Wasserstein Barycenter Convexity Detects Hilbertian Geometry

Bang-Xian Han, Deng-Yu Liu · 2026

Read Wasserstein Barycenter Convexity Detects Hilbertian Geometry by Bang-Xian Han, Deng-Yu Liu (2026) 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-2606-28213.txt.

Introduction

Bang-Xian Han and Deng-Yu Liu ask whether a space's geometry can be detected through optimal transport, the study of cheapest ways to move mass around. Their test uses Wasserstein barycenters, a kind of average of probability distributions, and asks whether a measure of disorder called Boltzmann entropy behaves convexly at those averages.

Han and Liu prove that if a finite-dimensional normed space passes this convexity test for every family of distributions, its norm must come from an inner product, making it a Hilbert space. Every such space passes a famous curvature condition, so their test is strictly sharper. Smooth Finsler manifolds passing the test must likewise have Riemannian tangent spaces. The proof handles flat faces of the unit ball without assuming strict convexity.

Catalog record

CREATORBang-Xian Han, Deng-Yu Liu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.28213

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Wasserstein Barycenter Convexity Detects Hilbertian Geometry Bang-Xian Han, Deng-Yu Liu Abstract We prove that convexity of the Boltzmann entropy at Wasserstein barycenters is strong enough to distinguish Hilbert spaces from general Banach spaces. Thus Wasserstein barycenters provide an intrinsic optimal-transport test for Hilbertian geometry. More precisely, we show that if a finite-dimensional normed vector space, equipped with Lebesgue measure, satisfies the Wasserstein Jensen’s inequality for the entropy at barycenters of arbitrary finite families of probability measures, then its norm must be induced by an inner product. This contrasts sharply with a well-known result: every finite-dimensional normed vector space satisfies the nonnegative Ricci curvature condition in the sense of Lott–Sturm–Villani, whereas barycenter convexity excludes all non-Hilbertian norms. As a consequence, smooth reversible Finsler manifolds satisfying the corresponding barycentric curvature-dimension condition have Riemannian tangent norms. The proof does not assume strict convexity of the norm. Its two main ingredients are a rank-one polarization argument, which yields the dual parallelogram identity in the strictly convex case, and a maximal-face trapping argument, which rules out flat faces of the unit ball. Keywords. Wasserstein barycenter; curvature-dimension condition; normed vector spaces; Finsler geometry; Hilbertianity; optimal transport. MSC 2020. Primary 53C23, 49Q22; Secondary 46B20, 53C60, 28A75. 1 Introduction The displacement convexity of entropy functionals along W_{2} geodesics, introduced by McCann in [ 15 ] , is an important theme in optimal transport. It relates optimal transport, lower Ricci curvature, and geometric functional inequalities. This point of view led Lott–Villani and Sturm to the synthetic curvature-dimension theory on metric-measure spaces [ 14 , 22 , 23 ] . The relation between curvature bounds, optimal transport, and Prékopa–Leindler type inequalities goes back to the Riemannian interpolation inequalities of Cordero-Erausquin–McCann–Schmuckenschläger [ 6 , 7 ] . In Lott–Sturm–Villani’s theory, curvature-dimension conditions are expressed through convexity of entropy along L^{2} -Wasserstein geodesics. This provides a formulation of lower Ricci bounds outside the smooth Riemannian setting. As Villani notes in “the last theorem in Old and New” [ 24 , p. 926] , every finite-dimensional normed vector space equipped with Lebesgue measure satisfies the standard synthetic non-negative Ricci condition in the weak Lott–Villani–Sturm sense. Thus Lott–Villani–Sturm’s \mathrm{CD} condition does not distinguish a Banach space from a Hilbert space. This is one reason for Ambrosio–Gigli–Savaré’s \mathrm{RCD} theory [ 2 , 8 ] . In that theory, Finsler structures are excluded by extra Riemannian assumptions: infinitesimal Hilbertianity, linearity of the heat flow, or equivalently the quadraticity of the Cheeger energy. For the Finsler side of the curvature-dimension theory, including weighted Ricci curvature and nonlinear analytic aspects, see also [ 16 , 17 , 18 , 20 ] . The basic question of this paper is whether one can detect the Riemannian, rather than merely Finsler, nature of a space using only entropy inequalities in optimal transport. For ordinary displacement convexity, the answer is negative: in the weak Lott–Villani–Sturm theory, finite-dimensional Banach spaces already satisfy the nonnegative curvature-dimension condition. Our result shows that the answer becomes positive once geodesic interpolation is replaced by Wasserstein barycenters. Motivated by Wasserstein barycenter geometry [ 1 , 13 ] , the barycenter curvature-dimension condition \mathrm{BCD} was introduced in [ 9 , 10 ] . In this condition displacement interpolation in the Lott–Sturm–Villani theory is replaced by Wasserstein barycenters, and the entropy functionals are