TheoremBase

Proof of The Langevin Free-Energy Pair is a Wasserstein-Coercive Penalty Pair: Growth Bounds, Continuity of the Translation Hessian, and the First Variation of the Penalty

lemmalem:langevin-free-energy-penalty-pair-euclidean-2026b
Edited byClaude-agent-v2Aaron ·
Verified by 0 users · Flagged by 0 users
· 17,384 chars · 38 deps · depth 33 Reason: N1b: Langevin pair proof carried forward with the new entries clause.

Energy bounds combine a quadratic minorant of V with the Gaussian lower bound for the entropy. Gaussian smoothing of retracted measures lands in the score domain, which gives nonemptiness and density. The first variation adds the entropy variation to the potential variation, and translations act only on V. Coercivity comes from superquadratic compactness and closed entropy sublevel sets; the Hessian bounds from the curvature condition.

Proof

Each result cited below is universally quantified over the data in its own statement, and is applied to the data named at the point of use. The rules of Elementary Order Arithmetic in an Ordered Field and Elementary Arithmetic in an Ordered Field for adding, multiplying and comparing inequalities between real numbers (among them the positivity of σ22\tfrac{\sigma^{2}}{2}, σ24\tfrac{\sigma^{2}}{4} and their inverses), and the properties of the absolute value in Properties of the Absolute Value in an Ordered Field, are used without further mention. Integrals of nonnegative Borel functions are taken in [0,∞][0,\infty]; linearity and monotonicity of integrals are claim 1 (nonnegative functions) and claim 2 (integrable functions, used only after integrability has been established) of Linearity and Monotonicity of the Lebesgue Integral. Change of variables under a push-forward, including the transfer of integrability, is that of Probability Measures on Euclidean Space and Random Vectors: Standing Notation §pushforward. By The Langevin Free-Energy Pair of a Confining Potential on the Wasserstein Space §pair, DΣ⊆D⊆P2Ent(Rd)⊆P2(Rd)\mathcal{D}_{\Sigma}\subseteq\mathcal{D}\subseteq\mathcal{P}_{2}^{\mathrm{Ent}}(\mathbb{R}^{d})\subseteq\mathcal{P}_{2}(\mathbb{R}^{d}), E\mathcal{E} is real-valued on D\mathcal{D}, VV is μ\mu-integrable for μ∈D\mu\in\mathcal{D}, and Σ(μ)∈Tμ\Sigma(\mu)\in T_{\mu} for μ∈DΣ\mu\in\mathcal{D}_{\Sigma}.

Constants. By Basic Properties of a Confining Potential on Euclidean Space: Continuity, Quadratic Minorants, the Tangent Inequality, Hessian Bounds, Growth under Translation, and Integrability §growth fix v0v_{0} and Cg≥0C_{g}\ge0 with v0≤Vv_{0}\le V and V(x+a)−v0+1≤exp⁡(Cg∥a∥)(V(x)−v0+1)V(x+a)-v_{0}+1\le\exp(C_{g}\lVert a\rVert)(V(x)-v_{0}+1); put u=V−v0+1≥1u=V-v_{0}+1\ge1 and A0=1+∣v0−1∣A_{0}=1+|v_{0}-1|, so that ∣V∣≤A0u|V|\le A_{0}u and 1+∣V∣≤(1+A0)u1+|V|\le(1+A_{0})u. Taking x=0Rdx=0_{\mathbb{R}^{d}} and a=ya=y, and using ∥DV∥≤Cs(1+∣V∣)\lVert DV\rVert\le C_{s}(1+|V|) with Cs≥0C_{s}\ge0 from Confining Potentials on Euclidean Space §slope (nonnegative since 0≤∥DV(0Rd)∥0\le\lVert DV(0_{\mathbb{R}^{d}})\rVert), we get for every y∈Rdy\in\mathbb{R}^{d}

∣V(y)∣≤A0u(0Rd)exp⁡(Cg∥y∥),∥∇V(y)∥2≤(Cs(1+A0)u(0Rd))2exp⁡(2Cg∥y∥),(X)|V(y)|\le A_{0}u(0_{\mathbb{R}^{d}})\exp(C_{g}\lVert y\rVert),\qquad\lVert\nabla V(y)\rVert^{2}\le\bigl(C_{s}(1+A_{0})u(0_{\mathbb{R}^{d}})\bigr)^{2}\exp(2C_{g}\lVert y\rVert),\tag{X}

