Ergodicity and time averages in forestry and wildfire

Ergodicity and time averages in forestry and wildfire

The strongest case against this whole framework is a case forestry practitioners already believe, because it is mostly true. Fuel physics does not drift. The moisture content at which a given pine litter ignites, the relationship between fine dead fuel and flame length, the way relative humidity governs desiccation rate — these are stable across decades and geographies. A fire behaviour analyst using equations calibrated in the 1970s Rothermel model is not fighting stale statistics. The chemistry of combustion has not moved. If a Large Language Model trained on forty years of fire science can reproduce that structure faithfully, then the entire complaint about frozen corpora looks overstated for the part of the domain that carries the most weight. An incident commander does not need a live feed to know that fuel below 8% moisture carries fire readily. That fact was true in 1988 at Yellowstone and it is true this afternoon. Calling this an "ensemble average frozen at a cutoff" makes it sound fragile when it is the opposite: the one part of the system immune to the calendar.

That objection deserves to be taken seriously before any counter-argument, because it is where a great deal of operational fire science actually lives, and dismissing it would misrepresent how much of the domain is genuinely stationary.

What actually moves

The trouble is that a wildfire incident is not run on combustion chemistry alone. It is run on the joint state of fuel-moisture sensors scattered across a landscape, satellite thermal passes arriving on a delay, wind models updated on their own cadence, and the real-time positions of crews moving through terrain that the map has not caught up with. None of those four streams is stationary over the timescale that matters to an incident commander, which is measured in minutes, not decades. Fuel moisture at a given sensor can swing several percentage points across a single afternoon as a frontal boundary passes. A satellite thermal pass gives an excellent picture of where the fire was at the moment of overpass — for a polar-orbiting sensor that can mean a gap of many hours between usable reads over the same ground. Wind direction near a fire is notoriously local, driven by the fire's own convection column as much as by synoptic forecasts, and a model tuned on yesterday's regime can be actively wrong about the plume-dominated windfield forming right now.

This is the part of the objection that survives only partially. The combustion physics is ergodic enough to trust as background structure. The situation the physics is operating on is not. A corpus of fire science, however well trained, describes the ensemble of fires as they have historically behaved under given conditions. It cannot tell an incident commander what the wind is doing on this ridge in the next twenty minutes, because that is not a property of the ensemble. It is a single trajectory, running once, right now, and the average over many past fires is not the average along this one.

The failure this produces

The characteristic failure in this domain is precise and recurs across after-action reports: an ignition — a spot fire, a slop-over, a new start from a firebrand carried ahead of the front — is detected after the wind has already shifted, not before. The detection systems are not broken. The satellite pass fires on schedule. The fuel-moisture sensor logs its reading. The wind model updates its forecast. What fails is the assumption, buried in how those readings get used, that the last known state is still the current state. A wind shift that changes fire behaviour from a backing fire to a head fire can happen in under fifteen minutes. A thermal satellite revisit cycle measured in hours cannot see that shift happening; it can only report, after the fact, that something has grown where the previous pass showed nothing.

This is a non-ergodic system in the strict sense the physics gives the word. There is drift — the diurnal humidity cycle, the approaching frontal boundary. There is an absorbing state — once a spot fire crosses containment lines, the system does not return to its earlier configuration. There is memory — the fire's own convective column reshapes the local wind, so the process generating tomorrow's readings is not the process that generated last week's training data on "typical" fire-wind interaction in that fuel type. Boltzmann's gas in a sealed box explores its states evenly enough that watching one molecule for a long time tells you what a snapshot of the whole box would tell you. A fireground with a moving boundary and its own weather does not explore states evenly. It has a direction.

The instrumentation already updates constantly — sensors report every few minutes, satellites revisit on a fixed schedule, wind models refresh on the hour. Calling this a snapshot problem misdescribes a system that is already streaming.

This objection is close to correct and needs to be answered on its own terms rather than waved off. Frequent refresh is real progress over a single briefing document read at the start of a shift. But a stream of fresh readings, each one replacing the last, is not the same thing as a time average. If the fuel-moisture reading at 14:00 simply overwrites the 13:45 reading with no record of the trajectory between them, the system has a sequence of snapshots, not a running belief about how fast moisture is dropping and under what forcing. The distinction matters operationally: an incident commander who knows moisture has dropped three points in the last two hours under rising wind is holding a rate, a first derivative, evidence of acceleration toward critical fire behaviour. An incident commander looking at the current reading alone is holding a fact with no trajectory attached, and cannot distinguish a slow, containable drying trend from a rapid one already crossing thresholds.

What a persistent, dated belief buys

The position this argument arrives at is not "ignore the sensors, watch the fire yourself" — no crew survives that. Nor is it "add more sensors" — resolution is not the missing ingredient. It is that the useful object is a belief about each stream that persists between readings, updates with each new observation, and carries when that observation arrived and from which source. A fuel-moisture estimate that is "6.2%, last confirmed at 13:45 from the ridge RAWS station, trending down 0.4 points/hour under a frontal approach" is a time average with provenance. It says something a raw current reading cannot: how much to trust it an hour from now, and what would falsify it. A wind forecast note that says "modelled at 210° 12mph as of the 11:00 run, but plume-dominated behaviour observed on the last thermal pass suggests local override" is the same move applied to a different stream — the belief survives the individual observation and gets corrected by the next one, rather than being replaced wholesale each time.

Where the ensemble still earns its place

None of this argues for discarding the historical corpus of fire behaviour science. The stationary structure — how a given fuel type responds to a given moisture and wind combination — is exactly what lets an incident commander interpret an incoming sensor reading at all. Without it, a raw number from a RAWS station is noise. The corpus supplies the model of how the system behaves in general; the continuous, dated, revisable stream supplies where this particular fire actually is on that model right now. Mistaking the general model for a substitute for the running observation is the error that produces late-detected ignitions. Mistaking the running observation for something that needs no general model behind it produces sensor readings nobody knows how to interpret.

The narrower claim

Fuel chemistry does not need a Large Universe Model; it is ergodic enough that a well-trained corpus serves it well, and pretending otherwise overstates the case. But the joint state of moisture, wind, fire position and crew position on an active fireground is not exploring its states evenly — it has an absorbing state called a burned area, a plume that manufactures its own weather, and a boundary that moves faster than the revisit cycle of the instruments watching it. The only intake structure suited to that condition is one that keeps taking the time average without a stopping point: every stream still running, each belief carrying the date and source of its last correction. That is not a claim that ergodicity is dead everywhere in forestry. It is a claim that the fireground is precisely where it was never alive, and that the incident commander's late-detected ignition is what non-ergodicity looks like when nobody kept a running record of the drift.