This means: After the transform, some coefficients are predictably greater, others smaller. Thus, some coefficients can be neglected or quantized (lossy compression) and/or entropy encoded (lossless compression).

For instance, JPEG performs a discrete cosine transform on 8-by-8 pixel blocks (in some pictures the chrominance is subsampled, so is transformed in 16-by-16 blocks with 3/4 of the results discarded), then the coefficients are quantized depending on their frequencies.

See also: