TheoremBase

Lower semicontinuity of the penalty makes the constant minus delta times the penalty an upper semicontinuous majorant on the open ball, which dominates the envelope. For the maximum, a maximising sequence has bounded penalty, so coercivity gives a convergent subsequence in the penalty domain; its limit stays in the closed ball and upper semicontinuity makes it a maximiser.

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. By Basic Properties of a Wasserstein-Coercive Penalty Pair (iii) the penalty E\mathcal{E} is lower semicontinuous on D\mathcal{D} relative to D\mathcal{D}, and by Basic Properties of a Wasserstein-Coercive Penalty Pair (ii) there is e0∈Re_{0}\in\mathbb{R} with e0≤E(ν)e_{0}\le\mathcal{E}(\nu) for every ν∈D\nu\in\mathcal{D}. The symmetry and the triangle inequality of W2W_{2} (The Quadratic Wasserstein Distance is a Metric on the Wasserstein Space (i), The Quadratic Wasserstein Distance is a Metric on the Wasserstein Space (iii)) are used without further mention.

Claim 1. Put O={ν∈P2(Rd):W2(ν,μ^)<r}O=\{\nu\in\mathcal{P}_{2}(\mathbb{R}^{d}):W_{2}(\nu,\hat{\mu})<r\}, which is open in (P2(Rd),W2)(\mathcal{P}_{2}(\mathbb{R}^{d}),W_{2}) by Open Ball in a Metric Space is Open, and S′=D∩OS'=\mathcal{D}\cap O, which contains μ^\hat{\mu} and is therefore nonempty. By The Delta-Envelopes of a Function on the Penalty Domain Relative to a Penalty Pair (i) the function u−δEu-\delta\mathcal{E} on D\mathcal{D} is bounded above near each point of D\mathcal{D} and uδ−=(u−δE)∗u^{-}_{\delta}=(u-\delta\mathcal{E})^{*}. By Local Bounds, Semicontinuous Envelopes and Local Extrema Are Unchanged on the Trace of an Open Set (i) its restriction (u−δE)∣S′(u-\delta\mathcal{E})|_{S'} is bounded above near each point of S′S', and by Local Bounds, Semicontinuous Envelopes and Local Extrema Are Unchanged on the Trace of an Open Set (ii)

((u−δE)∣S′)∗(ν)=uδ−(ν)(ν∈S′).\bigl((u-\delta\mathcal{E})|_{S'}\bigr)^{*}(\nu)=u^{-}_{\delta}(\nu)\qquad(\nu\in S').

Let h:S′→Rh:S'\to\mathbb{R} be h(ν)=c−δ E(ν)h(\nu)=c-\delta\,\mathcal{E}(\nu). We show that hh is upper semicontinuous on S′S'. Let ν0∈S′\nu_{0}\in S' and let ε∈R\varepsilon\in\mathbb{R} be positive; then εδ−1\varepsilon\delta^{-1} is positive by claims 7 and 5 of Elementary Order Arithmetic in an Ordered Field. By Lower Semicontinuous Function on a Subset of a Metric Space applied to E\mathcal{E} at ν0\nu_{0} with εδ−1\varepsilon\delta^{-1}, there is a positive θ\theta with E(ν0)−εδ−1<E(ν)\mathcal{E}(\nu_{0})-\varepsilon\delta^{-1}<\mathcal{E}(\nu) for every ν∈D\nu\in\mathcal{D} with W2(ν0,ν)<θW_{2}(\nu_{0},\nu)<\theta. For such ν\nu lying in S′S', multiplying by the positive δ\delta (claim 10 of Elementary Order Arithmetic in an Ordered Field) gives δ E(ν0)−ε<δ E(ν)\delta\,\mathcal{E}(\nu_{0})-\varepsilon<\delta\,\mathcal{E}(\nu), and hence, by claim 1 of that lemma applied twice, c−δ E(ν)<c−δ E(ν0)+εc-\delta\,\mathcal{E}(\nu)<c-\delta\,\mathcal{E}(\nu_{0})+\varepsilon, that is h(ν)<h(ν0)+εh(\nu)<h(\nu_{0})+\varepsilon. So hh is upper semicontinuous at ν0\nu_{0} relative to S′S' by Upper Semicontinuous Function on a Subset of a Metric Space.

