ENERGY BOUNDS FOR KANTOROVICH TRANSPORT DISTANCES WITH CONVEX COST FUNCTIONS
Sergey G. Bobkov 1,3, Friedrich Götze 2,3 · 2025
Read ENERGY BOUNDS FOR KANTOROVICH TRANSPORT DISTANCES WITH CONVEX COST FUNCTIONS by Sergey G. Bobkov 1,3, Friedrich Götze 2,3 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-18324.txt.
Introduction
Sergey Bobkov and Friedrich Götze develop energy bounds for Kantorovich transport distances when the cost of moving mass grows in a convex way. Their main results extend estimates first proved by M. Ledoux for the standard transport distances, widening the range of costs for which such energy controls hold.
Catalog record
| CREATOR | Sergey G. Bobkov 1,3, Friedrich Götze 2,3 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 68 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2512-18324
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
ENERGY BOUNDS FOR KANTOROVICH TRANSPORT DISTANCES WITH CONVEX COST FUNCTIONS Thanks: 1) School of Mathematics, University of Minnesota, Minneapolis, MN, USA, [email protected]. Thanks: 2) Faculty of Mathematics, Bielefeld University, Germany, [email protected]. Thanks: 3) Research supported by the GRF – SFB 1283/2 2021 – 317210226 Sergey G. Bobkov 1,3, Friedrich Götze 2,3 Abstract . Energy bounds for Kantorovich transport distances are developed for convex cost functions. The main results extend estimates due to M. Ledoux for the Kantorovich distances W p W_{p} . 2010 Mathematics Subject Classification Primary 60E, 60F 1. Introduction Let {\mathfrak{P}}({\mathbb{R}}^{n}) denote the collection of all (Borel) probability measures on the Euclidean space {\mathbb{R}}^{n} and let \Pi(\mu,\nu) denote the collection of all probability measures \pi on {\mathbb{R}}^{n}\times{\mathbb{R}}^{n} with given marginals \mu and \nu in {\mathfrak{P}}({\mathbb{R}}^{n}) , called couplings or transference (transport) plans. The Kantorovich transport problem amounts to describe or estimate the transportation cost {\mathcal{T}}_{c}(\mu,\nu)=\inf_{\pi\in\Pi(\mu,\nu)}\int\!\!\!\int c(x,y)\,d\pi(x,y), where c:{\mathbb{R}}^{n}\times{\mathbb{R}}^{n}\rightarrow[0,\infty) is a fixed Borel measurable function, called cost function. Here cost functions of power type, that is c_{p}(x,y)=|x-y|^{p} , p\geq 1 , give rise to the transport metrics W_{p}={\mathcal{T}}_{c_{p}}^{1/p} in the space of all probability measures \mu on {\mathbb{R}}^{n} with finite p -th absolute moment \int|x|^{p}\,d\mu(x) . As it turns out, W_{p} may be estimated in terms of the dual Sobolev norm \|\nu-\mu\|_{H^{-1,p}(\mu)}=\sup\Big\{\int f\,d\nu-\int f\,d\mu:\|\nabla f\|_{L^{q}(\mu)}\leq 1\Big\}, (1.1) where the supremum runs over all C^{\infty} -smooth, compactly supported functions f on {\mathbb{R}}^{n} such that \|\nabla f\|_{L^{q}(\mu)}=\Big(\int|\nabla f|^{q}\,d\mu\Big)^{1/q}\leq 1,\quad q=p/(p-1) (with usual understanding in the case p=1 ). Recently M. Ledoux has proposed the following remarkable bound on the power Kantorovich distance W_{p} called energy estimate (motivated by the bounds for convergence rates of empirical measures with respect to W_{2} ). Theorem 1.1 ( [ 7 ] ). For all \mu and \nu in {\mathfrak{P}}({\mathbb{R}}^{n}) with finite p -th absolute moment, p\geq 1 , W_{p}(\mu,\nu)\leq p\,\|\nu-\mu\|_{H^{-1,p}(\mu)}. (1.2) This bound was actually derived in