The escape law, in plain English

The formula, and why it looked like trouble

There is a formula at the centre of this framework. It says that the chance a system breaks out of the state it is stuck in falls off exponentially in a single number. That number, χ, is a ratio: how much energy the environment pours in during one “settling time” of the system, divided by how much energy it would take to get out.

χ greater than one means the environment is delivering more than enough. χ less than one means it isn’t.

The trouble is that this formula looks exactly like one of the most famous equations in chemistry. The Arrhenius equation says a reaction rate falls off exponentially in the activation energy divided by the temperature. Line the two up and the delivered energy is sitting precisely where the temperature sits.

So the formula appears to be saying: throughflow is a kind of temperature.

That is a claim with an unhappy history. Physicists have spent decades trying to define an “effective temperature” for systems that are being driven rather than sitting in equilibrium — a stirred fluid, a shaken pile of sand, a living cell. It sometimes works. It often doesn’t, and the way it fails is instructive.

Temperature is supposed to be one number. If you measure the temperature of a room with a mercury thermometer and an infrared one, you expect the same answer. That is not a convention; it is what makes temperature a physical quantity rather than a summary statistic.

In driven systems this breaks. You can define an effective temperature from how much the particles jiggle in position, or from how fast they move, or from how often they escape a trap — three reasonable definitions, three different answers, same system. There is a 2026 experiment that measures exactly this: three independently motivated definitions agree beautifully over a wide range of conditions, and then, in identifiable regimes, come apart.

So a framework that says “χ is a temperature” has walked into an argument it cannot win, on ground it did not need to fight on.

The way out: the exponential was never a thermal idea

Here is the paper’s first move, and it is a reinterpretation rather than a revision. Not one character of the formula changes.

The exponential does not come from thermal physics. It comes from a much more general and much older result about rare events, developed by Freidlin and Wentzell in the theory of large deviations. The result says, roughly: if something unlikely happens, it almost certainly happened by the least unlikely route.

Picture a mountain range, and a hiker being blown across it by random gusts of wind. Whether they get to the other side depends on two things: the height of the lowest pass — not the average height, not the summit — and how strong the wind is. The probability of crossing falls off exponentially in the pass height measured in units of wind strength.

Notice what that argument never needed. It never needed the wind to be thermal. It never needed a temperature. It never needed the mountain range to be in equilibrium with anything. It needed only that the wind be weak compared to the pass — because if the wind were strong enough to blow you over the summit routinely, none of this would be interesting.

That is the whole of it. The exponential form is inherited from the mathematics of rare events, not from thermodynamics. On this reading, the delivered energy in χ isn’t a temperature — it’s the strength of the wind. A noise scale. And a noise scale carries none of the obligations a temperature carries. It doesn’t have to be the same measured two ways, because it isn’t claiming to be a state variable of anything.

The part I find most satisfying

There is a further twist that makes this more than a lateral move, and it is the argument I’d defend hardest.

An escape law is only worth writing down when escapes are rare. If systems hopped out of their configurations constantly you wouldn’t need a probability; you’d just watch. The formula earns its keep in the rare regime.

But “rare” and “the barrier is large compared to the noise” are the same statement. And “the barrier is large compared to the noise” is precisely the condition the derivation requires.

The regime in which the law is useful and the regime in which it is valid are the same regime. You get the assumption for free, from the fact that you needed the formula at all.

Nothing like that is available for the temperature reading. Temperature consistency has to be established separately, for every system, for every pair of measurements, every time. There is no configuration in which it comes free.

What this costs, and it is a real cost

The reinterpretation secures the shape of the law. It does not secure what goes inside the exponent, and that is where the content lives.

In the general theory the exponent is a “minimum action” — the cost of the cheapest escape route. In the framework it is written as a free-energy barrier. Those coincide in nice cases and diverge in others, and the paper’s second half is a catalogue of four ways they diverge. All four are drawn from published work; the contribution is noticing which ones bite here.

One — the temperature isn’t one number

Already covered. Under the new reading this stops being a failure of the law and becomes a reason not to use the word. But a residue survives that isn’t merely verbal: if the quantity governing escape differs from the quantity governing jiggling, then a number you calibrated by measuring flux may not be the number that belongs in the exponent. Calibration route and application route have to match.

Two — when the push lasts too long

This one has the best physical picture.

There are two different things you can do to get a marble out of a dip in a table. You can shake the table, or you can tilt it.

Shaking is noise. More shaking, more escape, and it works symmetrically — the marble is as likely to go left as right.

Tilting is not noise, and it is not “more shaking.” A tilt lowers the barrier on one side and raises it on the other. It is asymmetric, and no amount of turning up a shaking-intensity dial will ever reproduce that asymmetry.

The distinction between the two is a matter of timescale. If the push changes direction many times before the marble could plausibly escape, it’s shaking. If the push holds steady for longer than the escape takes, it’s a tilt. And the published work shows exactly this: an effective-temperature treatment works while the drive decorrelates quickly, and has to be abandoned for a modified-barrier treatment when it doesn’t.

