Entropic Regularized Optimal Transport

• Entropic OT is a method that adds a relative entropy penalty to classical optimal transport, ensuring strict convexity and unique solutions.
• The formulation leverages dual Schrödinger systems and elliptic regularization, yielding smooth potential functions and robust convergence as the regularization vanishes.
• Its computational advantages enable efficient algorithms like Sinkhorn for large-scale numerical and mean-field planning problems.

Entropic regularized optimal transport (Entropic OT) incorporates an additional relative entropy penalty into the classical Kantorovich formulation, yielding a strictly convex, smoothly solvable optimal transport problem with strong regularity and favorable computational properties. Entropic regularization leads to elliptic smoothing of the potential functions and the couplings, convergence guarantees as the regularization vanishes, and is a fundamental building block in large-scale numerical and statistical optimal transport, as well as mean-field games and planning problems.

1. Mathematical Formulation: Static and Dynamic Entropic OT

Entropic regularization modifies the Kantorovich OT problem between two probability densities $m_0,m_1$ on a bounded convex domain $\Omega\subset\mathbb{R}^d$ by penalizing the negative (Boltzmann–Shannon) entropy of couplings relative to $m_0\otimes m_1$: $$\inf_{\gamma\in\Pi(m_0,m_1)} \int_{\Omega\times\Omega} c(x,y)\,d\gamma(x,y) + \varepsilon\, \mathrm{Ent}(\gamma\,\|\,m_0\otimes m_1)$$ where $$\mathrm{Ent}(\gamma\,\|\,m_0\otimes m_1) = \int \rho(x,y)\log\rho(x,y) d(m_0\otimes m_1)$$ for $\rho = \frac{d\gamma}{d(m_0\otimes m_1)}$. With $c(x,y)=|x-y|^2$, the standard quadratic cost, this functional is strictly convex for $\varepsilon > 0$ and admits a unique minimizer (Porretta, 2022).

Dynamic formulations based on the Benamou–Brenier approach regularize the kinetic action by an entropy term along the path: $$\inf_{(\rho_t, v_t)} \int_0^1 \int_\Omega \frac12 |v_t(x)|^2 \rho_t(x)\,dx\,dt + \varepsilon\int_0^1 \int_\Omega \rho_t(x)\log\rho_t(x)\,dx\,dt$$ subject to the continuity equation $\partial_t \rho_t + \nabla\cdot(\rho_t v_t)=0$ and boundary conditions $\rho_0=m_0,\, \rho_1=m_1$.

2. Duality, Schrödinger Systems, and Elliptic Regularization

The entropic cost yields a Schrödinger bridge dual of the form

$$L(\phi,\psi) = \int_\Omega \phi\,dm_0 + \int_\Omega \psi\,dm_1 - \varepsilon \int_{\Omega\times\Omega} e^{\frac{\phi(x)+\psi(y)-c(x,y)}{\varepsilon}}\,dm_0(x)dm_1(y)$$

whose stationarity provides the optimal potentials. For the quadratic cost $c(x,y) = |x-y|^2/2$, differentiation yields a quasilinear elliptic PDE for the Kantorovich potential: $$-\varepsilon \Delta \phi + \frac{1}{2}|\nabla\phi|^2 = \ln m + V$$ where $V$ encodes reference measure weights. The entropic $(\varepsilon \Delta \phi)$ term provides a vanishing-viscosity elliptic regularization of the Hamilton–Jacobi equation encountered in classical OT.

The corresponding Schrödinger system for the coupling recovers the Gibbs-type structure: $$\gamma(x, y) = e^{(\phi(x) + \psi(y) - |x-y|^2)/\varepsilon}\,m_0(x)m_1(y)$$ For the dynamic (path-space) entropic OT problem, duality induces an elliptic system in space-time for the value function $u(t,x)$ and density $m(t,x)$, yielding the coupled system: $$-u_t + \frac12|\nabla u|^2 = \ln m + V,\qquad \partial_t m - \nabla\cdot(m\nabla u) = 0$$ with no-flux (Neumann) boundary conditions (Porretta, 2022).

