Convex duality in nonlinear optimal transport (Q2421526): Difference between revisions

Let \(S_t\), \(t=0,\dots,T,\) be Polish spaces and \(S\) their topological product with the projections \(\pi_t:S\to S_t.\) For a continuous function \(\psi_t:S_t\to[1,\infty)\), denote by \(C_t\) the space of all continuous functions \(x_t:S_t\to \mathbb{R}^d\) such that \(x_t/\psi_t\) is bounded. Then \(C_t\) is a Banach space with respect to the norm \(\|x_t\|_{C_t}=\|x_t/\psi_t\|_\infty.\) The Banach space \(C\) is defined similarly by replacing \(\psi_t\) with \(\psi=\sum_{t=0}^T\psi_t\) and \(S_t\) with \(S\). Denote by \(M_t \; (M)\) the space of all finite \(\mathbb{R}^d\)-valued Borel measures under which \(\psi_t\) (\(\psi)\) is integrable. Then the dual spaces \(C_t^*\) \((C^*)\) satisfy \(C_t^*\subseteq M_t\) \((C^*\subseteq M)\) with equality if \(S_t\) \((S)\) is compact. The author considers the problem \[ \text{minimize } \sum_{t=0}^T G^*_t(\lambda_t)+H^*(\lambda) \text{ over }\lambda\in M, \tag{D} \] where \(G^*_t\) and \(H^*\) are convex functions on \(M_t \) and \(M\), respectively, and \(\lambda_t\) is the marginal of \(\lambda\) on \(S_t\). The problem (D) covers several optimization problems as particular cases: the Kantorovich-Monge optimal transport problem (when \(T=d=1\), \(G^*_t=\delta_{\{\mu_t\}}\), \(\mu_t\in M\), \(H^*_\lambda=\int_Sc\,d\lambda+\delta_{M_+}(\lambda)\) for a given lsc function \(c\ge 0\)), the Strassen problem of finding a probability measure with given marginals (for \(H^*=\delta_\Lambda\) where \(\Lambda\subset M\) is a set of probability measures), the classical problem of Schrödinger when \(H^*(\lambda) \) is the entropy with respect to a given reference measure \(\lambda\), as well as some problems in mathematical finance. Here, \( \delta_M\) denotes the indicator function of a set \(M\). The main idea of the paper is to interpret (D) (whith \(M\) replaced by \(C^*\)) as the conjugate dual problem (in the sense of Rockafellar) to the problem \[ \text{minimize }\sum_{t=0}^TG_t(x_t)+H\big(-\sum_{t=0}^Tx_t\circ\pi_t)\lambda\big)\text{ over } x=(x_t)_{t=0}^T\in \prod_{t=0}^TC_t,\tag{P} \] where \(G_t\) is a proper convex function on \(C_t\), \(t=0,\dots,T,\) and \(H\) is a proper convex function on \(C\). As the authors mention in the abstract: ``This provides a unified treatment of a large class of related problems in probability theory and allows for generalizations of the classical problem formulations.''