ft ∈{ } (ii) [Xk,Xl]c = akl t for any k,l 1,...,d , and t ∈{ }

~~(iii) E s (r,t] IB(∆Xs) r = ν(B) (t r) almost surely ∈ F − for any 0 r t and for any B (Rd 0 ), ≤ ≤ ∈ B \{ } 1 d then Xt =(Xt ,...,Xt ) is a Lévy process with characteristic exponent~~

1 ip y ψ(p) = p ap ip b + (1 e + ip y) ν(dy). (4.6) 2 − − For proving the theorem, we assume without proof that a local martingale is a semimartingale even if it is not strict, and that the stochastic integral of a bounded adapted left-continuous integrand w.r.t. a local martingale is again a local martingale, cf. e.g. [32].

~~UniversityofBonn WinterSemester2012/2013 160 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS~~

~~Proof of Theorem 4.3. We ﬁrst remark that (iii) implies~~

t E Gs f(∆Xs) r = E Gs f(y) ν(dy) ds r , a.s. for 0 r t F r F ≤ ≤ s (r,t] ∈ (4.7) for any bounded left-continuous adapted process G, and for any measurable function f : Rd 0 C satisfying f(y) const. ( y y 2). This can be verified \ { } → | | ≤ | |∧| | by first considering elementary functions of type f(y) = ci IBi (y) and Gs(ω) = A (ω) I (s) with c R, B (Rd 0 ), 0 t < t < < t , and A i (ti,ti+1] i ∈ i ∈ B \{ } ≤ 0 1 n i bounded and -measurable. Fti Now fix p Rd, and consider the semimartingale ∈ Z = exp(ip X + tψ(p)) = exp(ip M + t(ψ(p)+ ip b)). t t t

~~k l c k l c kl Noting that [M , M ]t = [X ,X ]t = a t by (ii), Itô’s formula yields~~

t t 1 kl Zt = 1+ Z ip dM + Z (ψ(p)+ ip b pkpla ) dt (4.8) 0 − 0 − − 2 k,l ip ∆X + Z e 1 ip ∆X . − − − (0,t] ip y 2 By (4.7) and since e 1 ip y const. ( y y ), the series on the right hand | − − | ≤ | |∧| | side of (4.8) can be decomposed into a martingale and the ﬁnite variation process

~~t ip y At = Zs (e 1 ip y) ν(dy) ds − − − 0 Therefore, by (4.8) and (4.6), (Z ) is a martingale for any p Rd. The assertion now t ∈ follows again by Lemma 4.1.~~

~~An interesting consequence of Theorem 4.3 is that a Brownian motion B and a Lévy process without diffusion part w.r.t. the same ﬁltration are always independent, because [Bk,Xl]=0 for any k,l.~~

Exercise (Independence of Brownian motion and Lévy processes). Suppose that ′ B : Ω Rd and X : Ω Rd are a Brownian motion and a Lévy process with t → t → ip y characteristic exponent ψ (p) = ip b + (1 e + ip y) ν(dy) deﬁned on the X − − Stochastic Analysis Andreas Eberle 4.1. LÉVY CHARACTERIZATIONS 161 same ﬁltered probability space (Ω, ,P, ( )). Assuming that ( y y 2) ν(dy) < , A Ft | |∧| | ∞ Rd d′ prove that (Bt,Xt) is a Lévy process on × with characteristic exponent

~~1 2 d d′ ψ(p, q) = p d + ψ (q), p R , q R . 2| |R X ∈ ∈ Hence conclude that B and X are independent.~~

~~Lévy characterization of weak solutions~~

~~Lévy’s characterization of Brownian motion can be extended to solutions of stochastic differential equations of type~~

dXt = b(t, Xt) dt + σ(t, Xt) dBt (4.9) driven by a d-dimensional Brownian motion (Bt). As a consequence, one can show that a process is a weak solution of (4.9) if and only if it solves the corresponding martingale d d d d d problem. We assume that the coefﬁcients b : R R R and σ : R R R × + × → + × → are measurable and locally bounded, i.e., bounded on [0, t] K for any t 0 and any × ≥ compact set K Rd. Let ⊂ 1 d ∂2 d ∂ = aij(t, x) + bi(t, x) (4.10) L 2 ∂xi∂xj ∂xi i,j=1 i=1 denote the corresponding generator where a(t, x)= σ(t, x)σ(t, x)T is a symmetric d d × matrix for any t and x.

Theorem 4.4 (Weak solutions and the martingale problem). If the matrix σ(t, x) is 1 invertible for any t and x, and (t, x) σ(t, x)− is a locally bounded function on → R Rd, then the following statements are equivalent: + × (i) (X ) is a weak solution of (4.9) on the setup (Ω, ,P, ( ), (B )). t A Ft t (ii) The processes M i := Xi Xi t bi(s,X ) ds, 1 i d, are continuous local t t − 0 − 0 s ≤ ≤ ( P ) martingales with covariations Ft t [M i, M j] = aij(s,X ) ds P -a.s. for any t 0. (4.11) t s ≥ 0

~~UniversityofBonn WinterSemester2012/2013 162 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS~~

~~(iii) The processes M [f] := f(X ) f(X ) t( f)(s,X ) ds, f C2(Rd), are t t − 0 − 0 L s ∈ continuous local ( P ) martingales. Ft~~

(iv) The processes Mˆ [f] := f(t, X ) f(0,X ) t ∂f + f)(s,X ) ds, t t − 0 − 0 ∂t L s f C2(R Rd , are continuous local ( P ) martingales. ∈ + × Ft Proof. (i) (iv) is a consequence of the Itô-Doeblin formula, cf. equation (2.57) above. ⇒ (iv) (iii) trivially holds. ⇒ (iii) (ii) follows by choosing for f polynomials of degree 2. Indeed, for f(x)= xi, ⇒ ≥ we obtain f = bi, hence L t M i = Xi X0 bi(s,X ) ds = M [f] (4.12) t t − t − s t 0 is a local martingale by (iii). Moreover, if f(x)= xixj then f = aij + xibj + xjbi by L the symmetry of a, and hence

~~t XiXj XiXj = M [f] + aij(s,X )+ Xi bj(s,X )+ Xj bi(s,X ) ds. (4.13) t t − 0 0 t s s s s s 0 On the other hand, by the product rule and (4.12),~~

t t i j i j i j j i i j Xt Xt X0X0 = Xs dXs + Xs dXs + [X ,X ]t (4.14) − 0 0 t i j j i i j = Nt + Xs b (s,Xs)+ Xs b (s,Xs) ds + [X ,X ]t 0 with a continuous local martingale N. Comparing (4.13) and (4.14) we obtain

~~t i j i j ij [M , M ]t = [X ,X ]t = a (s,Xs) ds 0 since a continuous local martingale of ﬁnite variation is constant.~~

~~(ii) (i) is a consequence of Lévy’s characterization of Brownian motion: If (ii) holds ⇒ then~~

~~dXt = dMt + b(t, Xt) dt = σ(t, Xt) dBt + b(t, Xt) dt~~

~~Stochastic Analysis Andreas Eberle 4.2. RANDOM TIME CHANGE 163~~

1 d t 1 where Mt = Mt ,...,Mt and Bt := 0 σ(s,Xs)− dMs are continuous local martin- Rd 1 gales with values in because σ− is locally bounded. To identify B as a Brownian motion it sufﬁces to note that t k l 1 1 i j [B , B ]t = σki− σlj− (s,Xs) d[M , M ] 0 i,j t 1 1 T = σ− a(σ− ) kl (s,Xs) ds = δkl t 0 for any k,l =1,...,d by (4.11).

Remark (Degenerate case). If σ(t, x) is degenerate then a corresponding assertion still holds. However, in this case the Brownian motion (Bt) only exists on an extension of the probability space (Ω, ,P, ( )). The reason is that in the degenerate case, the A Ft Brownian motion can not be recovered directly from the solution (Xt) as in the proof above, see [34] for details. The martingale problem formulation of weak solutions is powerful in many respects: It is stable under weak convergence and therefore well suited for approximation arguments, it carries over to more general state spaces (including for example Riemannian manifolds, Banach spaces, spaces of measures), and, of course, it provides a direct link to the theory of Markov processes. Do not miss to have a look at the classics by Stroock and Varadhan [36] and by Ethier and Kurtz [15] for much more on the martingale problem and its applications to Markov processes.

~~4.2 Random time change~~

Random time change is already central to the work of Doeblin from 1940 that has been discovered only recently [10]. Independently, Dambis and Dubins-Schwarz have developed a theory of random time changes for semimartingales in the 1960s [23], [33]. In this section we study random time changes with a focus on applications to SDE, in particular, but not exclusively, in dimension one. Throughout this section we ﬁx a right-continuous ﬁltration ( ) such that = P Ft Ft F for any t 0. Right-continuity is required to ensure that the time transformation is ≥ given by ( ) stopping times. Ft UniversityofBonn WinterSemester2012/2013 164 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS

~~Continuous local martingales as time-changed Brownian motions~~

Let (Mt)t 0 be a continuous local ( t) martingale w.r.t. the underlying probability mea- ≥ F sure P such that M0 =0. Our aim is to show that Mt can be represented as B[M]t with a one-dimensional Brownian motion (Ba). For this purpose, we consider the random time substitution a T where T = inf u :[M] > a is the ﬁrst passage time to the → a a { u } level u. Note that a T is the right inverse of the quadratic variation t [M] , i.e., → a → t [M] = a on T < , and, TA { a ∞} T = inf u :[M] > [M] = sup u :[M] = [M] [M]t { u t} { u t} 1 by continuity of [M]. If [M] is strictly increasing then T = [M]− . By right-continuity of ( ), T is an ( ) stopping time for any a 0. Ft a Ft ≥ Theorem 4.5 (Dambis, Dubins-Schwarz). If M is a continuous local ( ) martingale Ft with [M] = almost surely then the time-changed process Ba := MTa , a 0, is an ∞ ∞ ≥ ( ) Brownian motion, and FTa M = B for any t 0, almost surely. (4.15) t [M]t ≥ The proof is again based on Lévy’s characterization.

~~Proof. 1) Weﬁrstnotethat B[M]t = Mt almost surely. Indeed, by deﬁnition, B[M]t = M . It remains to verify that M is almost surely constant on the interval T[M]t~~

~~[t, T[M]t ]. This holds true since the quadratic variation [M] is constant on this interval, cf. the exercise below.~~

~~2) Next, we verify that Ba = MTa is almost surely continuous. Right-continuity holds since M and T are both right-continuous. To prove left-continuity note that for a> 0,~~

~~lim MTa−ε = MTa− for any a 0 ε 0 ↓ ≥~~

~~by continuity of M. It remains to show MTa = MTa almost surely. This again − holds true by the exercise below, because Ta and Ta are stopping times, and −~~

~~[M]Ta− = lim [M]Ta−ε = lim (a ε) = a = [M]Ta ε 0 ε 0 ↓ ↓ − by continuity of [M].~~

~~Stochastic Analysis Andreas Eberle 4.2. RANDOM TIME CHANGE 165~~

3) We nowshowthat (B ) is a square-integrable ( ) martingale. Since the random a FTa variables T are ( ) stopping times, (B ) is ( ) adapted. Moreover, for any a, a Ft a FTa Ta the stopped process Mt = Mt Ta is a continuous local martingale with ∧

~~Ta E [M ] = E [M]Ta = a < . ∞ ∞ Hence M Ta is inM 2 [0, ] , and c ∞~~

~~2 2 Ta 2 E[Ba] = E[MTa ] = E[(M ) ] = a for any a 0. ∞ ≥~~

~~This shows that (Ba) is square-integrable, and, moreover,~~

E[B ] = E[M ] = M = B for any 0 r a a|FTr Ta |FTr Tr r ≤ ≤ by the Optional Sampling Theorem applied to M Ta . Finally, we note that [B] = B = a almost surely. Indeed, by the Optional Sampling a a Theorem applied to the martingale (M Ta )2 [M Ta ], we have − E B2 B2 = E M 2 M 2 a − r |FTr Ta − Tr |FTr = E[M]T [M]T T = a r for 0 r a. a − r |F r − ≤ ≤ Hence B2 a is a martingale, and thus by continuity, [B] = B = a almost surely. a − a a We have shown that (B ) is a continuous square-integrable ( ) martingale with a FTa [B]a = a almost surely. Hence B is a Brownian motion by Lévy’s characterization.

Remark. The assumption [M] = in Theorem 4.5 ensures Ta < almost surely. ∞ ∞ ∞ If the assumption is violated then M can still be represented in the form (4.15) with a Brownian motion B. However, in this case, B is only deﬁned on an extended probability space and can not be obtained as a time-change of M for all times, cf. e.g. [33].

Exercise. Let M be a continuous local ( ) martingale, and let S and T be ( ) stop- Ft Ft ping times such that S T . Prove that if [M] = [M] < almost surely, then M ≤ S T ∞ is almost surely constant on the stochastic interval [S, T ]. Use this fact to complete the missing step in the proof above.

~~We now consider several applications of Theorem 4.5. Let (Wt)t 0 be a Brownian ≥ motion with values in Rd w.r.t. the underlying probability measure P .~~

~~UniversityofBonn WinterSemester2012/2013 166 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS~~

~~Time-change representations of stochastic integrals~~

By Theorem 4.5 and the remark below the theorem, stochastic integrals w.r.t. Brownian motions are time-changed Brownian motions. For any integrand G 2 (R , Rd), ∈ La,loc + there exists a one-dimensional Brownian motion B, possibly deﬁned on an enlarged probability space, such that almost surely,

t t for any Gs dWs = BR Gs 2 ds t 0. 0 | | ≥ 0 Example (Gaussian martingales). If G is a deterministic function then the stochastic integral is a Gaussian process that is obtained from the Brownian motion B by a deterministic time substitution. This case has already been studied in Section 8.3 in [13].