the setting of (complete connected) Riemannian manifolds M without boundary, assuming that \nu is absolutely continuous with respect to \mu on M (a condition that may be removed). Putting g=\varphi-1 , \varphi=d\nu/d\mu , M. Ledoux defined the dual Sobolev norm equivalently on functions g on M with \mu -mean zero by \|g\|_{H^{-1,p}(\mu)}=\Big(\int_{M}|\nabla(-T^{-1}g)|^{p}\,d\mu\Big)^{1/p}. Here T represents a second order linear differential operator which serves as the infinitesimal generator of the Markov semi-group associated to \mu , defined by \int_{M}u(-Tv)\,d\mu=\int_{M}\left<\nabla u,\nabla v\right>d\mu, and T^{-1} denotes the inverse of the non-negative operator -T . The aim of this note is to develop the “infimum-convolution” approach used in [ 7 ] and as a result – to extend (1.2) to more general cost functions of the form c(x,y)=L(x-y) , that is, for the optimal transportation cost {\mathcal{T}}_{L}(\mu,\nu)\,=\inf_{\pi\in\Pi(\mu,\nu)}\int\!\!\!\int L(x-y)\,d\pi(x,y). We assume that L is a convex function on {\mathbb{R}}^{n} such that L(0)=0 , L(x)>0 for x\neq 0 , and \frac{L(x)}{|x}\rightarrow\infty as |x|\rightarrow\infty . Note that the transport distance {\mathcal{T}}_{L}(\mu,\nu) does not need to be symmetric in (\mu,\nu) . It is also assumed that L satisfies a two-sided \Delta_{2} -condition L(2x)\leq cL(x) , valid for all x\in{\mathbb{R}}^{n} with some constant c . In this and only this case the associated Young function \Phi_{L}(r)\,=\sup_{x\neq 0}\,\frac{L(rx)}{L(x)},\quad r\geq 0, is finite. Necessarily it is convex on the positive half-axis (as the supremum of convex functions). In this setting the definition (1.1) has a natural generalization \|\nu-\mu\|_{H^{-1,L}(\mu)}=\sup\Big\{\int f\,d\nu-\int f\,d\mu:\int L^{*}(\nabla f)\,d\mu\leq 1\Big\}, (1.3) where L^{*}(x)=\sup_{y\in{\mathbb{R}}^{n}}\big[\left<x,y\right>-L(y)\big],\quad x\in{\mathbb{R}}^{n}, is the Legendre transform of L . The integrals in (1.3) are well-defined and finite (since f ’s are bounded), although in general the supremum may be infinite. Thus, in terms of the Luxemburg pseudo-norm generated by L^{*} , the value c=\|\nu-\mu\|_{H^{-1,L}(\mu)} represents an optimal constant in the inequality \int f\,d\nu-\int f\,d\mu\leq c\,\|\nabla f\|_{L^{*}(\mu)} in the class of all C^{\infty} -smooth functions f on {\mathbb{R}}^{n} with compact support. First we consider the “absolutely continuous” case. Theorem 1.2. Let \mu and \nu in {\mathfrak{P}}({\mathbb{R}}^{n}) satisfy \int\!\!\int L(x-y)\,d\mu(x)\,d\nu(y)<\infty . If \nu is absolutely continuous with respect to \mu which in turn is absolutely continuous with respect to the Lebesgue measure on {\mathbb{R}}^{n} , then {\mathcal{T}}_{L}(\mu,\nu)\leq A\,\Phi_{L}\big(\|\nu-\mu\|_{H^{-1,L}(\mu)}\big) (1.4) with constant A=\Phi_{L}(\Phi_{L}^{\prime}(1+)) . In particular, if L(x)=\|x\|^{p} for a norm \|\cdot\| on {\mathbb{R}}^{n} , we have \Phi_{L}(r)=r^{p} and A=p^{p} . In this case, (1.4) includes the Ledoux’ bound (1.2). Using a smoothing argument, one can give a slightly weakened version of Theorem 1.2, by removing the condition of absolute continuity, even without weakening of the constant for the homogeneous case of order p>1 where L(rx)=r^{p}L(x) for all x\in{\mathbb{R}}^{n} and r\geq 0 . Introduce the one-sided derivatives