Stability of Wasserstein projections in convex order via metric extrapolation
Jakwang Kim, Young-Heon Kim, Andrea Natale · 2025
Read Stability of Wasserstein projections in convex order via metric extrapolation by Jakwang Kim, Young-Heon Kim, Andrea Natale (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-2507-19221.txt.
Introduction
Jakwang Kim and co-authors prove stability bounds for a pair of projection problems in the geometry of probability. Their 2025 paper works with Wasserstein distance, the standard way to measure how far apart two distributions lie. Projecting a distribution onto those that dominate it in convex order has both backward and forward forms, each with uses in economics and probability.
Catalog record
| CREATOR | Jakwang Kim, Young-Heon Kim, Andrea Natale |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 35 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-19221
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Stability of Wasserstein projections in convex order via metric extrapolation Jakwang Kim, Young-Heon Kim, Andrea Natale Abstract . We build on recent work linking backward and forward W 2 W_{2} -projections in convex order with the recently introduced metric extrapolation problem to derive new quantitative stability estimates for both problems. 1. Introduction Given two probability measures with finite second moments \mu,\nu\in\mathcal{P}_{2}(\mathbb{R}^{d}) , we say that \mu is dominated by \nu in convex order, denoted \mu\preceq\nu , if for any convex function \varphi:\mathbb{R}^{d}\rightarrow\mathbb{R} , \int_{\mathbb{R}^{d}}\varphi\,\mathrm{d}\mu\leq\int_{\mathbb{R}^{d}}\varphi\,\mathrm{d}\nu\,. In this work we consider the backward and forward 2-Wasserstein projections in convex order, defined as follows: \text{(backward)}~\mathscr{P}^{cx}_{\preceq\nu}(\mu)\coloneqq\argmin_{\rho\preceq\nu}W_{2}^{2}(\rho,\mu)\,,\quad\text{(forward)}~\mathscr{P}^{cx}_{\mu\preceq}(\nu)\coloneqq\argmin_{\mu\preceq\rho}W_{2}^{2}(\rho,\nu)\,. Here, the 2-Wasserstein distance W_{2}(\rho,\nu) is defined as \displaystyle W_{2}^{2}(\rho,\nu)=\inf_{\pi\in\Pi(\rho,\nu)}\int_{\mathbb{R}^{d}\times\mathbb{R}^{d}}|x-y|^{2}\mathrm{d}\pi(x,y) where \Pi(\rho,\nu) denotes the set of couplings between \rho and \nu . The backward projection problem always admits a unique solution so \mathscr{P}^{cx}_{\preceq\nu}(\mu) is always well-defined [ 7 , 9 ] . In contrast, the forward problem admits a unique solution as long as the measure \nu being projected is absolutely continuous [ 9 ] . In this work we study the stability of \mathscr{P}^{cx}_{\preceq\nu}(\mu) and \mathscr{P}^{cx}_{\mu\preceq}(\nu) with respect to \mu and \nu in arbitrary dimensions. The main idea is to exploit a recently discovered duality between the backward projection problem and the so-called metric extrapolation problem. 1.1. Metric extrapolation The metric extrapolation problem is a variational problem introduced in [ 5 , 6 ] to define extensions of constant-speed length-minimizing geodesics in the Wasserstein space (referred to simply as minimizing geodesics in the following). These are curves \gamma:[0,t]\rightarrow\mathcal{P}_{2}(\mathbb{R}^{d}) satisfying W_{2}(\gamma(s_{1}),\gamma(s_{2}))=\frac{|s_{2}-s_{1}|}{t}W_{2}(\gamma(0),\gamma(t)) for all s_{1},s_{2}\in[0,t] . While for any \nu_{0},\nu_{1}\in\mathcal{P}_{2}(\mathbb{R}^{d}) there always exists a minimizing geodesic on [0,1] such that \gamma(0)=\nu_{0} and \gamma(1)=\nu_{1} , there might not be any minimizing geodesic with the same property defined on a