What Is Stochastic Resonance? Definitions, Misconceptions, Debates, and Its Relevance to Biology

For most of scientific history, noise was the enemy. Engineers spent careers eliminating it. Physicists built increasingly shielded instruments to escape it. The working assumption was simple and seemingly obvious: randomness corrupts signals, and the less of it you have, the better your system performs. Then, in nineteen eighty, a climate scientist named Roberto Benzi proposed something that broke that assumption entirely. It turns out the brain may have been exploiting the same trick for millions of years. The phenomenon is called stochastic resonance, and the core claim is genuinely strange: under the right conditions, adding random noise to a system can improve its ability to detect or transmit a signal. Not degrade it. Improve it. Mark McDonnell and Derek Abbott, in their comprehensive review of the field, define it practically as any situation where randomness produces a measurable noise benefit — where a system's output better represents an input signal at some nonzero level of noise than it does with no noise at all. Two ingredients are required. First, the system must be nonlinear. McDonnell and Abbott are emphatic on this point: noise cannot be beneficial in a linear system, because only nonlinear interactions between signal and fluctuation can produce the effect. Second, the system must be operating suboptimally in some sense — the signal on its own is too weak to trigger a response, and the random kicks from noise can, in combination with the system's nonlinear rules, push it into detecting or transmitting something it otherwise would have missed. The empirical signature is a curve with a single peak: plot a performance metric against noise intensity, and you find that performance is poor at very low noise, rises to an optimum at some intermediate level, then falls again as noise overwhelms everything. That inverted U is stochastic resonance. Benzi first proposed this mechanism to explain the periodicity of Earth's ice ages — a nonlinear, bistable climate system, a weak periodic forcing from orbital cycles, and atmospheric noise combining to produce the glacial rhythm we see in the geological record. The idea spread rapidly. By the time McDonnell and Abbott wrote their review, stochastic resonance had appeared in more than two thousand three hundred publications, and roughly twenty percent of those papers explicitly referenced neurons or neural systems. That's not a minor footnote. That's a field within a field. The classic performance metric is signal-to-noise ratio, or SNR — the ratio of output power at the signal frequency to the output noise floor. As interest expanded beyond simple periodic inputs, researchers adopted mutual information for aperiodic signals and Fisher information for estimation tasks, where the question is how precisely you can infer signal parameters from noisy output. White Gaussian noise — constant power across frequencies — is the most common model in these studies, though McDonnell and Abbott note that changing the noise distribution often does not eliminate the effect. The phenomenon is surprisingly robust across those variations. This brings us to the most contentious question in the field: what exactly counts as stochastic resonance? This might sound like a semantic dispute, but McDonnell and Abbott argue it's anything but. The answer determines whether the brain can be said to use the effect at all. The narrow definition, historically the original one, treats stochastic resonance as a genuine resonance phenomenon: a time-scale matching effect in bistable dynamical systems driven by periodic inputs, where the SNR curve peaks at a single nonzero noise level. Under this strict reading, stochastic resonance cannot produce information-theoretic gains, requires the signal to be weaker than noise, and is distinct from a technique engineers already use called dithering — the deliberate addition of a random signal before quantization to reduce harmonic distortion. The broader definition, which McDonnell and Abbott advocate, treats stochastic resonance as any case where noise measurably improves signal processing. Noise benefit, full stop. Under this reading, dithering is not separate from stochastic resonance — it's an example of it. SNR gains consistent with information theory become possible. And crucially, a variant called suprathreshold stochastic resonance, or SSR, comes into view. SSR involves parallel populations of neurons collectively encoding a stimulus that's already above threshold. Even there, noise can help, because the truly optimal configuration for such a population is, as McDonnell and Abbott put it, "extremely complex and not plausibly achievable by real neurons." When the system can't be perfectly tuned, randomness picks up the slack. So, does the brain actually exploit any of this? This is where the field has been frustratingly cautious. The biological evidence exists — and it's substantial. In nineteen ninety-three, researchers demonstrated stochastic resonance in crayfish mechanoreceptors, the sensory cells that detect water movement. Subsequent experiments showed it in the cercal sensory system of crickets and in the human proprioceptive system, which tells your body where your limbs are in space. Neurons are nonlinear dynamical systems, and these experiments showed they are fully capable of stochastic resonance when signal and noise are supplied externally. But McDonnell and Abbott draw a careful line. Those experiments prove neurons can exhibit stochastic resonance in the lab. They do not prove neurons use endogenous noise — the noise they generate internally — as part of the neural code in living organisms. Demonstrating intrinsic stochastic resonance would require something much harder: removing naturally occurring internal variability and showing that neural function is specifically impaired by its absence. That experiment has not been done. And that gap, between what's been demonstrated externally and what happens in vivo, is where much of the neuroscience community has stopped short. McDonnell and Abbott find this caution puzzling, and they say so directly. Neural tissue is extraordinarily noisy. Ion channels open and close stochastically. Synaptic transmission is unreliable. The brain is not a clean electronic device — it's a wet, variable, constantly fluctuating system. Given that, the authors argue the asymmetry cuts the other way: it would be more surprising if evolution had not found ways to use that randomness than if it had. The biomedical applications support this view indirectly. James Collins — work important enough that it contributed to his receiving a MacArthur Fellowship in October two thousand three — showed that electrically generated subthreshold stimuli could improve human balance control and somatosensation. For cochlear implants, Morse and Evans proposed in nineteen ninety-six that reintroducing controlled randomness into the electrical output could restore aspects of natural auditory-nerve variability that are absent in deafened ears — variability that, in healthy hearing, may help auditory fibers encode more information about sound than a clean, noise-free signal would allow. Mechanical ventilators have been modified to include random noise to more closely mimic natural breathing patterns, and those modifications improved performance in ways that were later interpreted as a form of stochastic resonance. None of these are laboratory curiosities. They are working devices, informed by the idea that noise can be a feature rather than a bug. McDonnell and Abbott close with a challenge to the field, and it's worth taking seriously. They offer six recommendations for biologists: don't treat signal-to-noise ratio as the only valid performance metric; design experiments that can manipulate naturally occurring variability to test for intrinsic stochastic resonance; use ecologically relevant stimuli rather than just periodic sine waves; recognize that noise benefits can appear even for suprathreshold signals and population codes; and when benefits are found, look for evolutionary or engineering constraints that make the non-noisy alternative infeasible. The broader point is that adopting a narrow definition of stochastic resonance artificially shrinks the question. If you only call it stochastic resonance when you see a specific curve shape in a bistable system with a periodic input, you will miss most of the ways noise might be doing useful work in biology. The brain is not fighting noise. It exists in noise, was shaped by noise, and runs on tissue that generates noise at every scale. The question McDonnell and Abbott are really asking is whether we've been looking at that noise the wrong way — treating as a problem what evolution may have long since turned into a solution. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.