Bit Depth

A standard display renders 256 levels per channel. A 16-bit camera image contains 65,536. Whatever is on screen, more than 99% of the levels in the file are not shown, and no display setting changes that. When a 16-bit image opens, the software applies a display window: a low value drawn as black and a high value drawn as white. Everything between them is mapped onto the 256 available shades, and everything outside the window is flattened to one end or the other. Moving the window shows different structure. The data are unchanged; the 256 shades now cover a different range. This is why a 14-bit and a 16-bit acquisition of the same field can look identical on screen while being measurably different files.

The opposite failure is visible. With too few acquired levels, quantization shows as posterization: smooth gradients break into flat bands with hard edges between them, because no value is available for what lies in between. Posterization in an acquired image cannot be undone, because the steps were made before the file was written. Posterization from a display window goes away when the window is moved. The two look identical on screen, so check which one you are seeing before concluding the acquisition was wrong.

The extra bits are for measuring, not for looking. Three things depend on them. Headroom: a bright structure that would clip at 8 bits still has room above it, so its intensity remains a number rather than a ceiling. Precision in arithmetic: ratios, background subtraction and unmixing all propagate the coarseness of their inputs, and a value known to one part in 256 cannot yield a ratio known any better. And deferral: because the display window is chosen after acquisition, keeping the full range means the decision about what to show is reversible, while acquiring at 8 bits makes it permanent at the moment of capture.

The relation has a closed form: for an ideal N-bit converter digitizing a full-scale sine wave, the signal-to-quantization-noise ratio is 6.02N + 1.76 dB, which is why each added bit is worth about 6 dB against quantization noise. In a camera that noise is small: with one count set at half to all of the read noise, quantization noise (one step divided by √12) is at most about 0.29 of the read noise, so the formula does not predict image SNR. There is a ceiling on the useful side too. The number of useful bits is set by the sensor (log₂ of full well divided by read noise), not by the converter: once one digitization step is smaller than the read noise, further bits describe noise with great precision. Sizing the converter comes down to matching it to the sensor's real dynamic range, full well divided by read noise.

Downstream, coarseness is what decides whether two populations can be separated at all. Thresholding and segmentation work by finding a value that divides one population from another, and if two populations sit only a few levels apart and each spreads over several levels, no threshold separates them cleanly. The algorithm is not at fault; the distinction was discarded at digitization. This is the practical reason to digitize finely enough that one step is no larger than the read noise, and it is a decision that cannot be revisited afterward.

Grayscale offers the eye one dimension to judge: lightness. That is a tighter limit than the 256 levels: shown a smooth gray ramp, most people resolve a few dozen distinct steps, not hundreds. A color lookup table changes the question by mapping intensity along a path that varies lightness and hue and saturation together, recruiting discrimination the gray ramp never used. Structure separated by a few hundred counts in a 16-bit file, invisible as two nearly identical grays, becomes plainly two different colors.

This holds only for a perceptually uniform map, where equal steps in value look like equal steps to the eye. Older rainbow maps are not uniform. They compress some ranges into a band of near-identical color and stretch others across a vivid transition, so real differences disappear in one place and a false sharp boundary appears in another.

A histogram alone says how many pixels hold each value. It cannot say what any of those values will look like, because that depends on a mapping the histogram does not draw. Put a color ribbon directly beneath it, on the same axis, and the two together show where the data sit and where the display spends its ability to distinguish.

Read it as a pair. A tall peak sitting over a stretch of ribbon that barely changes color signals a problem. Most of the pixels are in a range the display renders as one shade, so any structure among them is in the file but not on the screen. Wide bands of vivid ribbon over an empty stretch of histogram are the same waste in the other direction: discrimination spent where there is nothing to discriminate. Adjust the window until the ribbon's color changes fastest where the histogram is tallest. That puts the display's limited range of shades where the data are.