Probabilistic frames and Wasserstein distances
Dongwei Chen and Martin Schmoll · 2025
Read Probabilistic frames and Wasserstein distances by Dongwei Chen and Martin Schmoll (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-2501-02602.txt.
Introduction
Dongwei Chen and Martin Schmoll use transport distances to study probabilistic frames in a 2025 math paper. Frames are flexible stand-ins for bases in signal work: a probabilistic frame spreads that structure over a cloud of points with weights, rather than a fixed list of vectors.
The authors measure gaps between such clouds with Wasserstein distances, the cost of shoveling one pile into the shape of another. Borrowing classic results, they show that clouds sharing the same frame operator can be carried into each other by a best-possible linear push. That bridge lets them carry over recent theorems and prove the clouds form one connected family. They also sort which related transport plans truly come from such pushes.
Catalog record
| CREATOR | Dongwei Chen and Martin Schmoll |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 82 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-02602
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Probabilistic frames and Wasserstein distances Dongwei Chen and Martin Schmoll Abstract . Probabilities with finite moments, in particular those which are probabilistic frames, are characterized and studied using p-Wasserstein distances. Adapting results from [ 3 ] , [ 8 ] and [ 11 ] to frame operators, we show that the sets of probabilistic frames with given frame operator, or equivalently given frame ellipsoid, are homeomorphic via an optimal linear push-forward. Applied to frame operators of couplings we study and generalize transport duals. In particular we obtain a non-degeneracy inequality identifying pairs of transport duals as frames. Finally we describe duals that do not arise as push-forwards and characterize those that are push-forwards. 2020 Mathematics Subject Classification Primary 60A10, 42C15; Secondary 49Q22 1. Introduction and Main Results In recent work [ 5 , 10 , 9 , 16 ] probabilistic frames, a set of Borel probability measures on {\mathbb{R}}^{n} that generalize frames, have been considered. A frame in {\mathbb{R}}^{n} is a finite set of vectors spanning {\mathbb{R}}^{n} , for a general background see [ 2 ] . Probabilistic p-frames on {\mathbb{R}}^{n} are Borel probability measures with finite p-th moments whose support, interpreted as set of vectors, spans {\mathbb{R}}^{n} [ 5 ] . A key motivation to consider probabilistic frames is the possible use of analytic techniques and the advantage to compare frames of various cardinality by Wasserstein distances. Let us denote the set of Borel probability measures on {\mathbb{R}}^{n} by {\mathcal{P}}({\mathbb{R}}^{n}) and by {\mathcal{P}}_{p}({\mathbb{R}}^{n}) those with finite p-th moments, i.e. {\mathcal{P}}_{p}({\mathbb{R}}^{n})=\{\mu\in{\mathcal{P}}({\mathbb{R}}^{n}):\ \int\|{\bf x}\|^{p}\ d\mu({\bf x})<\infty\} . Definition 1.1 . \mu\in\mathcal{P}_{p}(\mathbb{R}^{n}) is called probabilistic p-frame if there exists 0<A\leq B such that for any {\bf x}\in\mathbb{R}^{n} , A\|{\bf x}\|^{p}\leq\int_{\mathbb{R}^{n}}|\left\langle{\bf x},{\bf y}\right\rangle|^{p}\ d\mu({\bf y})\leq B\|{\bf x}\|^{p}. If, in addition A=B we call \mu tight, and if A=B=1 then \mu is called Parseval (probabilistic) frame. This standard definition of (probabilistic) frames does not directly provide a geometric intuition. Below we reformulate it as set of degeneracy distances using p-Wasserstein metrics. Briefly, the p-Wasserstein distance W_{p}(\cdot,\cdot) provides a metric space