A stable system behaves like a ball in a bowl: give it a push, and it rolls back to the bottom. The spread of possible trajectories, uncertainty, shrinks over time. But the speed depends on the slope of the bowl's walls. In complex systems, this 'bowl' has sections of varying steepness. Previously, the recovery time was estimated using the flattest part — a safety margin that overestimated the forecast. The new method calculates the average slope where the system actually spends its time.
The authors showed that replacing a simple controller with a nonlinear one (akin to a smart suspension) dampens random jolts more effectively. While the maximum stiffness is the same, the bowl’s shape changes, nudging the system back to normal more actively. This improves machine learning algorithms, where the goal is to converge on the right data faster.
Like a toy spinning top that wobbles harder and harder, then suddenly settles, a simulation error swells to a peak and then abruptly drops, and the moment of that peak can be calculated ahead, if you know the average slope.
In Isaac Asimov’s stories, robots obey laws that ensure their predictability — much like contraction theory promises that a system will return to normal after any glitch.