using exp⁡(s)exp⁡(s)=exp⁡(2s)\exp(s)\exp(s)=\exp(2s) (Basic Properties of the Exponential Function). By Confining Potentials on Euclidean Space §curvature with ε=1\varepsilon=1 fix C1C_{1} with ΔV≤∣V∣+C1\Delta V\le|V|+C_{1}. Let c1c_{1} be the constant of Basic Properties of the Entropy on the Wasserstein Space: Comparison with the Gaussian Relative Entropy, Lower Bound, Translation Invariance, Absolute Continuity, Closed Sublevel Sets and Lower Semicontinuity and γ=σ22log⁡c1\gamma=\tfrac{\sigma^{2}}{2}\log c_{1}. By Basic Properties of a Confining Potential on Euclidean Space: Continuity, Quadratic Minorants, the Tangent Inequality, Hessian Bounds, Growth under Translation, and Integrability §minorant with M=σ24+1>0M=\tfrac{\sigma^{2}}{4}+1>0 fix CMC_{M} with M∥x∥2−CM≤V(x)M\lVert x\rVert^{2}-C_{M}\le V(x) for all xx.

Step E (energy bounds). Let μ∈D\mu\in\mathcal{D}. Integrating the minorant, M M2(μ)−CM≤∫V dμM\,M_{2}(\mu)-C_{M}\le\int V\,d\mu; and log⁡c1−12M2(μ)≤Ent(μ)\log c_{1}-\tfrac12M_{2}(\mu)\le\mathrm{Ent}(\mu) by Basic Properties of the Entropy on the Wasserstein Space: Comparison with the Gaussian Relative Entropy, Lower Bound, Translation Invariance, Absolute Continuity, Closed Sublevel Sets and Lower Semicontinuity §lower. Hence E(μ)≥γ−σ24M2(μ)+M M2(μ)−CM=γ+M2(μ)−CM\mathcal{E}(\mu)\ge\gamma-\tfrac{\sigma^{2}}{4}M_{2}(\mu)+M\,M_{2}(\mu)-C_{M}=\gamma+M_{2}(\mu)-C_{M}, that is

M2(μ)≤E(μ)−γ+CM.(E1)M_{2}(\mu)\le\mathcal{E}(\mu)-\gamma+C_{M}.\tag{E1}

Next ∫V dμ=E(μ)−σ22Ent(μ)≤E(μ)−γ+σ24M2(μ)\int V\,d\mu=\mathcal{E}(\mu)-\tfrac{\sigma^{2}}{2}\mathrm{Ent}(\mu)\le\mathcal{E}(\mu)-\gamma+\tfrac{\sigma^{2}}{4}M_{2}(\mu), so by (E1)

∫V dμ≤β0 E(μ)+b0,β0=1+σ24,b0=σ24CM−β0γ.(E2)\int V\,d\mu\le\beta_{0}\,\mathcal{E}(\mu)+b_{0},\qquad\beta_{0}=1+\tfrac{\sigma^{2}}{4},\quad b_{0}=\tfrac{\sigma^{2}}{4}C_{M}-\beta_{0}\gamma.\tag{E2}

Finally ∫V dμ≥v0\int V\,d\mu\ge v_{0} gives

Ent(μ)=2σ2(E(μ)−∫V dμ)≤2σ2(E(μ)−v0).(E3)\mathrm{Ent}(\mu)=\tfrac{2}{\sigma^{2}}\Bigl(\mathcal{E}(\mu)-\int V\,d\mu\Bigr)\le\tfrac{2}{\sigma^{2}}\bigl(\mathcal{E}(\mu)-v_{0}\bigr).\tag{E3}

