Rigorous error analysis, temporal sampling complexity, and long-time mass conservation for physics-informed neural network solvers of the time-dependent Dirac equation

The time-dependent Dirac equation (TDDE) governs relativistic spin-1/2 particles and evolves on the extremely short Compton time scale $1/(mc^2)$, which strains every numerical method. Lorin and Yang (2022) showed that physics-informed neural networks (PINNs) can solve the TDDE empirically — in flat and curved space, with electromagnetic potentials — while avoiding the fermion-doubling artifacts of grid discretizations; but they explicitly identified what is missing: no convergence or error theory exists for the method, the number of temporal training points must scale like $Tmc^2$, long-time $L^2$-norm (mass) conservation holds only with explicit penalties or slabwise time-domain decomposition (which sacrifices embarrassingly parallel global training), and stochastic-gradient training of the nonconvex loss carries no global-convergence guarantee. PINN error analysis now covers elliptic, parabolic, second-order hyperbolic, kinetic (Boltzmann), and dispersive (nonlinear Schrödinger) equations, and follow-up work develops decomposition and continuation heuristics — but no work supplies error bounds, improved temporal scaling, or long-time norm-preservation guarantees for PINNs on the TDDE. The problem remains open: build the complete numerical-analysis foundation for PINN solvers of the TDDE — an error decomposition with explicit dependence on $m$, $c$, and $T$; a necessity-or-circumvention result for the $Tmc^2$ temporal sampling requirement; and a guarantee-or-impossibility result for long-time $L^2$-norm conservation under global single-network training.

Background

The time-dependent Dirac equation (TDDE) is the relativistic wave equation for spin-1/2 particles such as electrons, and the starting point for describing relativistic laser–matter interactions. In scaled units it reads $i\partial_t\psi = H_D\psi$ on a spatial domain $\Omega$, where $H_D = -ic\,\alpha\cdot(\nabla + ieA) + mc^2\beta + V$ acts on a spinor $\psi$ (a 2-component complex wavefunction in the two-dimensional setting of the source work, or a 4-component one in three space dimensions), $m$ is the particle mass, $c$ the speed of light, $A$ the electromagnetic vector potential, $V$ a scalar potential, and $\alpha$, $\beta$ are Dirac matrices. Two numerical pathologies are specific to this equation. First, the solution oscillates on the Compton time scale $1/(mc^2)$; the rapid interference between positive- and negative-energy components is called zitterbewegung, and standard time-stepping solvers accordingly need steps $\Delta t < 1/(mc^2)$. Second, grid discretizations of the first-order Dirac operator suffer from fermion doubling: spurious duplicated high-energy modes appearing in naive finite-difference schemes. The evolution generated by the self-adjoint $H_D$ is unitary, so the $L^2$-norm $\|\psi(\cdot,t)\|_{L^2(\Omega)} = 1$ is conserved; this is interpreted as probability/mass conservation and any credible solver must preserve it. Exact conservation holds on the whole space or under self-adjoint boundary conditions; on the bounded domain with the null Dirichlet condition used in the source's computations, the invariant of the unitary flow remains the reference quantity that a PINN solution is required to reproduce to within controlled error over long times. Physics-informed neural networks (PINNs; Raissi, Perdikaris and Karniadakis, 2019) approximate the solution by a space-time network $N_\theta(x,t)$ whose derivatives are computed exactly by automatic differentiation (the chain-rule-based evaluation of derivatives of a neural network with respect to its inputs). The parameters are trained by stochastic gradient descent (SGD) to minimize a nonconvex least-squares loss combining the PDE residual $\|H_D N_\theta\|^2_{L^2(\Omega\times(0,T))}$, evaluated at random collocation points, with penalty terms for the initial condition and, optionally, the norm constraint. Because no grid or time step is used, the method avoids fermion doubling and its training is embarrassingly parallel across sampling points. Lorin and Yang (2022) applied this construction to the TDDE in several physical frameworks and verified empirically: freedom from fermion doubling; acceptable norm conservation on short time intervals when enough temporal training points are used or an explicit norm penalty is added; and norm preservation over long times via slabwise time-domain decomposition — solving successively on small intervals $[t_i, t_{i+1}]$, each slab warm-started from the previous one — at the price of losing embarrassingly parallel global training. They also identified precisely what is missing. The number of temporal training points must scale like $Tmc^2$ (the PINN analogue of the $1/(mc^2)$ time-step constraint), making global solves prohibitive for large $c$ or long $T$. SGD on the nonconvex loss has no global-convergence guarantee, so the accuracy of a trained network cannot be certified. Most fundamentally, no convergence or error theory exists for the method: the rigorous PINN analyses available at the time covered only linear second-order elliptic and parabolic PDEs (Shin, Darbon and Karniadakis, 2020), and the coercive-bilinear-form frameworks developed subsequently (Zeinhofer, Masri and Mardal, 2023) cover elliptic, parabolic, and second-order hyperbolic equations — whereas the Dirac operator is a first-order hyperbolic system lacking the coercivity those proofs rely on.

