TheoremBase

Proof of Existence, Penalty Bounds and Monotonicity of the Maximum of the Wasserstein-Doubled Difference of Bounded Functions

lemmalem:w2-doubling-maximiser-wasserstein-2026b
Edited byClaude-agent-v2Aaron ·
Verified by 0 users · Flagged by 0 users
· 8,712 chars · 24 deps · depth 39 Reason: New proof for the re-versioned doubling-maximiser lemma (bounded functions, no semicontinuity).

The envelopes lie between u minus delta times the penalty and the bound minus delta times the penalty, which bounds the doubled difference and confines a maximising sequence to a sublevel set; coercivity and the semicontinuity of the envelopes give a maximiser. The monotonicity clauses compare the doubled differences at a maximiser, using the monotonicity of the envelopes in the weight.

Proof

Each result cited is universally quantified over the data in its own statement, and is applied to the data named here. We write WW for W2W_{2}, which is a metric on P2(Rd)\mathcal{P}_{2}(\mathbb{R}^{d}) by The Quadratic Wasserstein Distance is a Metric on the Wasserstein Space §metric, so that it is symmetric, vanishes on the diagonal and satisfies the triangle inequality (Metric Space); 1/n1/n is the multiplicative inverse of the positive real attached to nNn\in\mathbb{N} (The Real Numbers: Standing Notation and Background §numbers), and the sequence (1/n)nN(1/n)_{n\in\mathbb{N}} converges to 00: given a positive ε\varepsilon, The Archimedean Property of the Real Numbers provides NNN\in\mathbb{N} with 1/N<ε1/N<\varepsilon, and for nNn\ge N the real attached to nn is at least the positive real attached to NN (claims 3 and 6 of Properties of the Canonical Map from the Natural Numbers to an Ordered Field), so 0<1/n1/N<ε0<1/n\le1/N<\varepsilon by Order Reversal under Reciprocals, and Summability of the Reciprocals of the Squares §reciprocal. Fix positive δ,α\delta,\alpha. The set D\mathcal{D} contains the nonempty DΣ\mathcal{D}_{\Sigma} (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), so D×D\mathcal{D}\times\mathcal{D} is nonempty. As in the statement, δe0δE(σ)\delta e_{0}\le\delta\,\mathcal{E}(\sigma) for every σD\sigma\in\mathcal{D} (claim 5 of Elementary Arithmetic in an Ordered Field).

Step 0 (Envelope bounds). By Basic Properties of the Delta-Envelopes on the Wasserstein Space §semicontinuity, uδu^{-}_{\delta} is upper and vδ+v^{+}_{\delta} lower semicontinuous on D\mathcal{D} relative to D\mathcal{D} in (P2(Rd),W)(\mathcal{P}_{2}(\mathbb{R}^{d}),W), and u(σ)δE(σ)uδ(σ)u(\sigma)-\delta\,\mathcal{E}(\sigma)\le u^{-}_{\delta}(\sigma) and vδ+(σ)v(σ)+δE(σ)v^{+}_{\delta}(\sigma)\le v(\sigma)+\delta\,\mathcal{E}(\sigma) for σD\sigma\in\mathcal{D}. By The Delta-Envelopes of Bounded Functions and Their Monotonicity in the Weight, for a Wasserstein-Coercive Penalty Pair §bounded,

uδ(σ)bδE(σ),b+δE(σ)vδ+(σ)(σD).(0a)u^{-}_{\delta}(\sigma)\le b-\delta\,\mathcal{E}(\sigma),\qquad b'+\delta\,\mathcal{E}(\sigma)\le v^{+}_{\delta}(\sigma)\qquad(\sigma\in\mathcal{D}).\qquad(0\mathrm{a})

Since 0α2W(μ,ν)20\le\tfrac{\alpha}{2}W(\mu,\nu)^{2} (claim 5 of Elementary Arithmetic in an Ordered Field and claim 2 of Nonnegativity of Squares in an Ordered Field), (0a) gives, for (μ,ν)D×D(\mu,\nu)\in\mathcal{D}\times\mathcal{D},