~~Doeblin [10] has developed a stochastic calculus based on time substitutions instead of Itô integrals. For example, an SDE in R1 of type~~

~~t t X X = σ(s,X ) dW + b(s,X ) ds t − 0 s s s 0 0 can be rephrased in the form~~

~~t t Xt X0 = BR σ(s,Xs)2 ds + b(s,Xs) ds − 0 0 with a Brownian motion B. The one-dimensional Itô-Doeblin formula then takes the form~~

~~t ∂f t ′ f(t, Xt) f(0,X0) = BR σ(s,Xs)2 f (s,Xs)2 ds + + f (s,Xs) ds − 0 ∂s L 0 1 2 with f = σ f ′′ + bf ′. L 2~~

~~Time substitution in stochastic differential equations~~

~~To see how time substitution can be used to construct weak solutions, we consider at ﬁrst an SDE of type~~

~~dYt = σ(Yt) dBt (4.16)~~

Stochastic Analysis Andreas Eberle 4.2. RANDOM TIME CHANGE 167 in R1 where σ : R (0, ) is a strictly positive continuous function. If Y is a weak → ∞ solution then by Theorem 4.5 and the remark below,

t 2 Yt = XAt with At = [Y ]t = σ(Yr) dr (4.17) 0 and a Brownian motion X. Note that A depends on Y , so at first glace (4.17) seems not 1 to be useful for solving the SDE (4.16). However, the inverse time substitution T = A− satisfies 1 1 1 T ′ = = = , A T σ(Y T )2 σ(X)2 ′ ◦ ◦ and hence a 1 T = du. a σ(X )2 0 u Therefore, we can construct a weak solution Y of (4.16) from a given Brownian motion 1 X by first computing T , then the inverse function A = T − , and finally setting Y = X A. More generally, the following result holds: ◦ Theorem 4.6. Suppose that (X ) on (Ω, ,P, ( )) is a weak solutionof an SDE of the a A Ft form

~~dXa = σ(Xa) dBa + b(Xa) da (4.18)~~

d d d d d with locally bounded measurable coefﬁcients b : R R and σ : R R × such → → 1 d that σ(x) is invertible for almost all x, and σ− is again locally bounded. Let ̺ : R → (0, ) be a measurable function such that almost surely, ∞ a Ta := ̺(Xu) du < a (0, ), and T = . (4.19) ∞ ∀ ∈ ∞ ∞ ∞ 0 Then the time-changed process deﬁned by

~~1 Yt := XAt , A := T − , is a weak solution of the SDE~~

~~σ b dYt = (Yt) dBt + (Yt) dt. (4.20) √̺ ̺ We only give a sketch of the proof of the theorem:~~

~~UniversityofBonn WinterSemester2012/2013 168 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS~~

Proof of 4.6. (Sketch). The process X is a solution of the martingale problem for the operator = 1 a (x) ∂2 + b(x) where a = σσT , i.e., L 2 ij ∂xi∂xj ∇ a M [f] = f(X ) F (X ) ( f)(X ) du a a − 0 − L u 0 is a local ( ) martingale for any f C2. Therefore, the time-changed process Fa ∈ At [f] MAt = f(XAt f(XA0 ) ( f)(Xu) du − − 0 L t = f(Y ) f(Y ) ( f)(Y )A′ dr t − 0 − L r r 0 is a local ( ) martingale. Noting that FAt 1 1 1 Ar′ = = = , T ′(Ar) ̺(XAr ) ̺(Yr) we see that w.r.t. the ﬁltration ( ), the process Y is a solution of the martingale FAt problem for the operator

~~1 1 a ∂2 b = = ij + . L ̺ L 2 ̺ ∂xi∂xj ̺ ∇ i,j a σ σT Since ̺ = ̺ ̺ , this implies that Y is a weak solution of (4.20).~~

~~In particular, the theorem shows that if X is a Brownian motion and condition (4.19) 1/2 holds then the time-changed process Y solves the SDE dY = ̺(Y )− dB.~~

~~Example (Non-uniqueness of weak solutions). Consider the one-dimensional SDE~~

dY = Y α dB ,Y =0, (4.21) t | t| t 0 with a one-dimensional Brownian motion (Bt) and α > 0. If α < 1/2 and x is a a Brownian motion with X0 = 0 then the time-change Ta = 0 ̺(Xu) du with ̺(x) = 2α x − satisﬁes | | a a 2α E[Ta] = E ̺(Xu) du = E[ Xu − ] du 0 0 | | a 2α α = E[ X − ] u− du < | 1| ∞ 0

~~Stochastic Analysis Andreas Eberle 4.2. RANDOM TIME CHANGE 169~~

1 for any a (0, ). Hence (4.19) holds, and therefore the process Y = X , A = T − , ∈ ∞ t At is a non-trivial weak solution of (4.21). On the other hand, Y 0 is also a weak t ≡ solution. Hence for α < 1/2, uniqueness in distribution of weak solutions fails. For α 1/2, the theorem is not applicable since Assumption (4.19) is violated. One can ≥ prove that in this case indeed, the trivial solution Y 0 is the unique weak solution. t ≡

~~Exercise (Brownian motion on the unit sphere). Let Yt = Bt/ Bt where (Bt)t 0 is a | | ≥ Brownian motion in Rn, n> 2. Prove that the time-changed process~~

t 1 2 Z = Y , T = A− with A = B − ds , a Ta t | s| 0 n 1 n is a diffusion taking values in the unit sphere S − = x R : x =1 with generator { ∈ | | } 2 1 ∂ f n 1 ∂f n 1 f(x) = ∆f(x) x x (x) − x (x), x S − . L 2 − i j ∂x ∂x − 2 i ∂x ∈ i,j i j i i One-dimensional SDE

~~By combining scale and time transformations, one can carry out a rather complete study of weak solutions for non-degenerate SDE of the form~~

dXt = σ(Xt) dBt + b(Xt) dt, X0 = x0, (4.22) on a real interval (α, β). We assume that the initial value X0 is contained in (α, β), and b, σ : (α, β) R are continuous functions such that σ(x) > 0 for any x (α, β). We → ∈ ﬁrst simplify (4.22) by a coordinate transformation Yt = s(Xt) where

s : (α, β) s(α),s(β) → 2 is C and satisﬁes s′(x) > 0 for all x. The scale function z y 2b(x) s(z) := exp 2 dx dy x − x σ(x) 0 0 1 2 has these properties and satisﬁes 2 σ s′′ + bs′ = 0. Hence by the Itô-Doeblin formula, the transformed process Yt = s(Xt) is a local martingale satisfying

~~dYt = (σs′)(Xt) dBt,~~

~~UniversityofBonn WinterSemester2012/2013 170 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS i.e., Y is a solution of the equation~~

~~dYt = σ(Yt) dBt,Y0 = s(x0), (4.23)~~

1 where σ := (σs′) s− . The SDE (4.23) is the original SDE in “natural scale”. It can ◦ be solved explicitly by a time change. By combining scale transformations and time change one obtains:

~~Theorem 4.7. The following statements are equivalent:~~

~~(i) The process (X ) on the setup (Ω, ,P, ( ), (B )) is a weak solution of (4.22) t t<ζ A Ft t deﬁned up to a stopping time ζ.~~

~~(ii) The process Yt = s(Xt), t<ζ, on the same setup is a weak solution of (4.23) up to ζ.~~

(iii) The process (Yt)s<ζ has a representation of the form Yt = BAt , where Bt is a 1 one-dimensional Brownian motion satisfying B0 = s(x0) and A = T − with r 2 Tr = ̺ Bu du, ̺(y) = 1/σ(y) . 0 Carrying out the details of the proof is left as an exercise. The measure m(dy) := ̺(y) dy is called the “speed measure” of the process Y although Y is moving faster if m is small. The generator of Y can be written in the form = 1 d d , and the L 2 dm dy generator of X is obtained from by coordinate transformation. For a much more L detailed discussion of one dimensional diffusions we refer to Section V.7 in [34]. Here we only note that 4.7 immediately implies existence and uniqueness of a maximal weak solution of (4.22):

~~Corollary 4.8. Under the regularity and non-degeneracy conditions on σ and b imposed above there exists a weak solution X of (4.22) deﬁned up to the ﬁrst exit time~~

ζ = inf t 0 : lim Xt a, b ≥ s t ∈{ } ↑ from the interval (α, β). Moreover, the distribution of any two weak solutions (Xt)t<ζ ¯ R and (Xt)t<ζ¯ on u>0 C([0,u), ) coincide. Stochastic Analysis Andreas Eberle 4.2. RANDOM TIME CHANGE 171

Remark. We have already seen above that uniqueness may fail if σ is degenerate. For example, the solution of the equationdY = Y α dB , Y = 0, is not unique in t | t| t 0 distribution for α (0, 1/2). ∈

Example (Bessel SDE). Suppose that (Rt)t<ζ is a maximal weak solution of the Bessel equation d 1 1 dRt = dWt + − dt, W BM(R ), 2Rt ∼ on the interval (α, β) = (0, ) with initial condition R = r (0, ) and the pa- ∞ 0 0 ∈ ∞ 1 d 1 rameter d R. The ODE s = s′′ + − s′ = 0 for the scale function has a strictly ∈ L 2 2r increasing solution 1 2 d 2 d r − for d =2, s(r) = −  for log r d =2 (More generally, cs + d is a strictly increasing solution for any c> 0 and d R).  ∈ Note that s is one-to-one from the interval (0, ) onto ∞ (0, ) for d< 2, ∞ (s(0),s( )) = ( , ) for d =2, ∞   −∞ ∞ ( , 0) for d> 2. −∞  By applying the scale transformation, we see that s(r ) s(a) P T R < T R = P T s(R) < T s(R) = 0 − b a s(b) s(a) s(b) s(a) − X for any a

1 for d 2, ≤ R R  P lim inf Rt =0 = P Ta < Tb = t ζ  ↑ a (0,r0) b (r0, )  ∈ ∈ ∞ 0 for d> 2,    1 for d 2, ≥ R R  P lim sup Rt = = P Tb < Ta = t ζ ∞  ↑ b (r0, ) a (0,r0)  ∈ ∞ ∈ 0 for d< 2.   UniversityofBonn WinterSemester2012/2013 172 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS

~~Note that d = 2 is the critical dimension in both cases. Rewriting the SDE in natural scale yields~~

1 ds(R) = σ s(R) dW with σ(y) = s′ s− (y) . y 2 2y In the critical case d = 2, s(r) = log r, σ(y) = e− , and hence ̺(y) = σ(y)− = e . 2y Thus the speed measure is m(dy)= e dy, and log Rt = BT −1(t), i.e., a Rt = exp BT −1(t) with Ta = exp 2Bu du 0 and a one-dimensional Brownian motion B.

~~4.3 Girsanov transformation~~

~~In Section 3.3-3.6 we study connections between two different ways of transforming a stochastic process (Y,P ):~~

~~1) Random transformations of the paths: For instance, mapping a Brownian motion~~

~~(Yt) to the solution (Xt) of s stochastic differential equation of type~~

~~dXt = b(t, Xt) dt + dYt (4.24)~~

~~corresponds to a random translation of the paths of (Yt):~~

~~t Xt(ω) = Yt(ω)+ Ht(ω) where Ht = b(Xs) ds. 0~~

2) Change of measure: Replace the underlying probability measure P by a modified probability measure Q such that P and Q are mutually absolutely continuous on for any t. Ft In this section we focus mainly on random transformations of Brownian motions and the corresponding changes of measure. To understand which kind of results we can expect in this case, we first look briefly at a simplified situation:

~~Stochastic Analysis Andreas Eberle 4.3. GIRSANOV TRANSFORMATION 173~~

~~Example (Translated Gaussian random variables in Rd). We consider the equation~~

X = b(X)+ Y,Y N(0,I ) w.r.t. P, (4.25) ∼ d for random variables X,Y : Ω Rd where b : Rd Rd is a “predictable” map → → in the sense that the i-th component bi(x) depends only on the ﬁrst i 1 components − i i 1 X ,...,X − of X. The predictability ensures in particular that the transformation deﬁned by (4.25) is invertible, with Y 1 = X1 b1, Y 2 = X2 b2(X1), Y 3 = X3 − − − 3 1 2 n n n 1 n 1 b (X ,X ), ... ,Y = X b (X ,...,X − ). − A random variable (X,P ) is a “weak” solution of the equation (4.25) if and only if Y := X b(X) is standard normally distributed w.r.t. P , i.e., if and only if the distribution − 1 P X− is absolutely continuous with density ◦ ∂(x b(x)) f P (x) = f P x b(x) det − X Y − ∂x d/2 x b(x) 2/2 = (2π)− e−| − | x b(x) b(x) 2/2 d = e −| | ϕ (x), where ϕd(x) denotes the standard normal density in Rd. Therefore we can conclude:

~~(X,P ) is a weak solution of (4.25) if and only if X N(0,I ) w.r.t. the unique proba- ∼ d bility measure Q on Rd satisfying P Q with ≪ dP = exp X b(X) b(X) 2/2 . (4.26) dQ −| |~~

~~In particular, we see that the distribution b of a weak solution of (4.25) is uniquely determined, and b satisﬁes~~

~~1 1 = P X− Q X− = N(0,I ) = b ◦ ≪ ◦ d 0 with relative density~~

~~db X b(X) b(x) 2/2 (X) = e −| | . d0~~

~~UniversityofBonn WinterSemester2012/2013 174 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS~~

The example can be extended to Gaussian measures on Hilbert spaces and to more general transformations, leading to the Cameron-Martin Theorem (cf. Theorem 4.17 below) and Ramer’s generalization [2]. Here, we study the more concrete situation where Y and X are replaced by a Brownian motion and a solution of the SDE (4.24) respectively. 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 ( ) be a ﬁltration on a measurable space (Ω, ), and ﬁx t (0, ). We consider Ft A 0 ∈ ∞ two probability measures P and Q on (Ω, ) that are mutually absolutely continuous A on the σ-algebra with relative density Ft0 dP Zt0 = > 0 Q-almost surely. dQ t0 F Then P and Q are also mutually absolutely continuous on each of the σ-algebras , Ft t t , with Q- and P -almost surely strictly positive relative densities ≤ 0 dP dQ 1 Zt = = EQ Zt0 t and = . dQ t F dP t Zt F F 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 càdlàg version of these martingales.

Lemma 4.9. 1) For any 0 s t t , and for any -measurable random vari- ≤ ≤ ≤ 0 Ft able X : Ω [0, ], → ∞ E [XZ ] E [XZ ] E [X ] = Q t|Fs = Q t|Fs P -a.s. and Q-a.s. (4.27) P |Fs E [Z ] Z Q t|Fs s

~~2) Suppose that (Mt)t t0 is an ( t) adapted càdlàg stochastic process. Then ≤ F~~

(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. ⇔ Stochastic Analysis Andreas Eberle 4.3. GIRSANOV TRANSFORMATION 175

Proof. 1) The right-hand side of (4.27) is -measurable. Moreover, for any A , Fs ∈Fs E [E [XZ ]/Z ; A] = E [E [XZ ]; A] P Q t|Fs s Q Q t|Fs = EQ[XZt ; A] = EQ[X ; A]. 2) (i) is a direct consequence of 1). (ii) By symmetry, it is enough to prove the implication " ". Hence suppose that ⇐ M Z is a local Q-martingale with localizing sequence (T ). We show that M Tn is a n P -martingale, i.e.,

~~EP [Mt Tn ; A] = EP [Ms Tn ; A] for any A s, 0 s t t0. (4.28) ∧ ∧ ∈F ≤ ≤ ≤ To verify (4.28), we ﬁrst note that~~

EP [Mt Tn ; A Tn s ] = EP [Ms Tn ; A Tn s ] (4.29) ∧ ∩{ ≤ } ∧ ∩{ ≤ } since t T = T = s T on T s . Moreover, one veriﬁes from the deﬁnition of ∧ n n ∧ n { n ≤ } the σ-algebra s Tn that for any A s, the event A Tn >s is contained in s Tn , F ∧ ∈F ∩{ } F ∧ and hence in t Tn . Therefore, F ∧

~~EP [Mt Tn ; A Tn >s ] = EQ[Mt Tn Zt Tn ; A Tn >s ] (4.30) ∧ ∩{ } ∧ ∧ ∩{ }~~

~~= 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 t Tn with relative density Zt Tn . (4.28) follows from (4.29) and (4.30). ≪ F ∧ ∧ If the probability measures P and Q are mutually absolutely continuous on the σ-algebra dP t, then the Q-martingale Zt = dQ of relative densities is actually an exponential F t martingale. Indeed, to obtain a correspondingF representation let t 1 Lt := dZs 0 Zs − denote the stochastic "logarithm" of Z. Note that (Lt)t t0 is a well-deﬁned local ≤ martingale w.r.t. Q since Q-a.s., (Zt) is càdlàg and strictly positive. Moreover, by the associative law,

dZt = Zt dLt, Z0 = 1, − so Zt is the stochastic exponential of the local Q-martingale (Lt): Z = L. t Et UniversityofBonn WinterSemester2012/2013 176 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS

