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 .

Isonormal Gaussian process

Definition 1.1.1. (Gaussian process on H or isonormal Gaussian process)

Stochastic process W={W(h):h∈H} defined on a complete probability space (Ω,F,P) is a Gaussian process on H if W is a centered Gaussian family r.v. and for all h,g∈H , we have E(W(h)W(g))=⟨h,g⟩H

The mapping h↦W(h) can be shown linear (a.s.). As one can expand and compute E[(W(λh+μg)−λW(h)−μW(g))2]=0 using E(W(h)W(g))=⟨h,g⟩H . Since the bilinearity of inner product ⟨⋅,⋅⟩H , we have E(W(h+k)W(g))=⟨h+k,g⟩H=E((W(h)W(g)))+W(W(k)W(g)) , and E(W(h)W(g+k))=E(W(h)W(g))+E(W(h)W(k)) , thus we have

E[(W(λh+μg)−λW(h)−μW(g))2]=E[W(λh+μg)2]+E[(λW(h)+μW(g))2]−2E[(W(λh+μg))(λW(h)+μW(g))]=0.

This shows that h↦W(h) is an linear isometry from H onto a closed subspace of L2(Ω,F,P) , since linear isometry preserves completeness, L2 space is complete, and thus yields W(H) is closed.

Existance of such probability space. The Kolmogrov extention theorem states that given some probability space (Ω,F,P) , for all k∈N and finite sequence h1,…,hk∈H , then there exists a probability measure on product σ -algebra of ΩH if the measure νh1,…,hk on Ωk satisfies

  1. (Permutation Invariant) Given arbitrary permutation π , k∈N and measurable sets Fhi⊂Ω,i=1,…,k , we have
νh1,…,hk(Fh1,…,Fhk)=νπ(h1),…,π(hk)(Fπ(h1),…,Fπ(hk)),

this means the measure persists when input indices and subscripts are jointly permuted. 2. (Marginal Consistency) Given k,m∈N and measurable sets Fhi⊂Ω,i=1,…,k , we have

νh1,…,hk(Fh1,…,Fhk)=νh1,…,hk,hk+1,…,hk+m(Fh1,…,Fhk,Ω,…,Ω⏟m times),

this tells that the measure is consistent when marginalized out some of its indices.

Then we check whether the isonormal Gaussian process matches the requirements of Kolmogrov extention theorem. Arbitrarily choose h1,…,hk∈H , construct the Gram matrix Σ whose indices Σij=⟨hi,hj⟩H . The Gram matrix Σ is symmetric and positive-semidefinite, which corresponds to a Gaussian distribution N(0,Σ) , and we use this distribution to denote the distribution of (W(h1),…,W(hk)) . We check symmetry conditions first. Given permutation π , there is a corresponding permutation matrix Pπ and the permutated distribution on permutated bases is N(0,PπΣPπ⊤) . To check the equivalence of two Gaussian measures, it suffices to prove the two covariance matrices is identical under the same set of bases. For permuted distribution we have x′⊤PπΣPπ⊤x′=x⊤Pπ⊤PπΣPπ⊤Pπx=x⊤Σx , this shows that the measure is invariant under permutation. For marginal consistency, it suffices to prove that given k,m∈N and h1,…,hk+m , the top left sub-matrix Σk of the big covariance matrix Σk+m corresponds to the very covariance matrix of (W(h1),…,h(hk)) , which is obvious. By Kolmogrov extention theorem, such a measure ν DOES exists.

We use Kolmogrov extention theorem to check the well-definiteness of isonormal Gaussian process. A core block of the definition is the covariance structure.

Relation to Ito Brownian Motion and Ito Stochastic Integral. Recall that for one-dimensinoal Brownian motion W(⋅) , we have E(W(s)W(t))=s∧t , which is symmetric and semidefinite and corresponds to the Cameron-Martin space HCM , which is defined as

HCM={f:[0,T]→R:f(0)=0,f abs. cont.,f′∈L2[0,T]}

and the inner product is ⟨f,g⟩HCM:=∫0Tf′(s)g′(s)ds . Consider the reproducing kernel K(s,t)=s∧t , construct Kt(⋅)=K(⋅,t) , whose weak derivative is 1[0,t] . Introduce the Ito stochastic integral operator W˙:L2[0,T]→L2(R) , Brownian motion B(t) could be expressed as the stochastic integral of identity function 1[0,t] :

W˙(h)=∫0ThdBt,

therefore one could rewrite B(t) as W˙(1[0,t]) . This setting corresponds with the definition of isonormal Gaussian process:

E[W(s)W(t)]=E[W˙(1[0,s])W˙(1[0,t])]=⟨1[0,s],1[0,t]⟩L2=s∧t.

This reveals that W˙ is the isonormal Gaussian process on [0,T] , it also behaves as the isometry, which is the well-known Ito isometry:

E[W˙2(h)]=‖h‖L2.

We seek to decompose L2(Ω,F,P) into orthogonal sum of subspaces, as Taylor expansion allows us to rewrite analytic function into series of functions, we look into Hermite polynomials Hn(x) for similar building blocks in the sense of Ito integration.