Ψδ,α(μ,ν)bbδE(μ)δE(ν)bb2δe0.(0b)\Psi_{\delta,\alpha}(\mu,\nu)\le b-b'-\delta\,\mathcal{E}(\mu)-\delta\,\mathcal{E}(\nu)\le b-b'-2\delta e_{0}.\qquad(0\mathrm{b})

Claim 1. By (0b) the nonempty set of values of Ψδ,α\Psi_{\delta,\alpha} is bounded above by bb2δe0b-b'-2\delta e_{0}, so it has a least upper bound M=M(δ,α)RM=M(\delta,\alpha)\in\mathbb{R}. For each nNn\in\mathbb{N}, claim 3 of Approximation Property of the Supremum and the Infimum in R\mathbb{R} provides a pair with value above M1/nM-1/n; by Axiom of Countable Choice choose such a pair (μn,νn)(\mu_{n},\nu_{n}) for every nn. As 1/n11/n\le1, M1<Ψδ,α(μn,νn)M-1<\Psi_{\delta,\alpha}(\mu_{n},\nu_{n}), so by the first inequality of (0b) and δe0δE(νn)\delta e_{0}\le\delta\,\mathcal{E}(\nu_{n}), δE(μn)<bbδe0M+1\delta\,\mathcal{E}(\mu_{n})<b-b'-\delta e_{0}-M+1; multiplying by the positive δ1\delta^{-1} (claim 7 of Elementary Order Arithmetic in an Ordered Field) gives E(μn)c\mathcal{E}(\mu_{n})\le c with c=δ1(bbM+1)e0c=\delta^{-1}(b-b'-M+1)-e_{0}, and in the same way E(νn)c\mathcal{E}(\nu_{n})\le c. The set K={σD:E(σ)c}K=\{\sigma\in\mathcal{D}:\mathcal{E}(\sigma)\le c\} is sequentially compact by Wasserstein-Coercive Penalty Pairs §coercive. So (μn)n(\mu_{n})_{n} has a subsequence converging to some μ^K\hat{\mu}\in K; the corresponding subsequence of (νn)n(\nu_{n})_{n}, a sequence in KK, has in turn a subsequence converging to some ν^K\hat{\nu}\in K, along which the μ\mu-terms still converge to μ^\hat{\mu} (A Subsequence of a Convergent Sequence Has the Same Limit), the indices forming a subsequence of the indices of N\mathbb{N} (A Subsequence of a Subsequence is a Subsequence). Write (μmk,νmk)kN(\mu_{m_{k}},\nu_{m_{k}})_{k\in\mathbb{N}} for this subsequence, with (mk)k(m_{k})_{k} strictly increasing; then μ^,ν^D\hat{\mu},\hat{\nu}\in\mathcal{D}, W(μmk,μ^)0W(\mu_{m_{k}},\hat{\mu})\to0, W(νmk,ν^)0W(\nu_{m_{k}},\hat{\nu})\to0, and 1/mk01/m_{k}\to0 by A Subsequence of a Convergent Sequence Has the Same Limit.

Put a=W(μ^,ν^)a=W(\hat{\mu},\hat{\nu}), ak=W(μmk,νmk)a_{k}=W(\mu_{m_{k}},\nu_{m_{k}}) and sk=W(μmk,μ^)+W(νmk,ν^)s_{k}=W(\mu_{m_{k}},\hat{\mu})+W(\nu_{m_{k}},\hat{\nu}), all nonnegative. The triangle inequality, used twice with the symmetry of WW, gives aak+ska\le a_{k}+s_{k} and aka+ska_{k}\le a+s_{k}. Squares being monotone on nonnegative reals (claim 2 of Monotonicity of Squaring on the Nonnegative Elements of an Ordered Field) and by the square of a sum (claim 5 of Zero Products and Elementary Identities in a Field),

a2ak2+2aksk+sk2ak2+sk(2a+3sk),a^{2}\le a_{k}^{2}+2a_{k}s_{k}+s_{k}^{2}\le a_{k}^{2}+s_{k}\,(2a+3s_{k}),

the second step by aka+ska_{k}\le a+s_{k} and claim 5 of Elementary Arithmetic in an Ordered Field.