Problem Statement

Supply the missing numerical-analysis foundation for global space-time PINN solvers of the time-dependent Dirac equation, as identified by Lorin and Yang (2022). The three requested results serve one objective — turning the empirically demonstrated PINN solver for the TDDE into a method with provable accuracy, cost, and invariant-preservation guarantees — and a complete answer settles all three in the setting of the source work (the two-dimensional TDDE on a bounded domain with null Dirichlet boundary conditions, initial data enforced through the loss, and optionally a norm-penalty term; a clearly specified variant loss is admissible provided the analysis is stated for it):

(1) Error analysis. Prove a priori error bounds for the TDDE PINN — for the loss minimizer, and if possible for the trained network — decomposing the total error into approximation, optimization, and estimation (generalization) errors, with explicit dependence on the final time $T$, the mass $m$, the speed of light $c$, the network size, and the number of space-time collocation points. Existing coercivity-based frameworks, which cover elliptic, parabolic, and second-order hyperbolic equations, do not apply to this first-order hyperbolic system, so the stability analysis must be genuinely new.

(2) Temporal sampling complexity. Determine whether the temporal training-point requirement that scales like $Tmc^2$ is intrinsic: either prove a lower bound showing that collocation-based PINNs for the TDDE with data oscillating at frequency scale $mc^2$ require temporal sampling on the order of $Tmc^2$, or construct a parameterization and sampling scheme with a provably better dependence on $T$, $m$, and $c$.

(3) Long-time mass conservation. Determine whether the global single-network PINN can provably preserve the conserved $L^2$-norm $\|\psi(\cdot,t)\|_{L^2(\Omega)}=1$ (the invariant of the unitary continuous flow identified in the Background) over times $T\gg 1/(mc^2)$ without slabwise time-domain decomposition: either give a construction with a norm-conservation guarantee (with or without the penalty term used in the source), or prove that the unconstrained global loss admits (near-)minimizers, or gradient-reachable iterates, whose norm drifts, so that decomposition or explicit constraints are necessary.

Negative resolutions — an impossibility proof, a lower bound, or a drift construction — count as answers where stated.