required to satisfy the Wasserstein Jensen’s inequality at Wasserstein barycenters of finite weighted families of measures. Definition 1.1 (BCD condition) . Let (X,\mathsf{d},\mathfrak{m}) be a geodesic metric measure space. We say that it satisfies \mathrm{BCD}(K,\infty) if for every finite family \mu_{i}\in\mathcal{P}_{2}(X) and every set of weights \lambda_{i}>0 , \sum_{i}\lambda_{i}=1 , there exists \bar{\mu}\in\operatorname*{argmin}_{\eta\in\mathcal{P}_{2}(X)}\sum_{i}\lambda_{i}W_{2}^{2}(\eta,\mu_{i}) (called a Wasserstein barycenter) such that \operatorname{Ent}_{\mathfrak{m}}(\bar{\mu})\leq\sum_{i}\lambda_{i}\operatorname{Ent}_{\mathfrak{m}}(\mu_{i})-\frac{K}{2}\sum_{i}\lambda_{i}W_{2}^{2}(\bar{\mu},\mu_{i}), (1.1) where the Boltzmann entropy {\rm Ent}_{\mathfrak{m}}(\cdot) is defined by {\rm Ent}_{\mathfrak{m}}(\mu):=\left\{\begin{array}[]{ll}\int\rho\ln\rho\,\mathrm{d}\mathfrak{m}&\text{if }\mu\ll\mathfrak{m},\ \mu=\rho\,\mathfrak{m},\\ +\infty&\text{otherwise}.\end{array}\right. and W_{2}(\cdot,\cdot) is the 2-Wasserstein distance associated with \mathsf{d}^{2} . For two marginals, weighted Wasserstein barycenters are the corresponding points on W_{2} -geodesics. Thus the above condition gives \mathrm{BCD}(0,\infty)\Longrightarrow\mathrm{CD}(0,\infty) in the weak Lott–Villani sense, with the same entropy convention. On smooth Riemannian manifolds, this stronger condition has the expected meaning. In the smooth Riemannian setting, the implication from a Ricci lower bound to the Wasserstein Jensen’s inequality follows from the work of Kim–Pass [ 13 ] . Conversely, applying the two-marginal case recovers the usual displacement convexity of entropy, and the theorem of von Renesse–Sturm [ 25 ] gives the Ricci lower bound. The main theorem below states that, outside the Riemannian category, BCD has a rigidity that CD does not have. In particular, the barycenter curvature-dimension condition excludes the Finsler examples. A vivid analogy: “Finsler geometry may masquerade as Riemannian geometry along a single geodesic, but it inevitably betrays its non-quadratic nature at the barycenter under multi-directional tension.” Theorem 1.2 (Barycenter convexity rigidity) . If (E,\|\cdot\|,\mathcal{L}^{n}) is a finite-dimensional normed vector space with Lebesgue measure and satisfies \mathrm{BCD}(0,\infty) , then \|\cdot\| is induced by an inner product. Thus the passage from geodesics to barycenters is not a cosmetic strengthening of convexity: it changes the class of infinitesimal models. Together with the stability of the barycenter curvature-dimension condition under measured blow-up, the following corollary rules out non-Hilbertian reversible Finsler tangent norms. Its proof is given in Section 4 . Corollary 1.3 (Exclusion of non-Riemannian reversible Finsler tangents) . Let (M,F,\mathfrak{m}) be a smooth reversible Finsler manifold with a smooth positive reference measure. Assume that it satisfies \mathrm{BCD}(K,\infty) . Then F_{x} is induced by an inner product on T_{x}M for every x\in M . In particular, the Finsler structure is Riemannian. In finite-dimensional normed spaces and in the corresponding Finsler setting, BCD detects the Riemannian case. Thus the barycenter curvature-dimension condition gives a new synthetic Riemannian curvature-dimension condition: it implies the usual CD condition, but excludes the Finsler examples allowed by CD. Although RCD spaces satisfy BCD (cf. [ 9 ] ), the comparison with the RCD theory would be understood as a motivation rather than as an equivalence statement. In the RCD framework, the Riemannian character of the space is imposed through infinitesimal Hilbertianity. The present theorem points to another route: the Wasserstein Jensen’s inequality already carries a Riemannianity constraint. This suggests studying a curvature-dimension theory in which Wasserstein barycenters, rather than only Wasserstein geodesics, are the testing objects. In that direction, one may ask which analytic and geometric consequences follow directly from the Wasserstein Jensen’s inequality, and whether it yields stability, localization, rigidity, or regularity results that are not visible from two-point displacement convexity alone. The theorem also gives a Hilbertian characterization. Kakutani’s theorem [ 12 ] detects the inner product by requiring every two-dimensional section of the unit ball to be an ellipse; the Jordan–von Neumann theorem [ 11 ] does so through the parallelogram law. For background on Hilbertian characterizations, see [ 3 ] . The characterization here comes from the Wasserstein Jensen’s inequality. Proof strategy The proof has two main ingredients. In the strictly convex case, the argument works directly with calibrated rank-one transport branches; in the non-strictly convex case, it exploits the convex geometry of the unit ball. The rank-one polarization method. For every covector r , the square potential \alpha r(x)^{2}/2 generates the affine rank-one map x\longmapsto x-\alpha r(x)J(r). For small \alpha this potential is c -concave by a one-dimensional dual-norm estimate. The four-branch identity

Read the complete text