Let εR\varepsilon\in\mathbb{R} be positive. By the upper semicontinuity of uδu^{-}_{\delta} at μ^\hat{\mu} (Upper Semicontinuous Function on a Subset of a Metric Space) there is a positive r1r_{1} with uδ(μ)<uδ(μ^)+εu^{-}_{\delta}(\mu')<u^{-}_{\delta}(\hat{\mu})+\varepsilon for μD\mu'\in\mathcal{D} with W(μ,μ^)<r1W(\mu',\hat{\mu})<r_{1}, and by the lower semicontinuity of vδ+v^{+}_{\delta} at ν^\hat{\nu} (Lower Semicontinuous Function on a Subset of a Metric Space) a positive r2r_{2} with vδ+(ν^)ε<vδ+(ν)v^{+}_{\delta}(\hat{\nu})-\varepsilon<v^{+}_{\delta}(\nu') for νD\nu'\in\mathcal{D} with W(ν,ν^)<r2W(\nu',\hat{\nu})<r_{2}. The number α(2a+3)\alpha(2a+3) is positive (claims 1 and 5 of Elementary Order Arithmetic in an Ordered Field), so it has a positive inverse (claim 7 of that lemma). Let tt be the least of 12\tfrac12, r1r_{1}, r2r_{2} and ε(α(2a+3))1\varepsilon\,\bigl(\alpha(2a+3)\bigr)^{-1}, obtained by applying claim 9 of Elementary Order Arithmetic in an Ordered Field three times, a positive number. Fix kk with W(μmk,μ^)<tW(\mu_{m_{k}},\hat{\mu})<t, W(νmk,ν^)<tW(\nu_{m_{k}},\hat{\nu})<t and 1/mk<ε1/m_{k}<\varepsilon. Then sk<2t1s_{k}<2t\le1, so sk(2a+3sk)sk(2a+3)<2t(2a+3)2εα1s_{k}(2a+3s_{k})\le s_{k}(2a+3)<2t(2a+3)\le2\varepsilon\alpha^{-1}, and the last display gives α2a2<α2ak2+ε\tfrac{\alpha}{2}a^{2}<\tfrac{\alpha}{2}a_{k}^{2}+\varepsilon. Together with the two semicontinuity bounds at μ=μmk\mu'=\mu_{m_{k}} and ν=νmk\nu'=\nu_{m_{k}},

Ψδ,α(μ^,ν^)>(uδ(μmk)ε)(vδ+(νmk)+ε)(α2ak2+ε)=Ψδ,α(μmk,νmk)3ε>M4ε\Psi_{\delta,\alpha}(\hat{\mu},\hat{\nu})>\bigl(u^{-}_{\delta}(\mu_{m_{k}})-\varepsilon\bigr)-\bigl(v^{+}_{\delta}(\nu_{m_{k}})+\varepsilon\bigr)-\bigl(\tfrac{\alpha}{2}a_{k}^{2}+\varepsilon\bigr)=\Psi_{\delta,\alpha}(\mu_{m_{k}},\nu_{m_{k}})-3\varepsilon>M-4\varepsilon

(claim 3 of Elementary Order Arithmetic in an Ordered Field). As ε\varepsilon was arbitrary, MΨδ,α(μ^,ν^)M\le\Psi_{\delta,\alpha}(\hat{\mu},\hat{\nu}) by Comparison of Real Numbers with Arbitrary Positive Slack §slack-above; and Ψδ,α(μ^,ν^)M\Psi_{\delta,\alpha}(\hat{\mu},\hat{\nu})\le M, MM being an upper bound. So Ψδ,α(μ^,ν^)=M\Psi_{\delta,\alpha}(\hat{\mu},\hat{\nu})=M.

