Parameterized Wasserstein Hamiltonian Flow

Hao Wu,Shu Liu,Xiaojing Ye,Haomin Zhou

2023-07-16

Abstract:In this work, we propose a numerical method to compute the Wasserstein Hamiltonian flow (WHF), which is a Hamiltonian system on the probability density manifold. Many well-known PDE systems can be reformulated as WHFs. We use parameterized function as push-forward map to characterize the solution of WHF, and convert the PDE to a finite-dimensional ODE system, which is a Hamiltonian system in the phase space of the parameter manifold. We establish error analysis results for the continuous time approximation scheme in Wasserstein metric. For the numerical implementation, we use neural networks as push-forward maps. We apply an effective symplectic scheme to solve the derived Hamiltonian ODE system so that the method preserves some important quantities such as total energy. The computation is done by fully deterministic symplectic integrator without any neural network training. Thus, our method does not involve direct optimization over network parameters and hence can avoid the error introduced by stochastic gradient descent (SGD) methods, which is usually hard to quantify and measure. The proposed algorithm is a sampling-based approach that scales well to higher dimensional problems. In addition, the method also provides an alternative connection between the Lagrangian and Eulerian perspectives of the original WHF through the parameterized ODE dynamics.

Numerical Analysis

What problem does this paper attempt to address?

### The problems the paper attempts to solve

This paper aims to solve the numerical computation problem of Wasserstein Hamiltonian Flow (WHF). Specifically, WHF is a Hamiltonian system defined on the probability density manifold, and many well - known partial differential equations (PDEs) can be reformulated in the form of WHF. However, computing WHF in high - dimensional spaces remains a challenging problem. To this end, the author proposes a new numerical method to solve this problem through the following steps:

1. **Parameterized function as forward mapping**: Use a parameterized function as forward mapping to characterize the solution of WHF and convert the PDE into a finite - dimensional ordinary differential equation (ODE) system. This ODE system is a Hamiltonian system in the phase space.

2. **Theoretical error bound**: Establish the theoretical error bound of the continuous - time approximation scheme under the Wasserstein metric.

3. **Numerical implementation**: Use a neural network as forward mapping and design an effective symplectic scheme to solve the derived Hamiltonian ODE system, thus preserving some important quantities such as the Hamiltonian.

4. **Training - free optimization**: The computation process is completely completed by a deterministic symplectic integrator, without the need for direct optimization of the neural network, thus avoiding the difficult - to - quantify and - measure errors introduced by stochastic gradient descent (SGD) or similar methods.

5. **Scalability for high - dimensional problems**: This method is a sampling - based method and is suitable for high - dimensional problems.

### Main contributions

1. **Parameterized WHF (PWHF)**: Derive PWHF by projecting the Lagrangian onto the parameter space. This is a finite - dimensional ODE system that can effectively approximate the infinite - dimensional WHF.

2. **Symplectic numerical scheme**: Design a symplectic numerical scheme to solve PWHF. This scheme can preserve the symplectic structure of the system, thereby preserving the conservation of the Hamiltonian over a long period of time.

3. **Simplified Wasserstein metric**: Introduce a simplified Wasserstein metric, which significantly reduces the computational cost and can theoretically guarantee the convergence of the approximation error.

4. **Efficient computation for high - dimensional problems**: This method performs well in high - dimensional problems and is suitable for large - scale computation tasks.

### Conclusion

This paper proposes a new numerical method to compute Wasserstein Hamiltonian Flow. Through parameterization and symplectic numerical schemes, it effectively solves the computational challenges in high - dimensional problems. This method not only has good convergence theoretically but also performs well in practical applications, providing new tools and perspectives for solving complex PDE problems.