TheoremBase

The lower bound combines the moment bound on the zero sublevel set with the defining lower bound of a penalty pair; lower semicontinuity and completeness follow from the closed sublevel sets (the latter via completeness of the Wasserstein space), the distance bound from the product coupling, and coercive pairs are closed by uniqueness of limits and bounded by the extreme value theorem applied to the square root of the second moment.

Proof

Each result cited is universally quantified over the data in its own statement.

We work in the setting of The Intrinsic Calculus on the Wasserstein Space: Standing Notation, and elementary order and arithmetic of real numbers (including square roots and squares of nonnegative reals, and the Archimedean property) is carried by The Real Numbers: Standing Notation and Background §background, which we adopt for that purpose. Throughout, (P2(Rd),W2)(\mathcal{P}_{2}(\mathbb{R}^{d}),W_{2}) is a metric space by The Quadratic Wasserstein Distance is a Metric on the Wasserstein Space §metric, and W2W_{2} is symmetric by The Quadratic Wasserstein Distance is a Metric on the Wasserstein Space §symmetry. By Penalty Pairs on the Wasserstein Space: the Penalty, Its Score, and Their Domains §pair, D⊆P2(Rd)\mathcal{D}\subseteq\mathcal{P}_{2}(\mathbb{R}^{d}) and E:D→R\mathcal{E}:\mathcal{D}\to\mathbb{R}, so Dc⊆P2(Rd)\mathcal{D}_{c}\subseteq\mathcal{P}_{2}(\mathbb{R}^{d}) for every c∈Rc\in\mathbb{R}.

Claim 1 (bounded below). Step 1. Apply condition 2 of the Wasserstein-closed property with level 00: there is B0∈RB_{0}\in\mathbb{R} with M2(μ)≤B0M_{2}(\mu)\le B_{0} for every μ∈D\mu\in\mathcal{D} with E(μ)≤0\mathcal{E}(\mu)\le0. By Penalty Pairs on the Wasserstein Space: the Penalty, Its Score, and Their Domains §bound there is a nonnegative C∈RC\in\mathbb{R} with −C(1+M2(μ))≤E(μ)-C(1+M_{2}(\mu))\le\mathcal{E}(\mu) for every μ∈D\mu\in\mathcal{D}. Put e0=−C(1+∣B0∣)e_{0}=-C(1+|B_{0}|); since C≥0C\ge0 and 1+∣B0∣>01+|B_{0}|>0, we have e0≤0e_{0}\le0.

Step 2. Let μ∈D\mu\in\mathcal{D}. If E(μ)≤0\mathcal{E}(\mu)\le0, then M2(μ)≤B0≤∣B0∣M_{2}(\mu)\le B_{0}\le|B_{0}| by Step 1, so 1+M2(μ)≤1+∣B0∣1+M_{2}(\mu)\le1+|B_{0}| and, CC being nonnegative, e0=−C(1+∣B0∣)≤−C(1+M2(μ))≤E(μ)e_{0}=-C(1+|B_{0}|)\le-C(1+M_{2}(\mu))\le\mathcal{E}(\mu). If E(μ)>0\mathcal{E}(\mu)>0, then e0≤0<E(μ)e_{0}\le0<\mathcal{E}(\mu). In both cases e0≤E(μ)e_{0}\le\mathcal{E}(\mu).

Claim 2 (lower semicontinuity). We verify Lower Semicontinuous Function on a Subset of a Metric Space for the metric space (P2(Rd),W2)(\mathcal{P}_{2}(\mathbb{R}^{d}),W_{2}), the subset A=DA=\mathcal{D} and u=Eu=\mathcal{E}. Let x∈Dx\in\mathcal{D} and let ε∈R\varepsilon\in\mathbb{R} be positive; we argue by contradiction and suppose that no positive δ\delta has the required property.

Step 1. Then for every n∈Nn\in\mathbb{N}, applying the supposition to δ=1n\delta=\tfrac1n (positive, as nn is positive in R\mathbb{R} by The Real Numbers: Standing Notation and Background §numbers), there is yn∈Dy_{n}\in\mathcal{D} with W2(x,yn)<1nW_{2}(x,y_{n})<\tfrac1n and E(yn)≤E(x)−ε\mathcal{E}(y_{n})\le\mathcal{E}(x)-\varepsilon; we choose one such yny_{n} for each nn, obtaining a sequence (yn)n∈N(y_{n})_{n\in\mathbb{N}} in D\mathcal{D}.

Step 2. The sequence (yn)(y_{n}) converges to xx in (P2(Rd),W2)(\mathcal{P}_{2}(\mathbb{R}^{d}),W_{2}) in the sense of Convergent Sequence in a Metric Space: given a positive ε′\varepsilon', claim 3 of The Archimedean Property of the Real Numbers gives N∈NN\in\mathbb{N} with 0<1N<ε′0<\tfrac1N<\varepsilon', and for every n≥Nn\ge N we have 1n≤1N\tfrac1n\le\tfrac1N, hence W2(yn,x)=W2(x,yn)<1n<ε′W_{2}(y_{n},x)=W_{2}(x,y_{n})<\tfrac1n<\varepsilon' by symmetry of W2W_{2}.

