> Chroma subsampling isn't going anywhere. You'll usually get subjectively better quality with 4:2:0 chroma compared to 4:4:4 at the same bitrate. And this means you can't have everything in RGB, so all the colorspace conversion complexity can't be ignored.
What's more, YCbCr is more efficiently compressed than RGB even if you don't subsample, for the same reason that a DCT saves bits even if you don't quantize: Linearly dependent or redundant information is moved into fewer components, in this case most of the information moves into the Y channel with the Cb and Cr both being very flat in comparison. (Just look at a typical YCbCr image reinterpreted as grayscale to see what I meant)
isn't it the case that amount of data required to store the result of a lossless DCT is bounded below by the size of the data, and this is why lossless JPG compression does not use such a scheme?
I'm not actually sure. In retrospect, I'm not sure what ‘DCT without quantizing’ really means, since the output of the cosines are probably real numbers? I guess the interpretation would be quantized to however many steps to reproduce the original result when inverted (and rounded).
In lossless JPEG it seems they omitted the DCT primarily for this reason: It not being a lossless operation to begin with, if you actually want to store the result. What other lossless codecs often do is store a lossy version such as that produced by a DCT, alongside a compressed residual stream coding the difference (error).
In either case, it's important to note the distinction between reordering and compressing; reordering tricks like DCT can reorder entropy without affecting the number of bits required to store them, but the simple fact of having reordered data can make the resultant stream much easier to predict.
For example, compare an input signal like this one:
FF 00 FF 01 FF 02 FF 03 FF 04 ...
By applying a reordering transformation to move all of the low and high bytes together, you turn it into
FF FF FF FF FF .. 00 01 02 03 04 ..
which is much more easily compressed. As for whether that's the case for (some suitable definition of) lossless DCT, I'm not sure.
What's more, YCbCr is more efficiently compressed than RGB even if you don't subsample, for the same reason that a DCT saves bits even if you don't quantize: Linearly dependent or redundant information is moved into fewer components, in this case most of the information moves into the Y channel with the Cb and Cr both being very flat in comparison. (Just look at a typical YCbCr image reinterpreted as grayscale to see what I meant)