Abstract:We introduce and study the class of linear transfers between probability distributions and the dual class of Kantorovich operators between function spaces. Linear transfers can be seen as an extension of convex lower semi-continuous energies on Wasserstein space, of cost minimizing mass transports, as well as many other couplings between probability measures to which Monge-Kantorovich theory does not readily apply. Basic examples include balayage of measures, martingale transports, optimal Skorokhod embeddings, and the weak mass transports of Talagrand, Marton, Gozlan and others. The class also includes various stochastic mass transports such as the Schrödinger bridge associated to a reversible Markov process, and the Arnold-Brenier variational principle for the incompressible Euler equations. We associate to most linear transfers, a critical constant, a corresponding effective linear transfer and additive eigenfunctions to their dual Kantorovich operators, that extend Mané's critical value, Aubry-Mather invariant tori, and Fathi's weak KAM solutions for Hamiltonian systems. This amounts to studying the asymptotic properties of the nonlinear Kantorovich operators as opposed to classical ergodic theory, which deals with linear Markov operators. This allows for the extension of Mather theory to other settings such as its stochastic counterpart. We also introduce the class of convex transfers, which includes $p$-powers ($ p \geq 1$) of linear transfers, the logarithmic entropy, the Donsker-Varadhan information, optimal mean field plans, and certain free energies as functions of two probability measures, i.e., where the reference measure is also a variable. Duality formulae for general transfer inequalities follow in a very natural way. This paper is an expanded version of a previously posted but not published work by the authors.

A Cantor-Kantorovich Metric Between Markov Decision Processes with Application to Transfer Learning

Measuring the Distance Between Finite Markov Decision Processes

Data-driven memory-dependent abstractions of dynamical systems via a Cantor-Kantorovich metric

Measurized Markov Decision Processes Part I: The Discounted Infinite Horizon Criterion

Data-driven abstractions via adaptive refinements and a Kantorovich metric [extended version]

A Kernel Perspective on Behavioural Metrics for Markov Decision Processes

Sufficient Markov Decision Processes with Alternating Deep Neural Networks

Off-Policy Evaluation in Markov Decision Processes under Weak Distributional Overlap

Mather Measures and Ergodic Properties of Kantorovich Operators associated to General Mass Transfers

Bayesian Learning of Optimal Policies in Markov Decision Processes with Countably Infinite State-Space

Learning Temporal Distances: Contrastive Successor Features Can Provide a Metric Structure for Decision-Making

A metric on directed graphs and Markov chains based on hitting probabilities

Learning Constrained Markov Decision Processes With Non-stationary Rewards and Constraints

Variational Approach for Learning Markov Processes from Time Series Data

Robust Anytime Learning of Markov Decision Processes

An invitation to adaptive Markov chain Monte Carlo convergence theory

A Fenchel-Moreau-Rockafellar type theorem on the Kantorovich-Wasserstein space with Applications in Partially Observable Markov Decision Processes

Robust and Efficient Transfer Learning with Hidden-Parameter Markov Decision Processes

Markov Decision Processes with Time-Varying Geometric Discounting

Quantitative contraction rates for Markov chains on general state spaces

A Riemannian Approach to Ground Metric Learning for Optimal Transport