TheoremBase

Differentiability of a Function on the Noise-Connected Measures Along Noise Couplings, and Its Gradient

A real function on the measures noise-connected to the reference measure is differentiable along noise couplings at a measure if its increments are approximated, to first order in the square root of the noise cost, by the noise displacement pairing with a noise field; that field is unique and is called the gradient along noise couplings.

Statement

In the setting of Probability Measures on a Hilbert Space Transported in the Noise Norm: Standing Notation, let Pρa\mathcal{P}^{a}_{\rho} be the set of The Measures Noise-Connected to the Reference Measure §space, let φ:Pρa→R\varphi:\mathcal{P}^{a}_{\rho}\to\mathbb{R} and let μ∈Pρa\mu\in\mathcal{P}^{a}_{\rho}. Let L2(μ;Xa)L^{2}(\mu;X^{a}) be the space of Probability Measures on a Hilbert Space Transported in the Noise Norm: Standing Notation §fields. For ν∈Pρa\nu\in\mathcal{P}^{a}_{\rho}, Πa(μ,ν)\Pi^{a}(\mu,\nu) is the set of Couplings of Finite Noise Cost and Their Noise Cost §couplings, and for π∈Πa(μ,ν)\pi\in\Pi^{a}(\mu,\nu) the noise cost Ia(π)I^{a}(\pi) is a nonnegative real number by Couplings of Finite Noise Cost and Their Noise Cost §cost and Ja(η,π)\mathcal{J}^{a}(\eta,\pi) is the noise displacement pairing of η∈L2(μ;Xa)\eta\in L^{2}(\mu;X^{a}) along π\pi. Square roots are those of Existence and Uniqueness of the Nonnegative Square Root.

1. (Differentiability along noise couplings) Let η∈L2(μ;Xa)\eta\in L^{2}(\mu;X^{a}). The function φ\varphi is differentiable along noise couplings at μ\mu with gradient η\eta if for every ε∈R\varepsilon\in\mathbb{R} with 0<ε0<\varepsilon there is θ∈R\theta\in\mathbb{R} with 0<θ0<\theta such that

∣φ(ν)−φ(μ)−Ja(η,π)∣≤ε Ia(π)\bigl|\varphi(\nu)-\varphi(\mu)-\mathcal{J}^{a}(\eta,\pi)\bigr|\le\varepsilon\,\sqrt{I^{a}(\pi)}

for every ν∈Pρa\nu\in\mathcal{P}^{a}_{\rho} and every π∈Πa(μ,ν)\pi\in\Pi^{a}(\mu,\nu) with Ia(π)<θ2I^{a}(\pi)<\theta^{2}. The function φ\varphi is differentiable along noise couplings at μ\mu if there is an η∈L2(μ;Xa)\eta\in L^{2}(\mu;X^{a}) with this property.

2. (The gradient along noise couplings) Suppose that φ\varphi is differentiable along noise couplings at μ\mu. Then exactly one η∈L2(μ;Xa)\eta\in L^{2}(\mu;X^{a}) has the property of clause 1. Indeed, one exists by that hypothesis. Suppose η\eta and η′\eta' both have it and let ε∈R\varepsilon\in\mathbb{R} be positive, so that ε⋅2−1\varepsilon\cdot2^{-1} is positive by Elementary Order Arithmetic in an Ordered Field §halving. Let θ\theta and θ′\theta' be as in clause 1 for η\eta and for η′\eta' respectively, with ε⋅2−1\varepsilon\cdot2^{-1} there in place of ε\varepsilon, and let θ′′=min⁡{θ,θ′}\theta''=\min\{\theta,\theta'\} be their minimum, which is positive because it equals θ\theta or θ′\theta' by claim 2 of Elementary Properties of the Minimum of Two Elements. Let ν∈Pρa\nu\in\mathcal{P}^{a}_{\rho} and π∈Πa(μ,ν)\pi\in\Pi^{a}(\mu,\nu) satisfy Ia(π)<(θ′′)2I^{a}(\pi)<(\theta'')^{2}. Since θ′′≤θ\theta''\le\theta and θ′′≤θ′\theta''\le\theta' by claim 1 of that lemma, and both sides of each of these inequalities are nonnegative, claim 2 of Monotonicity of Squaring on the Nonnegative Elements of an Ordered Field gives (θ′′)2≤θ2(\theta'')^{2}\le\theta^{2} and (θ′′)2≤(θ′)2(\theta'')^{2}\le(\theta')^{2}, whence Ia(π)<θ2I^{a}(\pi)<\theta^{2} and Ia(π)<(θ′)2I^{a}(\pi)<(\theta')^{2} by Elementary Order Arithmetic in an Ordered Field §mixed-transitivity. Both estimates of clause 1 are therefore available at π\pi, and, by the linearity of the noise displacement pairing (Noise Displacement and Cross Pairings Along Couplings: Bounds, Linearity, Displacement Couplings, Polarisation and a Vanishing Criterion §linear, applied with s=1s=1 and t=−1t=-1),

Ja(η−η′,π)=(φ(ν)−φ(μ)−Ja(η′,π))−(φ(ν)−φ(μ)−Ja(η,π)),\mathcal{J}^{a}(\eta-\eta',\pi)=\bigl(\varphi(\nu)-\varphi(\mu)-\mathcal{J}^{a}(\eta',\pi)\bigr)-\bigl(\varphi(\nu)-\varphi(\mu)-\mathcal{J}^{a}(\eta,\pi)\bigr),

so that the triangle inequality Properties of the Absolute Value in an Ordered Field §triangle, together with the symmetry Properties of the Absolute Value in an Ordered Field §symmetry, bounds ∣Ja(η−η′,π)∣|\mathcal{J}^{a}(\eta-\eta',\pi)| by ε⋅2−1Ia(π)+ε⋅2−1Ia(π)\varepsilon\cdot2^{-1}\sqrt{I^{a}(\pi)}+\varepsilon\cdot2^{-1}\sqrt{I^{a}(\pi)}, which equals εIa(π)\varepsilon\sqrt{I^{a}(\pi)} by the distributive law of Field and the halving identity of Elementary Order Arithmetic in an Ordered Field §halving. Thus

∣Ja(η−η′,π)∣≤ε Ia(π).\bigl|\mathcal{J}^{a}(\eta-\eta',\pi)\bigr|\le\varepsilon\,\sqrt{I^{a}(\pi)} .

So η−η′\eta-\eta' satisfies the hypothesis of Noise Displacement and Cross Pairings Along Couplings: Bounds, Linearity, Displacement Couplings, Polarisation and a Vanishing Criterion §vanishing, whence η−η′\eta-\eta' is the zero element of L2(μ;Xa)L^{2}(\mu;X^{a}) by that claim and η=η′\eta=\eta'. This unique η\eta is written ∇φ(μ)\nabla\varphi(\mu) and called the gradient along noise couplings of φ\varphi at μ\mu. This notion is not to be confused with the noise gradient ∇aψ\nabla_{a}\psi of a bounded C1C^{1} cylindrical function ψ\psi on XX (The Noise Gradient of a Cylindrical Function and the Noise Tangent Space at a Probability Measure on a Hilbert Space §gradient), which is a map on XX.

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…