Step 3. Put c=E(x)−εc=\mathcal{E}(x)-\varepsilon. Every yny_{n} lies in D\mathcal{D} with E(yn)≤c\mathcal{E}(y_{n})\le c, and (yn)(y_{n}) converges to x∈P2(Rd)x\in\mathcal{P}_{2}(\mathbb{R}^{d}), so condition 1 of the Wasserstein-closed property at level cc gives E(x)≤c=E(x)−ε\mathcal{E}(x)\le c=\mathcal{E}(x)-\varepsilon, that is ε≤0\varepsilon\le0, contradicting 0<ε0<\varepsilon. Hence some positive δ\delta has the property that every y∈Dy\in\mathcal{D} with W2(x,y)<δW_{2}(x,y)<\delta satisfies E(x)−ε<E(y)\mathcal{E}(x)-\varepsilon<\mathcal{E}(y): indeed the negation of this property for δ=1n\delta=\tfrac1n is exactly the existence of yny_{n} in Step 1, the negation of E(x)−ε<E(y)\mathcal{E}(x)-\varepsilon<\mathcal{E}(y) being E(y)≤E(x)−ε\mathcal{E}(y)\le\mathcal{E}(x)-\varepsilon by totality of the order. As x∈Dx\in\mathcal{D} and ε\varepsilon were arbitrary, E\mathcal{E} is lower semicontinuous on D\mathcal{D} relative to D\mathcal{D}.

Claim 3 (complete sublevel sets). Fix c∈Rc\in\mathbb{R}.

Step 1. By claim 1 of The Restriction of a Metric to a Subset Induces the Subspace Topology, applied to the metric space (P2(Rd),W2)(\mathcal{P}_{2}(\mathbb{R}^{d}),W_{2}) and the subset Dc\mathcal{D}_{c}, the restriction of W2W_{2} to Dc×Dc\mathcal{D}_{c}\times\mathcal{D}_{c} is a metric on Dc\mathcal{D}_{c}, so (Dc,W2)(\mathcal{D}_{c},W_{2}) is a metric space.

Step 2. Let (μn)n∈N(\mu_{n})_{n\in\mathbb{N}} be a Cauchy sequence in (Dc,W2)(\mathcal{D}_{c},W_{2}). Since the restricted metric takes the same values W2(μm,μℓ)W_{2}(\mu_{m},\mu_{\ell}), the sequence is also a Cauchy sequence in (P2(Rd),W2)(\mathcal{P}_{2}(\mathbb{R}^{d}),W_{2}), so by Completeness of the Quadratic Wasserstein Space over Euclidean Space §complete it converges in (P2(Rd),W2)(\mathcal{P}_{2}(\mathbb{R}^{d}),W_{2}) to some μ∈P2(Rd)\mu\in\mathcal{P}_{2}(\mathbb{R}^{d}).

Step 3. Each μn\mu_{n} lies in D\mathcal{D} with E(μn)≤c\mathcal{E}(\mu_{n})\le c, so condition 1 of the Wasserstein-closed property gives μ∈D\mu\in\mathcal{D} and E(μ)≤c\mathcal{E}(\mu)\le c, that is μ∈Dc\mu\in\mathcal{D}_{c}. The distances W2(μn,μ)W_{2}(\mu_{n},\mu) are the same in (Dc,W2)(\mathcal{D}_{c},W_{2}), so by Convergent Sequence in a Metric Space the sequence converges to the point μ\mu of Dc\mathcal{D}_{c} in (Dc,W2)(\mathcal{D}_{c},W_{2}). Hence (Dc,W2)(\mathcal{D}_{c},W_{2}) is complete.

Claim 4 (bounded distances). Let c,Bc,B be as in the claim and let μ,ν∈Dc\mu,\nu\in\mathcal{D}_{c}. Both lie in P2(Rd)⊆P(Rd)\mathcal{P}_{2}(\mathbb{R}^{d})\subseteq\mathcal{P}(\mathbb{R}^{d}), so the product coupling π=μ⊠ν\pi=\mu\boxtimes\nu belongs to Π(μ,ν)\Pi(\mu,\nu) by Couplings on Euclidean Space: Product Coupling, Swap, Finiteness of the Cost, Push-Forward Couplings, Modifying One Marginal, Quantisation, Gluing over a Finitely Supported Measure, and the Lipschitz Bound §product, and Couplings on Euclidean Space: Product Coupling, Swap, Finiteness of the Cost, Push-Forward Couplings, Modifying One Marginal, Quantisation, Gluing over a Finitely Supported Measure, and the Lipschitz Bound §cost-finite gives I(π)≤2M2(μ)+2M2(ν)≤2B+2B=4BI(\pi)\le2M_{2}(\mu)+2M_{2}(\nu)\le2B+2B=4B. By The Quadratic Wasserstein Distance on Euclidean Space §distance, W2(μ,ν)2≤I(π)W_{2}(\mu,\nu)^{2}\le I(\pi), hence W2(μ,ν)2≤4BW_{2}(\mu,\nu)^{2}\le4B.

