Introduction
Simulating the transport of physical quantities—such as mass, gas, or plasma—is a cornerstone of fluid and plasma dynamics. A fundamental requirement of these simulations is the preservation of positivity: densities must remain non-negative to be physically meaningful. However, standard numerical tools, particularly Galerkin methods, often produce "spurious non-positivity." In regions of low density, these methods can predict negative values, necessitating the use of "limiters" that manually zero out negative results. While limiters prevent simulation crashes, they often degrade accuracy in "rarefied" regions, which is critical for calculating chemical reaction rates or pressures.
A Statistical Lens on Discretization
This research proposes a paradigm shift by viewing numerical discretization through the lens of information geometry. The authors observe that standard Galerkin methods are mathematically equivalent to the Method of Moments (MoM) in statistics. Just as MoM can produce estimates inconsistent with observed data (e.g., negative probabilities), standard Galerkin projections can fall outside the space of valid probability densities.
To address this, the authors replace MoM with Maximum Likelihood Estimation (MLE). This leads to Maximum Likelihood Discretization (MLD), specifically the Fisher-Rao Galerkin (FRG) method. Instead of the standard $L^2$ inner product, which treats all errors equally regardless of the underlying density, FRG utilizes the Fisher-Rao metric. This metric is naturally suited for probability densities and ensures that the numerical solution remains within the manifold of valid densities.
Algorithmic Framework
The method is extended to a discontinuous framework, Discontinuous Fisher-Rao Galerkin (DFRG), which is essential for local mass conservation and handling steep gradients or shocks. The process involves alternating between time evolution ($Δt$) and a projection step (Proj) that maps the evolved state back onto the Galerkin space using the Fisher-Rao metric.
Key Insights: Geometry and Relative Error
By switching from the "flat" $L^2$ geometry to the "curved" Fisher-Rao geometry, positivity becomes an intrinsic property of the discretization space. A crucial advantage of this approach is its focus on relative error rather than absolute error. While standard methods minimize Euclidean distance, FRG minimizes Kullback-Leibler (KL) divergence. This makes the method significantly more sensitive to errors in low-density regions, where small absolute errors can represent massive relative discrepancies.
Results and Performance
The DFRG method was evaluated on several synthetic benchmarks, including 1D sinusoidal transport, bump advection, and 2D swirl problems. Key findings include:
- Strict Positivity: DFRG maintained strictly positive densities in all test cases, even where standard Discontinuous Galerkin (DG) methods failed.
- Superior Relative Accuracy: While $L^1$ and $L^2$ errors were comparable to standard methods, the KL divergence error for DFRG was orders of magnitude lower, confirming its effectiveness in rarefied zones.
- Robustness: In "extreme compression" scenarios where mass piles up, standard methods often diverged due to non-positive values, whereas DFRG remained stable.
Conclusion and Future Work
The Fisher-Rao Galerkin method provides a robust, statistically grounded framework for solving the transport equation while guaranteeing physical consistency. While the method introduces some computational complexity—specifically a solution-dependent mass matrix—it offers unparalleled accuracy for sensitive low-density simulations. Future work aims to extend this framework to more complex systems, including the Euler, Navier-Stokes, and Vlasov equations.
Summary updated