Noise is the ever-present source of headaches on many engineering designs, and often is one of the biggest underlying challenges a designer team faces. (See: Noise annoys, but do we secretly love it?). That's why I cringe when I hear an engineer casually remark, “Assuming that the noise is random and Gaussian (white)…” in any analysis or discussion.
There are two reasons for my concern. First, just because noise is random doesn't mean it is Gaussian. Second, assuming that it is Gaussian is actually a pretty big assumption to make. Lots of noise in the real world just isn't nice like that.
To cite a few examples, there's pink noise, impulse noise (of many types), Rician noise, cosmic noise, and one of my personal favorites, cyclostationary noise (where the mean and other statistical moments are periodic, as from a rotating radar antenna or a machine bearing). Note that you have to distinguish between the noise source , and the resultant noise type , as there are many possible sources for each type of noise.
Why do we make that Gaussian-noise assumption? It's an easy one to make, especially if you don't have the ability or opportunity to look at the noise of your situation in detail.
Even better, it simplifies the mathematical analysis, plus the hardware filtering and signal-analysis algorithms you may be planning to use. So we take a gamble, assume it is Gaussian, and hope it all works out in the end.
That's why I think that increased use of application-specific standard products (ASSPs) or application-specific ICs (ASICs) may help us overcome the result of this noise simplification. In principle, these ICs are designed with a specific application in mind, by their nature. Hopefully, the design and application teams have fully studied the noise of this application, and embedded filtering and signal-processing algorithms that are optimal for the situation at hand — though they may be quite wrong for others.
The design may need a low-pass, band-pass, time-varying, non-linear, or matched filter, to cite just a few options. The signal processing may also need to have algorithms that are tailored to the realities of the noise, such as multiple time constants and time bases, or special types of averaging.
These can truly make the ASSP (or ASIC) work where simplistic approaches may not be the right choice. After all, the noise seen by an electrocardiogram's analog front end (AFE) which is dealing with microvolt signals is quite different than the noise in a motor-control loop with MOSFETs switching nearby.
Or look at one of Jim Williams's earliest articles for EDN (“This 30-ppm scale proves that analog designs aren’t dead yet,” EDN , Oct 5, 1976), where he built a scale with dual time constants in the strain-gage front end: a fast one to accommodate mechanical noise from vibration, then switching to a slower one to deal when a sample (a baby!) was placed on the scale. It's unlikely that either “noise” source was Gaussian.
Noise is always there; how you handle it depends on what you know about it. Hopefully the developers of the latest ASSPs/ASICs understand it and can put that knowledge into their chip (or chipset), but only after verifying their analysis and approach to dealing with it.
What sort of noise annoys you? Have you seen circuits and algorithms which were well-tailored to non-Gaussian noise?