structure on {\mathcal{P}}_{p}({\mathbb{R}}^{n}) with convergence \mu_{n}\rightarrow\mu being equivalent to weak- \ast convergence together with convergence of the p-th moments: \int\|{\bf x}\|^{p}\ d\mu_{n}\rightarrow\int\|{\bf x}\|^{p}\ d\mu . Background on Wasserstein distances can be found in [ 7 , 13 , 14 ] . To reformulate the frame condition for a Borel probability let \pi_{{\bf x}^{\perp}} denote the orthogonal projection to the plane {\bf x}^{\perp} perpendicular to {\bf x} and (\pi_{{\bf x}^{\perp}})_{\#} be the associated push-forward on measures. Letting {\mathcal{P}}_{p}({\bf x}^{\perp}) denote the set of measures supported in {\bf x}^{\perp} with finite p -th moment. Proposition 1.2 . For any unit-vector {\bf x}\in S^{n-1} and any p\geq 1 W_{p}(\mu,(\pi_{{\bf x}^{\perp}})_{\#}\mu)=\left(\int_{\mathbb{R}^{n}}|\langle{\bf x},{\bf v}\rangle|^{p}d\mu({\bf v})\right)^{1/p}=\inf_{\nu\in{\mathcal{P}}_{p}({\bf x}^{\perp})}W_{p}(\mu,\nu). Since a probabilistic p-frame spans {\mathbb{R}}^{n} , its support cannot lie in a proper linear subspace, so it must have positive p-Wasserstein distance to measures supported in any linear subspace: Proposition 1.3 . A measure \mu\in{\mathcal{P}}_{p}({\mathbb{R}}^{n}) is a probabilistic p-frame if and only if W_{p}(\mu,(\pi_{{\bf x}^{\perp}})_{\#}\mu)>0 for all unit vectors {\bf x}\in S^{n-1} . It is a tight frame if and only if W_{p}(\mu,(\pi_{{\bf x}^{\perp}})_{\#}\mu)=C>0 for all {\bf x}\in S^{n-1} and a Parseval frame if and only if W_{p}(\mu,(\pi_{{\bf x}^{\perp}})_{\#}\mu)=1 for all {\bf x}\in S^{n-1} . Both propositions together (for proofs see section 3 ) indicate that the p-Wasserstein metrics capture the frame property and supply a geometric interpretation. Particularly interesting is the case p=2 where the respective Wasserstein distances are the eigenvalues of the frame operator . More precisely, if \mu\in{\mathcal{P}}_{2}({\mathbb{R}}^{n}) and {\bf x},{\bf y}\in\mathbb{R}^{n} then \langle{\bf x},{\bf S}_{\mu}{\bf y}\rangle:=\int_{\mathbb{R}^{n}}\langle{\bf x},{\bf v}\rangle\langle{\bf y},{\bf v}\rangle\ d\mu({\bf v}) is a linear operator. Given a basis of R^{n} , the matrix {\bf S}_{\mu}=\int_{\mathbb{R}^{n}}{\bf v}{\bf v}^{t}\ d\mu is a positive semi-definite, so that for {\bf x}\in S^{n-1} , (1.1) W^{2}_{2}(\mu,(\pi_{{\bf x}^{\perp}})_{\#}\mu)={\bf x}^{t}\int_{\mathbb{R}^{n}}{\bf v}{\bf v}^{t}\ d\mu({\bf v})\ {\bf x}={\bf x}^{t}\ {\bf S_{\mu}}\ {\bf x}. In particular {\bf S}_{\mu} is positive definite, if and only if \mu is a (probabilistic) frame. Should \mu not be a frame we still refer to {\bf S}_{\mu} as the frame operator. The frame ellipsoid {\mathcal{E}}_{\mu}:=\{{\bf S}^{1/2}_{\mu}\ {\bf x}:\ \|{\bf x}\|=1\}\subset\mathbb{R}^{n} associated with the root {\bf S}^{1/2}_{\mu} of {\bf S}_{\mu} , is a hyperellipsoid exactly if {\bf S}_{\mu} is definite, that is, if \mu is a probabilistic frame. The frame ellipsoid provides the 2-Wasserstein distance of a given (probabilistic) frame to the closest non-frame in any given direction. It can be seen as a generalized version of the Legendre ellipsoid as defined in [ 12 ] for symmetric, convex and compact bodies in {\mathbb{R}}^{n} , even though we do not represent the ellipsoid as a body or mass distribution in {\mathbb{R}}^{n} . Let {\mathbb{S}}^{n}_{+} be the set of non-negative definite n\times n symmetric matrices and {\mathbb{S}}^{n}_{++}\subset{\mathbb{S}}^{n}_{+} the subset of positive definite matrices. Further, let {\mathcal{P}}_{{\bf S}}\subset{\mathcal{P}}_{2}({\mathbb{R}}^{n}) denote the set of probabilities having frame operator {\bf S}\in{\mathbb{S}}^{n}_{+} and define W_{2}(\mu,{\mathcal{P}}_{{\bf