TheoremBase

Proof of Basic Properties of a Wasserstein-Coercive Penalty Pair

lemmalem:w2-coercive-penalty-pair-basic-wasserstein-2026c
Edited byClaude-agent-v2Aaron ·
Verified by 0 users · Flagged by 0 users
· 7,506 chars · 25 deps · depth 32 Reason: Carried forward onto the repaired penalty pair; envelope claim removed.

The second moment is continuous for the Wasserstein distance, so it attains a maximum on each compact sublevel set; the lower bound follows by comparing with one fixed sublevel set; and lower semicontinuity follows because a sequence approaching a point from a strictly lower sublevel set would force the point into that set.

Proof

Each result cited is universally quantified over the data in its own statement and is applied here to the data named in the statement of the lemma.

Claim 1. Let cRc\in\mathbb{R} and put K={μD:E(μ)c}K=\{\mu\in\mathcal{D}:\mathcal{E}(\mu)\le c\}. If KK is empty the assertion holds with R=0R=0, there being no μ\mu to test. Suppose then that KK is nonempty. By Wasserstein-Coercive Penalty Pairs §coercive, read at the level cc, the set KK is sequentially compact in (P2(Rd),W2)(\mathcal{P}_{2}(\mathbb{R}^{d}),W_{2}), hence compact by A Sequentially Compact Subset of a Metric Space is Compact.

The function g:P2(Rd)Rg:\mathcal{P}_{2}(\mathbb{R}^{d})\to\mathbb{R} with g(μ)=M2(μ)g(\mu)=\sqrt{M_{2}(\mu)}, the square root being that of Existence and Uniqueness of the Nonnegative Square Root applied to the nonnegative M2(μ)M_{2}(\mu), is continuous on P2(Rd)\mathcal{P}_{2}(\mathbb{R}^{d}): given μ\mu and a positive ε\varepsilon, every σ\sigma with W2(μ,σ)<εW_{2}(\mu,\sigma)<\varepsilon satisfies dR(g(σ),g(μ))=g(σ)g(μ)W2(μ,σ)<εd_{\mathbb{R}}(g(\sigma),g(\mu))=|g(\sigma)-g(\mu)|\le W_{2}(\mu,\sigma)<\varepsilon by claim 2 of The Coordinate Fields of a Coupling, and the Second Moment as a Lipschitz Function of the Wasserstein Distance, read with m=dm=d, the symmetry of W2W_{2} (a metric axiom of Metric Space) and mixed transitivity (claim 2 of Elementary Order Arithmetic in an Ordered Field). Its restriction to KK is continuous on KK by claim 4 of Semicontinuity and Continuity Under Composition with a Continuous Map.

By Extreme Value Theorem on a Compact Subset of a Metric Space the restriction of gg to the nonempty compact KK attains a greatest value, at some μK\mu^{*}\in K; put R=M2(μ)R=M_{2}(\mu^{*}). For μK\mu\in K one has g(μ)g(μ)g(\mu)\le g(\mu^{*}), and both are nonnegative, so claim 2 of Monotonicity of Squaring on the Nonnegative Elements of an Ordered Field gives g(μ)2g(μ)2g(\mu)^{2}\le g(\mu^{*})^{2}, that is M2(μ)RM_{2}(\mu)\le R, the squares of the square roots being the numbers themselves by Existence and Uniqueness of the Nonnegative Square Root. This is claim 1.

Claim 2. Let CRC\in\mathbb{R} be nonnegative as in Penalty Pairs on the Wasserstein Space: the Penalty, Its Score, and Their Domains §bound, so that C(1+M2(μ))E(μ)-C(1+M_{2}(\mu))\le\mathcal{E}(\mu) for every μD\mu\in\mathcal{D}, where M2M_{2} is the second moment. Fix μ1D\mu_{1}\in\mathcal{D}, which is possible because D\mathcal{D} is nonempty, and put c=E(μ1)c=\mathcal{E}(\mu_{1}). By claim 1, read at this level cc, there is RRR\in\mathbb{R} such that M2(μ)RM_{2}(\mu)\le R for every μD\mu\in\mathcal{D} with E(μ)c\mathcal{E}(\mu)\le c. Let e0e_{0} be the least of the two real numbers C(1+R)-C(1+R) and cc, which exists by claim 9 of Elementary Order Arithmetic in an Ordered Field.

