Ornstein–Uhlenbeck operator should not be confused with Ornstein–Uhlenbeck process.
In mathematics, the Ornstein–Uhlenbeck operator is a generalization of the Laplace operator to an infinite-dimensional setting. The Ornstein–Uhlenbeck operator plays a significant role in the Malliavin calculus.
Consider the gradient operator ∇ acting on scalar functions f : Rn → R; the gradient of a scalar function is a vector field v = ∇f : Rn → Rn. The divergence operator div, acting on vector fields to produce scalar fields, is the adjoint operator to ∇. The Laplace operator Δ is then the composition of the divergence and gradient operators:
\Delta=div\circ\nabla
acting on scalar functions to produce scalar functions. Note that A = -Δ is a positive operator, whereas Δ is a dissipative operator.
Using spectral theory, one can define a square root (1 - Δ)1/2 for the operator (1 - Δ). This square root satisfies the following relation involving the Sobolev H1-norm and L2-norm for suitable scalar functions f:
\|f
| 2 | |
| \| | |
| H1 |
=\|(1-\Delta)1/2f
| 2 | |
| \| | |
| L2 |
.
Often, when working on Rn, one works with respect to Lebesgue measure, which has many nice properties. However, remember that the aim is to work in infinite-dimensional spaces, and it is a fact that there is no infinite-dimensional Lebesgue measure. Instead, if one is studying some separable Banach space E, what does make sense is a notion of Gaussian measure; in particular, the abstract Wiener space construction makes sense.
To get some intuition about what can be expected in the infinite-dimensional setting, consider standard Gaussian measure γn on Rn: for Borel subsets A of Rn,
\gamman(A):=\intA(2\pi)-n/2\exp(-|x|2/2)dx.
This makes (Rn, B(Rn), γn) into a probability space; E will denote expectation with respect to γn.
The gradient operator ∇ acts on a (differentiable) function φ : Rn → R to give a vector field ∇φ : Rn → Rn.
The divergence operator δ (to be more precise, δn, since it depends on the dimension) is now defined to be the adjoint of ∇ in the Hilbert space sense, in the Hilbert space L2(Rn, B(Rn), γn; R). In other words, δ acts on a vector field v : Rn → Rn to give a scalar function δv : Rn → R, and satisfies the formula
E[\nablaf ⋅ v]=E[f\deltav].
On the left, the product is the pointwise Euclidean dot product of two vector fields; on the right, it is just the pointwise multiplication of two functions. Using integration by parts, one can check that δ acts on a vector field v with components vi, i = 1, ..., n, as follows:
\deltav(x)=
| n | |
| \sum | |
| i=1 |
\left(xivi(x)-
| \partialvi | |
| \partialxi |
(x)\right).
The change of notation from "div" to "δ" is for two reasons: first, δ is the notation used in infinite dimensions (the Malliavin calculus); secondly, δ is really the negative of the usual divergence.
The (finite-dimensional) Ornstein–Uhlenbeck operator L (or, to be more precise, Lm) is defined by
L:=-\delta\circ\nabla,
with the useful formula that for any f and g smooth enough for all the terms to make sense,
\delta(f\nablag)=-\nablaf ⋅ \nablag-fLg.
The Ornstein–Uhlenbeck operator L is related to the usual Laplacian Δ by
Lf(x)=\Deltaf(x)-x ⋅ \nablaf(x).
Consider now an abstract Wiener space E with Cameron-Martin Hilbert space H and Wiener measure γ. Let D denote the Malliavin derivative. The Malliavin derivative D is an unbounded operator from L2(E, γ; R) into L2(E, γ; H) - in some sense, it measures "how random" a function on E is. The domain of D is not the whole of L2(E, γ; R), but is a dense linear subspace, the Watanabe-Sobolev space, often denoted by
D1,2
Again, δ is defined to be the adjoint of the gradient operator (in this case, the Malliavin derivative is playing the role of the gradient operator). The operator δ is also known the Skorokhod integral, which is an anticipating stochastic integral; it is this set-up that gives rise to the slogan "stochastic integrals are divergences". δ satisfies the identity
E[\langleDF,v\rangleH]=E[F\deltav]
for all F in
D1,2
Then the Ornstein–Uhlenbeck operator for E is the operator L defined by
L:=-\delta\circD.