Then it gets more interesting. Escape is fastest at some intermediate persistence. Not slowest. Not steadily increasing. There is a sweet spot, from a resonance between how long the push lasts and the system’s own natural timescale — the same reason a swing goes higher when you push in time rather than harder. This shows up across three different shapes of trap, at almost the same optimum.

And here is the sharpest consequence in the entire paper: a formula with a single number in it cannot produce a resonance. A resonance is a relationship between two timescales. One number cannot hold two timescales. So the existence of the resonance is, by itself, proof that something is missing from the formula.

Three — jumps are not jitter

Take a fixed amount of energy and deliver it two ways: as constant small jitter, or as occasional enormous kicks. Same total, same variance. The escape rates are not similar.

The reason is beautiful. Under jitter, escape happens by an improbable run of small pushes all lining up. Under kicks, the system doesn’t bother — it waits for one big one. The mathematics shows this as a discontinuity in the optimal escape path: a finite jump, rather than a smooth climb.

The size of the error from ignoring this is not a correction. In one measured case — a tracer particle in a suspension of swimming bacteria — the observed escape rate was about one in a million. A calculation based on the particle’s variance gave one in 10³². That’s off by twenty-six orders of magnitude. It is the difference between predicting an event once in the lifetime of the universe and observing it happen every few seconds.

Four — two numbers aren’t enough to describe a push

This is the hardest one, because it doesn’t tell you where the safe ground is.

You can build three physically different drives — one intermittent source, five independent ones, a smooth random one — and tune them so they have identical variance and identical persistence time. They produce different escape rates, different distributions, different everything.

Which means: those two numbers do not determine the escape. Not approximately — at all. And no refinement of either one fixes it.

The reply the framework can make is a narrowing, and the paper makes it explicitly. χ needn’t claim to give you an escape rate. It can claim to tell you whether escape is affordable, and to rank systems by that. The first claim dies; the second survives. What isn’t allowed is using the ranking claim in argument and the rate claim in calculation.

What the paper actually adds

Two things, and both are small on purpose.

A missing parameter. The “settling time” inside χ is a property of the system — how fast it recovers when you nudge it. But every regime boundary above is set by the persistence of the drive — how long the push lasts before it changes. Those are different quantities, and no combination of what’s already in χ produces the second from the first.

So the law as written can’t say which of two qualitatively different descriptions it’s asserting. The fix is a second number: ψ, the drive’s persistence divided by the system’s settling time.

χ tells you whether escape is affordable. ψ tells you which law applies.

The analogy I’d use: knowing how much money you have doesn’t tell you whether you’re standing in a shop or at an auction. Same money, completely different rules about what it gets you.

A disagreement the framework didn’t know it had. The framework contains two statements that quietly conflict. One says χ = 1 is a clean line: above it systems escape, below it they can’t. The other — the exponential — says that at χ = 1 the chance of escape is 37%. Not certainty.

Neither is wrong. They’re the noiseless and the noisy versions of the same threshold, and the sharp one is the limit of the soft one. But they can’t both be exact, and results proved with the sharp version now have to carry that label. It is the sort of thing that only shows up when you write the formalism out carefully and hold the two halves next to each other, which is the argument for having written the note at all.

Where this connects

The paper deliberately claims nothing outside itself. But the structure it is about shows up in a lot of places, and the same four failure modes will apply wherever it does.

Extinction and evolution. The barrier is how deeply a lineage is committed to its way of making a living. Specialists sit in deep basins, generalists in shallow ones. Push everything at once and the specialists die where they stand while the generalists move — which is what the fossil record shows, and it falls out of the law rather than being assumed. The jumps-versus-jitter distinction then has real teeth: an asteroid and a long climate excursion are not the same driver scaled differently. They are different kinds of driver, and the mathematics says the jump-like one can produce escapes the gradual one never would at identical total energy.

Machine learning. Deep networks have a genuine order/chaos phase transition governed by how their weights are initialised: too ordered and every input collapses to the same representation, too chaotic and nearby inputs decorrelate and nothing propagates. Trainability of very deep networks is bounded by how close you sit to the critical line. Learning-rate schedules are annealing schedules and the failure mode has a name — decay too fast and you’re stuck in a bad basin. Networks in continual learning lose the ability to change at all, and the fix is injected noise, which is reheating. Critical learning periods, where a network becomes permanently unable to learn something it missed early, are canalisation measured in silico.

Closer to home: in my own sweeps of a repetition-penalty parameter, one model showed a clean sharp step and two others showed non-monotonic behaviour with a resonant optimum. That is exactly the two-regime structure in §5.2. I had the data before I had the reason.

Brains. Cortical activity sits near a critical point where avalanches of firing follow a power law. Below it, signals die out. Above it is a seizure. The tuning parameter tracks arousal — closest to critical at intermediate alertness, drifting ordered when drowsy and chaotic when overwrought. Same three-regime shape, with a pathology at each end.

Metallurgy, and a warning about method. Cool a metal slowly and you anneal out the accidents, leaving something close to optimal. Cool it fast and you freeze in whatever configuration happened to be there — which is what makes quenched steel hard. This matters for anyone reading a genome as a record of frozen optima: that reading is only valid under slow recovery. After a fast one, conserved features are frozen accidents, and solving backwards returns a confident wrong answer.

---