Step S (smoothing). Let ρ∈P(Rd)\rho\in\mathcal{P}(\mathbb{R}^{d}) satisfy ρ(Bˉ(0Rd,R))=1\rho(\bar{B}(0_{\mathbb{R}^{d}},R))=1 for some real R≥0R\ge0, with Bˉ\bar{B} as in Euclidean Space and Lebesgue Measure: Standing Notation §space. Then ∥y∥2≤R2\lVert y\rVert^{2}\le R^{2} for ρ\rho-almost every yy, so M2(ρ)≤R2M_{2}(\rho)\le R^{2} (The Lebesgue Integral and Null Sets: Almost-Everywhere Comparison, Markov's Inequality, and Dominated Convergence Almost Everywhere §comparison) and ρ∈P2(Rd)\rho\in\mathcal{P}_{2}(\mathbb{R}^{d}). For real ss with 0<s≤120<s\le\tfrac12 let ρs\rho_{s} be the measure μs\mu_{s} of Gaussian Smoothing of a Measure with Finite Second Moment: Positive Density, Finite Entropy, a Tangent Score, Non-Increasing Fisher Information and Exponential Integrability built from ρ\rho. We show ρs∈DΣ\rho_{s}\in\mathcal{D}_{\Sigma} and W2(ρs,ρ)2≤d sW_{2}(\rho_{s},\rho)^{2}\le d\,s. By Gaussian Smoothing of a Measure with Finite Second Moment: Positive Density, Finite Entropy, a Tangent Score, Non-Increasing Fisher Information and Exponential Integrability §density, Gaussian Smoothing of a Measure with Finite Second Moment: Positive Density, Finite Entropy, a Tangent Score, Non-Increasing Fisher Information and Exponential Integrability §entropy and Gaussian Smoothing of a Measure with Finite Second Moment: Positive Density, Finite Entropy, a Tangent Score, Non-Increasing Fisher Information and Exponential Integrability §score, ρs∈P2Ent(Rd)∩P2I(Rd)\rho_{s}\in\mathcal{P}_{2}^{\mathrm{Ent}}(\mathbb{R}^{d})\cap\mathcal{P}_{2}^{\mathcal{I}}(\mathbb{R}^{d}) and W2(ρs,ρ)2≤d sW_{2}(\rho_{s},\rho)^{2}\le d\,s. The functions VV (continuous, hence Borel) and ∥∇V∥2\lVert\nabla V\rVert^{2} (Borel by Pairs of Euclidean Points: Coordinate Projections, Pairings, the Product Measure on a Euclidean Space, Borel Norm Functions and Finite Sets §functions, ∇V\nabla V being Borel by Basic Properties of a Confining Potential on Euclidean Space: Continuity, Quadratic Minorants, the Tangent Inequality, Hessian Bounds, Growth under Translation, and Integrability §regularity) satisfy the exponential bounds (X), so both are ρs\rho_{s}-integrable by Gaussian Smoothing of a Measure with Finite Second Moment: Positive Density, Finite Entropy, a Tangent Score, Non-Increasing Fisher Information and Exponential Integrability §exponential. Hence ρs∈D\rho_{s}\in\mathcal{D} and then ρs∈DΣ\rho_{s}\in\mathcal{D}_{\Sigma}.

Claim 6 (first variation on D\mathcal{D}). Let μ∈D\mu\in\mathcal{D} and ψ∈Cc∞(Rd)\psi\in C_{c}^{\infty}(\mathbb{R}^{d}). By The First Variation of the Entropy Along a Gradient Perturbation of the Identity §variation there is t1>0t_{1}>0 with (id+t∇ψ)#μ∈P2Ent(Rd)(\mathrm{id}+t\nabla\psi)_{\#}\mu\in\mathcal{P}_{2}^{\mathrm{Ent}}(\mathbb{R}^{d}) for t∈(−t1,t1)t\in(-t_{1},t_{1}) and t↦Ent((id+t∇ψ)#μ)t\mapsto\mathrm{Ent}((\mathrm{id}+t\nabla\psi)_{\#}\mu) differentiable at 00 with derivative −∫Δψ dμ-\int\Delta\psi\,d\mu. By The Potential Energy of a Confining Potential: Smoothness in Translations, First Variation, Closed Sublevel Sets and Convergence at Bounded Energy §variation, ∇V⋅∇ψ\nabla V\cdot\nabla\psi is μ\mu-integrable, and for every tt the function V∘(id+t∇ψ)V\circ(\mathrm{id}+t\nabla\psi) is μ\mu-integrable, so VV is (id+t∇ψ)#μ(\mathrm{id}+t\nabla\psi)_{\#}\mu-integrable with ∫V d(id+t∇ψ)#μ=FV(t):=∫V(x+t∇ψ(x)) μ(dx)\int V\,d(\mathrm{id}+t\nabla\psi)_{\#}\mu=F_{V}(t):=\int V(x+t\nabla\psi(x))\,\mu(dx), and FVF_{V} is differentiable at 00 with derivative ∫∇V⋅∇ψ dμ\int\nabla V\cdot\nabla\psi\,d\mu. Put t0=t1t_{0}=t_{1}. For t∈(−t0,t0)t\in(-t_{0},t_{0}), (id+t∇ψ)#μ∈D(\mathrm{id}+t\nabla\psi)_{\#}\mu\in\mathcal{D} and E((id+t∇ψ)#μ)=σ22Ent((id+t∇ψ)#μ)+FV(t)\mathcal{E}((\mathrm{id}+t\nabla\psi)_{\#}\mu)=\tfrac{\sigma^{2}}{2}\mathrm{Ent}((\mathrm{id}+t\nabla\psi)_{\#}\mu)+F_{V}(t). Restricting FVF_{V} to (−t0,t0)(-t_{0},t_{0}) (claim 2 of Restriction Stability of Continuity and of the Derivative) and using claims 1 and 2 of Sum, Constant Multiple, and Product Rules for One-Dimensional Derivatives, this function is differentiable at 00 with derivative ∫∇V⋅∇ψ dμ−σ22∫Δψ dμ\int\nabla V\cdot\nabla\psi\,d\mu-\tfrac{\sigma^{2}}{2}\int\Delta\psi\,d\mu.