For ν∈S′\nu\in S' we have u(ν)≤cu(\nu)\le c by hypothesis, so u(ν)−δ E(ν)≤h(ν)u(\nu)-\delta\,\mathcal{E}(\nu)\le h(\nu) by the compatibility of the order with addition. Properties of the Upper Semicontinuous Envelope (iii), applied on S′S' to (u−δE)∣S′(u-\delta\mathcal{E})|_{S'} and hh, gives ((u−δE)∣S′)∗(ν)≤h(ν)\bigl((u-\delta\mathcal{E})|_{S'}\bigr)^{*}(\nu)\le h(\nu) for ν∈S′\nu\in S'. With the identity displayed above, uδ−(ν)≤c−δ E(ν)u^{-}_{\delta}(\nu)\le c-\delta\,\mathcal{E}(\nu) for every ν∈D\nu\in\mathcal{D} with W2(ν,μ^)<rW_{2}(\nu,\hat{\mu})<r.

Claim 2. For ν∈K\nu\in K we have δe0≤δ E(ν)\delta e_{0}\le\delta\,\mathcal{E}(\nu) by claim 5 of Elementary Arithmetic in an Ordered Field, hence g(ν)≤c−δ E(ν)≤c−δe0g(\nu)\le c-\delta\,\mathcal{E}(\nu)\le c-\delta e_{0}. The set {g(ν):ν∈K}\{g(\nu):\nu\in K\} is therefore nonempty, since μ^∈K\hat{\mu}\in K, and bounded above by c−δe0c-\delta e_{0}; let ss be its least upper bound, which exists by The Real Numbers: Standing Notation and Background (iv). Then g(μ^)≤sg(\hat{\mu})\le s.

By Existence of a Sequence of Positive Real Numbers with Limit Zero there is a sequence (hn)n∈N(h_{n})_{n\in\mathbb{N}} of positive real numbers with limit 00; replacing hnh_{n} by min⁡{hn,1}\min\{h_{n},1\} we may assume hn≤1h_{n}\le1 for every nn, the new sequence being positive by claim 9 of Elementary Order Arithmetic in an Ordered Field and still having limit 00 by claim 2 of Order Properties of Limits of Real Sequences, squeezed between the constant sequence 00 and (hn)(h_{n}). For each n∈Nn\in\mathbb{N}, claim 3 of Approximation Property of the Supremum and the Infimum in R\mathbb{R}, applied with ε=hn\varepsilon=h_{n}, provides νn∈K\nu_{n}\in K with

s−hn<g(νn)≤s.s-h_{n}<g(\nu_{n})\le s .

Then g(μ^)−1≤s−hn<g(νn)g(\hat{\mu})-1\le s-h_{n}<g(\nu_{n}), so, using g(νn)≤c−δ E(νn)g(\nu_{n})\le c-\delta\,\mathcal{E}(\nu_{n}) and claim 1 of Elementary Order Arithmetic in an Ordered Field, δ E(νn)<c−g(μ^)+1\delta\,\mathcal{E}(\nu_{n})<c-g(\hat{\mu})+1, and multiplying by δ−1\delta^{-1} (claims 7 and 10 of that lemma),

E(νn)≤c′withc′=δ−1(c−g(μ^)+1)(n∈N).\mathcal{E}(\nu_{n})\le c'\qquad\text{with}\qquad c'=\delta^{-1}\bigl(c-g(\hat{\mu})+1\bigr)\qquad(n\in\mathbb{N}).

