Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
We analyze the first-order $\Gamma$-expansion of entropy-regularized optimal transport (EOT) as the regularization parameter $\varepsilon$ approaches zero, by reformulating EOT as a relative entropy minimization against a Gibbs measure. Under mild assumptions on the duality gap, contact set, and marginals, we prove convergence of the Gibbs measures and derive this expansion for non-degenerate ground costs. This provides an entropic selection principle and an $o(\varepsilon)$ expansion of the EOT cost, generalizing prior work.
The vanishing regularization limit of EOT reveals an entropic selection principle and an $o(\varepsilon)$ cost expansion for non-degenerate costs.
Entropy-regularized optimal transport (EOT) has become a key tool in optimal transport, providing both theoretical insight and efficient computation via the Sinkhorn algorithm. We study the first-order \(Γ\)-expansion of EOT as the regularization parameter \(\varepsilon\) vanishes, by rewriting EOT as the minimization of a relative entropy with respect to a Gibbs measure \(\gibbs\) associated with the duality gap energy. Using a mild extension of the Laplace method, we prove the convergence of the Gibbs measures and derive this \(Γ\)-expansion for non-degenerate ground costs, under suitable assumptions on the duality gap, the contact set and the marginals. This yields at the same time an entropic selection principle and an expansion of the EOT cost up to \(o(\varepsilon)\) for this class of ground costs, generalizing existing results.