An L-Stable Sequential Two-Stage Fourth-Order Method with ADER Trajectory Derivatives for Stiff Transport--Relaxation Systems

What the math gives to ML

The paper develops a stiffly stable time integrator whose key transferable asset is the trajectory derivative \(\mathcal{G}^{\rm tr}=D\mathcal{L}(U)\mathcal{L}(U)\), the directional derivative of a vector field along its own trajectory. In neural-network training, this becomes a Hessian-vector product \(H(\theta)g(\theta)\), capturing local curvature evolution without forming a full Hessian. The paper's L-stability and cancellation of the leading inverse-power term suggest an optimizer that damps unresolved high-curvature modes instead of taking unstable explicit steps. A practical transfer is an implicit trajectory-derivative optimizer for small and medium models, using conjugate gradients and automatic-differentiation Hessian-vector products, with stiff-decay behavior tested against AdamW and implicit-gradient baselines.