~~Translated Brownian motions~~

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 d Brownian motion in R with X0 = 0 w.r.t. the probability measure Q and the ﬁltration ( ), and ﬁx t [0, ). Let Ft 0 ∈ ∞ t L = G dX , t 0, t s s ≥ 0 with G 2 R , Rd . Then [L] = t G 2 ds, and hence ∈La,loc + t 0 | s| t t 1 2 Zt = exp Gs dXs Gs ds (4.31) 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 t if and only if E [Z ]=1 (Exercise). In order to use Z for changing the ≤ 0 Q t0 t0 underlying probability measure on we have to assume the martingale property: Ft0

Assumption. (Zt)t t0 is a martingale w.r.t. Q. ≤ If the assumption holds then we can consider a probability measure P on with A dP = Zt0 Q-a.s. (4.32) dQ t0 F Note that P and Q are mutually absolutely continuous on for any t t with Ft ≤ 0 dP dQ 1 = Zt and = dQ t dP t Zt F F both P - and Q-almost surely. We are now ready to prove one of the most important results of stochastic analysis:

Theorem 4.10 (Maruyama 1954, Girsanov 1960). Suppose that X is a d-dimensional 2 R Rd Brownian motion w.r.t. Q and (Zt)t t0 is deﬁned by (4.31) with G a,loc( +, ). If ≤ ∈ L EQ[Zt0 ]=1 then the process

~~t B := X G ds, t t , t t − s ≤ 0 0 is a Brownian motion w.r.t. any probability measure P on satisfying (4.32). A Stochastic Analysis Andreas Eberle 4.3. GIRSANOV TRANSFORMATION 177~~

Rd Proof. By Lévy’s characterization, it sufﬁces to show that (Bt)t t0 is an -valued P - ≤ martingale with [Bi, Bj] = δ tP -almost surely for any i, j 1,...,d . Furthermore, t ij ∈{ } by Lemma 4.9, and since P and Q are mutually absolutely continuous on , this holds Ft0 i j true provided (BtZt)t t0 is a Q-martingale and [B , B ] = δijt 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(Bi Z) = Bi dZ + Z dBi + d[Bi,Z] (4.33) = BiZG dX + Z dXi Z Gidt + ZGi dt, − where we have used that

~~d[Bi,Z] = ZG d[Bi,X] = ZGi dt Q- a.s. The right-hand side of (4.33) 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 (4.31) is a Q-martingale, then X satisﬁes~~

dXt = Gt dt + dBt. with a P -Brownian motion B. It generalizes the Cameron-Martin Theorem to non- deterministic adapted translation t X (ω) X (ω) H (ω), H = G ds, t −→ t − t t s 0 of a Brownian motion X. Remark (Assumptions in Girsanov’s Theorem). 1) Absolute continuity and adapt- t edness of the “translation process” Ht = 0 Gs ds are essential for the assertion of Theorem 4.10.

2) The assumption Ea[Zt0 ]=1 ensuring that (Zt)t t0 is a Q-martingale is not always ≤ satisfied a sufficient condition is given in Theorem 4.12 below. If (Z ) is not a martin- − t 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 . However, in this case, P [Ω] < 1. The sub-probability measures Ft t Pt correspond to a transformed process with finite life-time.

~~UniversityofBonn WinterSemester2012/2013 178 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS~~

~~First applications to SDE~~

~~The Girsanov transformation can be used to construct weak solutions of stochastic differential equations. For example, consider an SDE~~

~~dX = b(t, X ) dt + dB , X = o, B BM(Rd), (4.34) t t t 0 ∼ where b : R Rd Rd is continuous, and o Rd is a ﬁxed initial value. Let + × → ∈ Xt(x)= xt denote the canonical process on~~

Ω := x C [0, ), Rd : x = o , ∈ ∞ 0 and let 0 denote Wiener measure on (Ω, X ). Then (X,0) is a Brownian motion. By F∞ changing measure, we will transform it into a weak solution of (4.34).

~~Assumption (A). The process~~

~~t 1 t Z = exp b(s,X ) dX b(s,X ) 2 ds , t 0, t s s − 2 | s | ≥ 0 0 is a martingale w.r.t. 0.~~

We will see later that the assumption is always satisﬁed if b is bounded, or, more generally, growing at most linearly in x. If (A) holds then by the Kolmogorov extension theorem, there exists a probability measure b on X such that b is mutually absolutely F∞ continuous w.r.t. 0 on each of the σ-algebras X , t [0, ), with relative densities Ft ∈ ∞ db = Z 0-a.s. 0 X t d t F By Girsanov’s Theorem, the process

~~t B = X b(s,X ) ds, t 0, t t − s ≥ 0 is a Brownian motion w.r.t. b. Thus we have shown:~~

~~Corollary 4.11. Suppose that (A) holds. Then:~~

~~1) The process (X,b) is a weak solution of (4.34).~~

~~Stochastic Analysis Andreas Eberle 4.3. GIRSANOV TRANSFORMATION 179~~

~~b 1 2) Forany t [0, ), the distribution X− is absolutely continuous w.r.t. Wiener ∈ ∞ ◦ measure 0 on X with relative density Z . Ft t~~

b 1 b The second assertion holds since X− = . It yields a rigorous path integral ◦ b b representation for the solution (X, ) of the SDE (4.34): If t denotes the distribution d b of (Xs)s t on C0 [0, t], R w.r.t. then ≤ t 1 t b(dx) = exp b(s, x ) dx b(s, x ) 2 ds 0(dx). (4.35) t s s − 2 | s | t 0 0 By combining (4.35) with the heuristic path integral representation

~~t 0 1 1 2 “ (dx) = exp x′ ds δ (dx ) dx ” t −2 | s| 0 0 s 0 0~~

~~t b 1 1 2 “ (dx) = exp x′ b(s, x ) ds δ (dx ) dx ” (4.36) t −2 | s − s | 0 0 s 0 0~~

xs′ = b(s, xs) obtained by setting the noise term in the SDE (4.34) equal to zero. Of course, these arguments do not hold rigorously, because I(x) = for 0- and b- almost every x. ∞ t t Nevertheless, they provide an extremely valuable guideline to conclusions that can then be veriﬁed rigorously, e.g. via (4.35).

Example (Likelihood ratio test for non-linear ﬁltering). Suppose that we are observ- d ing a noisy signal (Xt) taking values in R with X0 = o. We would like to decide if there is only noise, or if the signal is coming from an object moving with law of motion dx/dt = H(x) where H C2(Rd). The noise is modelled by the increments of a −∇ ∈ UniversityofBonn WinterSemester2012/2013 180 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS

Brownian motion (white noise). This is a simpliﬁed form of models that are used frequently in nonlinear ﬁltering (in realistic models often the velocity or the acceleration is assumed to satisfy a similar equation). In a hypothesis test, the null hypothesis and the alternative would be

~~H0 : Xt = Bt,~~

H1 : dXt = b(Xt) dt + dBt, where (B ) is a d-dimensional Brownian motion, and b = H. In a likelihood ratio t −∇ test based on observations up to time t, the test statistic would be the likelihood ratio b 0 dt /dt which by (4.35) and Itô’s formula is given by db t 1 t = exp b(X ) dX b(X ) 2 ds d0 s s − 2 | s | t 0 0 1 t = exp H(X ) H(X )+ (∆H H 2)(X ) ds (4.37) 0 − t 2 − |∇ | s 0 The null hypothesis H0 would then be rejected if this quantity exceeds some given value c for the observed signal x, i.e. , if 1 t H(x ) H(x )+ (∆H H 2)(x ) ds > log c. (4.38) 0 − t 2 − |∇ | s 0 Note that the integration by parts in (4.37) shows that the estimation procedure is quite stable, because the log likelihood ratio in (4.38) is continuous w.r.t. the supremum norm d on Co([0, t], R ).

~~Novikov’s condition~~

~~We ﬁnally derive a sufﬁcient condition due to Novikov for ensuring that the exponential~~

Z = exp L 1/2 [L] t t − t of a continuous local ( ) martingale is a martingale. Recall that Z is always a non- Ft negative local martingale, and hence a supermartingale w.r.t. ( ). Ft Theorem 4.12 (Novikov 1971). Let t R . If E[exp p [L] ] < for some p > 1 0 ∈ + 2 t0 ∞ then (Zt)t t0 is an ( t) martingale. ≤ F Stochastic Analysis Andreas Eberle 4.4. DRIFT TRANSFORMATIONS FOR ITÔ DIFFUSIONS 181

Remark (p =1). The Novikov criterion also applies if the condition above is satisﬁed for p = 1. Since the proof in this case is slightly more difﬁcult, we only prove the restricted form of Novikov’s criterion stated above.

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 ∈ N is enough to show that the random variables Zt Tn , n , are uniformly integrable for ∧ ∈ 1 1 each ﬁxed t t . However, for c> 0 and p, q (1, ) with p− + q− =1, we have ≤ 0 ∈ ∞

E[Zt Tn ; Zt Tn c] ∧ ∧ ≥ p p 1 = E exp Lt Tn [L]t Tn exp − [L]t Tn ; Zt Tn c (4.39) ∧ − 2 ∧ 2 ∧ ∧ ≥ 2 p 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 ∧ ≥ N for any n . Here we have used Hölder’s inequality and the fact that exp pLt Tn ∈ ∧ − p2 p [L]t Tn is an exponential supermartingale. If exp [L]t is integrable then the right 2 ∧ 2 hand side of (4.39) converges to 0 uniformly in n as c , because →∞ 1 1 P [Zt Tn 0] c− E[Zt Tn ] c− 0 ∧ ≥ ≤ ∧ ≤ −→ N uniformly in n as c . Hence Zt Tn n is uniformly integrable. →∞ { ∧ | ∈ } Example. If L = t G dX with a Brownian motion (X ) and an adapted process t 0 s s t (Gt) that is uniformly bounded on [0, t] for any ﬁnite t the quadratic variation then the quadratic variation [L] = t G2 ds is also bounded for ﬁnite t. Hence exp(L 1 [L]) t 0 | s| − 2 is an ( ) martingale for t [0, ). Ft ∈ ∞ A more powerful application of Novikov’s criterion is considered in the beginning of Section 3.4.

~~4.4 Drift transformations for Itô diffusions~~

~~We now consider an SDE~~

~~dX = b(X ) dt + dB , X = o, B BM(Rd) (4.40) t t t 0 ∼ UniversityofBonn WinterSemester2012/2013 182 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS~~

d d d with initial condition o R and b ∞ (R , R ). We will show that the solution ∈ ∈ Lloc constructed by Girsanov transformation is a Markov process, and we will study its transition function, as well as the bridge process obtained by conditioning on a given value 0 X d at a ﬁxed time. Let denote Wiener measure on (Ω, ) where Ω = C0([0, ), R ) F∞ ∞ and Xt(x)= xt is the canonical process on Ω. Similarly as above, we assume: Assumption (A). The exponential Z = exp t b(X ) dX 1 t b(X ) 2 ds is a t 0 s s − 2 0 | s | 0 martingale w.r.t. . We note that by Novikov’s criterion, the assumption always holds if

~~b(x) c (1 + x ) for some ﬁnite constant c> 0 : (4.41) | | ≤ | | Exercise (Martingale property for exponentials).~~

~~a) Prove that a non-negative supermartingale (Z ) satisfying E[Z ]=1 for any t 0 t t ≥ is a martingale.~~

~~b) Now consider t 1 t Z = exp b(X ) dX b(X ) 2 ds , t s s − 2 | s | 0 0 where b : Rd Rd is a continuous vector ﬁeld, and (X ) is a Brownian motion → t w.r.t. the probability measure P .~~

~~(i) Show that (Zt) is a supermartingale.~~

~~(ii) Prove that (Zt) is a martingale if (4.41) holds.~~

~~Hint: Prove ﬁrst that E[exp ε b(X ) 2 ds] < for ε > 0 sufﬁciently 0 | s | ∞ small, and conclude that E[Zε]=1. Then show by induction that E[Zkε]= 1 for any k N. ∈~~

~~The Markov property~~

Recall that if (A) holds then there exists a (unique) probability measure b on (Ω, X ) F∞ such that b[A] = E0[Z ; A] for any t 0 and A X . t ≥ ∈Ft Stochastic Analysis Andreas Eberle 4.4. DRIFT TRANSFORMATIONS FOR ITÔ DIFFUSIONS 183

~~Here E0 denotes expectation w.r.t. 0. By Girsanov’s Theorem, the process (X,b) is a weak solution of (4.40). Moreover, we can easily verify that (X,b) is a Markov process:~~

~~Theorem 4.13 (Markov property). If (A) holds then (X,b) is a time-homogeneous Markov process with transition function~~

pb(o,C) = b[X C] = E0[Z ; X C] C (Rd). t t ∈ t t ∈ ∀ ∈ B Proof. Let 0 s t, and let f : Rd R be a non-negative measurable function. ≤ ≤ → + Then, by the Markov property for Brownian motion,

~~Eb[f(X ) X] = E0[f(X )Z X]/Z t |Fs t t|Fs s t t 0 1 2 X = E f(Xt) exp b(Xr) dXr b(Xr) dr s s − 2 s | | F 0 b = EXs [f(Xt s)Zt s] = (pt sf)(Xs) − − −~~

~~0 b 0 - and -almost surely where Ex denotes the expectation w.r.t. Wiener measure with start at x.~~

Remark. 1) If b is time-dependent then one veriﬁes in the same way that (Xb,) is a time-inhomogeneous Markov process. 2) It is not always easy to prove that solutions of SDE are Markov processes. If the solution is not unique then usually, there are solutions that are not Markov processes.

~~Bridges and heat kernels~~

We now restrict ourselves to the time-interval [0, 1], i.e., we consider a similar setup as before with Ω = C ([0, 1], Rd). Note that X is the Borel σ-algebra on the Banach 0 F1 b space Ω. Our goal is to condition the diffusion process (X) on a given terminal value X = y, y Rd. More precisely, we will construct a regular version y → µb of the 1 ∈ y b conditional distribution µ [|X1 = y] in the following sense:

~~(i) For any y Rd, b is a probability measure on (Ω) such that b [X = y]=1. ∈ y B y 1 UniversityofBonn WinterSemester2012/2013 184 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS~~

~~(ii) Disintegration: For any A (Ω), the function y b [A] is measurable, and ∈ B → y~~

b b b [A] = y[A]p1(o,dy). Rd (iii) The map y b is continuous w.r.t. weak convergence of probability measures. → y Example (Brownian bridge). For b = 0, a regular version y 0 of the conditional → y distribution 0[ X = y] w.r.t. Wiener measure can be obtained by linearly transform- | 1 ing the paths of Brownian motion, cf. Theorem 8.11 in [13]: Under 0, the process Xy := X tX + ty, 0 t 1, is independent of X with terminal value y, and the t t − 1 ≤ ≤ 1 0 y 0 o 0 law y of (Xt )t [0,1] w.r.t. is a regular version of [ X1 = y]. The measure y is ∈ | called “pinned Wiener measure”.

The construction of a bridge process described in the example only applies for Brow- nian motion, which is a Gaussian process. For more general diffusions, the bridge can not be constructed from the original process by a linear transformation of the paths. For perturbations of a Brownian motion by a drift, however, we can apply Girsanov’s The- orem to construct a bridge measure.

