Tensor Trains Cut Vlasov-Maxwell DG Cost

Deterministic Vlasov-Maxwell solvers are attractive because they avoid particle noise, but their phase-space grids grow quickly with dimension and resolution. This paper targets that cost directly: keep the discontinuous Galerkin discretization, but represent the high-dimensional Vlasov distribution and its operators in tensor-train form. The result is a compressed TT-DG formulation for 1D2V kinetic plasma simulations, compared against a full-grid modal DG solver using the same discretization and explicit time integration.

Core Contribution

The main contribution is not a new plasma model; it is a compressed implementation of the existing DG machinery. The authors formulate quadrature, differentiation, nonlinear upwind flux evaluation, and time stepping so that the Vlasov part of the computation can be carried out directly on tensor trains. Maxwell’s equations remain in the full DG representation, which keeps the electromagnetic update simple while focusing compression where the phase-space tensor is largest.

That design choice matters. A low-rank representation is only useful if the solver can apply the numerical operators without constantly expanding back to the full grid. The paper’s TT-DG method follows a step-and-truncate pattern: after explicit Vlasov updates, the distribution is rounded to a prescribed tolerance, with $\epsilon_{TT}=10^{-6}$ used in the reported benchmark runs. This controls rank growth while introducing truncation error that must be checked against the full-grid reference.

Technical Approach

The implementation uses a modal DG discretization of polynomial degree $p=1$ on tensor-product grids. The TT solver is built with TT-Toolbox cross routines and THOR-accelerated tensor operations; the full-grid solver parallelizes the expensive Vlasov update across 44 Intel Xeon cores, while the Maxwell update is one-dimensional and not parallelized. This makes the comparison conservative in one sense: the TT implementation is tested against a parallel full-grid code, but the paper does not present a production GPU or distributed-memory implementation.

A practical detail receives real attention: tensor layout. The authors compare interleaved and grouped layouts, plus all six permutations of the phase-space variables $(x_2, v_{x1}, v_{x2})$, across streaming Weibel, weak Landau damping, and two-stream instability tests. Table 1 shows that interleaving consistently beats grouped layouts. Depending on problem and grid, the best layout is 1.91× to 4.73× faster than the worst tested layout for streaming Weibel and weak Landau damping, and 2.00× to 16.45× faster for two-stream instability. Compression also changes materially with ordering, reaching 58.8× higher mean compression for weak Landau damping in the layout study.

Figure 1 shows the clearest scaling result: for the streaming Weibel time-reversal benchmark at $T=1$, the TT solver becomes faster and more compressed as the uniform phase-space grid is refined.

Results and Analysis

The streaming Weibel benchmark tests both accuracy and reversibility. The setup uses periodic spatial length $\Omega=2\pi/k_0$ with $k_0=0.2$, velocity domain $[-1.2,1.2]^2$, perturbation parameters $\beta=0.01$, $\delta=0.5$, $v_{x1,0}=v_{x2,0}=0.3$, magnetic perturbation $b=10^{-3}$, and final time $T=1$. In the time-reversal test, the solver evolves forward, reverses velocity and magnetic-field sign, then evolves for another interval of length $T$.

The reversal errors in Table 2 are close between the full-grid and TT solvers. At $N=30$, the full-grid error in $f$ is 1.05E-02 and the TT error is 1.00E-02; at $N=160$, they are 5.89E-04 and 4.89E-04. Both methods show roughly second-order convergence. The authors correctly caution that exact reversibility is not expected after time discretization, TT truncation, and cross-interpolation error, but the numbers support the narrower claim that compression does not destroy the long-time reversibility behavior in this test.

The cost reduction is larger than the accuracy difference. For streaming Weibel, the TT speedup rises from 6.5× at $N=30$ to $3.84\times10^2$ at $N=160$, while compression increases from $9.53\times10^1$ to $1.56\times10^4$. For weak Landau damping, the fitted damping rates remain close to the theoretical rate of -0.153359. The full-grid run completes through $N_{v_{x1}}=N_{v_{x2}}=80$, with rates from -0.15468 to -0.15449; the TT run continues to 256 velocity cells, with rates from -0.15475 to -0.15771. That is a useful result: the compressed solver tracks the linear damping dynamics at the resolutions shared with the full-grid solver, then reaches finer velocity grids that the full-grid study did not complete.

Caveats

The evidence is strongest for the paper’s stated 1D2V benchmarks, not for general Vlasov-Maxwell workloads. Compression depends on phase-space structure; the abstract reports that the strongly nonlinear two-stream problem remains effective but becomes less compressible because of fine-scale filamentation. The solver is also a MATLAB implementation using external TT infrastructure, and the reported experiments keep Maxwell fields uncompressed. The paper therefore supports a focused conclusion: tensor trains can cut the cost of DG-based kinetic plasma simulations when the Vlasov distribution has exploitable low-rank structure, but the gains are problem- and layout-dependent.