A geometric perspective on regularized optimal transport (Q2333283)

The variational problem to minimize \(\int_0^1\int\frac12|b_t(x)|^2\rho_t(x)dx dt\) over \(\rho\) and \(b\) subject to \(\partial_t\rho_t+\operatorname{div}(\rho_tb_t)=\gamma\Delta\rho_t\) for \(0 < t<1\) with \(\rho_0=\mu\) and \(\rho_1=\nu\), where \(\rho\) is a time-dependent density \(\rho_t(x)=\rho(t,x)\) in the space of probability measures \(\mathcal{P}(\mathbb{R}^n)\), \(b=b(t)(x)\) is the vector field viewed as a control used to drive \(\rho_t\) from an initial density \(\mu\) to a final density \(\nu\), and the parameter \(\gamma>0\) is a fixed diffusion coefficient, is called the Schrödinger bridge problem. The Yasue problem is to minimize \(\int_0^1\int\frac12|v_t(x)|^2\rho_t(x)+\frac{\gamma^2}2\frac{|\nabla\rho_t(x)|^2}{\rho_t(x)}dx dt\) over \(\rho\) and \(v\). The Yasue problem is not an optimal control problem, comparing to the Schrödinger bridge problem, but rather a Lagrangian mechanics problem, however both problems can be seen as fluid-based formulations of regularized optimal transport. In this paper, the author presents new geometric intuition on dynamical versions of regularized optimal transport. First, the author shows that the Schrödinger and Yasue problems are equivalent and, in terms of minimizers, the densities \(\rho\) are the same. Next, the author introduces two families of variational problems on manifolds which contain a direct analogue of the Schrödinger bridge and the Yasue problems. The first family consists of optimal control problems, while the second one comprises Lagrangian mechanics problems. The geometric problems are defined on any Riemannian manifold and involve a potential function. By considering the pseudo-Riemannian structure of the Wasserstein space and taking the potential to be the entropy, then the geometric problems formally become the optimal transport problems. Finally, the author shows that a striking feature of the two optimal transport variants is that their Euler-Lagrange equations can be reduced to a linear PDE via a Hopf-Cole transformation.