Abstract
We apply the level set method to compute the three dimensional multivalued geometrical optics term in a paraxial formulation. The paraxial formulation is obtained from the 3-D stationary eikonal equation by using one of the spatial directions as the artificial evolution direction. The advection velocity field used to move level sets is obtained by the method of characteristics; therefore the motion of level sets is defined in phase space. The multivalued travel-time and amplitude-related quantity are obtained from solving advection equations with source terms. We derive an amplitude formula in a reduced phase space which is very convenient to use in the level set framework. By using a semi-Lagrangian method in the paraxial formulation, the method has $O(N^2)$ rather than $O(N^4)$ memory storage requirement for up to $O(N^2)$ multiple point sources in the five dimensional phase space, where $N$ is the number of mesh points along one direction. Although the computational complexity is still $O(MN^4)$, where $M$ is the number of steps in the ODE solver for the semi-Lagrangian scheme, this disadvantage is largely overcome by the fact that up to $O(N^2)$ multiple point sources can be treated simultaneously.
Introduction
In high-frequency wave propagation, such as seismic imaging or underwater acoustics, the geometrical optics approximation is a standard tool for solving the Eikonal equation for travel time ($\tau$) and the Transport equation for amplitude ($A$). A persistent challenge in complex media is the occurrence of "multivalued" solutions. When rays cross, they create caustics where multiple waves arrive at a single point $(x, y, z)$ at different times. Standard PDE solvers typically find only the first arrival, discarding later arrivals that are often physically critical.
Existing methods to track these multiple arrivals in 3D are often computationally prohibitive, requiring massive memory to store five-dimensional phase space data ($x, y, z$ plus two takeoff angles). This paper presents a numerical framework that combines level set methods with a paraxial formulation and a semi-Lagrangian solver to handle these complexities efficiently.
Methodology
Paraxial Formulation and Phase Space
The authors tackle the multivalued problem through three key innovations. First, they employ a paraxial formulation, assuming waves generally travel along a primary direction (e.g., the $z$-axis). This allows $z$ to be treated as an "artificial time" variable, transforming the stationary 3D Eikonal equation into an evolution problem.
To handle multivaluedness, the problem is "lifted" into a 5D reduced phase space $(x, y, z, \theta, \psi)$. Rays are represented as the intersection of the zero level sets of two functions, $\phi^1$ and $\phi^2$. This transformation ensures that while the solution may be complex and multivalued in physical space, it remains a smooth, single-valued surface in phase space.
Global Semi-Lagrangian Solver
Instead of a standard grid-based Eulerian scheme, the authors use a semi-Lagrangian approach. They trace characteristics (rays) backward from the current $z$-plane to the source plane. Because this approach decouples the grid points in the $x-y$ plane, there is no need to store the full 4D grid at every step. This reduces the memory requirement from $O(N^4)$ to $O(N^2)$.
Key Insights: Simultaneous Sources and Efficiency
A major breakthrough of this method is the realization that the entire level set function contains useful data. The values of $\phi^1$ and $\phi^2$ at any point in phase space identify the specific source point $(x_s, y_s)$ from which a ray originated. Consequently, a single computation can resolve rays from up to $O(N^2)$ different point sources simultaneously without increasing the advection cost.
For a grid of $N=100$, the jump from $O(N^4)$ to $O(N^2)$ memory represents a reduction from 100 million values to just 10,000—a massive practical advantage for 3D engineering applications.
Results and Limitations
The method was evaluated using synthetic benchmarks, including a constant velocity model, a waveguide model with "swallowtail" caustics, and Vinje’s Gaussian model. The results demonstrated second-order convergence ($O(\Delta x^2)$) and successfully captured multiple arrival sheets and amplitude blow-ups at caustic points.
However, the method is constrained by the paraxial assumption; it cannot handle "turning waves" that travel backward relative to the $z$-axis. While the time complexity remains $O(MN^4)$, the ability to process thousands of sources at once significantly lowers the "cost per source," making it a highly efficient tool for large-scale wave physics problems.
Summary updated