After the little interaction Paul Butzi and I had on an earlier post, I found this article on noise via Luminous Landscape. All you ever wanted to know about noise, and then some (and maybe a bit more besides). WARNING: it's pretty technical and I've not fully absorbed all the stuff.
A few interesting points it raises:
Ideally, the noise should slightly exceed the quantization step, in order
that roundoff errors introduced by quantization are negligible, and that no bits
are wasted in digitizing the noise...
Curiously, all the 14-bit cameras on the market (as of this writing) do not
merit 14-bit recording...
Twelve bits are perfectly adequate to record the image data without any
loss of image quality, for any of these cameras.
The proper reason to expose to the right comes ... the rise in signal-to-noise
ratio with increasing exposure.
Somewhat counter-intuitively, for fixed aperture/shutter speed, it is best to use the highest ISO (without clipping highlights); this result is consistent
with the ETTR philosophy, since using higher ISO pushes the histogram to the
right. However, the benefit from the use of higher ISO comes in the shadows, not
in the highlights where "there are more levels".
The fact that a digicam's performance is in the same ballpark as the best
DSLR's when referred to fixed spatial scale, suggests that the problems with
noise in digicams is not due to their ever smaller pixels, but rather it is due
to their small sensors.
Selected quotes only, go read the whole thing. Fascinating.