Fast and Slow Limits of Linear-Threshold Networks Are Stable

If you want neural circuits that settle down instead of wandering forever, this paper points to two extreme time scales. A linear-threshold network is a simple model of interacting neuron populations, with a dissipative part that pulls activity back and a recurrent part that feeds signals around the network. The authors build a one-parameter family of these networks that keeps the same equilibrium set and preserves a structural condition called Lyapunov diagonal stability. In the fast limit, the family becomes a projected dynamical system, and the paper proves that this limit is globally exponentially stable. In the slow limit, it becomes a discontinuous hard-selector system, and the paper proves that limit is globally asymptotically stable. The message is that the endpoints are not just mathematical curiosities: they capture the stability mechanisms of the whole family. The authors combine these proofs with numerical evidence and argue that checking the fast and slow limits may offer a structured path toward global stability for LTNs with asymmetric interactions and heterogeneous dissipation.

A neural circuit that keeps changing its mind is hard to trust. If you have ever watched a thermostat overshoot, you know the feeling. Linear-threshold networks, or LTNs, are brain-inspired systems that mix pull-back forces with looped signals. One part drains activity away. Another part feeds it back around the circuit. Those loops can make a circuit settle, switch, or wander. The surprise here is that two extreme time scales reveal the main stability story. At the fast end, the model behaves like a projected dynamical system, a rule that keeps motion inside limits. At the slow end, it behaves like a hard-selector system, where one choice takes over. That makes the far ends easier to study than the middle.

Two extremes, one steady destination

The key move is a one-parameter family of LTNs. A parameter is a knob that changes one balance without changing the steady states. The family keeps Lyapunov diagonal stability, a matrix test that says each part has enough built-in damping. That matters because the equilibrium set stays the same as the knob turns. In the fast limit, the family becomes a projected dynamical system. That limit is globally exponentially stable. Every path moves toward equilibrium at a fixed shrinking rate. In the slow limit, the family becomes a hard-selector system. That limit is globally asymptotically stable. Every path still settles down. The two endpoints therefore agree on the main outcome. They both point to stable behavior.

How the family is built

The construction turns one knob and leaves the destination fixed. That knob changes how fast the recurrent part acts compared with the dissipative part. The dissipative part is the piece that pulls activity back. The recurrent part is the loop that sends signals around again. By tuning that balance, the same equilibrium set stays in view. The fast version tightens the system into a projected dynamical system. The slow version loosens it into a hard-selector system. The point is not just to build two limits. The point is to keep the same structural test, Lyapunov diagonal stability, across the whole path. That makes the endpoints a fair guide to the family.

Why the endpoints matter

This result changes where you look first. Instead of attacking the full network in one step, the stability test can start at the two ends. If the fast end and the slow end both behave well, they point to the same stable core. That is useful for LTNs with asymmetric links, where symmetry tricks do not help. It is also useful when damping is uneven across units. The paper's message is modest but sharp. Resolve the endpoints, and you may learn the shape of the whole family. The numerical checks support that view. They show the limiting systems are not random edge cases. They are the right bookends for the full model.

What to test next

The surprise is that the most useful clues sit at the edges. The fast limit gives exponential pull. The slow limit still gives convergence. Together, they turn stability into a two-end test. The next check is the full one-parameter family itself, still under Lyapunov diagonal stability, across asymmetric links and uneven damping. If those endpoints keep matching the middle, this route could turn a hard global proof into a structured chain of smaller ones. That is a practical shift. It tells control theory where to spend its effort first.