Claim 5 (coercive pairs). Let (D′,DΣ′,E′,Σ′)(\mathcal{D}',\mathcal{D}'_{\Sigma},\mathcal{E}',\Sigma') be a Wasserstein-coercive penalty pair, not assumed to be Wasserstein-closed, and fix c∈Rc\in\mathbb{R}; we verify the two conditions of Wasserstein-Closed Penalty Pairs §w2-closed for this pair at level cc. Put Dc′={μ∈D′:E′(μ)≤c}\mathcal{D}'_{c}=\{\mu\in\mathcal{D}':\mathcal{E}'(\mu)\le c\}. By Wasserstein-Coercive Penalty Pairs §coercive, K=Dc′K=\mathcal{D}'_{c} is sequentially compact in (P2(Rd),W2)(\mathcal{P}_{2}(\mathbb{R}^{d}),W_{2}).

Step 1 (closed sublevel sets). Let (μn)n∈N(\mu_{n})_{n\in\mathbb{N}} be a sequence in D′\mathcal{D}' with E′(μn)≤c\mathcal{E}'(\mu_{n})\le c for every nn, converging to some μ∈P2(Rd)\mu\in\mathcal{P}_{2}(\mathbb{R}^{d}). Every μn\mu_{n} lies in KK, so by Sequentially Compact Subset of a Metric Space there are x∈Kx\in K and a strictly increasing sequence (nk)k∈N(n_{k})_{k\in\mathbb{N}} in N\mathbb{N} such that (μnk)k∈N(\mu_{n_{k}})_{k\in\mathbb{N}} converges to xx. By A Subsequence of a Convergent Sequence Has the Same Limit the subsequence (μnk)(\mu_{n_{k}}) also converges to μ\mu, so x=μx=\mu by Uniqueness of Limits in a Metric Space. Hence μ∈K\mu\in K, that is μ∈D′\mu\in\mathcal{D}' and E′(μ)≤c\mathcal{E}'(\mu)\le c.

Step 2 (bounded sublevel sets). If KK is empty, B=0B=0 satisfies condition 2 vacuously. Otherwise KK is nonempty and, by A Sequentially Compact Subset of a Metric Space is Compact, compact in P2(Rd)\mathcal{P}_{2}(\mathbb{R}^{d}) with the topology of open sets of (P2(Rd),W2)(\mathcal{P}_{2}(\mathbb{R}^{d}),W_{2}). Define f:K→Rf:K\to\mathbb{R} by f(μ)=M2(μ)f(\mu)=\sqrt{M_{2}(\mu)}, the nonnegative square root of the nonnegative real number M2(μ)M_{2}(\mu); here M2(μ)M_{2}(\mu) is a nonnegative real number since μ∈K⊆P2(Rd)\mu\in K\subseteq\mathcal{P}_{2}(\mathbb{R}^{d}), by The Second Moment of a Probability Measure on Euclidean Space and the Probability Measures with Finite Second Moment §moment. For x∈Kx\in K and positive ε\varepsilon take δ=ε\delta=\varepsilon: if y∈Ky\in K and W2(x,y)<δW_{2}(x,y)<\delta, then The Coordinate Fields of a Coupling, and the Second Moment as a Lipschitz Function of the Wasserstein Distance §lipschitz, read with m=dm=d and applied to yy and xx, together with symmetry of W2W_{2} gives ∣f(y)−f(x)∣≤W2(y,x)=W2(x,y)<ε|f(y)-f(x)|\le W_{2}(y,x)=W_{2}(x,y)<\varepsilon. Thus ff has the continuity property of Extreme Value Theorem on a Compact Subset of a Metric Space, which gives xmax⁡∈Kx_{\max}\in K with f(μ)≤f(xmax⁡)f(\mu)\le f(x_{\max}) for every μ∈K\mu\in K. Put B=M2(xmax⁡)B=M_{2}(x_{\max}). For μ∈K\mu\in K, squaring the inequality 0≤f(μ)≤f(xmax⁡)0\le f(\mu)\le f(x_{\max}) between nonnegative reals gives M2(μ)=f(μ)2≤f(xmax⁡)2=BM_{2}(\mu)=f(\mu)^{2}\le f(x_{\max})^{2}=B. This is condition 2 at level cc.

As cc was arbitrary, the pair (D′,DΣ′,E′,Σ′)(\mathcal{D}',\mathcal{D}'_{\Sigma},\mathcal{E}',\Sigma') is Wasserstein-closed. ■\blacksquare

Citations

Loading…

Dependencies

Uses0

Loading…

Comments

Log in to comment.

Loading…