~~We assume for simplicity that b is the gradient of a C2 function:~~

~~b(x) = H(x) with H C2(Rd). −∇ ∈~~

Then the exponential martingale (Zt) takes the form 1 t Z = exp H(X ) H(X )+ (∆H H 2)(X ) ds , t 0 − t 2 − |∇ | s 0 0 cf. (4.37). Note that the expression on the right-hand side is deﬁned y-almost surely for any y. Therefore, (Zt) can be used for changing the measure w.r.t. the Brownian bridge.

~~Theorem 4.14 (Heat kernel and Bridge measure). Suppose that (A) holds. Then:~~

~~b 1) The measure p1(o,dy) is absolutely continuous w.r.t. d-dimensional Lebesgue measure with density~~

~~pb (o,y) = p0(o,y) E0[Z ]. 1 1 y 1~~

~~Stochastic Analysis Andreas Eberle 4.4. DRIFT TRANSFORMATIONS FOR ITÔ DIFFUSIONS 185~~

2) A regular version of b[ X = y] is given by | 1 p0(o,y) exp H(o) 1 1 b (dx) = 1 exp (∆H H 2)(x ) ds 0(dx). y pb (o,y) exp H(y) 2 − |∇ | s y 1 0 The theorem yields the existence and a formula for the heat kernel p1(o,y), as well as a b path integral representation for the bridge measure y: 1 1 b (dx) exp (∆H H 2)(x ) ds 0(dx). (4.42) y ∝ 2 − |∇ | s y 0 Proof of 4.14. Let F : Ω R and g : R R be measurable functions. By the → + → + disintegration of Wiener measure into pinned Wiener measures,

Eb[F g(X )] = E0[Fg(X )Z ] = E0[FZ ] g(y) p0(o,y) dy. 1 1 1 y 1 1 Choosing F 1, we obtain ≡ b 0 0 g(y) p1(o,dy) = g(y) Ey [Z1] p1(x, y) dy for any non-negative measurable function g, which implies 1). Now, by choosing g 1, we obtain ≡ 0 E [FZ1] Eb[F ] = E0[FZ ] p0(o,y) dy = y pb (o,dy) (4.43) y 1 1 E0[Z ] 1 y 1 b b = Ey[F ] p1(0,dy) (4.44) by 1). This proves 2), because X = yb -almost surely, and y b is weakly contin- 1 y → y uous.

Remark (Non-gradient case). If b is not a gradient then things are more involved because the expressions for the relative densities Zt involve a stochastic integral. One 0 can proceed similarly as above after making sense of this stochastic integral for <- almost every x. Example (Reversibility in the gradient case). The representation (4.42) immediately implies the following reversibility property of the diffusion bridge when b is a gradient: d d If R : C([0, 1], R ) C([0, 1], R ) denotes the time-reversal deﬁned by (Rx)t = x1 t, → − b 1 then the image R− of the bridge measure from o to y coincides with the bridge y ◦ measure from y to o. Indeed, this property holds for the Brownian bridge, and the relative density in (4.42) is invariant under time reversal.

~~UniversityofBonn WinterSemester2012/2013 186 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS~~

~~Drift transformations for general SDE~~

~~We now consider a drift transformation for an SDE of the form~~

~~dXt = σ(t, Xt) dWt + b(t, Xt) dt (4.45) where W is an Rd-valued Brownian motion w.r.t. the underlying probability measure Q. We change measure via an exponential martingale of type~~

~~t 1 t Z = exp β(s,X ) dW β(s,X ) 2 ds t s s − 2 | s | 0 0 n n n n d where b, β : R R R and σ : R R R × are continuous functions. + × → + × →~~

~~Corollary 4.15. Suppose that (X, Q) is a weak solution of (4.45). If (Zt)t t0 is a Q- ≤ martingale and P Q on t0 with relative density Zt0 then ((Xt)t t0 ,P ) is a weak ≪ F ≤ solution of~~

~~dX = σ(t, X ) dB +(b + σβ)(t, X ) dt, B BM(Rd). (4.46) t t t t ∼ Proof. By (4.45), the equation (4.46) holds with~~

~~t B = W β(s,X ) ds. t t − s 0 Girsanov’s Theorem implies that B is a Brownian motion w.r.t. P .~~

~~Note that the Girsanov transformation induces a corresponding transformation for the martingale problem: If (X, Q) solves the martingale problem for the operator~~

~~1 ∂2 = aij + b , a = σσT , L 2 ∂xi∂xj ∇ i,j then (X,P ) is a solution of the martingale problem for~~

~~= + β σT . L L ∇ This “Girsanov transformation for martingale problems” carries over to diffusion processes with more general state spaces than Rn.~~

~~Stochastic Analysis Andreas Eberle 4.4. DRIFT TRANSFORMATIONS FOR ITÔ DIFFUSIONS 187~~

~~Doob’s h-transform and diffusion bridges~~

In the case of Itô diffusions, the h-transform for Markov processes is a special case of the drift transform studied above. Suppose that h C2(R Rn) is a strictly positive ∈ + × space-time harmonic function for the generator of the Itô diffusion (X, Q) (deﬁned L by (4.41)) with h(0,o)=1: ∂h + h = 0, h(0,o)=1. (4.47) ∂t L Then, by Itô’s formula, the process

Z = h(t, X ), t 0, t t ≥ is a positive local Q-martingale satisfying Z0 = 1 Q-almost surely. We can therefore try to change the measure via (Zt). For this purpose we write Zt in exponential form. By the Itô-Doeblin formula and (4.47),

dZ = (σT h)(t, X ) dW . t ∇ t t Hence Z = L = exp(L 1 [L] ) where t Et t − 2 t t 1 t L = dZ = (σT log h)(s,X ) dW t Z s ∇ s s 0 s 0 dP is the stochastic logarithm of Z. Thusif Z is a martingale, and P Q with = Zt dQ t ≪ F then (X,P ) solves the SDE (4.45) with β = σT log h. ∇

~~Example. If Xt = Wt is a Brownian motion w.r.t. Q then~~

~~dX = log h(t, X ) dt + dB , B BM(Rd) w.r.t. P. t ∇ t t ∼ For example, choosing h(t, x) = exp(α x 1 α 2t),α Rd, (X,P ) is a Brownian − 2 | | ∈ motion with constant drift α.~~

A more interesting application of the h-transform is the interpretation of diffusion bridges by a change of measure w.r.t. the law of the unconditioned diffusion process (X,b) on d C0([0, 1], R ) satisfying

~~dXt = dBt + b(Xt) dt, X0 = o,~~

UniversityofBonn WinterSemester2012/2013 188 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS with an Rd-valued Brownian motion B. We assume that the transition density (t, x, y) → pb(x, y) is smooth for t > 0 and bounded for t ε for any ε > 0. Then for y R, t ≥ ∈ pb( ,y) satisﬁes the Kolmogorov backward equation t ∂ pb( ,y) = bpb( ,y) for any t> 0, ∂t t L t where b = 1 ∆+ b is the corresponding generator. Hence L 2 ∇ b b h(t, z) = p1 t(z,y)/p1(o,y), t< 1, − is a space-time harmonic function with h(0,o)=1. Since h is bounded for t 1 ε ≤ − for any ε> 0, h(t, X ) is a martingale for t< 1. Now suppose that b b on with t y ≪ Ft relative density h(t, Xt) for any t < 1. Then the marginal distributions of the process b b (Xt)t<1 under , y respectively are

b b b k b (Xt1 ,...,Xtk ) pt1 (o, x1)pt2 t1 (x1, x2) pt t − (xk 1, xk) dx w.r.t. , ∼ − k− k 1 − b b b b pt1 (o, x1)pt2 t1 (x1, x2) pt t − (xk 1, xk)p1 t (xk,y) − k− k 1 − − k k b b dx w.r.t. y. ∼ p1(o,y) This shows that y b coincides with the regular version of the conditional distribution → y b b of given X1, i.e., y is the bridge measure from o to y. Hence, by Corollary 4.15, we have shown:

~~b Theorem 4.16 (SDE for diffusion bridges). The diffusion bridge (X,y) is a weak solution of the SDE~~

b dXt = dBt + b(Xt) dt + ( log p1 t( ,y))(Xt) dt, t< 1. (4.48) ∇ − b Note that the additional drift β(t, x) = log p1 t( ,y)(x) is singular as t 1. Indeed, ∇ − ↑ if at a time close to 1 the process is still far away from y, then a strong drift is required to force it towards y. On the σ-algebra , the measures b and b are singular. F1 y Remark (Generalized diffusion bridges). Theorem 4.16 carries over to bridges of diffusion processes with non-constant diffusion coefﬁcients σ. In this case, the SDE (4.48) is replaced by

T dXt = σ(Xt) dBt + b(Xt) dt + σσ log p1 t( ,y) (Xt) dt. (4.49) ∇ − The last term can be interpreted as a gradient of the logarithmic heat kernel w.r.t. the T 1 Riemannian metric g =(σσ )− induced by the diffusion process.

~~Stochastic Analysis Andreas Eberle 4.5.LARGEDEVIATIONSONPATHSPACES 189~~

~~4.5 Large deviations on path spaces~~

~~In this section, we apply Girsanov’s Theorem to study random perturbations of a dynamical system of type~~

ε ε ε dXt = b(Xt ) dt + √ε dBt, X0 = 0, (4.50) asymptotically as ε 0. We show that on the exponential scale, statements about the ↓ probabilities of rare events suggested by path integral heuristics can be put in a rigorous form as a large deviation principle on path space. Before, we give a complete proof of the Cameron-Martin Theorem.

~~Let Ω= C ([0, 1], Rd) endowed with the supremum norm ω = sup ω(t) : t [0, 1] , 0 || || {| | ∈ } let denote Wiener measure on (Ω), and let W (ω)= ω(t). B t~~

~~Translations of Wiener measure~~

~~For h Ω, we consider the translation operator τ : Ω Ω, ∈ h →~~

~~τh(ω) = ω + h,~~

~~1 and the translated Wiener measure := τ − . h ◦ h Theorem 4.17 (Cameron, Martin 1944). Let h Ω. Then if and only if h is ∈ t ≪ contained in the Cameron-Martin space~~

~~2 d H = h Ω: h is absolutely contin. with h′ L ([0, 1], R ) . CM ∈ ∈ In this case, the relative density of h w.r.t. is~~

t t dh 1 2 = exp hs′ dWs hs′ ds . (4.51) d 0 − 2 0 | | Proof. “ ” is a consequence of Girsanov’s Theorem: For h H , the stochastic ⇒ ∈ CM 1 2 integral h′ dW has ﬁnite deterministic quadratic variation [ h′ dW ] = h′ ds. 1 0 | | Hence by Novikov’s criterion, t t 1 2 Zt = exp h′ dW h′ ds 0 − 2 0 | | UniversityofBonn WinterSemester2012/2013 190 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS is a martingale w.r.t. Wiener measure . Girsanov’s Theorem implies that w.r.t. the measure ν = Z , the process (W ) is a Brownian motion translated by (h ). Hence 1 t t

~~1 1 = (W + h)− = ν W − = ν. h ◦ ◦~~

“ ” To prove the converse implication let h Ω, and suppose that . Since W ⇐ ∈ h ≪ is a Brownian motion w.r.t. , W h is a Brownian motion w.r.t. . In particular, it − h is a semimartingale. Moreover, W is a semimartingale w.r.t. and hence also w.r.t. h. Thus h = W (W h) is also a semimartingale w.r.t. . Since h is deterministic, this − − h implies that h has ﬁnite variation. We now show:

Claim. The map g 1 g dh is a continuous linear functional on L2([0, 1], Rd). → 0 The claim implies h H . Indeed, by the claim and the Riesz Representation Theo- ∈ CM rem, there exists a function f L2([0, 1], Rd) such that ∈ 1 1 g dh = g fds for any g L2([0, 1], Rd). ∈ 0 0 2 d Hence h is absolutely continuous with h′ = f L ([0, 1], R ). To prove the claim ∈ 2 d let (g ) be a sequence in L ([0, 1], R ) with g 2 0. Then by Itô’s isometry, n || n||L → g dW 0 in L2(), and hence - and -almost surely along a subsequence. Thus n → h also g dh = g d(W + h) g dW 0 n n − n −→ -almost surely along a subsequence. Applying the same argument to a subsequence of (g ), we see that every subsequence (g ) has a subsequence (ˆg ) such that gˆ dh 0. n n n n → This shows that g dh converges to 0 as well. The claim follows, since (g ) was an n n 2 Rd arbitrary null sequence in L ([0, 1], ). A ﬁrst consequence of the Cameron-Martin Theorem is that the support of Wiener mea- d sure is the whole space Ω= C0([0, 1], R ):

~~Corollary 4.18 (Support Theorem). For any h Ω and δ > 0, ∈ ω Ω: ω h < δ > 0. { ∈ || − || } Stochastic Analysis Andreas Eberle 4.5.LARGEDEVIATIONSONPATHSPACES 191~~

~~Proof. Since the Cameron-Martin space is dense in Ω w.r.t. the supremum norm, it is enough to prove the assertion for h H . In this case, the Cameron-Martin Theorem ∈ CM implies~~

~~W h < δ = h W < δ > 0. || − || − || || as [ W < δ] > 0 and h . || || − ≪ Remark (Quantitative Support Theorem). More explicitly,~~

W h < δ = h W < δ || − || − || || 1 1 1 2 = E exp h′ dW h′ ds ; max Ws < δ − 0 − 2 0 | | s 1 | | ≤ where the expectation is w.r.t. Wiener measure. This can be used to derive quantitative estimates.

~~Schilder’s Theorem~~

~~We now study the solution of (4.50) for b 0, i.e., ≡ Xε = √ε B , t [0, 1], t t ∈ with ε> 0 and a d-dimensional Brownian motion (Bt). Path integral heuristics suggests that for h H , ∈ CM~~

~~ε h I(h/√ε) I(h)/ε “ P [X h] = W e− = e− ” ≈ ≈ √ε ∼ where I : Ω [0, ] is the action functional deﬁned by → ∞~~

1 1 2 ω′(s) ds if ω HCM , I(ω) = 2 0 | | ∈ + otherwise.  ∞ The heuristics can be turned into a rigorous statement asymptotically as ε 0 on the → exponential scale. This is the content of the next two results that together are know as Schilder’s Theorem:

~~UniversityofBonn WinterSemester2012/2013 192 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS~~

~~Theorem 4.19 (Schilder’s large derivation principle, lower bound).~~

1) For any h H and δ > 0, ∈ CM lim inf ε log √εW B(h, δ)] I(h). ε 0 ↓ ∈ ≥ − 2) For any open subset U Ω, ⊆ lim inf ε log √εW U inf I(ω). ε 0 ω U ↓ ∈ ≥ − ∈ Here B(h, δ)= ω Ω: ω h < δ denotes the ball w.r.t. the supremum norm. { ∈ || − || } Proof. 1) Let c = 8I(h). Then for ε> 0 sufﬁciently small,

√εW B(h, δ) = W B(h/√ε,δ/√ε) ∈ ∈ = h/√ε B(0,δ/√ε) − 1 1 1 1 2 δ = E exp h′ dW h′ ds ; B 0, − √ε 0 − 2ε 0 | | √ε 1 c 1 δ exp I(h) h′ dW c B(0, ) ≥ − ε − √ε 0 ≤ ∩ √ε 1 1 8I(h) exp I(h) ≥ 2 −ε − ε where E stands for expectation w.r.t. Wiener measure. Here we have used that

1 1 2 2 2 h′ dW >c c− E h′ dW = 2I(h)/c 1/4 0 ≤ 0 ≤ by Itô’s isometry and the choice of c. 2) Let U be an open subset of Ω. For h U H , there exists δ > 0 such that ∈ ∩ CM B(h, δ) U. Hence by 1), ⊂ lim inf ε log [√εW U] I(h). ε 0 ↓ ∈ ≥ − Since this lower bound holds for any h U H , and since I = on U H , we ∈ ∩ CM ∞ \ CM can conclude that

~~lim inf ε log [√εW U] inf I(h) = inf I(ω). ε 0 h U HCM ω U ↓ ∈ ≥ − ∈ ∩ − ∈~~

~~Stochastic Analysis Andreas Eberle 4.5.LARGEDEVIATIONSONPATHSPACES 193~~

To prove a corresponding upper bound, we consider linear approximations of the Brow- nian paths. For n N let ∈ W (n) := (1 s)W + sW t − k/n k+1/n whenever t =(k + s)/n for k 0, 1,...,n 1 and s [0, 1]. ∈{ − } ∈ Theorem 4.20 (Schilder’s large deviations principle, upper bound).