By Wasserstein-Coercive Penalty Pairs (i) the set {μ∈D:E(μ)≤c′}\{\mu\in\mathcal{D}:\mathcal{E}(\mu)\le c'\} is sequentially compact in (P2(Rd),W2)(\mathcal{P}_{2}(\mathbb{R}^{d}),W_{2}), so by Sequentially Compact Subset of a Metric Space there are ν∗\nu^{*} in that set, in particular ν∗∈D\nu^{*}\in\mathcal{D}, and a strictly increasing sequence (nk)k∈N(n_{k})_{k\in\mathbb{N}} in N\mathbb{N} such that (νnk)k∈N(\nu_{n_{k}})_{k\in\mathbb{N}} converges to ν∗\nu^{*}.

ν∗\nu^{*} lies in KK. Let ε∈R\varepsilon\in\mathbb{R} be positive. By Convergent Sequence in a Metric Space there is kk with W2(νnk,ν∗)<εW_{2}(\nu_{n_{k}},\nu^{*})<\varepsilon, and then W2(ν∗,μ^)≤W2(ν∗,νnk)+W2(νnk,μ^)≤ε+rW_{2}(\nu^{*},\hat{\mu})\le W_{2}(\nu^{*},\nu_{n_{k}})+W_{2}(\nu_{n_{k}},\hat{\mu})\le\varepsilon+r, since νnk∈K\nu_{n_{k}}\in K. As ε\varepsilon was arbitrary, W2(ν∗,μ^)≤rW_{2}(\nu^{*},\hat{\mu})\le r by Comparison of Real Numbers with Arbitrary Positive Slack (i), so ν∗∈K\nu^{*}\in K and g(ν∗)≤sg(\nu^{*})\le s.

ν∗\nu^{*} is a maximiser. The real sequence (hn)(h_{n}) converges to 00 in the metric space (R,dR)(\mathbb{R},d_{\mathbb{R}}), dR(x,y)=∣x−y∣d_{\mathbb{R}}(x,y)=|x-y| being the metric of The Absolute Value Metric on the Real Numbers, so that convergence in (R,dR)(\mathbb{R},d_{\mathbb{R}}) is exactly the condition of Limit of a Sequence of Real Numbers; by A Subsequence of a Convergent Sequence Has the Same Limit the subsequence (hnk)k∈N(h_{n_{k}})_{k\in\mathbb{N}} converges to 00 as well. Let ε∈R\varepsilon\in\mathbb{R} be positive. By Upper Semicontinuous Function on a Subset of a Metric Space there is a positive θ\theta with g(ν)<g(ν∗)+εg(\nu)<g(\nu^{*})+\varepsilon for every ν∈D\nu\in\mathcal{D} with W2(ν∗,ν)<θW_{2}(\nu^{*},\nu)<\theta. Choose kk so large that W2(νnk,ν∗)<θW_{2}(\nu_{n_{k}},\nu^{*})<\theta and hnk<εh_{n_{k}}<\varepsilon, which is possible because both conditions hold from some index on and the larger of the two indices serves. Then

s−ε<s−hnk<g(νnk)<g(ν∗)+ε,s-\varepsilon<s-h_{n_{k}}<g(\nu_{n_{k}})<g(\nu^{*})+\varepsilon ,

so s≤g(ν∗)+2εs\le g(\nu^{*})+2\varepsilon. As ε\varepsilon was arbitrary, s≤g(ν∗)s\le g(\nu^{*}) by Comparison of Real Numbers with Arbitrary Positive Slack (i) (applied with the positive numbers 2ε2\varepsilon, which exhaust the positive reals by claim 8 of Elementary Order Arithmetic in an Ordered Field). Together with g(ν∗)≤sg(\nu^{*})\le s this gives g(ν∗)=sg(\nu^{*})=s, and so g(ν)≤s=g(ν∗)g(\nu)\le s=g(\nu^{*}) for every ν∈K\nu\in K.

Citations

Loading…

Dependencies

Uses0

Loading…

Comments

Log in to comment.

Loading…