011 / RANDOMNESS
Noise that helps
Noise usually gets in the way. But when a signal is too weak to cross a threshold, a little noise helps to notice it. Too much noise hides everything again. Turn the noise yourself: on a photo and on balls in two wells.
Theory behind the experimentWhy noise helps to notice a weak signal, where that help peaks, and what we checked ourselvesPlain explanation, equations and how the model works
In plain words
Imagine a sensor that only answers yes or no: is the signal above a threshold or below it. If the signal is weak and always below, the sensor stays silent and tells you nothing about it.
Now add noise. It randomly pushes the signal up and down, and sometimes the sum crosses the threshold. Where the signal is stronger, the threshold is closer and the sensor fires more often. So the signal can be read again from how often it fires. On a photo this is how the bright parts emerge: a forehead, grey hair, a white collar.
With too much noise almost everything fires, and the difference between bright and dark drowns. The result is a hill: no noise is bad, strong noise is bad, and somewhere in between is the best value. This is called stochastic resonance.
The two wells work the same way, only the threshold is the hump between them. The wells tilt slowly, but weakly: the tilt alone cannot roll a ball over. Noise gives the balls random kicks, and climbing over is easier towards the side that is lower right now. It works best when a ball waits about half a swing on average before it hops: then the hops fall into step.
Note: noise does not make the signal itself stronger. It helps a system with a threshold to notice it. In an ordinary linear device without a threshold noise only gets in the way.
Try it yourself
- On the Photo in noise tab start without noise and slowly move the slider right. The picture appears and then drowns in grain again.
- Open More experiments and press As the eye sees it. The flicker is averaged and at the best noise the picture becomes almost clean.
- Raise the signal brightness above 100 %. The bright parts show without noise and the hill on the graph disappears: now noise only spoils things.
- Upload your own photo. The best noise differs a little from picture to picture.
- Tap a bright spot of the picture, then a dark one. The One pixel panel shows how noise pushes the pixel across the threshold and why the bright spot fires more often.
- In the same place turn on No threshold. The visibility curve now only falls: without a threshold noise does not help at all.
- On the Two wells tab press Little, then Best, then A lot, and watch the counter of balls in the lower well.
- Raise the tilt above 0.54. The hump disappears, the balls roll over on their own and noise is no longer needed.
Physics and equations
Two wells. A particle in a viscous medium where inertia can be neglected (the overdamped limit of the Langevin equation):
Minima , hump height . The term tilts the potential. As long as the potential keeps two wells and without noise there are no hops. sets the noise strength; for thermal noise .
Without the tilt a ball hops over the hump rarely and at random. The hop rate is given by Kramers' formula:
Resonance appears when the mean waiting time is comparable to half the period of the tilt:
This is an estimate, not the exact position of the peak. Different measures of the response peak at different noise. For instance, the signal-to-noise ratio in the two-state theory of McNamara and Wiesenfeld
peaks at regardless of frequency, while the synchrony of the hops with the tilt does depend on frequency.
Threshold. A pixel of brightness lights up when , where is normal noise. The firing probability is
As every pixel below the threshold goes dark; for large the probability tends to 1/2 for all of them. The link between the output and the original picture peaks at intermediate noise.
How the model works
Two wells: , period , step , Euler–Maruyama method, 200 balls. A ball changes wells only when it crosses : this removes false hops from jitter at the top of the hump. The counter shows the share of balls in the lower well over the last two swings; without hops it is 50 %.
The noise generator is the same as in the Python version used for the video: a 32-bit hash and the Box–Muller transform. The site tests check that the noise matches Python bit for bit and the trajectories agree to . The Python version was checked separately: the Boltzmann distribution without tilt, the exact mean crossing time, convergence in the step size.
Photo: threshold , signal brightness is set by the slider. How visible is the correlation between the original picture and the sensor output smoothed over a 5×5 pixel window (and averaged over 8 frames in As the eye sees it mode). The curve comes from a formula; the tests compare it with direct simulation. This is a rough model of vision, not an exact model of the eye.
The portrait of Albert Einstein was taken by Ferdinand Schmutzer in 1921 and is in the public domain. We cropped it and converted it to greyscale.
Sources and further reading
- R. Benzi, A. Sutera, A. Vulpiani. The mechanism of stochastic resonance. J. Phys. A 14, L453 (1981). doi ↗
- B. McNamara, K. Wiesenfeld. Theory of stochastic resonance. Phys. Rev. A 39, 4854 (1989). doi ↗
- K. Wiesenfeld, F. Moss. Stochastic resonance and the benefits of noise: from ice ages to crayfish and SQUIDs. Nature 373, 33 (1995). doi ↗
- L. Gammaitoni, P. Hänggi, P. Jung, F. Marchesoni. Stochastic resonance. Rev. Mod. Phys. 70, 223 (1998). doi ↗
- J. K. Douglass, L. Wilkens, E. Pantazelou, F. Moss. Noise enhancement of information transfer in crayfish mechanoreceptors by stochastic resonance. Nature 365, 337 (1993). doi ↗
- D. F. Russell, L. A. Wilkens, F. Moss. Use of behavioural stochastic resonance by paddle fish for feeding. Nature 402, 291 (1999). doi ↗
- E. Simonotto et al. Visual perception of stochastic resonance. Phys. Rev. Lett. 78, 1186 (1997). doi ↗
- A. A. Priplata et al. Vibrating insoles and balance control in elderly people. Lancet 362, 1123 (2003). doi ↗