TL;DR
This paper introduces "backward Monge" martingale transports, which are surprising constructions that allow for deterministic maps in the context of martingale couplings. This breakthrough demonstrates that these backward Monge transports are dense in the space of all martingale transports, paralleling a known result in classical optimal transport theory.
What Problem Does This Paper Address?
In optimal transport theory, a "Monge map" provides a deterministic way to move mass from one distribution to another. However, in the related field of "martingale transport," where the transport must also satisfy a martingale property (meaning the expected future value equals the current value), it's generally understood that Monge maps are not possible for non-trivial cases. This is because a deterministic martingale must be constant over time, implying the transport map would have to be trivial (identity) if it were a forward-looking Monge map. This paper challenges this understanding by exploring a "backward" perspective. Instead of mapping from the first marginal (forward in time), they consider maps from the second marginal (backward in time). This shift in perspective opens up the possibility of non-trivial Monge maps in the martingale setting, a concept previously thought to be uninteresting or impossible.
Methods
The paper introduces and utilizes several key techniques:
- Barcode Transport: A novel construction for a specific type of martingale transport that is "Monge outside the atoms" of the second marginal. This construction is built by decomposing the marginals into "bars" and using a "left-curtain transport" for each part.
- Left-Curtain Transport: A specific type of martingale transport, analyzed for its structural properties, particularly its support.
- Shadow Measures: A concept from optimal transport theory used to characterize the relationship between marginals and the properties of transports.
- Generalized Optimal Transport: The paper characterizes Monge martingale transports as minimizers of a specific type of generalized optimal transport problem, which involves costs dependent on conditional distributions.
- Backward Deterministic Martingales: A new class of stochastic processes is introduced, where the entire history up to time n is encoded in the state at time n. This class is shown to be rich enough to mimic any given martingale with continuous marginals.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Reversing the Monge Map Perspective: The core insight is that by considering Monge maps from the second marginal (backward in time), non-trivial deterministic maps become possible within the martingale transport framework. This fundamentally changes the understanding of Monge maps in this context.
- Density of Backward Monge Transports: A major result is that these "backward Monge" martingale transports are not just isolated examples but are dense in the space of all martingale transports. This is a significant parallel to a known result in classical optimal transport and implies that optimal transport problems can be solved using these Monge maps.
- Existence of Monge Martingale Transports: The paper proves the existence of Monge martingale transports when the second marginal is atomless, constructing a specific example called the "barcode transport."
- Backward Deterministic Martingales as Mimics: The introduction of backward deterministic martingales provides a powerful tool for mimicking any given martingale, offering a new perspective on process construction.
- Uniqueness Characterization: The paper provides a precise characterization of when martingale transports (and Monge martingale transports) are unique, linking it to the properties of the marginals and their shadows.
Key Results
- Existence Theorem (Theorem 2.1): For any two probability measures $\mu, \nu$ in convex order ($\mu \leq_{cx} \nu$) where $\nu$ is atomless, there exists a Monge martingale transport from $\mu$ to $\nu$.
- Denseness Theorem (Theorem 2.3): The set of Monge martingale transports is weakly dense in the set of all martingale transports when $\nu$ is atomless.
- Equivalence of Formulations (Corollary 2.4): The infimum of a cost function over all martingale transports is the same as the infimum over Monge martingale transports, provided $\nu$ is atomless.
- Refinement of Strassen's Theorem (Theorem 3.1): A new version of Strassen's theorem is presented, showing that if $X \leq_{cx} Y$, then there exists a random variable $X'$ on the same probability space such that $X'$ has the same law as $X$ and $X' = E[Y|X']$, which is a martingale.
- Mimicking Theorem (Corollary 3.6): Any martingale with atomless marginals can be mimicked by a backward deterministic martingale.
- Uniqueness Characterization (Theorem 2.5): The uniqueness of martingale transports and Monge martingale transports is characterized in terms of the marginals and their shadows.
Limitations and Future Work
- Atomless Assumption: Many of the key results (existence, denseness, equivalence of formulations) rely on the assumption that the second marginal $\nu$ is atomless. The paper explores cases where $\nu$ has atoms, but the general denseness result does not hold.
- Higher Dimensions: The paper focuses on the one-dimensional case ($\mathbb{R}$). Extending the results to higher dimensions ($\mathbb{R}^d$) is identified as a significant open problem, with Conjecture 5.1 proposing a potential analogue of the denseness result.
- Other Constraints: The paper speculates on analogues of its results for other transport constraints, such as supermartingale couplings and directional constraints ($X \leq Y$), suggesting these as avenues for future research (Conjectures 5.3 and 5.4).
- Backward Deterministic Martingales: The paper notes that the class of backward deterministic martingales has not been extensively studied and suggests a deeper investigation for future work.
Summary updated