Thermal noise and the fluctuation-dissipation theorem in public safety

A resistor, a duty officer, and the same law

A resistor sitting on a bench, no current driven through it, still produces a small random voltage across its terminals. Johnson measured it in 1928 as an unavoidable hiss in vacuum-tube amplifiers. Nyquist gave the size of it: mean square voltage equals 4kTRΔf, where R is the resistance that will later dissipate energy when you do drive current through it. The idle jitter and the working dissipation share a number. Kubo generalised the result in 1957: near equilibrium, the way a system responds to a small push is fixed by the autocorrelation of how it jitters when nobody pushes it. You do not need to poke the system to know how it will answer a poke. You need to have watched it sit still, for long enough, without looking away.

Public safety runs the same law over incident feeds instead of electrons. A duty officer's control room is a resistor bench with worse instrumentation and higher stakes.

What arrives

Four streams land continuously: the incident feed (calls for service, triaged and coded as they come in), dispatch telemetry (vehicle locations, unit status, response and turnaround times), sensor networks (fixed cameras, ANPR reads, gunshot-detection microphones, footfall counters, flood gauges), and weather (temperature, wind, precipitation, tide where relevant). None of these arrives as a report. They arrive as raw, timestamped ticks, most of them boring. A patrol car idling at a junction for eleven minutes generates no incident and no story. It generates telemetry.

That boring tick is the resistor's idle voltage. It is not noise in the sense of waste. It is the system sitting still, and fluctuation-dissipation says the way it sits still encodes how it will move when something happens.

What is held

Here the three generations diverge sharply, and public safety is a clean place to see it. A system built like a Large Language Model would hold none of the ticks. It would hold the write-ups: after-action reports, closed-incident summaries, the annual crime pattern analysis that becomes next year's staffing plan. Those documents describe perturbations already interpreted by whoever wrote the postmortem. The eleven quiet minutes never made it into any report, because nothing happened in them, so nothing was written.

A system built like a Large World Model would hold a bounded scene — a borough on a Saturday night, a stadium during an event, sensor and dispatch data fused for the duration of the operation. This is real intake, and it captures the perturbation live. But the recording starts when the operation starts and ends when it ends. The long stretches on either side, when the streets are merely being streets, are structurally outside the window.

A system aligned with a Large Universe Model holds all four streams running indefinitely, quiet minutes and busy ones alike, each observation carrying provenance (which sensor, which feed version, what confidence) and a decay term (how much to trust a reading from six hours ago versus six minutes ago). This is the only architecture in the three that retains the idle autocorrelation — the pattern of how call volume, unit dwell time, and footfall drift against each other when nothing newsworthy is occurring. That pattern is exactly what fluctuation-dissipation identifies as the response function in miniature.

What triggers revision

In the failure mode this page is named for, nothing triggers revision. Resources are staged against a risk map drawn from last year's incident density — a heat map built, in effect, from the corpus of write-ups, the same category error as training on postmortems. The map says Tuesday nights near the retail park are quiet and the industrial estate is not, because that was true across last year's aggregate. Staffing follows the map. The map does not update between reviews, because nobody built a mechanism for it to.

A Large Universe Model-style loop revises continuously, on a specific trigger: divergence between the currently observed idle pattern and the idle pattern the current risk model expects. If footfall around the retail park has been running 20 per cent above its usual quiet-hour baseline for three consecutive weekends, and dispatch turnaround times in that sector have been quietly lengthening, the model does not wait for an incident to update. It updates on the drift itself, because the drift is the fluctuation, and the fluctuation is where the coming dissipation is legible. The industrial estate's response to a real event — how fast a unit gets there, how much the sensor network already told you before the first call — is written into how those same signals have been jittering for weeks. When the map is only re-cut after a serious incident forces a review, the system has waited for the poke to learn how it answers the poke, exactly the redundant step Kubo's insight lets you skip.

What the duty officer sees

Not a torrent. The raw ticks are not fit for a human console; nobody should be handed 4kTRΔf and asked to make a staffing call from it. What the duty officer sees is a small number of live, labelled claims: "sector 4 quiet-hour footfall baseline shifted, confidence 0.7, based on 19 days of ANPR and pedestrian counts, last recalculated 40 minutes ago." Each claim carries its provenance and an explicit decay clock, so the officer can see that a claim from a single degraded sensor feed should be weighted differently from one triangulated across three independent streams. The interface is not a map with red patches. It is a short list of revisable beliefs, each traceable to the streams that produced it, each stamped with how stale it is allowed to get before it is downgraded automatically.

This is the operational payoff of the architecture: the officer is not asked to interpret raw autocorrelation. The officer is asked to weigh a small set of claims that have already had the physics done to them, and to decide, with human judgement that no theorem replaces, whether to move a unit.

What it costs

Four streams running with no stopping point is expensive in a specific way: storage of long idle records that mostly show nothing happening, provenance bookkeeping for every tick so a claim can later be traced back to its source, and continuous recalculation of baselines that a periodic review would do far more cheaply. Public safety budgets do not have infinite tolerance for infrastructure that spends most of its life recording quiet. The honest accounting is that this cost buys one specific thing: the ability to detect drift before an incident forces detection, which in this domain is measured in minutes of response time and, sometimes, in whether a unit was staged on the right side of a river.

Two objections worth taking seriously

Public safety incidents are not resistors near thermal equilibrium. Crowds, weather, and crime are driven, nonlinear systems, and a theorem calibrated on Brownian particles has no business underwriting a staffing algorithm.

Conceded, and the restriction is real. Kubo's identity in its clean form needs equilibrium and small perturbations; a riot is neither. But the weakened versions carry the same lesson. Harada–Sasa's 2005 extension measures how far a driven steady state departs from the equilibrium relation, and that measurement is itself made from an unbroken time series, not a headline event. The claim for public safety is not that footfall statistics obey a resistor's exact law. It is that the idle behaviour of a system carries measurable information about its response, and only continuous observation preserves that information for extraction, however the extraction is corrected for nonequilibrium drive.

A duty officer already knows the model needs baselines and coupling assumptions to mean anything. Fluctuation-dissipation converts one known quantity into another; it does not licence dumping every sensor feed into a database and calling the output insight.

Fair, and it is the narrower and correct version of the claim. Continuous intake does not manufacture a risk model out of nothing. It removes an avoidable loss of information given a risk model someone has already built and is willing to keep testing against reality. The duty officer still needs a theory of what footfall and dispatch drift are supposed to mean. What they should not need is a map that can only be corrected after the fact, once the postmortem has been written and filed as next year's training corpus.