The DCT is a relative of the Fourier transform and likewise gives a frequency map, with 8x8 pixel-blocks. Thus numbers are
representing the average value in each block and successively higher-frequency changes within the block. The motivation for doing this is that high-frequency information can now be thrown away without
affecting low-frequency information. (The DCT transform itself is reversible except for round off errors.) In each block, each of the 64 frequency components is divided by a separate "humanisation
coefficient", and the results are rounded to integers. This is the fundamental information-losing step. The larger the quantization coefficients, the more data is discarded. Note that even the minimum
possible quantization coefficient, 1, loses some information, because the exact DCT outputs are typically not integers. Higher frequencies are always quantized less accurately (given larger coefficients)
than lower, since they are less visible to the eye.