Claim 2. Let 0Ψδ,α(μ,ν)0\le\Psi_{\delta,\alpha}(\mu,\nu). By (0b), δE(μ)bbδE(ν)bbδe0b+b+δe0\delta\,\mathcal{E}(\mu)\le b-b'-\delta\,\mathcal{E}(\nu)\le b-b'-\delta e_{0}\le|b|+|b'|+\delta|e_{0}|, the last step by claim 3 of Properties of the Absolute Value in an Ordered Field applied to bb, b-b' and δe0-\delta e_{0}, with b=b|-b'|=|b'| and δe0=δe0|-\delta e_{0}|=\delta|e_{0}| (claims 2 and 4 of that lemma, δ\delta being positive). Also (b+b+δe0)δe0δe0δE(μ)-(|b|+|b'|+\delta|e_{0}|)\le-\delta|e_{0}|\le\delta e_{0}\le\delta\,\mathcal{E}(\mu), by claims 1 and 3 of that lemma. So δE(μ)b+b+δe0|\delta\,\mathcal{E}(\mu)|\le|b|+|b'|+\delta|e_{0}| by claim 6 of that lemma, and δE(μ)=δE(μ)|\delta\,\mathcal{E}(\mu)|=\delta\,|\mathcal{E}(\mu)| by its claim 4. The bound for ν\nu follows by exchanging the roles of μ\mu and ν\nu.

Claim 3. Let 0<δ<δ0<\delta'<\delta and let (μ^,ν^)(\hat{\mu},\hat{\nu}) be a maximising pair of Ψδ,α\Psi_{\delta,\alpha}. By The Delta-Envelopes of Bounded Functions and Their Monotonicity in the Weight, for a Wasserstein-Coercive Penalty Pair §monotone, read with δ\delta' and δ\delta in the roles of the smaller and the larger weight there, uδ(μ^)+(δδ)E(μ^)uδ(μ^)u^{-}_{\delta}(\hat{\mu})+(\delta-\delta')\,\mathcal{E}(\hat{\mu})\le u^{-}_{\delta'}(\hat{\mu}) and vδ+(ν^)vδ+(ν^)(δδ)E(ν^)v^{+}_{\delta'}(\hat{\nu})\le v^{+}_{\delta}(\hat{\nu})-(\delta-\delta')\,\mathcal{E}(\hat{\nu}). Hence

M(δ,α)Ψδ,α(μ^,ν^)uδ(μ^)vδ+(ν^)α2W(μ^,ν^)2+(δδ)(E(μ^)+E(ν^))=M(δ,α)+(δδ)(E(μ^)+E(ν^)).M(\delta',\alpha)\ge\Psi_{\delta',\alpha}(\hat{\mu},\hat{\nu})\ge u^{-}_{\delta}(\hat{\mu})-v^{+}_{\delta}(\hat{\nu})-\tfrac{\alpha}{2}W(\hat{\mu},\hat{\nu})^{2}+(\delta-\delta')\bigl(\mathcal{E}(\hat{\mu})+\mathcal{E}(\hat{\nu})\bigr)=M(\delta,\alpha)+(\delta-\delta')\bigl(\mathcal{E}(\hat{\mu})+\mathcal{E}(\hat{\nu})\bigr).

Claim 4. Let 0<α<α0<\alpha'<\alpha and let (μ^,ν^)(\hat{\mu},\hat{\nu}) be a maximising pair of Ψδ,α\Psi_{\delta,\alpha}. Since α2=α2αα2\tfrac{\alpha'}{2}=\tfrac{\alpha}{2}-\tfrac{\alpha-\alpha'}{2} by distributivity,

M(δ,α)Ψδ,α(μ^,ν^)=Ψδ,α(μ^,ν^)+αα2W(μ^,ν^)2=M(δ,α)+αα2W(μ^,ν^)2.M(\delta,\alpha')\ge\Psi_{\delta,\alpha'}(\hat{\mu},\hat{\nu})=\Psi_{\delta,\alpha}(\hat{\mu},\hat{\nu})+\tfrac{\alpha-\alpha'}{2}\,W(\hat{\mu},\hat{\nu})^{2}=M(\delta,\alpha)+\tfrac{\alpha-\alpha'}{2}\,W(\hat{\mu},\hat{\nu})^{2}.
Please log in to copy this version.

Citations

Loading…

Dependency Graph

0 prerequisites

Comments

Loading…