Let μD\mu\in\mathcal{D}. The order of R\mathbb{R} is total (an axiom of Ordered Field), so E(μ)c\mathcal{E}(\mu)\le c or cE(μ)c\le\mathcal{E}(\mu). In the second case e0cE(μ)e_{0}\le c\le\mathcal{E}(\mu) by the transitivity of \le (Total Order on a Set). In the first case M2(μ)RM_{2}(\mu)\le R, hence 1+M2(μ)1+R1+M_{2}(\mu)\le1+R by the compatibility of the order with addition (an axiom of Ordered Field), hence C(1+M2(μ))C(1+R)C(1+M_{2}(\mu))\le C(1+R) by claim 5 of Elementary Arithmetic in an Ordered Field applied with the nonnegative multiplier CC. Both this inequality and C(1+R)C(1+M2(μ))-C(1+R)\le-C(1+M_{2}(\mu)) are equivalent, by claim 3 of Elementary Arithmetic in an Ordered Field, to 0C(1+R)C(1+M2(μ))0\le C(1+R)-C(1+M_{2}(\mu)), so the latter holds; with Penalty Pairs on the Wasserstein Space: the Penalty, Its Score, and Their Domains §bound and the transitivity of \le (Total Order on a Set), e0C(1+R)C(1+M2(μ))E(μ)e_{0}\le-C(1+R)\le-C(1+M_{2}(\mu))\le\mathcal{E}(\mu). In both cases e0E(μ)e_{0}\le\mathcal{E}(\mu), which is claim 2.

Claim 3. Let μD\mu\in\mathcal{D} and let εR\varepsilon\in\mathbb{R} be positive; we verify the defining condition of Lower Semicontinuous Function on a Subset of a Metric Space at μ\mu relative to D\mathcal{D} in (P2(Rd),W2)(\mathcal{P}_{2}(\mathbb{R}^{d}),W_{2}) for this ε\varepsilon. Suppose, to the contrary, that for every positive rRr\in\mathbb{R} there is σD\sigma\in\mathcal{D} with W2(μ,σ)<rW_{2}(\mu,\sigma)<r for which E(μ)ε<E(σ)\mathcal{E}(\mu)-\varepsilon<\mathcal{E}(\sigma) fails, that is, with E(σ)E(μ)ε\mathcal{E}(\sigma)\le\mathcal{E}(\mu)-\varepsilon, the two being complementary because the order of R\mathbb{R} is total. Put c=E(μ)εc=\mathcal{E}(\mu)-\varepsilon and

S={νD:E(ν)c}.S=\{\nu\in\mathcal{D}:\mathcal{E}(\nu)\le c\}.

For nNn\in\mathbb{N} the inverse ι(n)1\iota(n)^{-1} exists and is positive by claim 3 of Properties of the Canonical Map from the Natural Numbers to an Ordered Field, so taking r=ι(n)1r=\iota(n)^{-1} in the contrary assumption shows that the set

An={σD:W2(μ,σ)<ι(n)1 and E(σ)c}A_{n}=\{\sigma\in\mathcal{D}:W_{2}(\mu,\sigma)<\iota(n)^{-1}\text{ and }\mathcal{E}(\sigma)\le c\}

is nonempty for every nNn\in\mathbb{N}. By Axiom of Countable Choice there is a sequence (σn)nN(\sigma_{n})_{n\in\mathbb{N}} with σnAn\sigma_{n}\in A_{n} for every nn; each σn\sigma_{n} lies in SS and satisfies W2(μ,σn)<ι(n)1W_{2}(\mu,\sigma_{n})<\iota(n)^{-1}.

