If an image, which only has 8 bit color, is scaled up, the output of the filtering calculations will have more then 8 bits of precision.
As an illustration, imagine that there are two adjacent pixels, one with the value 0, and another with the value 1. This two pixel image is scale dup horizontally by a factor of 2. The simplest linear filter that you could use would yield 0.5 for the pixel half way between them. i.e., there are now more fractional bits worth displaying then there were in the original output.
Even without scaling, the output of a decompression algorithm can yield more bits than the original input, as can any adjustments such as brightness, contrast, etc.
Originally Posted by noah katz
"The 10 bit color extension would be useful even without extended gamut because even the traditional gamut would have more color resolution - i.e. more in-between shades for existing the color spaces."
Where are those in-between shades going to come from?
Besides, a good display doesn't suffer from banding.