TheoremBase

Relative Entropy along a Noise-Gradient Perturbation of the Identity: the Formula, the Second-Order Expansion and the First Variation

Pushing a measure of finite relative entropy with respect to the diagonal Gaussian measure forward by the identity plus a small multiple of the noise gradient of a bounded C2C^2 cylindrical function keeps the relative entropy finite, changes it by an explicit integral, and differentiates at zero to an Ornstein-Uhlenbeck functional.

Statement

In the setting of A Diagonal Gaussian Reference Measure on the Noise Wasserstein Space, Rescaled Heads and Gaussian Tails: Standing Notation, let μ∈P(X)\mu\in\mathcal{P}(X) have finite relative entropy with respect to γc\gamma_{c}; then μ∈P2(X)\mu\in\mathcal{P}_{2}(X), the set of The Second Moment of a Borel Probability Measure on a Hilbert Space and the Probability Measures with Finite Second Moment §space, by Relative Entropy with Respect to a Diagonal Gaussian Measure on a Hilbert Space: the Moment Bound, the Cutoff Projections, and Bounded, Tight, Weakly Closed, Wasserstein-Closed and Weakly Sequentially Compact Sublevel Sets §moment. Let n∈Nn\in\mathbb{N} and let g∈Cb2(Rn)g\in C^{2}_{b}(\mathbb{R}^{n}), the set of Bounded Twice Continuously Differentiable Functions with Bounded First and Second Partial Derivatives on Euclidean Space §bounded with q=nq=n, with partial derivatives ∂kg\partial_{k}g and ∂j∂ig\partial_{j}\partial_{i}g as in Differential Calculus and Convexity on Euclidean Open Sets: Standing Notation §derivatives, and let Lμa(g)L^{a}_{\mu}(g) be the noise Ornstein-Uhlenbeck functional of μ\mu at gg. A function of class C2C^{2} on Rn\mathbb{R}^{n} is of class C1C^{1} there by clause 2 of C^k Maps on a Euclidean Open Set, so gg belongs to the set Cb1(Rn)C^{1}_{b}(\mathbb{R}^{n}) of Bounded Continuously Differentiable Functions with Bounded Partial Derivatives on Euclidean Space §bounded, and ψ=g∘pn\psi=g\circ p_{n} belongs to FCb2(X)\mathcal{F}C^{2}_{b}(X) and to FCb1(X)\mathcal{F}C^{1}_{b}(X), the sets of Bounded C^2 Cylindrical Functions on a Hilbert Space with an Orthonormal Basis §cylindrical and Bounded C^1 Cylindrical Functions on a Hilbert Space with an Orthonormal Basis §cylindrical; its noise gradient is ∇aψ(x)=∑k=1nak ∂kg(pn(x)) ek\nabla_{a}\psi(x)=\sum_{k=1}^{n}a_{k}\,\partial_{k}g(p_{n}(x))\,e_{k} by that clause and Bounded C^1 Cylindrical Functions: Linear Structure, the Gradient and the Partial Derivatives, Bounds and Integrability, and Density in the Square-Integrable Functions §partial. For t∈Rt\in\mathbb{R}, id+t∇aψ\mathrm{id}+t\nabla_{a}\psi is the map x↦x+t ∇aψ(x)x\mapsto x+t\,\nabla_{a}\psi(x). Let B,b∈RB,b\in\mathbb{R} be nonnegative with ∣∂kg(u)∣≤B|\partial_{k}g(u)|\le B and ai1/2aj1/2∣∂j∂ig(u)∣≤ba_{i}^{1/2}a_{j}^{1/2}|\partial_{j}\partial_{i}g(u)|\le b for all u∈Rnu\in\mathbb{R}^{n} and i,j,k∈[n]i,j,k\in[n]. For x∈Xx\in X let M(x)M(x) be the real n×nn\times n matrix with entry ai1/2aj1/2 ∂j∂ig(pn(x))a_{i}^{1/2}a_{j}^{1/2}\,\partial_{j}\partial_{i}g(p_{n}(x)) in row ii and column jj, and let InI_{n} be the identity matrix; det⁡\det is the determinant, log⁡\log the natural logarithm, and θn\theta_{n} and KnK_{n} are fixed constants as provided by Determinants of Positive Definite Matrices: Positivity, the Bound log⁡det⁡A≤tr A−d\log\det A\le\mathrm{tr}\,A-d, Bounds under Pinching, and the Expansion of det⁡(I+tB)\det(I+tB) §expansion with d=nd=n, where they are written cdc_{d} and KdK_{d}.