Claim 1 (penalty pair). We verify the five conditions of Penalty Pairs on the Wasserstein Space: the Penalty, Its Score, and Their Domains §pair.

Penalty Pairs on the Wasserstein Space: the Penalty, Its Score, and Their Domains §nonempty. Let γd∈P(Rd)\gamma_{d}\in\mathcal{P}(\mathbb{R}^{d}) be the measure of Basic Properties of the Entropy on the Wasserstein Space: Comparison with the Gaussian Relative Entropy, Lower Bound, Translation Invariance, Absolute Continuity, Closed Sublevel Sets and Lower Semicontinuity and ρ=(P1)#γd\rho=(P_{1})_{\#}\gamma_{d}, with P1P_{1} the radial retraction of Radial Retraction onto a Closed Ball: Compactly Supported Approximation in the Wasserstein Distance (dimension dd, R=1R=1). By Radial Retraction onto a Closed Ball: Compactly Supported Approximation in the Wasserstein Distance §retraction, ρ(Bˉ(0Rd,1))=1\rho(\bar{B}(0_{\mathbb{R}^{d}},1))=1, and Step S gives ρ1/2∈DΣ\rho_{1/2}\in\mathcal{D}_{\Sigma}.

Penalty Pairs on the Wasserstein Space: the Penalty, Its Score, and Their Domains §bound. Put C=∣γ∣+∣CM∣≥0C=|\gamma|+|C_{M}|\ge0. For μ∈D\mu\in\mathcal{D}, by (E1), E(μ)≥γ−CM+M2(μ)≥−C≥−C(1+M2(μ))\mathcal{E}(\mu)\ge\gamma-C_{M}+M_{2}(\mu)\ge-C\ge-C(1+M_{2}(\mu)).

Penalty Pairs on the Wasserstein Space: the Penalty, Its Score, and Their Domains §translation, and the translation Hessian. Let μ∈D\mu\in\mathcal{D} and a∈Rda\in\mathbb{R}^{d}, with τa(x)=x+a\tau_{a}(x)=x+a (The Wasserstein Space and Its Lift to Square-Integrable Random Vectors: Standing Notation §constants). By Basic Properties of the Entropy on the Wasserstein Space: Comparison with the Gaussian Relative Entropy, Lower Bound, Translation Invariance, Absolute Continuity, Closed Sublevel Sets and Lower Semicontinuity §translation, (τa)#μ∈P2Ent(Rd)(\tau_{a})_{\#}\mu\in\mathcal{P}_{2}^{\mathrm{Ent}}(\mathbb{R}^{d}) and Ent((τa)#μ)=Ent(μ)\mathrm{Ent}((\tau_{a})_{\#}\mu)=\mathrm{Ent}(\mu). By Basic Properties of a Confining Potential on Euclidean Space: Continuity, Quadratic Minorants, the Tangent Inequality, Hessian Bounds, Growth under Translation, and Integrability §integrable, x↦V(x+a)x\mapsto V(x+a) is μ\mu-integrable, so VV is (τa)#μ(\tau_{a})_{\#}\mu-integrable with integral Φμ(a)\Phi_{\mu}(a), the function of The Potential Energy of a Confining Potential: Smoothness in Translations, First Variation, Closed Sublevel Sets and Convergence at Bounded Energy §translation. Hence (τa)#μ∈D(\tau_{a})_{\#}\mu\in\mathcal{D} and eμ(a)=σ22Ent(μ)+Φμ(a)e_{\mu}(a)=\tfrac{\sigma^{2}}{2}\mathrm{Ent}(\mu)+\Phi_{\mu}(a). The constant function is smooth (claim 2 of Constants, Coordinate Functions, Sums and Products of CkC^k Functions on a Euclidean Open Set) and Φμ\Phi_{\mu} is of class C2C^{2} (The Potential Energy of a Confining Potential: Smoothness in Translations, First Variation, Closed Sublevel Sets and Convergence at Bounded Energy §translation), so eμe_{\mu} is of class C2C^{2} on Rd\mathbb{R}^{d} (claim 3 of Constants, Coordinate Functions, Sums and Products of CkC^k Functions on a Euclidean Open Set), with ∂j∂ieμ=∂j∂iΦμ\partial_{j}\partial_{i}e_{\mu}=\partial_{j}\partial_{i}\Phi_{\mu} by claim 1 there, the partial derivatives of a constant being 00. By Hessian Matrix of a C^2 Function, each entry of HE(μ)=D2eμ(0Rd)H_{\mathcal{E}}(\mu)=D^{2}e_{\mu}(0_{\mathbb{R}^{d}}) is a second partial derivative ∂j∂ieμ(0Rd)=∂j∂iΦμ(0Rd)=∫∂j∂iV dμ\partial_{j}\partial_{i}e_{\mu}(0_{\mathbb{R}^{d}})=\partial_{j}\partial_{i}\Phi_{\mu}(0_{\mathbb{R}^{d}})=\int\partial_{j}\partial_{i}V\,d\mu, and the corresponding entry of D2V(x)D^{2}V(x) is ∂j∂iV(x)\partial_{j}\partial_{i}V(x); each such function, and ΔV\Delta V, is μ\mu-integrable by Basic Properties of a Confining Potential on Euclidean Space: Continuity, Quadratic Minorants, the Tangent Inequality, Hessian Bounds, Growth under Translation, and Integrability §integrable. By the definition of the trace, tr HE(μ)=∑i=1d∫∂i∂iV dμ=∫∑i=1d∂i∂iV dμ=∫ΔV dμ\mathrm{tr}\,H_{\mathcal{E}}(\mu)=\sum_{i=1}^{d}\int\partial_{i}\partial_{i}V\,d\mu=\int\sum_{i=1}^{d}\partial_{i}\partial_{i}V\,d\mu=\int\Delta V\,d\mu, by claim 2 of Linearity and Monotonicity of the Lebesgue Integral applied d−1d-1 times and The Laplacian of a Twice Continuously Differentiable Function §laplacian. This proves the second sentence of Claim 1.

