The back-and-forth method for Wasserstein gradient flows

This page illustrates the method proposed in (Jacobs et al., 2021). The back-and-forth method was initially introduced to compute optimal transport maps (Jacobs & Léger, 2020), and it can be generalized to compute Wasserstein gradient flows, via the JKO scheme (Jordan et al., 1998). This allows to efficiently solve PDEs of the type

\[\begin{align*} \partial_t\rho - \mathrm{div}(\rho\nabla\phi) = 0, \\ \phi = \delta U(\rho), \end{align*}\]

for a large class of interesting energies $U(\rho)$. We consider the following energies:

  • (The porous medium equation) Slow diffusion energies $$ U(\rho) = \int_{\mathbb{R}^d} u_m(\rho(x)) + V(x)\,dx, $$ where $u_m(\rho) = \gamma \rho(x)^m$ for $m>1$ and $\gamma>0$. Here $V\colon\mathbb{R}^d\to[0,\infty]$ is a potential function.
  • (The incompressible limit) Congestion energies $$ U(\rho) = \int_{\mathbb{R}^d} u_{\infty}(\rho(x)) + V(x)\,dx, $$ where $u_\infty(\rho)=0$ if $0\le \rho \le 1$ and $\infty$ otherwise.
  • (An aggregation-diffusion equation) Convex-concave energies $$ U(\rho) = \int_{\mathbb{R}^d} u_m(\rho(x))\,dx + \iint_{\mathbb{R}^d\times \mathbb{R}^d} \lvert x-y\rvert^2\rho(x)\rho(y)\,dxdy. $$ Note that the first term in convex in $\rho$ while the second term is concave in $\rho$.

Code

The code used in the paper is available on GitHub. The documentation is available here.

Porous medium equation

The Wasserstein gradient flow of $U(\rho)=\int_{\mathbb{R}^d} u_m(\rho(x)) + V(x)\,dx$ is the advection slow diffusion equation \(\partial_t\rho +\mathrm{div}(\rho (-\nabla V))= \gamma\Delta \rho^m,\) together with Neumann boundary conditions and an initial condition $\rho(0,\cdot) = \rho_0$.

A star-shaped obstacle (in white). Here $m=4$ and $V(x)=\lVert x-a\rVert^2$, $a=(0.9, 0.9)$.
A nonconvex potential $V(x_1, x_2)=\sin(5\pi x_1)\sin(3\pi x_2)$, with $m=2$.

Incompressible limit

A disk obstacle (in white), with $V(x)=\lVert x-a\rVert^2$, $a=(0.9, 0.9)$.
A four-disk obstacle, with $V(x)=\lVert x-a\rVert^2$, $a=(0.9, 0.9)$.

Aggregation-diffusion equations

Consider the energy \(U(\rho) = \int_{\mathbb{R}^d} u_m(\rho(x))\,dx + \iint_{\mathbb{R}^d\times \mathbb{R}^d} \lVert x-y\rVert^2\rho(x)\rho(y)\,dxdy.\) The second term encourages aggregation and is concave with respect to $\rho$.

Here $m=2$.
3D view.

References

2021

  1. The back-and-forth method for Wasserstein gradient flows
    Matt Jacobs, Wonjun Lee, and Flavien Léger
    ESAIM Control Optim. Calc. Var., 2021

2020

  1. A fast approach to optimal transport: the back-and-forth method
    Matt Jacobs, and Flavien Léger
    Numer. Math., 2020

1998

  1. The variational formulation of the Fokker–Planck equation
    Richard Jordan, David Kinderlehrer, and Felix Otto
    SIAM J. Math. Anal., 1998