p_{+}=\Phi^{\prime}_{L}(1+) , p_{-}=\Phi^{\prime}_{L}(1-) and the associated quantity \gamma=\sup\big\{a+b:a+br\leq\min(r^{p_{+}},r^{p_{-}})\ {\rm for\ all}\ r\geq 0\big\}. In particular, 0<\gamma\leq 1 , and \gamma=1 if and only if p_{+}=p_{-}=p . The latter may be shown to be equivalent to the property that L is positive homogeneous of order p . Note that since \Phi_{L}(1)=1 , we have p_{+}\geq p_{-}\geq 1 . If p_{-}>1 , then necessarily L satisfies the super-linear growth condition L(x)/|x|\rightarrow\infty as |x|\rightarrow\infty . Theorem 1.3. If p_{-}>1 , then for all probability measures \mu and \nu on {\mathbb{R}}^{n} such that \int\!\!\int L(x-y)\,d\mu(x)\,d\nu(y)<\infty , the inequality (1.4) holds true with constant A=\Phi_{L}(p_{+})\,\Phi_{L}(1/\gamma). The paper is a organized as follows. First we recall basic results on Kantorovich duality, infimum-convolution operators and Hamilton-Jacobi equations (Sections 2-6). The two-sided \Delta_{2} -condition is discussed in Sections 7-8, including one special boundary value problem about the Young functions \Phi_{L} . Basic definitions and results about the Luxemburg and Orlicz pseudo-norms for vectors valued-functions are collected in Section 9. Sections 10-11 are devoted to properties of dual Sobolev norms (the lower semi-continuity and continuity along convolutions needed in the smoothing argument). Proofs of Theorems 1.2 and 1.3 are completed in Section 12. 2. Kantorovich Duality For reader’s convenience, here we collect several general facts about the Kantorovich duality. Given a separable metric space M , denote by {\mathfrak{P}}(M) the collection of all Borel probability measures \mu on M . The product space M\times M^{\prime} of two separable metric spaces is metrizable and separable. Fix a lower semi-continuous (cost) function c:M\times M^{\prime}\rightarrow[0,\infty) . Given \mu\in{\mathfrak{P}}(M) and \nu\in{\mathfrak{P}}(M^{\prime}) , denote by \Pi(\mu,\nu) the collection of all Borel probability measures \pi on M\times M^{\prime} with marginals \mu and \nu . Like in the Euclidean space, the problem is how to describe or estimate the transportation cost {\mathcal{T}}_{c}(\mu,\nu)=\inf_{\pi\in\Pi(\mu,\nu)}\int\!\!\!\int c(x,y)\,d\pi(x,y). (2.1) The set \Pi(\mu,\nu) is non-empty, since it contains, for example, the product measure \pi=\mu\otimes\nu , which represents a transference transport plan. Hence {\mathcal{T}}_{c}(\mu,\nu) is well-defined. If both M and M^{\prime} are complete separable metric spaces, \Pi(\mu,\nu) is compact in {\mathfrak{P}}(M\times M^{\prime}) in the weak topology, by Prokhorov’s compactness criterion. Since the cost function is lower semi-continuous, the infimum in (2.1) is attained for some plan \pi , called an optimal plan. The next theorem is known as the Kantorovich duality, cf. Villani [ 13 ] , Theorem 1.3. Proposition 2.1. Let M,M^{\prime} be complete separable metric spaces. For all \mu\in{\mathfrak{P}}(M) and \nu\in{\mathfrak{P}}(M^{\prime}) , {\mathcal{T}}_{c}(\mu,\nu)=\sup\bigg[\int_{M}f\,d\mu+\int_{M^{\prime}}g\,d\nu\bigg], (2.2) where the supremum is running over all functions f\in L^{1}(\mu) and g\in L^{1}(\nu) such that f(x)+g(y)\leq c(x,y) (2.3) for \mu -almost all x\in M and \nu -almost all y\in M^{\prime} . Subject to (2.3) , the supremum in (2.2) may be restricted to functions f,g that are