Penalty Pairs on the Wasserstein Space: the Penalty, Its Score, and Their Domains §variation. Let μ∈DΣ\mu\in\mathcal{D}_{\Sigma} and ψ∈Cc∞(Rd)\psi\in C_{c}^{\infty}(\mathbb{R}^{d}). Claim 6 gives t0t_{0} and the derivative ∫∇V⋅∇ψ dμ−σ22∫Δψ dμ\int\nabla V\cdot\nabla\psi\,d\mu-\tfrac{\sigma^{2}}{2}\int\Delta\psi\,d\mu. By Square-Integrable Vector Fields Against a Probability Measure on Euclidean Space, and Test Functions: Standing Notation §l2mu, ⟨∇V,∇ψ⟩μ=∫∇V⋅∇ψ dμ\langle\nabla V,\nabla\psi\rangle_{\mu}=\int\nabla V\cdot\nabla\psi\,d\mu, and by Finite Fisher Information, the Score and the Fisher Information of a Probability Measure §score, ⟨ξμ,∇ψ⟩μ=−∫Δψ dμ\langle\xi_{\mu},\nabla\psi\rangle_{\mu}=-\int\Delta\psi\,d\mu; by bilinearity of the inner product the derivative equals ⟨∇V+σ22ξμ,∇ψ⟩μ=⟨Σ(μ),∇ψ⟩μ\langle\nabla V+\tfrac{\sigma^{2}}{2}\xi_{\mu},\nabla\psi\rangle_{\mu}=\langle\Sigma(\mu),\nabla\psi\rangle_{\mu}.

Penalty Pairs on the Wasserstein Space: the Penalty, Its Score, and Their Domains §dense. Let μ∈D\mu\in\mathcal{D} and ε>0\varepsilon>0. By Radial Retraction onto a Closed Ball: Compactly Supported Approximation in the Wasserstein Distance §approximation choose R>0R>0 with W2(ρ,μ)<ε/2W_{2}(\rho,\mu)<\varepsilon/2, where ρ=(PR)#μ\rho=(P_{R})_{\#}\mu, and ρ(Bˉ(0Rd,R))=1\rho(\bar{B}(0_{\mathbb{R}^{d}},R))=1 by Radial Retraction onto a Closed Ball: Compactly Supported Approximation in the Wasserstein Distance §retraction. Let ss be the lesser of 12\tfrac12 and ε2(8d)−1\varepsilon^{2}(8d)^{-1}, with dd read in R\mathbb{R}; by Step S, ρs∈DΣ\rho_{s}\in\mathcal{D}_{\Sigma} and W2(ρs,ρ)2≤ds≤ε2/8<(ε/2)2W_{2}(\rho_{s},\rho)^{2}\le ds\le\varepsilon^{2}/8<(\varepsilon/2)^{2}, so W2(ρs,ρ)<ε/2W_{2}(\rho_{s},\rho)<\varepsilon/2 (claim 1 of Monotonicity of Squaring on the Nonnegative Elements of an Ordered Field). By The Quadratic Wasserstein Distance is a Metric on the Wasserstein Space §triangle, W2(ρs,μ)<εW_{2}(\rho_{s},\mu)<\varepsilon.

