Inference via low-dimensional couplings

Alessio Spantini, Daniele Bigoni, Youssef Marzouk

Abstract

We investigate the low-dimensional structure of deterministic transformations between random variables, specifically transport maps between probability measures. In the context of statistics and machine learning, these transformations couple a tractable "reference" measure (e.g., a standard Gaussian) with a target measure of interest. While direct simulation from a target measure can be achieved by pushing forward reference samples through such a map, representing and evaluating these maps becomes increasingly difficult in high dimensions. This paper establishes a formal link between the Markov properties of a target measure and the existence of low-dimensional couplings induced by sparse and/or decomposable transport maps. This analysis facilitates the construction of transformations in high-dimensional settings and suggests new inference methodologies for continuous non-Gaussian graphical models, including variational algorithms for filtering and smoothing that generalize the classic square-root Rauch-Tung-Striebel Gaussian smoother.

Introduction

In modern statistics and machine learning, characterizing complex, high-dimensional "target" distributions—such as the posterior distribution of hidden states in a geophysical model—is a fundamental challenge. One elegant approach involves transport maps: deterministic functions that transform a simple, known "reference" distribution into the target distribution. Once a map is defined, sampling becomes trivial: one simply draws from the reference and pushes the samples through the map.

However, the "curse of dimensionality" often renders this approach impractical. Without structural constraints, the number of parameters required to represent a high-dimensional map grows exponentially. This paper addresses this bottleneck by demonstrating that the graphical model structure of a distribution (its conditional independence properties) can be directly translated into structural simplifications of the transport map.

The Knothe-Rosenblatt Rearrangement and Sparsity

The authors focus on a specific class of transformations known as the Knothe-Rosenblatt (KR) rearrangement. This is a triangular transport map where each component depends only on a subset of the preceding variables. The central theoretical contribution of the paper is proving that the sparsity of the target distribution's graph is inherited by the transport map.

The "Inverse" Advantage

A critical insight of the research is the distinction between direct and inverse transport maps. The authors prove (Theorem 5.1) that the components of the inverse transport map inherit the sparsity of the target distribution's Markov graph. This is a significant advantage because sparsity in the inverse map is tied to conditional independence (the edges in a Markov network), which is common in physical and biological systems. In contrast, sparsity in the direct map requires marginal independence, a much rarer property in practice.

Furthermore, the authors show that complex, high-dimensional maps can be broken down, or decomposed, into a sequence of simpler, low-dimensional maps (Theorem 6.1). This allows researchers to solve many small, local optimization problems rather than one intractable global problem.

Generalizing the Rauch–Tung–Striebel Smoother

By applying these theories to state-space models, the authors derive new variational algorithms for filtering, smoothing, and sequential parameter inference. These algorithms represent a "non-Gaussian generalization" of the classic square-root Rauch–Tung–Striebel (RTS) Gaussian smoother. Unlike traditional Particle Filters, which represent distributions as a discrete cloud of points, this method provides a functional representation. This functional approach allows for easier density evaluation, gradient computation, and the ability to reuse the map for different tasks without re-sampling.

Empirical Validation

The framework was validated using several benchmarks, ranging from synthetic tests to real-world financial data:

  • Stochastic Volatility Modeling: In tests using daily Pound/Dollar exchange rate data (N=945 days), the transport map approach produced results nearly indistinguishable from gold-standard Markov Chain Monte Carlo (MCMC) methods.
  • Long-Horizon Stability: To test the robustness of the recursive algorithms, the authors ran synthetic benchmarks with up to 9,009 time steps. The variational algorithms remained stable, suggesting that approximation errors in the maps do not necessarily accumulate or explode over long time horizons.
  • Computational Trade-offs: The research highlighted a clear hierarchy in map complexity. While high-order polynomial maps (e.g., degree 7) provided extreme accuracy at the cost of significant computation time, even linear maps provided remarkably accurate marginals for long-time smoothing in a fraction of the time.

Conclusion and Future Directions

This work provides a rigorous bridge between the world of graphical models and the world of measure transport. By showing that we can determine the sparsity pattern of a map a priori—simply by looking at the graph of the target distribution—the authors enable more efficient high-dimensional inference.

Future work may extend these results to "approximate" Markov properties, where dependencies are weak but non-zero, or explore low-rank structures where connections exist only along specific linear combinations of variables. While the current optimization (minimizing KL divergence) can be challenging for multi-modal distributions, the structural foundations laid here provide a clear path forward for non-Gaussian inference in complex systems.

Summary updated

Abstract

We investigate the low-dimensional structure of deterministic transformations between random variables, i.e., transport maps between probability measures. In the context of statistics and machine learning, these transformations can be used to couple a tractable "reference" measure (e.g., a standard Gaussian) with a target measure of interest. Direct simulation from the desired measure can then be achieved by pushing forward reference samples through the map. Yet characterizing such a map---e.g., representing and evaluating it---grows challenging in high dimensions. The central contribution of this paper is to establish a link between the Markov properties of the target measure and the existence of low-dimensional couplings, induced by transport maps that are sparse and/or decomposable. Our analysis not only facilitates the construction of transformations in high-dimensional settings, but also suggests new inference methodologies for continuous non-Gaussian graphical models. For instance, in the context of nonlinear state-space models, we describe new variational algorithms for filtering, smoothing, and sequential parameter inference. These algorithms can be understood as the natural generalization---to the non-Gaussian case---of the square-root Rauch-Tung-Striebel Gaussian smoother.
BibTeX
@article{spantini2018inference,
  author = {Alessio Spantini and Daniele Bigoni and Youssef Marzouk},
  title = {{Inference via low-dimensional couplings}},
  journal = {J. Mach. Learn. Res.},
  year = {2018},
  url = {https://openreview.net/forum?id=cawP5oKLZs},
}

Loading related content...

Figure