Information Convexity and Contraction along Diffusion Flows

30 concepts11 papers1 proposal

When a complex probability distribution is subjected to a diffusion process—such as the heat flow or the Ornstein-Uhlenbeck (OU) process—it gradually sheds its intricate features and relaxes toward a smooth, symmetric equilibrium. Intuitively, this is a process of information loss and structural erasure. Imagine a drop of ink dispersing in water: over time, the sharp boundaries blur, and the distribution of ink molecules becomes increasingly uniform.

However, recent mathematical insights reveal that this erasure is governed by deep structural regularities. Instead of decaying chaotically, information-theoretic quantities (like entropy, mutual information, and Fisher information) decrease along highly predictable, convex trajectories. Concurrently, the geometric transformations that map these smoothed distributions back to their sharper origins behave as contractions—maps that never increase the distance between points. By studying the interplay between information convexity and contractive measure transport, researchers have unlocked a powerful framework for proving fundamental functional inequalities across probability, statistical mechanics, and information theory.

The Smooth Trajectory of Information Loss

To understand why diffusion behaves predictably, we can look at how mutual information decays when a signal is corrupted by progressive Gaussian noise. In Convexity of mutual information along the heat flow, Wibisono et al. (Andre Wibisono, Varun S. Jog) demonstrate that if an initial distribution is log-concave, the mutual information between the source and its diffused state remains strictly convex over time. This occurs because the heat equation preserves the log-concavity of the distribution, ensuring that the second time-derivative of mutual information stays positive.

What if the starting distribution is highly irregular, such as a mixture of discrete point masses or narrow Gaussians? In the short term, the mutual information can exhibit non-convex fluctuations as the noise first begins to blur the distinct peaks. However, the authors show that both the heat flow and the OU process (Convexity of mutual information along the Ornstein-Uhlenbeck flow) act as powerful geometric regularizers. Once the elapsed time crosses a specific threshold determined by the initial distribution's geometry (such as its diameter or Fisher information), the noise completely dominates the initial structure. From that point onward, the system enters an eventual convexity regime, where the rate of information loss stabilizes into a predictable geometric decay.

This smoothing effect extends to other fundamental quantities. For instance, Junliang Liu establishes in Square Root Convexity of Fisher Information along Heat Flow in Dimension Two that the square root of Fisher information evolves convexly under two-dimensional heat flow, providing a verifiable structural property for multi-dimensional diffusion. Going even further, Guillaume Wang tackles the long-standing Gaussian Completely Monotone conjecture in A higher-order Otto calculus approach to the Gaussian completely monotone conjecture. By developing a "convective calculus" framework that treats the curved space of probability measures as flat, Wang proves that the first five time-derivatives of entropy along a heat flow alternate in sign for log-concave measures, supporting the hypothesis that entropy decay is completely monotone.

Reversing Diffusion via Contractive Maps

If diffusion smooths out a distribution, can we construct a mathematical map that reverses this process, pushing the smoothed, noisy measure back onto its original, concentrated form? More importantly, what geometric properties would such a map possess?

A foundational answer comes from Caffarelli's contraction theorem, which states that the optimal transport map from a standard Gaussian to a log-concave measure is a contraction. In A generalization of Caffarelli’s contraction theorem via (reverse) heat flow, Milman et al. (Emanuel Milman, Young-Heon Kim) generalize this result by using a heat flow process to construct the transport map. By analyzing parabolic partial differential equations, the authors prove that because diffusion preserves log-concavity, the resulting transport map inherits strict contractive properties.

This contractive property is incredibly useful because contractions do not increase distances between points in a distribution. Consequently, when you push a smoothed measure back to its original state via a contraction, geometric and functional properties are preserved or tightened.

In Spectral monotonicity under Gaussian convolution, Klartag et al. (Bo’az Klartag, Eli Putterman) exploit this exact mechanism. They construct a contractive transport map to prove that the Poincaré constant (which measures the "spectral gap," or how fast a system returns to equilibrium) increases monotonically as a log-concave measure is smoothed by Gaussian noise. The existence of this contraction ensures that functional inequalities follow a monotonic path during the smoothing process.