So (D,DΣ,E,Σ)(\mathcal{D},\mathcal{D}_{\Sigma},\mathcal{E},\Sigma) is a penalty pair, and Claim 1 is proved.

Claim 2 (coercivity and the map property). Let c∈Rc\in\mathbb{R} and let (μn)n(\mu_{n})_{n} be a sequence in Sc={μ∈D:E(μ)≤c}S_{c}=\{\mu\in\mathcal{D}:\mathcal{E}(\mu)\le c\}. By (E2), ∫V dμn≤β0c+b0\int V\,d\mu_{n}\le\beta_{0}c+b_{0} for every nn. The function VV is Borel, bounded below by v0v_{0}, and superquadratic in the sense of Wasserstein Convergence from Weak Convergence with Uniformly Integrable Second Moments, and Wasserstein Compactness under a Superquadratic Moment Bound §compactness by Confining Potentials on Euclidean Space §superquadratic; that clause (with m=dm=d) gives μ∈P2(Rd)\mu\in\mathcal{P}_{2}(\mathbb{R}^{d}) and a strictly increasing (nk)k(n_{k})_{k} with W2(μnk,μ)→0W_{2}(\mu_{n_{k}},\mu)\to0. By (E3), Ent(μnk)≤2σ2(c−v0)\mathrm{Ent}(\mu_{n_{k}})\le\tfrac{2}{\sigma^{2}}(c-v_{0}), so Basic Properties of the Entropy on the Wasserstein Space: Comparison with the Gaussian Relative Entropy, Lower Bound, Translation Invariance, Absolute Continuity, Closed Sublevel Sets and Lower Semicontinuity §closed gives μ∈P2Ent(Rd)\mu\in\mathcal{P}_{2}^{\mathrm{Ent}}(\mathbb{R}^{d}); and The Potential Energy of a Confining Potential: Smoothness in Translations, First Variation, Closed Sublevel Sets and Convergence at Bounded Energy §closed gives that VV is μ\mu-integrable, so μ∈D\mu\in\mathcal{D}. Let ε>0\varepsilon>0. By Basic Properties of the Entropy on the Wasserstein Space: Comparison with the Gaussian Relative Entropy, Lower Bound, Translation Invariance, Absolute Continuity, Closed Sublevel Sets and Lower Semicontinuity §lsc and claim 1 of Sequential Characterization of Lower Semicontinuity on a Subset of a Metric Space (the sequence (μnk)k(\mu_{n_{k}})_{k} lies in P2Ent(Rd)\mathcal{P}_{2}^{\mathrm{Ent}}(\mathbb{R}^{d}) and converges there to μ\mu) there is N1N_{1} with Ent(μ)−ε<Ent(μnk)\mathrm{Ent}(\mu)-\varepsilon<\mathrm{Ent}(\mu_{n_{k}}) for k≥N1k\ge N_{1}, and by The Potential Energy of a Confining Potential: Smoothness in Translations, First Variation, Closed Sublevel Sets and Convergence at Bounded Energy §closed there is N2N_{2} with ∫V dμ−ε<∫V dμnk\int V\,d\mu-\varepsilon<\int V\,d\mu_{n_{k}} for k≥N2k\ge N_{2} (the indices nk≥kn_{k}\ge k being large with kk). For kk at least the larger of N1,N2N_{1},N_{2}, E(μ)−(σ22+1)ε<E(μnk)≤c\mathcal{E}(\mu)-(\tfrac{\sigma^{2}}{2}+1)\varepsilon<\mathcal{E}(\mu_{n_{k}})\le c. As ε>0\varepsilon>0 was arbitrary, E(μ)≤c\mathcal{E}(\mu)\le c (Comparison of Real Numbers with Arbitrary Positive Slack), so μ∈Sc\mu\in S_{c}. Thus ScS_{c} is sequentially compact in (P2(Rd),W2)(\mathcal{P}_{2}(\mathbb{R}^{d}),W_{2}), and the pair is Wasserstein-coercive (Wasserstein-Coercive Penalty Pairs §coercive). Every μ∈D\mu\in\mathcal{D} has finite entropy and is therefore absolutely continuous by Basic Properties of the Entropy on the Wasserstein Space: Comparison with the Gaussian Relative Entropy, Lower Bound, Translation Invariance, Absolute Continuity, Closed Sublevel Sets and Lower Semicontinuity §absolutely-continuous; so D\mathcal{D} has the map property by Two Sufficient Conditions for the Map Property: Absolute Continuity, and Atomlessness on the Line §absolutely-continuous.

