DISCLAIMER

This note is based on Malliavin Calculus and Related Topics by David Nualart and is intended solely for personal academic study and should not be regarded as a substitute for the original text.

1 Wiener Chaos Decomposition

We seek a generalized form of Gaussian process, say changing the set of time indices T to a "more abstract" space H . Suppose H is a real separable Hilbert space with innrt product ⟨⋅,⋅⟩H and norm ‖h‖H for h∈H .

Hermite Polynomials

Definition 1.1.2. Hermite Polynomials Hn(x) Hermite Polynomialsn show up when we expand F(x,t)=exp⁡(tx−t22) w.r.t. t :

F(x,t)=exp⁡[x22−12(x−t)2]=exp⁡(x22)∑n=0∞tnn![dndtnexp⁡(−(x−t)22)]t=0=exp⁡(x22)∑n=0∞tn⋅(−1)nn!dndxnexp⁡(−x22)x=x−t=∑n=0∞tn⋅Hn(x).

Lemma 1.1.3. Properties of Hermite Polynomials The Hermite polynomials admit the following properties:

Hn′(x)=Hn−1(x)n⩾1,(n+1)Hn+1(x)=xHn(x)−Hn−1(x),n⩾1,Hn(−x)=(−1)nHn(x),n⩾1.

Proof.

We first prove Hn′(x)=Hn−1(x)

Hn′(x)=ddx{(−1)nn!exp⁡(x22)dndxnexp⁡(−x22)}=(−1)nn!exp⁡(x22)[x⋅dndxnexp⁡(−x22)+dn+1dxn+1exp⁡(−x22)]=(−1)nn!exp⁡(x22)[x⋅dndxnexp⁡(−x22)−xdndxnexp⁡(−x22)−ndn−1dxn−1exp⁡(−x22)]=(−1)n−1(n−1)!dn−1dxn−1exp⁡(−x22)=Hn−1(x).

Note that when proving Hn′(x)=Hn−1(x) , by Leibniz formula, we have

(−1)nn!exp⁡(x22)dn+1dxn+1exp⁡(−x22)=(−1)nn!exp⁡(x22)[−xdndxnexp⁡(−x22)−ndn−1dxn−1exp⁡(−x22)]⟺−(n+1)Hn+1(x)=−xHn(x)+Hn−1(x).

The last statement is straight forward, by performing y=−x , we have ∂y∂x=−1 , therefore ∂∂x=−∂∂y . Applying n times yields ∂n∂xn=(−1)n∂n∂yn and Hn(−x)=(−1)nHn(x) is immediate. ◻

One may also prove using properties of F in definition 1.1.2, as ∂F∂x=tF,∂F∂t=(x−t)F and F(−x,t)=F(x,−t) .

The following lemma shows that Hermite polynomial acting on two standard centered Gaussian variable yields orthogonal-like result.

Lemma 1.1.4. (Orthogonality of centered Gaussian variables acted by Hermite polynomials) Let X,Y be two random variable with joint Gaussian distribution s.t. E(X)=E(Y)=0 and E(X2)=E(Y2)=1 . Then for all n,m⩾0 , we have

E(Hn(X)Hm(Y))=δmn1n![E(XY)]n

Proof. Note that F(x,t)=exp⁡(tx−t22) could be rewritten to series related to Hn(x) , we consider taking expectation of F(X,t) and F(Y,t) . On the one hand, we have

E[F(X,s)F(Y,t)]=E[exp⁡(sX−s22)exp⁡(tY−t22)]=exp⁡(−s2+t22)E[exp⁡(sX+tY)]Z=sX+tY is centered Gaussian=exp⁡(−s2+t22)∫−∞+∞12πσexp⁡(z−z22σ)dz=exp⁡(−s2+t22+σ2)∫−∞+∞12πσexp⁡(−(z−σ)22σ)dz=exp⁡(−s2+t22)exp⁡(stE[XY])σ=E[Z2]=2stE[XY]=exp⁡(−s2+t22)∑k=0∞(stE[XY])nn!

On the other hand, we have

E[F(X,s)F(Y,t)]=E{[∑n=0∞sn⋅Hn(X)][∑m=0∞tm⋅Hm(Y)]}=E[∑m,n=0∞smtnHn(X)Hm(Y)]=∑m,n=0∞smtnE[Hn(X)Hm(Y)]

Comparing the order of smtn results in what we need to prove. ◻

In the following we will show that random variables of the form Hn(W(h)) spans the space L2(Ω,F,P) . Let Hn be the closed linear subspace generated by the random variables {Hn(W(h)):h∈H,‖h‖H=1} . We force the norm of h to 1 , since by the linearity of W(⋅) any W(h),h∈H is equal to cW(h′) in which c is a constant and h′ has norm 1 . As W(h) is a centered Gaussian variable, using lemma 1.1.4, we have Hn⊥Hm if n≠m . To show L2(Ω,F,P)=⨁n=1∞Hn , we only need to prove that "if X is perpendicular to all subspaces Hn , X shall only be zero", otherwise X must be a linear combination of the vectors in Hn s. Formally, "pick X∈L2(Ω,F,P) such that for all n⩾0 and all uniform vector h∈H , we have E[XHn(W(h))]=0 , this would result in X=0 ". This is equivalent to " E[X(W(h))n]=0 implies x=0 ". Further more, we shall prove " E[Xexp⁡(W(h))]=0 implies X=0 ", and this gives up to the following lemma.

Lemma 1.1.5. Random variables {exp⁡(W(h)):h∈H} form a total(dense) subset of L2(Ω,G,P) , in which G is the σ -algebra generated by {W(h):h∈H} .

Proof. To prove {exp⁡(W(h)):h∈H} forms a dense subset of the Hilbert space L2(Ω,G,P) , we only need to prove the set of vectors spans the total space. This is equivalent to the statement:

⟨X,exp⁡(W(h))⟩=0 for all h∈H⟺X=0.

Therefore pick X∈L2(Ω,G,P) such that for all h∈H , E[Xexp⁡(W(h))]=0 . Fix m and pick h1,…,hm∈H and let h∈∑i=1mtihi , the linearity tells E[Xexp⁡(∑i=1mtiW(hi))]=0 . To prove X=0 , it suffices to prove ∫GXdP=E[X1G]=0 for all G∈G , which is equivalent to showing E[X1B(W(h1),…,W(hm))]=0 in which B∈B(Rm) , as G are generated by cylinder sets {ω∈Ω:(W(h1)(ω),…,W(hm)(ω))∈B,B∈B(Rm)} . Note that E[X1B(W(h1),…,W(hm))] could be interpreted as a signed measure on (Rm,B(Rm)) :

ν(B)=E[X1B(W(h1),…,W(hm))]

note that ∫1Bdν=ν(B)=E[X1B(W(h1),…,W(hm))] one could prove that ∫fdν=E[Xf(W(h1),…,W(hm))] by linearity and monotonic convergence theorem. Therefore we consider the Laplace transform to get an exponential in our integral and exactly meets the result we just arrived at:

ν~(t)=∫Rme⟨t,y⟩dν(y)=E[Xexp⁡(∑i=1mtiW(hi))]=0.

As B∈B is randomly chosen, by the injectivity of Laplace transform, we have ν(B)=E[X1B(W(h1),…,W(hm))]≡0 and this results holds for all m and all h1,…,hm∈H . Therefore E[X1A]≡0 for all cylinder set A . Finally by the monotone class theorem, we have E[X,1G]=0 for all G∈G . ◻

With lemma 1.1.5, we are able to prove the Wiener chaos expansion theorem.