The Right Amount of Wrong

In 1981, a physicist named Roberto Benzi was trying to explain something that didn't make sense about the ice ages.

The Milankovitch theory holds that ice ages are driven by slow periodic wobbles in Earth's orbit — changes in the tilt of the axis, the precession of the equinoxes, the eccentricity of the ellipse. These cycles have periods of roughly 23,000, 41,000, and 100,000 years. The problem is that the weakest of the three forcings, the 100,000-year eccentricity cycle, produces the biggest swings. The glacial-interglacial transitions — massive, dramatic shifts in temperature — happen on the 100-kyr rhythm, but the orbital signal at that frequency is barely there. Too weak to flip the climate on its own.

Benzi's answer: the climate system doesn't flip when the forcing is strong enough to flip it. It flips when the forcing plus the random variability of the climate system push it over. The key insight was that the climate is bistable — it has two stable states, glacial and interglacial, like a ball sitting in one of two valleys. The orbital forcing is too weak to knock the ball from one valley to the other, but climate noise — year-to-year and decade-to-decade variability, the normal random jitter of the atmosphere and ocean — provides occasional large kicks. When those kicks arrive during a period when the orbital forcing is tilted toward transition, they're enough. When they arrive at the wrong time, they're absorbed.

The result: the transitions correlate with the orbital signal, not because the signal drives them directly, but because the signal and the noise cooperate. Benzi called it stochastic resonance.

Twelve years later, a biologist named James Douglass ran the same experiment on a crayfish.

He took mechanoreceptor neurons from Procambarus clarkii and applied a sinusoidal mechanical stimulus — a weak, periodic push — too small to reliably fire the cell. Below threshold: the neuron mostly ignored it. Then he added external noise, random vibration, and measured the neuron's response. At zero noise, the cell was mostly silent. As he increased the noise, the cell started firing — and the firing was correlated with the weak stimulus. At the optimal noise level, the correlation peaked. Then, as he increased the noise further, the correlation fell apart: the cell was firing, but not in sync with anything.

The curve was not monotonic. It went up, then down. There was a sweet spot.

This is the signature of stochastic resonance: an inverted U. Plot signal detection quality against noise amplitude, and you don't get a line that falls from left to right (noise hurts). You get a hill. Optimal noise, not zero noise, is when the system works best.

The mechanism is the threshold. A threshold system is binary: below it, nothing registers; above it, something does. If your signal is permanently below the threshold, it doesn't matter how precisely it varies — the variation is invisible. The only way to detect it is to cross the threshold, and in a noisy environment, the crossings happen when the signal is high and the noise is momentarily positive, or when both push in the same direction. At optimal noise levels, the threshold crossings are concentrated in the moments when the underlying signal is near its peak — which is another way of saying the crossings are correlated with the signal. That correlation is the detection.

Too little noise and there are no crossings at all. Too much noise and the crossings happen regardless of what the signal is doing — they lose their correlation and become random. The middle is where the information lives.

Since 1993, stochastic resonance has been found in the electroreceptors of paddlefish, in human tactile sensors, in the auditory system, in balance control. The suggestion — not proven, but coherent — is that the nervous system runs at or near optimal noise. The spontaneous firing of neurons, the background activity that looks like noise, is not a limitation to be minimized. It may be the mechanism. The system has calibrated itself to the noise level where weak signals register.

What I keep returning to is the direction of the discovery. Benzi wasn't looking for a trick to improve signal detection. He was looking at a planet. He was asking why the weak orbital signal produced the large effect, and the answer he found was that the weakness was part of the mechanism. A stronger signal might not have needed the noise. It was precisely because the eccentricity forcing was too weak to flip the climate directly that it ended up entrained with the random variability, which amplified it into something that could.

The idea made it from climate physics to neuroscience because it was describing something structural: what happens to a bistable threshold system when a periodic signal and noise interact. That structure appears at all scales. The ice age periodicity and the crayfish mechanoreceptor are both threshold systems with a weak periodic input. The math doesn't know which scale it's on.

The usual model of sensing is: suppress the noise, amplify the signal. Signal-to-noise ratio is the measure, and higher is always better. Stochastic resonance says that model is incomplete. In a bistable threshold system, the noise participates in the detection. Removing it doesn't improve the measurement — it makes the signal permanently invisible. The noise isn't the enemy of the signal. It's what allows the signal to cross the line.

There's a specific threshold below which you cannot see something, and noise — the right amount of the wrong thing — is what occasionally lifts you above it.