1. (The relative entropy of the push-forward) Let t∈Rt\in\mathbb{R} satisfy 2n ∣t∣ b≤12n\,|t|\,b\le1. Then 0<det⁡(In+tM(x))0<\det(I_{n}+tM(x)) for every x∈Xx\in X, the function x↦log⁡det⁡(In+tM(x))x\mapsto\log\det(I_{n}+tM(x)) is Borel and bounded, the map id+t∇aψ\mathrm{id}+t\nabla_{a}\psi is Borel, the measure (id+t∇aψ)#μ(\mathrm{id}+t\nabla_{a}\psi)_{\#}\mu has finite relative entropy with respect to γc\gamma_{c}, and

H((id+t∇aψ)#μ ∣ γc)=H(μ ∣ γc)+∑k=1n∫X(t ak xkck ∂kg(pn(x))+t2ak22ck ∂kg(pn(x))2)μ(dx)−∫Xlog⁡det⁡(In+tM(x)) μ(dx).H\bigl((\mathrm{id}+t\nabla_{a}\psi)_{\#}\mu\,\big|\,\gamma_{c}\bigr)=H(\mu\,|\,\gamma_{c})+\sum_{k=1}^{n}\int_{X}\Bigl(\frac{t\,a_{k}\,x_{k}}{c_{k}}\,\partial_{k}g(p_{n}(x))+\frac{t^{2}a_{k}^{2}}{2c_{k}}\,\partial_{k}g(p_{n}(x))^{2}\Bigr)\mu(dx)-\int_{X}\log\det\bigl(I_{n}+tM(x)\bigr)\,\mu(dx).

2. (Second-order expansion) Let t∈Rt\in\mathbb{R} satisfy 2n ∣t∣ b≤12n\,|t|\,b\le1 and ∣t∣ b≤θn|t|\,b\le\theta_{n}. Then

∣H((id+t∇aψ)#μ ∣ γc)−H(μ ∣ γc)−t Lμa(g)∣≤t2(Kn b2+B2∑k=1nak22ck).\Bigl|H\bigl((\mathrm{id}+t\nabla_{a}\psi)_{\#}\mu\,\big|\,\gamma_{c}\bigr)-H(\mu\,|\,\gamma_{c})-t\,L^{a}_{\mu}(g)\Bigr|\le t^{2}\Bigl(K_{n}\,b^{2}+B^{2}\sum_{k=1}^{n}\frac{a_{k}^{2}}{2c_{k}}\Bigr).

3. (First variation) There is a positive real number t0t_{0} such that (id+t∇aψ)#μ(\mathrm{id}+t\nabla_{a}\psi)_{\#}\mu has finite relative entropy with respect to γc\gamma_{c} for every tt in the open interval (−t0,t0)(-t_{0},t_{0}), and the function t↦H((id+t∇aψ)#μ ∣ γc)t\mapsto H\bigl((\mathrm{id}+t\nabla_{a}\psi)_{\#}\mu\,|\,\gamma_{c}\bigr) on (−t0,t0)(-t_{0},t_{0}) is differentiable at 00, in the sense of Derivative at an Interior Point, with derivative Lμa(g)L^{a}_{\mu}(g).

Proofs

Log in to submit a proof.

Loading...

Citations

Loading…

Dependencies

Loading…

Related

0 relations

Curated associations between results. These are editable and subjective — they do not replace the dependency graph, which is derived from the references in the text.

No relations recorded yet.

Comments

Log in to comment.

Loading…