1) For any n N and λ 0, ∈ ≥ lim sup ε log [I(√εW (n)) λ] λ. ε 0 ≥ ≤ − ↓ 2) For any closed subset A Ω, ⊆ lim sup ε log [√εW A] inf I(ω). ε 0 ∈ ≤ − ω A ↓ ∈ Proof. 1) Let ε> 0 and n N. Then ∈ n (n) 1 2 I(√εW ) = ε n (Wk/n W(k 1)/n) . 2 − − k=1

~~Since the random variables ηk := √n (Wk/n W(k 1)/n) are independent and standard − − normally distributed, we obtain~~

~~[I(√εW (n)) λ] = η 2 2λ/ε ≥ | k| ≥ exp(2λc/ε) E exp c η 2 , ≤ − | k| where the expectation on the right hand side is ﬁnite for c< 1/2. Hence for any c< 1/2,~~

lim sup ε log [I(√εW (n)) λ] 2cλ. ε 0 ≥ ≤ − ↓ The assertion now follows as c 1/2. ↑ 2) Now ﬁx a closed set A Ω and λ < inf I(ω): ω A . To prove the second ⊆ { ∈ } assertion it sufﬁces to show

~~lim sup ε log [√εW A] λ. (4.52) ε 0 ∈ ≤ − ↓ UniversityofBonn WinterSemester2012/2013 194 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS~~

~~By the Theorem of Arzéla-Ascoli, the set I λ is a compact subset of the Banach { ≤ } space Ω. Indeed, by the Cauchy-Schwarz inequality,~~

t ω(t) ω(s) = ω′(u) du √2λ √t s s, t [0, 1] | − | s ≤ − ∀ ∈ holds for any ω Ω satisfying I(ω) λ. Hence the paths in I λ are equicontinu- ∈ ≤ { ≤ } ous, and the Arzéla-Ascoli Theorem applies. Let δ denote the distance between the sets A and I λ w.r.t. the supremum norm. { ≤ } Note that δ > 0, because A is closed, I λ is compact, and both sets are disjoint by { ≤ } the choice of λ. Hence for ε> 0, we can estimate

~~[√εW A] [I(√εW (n)) >λ]+ [ √εW √εW (n) > δ]. ∈ ≤ || − ||sup The assertion (4.52) now follows from~~

lim sup ε log [I(√εW (n)) >λ] λ, and (4.53) ε 0 ≤ − ↓ (n) lim sup ε log [ W W sup > δ/√ε] λ. (4.54) ε 0 || − || ≤ − ↓ The bound (4.53) holds by 1) for any n N. The proof of (4.54) reduces to an estimate ∈ of the supremum of a Brownian bridge on an interval of length 1/n. We leave it as an exercise to verify that (4.54) holds if n is large enough.

Remark (Large deviation principle for Wiener measure). Theorems 4.19 and 4.20 show that 1 [√εW A] exp inf I(ω) ∈ ≃ − ε ω A ∈ holds on the exponential scale in the sense that a lower bound holds for open sets and an upper bound holds for closed sets. This is typical for large deviation principles, see e.g. [8] or [9]. The proofs above based on “exponential tilting” of the underlying Wiener measure (Girsanov transformation) for the lower bound, and an exponential estimate combined with exponential tightness for the upper bound are typical for the proofs of many large deviation principles.

~~Stochastic Analysis Andreas Eberle 4.5.LARGEDEVIATIONSONPATHSPACES 195~~

~~Random perturbations of dynamical systems~~

~~We now return to our original problem of studying small random perturbations of a dynamical system~~

~~ε ε ε dXt = b(Xt ) dt + √ε dBt, X0 = 0. (4.55)~~

~~This SDE can be solved pathwise:~~

Lemma 4.21 (Control map). Suppose that b : Rd Rd is Lipschitz continuous. Then: → 1) For any function ω C([0, 1], Rd) there exists a unique function x C([0, 1], Rd) ∈ ∈ such that t x(t) = b(x(s)) ds + ω(t) t [0, 1]. (4.56) ∀ ∈ 0 The function x is absolutely continuous if and only if ω is absolutely continuous, and in this case,

~~x′(t) = b(x(t)) + ω′(t) for a.e. t [0, 1]. (4.57) ∈~~

2) The control map : C([0, 1], Rd) C([0, 1], Rd) that maps ω to the solution J → (ω)= x of (4.56) is continuous. J Proof. 1) Existence and uniqueness holds by the classical Picard-Lindelöf Theorem. 2) Suppose that x = (ω) and x = (ω) are solutions of (4.56) w.r.t. driving paths J J ω, ω C[0, 1], Rd). Then for t [0, 1], ∈ ∈ t x(t) x(t) = (b(ω(s)) b(ω(s))) ds + √ε(ω(t) ω(t)) | − | 0 − − t L ω(s) ω(s) ds + √ε (t)m ω(t) . ≤ | − | | − | 0 where L R is a Lipschitz constant for b. Gronwall’s Lemma now implies ∈ + x(t) x(t) exp(tL) √ε ω ω t [0, 1], | − | ≤ || − ||sup ∀ ∈ and hence x x exp(L) √ε ω ω . || − ||sup ≤ || − ||sup This shows that the control map is even Lipschitz continuous. J UniversityofBonn WinterSemester2012/2013 196 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS

~~For ε> 0, the unique solution of the SDE (4.55) on [0, 1] is given by~~

Xε = (√εB). J Since the control map is continuous, we can apply Schilder’s Theorem to study the J large deviations of Xε as ε 0: ↓ Theorem 4.22 (Fredlin & Wentzel 1970, 1984). If b is Lipschitz continuous then the large deviations principle

ε lim inf ε log P [X U] inf Ib(x) for any open set U Ω, ε 0 x U ↓ ∈ ≥ − ∈ ⊆ ε lim inf ε log P [X A] inf Ib(x) for any closed set A Ω, ε 0 x A ↓ ∈ ≥ − ∈ ⊆ holds, where the rate function I : Ω [0, ] is given by b → ∞ 1 1 2 2 0 x′(s) b(x(s)) ds for x HCM , Ib(x) = | − | ∈ + for x Ω H .  ∞ ∈ \ CM Proof. For any set A Ω, we have ⊆ ε 1 1 P [X A] = P [√εB − (A)] = [√εW − (A)]. ∈ ∈J ∈J 1 If A is open then − (A) is open by continuity of , and hence J J lim inf ε log P [Xε A] inf I(ω)) ε 0 ∈ ≥ − ω −1(A) ↓ ∈J 1 by Theorem 4.19. Similarly, if A is closed then − (A) is closed, and hence the corre- J sponding upper bound holds by Theorem 4.20. Thus it only remains to show that

inf I(ω) = inf Ib(x). ω −1(A) x A ∈J ∈ 1 To this end we note that ω − (A) if and only if x = (ω) A, and in this case ∈ J J ∈ ω′ = x′ b(x). Therefore, − 1 1 2 inf I(ω) = inf ω′(s) ds −1 −1 ω (A) ω (A) HCM 2 | | ∈J ∈J ∩ 0 1 2 = inf x′(s) b(x(s)) ds = inf Ib(x). x A HCM x A ∈ ∩ 2| − | ∈

~~Stochastic Analysis Andreas Eberle 4.6. CHANGEOFMEASUREFORJUMPPROCESSES 197~~

Remark. The large deviation principle in Theorem 4.22 generalizes to non-Lipschitz continuous vector ﬁelds b and to SDEs with multiplicative noise. However, in this case, there is no continuous control map that can be used to reduce the statement to Schilder’s Theorem. Therefore, a different proof is required, cf. e.g. [8].

~~4.6 Change of measure for jump processes~~

A change of the underlying probability measure by an exponential martingale can also be carried out for jump processes. In this section, we ﬁrst consider absolutely continuous measure transformations for Poisson point processes. We then apply the results to Lévy processes, and ﬁnally we prove a general result for semimartingales.

~~Poisson point processes~~

Let (Nt)t 0 be a Poisson point process on an σ-finite measure space (S, , ν) that ≥ S is defined and adapted on a filtered probability space (Ω, , Q, ( )). Suppose that A Ft (ω,t,y) H (y)(ω) is a predictable process in 2 (P λ ν). Our goal is to change → t Lloc ⊗ ⊗ the underlying measure Q to a new measure P such that w.r.t. P , (Nt)t 0 is a point ≥ process with intensity of points in the infinitesimal time interval [t, t + dt] given by

~~(1 + Ht(y)) dt ν(dy).~~

Note that in general, this intensity may depend on ω in a predictable way. Therefore, under the new probability measureP , the process (Nt) is not necessarily a Poisson point process. We deﬁne a local exponential martingale by

~~L Zt := t where Lt := (H N)t. (4.58) E • Lemma 4.23. Suppose that inf H > 1, and let G := log (1 + H). Then for t 0, − ≥~~

~~L t = exp Gs(y) N(ds dy) (Hs(y) Gs(y)) ds ν(dy) . E (0,t] S − (0,t] S − × × UniversityofBonn WinterSemester2012/2013 198 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS~~

~~Proof. The assumption inf H > 1 implies inf ∆L > 1. Since, moreover, [L]c = 0, − − we obtain~~

~~L L [L]c/2 ∆L = e − (1+∆L)e− E = exp L + (log(1 + ∆L) ∆L) − = exp G N + (G H) ds ν(dy) . • − Here we have used that~~

~~(log(1 ∆L) ∆L) = log (1 + H (y)) H (y) N(ds dy) − − s − s holds, since log(1 + H (y)) H (y) const. H (y) 2 is integrable on ﬁnite time | s − s | ≤ | s | intervals.~~

~~The exponential Z = L is a strictly positive local martingale w.r.t. Q, and hence a t Et supermartingale. As usual, we ﬁx t R and assume: 0 ∈ +~~

Assumption. (Zt)t t0 is a martingale w.r.t. Q, i.e. EQ[Zt0 ]=1. ≤ Also tThen there is a probability measure P on such that Ft0 dP = Zt for any t t0. dQ t ≤ F In the deterministic case Ht(y)(ω)= h(y), we can prove that w.r.t. P , (Nt) is a Poisson point process with changed intensity measure

~~(dy) = (1+ h(y)) ν(dy):~~

Theorem 4.24 (Change of measure for Poisson point processes). Let (Nt, Q) be a Poisson point process with intensity measure ν, andlet g := log (1+h) where h 2(ν) ∈L satisﬁes inf h> 1. Suppose that the exponential − N h g Z = e = exp N + t (g h) dν t Et t − dP is a martingale w.r.t. Q, and assume that P Q on t with relative density dQ = Zt ≪ F t for any t 0. Then the process (N ,P ) is a Poisson point process with intensityF ≥ t measure d = (1+ h) dν.

~~Stochastic Analysis Andreas Eberle 4.6. CHANGEOFMEASUREFORJUMPPROCESSES 199~~

~~Proof. By the Lévy characterization for Poisson point processes (cf. the exercise below Lemma 4.1) it sufﬁces to show that the process~~

M f = exp iN f + tψ(f) , ψ(f) = 1 eif d, t t − is a local martingale w.r.t. P for any elementary functions f 1(S, , ν). Further- ∈ L S more, by Lemma 4.9, M f is a local martingale w.r.t. P if and only if M f Z is a local martingale w.r.t. Q. The local martingale property for (M f Z, Q) can be veriﬁed by a computation based on Itô’s formula.

Remark (Extension to general measure transformations). The approach in Theo- rem 4.24 can be extended to the case where the function h(y) is replaced by a general predictable process Ht(y)(ω). In that case, one veriﬁes similarly that under a new measure P with local densities given by (4.58), the process

M f = exp iN f + (1 eif(y))(1 + H (y)) dy t t − t is a local martingale for any elementary functions f 1(ν). This property can be used ∈L as a deﬁnition of a point process with predictable intensity (1 + Ht(y)) dt ν(dy).

~~Application to Lévy processes~~

Since Lévy processes can be constructed from Poisson point processes, a change of measure for Poisson point processes induces a corresponding transformation for Lévy processes. Suppose that ν is a σ-ﬁnite measure on Rd 0 such that ( y y 2) ν(dy) < \{ } | |∧| | , and let ∞ (dy) = (1+ h(y)) ν(dy).

~~Recall that if (Nt, Q) is a Poisson point process with intensity measure ν, then~~

~~X = y N (dy), N = N tν, t t t t − is a Lévy martingale with Lévy measure w.r.t. Q.~~

~~UniversityofBonn WinterSemester2012/2013 200 CHAPTER 4. SDE II: TRANSFORMATIONS AND WEAK SOLUTIONS~~

~~Corollary 4.25. Suppose that h 2(ν) satisﬁes inf h > 1 and sup h < . Then ∈ L − ∞ the process~~

X = y N dy + t y h(y) ν(dy), N = N t, t t t t − is a Lévy martingale with Lévy measure w.r.t. P provided P Q on with relative ≪ Ft density Z for any t 0. t ≤ Example. Suppose that (X, Q) is a compound Poisson process with ﬁnite jump intensity measure ν, and let h Nt = h(∆Xs). s t ≤ with h as above. Then (X,P ) is a compound Poisson process with jump intensity measure d =(1+ h) dν provided

~~dP N h t R h dν = te = e− (1 + h(∆Xs)). dQ t E F s t ≤ A general theorem~~

~~We ﬁnally state a general change of measure theorem for possibly discontinuous semimartingales:~~

Theorem 4.26 (P.A. Meyer). Suppose that the probability measures P and Q are equiv- dP alent on t for any t 0 with relative density dQ = Zt. If M is a local martingale F ≥ t w.r.t. Q then M 1 d[Z, M] is a local martingaleF w.r.t. P . Z − The theorem shows that w.r.t. P , (Mt) is again a semimartingale, and it yields an explicit semimartingale decomposition for (M,P ). For the proof we recall that (Zt) is a local martingale w.r.t. Q and (1/Zt) is a local martingale w.r.t. P .

Proof. The process ZM [Z, M] is a local martingale w.r.t. Q. Hence by Lemmy 4.9, − the process M 1 [Z, M] is a local martingale w.r.t. P . It remains to show that 1 [Z, M] − Z Z 1 differs from Z d[Z, M] by a local P -martingale. This is a consequence of the Itô product rule: Indeed, 1 1 1 1 [Z, M] = [Z, M] d + d[Z, M]+ , [Z, M] . Z − Z Z Z − Stochastic Analysis Andreas Eberle 4.6. CHANGEOFMEASUREFORJUMPPROCESSES 201

~~The ﬁrst term on the right-hand sideis a local Q-martingale, since 1/Z is a Q-martingale. 1 The remaining two terms add up to Z d[Z, M], because 1 1 , [Z, M] = ∆ ∆[Z, M]. Z Z~~

~~1 Remark. Note that the process Z d[Z, M] is not predictable in general. For a predictable counterpart to Theorem 4.26 cf. e.g. [32].~~

~~UniversityofBonn WinterSemester2012/2013 Chapter 5~~

~~Variations of stochastic differential equations~~

This chapter contains a ﬁrst introduction to basic concepts and results of Malliavin calculus. For a more thorough introduction to Malliavin calculus we refer to [31], [30], [37], [21], [29] and [7].

