Oscillation, and the limit on gain

How much correction is too much, and what happens past that point.

There is a hard limit on how much correction, or gain, a control loop can apply before it stops settling calmly toward its target and starts oscillating instead, because past that limit each correction overshoots by more than the error it was meant to fix, handing the next correction a bigger error to respond to than the one it had only just finished correcting, and a loop that makes its own error larger with every pass around it has stopped controlling the system entirely and started amplifying its own mistakes instead.

What is actually happening

A well-tuned control loop applies a correction that closes most, but not quite all, of the current error, leaving a smaller error behind for the next correction to close a little further, each pass around the loop shrinking the gap until it disappears. Turning up the gain, the strength of that correction relative to the error it is responding to, makes each pass close more of the gap faster, which sounds on the surface like it should only ever help. Past a certain point, though, the correction stops merely closing the gap and starts overshooting past the target by more than the original error, leaving behind a new error on the opposite side that is actually larger than the one it just fixed. The next correction, responding to that larger error with the same excessive gain, overshoots even further the other way, and the error grows with every pass rather than shrinking, a runaway that has nothing to do with the target moving and everything to do with the loop's own corrections feeding on themselves.

The microphone-feedback comparison

Point a microphone at the speaker it is connected to and the result, within seconds, is the familiar rising electronic squeal every public address system has produced at least once. The microphone picks up a faint sound, the amplifier makes it louder, the speaker plays that louder sound back out, the microphone picks up the now-louder sound again, and the amplifier makes it louder still, each trip around that loop multiplying the signal rather than correcting anything toward a target at all. With the amplifier's gain turned down low enough, the same loop is perfectly harmless, since each trip around it makes the sound only slightly louder before it fades away naturally, no different from an ordinary echo dying out in a quiet room, but past a certain gain the loop starts adding more sound on each pass than is lost in between, and the signal grows without any real limit until it slams into the loudest sound the speaker can physically produce. A control loop with too much gain runs the same escalating loop, just with a temperature or a position standing in for the sound, and an overcorrection standing in for the amplifier's excess.

Why the limit is not a fixed number

The gain at which a particular loop tips over from settling into oscillation depends on the whole system, not on the controller alone, since a system with more delay between a correction being applied and its effect being measured tolerates far less gain before crossing that limit than a system where cause and effect are nearly instantaneous. This is why the same controller, with the same gain setting, can behave perfectly well on one machine and oscillate wildly on another, if the second machine simply takes longer for a correction to actually show up in what is being measured, the delay problem explored in more depth in the final article of this set, and it is why a controller copied unchanged from one machine to a superficially similar one is never a safe assumption.

One figure worth keeping in mind

A control loop tuned right at the edge of this limit can appear stable indefinitely under calm conditions and then begin oscillating with growing amplitude the moment a small disturbance nudges it, since a system sitting exactly on the boundary between settling and diverging has no real margin left to absorb anything extra, and even a tiny push, a gust, a bump, a moment's extra friction, is enough to reveal how little safety margin was actually built into the tuning in the first place.

Where low gain causes its own kind of trouble

The instinct to simply keep gain low, well clear of the limit that causes oscillation, is not free of cost either, since a loop with very low gain corrects errors sluggishly, taking a long time to close even a small gap and responding weakly to any real disturbance that pushes the system away from its target. This is the same speed-versus-restraint trade-off already covered when overshoot was discussed earlier in this set, restated here in terms of gain rather than damping, a controller tuned too cautiously can technically be stable while still performing poorly, drifting noticeably off target for long stretches simply because it was never given enough authority to correct itself promptly.

Why this matters in practice

Recognising oscillation as a symptom of gain pushed past what the system's own delay and momentum can tolerate, rather than as some unrelated fault, changes how a real controller gets tuned, since the fix for a system that has started oscillating is very often to turn the gain down rather than to search for some other unrelated problem. Engineers tuning a real loop deliberately leave a real margin below this limit precisely because the resonance-like escalation past it happens abruptly rather than gradually, a system that behaves perfectly well at ninety percent of the critical gain can tip into runaway oscillation at only a slightly higher setting, with very little warning offered in between, which is why the search for the right gain is usually conducted carefully upward from a setting known to be safely sluggish, rather than downward from a setting suspected of already being too aggressive.