Hysteretic Gain-Update Filter¶
Update guardrail — instantiates Adaptive Gain Retuning
A dwell requirement on the retuning loop: gain moves only after the evidence for moving it has persisted, so transient spikes can't rewrite sensitivity and the loop stops chasing noise.
An adaptive loop that reacts to every wiggle in its input will spend its life thrashing — nudging gain up on a momentary spike, back down on the lull that follows, never settling. The Hysteretic Gain-Update Filter puts a deliberate stickiness between the evidence and the adjustment. It requires that the case for a gain change persist — across enough time, enough samples, or a wide enough margin — before the change is allowed through, and it makes reversing a recent change harder than making it, so the loop can't oscillate around a boundary. Its distinguishing move is that it governs when gain may move, not how far or within what bounds: it is the low-pass filter on the retuning decision itself, the thing that separates a real regime change from a transient the loop should ignore.
Example¶
A cloud service autoscales, and the sensitivity of its scaling response — how strongly a rise in queue latency translates into adding capacity — is itself retuned as traffic patterns change. Without hysteresis the system flaps: a ten-second latency blip provokes a scale-out, the added capacity plus the passing blip makes latency look great, so it scales back in — just in time for the next blip. Each flap costs money and churns live connections. The Hysteretic Gain-Update Filter breaks the cycle by demanding sustained evidence: the responsiveness gain is only raised after latency stays elevated for, say, three consecutive intervals, and — the asymmetric part — it won't be lowered again until latency has stayed healthy for noticeably longer than it took to trigger the rise. A five-minute promotional spike no longer rewrites the system's sensitivity; a genuine, sustained shift in traffic still does, just a few minutes later. The cost is that deliberate lag; the payoff is a loop that stops chasing its own tail.
How it works¶
- Require persistence before acting. A proposed gain change must clear a dwell time, a sample count, or a confidence threshold before it takes effect — a brief excursion decays before it can move anything.
- Make reversal harder than action (the hysteresis). The threshold to undo a recent change is set wider than the threshold that made it, so gain can't oscillate across a single boundary.
- Distinguish transient from regime. Short, self-reversing excursions are absorbed; only changes that outlast the dwell are treated as real shifts in the world.
- Pass the survivors downstream. What clears the filter goes to the update rule and the bounds; what doesn't is simply forgotten.
Tuning parameters¶
- Dwell / persistence window — how long evidence must hold before gain moves. Longer rejects more noise but delays adaptation to real regime changes; this is the core speed-vs-stability dial.
- Hysteresis gap — how much wider the reversal threshold is than the trigger. A larger gap kills oscillation but leaves stale gain in place longer after conditions genuinely reverse.
- Evidence measure — consecutive intervals, a moving average, or a statistical change-point test. Stronger tests reject noise better at more cost and lag.
- Asymmetry — whether raising and lowering gain get the same dwell. Often you want to raise sensitivity fast (safety) and lower it slowly (stability), or the reverse.
When it helps, and when it misleads¶
Its strength is stability: it stops the adaptive loop from amplifying its own noise into a flapping, expensive, connection-churning mess, and it draws a principled line between a transient the loop should ignore and a regime change it should track. This is the same idea as a Schmitt trigger's two-threshold switching, where hysteresis turns a noisy signal crossing a level into one clean transition instead of a burst of chatter.[n1]
Its failure mode is lag, and lag can bite. Set the dwell too long and the filter faithfully protects a gain that is now wrong, holding old sensitivity through a real regime change until the miscalibration has already done its damage — numbness or clipping persisting precisely because the guardrail is doing its job too well. The classic misuse is turning the dwell up until the loop effectively never adapts, freezing gain under the respectable cover of "stability" — fixed-gain by the back door, minus the honesty of declaring it. The discipline is to size the dwell to the fastest regime change you must still catch, not to the noise you want to suppress, and to let the Saturation Occupancy Dashboard reveal when the filter is holding a stale gain too long.
How it implements the components¶
noise_chasing_guardrail— it is this component: the barrier that stops transient artifacts from rewriting sensitivity.adaptation_timescale— by setting how long evidence must persist, it sets the loop's effective adaptation timescale: deliberately slower than ordinary response, faster than the regimes it must track.
It decides only the timing of a change; the size and direction of the update come from the Automatic Gain Control Loop's rule, and the hard limits a change must respect come from the Gain Floor/Ceiling Rule. It suppresses noise but does not measure it — the input and occupancy monitors on the Saturation Occupancy Dashboard do that.
Related¶
- Instantiates: Adaptive Gain Retuning — the stability guardrail on the retuning loop.
- Consumes: Automatic Gain Control Loop — it filters that loop's proposed gain changes before they take effect.
- Sibling mechanisms: Automatic Gain Control Loop · Gain Floor/Ceiling Rule · Saturation Occupancy Dashboard · Contextual Gain-Scheduling Table · Gain-Change Review Log
Editorial Notes¶
Form Classification¶
Form family: Control, Automation & Runtime
Rationale: Hysteretic Gain-Update Filter operates as a live operational control that automatically routes, enforces, adapts, or responds during execution because it a dwell requirement on the retuning loop: gain moves only after the evidence for moving it has persisted, so transient spikes can't rewrite sensitivity and the loop stops chasing noise
Independent corroboration: The frozen evidence defines Hysteretic Gain-Update Filter as 'A dwell requirement on the retuning loop: gain moves only after the evidence for moving it has persisted, so transient spikes can't rewrite sensitivity and the loop stops chasing noise', so its operative form is Control, Automation & Runtime.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Engineering & Design
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: History-dependent gain updates using separate thresholds descend from control engineering and Schmitt-trigger logic.
Related originating lineages:
- Computer Science & Software Engineering — Software configuration and signal-processing filters materially implement stateful gain changes.
- Systems Thinking & Cybernetics — Adaptive feedback provides the general regulator-update framing.
Review resolution: Both reviewers independently assign engineering_design as the primary originating domain, so that shared primary is retained. Alternate domains are the union of reviewer-identified formative or independently originating lineages; later application settings alone are excluded. The final form materially composes methods or concepts from more than one formative domain. It has established independent use across several domains, but that does not make it domain-free. The encyclopedia entry makes that composition explicit.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
[n1] Hysteresis — making a system's response depend on its recent history, so the threshold to switch back differs from the threshold to switch. The Schmitt trigger is the canonical example: two switching thresholds convert a noisy input crossing a level into a single clean transition instead of repeated chatter. ↩