Let denote Wiener measure on the Borel σ-algebra (Ω) over the Banach space B Ω = C ([0, 1], Rd) endowed with the supremum norm ω = sup ω(t) : t [0, 1] . 0 || || {| | ∈ } We consider an SDE of type

~~dXt = b(Xt) dt + σ(Xt) dWt, X0 = x, (5.1)~~

driven by the canonical Brownian motion Wt(ω) = ω(t). In this chapter, we will be interested in variations of the SDE and its solutions respectively. We will study the relations between different types of variations of (5.1):

Variations of the initial condition: x x(ε) • → Variations of the coefﬁcients: b(x) b(ε, x), σ(x) σ(ε, x) • → → Variations of the driving paths: W W + εH , (H ) adapted • t → t t t Variations of the underlying probability measure: ε = Zε • → 202 5.1.VARIATIONSOFPARAMETERSINSDE 203

We first prove differentiability of the solution w.r.t. variations of the initial condition and the coefficients, see Section 5.1. In Section 5.2, we introduce the Malliavin gradient which is a derivative of a function on Wiener space (e.g. the solution of an SDE) w.r.t. variations of the Brownian path. Bismut’s integration by parts formula is an infinitesimal version of the Girsanov Theorem, which relates these variations to variations of Wiener measure. After a digression to representation theorems in Section 5.3, Sec- tion 5.4 discusses Malliavin derivatives of solutions of SDE and their connection to variations of the initial condition and the coefficients. As a consequence, we obtain first stability results for SDE from the Bismut integration by parts formula. Finally, Section 5.5 sketches briefly how Malliavin calculus can be applied to prove existence and smoothness of densities of solutions of SDE. This should give a first impression of a powerful technique that eventually leads to impressive results such as Malliavin’s stochastic proof of Hörmander’s theorem, cf. [19], [30].

~~5.1 Variations of parameters in SDE~~

~~We now consider a stochastic differential equation~~

d ε ε ε k ε dXt = b(ε,Xt ) dt + σk(ε,Xt ) dWt , X0 = x(ε), (5.2) k=1 on Rn with coefﬁcients and initial condition depending on a parameter ε U, where ∈ U is a convex neighbourhood of 0 in Rm, m N. Here b, σ : U Rn Rn are ∈ k × → functions that are Lipschitz continuous in the second variable, and x : U Rn. We → ε already know that for any ε U, there exists a unique strong solution (Xt )t 0 of (5.2). ∈ ≥ For p [1, ) let ∈ ∞ 1/p ε ε p X p := E sup Xt . || || t [0,1] | | ∈ Exercise (Lipschitz dependence on ε). Prove that if the maps x, b and σk are all Lip- schitz continuous, then ε Xε is also Lipschitz continuous w.r.t. , i.e., there exists → ||||p a constant L R such that p ∈ + Xε+h Xε L h , for any ε, h Rm with ε,ε + h U. || − ||p ≤ p | | ∈ ∈ We now prove a stronger result under additional regularity assumptions.

~~UniversityofBonn WinterSemester2012/2013 204 CHAPTER 5. VARIATIONS OF SDE~~

~~Differentation of solutions w.r.t. a parameter~~

Theorem 5.1. Let p [2, ), and suppose that x, b and σ are C2 with bounded ∈ ∞ k derivatives up to order 2. Then the function ε Xε is differentiable on U w.r.t. , → ||||p ε dXε and the differential Y = dε is the unique strong solution of the SDE

∂b ∂b dY ε = (ε,Xε)+ (ε,Xε)Y ε dt (5.3) t ∂ε t ∂x t t d ∂σ ∂σ + k (ε,Xε)+ k (ε,Xε)Y ε dW k, ∂ε t ∂x t t t k=1 ε Y0 = x′(ε), (5.4) that is obtained by formally differentiating (5.2) w.r.t. ε.

~~∂ ∂ Here and below ∂ε and ∂x denote the differential w.r.t. the ε and x variable, and x′ denotes the (total) differential of the function x.~~

ε ε Remark. Note that if (Xt ) is given, then (5.3) is a linear SDE for (Yt ) (with multiplicative noise). In particular, there is a unique strong solution. The SDE for the derivative process Y ε is particularly simple if σ is constant: In that case, (5.3) is a deterministic ODE with coefﬁcients depending on Xε.

~~Proof of 5.1. We prove the stronger statement that there is a constant M (0, ) such p ∈ ∞ that~~

Xε+h Xε Y εh M h 2 (5.5) − − p ≤ p | | holds for any ε, h Rmwith ε,ε + h U, where Y ε is the unique strong solution of ∈ ∈ (5.3). Indeed, by subtracting the equations satisﬁed by Xε+h, Xε and Y εh, we obtain for t [0, 1]: ∈

~~t d t Xε+h Xε Y εh I + II ds + III dW k , t − t − t ≤ | | | | k 0 k=1 0 Stochastic Analysis Andreas Eberle 5.1.VARIATIONSOFPARAMETERSINSDE 205 where~~

~~I = x(ε + h) x(ε) x′(ε)h, − − ε+h ε ε h II = b(ε + h, X ) b(ε,X ) b′(ε,X ) , and − − Y εh~~

~~ε+h ε ε h IIIk = σk(ε + h, X ) σk(ε,X ) σk′ (ε,X ) . − − Y εh~~

~~Hence by Burkholder’s inequality, there exists a ﬁnite constant Cp such that~~

~~t d E (Xε+h Xε Y εh)⋆ p C I p + E II p + III p ds . (5.6) − − t ≤ p | | | | | k| 0 k=1 2 Since x, b and σk are C with bounded derivatives, there exist ﬁnite constants CI, CII,~~

~~CIII such that~~

I C h 2, (5.7) | | ≤ I| | ∂b II C h 2 + (ε,Xε)(Xε+h Xε Y εh) , (5.8) | | ≤ II| | ∂x − − 2 ∂σk ε ε+h ε ε IIIk CIII h + (ε,X )(X X Y h) . (5.9) | | ≤ | | ∂x − − Hence there exist ﬁnite constants Cp, Cˆp such that

~~d p E[ II p + III p] C h 2p + E Xε+h Xε Y εh , | | | k| ≤ p | | − − k=1 and thus, by (5.6) and (5.7),~~

~~t E (Xε+h Xε Y εh)⋆p Cˆ h 2p + C C E (Xε+h Xε Y εh)⋆p ds − − t ≤ p| | p p − − s 0 for any t 1. The assertion (5.5) now follows by Gronwall’s lemma. ≤~~

~~Derivative ﬂow and stability of stochastic differential equations~~

We now apply the general result above to variations of the initial condition, i.e., we consider the ﬂow d x x x k x dξt = b(ξt ) dt + σk(ξt ) dWt , ξ0 = x. (5.10) k=1 UniversityofBonn WinterSemester2012/2013 206 CHAPTER 5. VARIATIONS OF SDE

~~2 Assuming that b and σk (k = 1,...,d) are C with bounded derivatives, Theorem 5.1 shows that the derivative ﬂow~~

x ∂ x,l ′ Yt := ξt(x) = k ξt ∂x 1 k,l n ≤ ≤ x exists w.r.t. p and (Yt )t 0 satisﬁes the SDE |||| ≥ d x x x x x k x dYt = b′(ξt ) Yt dt + σk′ (ξt ) Yt dWt ,Y0 = In. (5.11) k=1 Note that again, this is a linear SDE for Y if ξ is given, and Y is the fundamental solution of this SDE.

~~Remark (Flow of diffeomorphisms). One can prove that x ξx(ω) is a diffeomor- → t phism on Rn for any t and ω, cf. [24] or [14].~~

~~In the sequel, we will denote the directional derivative of the ﬂow ξ in direction v Rn t ∈ by Yv,t: x x x Yv,t = Yv,t = Yt v = ∂vξt .~~

(i) Constant diffusion coefficients. Let us now first assume that d = n and σ(x) = In for any x Rn. Then the SDE reads ∈ x x x dξ = b(ξ ) dt + dW, ξ0 = x; and the derivative flow solves the ODE

~~x x dY = b′(ξ )Ydt, Y0 = In.~~

~~This can be used to study the stability of solutions w.r.t. variations of initial conditions pathwise:~~

~~Theorem 5.2 (Exponential stability I). Suppose that b : Rn Rn is C2 with bounded → derivatives, and let~~

~~κ = sup sup v b′(x)v. x Rn v Rn ∈ v∈=1 | | Then for any t 0 and x, y, v Rn, ≥ ∈ ∂ ξx eκt v , and ξx ξy eκt x y . | v t | ≤ | | | t − t | ≤ | − |~~

~~Stochastic Analysis Andreas Eberle 5.1.VARIATIONSOFPARAMETERSINSDE 207~~

~~The theorem shows in particular that exponential stability holds if κ< 0.~~

~~x x Proof. The derivative Yv,t = ∂vξt satisﬁes the ODE~~

~~dYv = b′(ξ)Yv dt.~~

~~Hence 2 2 d Y = 2Y b′(ξ)Y dt 2κ Y dt, | v| v v ≤ | v| which implies~~

~~∂ ξx 2 = Y x 2 e2κt v 2, and thus | v t | | v,t| ≤ | | 1 x y (1 s)x+sy κt ξt ξt = ∂x yξt − ds e x y . | − | 0 − ≤ | − |~~

~~n n Example (Ornstein-Uhlenbeck process). Let A R × . The generalized Ornstein- ∈ Uhlenbeck process solving the SDE~~

~~dξt = Aξt dt + dWt~~

n 1 is exponentially stable if κ = sup v Av : v S − < 0. { ∈ } (ii) Non-constant diffusion coefficients. If the diffusion coefficients are not constant, the noise term in the SDE for the derivative flow does not vanish. Therefore, the derivative flow can not be bounded pathwise. Nevertheless, we can still obtain stability in an L2 sense.

Lemma 5.3. Suppose that b, σ ,...,σ : Rn Rn are C2 with bounded derivatives. 1 d → Then for any t 0 and x, v Rn, the derivative ﬂow Y x = ∂ ξx is in L2(Ω, ,P ), ≥ ∈ v,t v t A and d E[ Y x 2] = 2E[Y x K(ξx)Y x ] dt | v,t| v,t t v,t where d 1 T K(x) = b′(x)+ σ′ (x) σ′ (x). 2 k k k=1 UniversityofBonn WinterSemester2012/2013 208 CHAPTER 5. VARIATIONS OF SDE

~~(k) Proof. Let Yv denote the k-the component of Yv. The Itô product rule yields~~

d Y 2 = 2Y dY + d[Y (k)] | v| v v v k (5.11) k 2 = 2Y b′(ξ)Y dt +2 Y σ′ (ξ) dW + σ′ (ξ)Y dt. v v v k | k v| k k Noting that the stochastic integrals on the right-hand side stopped at T = inf t 0 : Y n are martingales, we obtain n { ≥ | v,t| ≥ }

t Tn 2 2 ∧ E Yv,t Tn = v +2E Yv K(ξ)Yv ds . | ∧ | | | 0 The assertion follows as n . →∞ Theorem 5.4 (Exponential stability II). Suppose that the assumptions in Lemma 5.3 hold, and let κ := sup sup v K(x)v. (5.12) x Rn v Rn ∈ v∈=1 | | Then for any t 0 and x, y, v Rn, ≥ ∈

~~E[ ∂ ξx 2] e2κt v 2, and (5.13) | v t | ≤ | |~~

E[ ξx ξy 2]1/2 eκt x y . (5.14) | t − t | ≤ | − | Proof. Since K(x) κI holds in the form sense for any x, Lemma 5.3 implies ≤ n d E[ Y 2] 2κE[ Y 2]. dt | v,t| ≤ | v,t| (5.13) now follows immediately by Gronwell’s lemma, and (5.14) follows from (5.13) x y 1 (1 s)x+sy since ξt ξt = ∂x yξt − ds. − 0 − Remark. (Curvature) The quantity κ can be viewed as a lower curvature bound − w.r.t. the geometric structure deﬁned by the diffusion process. In particular, exponential stability w.r.t. the L2 norm holds if κ < 0, i.e., if the curvature is bounded from below by a strictly positive constant.

~~Stochastic Analysis Andreas Eberle 5.1.VARIATIONSOFPARAMETERSINSDE 209~~

~~Consequences for the transition semigroup~~

~~x We still consider the ﬂow (ξt ) of the SDE (5.1) with assumptions as in Lemma 5.3 and Theorem 5.4. Let~~

~~p (x, B) = P [ξx B], x Rn, B (Rn), t t ∈ ∈ ∈ B denote the transition function of the diffusion process on Rn. For two probability measures , ν on Rn, we deﬁne the L2 Wasserstein distance~~

2 1/2 2(, ν) = inf E[ X Y ] W (X,Y ) | − | X ,Y ν ∼ ∼ as the infimum of the L2 distance among all couplings of and ν. Here a coupling of and ν is defined as a pair (X,Y ) of random variables on a joint probability space with distributions X and Y ν. Let κ be defined as in (5.12). ∼ ∼ Corollary 5.5. For any t 0 and x, y Rn, ≥ ∈ p (x, ),p (y, ) eκt x y . W2 t t ≤ | − | Proof. The flow defines a coupling betweenp (x, ) and p (y, ) for any t, x and y: t t ξx p (x, ), ξy p (y, ). t ∼ t t ∼ t Therefore, p (x, ),p (y, ) 2 E ξx ξy 2 . W2 t t ≤ | t − t | The assertion now follows from Theorem 5.4.

Exercise (Exponential convergence to equilibrium). Suppose that is a stationary distribution for the diffusion process, i.e., is a probability measure on (Rn) satisfying B p = for every t 0. Prove that if κ < 0 and x 2 (dx) < , then for any t ≥ | | ∞ x Rd, p (x, ), 0 exponentially fast with rate κ as t . ∈ W2 t → →∞ Besides studying convergence to a stationary distribution, the derivative ﬂow is also useful for computing and controlling derivatives of transtion functions. Let

~~x (ptf)(x)= pt(x,dy)f(y)= E[f(ξt )] denote the transition semigroup acting on functions f : Rn R. We still assume the → conditions from Lemma 5.3.~~

~~UniversityofBonn WinterSemester2012/2013 210 CHAPTER 5. VARIATIONS OF SDE~~

Exercise (Lipschitz bound). Prove that for any Lipschitz continuous function f : Rn R, → p f eκt f t 0, || t ||Lip ≤ || ||Lip ∀ ≥ where f = sup f(x) f(y) / x y : x, y Rn s.t. x = y . || ||Lip {| − | | − | ∈ } For continuously differentiable functions f, we even obtain an explicit formula for the gradient of ptf:

Corollary 5.6 (First Bismut-Elworthy Formula). For any function f C1(Rn) and ∈ b t 0, p f is differentiable with ≥ t x n v p f = E Y x f x, v R . (5.15) ∇x t v,t ∇ξt ∀ ∈ x Here p f denotes the gradient evaluated at x. Note that Y x f is the directional ∇x t t,v ∇ξt x derivative of f in the direction of the derivative ﬂow Yt,v.

