1) Using the upscaling integer number (1x, 2x, 3x etc) and downscaling to same resolution with same algorithm should, when properly implemented, yield zero change.
2) It should be done after switching to 16 bit space - there won't be any profit in upscaling if image cannot take more than 256 tonal values.
3) I have just tried all three bicubic algorithms from Photoshop CS5 and none of them had given any increase of output tonal depth.
I have used folowing steps: 1) make a 8 bit image 2) put a white-to-black gradient on it 3) switch it to 16 bit 4) upscale 2x with different algorithms 5) apply tone curve {(33,0), (160,255)} (127 output values)
With either alrogythm I observed the following: I could not find every second tonal value on the image.
However: - if I scaled up using non-integer value, I got some smoothing (effectively loosing details) - if I applied curves to 8 bit image, Photoshop applied some dithering around gradient steps
