[Radio Wrangler]

Theoretically, all noise levels will have peaks which can reach infinite voltage, in both directions. This means that ALL real-world noise has limited crest factor, and it also means that you cannot assess the level of a noise signal by looking at the peak voltage. Only statistical methods can get a grip on it.

Where the noise creation mechanism is the sum of a great many small, similarly sized contributors, the central limit theorem says the probability density function will tend to Gaussian. So, everyone always assumes noise is Gaussian, which leads to the rather interesting circumstances when it isn't, being overlooked.

Human perception of white noise in analogue systems seems to follow the RMS level.

As Jeremy says, in digital systems, it's the peak values which do the greatest damage. In something like a QAM modulated transmission, if peaks aren't big enough to cross into the next state, you get no errors, so crest factor limitation makes the Eb/N0 curve stop dead to zero errors instead of the theoretical decline. Noise peaks, if not clipped can generate multi-bit errors on occasions and the classic text books on digital radio give the standard analyses... but limited crest factor makes things massively different.

The HP 3708A is one of their more obscure instruments. It would take the IF of a microwave radio link receiver, measure its level (true RMS thermal converter) and create band limited noise of a level that tracked it, giving a servo-controlled signal to noise ratio for assessing the liability of demodulators to create errors versus noise level. In designing it, we had to be very, very careful to always maintain good crest factor. I had to design filters which maintained constant noise-bandwidth over temperature for this.

David