Abstract:The classical (overdamped) Langevin dynamics provide a natural algorithm for sampling from its invariant measure, which uniquely minimizes an energy functional over the space of probability measures, and which concentrates around the minimizer(s) of the associated potential when the noise parameter is small. We introduce analogous diffusion dynamics that sample from an entropy-regularized optimal transport, which uniquely minimizes the same energy functional but constrained to the set $\Pi(\mu,\nu)$ of couplings of two given marginal probability measures $\mu$ and $\nu$ on $\mathbb{R}^d$, and which concentrates around the optimal transport coupling(s) for small regularization parameter. More specifically, our process satisfies two key properties: First, the law of the solution at each time stays in $\Pi(\mu,\nu)$ if it is initialized there. Second, the long-time limit is the unique solution of an entropic optimal transport problem. In addition, we show by means of a new log-Sobolev-type inequality that the convergence holds exponentially fast, for sufficiently large regularization parameter and for a class of marginals which strictly includes all strongly log-concave measures. By studying the induced Wasserstein geometry of the submanifold $\Pi(\mu,\nu)$, we argue that the SDE can be viewed as a Wasserstein gradient flow on this space of couplings, at least when $d=1$, and we identify a conjectural gradient flow for $d \ge 2$. The main technical difficulties stems from the appearance of conditional expectation terms which serve to constrain the dynamics to $\Pi(\mu,\nu)$.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to design an effective stochastic dynamics method to efficiently draw samples from probability measures that are minimized under a given energy functional and are constrained to the coupling space $\Pi(\mu, \nu)$ of two given marginal probability measures $\mu$ and $\nu$. Specifically, the authors introduce a new diffusion dynamics - Projected Langevin Dynamics, which can converge to the unique solution of the Entropic Optimal Transport Problem. The Entropic Optimal Transport Problem adds a regularization term to the optimal transport problem, making the problem more tractable, especially in high - dimensional cases.
### Main Contributions
1. **Introduction of Projected Langevin Dynamics**:
- This dynamics is described by a stochastic differential equation (SDE) in the form:
\[
dX_t = \left( \mathbb{E}[\nabla_x c(X_t, Y_t) \mid X_t] - \nabla_x c(X_t, Y_t) - \epsilon \nabla U(X_t) \right) dt + \sqrt{2\epsilon} dW_t,
\]
\[
dY_t = \left( \mathbb{E}[\nabla_y c(X_t, Y_t) \mid Y_t] - \nabla_y c(X_t, Y_t) - \epsilon \nabla V(Y_t) \right) dt + \sqrt{2\epsilon} dB_t,
\]
where $W$ and $B$ are independent $d$-dimensional Brownian motions, and $\epsilon$ is a regularization parameter.
2. **Theoretical Analysis**:
- It is proved that under appropriate conditions, the above SDE has a unique weak solution, and the distribution of its solution always remains in $\Pi(\mu, \nu)$.
- In the long - time behavior, the distribution of the solution will converge to the unique solution $\pi$ of the Entropic Optimal Transport Problem.
- Through a new log - Sobolev inequality, it is proved that under a large enough regularization parameter and a class of marginal measures that strictly include all strongly log - concave measures, the convergence rate is exponential.
3. **Geometric Perspective**:
- By studying the Wasserstein geometry of $\Pi(\mu, \nu)$, the authors point out that when $d = 1$, the SDE can be regarded as a gradient flow on the Wasserstein space. For the case of $d\geq2$, a conjectural gradient flow is proposed.
### Special Cases and Extensions
- **Quadratic Cost Function and Gaussian Marginal Measures**:
- When the cost function $c(x, y)=\frac{|x - y|^2}{2}$ and the marginal measures are Gaussian measures, the SDE has an explicit solution, and the dynamic behavior of the solution can be described by the matrix Riccati equation.
- **Non - stationary Marginal Measures**:
- The case where the initial distribution $P_0$ does not belong to $\Pi(\mu, \nu)$ is discussed, and an open problem is proposed, that is, whether $P_t$ still converges to $\pi$ in this case.
- **Multi - marginal Measure Case**:
- The results are extended to the case of multi - marginal measures, that is, the optimal transport problem of multiple marginal measures $\mu_1,\ldots,\mu_m$ and its entropic regularization version are considered.
In summary, this paper provides an effective method to solve the Entropic Optimal Transport Problem by introducing Projected Langevin Dynamics and conducts in - depth analysis of its theoretical properties. This is not only of great theoretical significance but also provides new tools for practical applications.