Proof of 5.6. For λ R 0 , ∈ \{ } λ (ptf)(x + λv) (ptf)(x) 1 x+λv x 1 x+sv − = E f(ξ ) f(ξ ) = E Y x+sv f ds. λ λ t − t λ v,t ∇ξt 0 The assertion now follows since x ξx and x Y x are continuous, f is continuous → t → v,t ∇ and bounded, and the derivative ﬂow is bounded in L2.

The ﬁrst Bismut-Elworthy Formula shows that the gradient of ptf can be controlled by the gradient of f for all t 0. In Section 5.4, we will see that by applying an integration ≥ by parts on the right hand side of (5.15), for t> 0 it is even possible to control the gradient of ptf in terms of the supremum norm of f, provided a non-degeneracy condition holds, cf. (??).

~~5.2 Malliavin gradient and Bismut integration by parts formula~~

d Let Wt(ω)= ωt denote the canonical Brownian motion on Ω= C0([0, 1], R ) endowed with Wiener measure. In the sequel, we denote Wiener measure by P , expectation values w.r.t. Wiener measure by E[ ], and the supremum norm by . |||| Stochastic Analysis Andreas Eberle 5.2. MALLIAVIN GRADIENT AND BISMUT INTEGRATION BY PARTS FORMULA 211

Deﬁnition. Let ω Ω. A function F : Ω R is called Fréchet differentiable at ω iff ∈ → there exists a continuous linear functional d F : Ω R such that ω → F (ω + h) F (ω) (d F )(h) = o( h ) for any h Ω. || − − ω || || || ∈ If a function F is Fréchet differentiable at ω then the directional derivatives ∂F F (ω + εh) F (ω) (ω) = lim − = (dωF )(h) ε 0 ∂h → ε exist for all directions h Ω. For applications in stochastic analysis, Fréchet differ- ∈ entiability is often too restrictive, because Ω contains “too many directions”. Indeed, solutions of SDE are typically not Fréchet differentiable as the following example indicates:

1 1 2 1 2 Example. Let F = 0 Wt dWt where Wt =(Wt , Wt ) is a two dimensional Brownian motion. A formal computation of the derivative of F in a direction h = (h1, h2) Ω ∈ yields ∂F 1 1 = h1 dW 2 + W 1 dh2. ∂h t t t t 0 0 Clearly, this expression is NOT CONTINUOUS in h w.r.t. the supremum norm.

~~A more suitable space of directions for computing derivatives of stochastic integrals is the Cameron-Martin space~~

~~d 2 d H = h : [0, 1] R : h =0, h abs. contin. with h′ L ([0, 1], R )] . CM → 0 ∈ Recall that HCM is a Hilbert space with inner product~~

~~1 (h, g) = h′ g′ dt, h, g H . H t t ∈ CM 0 2 d The map h h′ is an isometry from H onto L ([0, 1], R ). Moreover, H is → CM CM continuously embedded into Ω, since~~

~~1 1/2 h = sup ht ht′ dt (h, h)H || || t [0,1] | | ≤ 0 | | ≤ ∈ for any h H by the Cauchy Schwarz inequality. ∈ CM UniversityofBonn WinterSemester2012/2013 212 CHAPTER 5. VARIATIONS OF SDE~~

As we will consider variations and directional derivatives in directions in HCM , it is convenient to think of the Cameron-Martin space as a tangent space to Ω at a given path ω Ω. We will now define a gradient corresponding to the Cameron-Martin inner ∈ product in two steps: at first for smooth functions F : Ω R, and then for functions → that are only weakly differentiable in a sense to be specified.

~~h 0 ω~~

~~Gradient and integration by parts for smooth functions~~

Let C1(Ω) denote the linear space consisting of all functions F : Ω R that are b → everywhere Fréchet differentiable with continuous bounded derivative dF : Ω Ω′, → ω d F . Here Ω′ denotes the space of continuous linear functionals l : Ω R → ω → endowed with the dual norm of the supremum norm, i.e.,

~~l ′ = sup l(h): h Ω with h 1 . || ||Ω { ∈ || || ≤ } Deﬁnition (Malliavin Gradient I). Let F C1(R) and ω Ω. ∈ b ∈~~

~~H 1) The H-gradient (D F )(ω) is the unique element in HCM satisfying~~

~~∂F (DHF )(ω), h = (ω) = (d F )(h) for any h H . H ∂h ω ∈ CM (5.16)~~

~~2) The Malliavin gradient (DF )(ω) is the function t (D F )(ω) in L2([0, 1], Rd) → t deﬁned by~~

~~d (D F )(ω) = (DH F )(ω)(t) for a.e. t [0, 1]. (5.17) t dt ∈~~

In other words, DH F is the usual gradient of F w.r.t. the Cameron-Martin inner product, and (DF )(ω) is the element in L2 [0, 1], Rd identiﬁed with (DH F )(ω) by the canoni- Stochastic Analysis Andreas Eberle 5.2. MALLIAVIN GRADIENT AND BISMUT INTEGRATION BY PARTS FORMULA 213

~~2 d cal isometry h h′ between H and L ([0, 1], R ). In particular, for any h H → CM ∈ CM and ω Ω, ∈~~

∂F H (ω) = h, (D F )(ω) = (h′, (DF )(ω)) 2 ∂h H L 1 = h′ (D F )(ω) dt, (5.18) t t 0 and this identity characterizes DF completely. The examples given below should help to clarify the deﬁnitions.

Remark. 1) The existence of the H-gradient is guaranteed by the Riesz Representation The- orem. Indeed, for ω Ω and F C1(Ω), the Fréchet differential d F is a ∈ ∈ b ω continuous linear functional on Ω. Since HCM is continuously embedded into

Ω, the restriction to HCM is a continuous linear functional on HCM w.r.t. the H- H norm. Hence there exists a unique element (D F )(ω) in HCM such that (5.16) holds. 2) By deﬁnition of the Malliavin gradient,

~~1 DH F (ω) 2 = D F (ω) 2 dt. || ||H | t | 0~~

~~3) Informally, one may think of DtF as a directional derivative of F in direction~~

~~I(t,1], because~~

~~1 d H H “ D F = D F (t) = (D F )′ I′ = ∂ F ”. t dt (t,1] I(t,1] 0~~

~~Of course, this is a purely heuristic representation, since I(t,1] is not even continuous.~~

~~Example (Linear functions on Wiener space).~~

1) Brownian motion: Consider the function F (ω) = W i(ω) = ωi , where s (0, 1] s s ∈ and i 1,...,d . Clearly, F is in C1(Ω) and ∈{ } b 1 ∂ i d i i i W = W + εh = h = h′ e I (t) dt ∂h s dε s s ε=0 s t i (0,s) 0 UniversityofBonn WinterSemester2012/2013 214 CHAPTER 5. VARIATIONS OF SDE

~~for any h H . Therefore, by the characterization in (5.18), the Malliavin ∈ CM gradient of F is given by~~

D W i (ω) = e I (t) for every ω Ω and a.e. t (0, 1). t s i (0,s) ∈ ∈ Since the function F : Ω R is linear, the gradient is deterministic. The H- → i gradient is obtained by integrating DWs :

~~t t DH W i = D W i dr = e I = (s t) e . t s r s i (0,s) ∧ i 0 0~~

~~2) Wiener integrals: More generally, let~~

1 F = g dW s s 0 where g : [0, 1] Rd is a C1 function. Integration by parts shows that → 1 F = g W g′ W ds almostsurely. (5.19) 1 1 − s s 0 The function on the right hand side of (5.19) is deﬁned for every ω, and it is Fréchet differentiable. Taking this expression as a pointwise deﬁnition for the stochastic integral F , we obtain

~~∂F 1 1 = g h g′ h ds = g h′ ds ∂h 1 1 − s s s s 0 0 for any h H . Therefore, by (5.18), ∈ CM t H DtF = gt and Dt F = gs ds. 0~~

Theorem 5.7 (Integration by parts, Bismut). Let F C1(Ω) and G 2(Ω ∈ b ∈ La × [0, 1] Rd,P λ). Then → ⊗ 1 1 E DtF Gt dt = E F Gt dWt . (5.20) 0 0 Stochastic Analysis Andreas Eberle 5.2. MALLIAVIN GRADIENT AND BISMUT INTEGRATION BY PARTS FORMULA 215

t To recognize (5.20) as an integration by parts identity on Wiener space let Ht = 0 Gs¸ds. Then 1 D F G dt = DH F,H = ∂ F. t t H H 0 Replacing F in (5.20) by F F with F, F C1(Ω), we obtain the equivalent identity ∈ b 1 E[F ∂H F ] = E[∂H F F ]+ E F F Gt dWt (5.21) − 0 by the product rule for the directional derivative.

~~Proof of Theorem 5.7. The formula (5.21) is an inﬁnitesimal version of Girsanov’s The- orem. Indeed, suppose ﬁrst that G is bounded. Then, by Novikov’s criterion,~~

~~t 1 t ε ε 2 Zt = exp ε Gs dWs Gs ds 0 − 2 0 | | is a martingale for any ε R. Hence for H = t G ds, ∈ t 0 s ε E[F (W + εH)] = E[F (W )Z1].~~

The equation (5.21) now follows formally by taking the derivative w.r.t. ε at ε = 0. Rigorously, we have F (W + εH) F (W ) Zε 1 E − = E F (W ) 1 − . (5.22) ε ε As ε 0, the right hand side in (5.22) converges to E F (W ) t G dW , since → 0 1 1 1 (Zε 1) = ZεG dW G dW in L2(P ). ε 1 − −→ 0 0 Similarly, by the Dominated Convergence Theorem, the left hand side in (5.22) converges to the left hand side in (5.21): 1 ε E (F (W + εH) F (W )) = E (∂H F )(W + sH) ds E[(∂H F )(W )] ε − 0 −→ as ε 0 since F C1(Ω). We have shown that (5.21) holds for bounded adapted G. → ∈ b Moreover, the identity extends to any G 2(P λ) because both sides of (5.21) are ∈ La ⊗ continuous in G w.r.t. the L2(P λ) norm. ⊗ Remark. Adaptedness of G is essential for the validity of the integration by parts identity.

~~UniversityofBonn WinterSemester2012/2013 216 CHAPTER 5. VARIATIONS OF SDE~~

~~Skorokhod integral~~

~~The Bismut integration by parts formula shows that the adjoint of the Malliavin gradient coincides with the Itô integral on adapted processes. Indeed, the Malliavin gradient~~

D : C1(Ω) L2(Ω, ,P ) L2(Ω [0, 1] Rd, ,P λ), b ⊆ A −→ × → A ⊗ B ⊗ F (DtF )0 t 1, −→ ≤ ≤ is a densely deﬁned linear operator from the Hilbert space L2(Ω, ,P ) to the Hilbert A space L2(Ω [0, 1] Rd, ,P λ). Let × → A ⊗ B ⊗ δ : Dom(δ) L2(Ω [0, 1] Rd, B,P λ) L2(Ω, ,P ) ⊆ × → A ⊗ ⊗ −→ A denote the adjoint operator (i.e., the divergence operator corresponding to the Malliavin gradient). By (5.21), any adapted process G 2(Ω [0, 1] Rd, ,P λ) is ∈ L × ∈ A ⊗ B ⊗ contained in the domain of δ, and

1 δG = G dW for any G 2. t t ∈La 0 Hence the divergence operator δ deﬁnes an extension of the Itô integral G 1 G dW → 0 t t to not necessarily adapted square integrable processes G : Ω [0, 1] Rd. This × → extension is called the Skorokhod integral .

Exercise (Product rule for divergence). Suppose that (Gt)t [0,1] is adapted and bounded, ∈ 1 and F Cb (Ω). Prove that the process (F Gt)t [0,1] is contained in the domain of δ, ∈ ∈ and 1 δ(FG) = F δ(G) D F G dt. − t t 0 Deﬁnition of Malliavin gradient II

~~So far we have deﬁned the Malliavin gradient only for continuously Fréchet differentiable functions F on Wiener space. We will now extend the deﬁnition to the Sobolev spaces D1,p, 1~~

~~In particular, we will be interested in the case p =2 where~~

~~1 2 2 2 F 1,2 = E F + DtF dt . || || 0 | | Theorem 5.8 (Closure of the Malliavin gradient).~~

~~1) There exists a unique extension of DH to a continuous linear operator~~

~~DH : D1,p LP (Ω H,P ) −→ → .~~

~~2) The Bismut integration by parts formula holds for any F D1,2. ∈~~

~~1,2 Proof for p =2. 1) Let F D and let (Fn)n N be a Cauchy sequence w.r.t. the (1, 2) ∈ ∈ 1 2 norm of functions in Cb (Ω) converging to F in L (Ω,P ). We would like to deﬁne~~

H H D F := lim D Fn (5.23) n →∞ w.r.t. convergence in the Hilbert space L2(Ω H,P ). The non-trivial fact to be → shown is that DH F is well-deﬁned by (5.23), i.e., independently of the approximating sequence. In functional analytic terms, this is the closability of the operator DH .

To verify closability, we apply the integration by parts identity. Let (Fn) and (Fn) be 2 approximating sequences as above, and let L = lim Fn and L = lim Fn in L (Ω,P ). We have to show L = L. To this end, it sufﬁces to show (L L, h) = 0 almost surely for any h H. (5.24) − H ∈ Hence ﬁx h H, and let ϕ C2(Ω). Then by (5.21), ∈ ∈ b

~~E[(L L, h)H ϕ] = lim E[∂h(Fn Fn) ϕ] n − →∞ − 1 = lim E (Fn Fn)ϕ h′ dW E (Fn Fn)∂hϕ n − − − →∞ 0 = 0~~

UniversityofBonn WinterSemester2012/2013 218 CHAPTER 5. VARIATIONS OF SDE since F F 0 in L2. As C1(Ω) is dense in L2(Ω, ,P ) we see that (5.24) holds. n − n → b A D1,2 2) To extend the Bismut integration by parts formula to functions F let (Fn) be ∈ 1 an approximating sequence of Cb functions w.r.t. the (1, 2) norm. Then for any process G 2 and H = t G ds, we have ∈La t 0 s

~~1 1 H E DtFn Gt dt = E (D Fn,H)H = E Fn G dW . 0 0~~

~~Clearly, both sides are continuous in Fn w.r.t. the (1, 2) norm, and hence the identity extends to F as n . →∞~~

~~The next lemma is often useful to verify Malliavin differentiability:~~

~~2 1,2 Lemma 5.9. Let F L (Ω, ,P ), and let (Fn)n N be a sequence of functions in D ∈ A ∈ converging to F w.r.t. the L2 norm. If~~

~~H 2 sup E[ D Fn H ] < (5.25) n N || || ∞ ∈ D1,2 then F is in , and there exists a subsequence (Fni )i N of (Fn) such that ∈~~

~~1 k F F w.r.t. the (1,2) norm. (5.26) k ni → i=1 The functional analytic proof is based on the theorems of Banach-Alaoglu and Banach- Saks, cf. e.g. the appendix in [27].~~

H 2 Proof. By (5.25), the sequence (D Fn)n N of gradients is bounded in L (Ω H; P ), ∈ → which is a Hilbert space. Therefore, by the Banach-Alaoglu theorem , there exists a H weakly convergent subsequence (D Fki )i N. Moreover, by the Banach-Saks Theorem, ∈ H there exists a subsequence (D Fni )i N of the ﬁrst subsequence such that the averages ∈ 1 k DH F are even strongly convergent in L2(Ω H; P ). Hence the correspond- k i=1 ni → ing averages 1 k F converge in D1,2. The limit is F since F F in L2 and the k i=1 ni ni → D1,2 2 norm is stronger than the L norm.