S}}):=\inf_{\nu\in{\mathcal{P}}_{{\bf S}}}W_{2}(\mu,\nu) . The lower estimate from [ 3 ] adapted to probabilistic frames shows that the characteristic Wasserstein distances given in Proposition 1.2 are useful. Namely, if \{{\bf v}_{1},...,{\bf v}_{n}\} is an orthonormal basis of {\mathbb{R}}^{n} and \mu,\nu\in{\mathcal{P}}_{2}({\mathbb{R}}^{n}) , then (1.2) W^{2}_{2}(\mu,\nu)\geq\sum^{n}_{i=1}\left(W_{2}(\mu,(\pi_{{{\bf v}_{i}}^{\perp}})_{\#}\mu)-W_{2}(\nu,(\pi_{{{\bf v}_{i}}^{\perp}})_{\#}\nu)\right)^{2}, and equality holds if and only if \nu={\bf T}_{\#}\mu where {\bf T}\in{\mathbb{S}}^{n}_{+} has eigenbasis \{{\bf v}_{i}\} . In the equality case, similar to [ 8 ] and [ 11 ] for covariance operators, we obtain: Theorem 1.4 . For any {\bf S},{\bf A}\in{\mathbb{S}}^{n}_{++} the push-forward map {\bf A}_{\#}:{\mathcal{P}}_{{\bf S}}\rightarrow{\mathcal{P}}_{{\bf A}{\bf S}{\bf A}} is a homeomorphism, so that (1.3) W^{2}_{2}(\mu,{\bf A}_{\#}\mu)=W^{2}_{2}(\mu,{\mathcal{P}}_{{\bf A}{\bf S}{\bf A}})={\operatorname{tr}}\ {\bf S}({\bf{\operatorname{\bf Id}}}-{\bf A})^{2} for all \mu\in{\mathcal{P}}_{{\bf S}} and W_{2}(\mu,\nu)>W_{2}(\mu,{\bf A}_{\#}\mu) for any \nu\in{\mathcal{P}}_{{\bf A}{\bf S}{\bf A}} so that \nu\neq{\bf A}_{\#}\mu . As for the Wasserstein distance between frames with given frame operators, say {\bf S},{\bf T}\in{\mathbb{S}}^{n}_{++} , one applies Theorem 1.4 to the unique {\bf A}\in{\mathbb{S}}^{n}_{++} solving {\bf T}={\bf A}{\bf S}{\bf A} (see Proposition 4.1 ) given by (1.4) {\bf A}={\bf A}({\bf S},{\bf T}):={\bf S}^{-1/2}({\bf S}^{1/2}{\bf T}{\bf S}^{1/2})^{1/2}{\bf S}^{-1/2}. Since W_{2}(\mu,{\mathcal{P}}_{{\bf T}}) depends only on {\bf S} and not on the particular \mu\in{\mathcal{P}}_{{\bf S}} , d_{W}({\bf S},{\bf T}):=W_{2}({\mathcal{P}}_{{\bf S}},{\mathcal{P}}_{{\bf T}})=W_{2}(\mu,{\mathcal{P}}_{{\bf T}}) is well defined. Proposition 1.5 . Given {\bf S},{\bf T}\in{\mathbb{S}}^{n}_{+} . Then d_{W} is a metric on {\mathbb{S}}^{n}_{+} . More precisely, we have (1.5) d_{W}({\bf S},{\bf T})={\operatorname{tr}}({\bf S}+{\bf T}-2({\bf S}^{1/2}{\bf T}{\bf S}^{1/2})^{1/2}) If \|\cdot\|_{op} denotes the operator norm and \|\cdot\|_{F} the Frobenius norm, then (1.6) \|{\bf S}^{1/2}-{\bf T}^{1/2}\|_{op}\leq W_{2}(\mathcal{P}_{\bf S},\mathcal{P}_{\bf T})=d_{W}({\bf S},{\bf T})\leq\|{\bf S}^{1/2}-{\bf T}^{1/2}\|_{F}. In particular, the topology generated by d_{W} is the standard (norm) topology on {\mathbb{S}}^{n}_{+} . The proposition implies that the frame map \mathcal{S}:{\mathcal{P}}_{2}({\mathbb{R}}^{n})\rightarrow{\mathbb{S}}^{n}_{+} given by \mathcal{S}(\mu)={\bf S}_{\mu} is continuous; for a different proof, see [ 16 ] . The closely related metric d({\bf S},{\bf T}):=\sqrt{d_{W}({\bf S}^{2},{\bf T}^{2})} for symmetric matrices {\bf S},{\bf T}\in{\mathbb{S}}^{n} is by estimate 1.6 equivalent to norm-induced metrics, however, it is not induced by a norm [ 3 ] . Corollary 1.6 . Let p\in[1,\infty) , then the set of probabilistic p-frames in {\mathcal{P}}_{p}({\mathbb{R}}^{n}) is open in the p-Wasserstein topology on {\mathcal{P}}_{p}({\mathbb{R}}^{n}) . For p=2 this is easy to see, just compose the (continuous) frame map \mathcal{S} with the determinant map \det:{\mathbb{S}}^{n}_{+}\rightarrow{\mathbb{R}}_{\geq 0} , so that \det\circ\mathcal{S}:{\mathcal{P}}_{2}(\mathbb{R}^{n})\rightarrow{\mathbb{R}}_{\geq 0} is continuous. It follows, that the set of probabilistic frames {\mathcal{P}}_{++}:=\{\mu\in{\mathcal{P}}_{2}(\mathbb{R}^{n}):\det\circ\mathcal{S}(\mu)>0\}=\mathcal{S}^{-1}{\mathbb{S}}^{n}_{++} is open. Hence, the frame map \mathcal{S}:{\mathcal{P}}_{2}({\mathbb{R}}^{n})\rightarrow{\mathbb{S}}^{n}_{+}