Implementation Status

Implemented

  • Differentiable one-parameter Mittag-Leffler evaluator.
  • Explicit custom backward for the truncated series branch.
  • History-free spectral layer with differentiable alpha and beta.
  • 1D and 2D Dirichlet spectral constructors.
  • Stable preset via MLSLConfig.stable() for broader negative-real spectral regimes.
  • 1D Neumann spectral constructor in addition to 1D/2D Dirichlet constructors.
  • 1D periodic and mixed-boundary spectral constructors.
  • 2D Neumann, periodic, and mixed-boundary tensor-product constructors.
  • Batched forward evaluation for dataset-level use.
  • Baselines and experiment suite covering FNO, DeepONet, fPINN, OOD alpha, sparse inverse recovery, 2D inverse recovery, and runtime scaling.
  • Unit tests for layer shapes, gradients, hybrid finite values, and 2D use.
  • Examples for quickstart, batched evaluation, and inverse alpha/beta recovery.
  • Primitive generality matrix across dimensions, fractional orders, initial conditions, batching, gradients, and inverse recovery.
  • Paper-grade extension suite covering long horizons, repeated seeds, batch and device profiling, stronger neural baselines, forcing, fPINN repetition, and proposition-level timing evidence.
  • Manufactured-solution validation for the forced MLSL component in a multi-mode spectral regime.
  • Boundary-condition generality checks for 1D Dirichlet and Neumann settings.
  • Extended boundary-condition generality checks for 1D Dirichlet, Neumann, periodic, and mixed settings.
  • High-precision reference checks for the hybrid two-parameter Mittag-Leffler evaluator.
  • Shifted-spectrum linear reaction-diffusion family experiment.
  • Semilinear cubic-reaction backbone experiment using MLSL as the linear history-free primitive.
  • Larger multi-seed neural baseline experiment.
  • CUDA validation on RTX 5070, including CPU/GPU consistency, batch profiling, and half-precision capability probes.
  • Diagnostic one-/two-parameter Mittag-Leffler evaluation with embedded truncation disagreement and branch counts.
  • True direct/stable algorithm reconfiguration plus automatic regime selection.
  • Generalized symmetric stiffness/mass spectral problems with mass projection.
  • Differentiable constant-order linear Caputo L1 history fallback.
  • Dense symmetric-PSD Lanczos Mittag-Leffler matrix-function action, including batched initial states, automatic size-based selection, and empirical subspace-disagreement diagnostics.
  • Sparse tensor and matrix-free self-adjoint operator paths, including differentiable operator parameters, N=4096 validation, and CUDA execution.
  • FFT-accelerated Caputo-L1 full-trajectory history convolution, including direct-reference agreement, analytic power-law convergence, gradients, N=65536 execution, and CPU/GPU consistency.
  • Controlled complex-argument Mittag-Leffler evaluation and Arnoldi actions for general operators, validated against mpmath, high-precision matrix series, finite differences, matrix exponentials, and CUDA complex128.
  • Domain-oriented application cases for regular-domain anomalous diffusion, assembled finite-element relaxation, graph memory diffusion, and controlled periodic fractional advection--diffusion. Each case reports its recommended algorithm, differentiable parameters, assumptions, and limitations.
  • Experimental H-actin SPT evidence chain with 3,112 trajectories, trajectory-level leakage control, traditional and neural baselines, five splits, and an MLSL-residual regression component. Raw-data redistribution remains pending because the source page does not state a license.
  • Semilinear mild-form Picard solver with convergence diagnostics, non-success retcodes, and autograd through fractional and nonlinear parameters.
  • Single-package dfsc source layout; the legacy mlsl distribution has been removed.
  • Machine-readable numerical reliability reports and strict validated-domain evaluation.
  • A 20-check pre-release internal readiness gate covering numerical, API, test, application, documentation, and packaging evidence.

Still Needed For Paper-Grade Primitive

  • More accurate stable Mittag-Leffler evaluator with reference validation over wider long-time and stiff regimes.
  • Contour or rational evaluation beyond the current conservative complex radius, including non-normal and pseudospectral error analysis.
  • Wider reference validation for the hybrid two-parameter Mittag-Leffler evaluator in stiffer forced dynamics and larger negative arguments.
  • Larger GPU profiling for full baseline training workloads.
  • Broader nonlinear or variable-coefficient PDE families beyond linear spectral cases.
  • Broader nonlinear solvers beyond the current fixed-point mild formulation.
  • Block/preconditioned Krylov actions, contour methods, and online/adaptive fast-memory CQ/SOE time-stepping methods.
  • Larger benchmark datasets with repeated seeds and confidence intervals.
  • Cleaner narrative comparing MLSL as a primitive against black-box operator learners and residual-only fPINNs.