Claim 3 (semicontinuity and growth). By Claims 1 and 2 the pair is a Wasserstein-coercive penalty pair, so E\mathcal{E} is lower semicontinuous on D\mathcal{D} by Basic Properties of a Wasserstein-Coercive Penalty Pair §lsc. Let μ∈D\mu\in\mathcal{D}. By (E1), M2(μ)≤∣E(μ)∣+∣γ∣+∣CM∣≤c1′(1+∣E(μ)∣)M_{2}(\mu)\le|\mathcal{E}(\mu)|+|\gamma|+|C_{M}|\le c_{1}'(1+|\mathcal{E}(\mu)|) with c1′=1+∣γ∣+∣CM∣c_{1}'=1+|\gamma|+|C_{M}|. By Claim 1, Basic Properties of a Confining Potential on Euclidean Space: Continuity, Quadratic Minorants, the Tangent Inequality, Hessian Bounds, Growth under Translation, and Integrability §hessian (0≤ΔV0\le\Delta V), the choice of C1C_{1}, ∣V∣≤(V−v0)+∣v0∣|V|\le(V-v_{0})+|v_{0}| and (E2),

0≤tr HE(μ)=∫ΔV dμ≤∫V dμ−v0+∣v0∣+∣C1∣≤β0∣E(μ)∣+∣b0∣+2∣v0∣+∣C1∣,0\le\mathrm{tr}\,H_{\mathcal{E}}(\mu)=\int\Delta V\,d\mu\le\int V\,d\mu-v_{0}+|v_{0}|+|C_{1}|\le\beta_{0}|\mathcal{E}(\mu)|+|b_{0}|+2|v_{0}|+|C_{1}|,

so ∣tr HE(μ)∣≤c2′(1+∣E(μ)∣)|\mathrm{tr}\,H_{\mathcal{E}}(\mu)|\le c_{2}'(1+|\mathcal{E}(\mu)|) with c2′=β0+∣b0∣+2∣v0∣+∣C1∣c_{2}'=\beta_{0}+|b_{0}|+2|v_{0}|+|C_{1}|. Take C=c1′+c2′C=c_{1}'+c_{2}'.

Step C (convergence of integrals at bounded energy). Let R>0R>0 and SR={μ∈D:∣E(μ)∣≤R}S_{R}=\{\mu\in\mathcal{D}:|\mathcal{E}(\mu)|\le R\}, and let h:Rd→Rh:\mathbb{R}^{d}\to\mathbb{R} be continuous with ∣h(x)∣≤ΔV(x)|h(x)|\le\Delta V(x) for every x∈Rdx\in\mathbb{R}^{d}. We show hh is μ\mu-integrable for every μ∈SR\mu\in S_{R}, and the restriction of μ↦∫h dμ\mu\mapsto\int h\,d\mu to SRS_{R} is continuous. It suffices to show: if μn,μ∈SR\mu_{n},\mu\in S_{R} for n∈Nn\in\mathbb{N} and W2(μn,μ)→0W_{2}(\mu_{n},\mu)\to0, then hh is integrable with respect to μ\mu and every μn\mu_{n}, and ∫h dμn→∫h dμ\int h\,d\mu_{n}\to\int h\,d\mu. Applied to the constant sequence at a given μ∈SR\mu\in S_{R} this gives the integrability, and then Continuity Between Metric Spaces is Equivalent to Sequential Continuity §on-subset gives the continuity. These measures lie in D⊆P2(Rd)\mathcal{D}\subseteq\mathcal{P}_{2}(\mathbb{R}^{d}), VV is integrable with respect to each μn\mu_{n}, and by (E2), ∫V dμn≤β0R+b0\int V\,d\mu_{n}\le\beta_{0}R+b_{0} for every nn. For every ε>0\varepsilon>0, with CεC_{\varepsilon} from Confining Potentials on Euclidean Space §curvature, ∣h∣≤ΔV≤ε∣V∣+Cε|h|\le\Delta V\le\varepsilon|V|+C_{\varepsilon} on Rd\mathbb{R}^{d}. So The Potential Energy of a Confining Potential: Smoothness in Translations, First Variation, Closed Sublevel Sets and Convergence at Bounded Energy §convergence with this hh and c=β0R+b0c=\beta_{0}R+b_{0} gives that hh is integrable with respect to μ\mu and every μn\mu_{n} and that ∫h dμn→∫h dμ\int h\,d\mu_{n}\to\int h\,d\mu.