3. Regularity and Limit $\varepsilon\to0$

Assuming smooth, positive initial and terminal densities, and uniformly convex Hamiltonian, the entropic regularization yields a unique smooth classical solution $u\in C^{2,\alpha}([0,1]\times \Omega)$ and $m>0$, with a full suite of a priori bounds:

• $L^\infty$, gradient, Hölder, and higher-order Schauder estimates
• Uniform positivity and regularity of $(m,u)$ over $[0,1]\times\Omega$

As $\varepsilon\downarrow 0$, the solutions converge (in strong Hölder spaces) to the classical optimal transport geodesic and Kantorovich potential, recovering the singular first-order PDEs of standard OT. The elliptic regularization provides smooth approximants for use in variational problems, mean-field control, and planning under congestion (Porretta, 2022).

4. Displacement Convexity and Application to Mean-Field Problems

The entropic functional allows new displacement convexity estimates in the Eulerian approach. For any convex function $U(m)$ with $P(m)=mU'(m)-U(m)\geq0$, displacement convexity implies: $$\frac{d^2}{dt^2}\int U(m(t)) \geq \int P'(m)f'(m)H_{pp}\nabla m\cdot\nabla m + \cdots \geq 0$$ In the entropic case, convexity arguments provide additional local $L^p$-bounds and control the supremum of density along the optimal flow. As $\varepsilon\to0$ these inequalities recover McCann's classical displacement convexity for OT (Porretta, 2022).

From a mean-field games/planning perspective, the entropic regularization principle originated in works by P.-L. Lions and enables smooth, strongly convex approximations to optimization problems with congestion penalization and mean-field interaction.

5. Key Algorithmic and Computational Features

The strict convexity and regularization provided by entropy enable efficient numerical solution by matrix scaling (Sinkhorn) algorithms and their modern variants. In the discrete case, the entropic OT problem yields a dual with smooth convex structure, solved efficiently with alternating row-column normalizations. Entropic regularization ensures well-conditioned iterations, uniform positivity of couplings, and facilitates convergence analysis:

• The entropy term prevents the degeneracy and support-sparsity of classic OT solutions, making all entries strictly positive.
• This ensures the applicability of high-performance first-order methods (Sinkhorn, Greenkhorn, APDAMD) and efficient Newton-type solvers exploiting the strict strong convexity/concavity of the regularized structure (Lin et al., 2019).

For small regularization, the entropic plan approximates the classical OT plan with vanishing error, while for large regularization, it interpolates toward the independent (product) coupling.

6. Summary Table: Key Properties of Entropic OT

Aspect	Property/Result	Reference
Static formulation	Adds $\varepsilon \mathrm{Ent}(\gamma\|\mu\otimes\nu)$ to cost	(Porretta, 2022)
Dual system	Schrödinger system for potentials; log-Gibbs kernel	(Porretta, 2022)
Regularity	Solutions $(u, m)$ smooth, elliptic regularity for potentials	(Porretta, 2022)
Limit $\varepsilon\to0$	Classical OT geodesic and potential recovered	(Porretta, 2022)
Displacement convexity	Derived via regularized functionals, density bounds	(Porretta, 2022)
Numerical aspects	Efficient, well-posed, uniform positivity, parallelizable	(Porretta, 2022, Lin et al., 2019)

Uniform elliptic regularization fundamentally improves theoretical and computational properties of OT, supporting further developments in analysis, numerical methods, and mean-field PDEs.

7. Further Directions and Applications

The entropic regularization of OT underpins much of the recent progress in large-scale computational optimal transport, statistical learning, and mean-field optimization. It provides a robust approximation framework for variational OT, JKO schemes for gradient flows, and dynamic mean-field game systems. The vanishing-viscosity interpretation is instrumental in regularizing degenerate PDEs in transport, enabling displacement convexity and strong a priori estimates in mean-field planning and control (Porretta, 2022).

The entropy regularized approach yields unique, smooth approximations that converge robustly to the sharp classical solution, allowing the exploitation of convex analytic, PDE-algorithmic, and computational advantages across a wide spectrum of OT-related fields.