Bounding adapted Wasserstein metrics

José Blanchet, Martin Larsson, Jonghwa Park, Johannes WieselarXiv 2024 · DOI · arXiv

TL;DR

This paper provides the first explicit upper bounds for the adapted Wasserstein distance ($\mathcal{AW}_p$) in terms of the standard Wasserstein distance ($\mathcal{W}_p$). By introducing a "smoothed" version of the adapted metric, the authors demonstrate that these two distances are nearly equivalent on sets of stochastic processes with sufficiently regular (smooth) kernels, bridging a significant gap in optimal transport theory for time-dependent problems.

What Problem Does This Paper Address?

In fields like mathematical finance and multi-stage optimization, we often compare the "distance" between two stochastic processes. The standard Wasserstein distance ($\mathcal{W}_p$) is the go-to tool for this, but it has a major flaw: it ignores the flow of information (the filtration). It allows "couplings" that can look into the future, which is physically impossible in real-time decision-making.

The adapted Wasserstein distance ($\mathcal{AW}_p$) fixes this by requiring couplings to be "bicausal," meaning they respect the temporal structure of the processes. While it is well-known that $\mathcal{W}_p \le \mathcal{AW}_p$, the reverse—bounding the adapted distance by the standard one—has remained an open challenge. Without such bounds, researchers cannot easily translate the vast library of results from standard optimal transport into the "adapted" world needed for optimal stopping or risk management.

Methods

The authors bridge this gap using a smoothing and decomposition strategy:

  1. Smooth Adapted Wasserstein Distance: They define a new metric, $\mathcal{AW}_p^{(\sigma)}$, by convolving the measures with independent noise (bandwidth $\sigma$).
  2. The Triangle Inequality: They decompose the target distance into two parts:
  3. The distance between the smoothed versions of the measures.
  4. The "smoothing error" (how much information is lost by adding noise).
  5. Explicit Bounding:
    • They prove that the smoothed adapted distance $\mathcal{AW}_1^{(\sigma)}$ can be bounded by the standard $\mathcal{W}_1$ distance, scaled by $1/\sigma$, plus a term for the "tails" of the distribution.
  6. They use Eder’s modulus of continuity to quantify the smoothing error. This measures the regularity of the "kernels" (the conditional distributions of the process over time).
  7. Topological Analysis: They investigate the topology induced by this smoothed metric, discovering it acts as an "interpolator" between the standard weak topology and the adapted weak topology.

Datasets

This is a theoretical probability paper and does not use empirical datasets. Instead, it utilizes a "Standard Example" (Example 3) involving simple discrete measures to illustrate why $\mathcal{W}_p$ and $\mathcal{AW}_p$ can diverge and to test the tightness of their new bounds.

Key Insights and Ideas

  • Regularity is the Bridge: The "gap" between standard and adapted transport is essentially a lack of regularity in how processes evolve. If a process has "Lipschitz kernels" (meaning its future distribution changes smoothly as its past changes), the adapted distance behaves much more like the standard distance.
  • The $\sqrt{\mathcal{W}_1}$ Bound: For measures with Lipschitz kernels, the authors derive a strikingly simple inequality: $\mathcal{AW}_1 \le C\sqrt{\mathcal{W}_1}$. This provides a concrete way to guarantee that if two processes are close in the standard sense, they are also close in the adapted sense.
  • Smoothing as a Multi-tool: Smoothing doesn't just make the math easier; it reveals a hierarchy of topologies. The authors show that as the noise level $\sigma$ changes, the resulting metric shifts its focus from the global distribution (standard OT) to the temporal structure (adapted OT).
  • Overcoming the Curse of Dimensionality: The paper suggests that smoothing allows for better empirical estimation of adapted distances, which usually suffer from severe dimensionality issues.

Key Results

  • Theorem 22 & Corollary 30: These provide the core explicit bounds. For processes on a bounded domain, the adapted distance is bounded by a sum of the standard distance (weighted by the noise level) and the modulus of continuity of the kernels.
  • Finite-Sample Guarantees: The authors provide convergence rates for empirical measures. They show that the smoothed adapted distance between a true distribution and its empirical sample converges at a rate of $n^{-1/2}$ (ignoring some log factors), which is much faster than the rates typically seen in high-dimensional adapted transport.
  • Topological Equivalence: Theorem 35 proves that for any fixed noise level $\sigma > 0$, the smoothed adapted Wasserstein distance induces the same topology as the standard Wasserstein distance.

Limitations and Future Work

  • Dependence on $T$: The constants in the bounds currently grow with the number of time steps $T$. While the authors cite recent work that improves this from exponential to linear growth, it remains a factor in very long-horizon problems.
  • Requirement for Regularity: The bounds are most useful when the measures have some degree of smoothness (equicontinuous kernels). For highly irregular or "jumpy" processes, the bounds may become uninformative.
  • Future Directions: The authors suggest that these bounds could be used to simplify the analysis of complex time-dependent optimization problems by allowing researchers to use standard OT estimates as a proxy for adapted ones.

Summary updated

Abstract

The Wasserstein distance $\mathcal{W}_p$ is an important instance of an optimal transport cost. Its numerous mathematical properties as well as applications to various fields such as mathematical finance and statistics have been well studied in recent years. The adapted Wasserstein distance $\mathcal{A}\mathcal{W}_p$ extends this theory to laws of discrete time stochastic processes in their natural filtrations, making it particularly well suited for analyzing time-dependent stochastic optimization problems. While the topological differences between $\mathcal{A}\mathcal{W}_p$ and $\mathcal{W}_p$ are well understood, their differences as metrics remain largely unexplored beyond the trivial bound $\mathcal{W}_p\lesssim \mathcal{A}\mathcal{W}_p$. This paper closes this gap by providing upper bounds of $\mathcal{A}\mathcal{W}_p$ in terms of $\mathcal{W}_p$ through investigation of the smooth adapted Wasserstein distance. Our upper bounds are explicit and are given by a sum of $\mathcal{W}_p$, Eder's modulus of continuity and a term characterizing the tail behavior of measures. As a consequence, upper bounds on $\mathcal{W}_p$ automatically hold for $\mathcal{AW}_p$ under mild regularity assumptions on the measures considered. A particular instance of our findings is the inequality $\mathcal{A}\mathcal{W}_1\le C\sqrt{\mathcal{W}_1}$ on the set of measures that have Lipschitz kernels. Our work also reveals how smoothing of measures affects the adapted weak topology. In fact, we find that the topology induced by the smooth adapted Wasserstein distance exhibits a non-trivial interpolation property, which we characterize explicitly: it lies in between the adapted weak topology and the weak topology, and the inclusion is governed by the decay of the smoothing parameter.
BibTeX
@misc{blanchet2024bounding,
  author = {José Blanchet and Martin Larsson and Jonghwa Park and Johannes Wiesel},
  title = {{Bounding adapted Wasserstein metrics}},
  year = {2024},
  eprint = {2407.21492},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2407.21492},
}
Figure