larger interval [0,t] , with t>1 . The metric extrapolation of a geodesic from \nu_{0} to \nu_{1} at any time t>1 is however always well-defined as the unique minimizer of the following functional: (1) \mathscr{E}^{t}(\nu_{0},\nu_{1})\coloneqq\argmin_{\rho}\left\{\frac{W_{2}^{2}(\rho,\nu_{1})}{2(t-1)}-\frac{W_{2}^{2}(\rho,\nu_{0})}{2t}\right\}\,. By elementary inequalities, if there exists a minimizing geodesic \gamma:[0,t]\rightarrow\mathcal{P}_{2}(\mathbb{R}^{d}) such that \gamma(0)=\nu_{0} and \gamma(1)=\nu_{1} then necessarily \mathscr{E}^{t}(\nu_{0},\nu_{1})=\gamma(t) . When this is not the case however, [ 6 ] found that there still exists a minimizing geodesic \tilde{\gamma}:[0,t]\rightarrow\mathcal{P}_{2}(\mathbb{R}^{d}) such that \tilde{\gamma}(1)=\nu_{1} and \tilde{\gamma}(t)=\mathscr{E}^{t}(\nu_{0},\nu_{1}) , but \tilde{\gamma}(0)\neq\nu_{0} is precisely \mathscr{P}_{\preceq\nu_{0}}^{cx}(\mu) , where \mu is a dilated version of \nu_{1} (see Theorem 1 for a precise statement). 1.2. Main results In this work we review and make explicit the correspondence between metric extrapolation and the (backward and forward) W_{2} -projections in convex order (Theorem 1 ), by assembling recent results established in [ 6 , 3 , 7 , 9 ] . We use such correspondence to establish quantitative stability estimates for both problems. First, exploiting the fact that the backward projection problem with \nu fixed can be lifted to a projection on a Hilbert space, we prove that the map \mu\mapsto\mathscr{P}_{\preceq\nu}^{cx}(\mu) is non-expansive (Lemma 2 ). In turn, we show that this fact implies a Lipschitz estimate also on the map \nu_{1}\mapsto\mathscr{E}^{t}(\nu_{0},\nu_{1}) (Proposition 1 ). We further show that the map \nu\mapsto\mathscr{P}_{\preceq\nu}^{cx}(\mu) is 1/2-Hölder (Proposition 4 ). While finalizing our manuscript, we found that Lemma 2 and Proposition 4 had also been independently established in a recent work [ 1 , Proposition 2.2, Proposition 2.9] by Alfonsi and Jourdain; Lemma 2 by a similar method and Proposition 4 by a different method. Their work further includes a detailed study of the special case where both \mu and \nu are Gaussians. Our approach is novel, for in particular, it can be directly applied to the forward projection problem to show that, upon some regularity assumption, the map (\mu,\nu)\mapsto\mathscr{P}_{\mu\preceq}^{cx}(\nu) is also Hölder continuous (Proposition 5 ). For this we use a recent stability result on Wasserstein barycenters [ 4 , Theorem 1.5] . Note that even if the latter results may not be sharp, the fundamental idea of this work is that in view of Theorem 1 , stability results on the extrapolation problem can be transferred to the projection problems and vice-versa. 2. Projections in convex order and metric extrapolation In this section, we recall the equivalence between the W_{2} -projections in convex order and the metric extrapolation problem. This will be central to establish the stability of the backward and forward projections. We start by recalling that if \nu_{0}\in\mathcal{P}_{2}^{ac}(\mathbb{R}^{d}) is absolutely continuous, then for any \nu_{1}\in\mathcal{P}_{2}(\mathbb{R}^{d}) there exists a uniquely defined minimizing geodesic \gamma:[0,1]\rightarrow\mathcal{P}_{2}(\mathbb{R}^{d}) such that \gamma(1)=\nu_{1} . This is given by \gamma(s)=((1-s)\mathrm{Id}+s\nabla u)_{\#}\nu_{0} for all s\in[0,1] , where \# denotes the push-forward of measures and \nabla u is the gradient of a convex function providing the unique optimal transport map from \nu_{0} to \nu_{1} , i.e., \pi=(\mathrm{Id},\nabla u)_{\#}\nu_{0} is the unique optimal coupling between \nu_{0} and \nu_{1} . For any \lambda>0 we denote by D^{\lambda}:\mathbb{R}^{d}\rightarrow\mathbb{R}^{d} the dilation map defined by D^{\lambda}(x)\coloneqq\lambda x\,. The following theorem summarizes some of the results from [ 6 ] , and combines them with those from [ 2 , 7 , 9 ] . Theorem 1 . Let t>1 , \theta\coloneqq(t-1)/t , and let \nu_{0},\nu_{1}\in\mathcal{P}_{2}(\mathbb{R}^{d}) be arbitrary measures. The following holds: (1) There exists a unique minimizing geodesic \gamma:[0,t]\rightarrow\mathcal{P}_{2}(\mathbb{R}^{d}) such that \gamma(1)=\nu_{1}~\text{ and }~\gamma(t)=\mathscr{E}^{t}(\nu_{0},\nu_{1}). Moreover \gamma(0)=\mathscr{P}^{cx}_{\preceq\nu_{0}}(D^{1/\theta}_{\#}\nu_{1}) and there exists a lower semi-continuous \theta -strongly convex potential u:\mathbb{R}^{d}\rightarrow\mathbb{R}\cup\{+\infty\} such that \gamma(s)=(s\mathrm{Id}+(1-s)\nabla u^{*})_{\#}\nu_{1} for all s\in[0,t] . Here u^{*} is the Legendre transform of u . (2) If \nu_{0} is absolutely continuous, there exists a unique minimizing geodesic \tilde{\gamma}:[0,t]\rightarrow\mathcal{P}_{2}(\mathbb{R}^{d}) such that \tilde{\gamma}(0)=\nu_{0}\quad\text{and}\quad\tilde{\gamma}(t)=\mathscr{E}^{t}(\nu_{0},\nu_{1})\,. Moreover, letting \mu\coloneqq{D_{\#}^{1/\theta}\nu_{1}} , \tilde{\gamma}(1)=D^{\theta}_{\#}\mathscr{P}^{cx}_{\mu\preceq}(\nu_{0}) and \tilde{\gamma}(s)=((1-s)\mathrm{Id}+s\nabla u)_{\#}\nu_{0} for all s\in[0,t] , and for the same potential u as in point (1) . Proof. The first point is a direct consequence of [ 6 , Theorem 3.3] stating the equivalence (up to a scaling factor) between the metric extrapolation problem and the weak optimal transport characterization of the backward projection (see [ 7 , Proposition 1.1] or [ 2 , Theorem 2.1] ); see also Remark 4.4 in [ 3 ] for a direct argument for the equivalence of the two problems in a semi-discrete setting. The expression for \gamma is provided in [ 6 , Theorem 2.6] . To prove the second point, let v^{*}\coloneqq u^{*}(\theta\,\cdot)/\theta , with u from point (1) and u^{*} be its Legendre dual. It is easy to check that \nabla v^{*} is 1-Lipschitz (since \nabla u^{*} is {1/\theta} -Lipschitz on \mathbb{R}^{d} from the \theta -strong convexity of u ) and from the equation for \gamma in point (1), (2) \nabla v^{*}_{\#}(D^{1/\theta}_{\#}\nu_{1})=\mathscr{P}^{cx}_{\preceq\nu_{0}}(D^{1/\theta}_{\#}\nu_{1}). By Theorem 8.3 and 8.5 in [ 9 ] , the Legendre dual of v^{*} , that is, v^{**}=u/\theta is the optimal potential (i.e., \nabla u/\theta is the optimal transport map) from \nu_{0} to \mathscr{P}^{cx}_{\mu\preceq}(\nu_{0}) , where \mu\coloneqq D_{\#}^{1/\theta}\nu_{1} . In particular, \nabla u_{\#}\nu_{0}=D^{\theta}_{\#}\mathscr{P}^{cx}_{\mu\preceq}(\nu_{0}). On the other hand, as a consequence of Theorem 2.6 in [ 6 ] (see in particular equation (2.26) therein), ((1-t)\mathrm{Id}+t\nabla u)_{\#}\nu_{0}=\mathscr{E}^{t}(\nu_{0},\nu_{1}). Finally, since u is \theta -strongly convex and \nu_{0} is absolutely continuous, \tilde{\gamma} , given by \tilde{\gamma}(s)=((1-s)\mathrm{Id}+s\nabla u)_{\#}\nu_{0},\quad s\in[0,t], is the unique minimizing geodesic connecting its endpoints. ∎ 3. Non-expansiveness of the backward projection in convex order In this section we prove that for any \nu\in\mathcal{P}_{2}(\mathbb{R}^{d}) , the map \mathscr{P}_{\preceq\nu}^{cx}(\cdot) is non-expansive. This will also imply a new Lipschitz stability estimate on the metric extrapolation. For a given