In a nutshell: smaller "raw" files aggregate the sensor values within blocks of pixels.

For instance, Canon's RAW format conveys information about individual "sensels." Each sensel (or "photosite") responds to a restricted range of frequencies (termed red, green, and blue). Each one of these, when later "developed," will be located at a single pixel site in the final image.

Canon's sRAW format, however, conveys summary information about 2 x 2 blocks of sensels. It reports brightness (luminance) data for each block, but "decimates" (skips over in a regular way) some of the color information. As such, several important things happen:

• The individual sensel data are no longer available. (The sRAW data are indeed "processed.")

• The resolution of the image is reduced (it is halved, implying there are a quarter as many pixels).

• The file size of the data is reduced approximately by two-thirds.

• The sRAW data are not a "subset" of the RAW data. They are a different encoding of the raw data, with less information. No sensels are "ignored."

(Normally, reducing the resolution of an image by a factor of two will decrease its size on disk to one quarter the original. Here, though, the original sensels deliver about 14 bits of information, amounting to 56 bits in each 2 x 2 block in the RAW format. In sRAW, each 2 x 2 block is encoded as three 8-bit pieces, or 24 bits. The resulting data stream is therefore only 24/56 = about 1/2 the size of the original, and is reduced by another 1/3 by the decimation of the chrominance data, for a net reduction of 2/3. Lossless compression is applied in sRAW, so the ratio may differ slightly.)

This information was obtained through extensive reverse engineering reported last year by Douglas Kerr, whose report I have very briefly summarized here (without too much distortion, I hope).