Solving partial differential equations (PDEs) that describe how probability densities evolve—such as the Fokker-Planck, Schrödinger, or continuity equations—is a fundamental challenge in physics and statistics. Traditionally, these problems are solved using grid-based methods that discretize space. However, as the number of dimensions increases, the number of grid points grows exponentially, a phenomenon known as the "curse of dimensionality."
A growing body of research proposes a shift in perspective: instead of approximating the density function directly, we can parameterize the pushforward map that transforms a simple reference distribution into the target density. By using neural networks to represent these maps, researchers can transform infinite-dimensional PDE problems into finite-dimensional optimization problems in parameter space.
From Densities to Pushforward Maps
The core intuition behind this research direction is the move from an Eulerian view (tracking density values at fixed points in space) to a Lagrangian view (tracking the trajectories of individual "probability mass" parcels).
In the framework established by Wuchen Li, Hongyuan Zha, and Haomin Zhou in the Parametric Fokker-Planck Equation, a probability distribution is represented by a map $T_\theta$ that pushes forward a static reference distribution (like a standard Gaussian) to the desired density.
This approach has two immediate advantages:
- Instantaneous Sampling: Once the parameters $\theta$ are found for a specific time, generating samples is as simple as passing reference points through the map $T_\theta$. This bypasses the need for expensive Markov Chain Monte Carlo (MCMC) methods, as noted in Learning to Sample Better.
- Intrinsic Constraints: Because the map $T_\theta$ merely moves mass around, the resulting density is guaranteed to be non-negative and integrate to one. As Zebang Shen and colleagues show in Self-Consistency of the Fokker Planck Equation, modeling the underlying velocity field ensures these physical properties are satisfied by construction, rather than through soft penalties.
Reducing Dimensions via Parameterized Paths
By restricting the evolution of a distribution to a manifold defined by the parameters of a neural network, the problem of solving a PDE becomes a problem of finding the "optimal" trajectory of parameters $\theta(t)$.
In Constrained dynamical optimal transport and its Lagrangian formulation, Wuchen Li demonstrates that this reduction transforms infinite-dimensional density optimization into a finite-dimensional problem compatible with standard gradient-based optimization. This is further refined in Parametric Density Path Optimization (PDPO) by Sebastian Gutierrez Hernandez and Peng Chen, which uses spline-interpolated maps to achieve 40–80% faster runtimes than traditional solvers in spaces with up to 1,000 dimensions (PDPO).
Wasserstein Gradient and Hamiltonian Flows
The geometry of these parameter spaces is often governed by the Wasserstein metric, which measures the distance between distributions based on the cost of transporting mass.
- Gradient Flows: For dissipative systems like the Fokker-Planck equation, the evolution follows a Parameterized Wasserstein Gradient Flow. Jiaojiao Fan and Yongxin Chen introduced JKO-Flow (Variational Wasserstein gradient flow), which uses the Jordan-Kinderlehrer-Otto (JKO) scheme to learn transport maps without needing explicit density estimation or Hessian calculations.
- Hamiltonian Flows: For conservative systems like the Schrödinger or Vlasov equations, the evolution follows a Parameterized Wasserstein Hamiltonian Flow. Hao Wu and Shu Liu developed symplectic integrators that preserve the energy of these high-dimensional systems while scaling to dimensions that are computationally inaccessible to grid-based solvers (A parameterized Wasserstein Hamiltonian flow approach for solving the Schrödinger equation).
Physics-Informed Architectures
A practical application of this framework is in fluid dynamics and conservation laws. Fabricio Arend Torres and Volker Roth introduced Lagrangian Flow Networks (LFlows) (Lagrangian Flow Networks for Conservation Laws). By using time-conditioned normalizing flows to represent the transformation of a base distribution, LFlows satisfy the continuity equation (mass conservation) by design. This allows the model to learn complex dynamics from sparse observations without requiring the "soft" penalty terms often used in physics-informed neural networks (PINNs).
By shifting the computational burden from grid-based discretization to the optimization of continuous neural transformations, this research direction provides a scalable path for simulating complex, high-dimensional physical systems and sampling from intricate probability distributions.