Mirror Descent Unification of Sampling and Optimization

This direction is split into other directions: primal mirror descent. Its last version stays readable and citable here.

Version 1 of 119 papers27 conceptsfrozen 2026-10-04

In machine learning, we often treat optimization (finding the best point) and sampling (finding the distribution of points) as two distinct worlds. Optimization uses gradients to descend a landscape, while sampling uses stochastic processes to explore it.

However, a growing body of research suggests these two tasks are actually ends of the same spectrum. By viewing iterative algorithms through the lens of Mirror Descent, researchers have shown that classic sampling techniques like tempering and transport algorithms like Sinkhorn are mathematically equivalent to optimization steps in the space of probability measures.

The Mirror Descent Perspective

To understand this connection, we first need to look at Mirror Descent (MD). Unlike standard Gradient Descent, which assumes a flat Euclidean geometry, MD adapts to the "shape" of the space using a mirror map.

Recent work has revealed that many workhorse algorithms in machine learning are secretly performing Mirror Descent. Mohammad Reza Karimi and Andreas Krause demonstrated that the Sinkhorn algorithm—the standard tool for Entropic Optimal Transport—is a specific instance of Mirror Descent where the mirror map is relative entropy [Paper]. Similarly, Pierre-Cyril Aubin-Frankowski and Anna Korba showed that iterative scaling algorithms like Sinkhorn and Expectation-Maximization (EM) can be interpreted as primal mirror descent steps in the space of joint distributions [Paper].

This is more than a mathematical curiosity. By framing these algorithms as optimization, we can apply standard optimization tools—like variable step sizes, momentum, and regret bounds—to make them faster and more stable.

Tempering as Optimization

One of the most striking results in this area concerns tempering. In Bayesian inference, tempering involves starting with a simple distribution (high temperature) and gradually "cooling" it until it concentrates on the complex posterior (low temperature).

Nicolas Chopin and Anna Korba proved that this process is mathematically equivalent to Entropic Mirror Descent applied to the Kullback-Leibler (KL) divergence [Paper]. In this view, the "temperature schedule" practitioners have used for decades is actually a sequence of step sizes in an optimization algorithm.

This connection allows us to analyze sampling convergence using the same tools we use for neural network training. For instance, Allan Douglas Jepson and Amir-massoud Farahmand leveraged this to show that temperature annealing in Optimal Transport is a sequence of Mirror Descent steps, allowing for principled "warm-starting" of transport problems [Paper].

Optimization as the Zero-Temperature Limit

If sampling is just optimization on a distribution, what happens when that distribution becomes infinitely sharp?

The research shows that standard point-estimate optimization (like SGD) is simply the zero-temperature limit of sampling. Samuel Duffield and Daniel Simpson developed a unified framework called posteriors that treats gradient descent as the limit of a stochastic differential equation [Paper]. By adjusting a single temperature parameter, a user can seamlessly transition from training a single "best" model to sampling a full Bayesian posterior.

This limit isn't just conceptual; it's reflected in the convergence rates. Austin J. Stromme proved that as the temperature vanishes, the rate at which Langevin dynamics (sampling) converges to a distribution becomes exactly equal to the rate at which gradient flow (optimization) converges to a minimum [Paper]. The efficiency of drawing samples is fundamentally bounded by the same geometric properties—like the Polyak-Łojasiewicz constant—that govern the speed of finding a function's minimum.

Why This Matters: Robustness and Stability

Unifying these fields leads to practical algorithmic breakthroughs. For example, Dong-Sig Han and colleagues introduced Variational Online Mirror Descent (VOMD) to solve Schrödinger Bridge problems [Paper]. By treating the generative modeling task as an online optimization problem, they achieved significantly higher stability and accuracy in high-dimensional tasks compared to traditional solvers.

By viewing the world through the mirror, researchers are finding that the tools we developed to find the "best" answer (optimization) are the same tools we need to understand the "uncertain" answer (sampling). This unification provides a rigorous foundation for heuristics like annealing and offers a path toward more robust, scalable Bayesian learning.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_mirror_descent_unification_of_sampling_a_bbcf8117,
  title = {Mirror Descent Unification of Sampling and Optimization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mirror-descent-unification-of-sampling-and-optimization/rdr_bbcf811722f44d4dbae28fd7f9fc2cc1},
  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_mirror_descent_unification_of_sampling_a_bbcf8117_v1,
  title = {Mirror Descent Unification of Sampling and Optimization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mirror-descent-unification-of-sampling-and-optimization/rdr_bbcf811722f44d4dbae28fd7f9fc2cc1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}