Abstract
Causal transport plans represent a relaxation of adapted processes, extending the classical Kantorovich transport framework to account for temporal constraints. In a discrete-time setting, this research establishes a Dynamic Programming Principle (DPP) that links causal transport to general transport problems. This recursive principle allows for the identification of the Knothe-Rosenblatt rearrangement as a causal analogue to the Brenier map. Furthermore, the framework provides transport-information inequalities for the nested distance between stochastic processes, offering a robust method to gauge discrepancies between stochastic programs driven by different noise distributions.
Introduction
Classical optimal transport, rooted in the Monge-Kantorovich problem, identifies the most efficient way to transform one probability distribution into another. While powerful, this framework is often inadequate for stochastic processes because it ignores the temporal flow of information. Standard transport plans can effectively "look into the future," allowing mass to be moved at time $t$ based on information that only becomes available at time $t+1$.
In fields such as finance, engineering, and multistage decision-making, such non-causal behavior is unrealistic. This paper addresses the causal optimal transport problem, where transport plans are constrained to be non-anticipative. This ensures that the movement of mass at any given time depends only on the history of the processes up to that point, preserving the filtration of the underlying probability spaces.
The Causal Framework and Dynamic Programming
The authors work within a discrete-time setup, defining causal and bicausal plans as relaxations of adapted processes. A plan is causal if the conditional distribution of the target given the source at time $t$ depends only on the source's history. Bicausal plans satisfy this constraint in both directions.
A central contribution of this work is the Dynamic Programming Principle (DPP) (Theorem 2.7). The authors prove that the complex, multi-dimensional causal transport problem can be decomposed into a sequence of one-dimensional transport problems. This recursive structure reduces the dimensionality of the problem, although it introduces non-linearities in the recursion. This decomposition is particularly effective for "causal quasi-Markov" transport plans, where the state space remains manageable.
The Causal Brenier Map
In classical transport theory, the Brenier map is the unique optimal gradient of a convex function that solves the Monge problem. This paper identifies the Knothe-Rosenblatt rearrangement (also known as the multivariate quantile transform) as the causal counterpart to the Brenier map.
Theorem 2.9 demonstrates that for separable convex costs (such as squared Euclidean distance) and source measures with independent marginals, the Knothe-Rosenblatt rearrangement is the optimal causal plan. This result provides a concrete, computable optimizer for a broad class of temporal transport problems, bridging the gap between abstract coupling theory and practical application.
Transport-Information Inequalities
The framework also extends to the measurement of distance between stochastic processes. The authors utilize the nested distance (or bicausal distance), which measures the discrepancy between processes while respecting their information structures.
By establishing a Bicausal $T_1$ Inequality (Theorem 2.11), the authors bound the nested distance using relative entropy: $$W_{1,bc}(\mu, \nu) \leq K \sqrt{Ent(\nu|\mu)}$$ This inequality is highly significant for stochastic programming. It allows researchers to use relative entropy—a measure often easier to calculate than transport distances—to bound the potential error introduced when a complex "true" noise distribution is replaced by a simplified approximation in a model.
Implications for Stochastic Programming
The paper introduces "lexicographical displacement interpolation," showing that many value functions in stochastic programming exhibit concavity along these specific paths. This provides a geometric bridge between optimal transport and multistage decision-making, suggesting that the causal transport framework is the natural geometry for analyzing the stability of stochastic programs.
Conclusion and Future Directions
While this work provides a rigorous foundation for causal transport in discrete time, several avenues for future research remain. The current results rely heavily on discrete-time machinery; extending these proofs to continuous-time processes requires more complex tools from stochastic calculus and filtered probability spaces. Additionally, the optimality of the Knothe-Rosenblatt rearrangement currently requires specific independence assumptions on the marginals. Future work may explore how to relax these conditions while maintaining the recursive tractability provided by the Dynamic Programming Principle.
Summary updated