Paperlane

no. 16 · Tuesday 17 November 2026

The Stability of Wasserstein Shadow Projection and its Application to Transfer Learning

Jakwang Kim

math.ST · v2 · 26 pp

Abstract

In this paper, we study the stability of the shadow, a projection of a probability measure onto the set of couplings with prescribed marginals in the Wasserstein space, and its application to transfer learning. The shadow was introduced by \citet{EcksteinNutz2022} to analyze the stability of the Sinkhorn algorithm and was recently revisited by \citet{kim2026extensioncouplingprojectionoptimal} for statistical applications, particularly for reconstructing joint distributions from unpaired marginal observations. Under suitable regularity assumptions on Brenier potentials, we establish quantitative continuity of the shadow. As a key ingredient in our analysis, we also establish a stability bound for optimal transport maps under a smoothness assumption based on the previous results \citet{Li_Nochetto_2021, balakrishnan2025stabilityboundssmoothoptimal}. Also, we show the uniqueness of optimal transport projection provided that source measure is absolutely continuous, and the separable cost satisfies some regular conditions introduced by \cite{gangbo1996geometry}, which includes the $p$-th power of distance for $1 < p <\infty$. As an application, we investigate transfer learning from the perspective of shadow projection. Specifically, we consider a transport-based domain adaptation model with unpaired target covariate and response observations. We bound the generalization error of it by the approximation error and the distance between the shadow and its empirical counterpart, and also show and the consistency of the proposed estimator under smooth the Brenier potentials transporting the source marginals to their target counterparts. Additionally, assuming their strong convexity, we derive sample-complexity bounds and concentration of the generalization error. In particular, our analysis applies to suitable families of strongly log-concave measures.

comments
26 pages; fix error in the previous version by assuming regular Brenier potentials; add application of Wasserstein projection to transfer learning
fields
Statistics Theory, Optimization and Control, Probability
pdf ↗ page arXiv:2604.17711 ↗