~~Stochastic Analysis Andreas Eberle 5.3. DIGRESSIONONREPRESENTATIONTHEOREMS 219~~

~~Product and chain rule~~

~~Lemma 5.9 can be used to extend the product and the chain rule to functions in D1,2.~~

~~Theorem 5.10. 1) If F and G are bounded functions in D1,2 then the product FG is again in D1,2, and~~

~~D(FG) = F DG + G DF a.s.~~

~~2) Let m N and F (1),...,F (m) D1,2. If ϕ : Rm R is continuously differen- ∈ ∈ → tiable with bounded derivatives then ϕ(F (1),...,F (m)) is in D1,2, and~~

m ∂ϕ Dϕ(F (1),...,F (m)) = (F (1),...,F (m)) DF (i). ∂x i=1 i Proof. We only prove the product rule, whereas the proof of the chain rule is left as 1 an exercise. Suppose that (Fn) and (Gn) are sequences of Cb functions converging to F and G respectively in D1,2. If F and G are bounded then one can show that the approximating sequences (Fn) and (Gn) can be chosen uniformly bounded. In particular, F G FG in L2. By the product rule for the Fréchet differential, n n → DH (F G ) = F DH G + G DH F for any n N, and (5.27) n n n n n n ∈ DH (F G ) F DHG + G DH F . || n n ||H ≤ | n| || n||H | n| || n||H

~~H 2 Thus the sequence (D (FnGn))n N is bounded in L (Ω H; P ). By Lemma 5.9, we ∈ → conclude that FG is in D1,2 and~~

~~k H 2 1 H D (FG) = L - lim D (Fni Gni ) k k →∞ i=1 for an appropriate subsequence. The product rule for FG now follows by (5.27).~~

~~5.3 Digression on Representation Theorems~~

~~We now prove basic representation theorems for functions and martingales on Wiener space. The Bismut integration by parts identity can then be applied to obtain a more~~

UniversityofBonn WinterSemester2012/2013 220 CHAPTER 5. VARIATIONS OF SDE explicit form of the classical Itô Representation Theorem. Throughout this section, W (ω)= ω denotes the canonical Brownian motion on Wiener space (Ω, ,P ), and t t A = σ(W : s [0, t])P , t 0, Ft s ∈ ≥ is the completed ﬁltration generated by (Wt).

~~Itôs Representation Theorem~~

Itô’s Representation Theorem states that functions on Wiener space that are measurable w.r.t. the Brownian filtration = W,P can be represented as stochastic integrals: Ft Ft Theorem 5.11 (Itô). For any function F 2(Ω, ,P ) there exists a unique process ∈ L F1 G L2(0, 1) such that ∈ a 1 F = E[F ]+ G dW P -almost surely. (5.28) s s 0 An immediate consequence of Theorem 5.11 is a corresponding representation for martingales w.r.t. the Brownian filtration = W,P : Ft Ft Corollary 5.12 (Itô representation for martingales). For any L2-bounded ( ) mar- Ft 2 tingale (Mt)t [0,1] there exists a unique process G La(0, 1) such that ∈ ∈ t M = M + G dW P -a.s. forany t [0, 1]. t 0 s 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 5.11, and then we prove the theorem:~~

2 Proof of Corollary 5.12. If (Mt)t [0,1] is an L bounded ( t) martingale then M1 ∈ F ∈ 2(Ω, ,P ), and L F1 M = E[M ] a.s. for any t [0, 1]. t 1|Ft ∈ Stochastic Analysis Andreas Eberle 5.3. DIGRESSIONONREPRESENTATIONTHEOREMS 221

~~Hence, by Theorem 5.11, there exists a unique process G L2(0, 1) such that ∈ a 1 1 M = E[M ]+ G dW = M + G dW a.s., 1 1 0 0 0 and thus~~

~~t M = E[M ] = M + G dW a.s. for any t 0. t 1|Ft 0 ≥ 0~~

~~Proof of Theorem 5.11. Uniqueness. Suppose that (5.28) holds for two processes G, G ∈ 2 La(0, 1). Then 1 1 G dW = G dW, 0 0 and hence, by Itô’s isometry,~~

~~G G L2(P λ) = (G G) dW = 0. || − || ⊗ − L2(P ) Hence Gt(ω)= Gt(ω) for almost every(t, ω).~~

Existence. We prove the existence of a representation as in (5.28) in several steps − starting with “simple” functions F . 1. Suppose that F = exp(ip (W W )) for some p Rd and 0 s t 1. By t − s ∈ ≤ ≤ ≤ Itô’s formula,

~~1 1 t 1 exp(ip W + p 2t) = exp(ip W + p 2s)+ exp ip W + p 2r ip dW . t 2| | s 2| | r 2| | r s Rearranging terms, we obtain an Itô representation for F with a bounded adapted integrand G. n~~

2. Now suppose that F = Fk where Fk = exp ipk (Wtk Wtk−1 ) for some k=1 − d n N, p1,...,pn R , and 0 t0 t1 tn 1. Denoting by Gk the bounded ∈ ∈ ≤ ≤ ≤≤ ≤ adapted process in the Itô representation for Fk, we have

~~n tk+1 k F = E[Fk]+ G dW . tk k=1 UniversityofBonn WinterSemester2012/2013 222 CHAPTER 5. VARIATIONS OF SDE~~

n We show that the right hand sidecan be written as the sum of k=1 E[Fk] and a stochastic integral w.r.t. W . For thispurpose, it sufﬁces toverify that the productof two stochastic integrals X = t G dW and Y = t H dW with bounded adapted processes G t 0 t 0 and H is the stochastic integral of a process in L2(0, 1) provided 1 G H dt =0. This a 0 t t holds true, since by the product rule,

~~1 1 1 X Y = X H dW + Y G dW + G H dt, 1 1 t t t t t t t t 0 0 0 and XH + YG 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 2(Ω, ,P ), we ﬁrst note L F1 that the linear combinations of the functions in Step 2 form a dense subspace of the Hilbert space L2(Ω, ,P ). Indeed, if ϕ is an element in L2(Ω, ,P ) that is orthogonal F1 F1 to this subspace then

n E ϕ exp(ip (W W )) = 0 k tk − tk−1 k=1 for any n N, p ,...,p Rd and 0 t t t 1. By Fourier inversion, ∈ 1 n ∈ ≤ 0 ≤ 1 ≤≤ n ≤ this implies E[ϕ σ(W W :1 k n)] = 0 a.s. | tk − tk−1 ≤ ≤ for any n N and 0 t t 1, and hence ϕ = 0 a.s. by the Martingale ∈ ≤ 0 ≤≤ n ≤ Convergence Theorem. Now ﬁx an arbitrary function F L2(Ω, ,P ). Then by Step 3, there existsa sequence ∈ F1 (F ) of functions in L2(Ω, ,P ) converging to F in L2 that have a representation of n F1 the form 1 F E[F ] = G(n) dW (5.29) n − n 0 with processes G(n) L2(0, 1). As n , ∈ a →∞

~~F E[F ] F E[F ] in L2(P ). n − n −→ − Stochastic Analysis Andreas Eberle 5.4. FIRST APPLICATIONS TO SDE 223~~

~~Hence, by (5.29) 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~~

~~1 F E[F ] = G dW − 0 by taking the L2 limit on both sides of (5.29).~~

~~Clark-Ocone formula~~

~~If F is in D1,2 then the process G in the Itô representation can be identiﬁed explicitly:~~

Theorem 5.13 (Clark-Ocone). For any F D1,2, ∈ 1 F E[F ] = G dW − 0 where G = E[D F ]. t t |Ft Proof. It remains to identify the process G in the Itô representation. We assume w.l.o.g. that E[F ]=0. Let H L1([0, 1], Rd). Then by Itô’s isometry and the integration by ∈ a parts identity,

~~1 1 1 1 E Gt Ht dt = E G dW H dW = E DtF Ht dt 0 0 0 0 1 = E E[DtF t] Ht dt 0 |F for all Setting H := G E[D F ] we obtain t t − t |Ft G (ω) = E[D F ](ω) P λ a.e. t t |Ft ⊗ −~~

~~5.4 First applications to stochasticdifferential equations~~

~~5.5 Existence and smoothness of densities~~

~~UniversityofBonn WinterSemester2012/2013 Index~~

α-stable, 35 geometric Lévy processe, 90 associative law, 79 infinitely divisible, 7 integration by parts, 42 Bismut-Elworthy Formula, 181 inverse Gaussian subordinator, 33 Brownian bridge, 133 Itô formula càdlàg, 6 for finite variation functions, 44 Cameron-Martin space, 138 Itô isometry, 78, 91, 93 chain rule, 44 Itô representation theorem, 191 change of variables, 44 Itô’s formula characteristic exponent, 11 for semimartingales, 80 Clark-Ocone formula, 193 Kunita-Watanabe characterization, 76 cocycle property, 170 Kunita-Watanabe inequality, 96 compensated Poisson point process, 23 compensated Poisson process, 9 Lévy measure, 16 compound Poisson distribution, 15 Lévy subordinator, 33 compound Poisson process, 13 Lévy-Itô decomposition, 30 Lévy-Khinchin formula, 30 derivative flow, 177 local martingale, 63 Doléans measure, 92 localizing sequence, 63 Doléans-Dade theorem, 88 Malliavin Gradient, 183 Fréchet differentiable, 182 mapping theorem, 20 Gamma process, 32 Markov property, 132 Gaussian Lévy processes, 31 martingale

~~224 INDEX 225~~

local martingale, 63 strict local, 67 strict local martingale, 63 pinned Wiener measure, 133 Poisson point process, 19 Poisson process, 9 Poisson random measure, 9 predictable, 77, 98 product rule, 42 quadratic variation, 68

Schilder’s large derivation principle lower bound, 141 upper bound, 142 semimartingale, 55 Skorokhod integral, 187 spatial Poisson process, 9 strict local martingale, 63 subordinator, 31 ucp-convergence, 67

~~Variance Gamma process, 34~~

~~UniversityofBonn WinterSemester2012/2013 Bibliography~~

~~[1]~~

~~[2]~~

~~[3] David Applebaum. Lévy Processes and Stochastic Calculus. Cambridge UP, Cam- bridge, 2004.~~

~~[4] Ole E. Barndorff-Nielsen. Processes of normal inverse Gaussian type. Finance Stoch., 2(1):41–68, 1998.~~

~~[5] Richard F. Bass. Stochastic Processes. Cambridge University Press, Cambridge, 2011.~~

~~[6] Jean Bertoin. Lévy Processes. Cambridge UP, Cambridge, 1998.~~

~~[7] Jean-Michel Bismut. Large deviations and the Malliavin calculus. Birkhaeuser, Boston, 1994.~~

~~[8] Amir Dembo and Ofur Zeitouni. Large Deviations Techniques and Applications. Springer, 1998, Berlin.~~

~~[9] Frank den Hollander. Large Deviations. American Mathematical Society, Provi- dence, 2008.~~

~~[10] Wolfgang Doeblin and Marc Yor. pli cacheté.~~

~~[11] Andreas Eberle. Markov processes. http://www.uni-bonn.de/~eberle/MarkovProcesses/MPSkript.pdf, 2009.~~

~~226 BIBLIOGRAPHY 227~~

~~[12] Andreas Eberle. Wahrscheinlichkeitstheorie und Stochastische Prozesse. http://www.uni-bonn.de/~eberle/StoPr/WthSkriptneu.pdf, 2010.~~

~~[13] Andreas Eberle. Introduction to Stochastic Analysis. http://www.uni-bonn.de/~eberle/StoAn1011/StoAnSkriptneu.pdf, 2011.~~

~~[14] David Elworthy. On the Geometry of Diffusion Operators and Stochastic Flows. Springer, 1999, Berlin.~~

~~[15] Stewart N. Ethier and Thomas G. Kurtz. Markov Processes. Wiley-Interscience, Hoboken, NJ, 2005.~~

~~[16] William Feller. Introduction to Probability Theory and Its Applications. Wiley, New York, 1957.~~

~~[17] Peter K. Friz and Nicolas Victoir. Multidimensional Stochastic Processes as Rough Paths. Cambridge UP, Cambridge, 2010.~~

~~[18] W. Hackenbroch and A. Thalmaier. Stochastic Analysis. Teubner, 1991.~~

~~[19] Martn Hairer. On malliavin’s proof of hörmander’s theorem. Bull. Sci. Math., 135(6-7):650–666, 2011.~~

~~[20] Elton Hsu. Stochastic Analysis on Manifolds. Oxford University Press, Oxford, 2002.~~

~~[21] Nobuyuki Ikeda and Takehiko Ed. Watanabe. Stochastic Differential Equations and Diffusion Processes. North-Holland, Amsterdam, 1981.~~

~~[22] Jean Jacod and Albert Shiryaev. Limit Theorems for Stochastic Processes. Springer, Berlin, 2003.~~

~~[23] Ioannis Karatzas and Steven E. Shreve. Brownian motion and stochastic calculus. Springer, New York, 2010.~~

~~UniversityofBonn WinterSemester2012/2013 228 BIBLIOGRAPHY~~

~~[24] Hiroshi Kunita. Stochastic Flows and Stochastic Differential Equations. Cam- bridge University Press, Cambridge, 1997.~~

~~[25] Andreas Kyprianou. Introductory Lectures on Fluctuations of Lévy Processes with Applications. Springer, Berlin, 2006.~~

~~[26] Terry J. Lyons. Differential Equations Driven by Rough Paths, Ecole d’Eté de Probabilités de Saint-Flour XXXIV. Springer, 2004.~~

~~[27] Zhi-Ming Ma and Michael Röckner. Introduction to the Theory of (Non- Symmetric) Dirichlet Forms. Springer, Berlin, 1992.~~

~~[28] Dilip B. Madan and Eugene Seneta. The variance gamma model for share market returns. Journal of Business, 63(4):511–524, 1990.~~

~~[29] Paul Malliavin. Stochastic Analysis. Springer, Berlin, 2002.~~

~~[30] David Nualart. The Malliavin Calculus and Related Topics. Springer, Berlin, 2006.~~

~~[31] Giuseppe Da Prato. Introduction to Stochastic Analysis and Malliavin Calculus. Edizioni della Normale, Pisa, 2009.~~

~~[32] Philip E. Protter. Stochastic Integration and Differential Equations. Springer, Berlin, 2005.~~

~~[33] Daniel Revuz and Marc Yor. Continuous martingales and Brownian motion. Springer, Berlin, 2005.~~

~~[34] L. C. G. Rogers and David Williams. Diffusions, Markov processes and martingales, Vol. 2: Ito calculus. Cambridge UP, Camebridge, 2008.~~

~~[35] Shreve. Stochastic calculus for Finance II. Springer, New York, 2004.~~

~~[36] Daniel W. Stroock and S.R. S. Varadhan. Multidimensional Diffusion Processes. Springer, Berlin, 2005.~~

~~Stochastic Analysis Andreas Eberle BIBLIOGRAPHY 229~~

~~[37] Shinzo Watanabe. Lectures on stochastic differential equations and Malliavin calculus. Springer, Berlin, 1984.~~

~~UniversityofBonn WinterSemester2012/2013~~