Thermal noise and the fluctuation-dissipation theorem in sports analytics

The hiss in the wire

In 1928, a Bell Telephone Laboratories engineer named John Johnson was trying to work out why vacuum-tube amplifiers had a noise floor no amount of careful construction could remove. Every resistor, left alone, sitting in a circuit with no current driven through it deliberately, produced a small random voltage across its terminals. It was not a flaw in manufacturing. It was thermal: the electrons inside the resistor jittering from heat, summing to a fluctuating signal that no shielding could eliminate. Harry Nyquist gave the size of it that same year — mean square voltage equals 4kTRΔf, where R is the resistance, T the temperature, k Boltzmann's constant. The striking part was which quantity showed up in the formula. Resistance is what dissipates energy when you push current through a circuit. It is also, exactly, what sets the size of the circuit's noise when you push nothing through it at all. The same number governs both.

Ryogo Kubo generalised this in 1957 into a statement with much wider reach: for a system near equilibrium, its linear response to a small applied force is fixed by the autocorrelation of its own spontaneous fluctuations. You do not need to perturb a system to know how it will respond to a perturbation. You need to watch it sit still, for long enough, and record the quiet honestly. The quiet is not empty. It is the calibration.

The same failure, on a different pitch

A performance analyst building a game plan works from three kinds of intake: tracking data on player and ball position, injury reports, transfer activity, and a file of opponent tendencies compiled from recent matches. The tendencies file is the load-bearing part. It says: this fullback overlaps on 68% of attacking sequences down the left, this striker drops deep against a low block, this team presses higher after conceding first. The plan is built to exploit the tendency.

The characteristic failure is depressingly familiar to anyone who has done this job: the plan is built on a tendency the opponent abandoned last month. The fullback stopped overlapping after an injury to the winger ahead of him. The coaching staff changed the press trigger three fixtures ago, in a match nobody flagged for review because it was a routine 1–0 and nothing "happened." The tendency was real. It was also stale, and the analyst had no way of knowing it, because the record that would have shown the change was never kept in a form anyone looked at.

This is not a failure of effort. It is a failure of what got recorded and when.

Three ways of watching a team

Consider the intake regimes available to analytics departments in turn, because the failure recurs in a structured way across them, and the structure is the argument.

A corpus-based approach — call it the Large Language Model of scouting — works from what has already been written up: match reports, post-match statistical summaries, transfer press releases, the tendencies file itself once compiled. This is perturbation already interpreted. Someone decided the fullback's overlap rate was worth recording as a headline number, at a moment when it was true. Nobody recorded the week the overlap rate began to fall, because that week produced no notable event, no goal, no headline. The corpus is a photograph album of moments editors thought mattered. It cannot show you decay it was never pointed at.

A scene-based approach — the Large World Model of scouting — improves on this by sensing directly: full tracking data, ingested live, covering the ninety minutes as they happen. This captures the perturbation in real time, which is a genuine gain over reading someone's summary of it. But it is still triggered by the match. It watches while something is happening. It does not watch the training ground in the weeks between matches, where the fullback's overlap habit was actually extinguished, quietly, without an audience, without a camera pointed at the right patch of grass. The interesting moment gets full resolution. The idle stretch either gets none, or gets discarded because nothing "happened" in it.

A universe-scale approach keeps every stream running with no stopping point: tracking data, yes, but also injury reports the moment they are filed, transfer chatter as it circulates, training-ground observations, minutes played in reserve fixtures, the long unremarkable weeks — all timestamped, all carrying provenance, all held as beliefs that can be revised and decayed rather than facts fixed at compilation time. Nothing about this regime requires an event to justify recording. The recording continues because the theorem says the quiet interval is where the answer to "would this tendency survive a change?" actually lives.

What the quiet actually contains

Nyquist's result says the resistor's idle jitter and its driven response are two readings of one microscopic mechanism. The sports analogue is not identical, but the structural claim transfers cleanly enough to be useful, and the honest version of the claim is narrower than it first sounds. A tendency — the fullback's overlap rate, the press trigger, the striker's drop-off distance — is a description of how a team responds to circumstance. Its autocorrelation over the quiet weeks, the fixtures where nothing dramatic happened, the training sessions nobody clips for a highlight reel, carries information about whether that response is stable or already eroding. A team that has quietly stopped rehearsing a pattern in low-stakes fixtures is a team whose "tendency" is a fossil, however cleanly it shows up in last month's headline statistics.

Event-triggered intake — watching only the match, or worse, reading only the report of the match — throws away exactly this. It captures the tendency at the moment it was notable and assumes it persists. Continuous intake, timestamped and decaying by design, is the only regime built to notice the tendency softening before it disappears from the highlight reel and reappears, absent, on matchday.

Where this overreaches, and where it holds

Football is not a resistor in thermal equilibrium. Teams are driven systems full of intention, coaching intervention, and deliberate misdirection. Applying a theorem built for Brownian particles to dressing rooms is analogy dressed as physics.

This is fair and should be conceded without hedging. Kubo's identity in its classical form needs equilibrium, small perturbations, and linear response. A team preparing a specific tactical ambush for one opponent is about as far from equilibrium as a system gets. The exact theorem does not port over.

But the weaker claim survives the concession, and it is the one actually being made. Harada and Sasa's 2005 extension of fluctuation-dissipation to nonequilibrium steady states shows that the size of the violation of the equilibrium relation is itself measurable from an unbroken time series — the departure from the clean formula is informative, not just the formula's success. Nonequilibrium work relations from Jarzynski and Crooks extract real quantities from fluctuation records in systems nowhere near equilibrium. The general lesson: noise carries information about response even when the neat identity fails outright, and you can only extract that information from continuous, unbroken observation. A team's deviation from its "typical" pressing pattern in low-stakes matches is itself a signal, whether the deviation is quiet drift or deliberate disguise. You cannot measure a deviation from a baseline you never recorded.

Watching a team doesn't tell you what happens if you press them differently. That's an intervention, and Pearl's hierarchy is clear that observation alone cannot resolve it. Citing thermodynamics to justify passive scouting is smuggling in physics that doesn't apply to opponents who know they're being watched.

Also fair, and the correct response is to accept the limit rather than argue around it. Nothing here claims that continuous tracking substitutes for the actual match — the intervention, the real test of whether the tendency holds under pressure. The claim is narrower: whatever inference is drawn from that match, whatever causal story the coaching staff builds from it, will be built on top of a baseline, and a truncated baseline cannot be repaired after the fact. If the training-ground weeks were never recorded, no amount of clever modelling recovers what happened in them. Continuous intake does not replace the intervention. It is the substrate the intervention gets interpreted against, and it is either there or it is permanently missing.

Why this is the last rung, not a bigger one

The lineage from corpus to scene to continuous multi-stream intake is not simply "more data each time." Each step removes a specific, nameable category of blindness. The corpus misses everything not written up. The scene misses everything outside the interesting moment. Once intake covers every stream, continuously, with provenance and decay attached so that stale beliefs are marked as stale rather than quietly retained as current — there is no further category of missing observation left to name. What remains after that is not a new kind of watching. It is more of it, over longer periods, with better-trusted provenance. That is why the third position is terminal on this particular axis: not because the analyst's job becomes easy, but because fluctuation-dissipation specifies, precisely, what the quiet interval was worth all along, and continuous intake is the only regime that keeps it.