Abstract
This paper presents a novel level set framework designed to compute multi-valued solutions for non-linear partial differential equations (PDEs), specifically quasilinear hyperbolic equations and Hamilton-Jacobi equations. By lifting the problem into a higher-dimensional phase space, the authors transform complex folding singularities into a smooth, linear evolution governed by the Liouville equation. This approach allows for the efficient capture of caustics and multi-valued wave fronts using Eulerian grids, overcoming the limitations of traditional viscosity solutions and Lagrangian ray-tracing methods.
Introduction
In physical systems such as geometric optics, semiclassical mechanics, and seismic wave propagation, solutions to underlying equations often become multi-valued. For instance, when a wave front passes through a lens, it may fold over itself, resulting in multiple waves arriving at the same spatial point from different directions. Standard numerical approaches typically seek the "viscosity solution," which selects a single value (analogous to a shock in gas dynamics) and discards others. However, in applications like quantum chemistry or optics, retaining all values is essential to account for interference patterns and caustics—regions of high intensity where wave fronts converge.
Existing methods for tracking these multi-valued solutions generally fall into two categories: Lagrangian and Eulerian. Lagrangian methods, such as ray tracing, track individual particles or rays. While they can represent multiple values, they often suffer from poor resolution as rays diverge, leaving "holes" in the solution. Traditional Eulerian methods use fixed grids and are easier to implement but struggle to represent multiple values at a single grid point. This paper bridges that gap by introducing a level set method that operates in phase space.
The Phase Space Framework
The core innovation of this work is the "unfolding" of singularities by moving the problem from physical space ($x$) to phase space ($x, p$), where $p$ represents the gradient or velocity.
1. Lifting to Phase Space
Instead of solving directly for a potentially multi-valued function $u(x,t)$, the authors define a level set function $\phi(t, x, p)$. The solution $u$ is then represented as the zero level set of $\phi$. In this higher-dimensional space, a "fold" that appears as a singularity in physical space becomes a smooth, continuous surface.
2. The Liouville Equation
A significant advantage of this lifting is that the original non-linear PDE is transformed into a linear Liouville equation in phase space. Because the equation is linear, it naturally supports the superposition of solutions. This allows different wave fronts to pass through one another without artificial merging or numerical dissipation of the multi-valued structure.
3. Vector Level Sets
For multi-dimensional problems (e.g., 2D or 3D physical space), the method employs an intersection of multiple level set functions. The multi-valued solution is identified where these functions simultaneously vanish, allowing for the representation of complex geometric structures in high-dimensional space.
Numerical Methodology
To solve the Liouville equation on a fixed grid, the authors utilize high-order accurate numerical schemes. Specifically, they employ Essentially Non-Oscillatory (ENO) and Weighted ENO (WENO) schemes for spatial discretization, paired with Total Variation Diminishing (TVD) Runge-Kutta methods for time integration.
To address the "curse of dimensionality"—the exponential increase in computational cost as dimensions increase—the authors implement a Local Level Set Method. This technique restricts calculations to a narrow band around the zero level set. Consequently, the computational complexity is reduced to approximately $O(N^d \log N)$ for a $d$-dimensional physical problem, making the approach feasible for 2D and 3D applications.
Results and Discussion
The method was validated using several classic mathematical benchmarks, including the 1D and 2D Burgers' equations, the harmonic oscillator, and the 2D Eikonal equation.
Key findings include:
- Caustic Capture: The framework successfully captures complex caustic structures and multi-valued velocity fields that are typically lost in viscosity-based solvers.
- Uniform Resolution: Unlike Lagrangian ray tracing, this Eulerian approach maintains a uniform grid resolution across the entire domain, ensuring that no regions are left undersampled.
- Handling Discontinuities: For problems with discontinuous initial data, such as Riemann problems, the authors demonstrated that using a signed distance function as the initial level set is vital for correctly capturing rarefaction waves.
Numerical experiments, particularly those involving "Z-shaped" folds in 2D surfaces, demonstrate the method's ability to track multiple layers of a solution simultaneously and accurately.
Conclusion and Future Work
The level set method in phase space provides a robust and efficient Eulerian framework for computing multi-valued solutions to non-linear PDEs. By leveraging the linearity of the Liouville equation and the efficiency of local level set methods, the authors provide a powerful tool for fields requiring high-fidelity wave front tracking.
However, challenges remain regarding dimensionality and memory. Even with local methods, a 3D physical problem requires a 6D phase space. Future research will focus on further reducing storage requirements and extending the framework to more complex systems, such as the Euler-Poisson equations.
Summary updated