A raster image samples a continuous projected scene on a discrete pixel grid. If geometry or shading varies more rapidly than that grid can represent, the result aliases: diagonal edges staircase, thin features flicker and repeated patterns produce moiré.

Anti-aliasing reduces this problem by estimating a pixel's average contribution over a finite footprint instead of treating the pixel as one ideal point sample.

One approach is supersampling: evaluate several samples per output pixel and average them. Multisample anti-aliasing applies a related idea primarily to geometric coverage while avoiding repeated evaluation of every shading operation.

Filtering before reduction is equally important when an already sampled image is resized. Simply discarding samples can fold high-frequency detail into false lower-frequency patterns.

Anti-aliasing cannot recover information after undersampling has destroyed it. The renderer must sample densely enough, or filter out detail that the final pixel grid cannot represent, before producing the final samples.