In supersampling we sample more than 1 sample per pixel. These samples can be at regularly spaced intervals. For example, we might compute an image at 2K by 2K or 4K by 4K points and and display at 1K by 1K pixel resolution. Sampling at 2K x 2K for a 1K x 1K image increases the number of samples and graphics computations (and Z-buffer requirements for scan-line graphics) by a factor of 4. An alternative method would be to sample at the corners and center of each pixel. This only increases the number of computations by a factor of two but requires increased overhead for bookkeeping (since the samples from the previous row must be stored). Supersampling has the effect of moving the Nyquist Limit to higher frequencies.

Rather than combining pixels with unweighted average (box filter) might use a weighted filter. We might even average pixel samples over several pixels , i.e. combine over a wider range with weighted average. This is an example of digital filtering. Even if we move the Nyquist Limit to higher frequencies, we will still have aliasing. Adaptive supersampling: additional sampling in region of high frequency, eg. near edges. We continue supersampling as long as |Ic - Ii| > threshold variance. It does better but requires large number of rays/samples and it is a more complicated algorithm.

From HyperGraph -- see details 4.12.1996