Sensor Shadows
	The dynamic range of the sensor in my Canon 5D is just over ten stops. Unfortunately getting the camera to give you ten usable stops of information is just about impossible, but that's the range of subject matter between the brightest value that the sensor can send out of its AtoD (analog to digital) converter and pretty much the lowest value. Every digital sensor, be it a PMT (photomultiplier tube, which drum scanners nearly always use), CCD, or CMOS chip, makes voltage which is proportional to the number of photons hitting it, which seems fair enough. Unfortunately, then taking the voltages and assigning them proportionally to RGB values (0 to 65,535 for a 16-bit scale) results in the brightest f-stop of the exposure being allocated across the entire top half of the RGB value scale (we usually think of that scale as 0 to 255, as it is with 8-bit per channel data). And the second brightest f-stop is then allocated across the range from 1/4 to half of the scale. You can see where this is going. Think of what the sensor sees as linear photons-to-voltage. But think of the way we see the world, with perceptually uniform steps of brightness between steps in a gray scale, as being perceptually linear. The two are entirely different, so when RGB values coming out of any sensor are mapped into any ordinary space like a monitor or a printer without a major adjustment, the result looks super dark. We see equal intervals, more or less, when each interval is brighter or dimmer in absolute terms by a given factor, such as a factor of two. If it weren't for this, our visual system would work very poorly. The deepest shadows which the sensor can perceive as different from noise are so terribly flat and have so few RGB levels resulting from the AtoD conversion, that if you don't have lots of bits (way more than 8) in the digital data, you get dreadful posterization in the shadows after the image has been opened way up, so that it looks normal to us, i.e. after it has been made more or less perceptually linear. If drum scanners didn't have 14 or 16 bits per channel, by the time the deep shadows of a transparency were opened up so you can see them better than you can on the chrome, they would be terribly posterized. The first f-stop of the 5D's response contains 2,048 steps. The second contains 1,024 steps. By the seventh stop, there are only 32 levels. The eighth has just 16, the ninth has 8, and the tenth has only four! (Prior to the creation of new,"false" levels by the Bayer interpolation and other processing steps.) This means that under the best of circumstances, the eighth stop can yield quite usably smooth tonality when it has about as much tonal separation as most of the tone scale does with respect to the subject matter, but the ninth and tenth stops are only usable as compressed, deep shadow detail, which is sort of OK, because most images need that anyway. Of course, we'd rather have more f-stops of usable dynamic range — I'd love to see at least two more. Part of what I am getting at here is that when you start with a file in 16-bit form, but which was really just a 12-bit file to begin with, and when you are working so as to actually make use of one of the darkest three stops in any of the three channels, the image isn't really much akin to a 16-bits per channel image, at least insofar as the shadows are concerned. Rather it's actually a little worse than a good quality 8-bit image from a fine 16-bit drum scan, despite being "16-bit". Therefore, quantization error that builds up in imaging processes can still affect the character of such images, and things like curves adjustments, levels adjustments, profile to profile conversions, and choice of working space volume all still have real effects on the quality of the image if the data have been converted down to 8-bits per channel. This argues against automatically using a space as large as DCam 4, ProPhoto or DCam 5, if you want the freedom to work in 8-bits and/or to be sure you avoid that quantization by unexpected reductions in bit depth by rude software. It also argues for using the smallest one that works for a given image. Files from the medium format backs, like the Phase One backs, are actually 16-bit files, having been generated from the sensor with a 16-bit per channel AtoD conversion. These files may all be just as well off stored in a giant space as in a smaller one into which they fit without clipping. May be.