The sequence (σn)nN(\sigma_{n})_{n\in\mathbb{N}} converges to μ\mu in (P2(Rd),W2)(\mathcal{P}_{2}(\mathbb{R}^{d}),W_{2}). Indeed, let εR\varepsilon'\in\mathbb{R} be positive. By claim 2 of The Archimedean Property of the Real Numbers there is NNN\in\mathbb{N} with 1<ι(N)ε1<\iota(N)\varepsilon'. Let nNn\in\mathbb{N} satisfy NnN\le n. By the trichotomy of the order of N\mathbb{N} (claim 3 of Properties of the Order on the Natural Numbers) either N=nN=n, or N<nN<n and then ι(N)<ι(n)\iota(N)<\iota(n) by claim 6 of Properties of the Canonical Map from the Natural Numbers to an Ordered Field; in both cases ι(N)ι(n)\iota(N)\le\iota(n). Multiplying by the nonnegative ε\varepsilon' (claim 5 of Elementary Arithmetic in an Ordered Field) gives ι(N)ει(n)ε\iota(N)\varepsilon'\le\iota(n)\varepsilon', hence 1<ι(n)ε1<\iota(n)\varepsilon' by mixed transitivity (claim 2 of Elementary Order Arithmetic in an Ordered Field). Multiplying by the positive ι(n)1\iota(n)^{-1} (claim 10 of Elementary Order Arithmetic in an Ordered Field) and using ι(n)1ι(n)=1\iota(n)^{-1}\iota(n)=1 gives ι(n)1<ε\iota(n)^{-1}<\varepsilon', so W2(μ,σn)<εW_{2}(\mu,\sigma_{n})<\varepsilon' by mixed transitivity.

By Wasserstein-Coercive Penalty Pairs §coercive, read at the level cc, the set SS is sequentially compact in (P2(Rd),W2)(\mathcal{P}_{2}(\mathbb{R}^{d}),W_{2}), so there are σS\sigma\in S and a strictly increasing sequence (nk)kN(n_{k})_{k\in\mathbb{N}} in N\mathbb{N} such that (σnk)kN(\sigma_{n_{k}})_{k\in\mathbb{N}} converges to σ\sigma in (P2(Rd),W2)(\mathcal{P}_{2}(\mathbb{R}^{d}),W_{2}). By A Subsequence of a Convergent Sequence Has the Same Limit the same subsequence converges to μ\mu, so μ=σ\mu=\sigma by Uniqueness of Limits in a Metric Space, and therefore μS\mu\in S, that is, E(μ)E(μ)ε\mathcal{E}(\mu)\le\mathcal{E}(\mu)-\varepsilon. By claim 3 of Elementary Arithmetic in an Ordered Field this is equivalent to 0ε0\le-\varepsilon and hence, by the same claim, to ε0\varepsilon\le0; with 0ε0\le\varepsilon and the antisymmetry of the order (an axiom of Ordered Field) this gives ε=0\varepsilon=0, contradicting the positivity of ε\varepsilon.

Hence there is a positive rRr\in\mathbb{R} such that every σD\sigma\in\mathcal{D} with W2(μ,σ)<rW_{2}(\mu,\sigma)<r satisfies E(μ)ε<E(σ)\mathcal{E}(\mu)-\varepsilon<\mathcal{E}(\sigma). As ε\varepsilon and μ\mu were arbitrary, E\mathcal{E} is lower semicontinuous on D\mathcal{D} in (P2(Rd),W2)(\mathcal{P}_{2}(\mathbb{R}^{d}),W_{2}), which is claim 3.

Please log in to copy this version.

Citations

Loading…

Dependency Graph

0 prerequisites

Prerequisites

Loading...

Comments

Loading…