Transferring Functional Inequalities

The true utility of contractive measure transport lies in its ability to transfer functional inequalities—such as Poincaré, log-Sobolev, or isoperimetric inequalities—from simple, well-understood reference distributions (like Gaussians) to highly complex target distributions.

Consider the challenge of analyzing interacting quantum or statistical field theories. These theories are notoriously difficult because their probability measures involve complex, non-local interactions. In Exact Renormalization Groups and Transportation of Measures, Yair Shenfeld bridges this gap by linking the exact renormalization group to transport theory.

[Simple Reference Measure (Gaussian)] 
               │
               ▼  (Diffusion Flow / Semigroup)
[Construct Contractive / Lipschitz Map]
               │
               ▼  (Pushes back / transfers properties)
[Complex Target Measure (Interacting Field)]

Shenfeld constructs a Langevin transport map derived from the Ornstein-Uhlenbeck semigroup to connect a simple Gaussian free field to an interacting target theory. By utilizing a technical condition known as the multiscale Bakry-Émery criterion, Shenfeld controls the derivatives of the potential, proving that the transport map is Lipschitz continuous. This allows known functional inequalities to be systematically transferred from the simple Gaussian free field directly to the complex interacting theory.

When traditional optimal transport maps fail to provide the necessary contraction properties, stochastic processes can step in. In The Brownian transport map, Dan Mikulincer introduces the "Brownian transport map" built via the Föllmer process. By leveraging the infinite-dimensional nature of the Wiener measure and applying Malliavin calculus, Mikulincer constructs a map that exhibits contraction properties even in settings where classical optimal transport breaks down. This contractive map immediately yields a suite of new functional inequalities, including $\Psi$-Sobolev and $q$-Poincaré inequalities.

Sharp Bounds and Control

Understanding these contractive and convex behaviors allows researchers to establish fundamental limits in information theory and control. For instance, in The strong data processing inequality under the heat flow, Klartag et al. (Bo'az Klartag, Or Ordentlich) show that the information divergence between two distributions evolving under heat flow is convex relative to the noise parameter. By linking this structural property to Poincaré and log-Sobolev constants, they derive tight upper and lower bounds on divergence decay, establishing sharper limits for estimation errors (MMSE) in Gaussian channels than classical bounds could provide.

Meanwhile, the interplay between diffusion and information metrics finds practical computational use in mean field control. In Generalized optimal transport and mean field control problems for reaction-diffusion systems with high-order finite element computation, Fu et al. (Guosheng Fu, Stanley Osher, Stanley J. Osher, Will Pazner, Wuchen Li) utilize Fisher information as a tunable regularization parameter. In their framework, the Fisher information directly dictates the smoothness of density transitions, demonstrating how the theoretical guarantees of information-theoretic flows can be engineered to compute optimal transitions between multiple interacting densities.

Finally, Backhoff‐Veraguas et al. (Julio Backhoff‐Veraguas, Grégoire Loeper, Jan Obloj) extend these geometric transport principles to the stochastic setting in Geometric martingale Benamou–Brenier transport and geometric Bass martingales, establishing dualities between different classes of martingales using optimal transport and change of measure techniques.

Ultimately, whether viewed through the lens of information decay or contractive transport maps, diffusion processes impose a strict geometric order on probability spaces. By treating smoothing not just as a source of noise, but as a tool to construct contractions, this paradigm transforms hard problems in high-dimensional analysis into manageable geometric exercises.

Key Foundational Work

Convexity of mutual information along the heat flow (2018) — contributed 6 concept elements to this direction.
Convexity of mutual information along the Ornstein-Uhlenbeck flow (2018) — contributed 5 concept elements to this direction.
A generalization of Caffarelli’s contraction theorem via (reverse) heat flow (2011) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_contractive_measure_transport_abe51a3e,
  title = {contractive measure transport},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/contractive-measure-transport/rdr_abe51a3e3cba4eec8a25b0210f20be9a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_contractive_measure_transport_abe51a3e_v1,
  title = {contractive measure transport},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/contractive-measure-transport/rdr_abe51a3e3cba4eec8a25b0210f20be9a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake