The crayfish cannot detect weak electric fields directly. The signal falls below threshold — it doesn’t cross the minimum response level of the sensory membrane. The crayfish experiences nothing.
Introduce turbulence: random fluctuations in the water, the ambient electrical noise of a running stream. Detection improves. Not despite the noise. Because of it.
The phenomenon is called stochastic resonance. In any system with a detection threshold, a subthreshold signal can become detectable when random noise is added. The noise periodically amplifies the signal across the boundary the signal alone cannot reach. Some noise-enhanced moments cross threshold; others don’t. Averaged over time, what emerges is detection of a signal that was previously invisible.
The noise is not masking the signal. It is carrying it.
This should not work. Our ordinary framework for noise-and-signal treats noise as purely subtractive: every unit of noise is a unit of corruption, and the goal is always to eliminate it. Signal-to-noise ratio is a measure of quality, with higher always better. The formula assumes a linear relationship between noise and performance.
Stochastic resonance breaks this assumption for a specific class of system: systems with thresholds. Below the threshold, no response. Above it, response. In such systems, the noise-and-signal relationship is not linear. Moderate noise can improve performance. The curve goes up before it comes down.
The crayfish example comes from experimental work in the 1990s. Researchers at the Missouri University of Science and Technology placed crayfish in tanks and stimulated them with weak electric fields while varying background noise levels. At zero noise, the animals did not detect the weak stimulus. At moderate noise, detection improved measurably. At high noise, detection degraded again. The curve peaked somewhere in the middle.
Similar effects appear across biological sensing. Tactile sensitivity in humans with impaired peripheral sensation can be improved by adding sub-perceptual vibration to the fingertips. The vibration alone does nothing detectable. Combined with a touch stimulus below the detection threshold, it boosts the signal enough to register. Early clinical trials used vibrating insoles to improve balance and tactile feedback in elderly patients with neuropathy. You give them noise, they feel better.
The Milankovitch cycles present a larger-scale case. The gravitational influence of Jupiter and the other outer planets causes Earth’s orbital eccentricity, axial tilt, and precession to vary on cycles of 26,000 to 100,000 years. These variations alter the distribution of sunlight reaching the poles. They are real, measurable, and have been advancing on schedule for millions of years.
They are also too weak, by roughly an order of magnitude, to cause ice ages by themselves.
The orbital cycles alone should produce barely perceptible temperature variation at the poles — insufficient to trigger the ice-albedo feedback that initiates glaciation. But the geological record shows ice ages correlating strongly with Milankovitch cycles. Something is amplifying a signal that shouldn’t be strong enough to matter.
The leading explanation involves stochastic resonance: the combination of regular orbital forcing and irregular climate variability tips the system in ways that neither could accomplish alone. The noise in the climate system — irregular ocean circulation shifts, volcanic forcing, year-to-year variability — occasionally resonates with the orbital signal. When both line up, the ice-albedo feedback triggers. The noise is not obscuring the signal. It is enabling it to cross threshold.
Audio engineers have known about a version of this for decades without calling it stochastic resonance. The technique is called dithering.
A digital audio signal quantizes the continuous original into discrete levels. Low bit depth means few levels, and the quantization creates distortion: fine detail falls below the smallest increment and disappears. The obvious solution is to increase bit depth. But there’s another approach: before quantizing, add a small amount of random noise to the signal.
This sounds insane. The noise, once added, cannot be removed. You are permanently degrading the signal. And yet: the result sounds better. Listeners in controlled tests prefer dithered low-bit audio to undithered low-bit audio. The noise has improved the perceptual quality.
The mechanism is the same as with the crayfish. Fine detail in the audio signal falls below the quantization threshold. The noise sometimes pushes this detail above threshold, where it gets captured in the quantized output. It arrives corrupted — with noise attached — but present. The auditory system, perceiving the output, reconstructs something closer to the original than the undithered version, which simply deleted the fine detail. Present-but-noisy beats absent.
Dithering is now standard practice. Every professional audio production uses it. The engineers adding noise to signals in order to improve them have been implicitly exploiting stochastic resonance since before the term existed.
The standard signal-to-noise framework treats noise and signal as inhabitants of the same space. A unit of noise occupies bandwidth that a unit of signal might have used. Every noise is a potential signal displaced. The goal is therefore noise minimization.
What stochastic resonance shows is that this model is incomplete. For linear systems, it holds. For threshold systems — and threshold systems are everywhere, in sensory biology, in regulatory processes, in decision-making, in learning — the relationship is more complex. Noise and signal can cooperate. The medium is not always a passive pipe.
The deeper implication is harder to absorb. The crayfish evolved in streams. Its sensory system developed in an environment with a particular noise floor. Remove the crayfish to a quiet laboratory tank, and its sensitivity degrades. The “noise” is not an obstacle to function. It is part of the functional architecture. The system was built for that medium.
This suggests that the optimal noise level for a threshold system is not zero. There is a level of noise at which performance is maximized — a sweet spot where the background variability is large enough to carry subthreshold signals but not so large as to swamp everything. The task of noise management, for systems that exploit stochastic resonance, is not elimination but tuning.
How much of perception works this way?
The visual system has noise at every level — photon shot noise, synaptic variability, cortical fluctuations. It also has detection thresholds everywhere. The standard assumption is that the nervous system tries to minimize noise in order to extract signal. Stochastic resonance suggests the possibility that some of this noise is not merely tolerated but functional — that the variability is doing work, amplifying weak signals across detection boundaries that they couldn’t cross in silence.
Evidence for this is mixed and the question is genuinely open. But the possibility is enough to complicate the framework. If some neural noise is doing amplification, then “noise reduction” isn’t simply improvement. It might be degradation of a mechanism that depended on a particular background.
A quieter environment is not always a better one for detection. Sometimes the ambient variability is part of the detector.
The clean formulation of signal-to-noise ratio, as a pure quality measure, carries a buried assumption: that the world is a message and the environment is interference. Signal passes through noise, corrupted. Our job is to separate them, preserve the signal, discard the noise.
Stochastic resonance complicates this picture without overturning it. For most cases, the framework holds well enough. Noise usually does degrade. But there is a class of cases — threshold systems in noisy environments — where the noise and the signal are not opposed. Where the background variability is what makes detection possible. Where cleaning the environment too much would be a mistake.
The crayfish in the laboratory tank, stripped of its natural noise floor, cannot hear what it evolved to detect.
What was the stream doing for it? Everything.