Exact discrete-adjoint optimization of trap timing and placement in a Stieltjes-time reaction-diffusion model: A Galicia case study

What the math gives to ML

The paper's transferable asset is a discrete, rather than continuous, adjoint for dynamics whose effective time measure contains both flat intervals and point-mass jumps. A neural state-space layer can use the same Stieltjes clock to represent dormancy, instantaneous resets, scheduled token or event updates, or heterogeneous update rates without forcing every operation into an ordinary smooth-time ODE. The exact reverse recursion through the discretized residual gives gradients consistent with the actual forward computation and avoids one forward linearized solve per control or parameter direction. The most promising experiment is an event-driven recurrent or state-space model comparing a Stieltjes-clock implementation against dense small-step Euler integration at matched accuracy and wall-clock cost.