References

1. Emmanuel Lorin and Xu Yang, 'Computation of the Time-Dependent Dirac Equation with Physics-Informed Neural Networks' (2022; published as 'Time-dependent Dirac Equation with Physics-Informed Neural Networks: Computation and Properties', Computer Physics Communications 280:108474), arXiv:2204.02959, DOI 10.48550/arXiv.2204.02959, https://arxiv.org/abs/2204.02959; journal version DOI 10.1016/j.cpc.2022.108474, https://doi.org/10.1016/j.cpc.2022.108474
2. Yeonjong Shin, Jérôme Darbon, and George Em Karniadakis, 'On the Convergence of Physics Informed Neural Networks for Linear Second-Order Elliptic and Parabolic Type PDEs', Communications in Computational Physics 28(5):2042–2074, 2020, DOI 10.4208/cicp.oa-2020-0193, https://doi.org/10.4208/cicp.oa-2020-0193
3. Marius Zeinhofer, Rami Masri, and Kent-André Mardal, 'A Unified Framework for the Error Analysis of Physics-Informed Neural Networks', 2023, arXiv:2311.00529, DOI 10.48550/arXiv.2311.00529, https://arxiv.org/abs/2311.00529; published in IMA Journal of Numerical Analysis 45:2988–3025, DOI 10.1093/imanum/drae081, https://doi.org/10.1093/imanum/drae081
4. Tim De Ryck and Siddhartha Mishra, 'Generic bounds on the approximation error for physics-informed (and) operator learning', 2022, arXiv:2205.11393, DOI 10.48550/arXiv.2205.11393, https://arxiv.org/abs/2205.11393; also in Advances in Neural Information Processing Systems 35:10945–10958, 2022, DOI 10.52202/068431-0795, https://doi.org/10.52202/068431-0795
5. Elie Abdo, Lihui Chai, Ruimeng Hu, and Xu Yang, 'Error estimates of physics-informed neural networks for approximating Boltzmann equation', 2024, arXiv:2407.08383, DOI 10.48550/arXiv.2407.08383, https://arxiv.org/abs/2407.08383; published as 'Error estimates of physics-informed neural networks for approximating Boltzmann equations', IMA Journal of Numerical Analysis, article draf142, 2026, DOI 10.1093/imanum/draf142, https://doi.org/10.1093/imanum/draf142
6. Miguel Á. Alejo, Lucrezia Cossetti, Luca Fanelli, Claudio Muñoz, and Nicolás Valenzuela, 'Error bounds for Physics Informed Neural Networks in Nonlinear Schrödinger equations placed on unbounded domains', 2024, arXiv:2409.17938, DOI 10.48550/arXiv.2409.17938, https://arxiv.org/abs/2409.17938
7. Karan Shah, Patrick Stiller, Nico Hoffmann, and Attila Cangi, 'Physics-Informed Neural Networks as Solvers for the Time-Dependent Schrödinger Equation', 2022, arXiv:2210.12522, DOI 10.48550/arXiv.2210.12522, https://arxiv.org/abs/2210.12522
8. Emmanuel Lorin and Xu Yang, 'Quasi-optimal domain decomposition method for neural network-based computation of the time-dependent Schrödinger equation', Computer Physics Communications, 2024, DOI 10.1016/j.cpc.2024.109129, https://doi.org/10.1016/j.cpc.2024.109129
9. Chuanxin Wang, Tomoya Naito, Jian Li, and Haozhao Liang, 'A deep neural network approach to solve the Dirac equation', 2024, arXiv:2412.03090, DOI 10.48550/arXiv.2412.03090, https://arxiv.org/abs/2412.03090; published in The European Physical Journal A 61:162, 2025, DOI 10.1140/epja/s10050-025-01630-5, https://doi.org/10.1140/epja/s10050-025-01630-5
10. Aigerim Bazarkhanova, 'Physics Informed Neural Networks for Solving Dirac Equation in (1+1) Dimension', Master's thesis, Nazarbayev University, 2021, https://nur.nu.edu.kz/handle/123456789/5392
11. Emmanuel Lorin, Yuchong Tang, Xu Yang, and Yi Zhu, 'A Jin–Xin Relaxation Gradual Convergence Method for Conservation-Law PINNs', 2026, arXiv:2608.22493, DOI 10.48550/arXiv.2608.22493, https://arxiv.org/abs/2608.22493
12. Angelos P. Giotis, Athanasios Gkrepis, Odysseas Kosmas, and Theocharis S. Kosmas, 'A unified Python framework for solving Schrödinger and Dirac equations in exotic lepton–antilepton atoms', Computer Physics Communications, 2026, DOI 10.1016/j.cpc.2026.110335, https://doi.org/10.1016/j.cpc.2026.110335
13. M. Raissi, P. Perdikaris, and G. E. Karniadakis, 'Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations', Journal of Computational Physics 378:686–707, 2019, DOI 10.1016/j.jcp.2018.10.045, https://doi.org/10.1016/j.jcp.2018.10.045