ADST — Asymmetric Discrete Sine Transform — is a mathematical sibling to the DCT that AV1 and VP9 use for specific block patterns the regular DCT handles poorly. Where the DCT assumes the residual (the small differences left after prediction) is symmetric within the block, ADST assumes one edge of the block has higher error than the other. That's exactly what happens at the edge of a prediction region: the side closest to the reference is well-predicted, the far side is less so, and the residual is asymmetric.
For these block patterns, ADST compresses the residual more compactly than the DCT — typically a few percent better, which adds up across the whole frame. AV1 actually uses a menu of transforms (DCT, ADST, flipped ADST, identity transform) and the encoder picks whichever one compresses each specific block best. The encoder spends extra computation evaluating the options and saves bits in the bitstream. That's a typical trade-off in modern codecs: more encoder complexity in exchange for smaller files.
For a product team, ADST is invisible plumbing that doesn't appear in any user-facing setting. It's part of why AV1 compresses better than HEVC, and why AV1's reference encoders (SVT-AV1, libaom) are noticeably slower than HEVC encoders on the same content — they're running more transform evaluation per block. The takeaway: when codec vendors claim X percent efficiency gains from "new transform tools", ADST is usually what they mean. The math has been around since the early 2000s but only recently became practical enough to deploy at scale.