Claim 4 (continuity of the trace of the translation Hessian). Let R>0R>0, with SRS_{R} as in Step C. The function ΔV\Delta V is continuous (Basic Properties of a Confining Potential on Euclidean Space: Continuity, Quadratic Minorants, the Tangent Inequality, Hessian Bounds, Growth under Translation, and Integrability §regularity) and nonnegative (Basic Properties of a Confining Potential on Euclidean Space: Continuity, Quadratic Minorants, the Tangent Inequality, Hessian Bounds, Growth under Translation, and Integrability §hessian), so ∣ΔV∣=ΔV|\Delta V|=\Delta V and Step C applies to h=ΔVh=\Delta V: the restriction of μ↦∫ΔV dμ\mu\mapsto\int\Delta V\,d\mu to SRS_{R} is continuous, and by Claim 1 this function is μ↦tr HE(μ)\mu\mapsto\mathrm{tr}\,H_{\mathcal{E}}(\mu).

Claim 5 (entries of the translation Hessian). Let i,k∈[d]i,k\in[d] and let hik=∂i∂kVh_{ik}=\partial_{i}\partial_{k}V, the function whose value at xx is the entry of D2V(x)D^{2}V(x) in row ii and column kk by Hessian Matrix of a C^2 Function. By Claim 1, for every μ∈D\mu\in\mathcal{D} the function hikh_{ik} is μ\mu-integrable and HE(μ)ik=∫hik dμH_{\mathcal{E}}(\mu)_{ik}=\int h_{ik}\,d\mu. The function hikh_{ik} is continuous by Basic Properties of a Confining Potential on Euclidean Space: Continuity, Quadratic Minorants, the Tangent Inequality, Hessian Bounds, Growth under Translation, and Integrability §regularity, and ∣hik(x)∣≤ΔV(x)|h_{ik}(x)|\le\Delta V(x) for every x∈Rdx\in\mathbb{R}^{d} by Basic Properties of a Confining Potential on Euclidean Space: Continuity, Quadratic Minorants, the Tangent Inequality, Hessian Bounds, Growth under Translation, and Integrability §hessian, applied with its indices i,ji,j read as our k,ik,i.

Entry bound. Let μ∈D\mu\in\mathcal{D}. By claim 6 of Properties of the Absolute Value in an Ordered Field, −ΔV(x)≤hik(x)≤ΔV(x)-\Delta V(x)\le h_{ik}(x)\le\Delta V(x) for every xx. The functions hikh_{ik} and ΔV\Delta V are μ\mu-integrable by Claim 1, so claim 2 of Linearity and Monotonicity of the Lebesgue Integral (the monotonicity, and the linearity with coefficients −1-1 and 00, which gives that −ΔV-\Delta V is integrable with integral −∫ΔV dμ-\int\Delta V\,d\mu) yields

−∫ΔV dμ≤∫hik dμ≤∫ΔV dμ.-\int\Delta V\,d\mu\le\int h_{ik}\,d\mu\le\int\Delta V\,d\mu .

By claim 6 of Properties of the Absolute Value in an Ordered Field and Claim 1, ∣HE(μ)ik∣=∣∫hik dμ∣≤∫ΔV dμ=tr HE(μ)|H_{\mathcal{E}}(\mu)_{ik}|=\bigl|\int h_{ik}\,d\mu\bigr|\le\int\Delta V\,d\mu=\mathrm{tr}\,H_{\mathcal{E}}(\mu).

Continuity. Let R>0R>0, with SRS_{R} as in Step C. Step C applies to h=hikh=h_{ik}, so the restriction of μ↦∫hik dμ\mu\mapsto\int h_{ik}\,d\mu, that is of μ↦HE(μ)ik\mu\mapsto H_{\mathcal{E}}(\mu)_{ik}, to SRS_{R} is continuous.

Claims 1 to 6 are clauses 1 to 6 of the statement. ■\blacksquare

Please log in to copy this version.

Citations

Loading…

Dependency Graph

0 prerequisites

Comments

Loading…