Lossy Data Compression
Lossy compression achieves large size reductions by discarding information judged unimportant, trading exactness for compactness.
Trading fidelity for size
Lossy compression shrinks data by permanently discarding some of it, keeping enough to reconstruct an acceptable approximation of the original. It reaches far higher ratios than lossless methods but the original cannot be recovered. It is the right tool for imagery, audio, and video meant for human perception, and a dangerous one for authoritative scientific data.
How it discards
- Transform coding: move data to a domain (frequency) where less-important components can be dropped.
- Quantization: reduce the precision of values, coarsening the representation.
- Perceptual models: discard detail humans cannot perceive.
- Error-bounded methods: guarantee each value is within a stated tolerance of the original.
Rate-distortion
The central trade is between rate (size) and distortion (how far the reconstruction departs from the original). Rate-distortion theory formalizes this frontier: for a given acceptable distortion, there is a minimum achievable size. Choosing an operating point means deciding how much error the use can tolerate.
Error-bounded scientific compression
For large simulation outputs, error-bounded lossy compressors let a scientist set a strict tolerance, for example a maximum relative error per value, and then compress aggressively within it. This can shrink data enough to make otherwise impossible archives feasible, but the tolerance becomes part of the data's metadata and must be reported honestly.
When it is and is not acceptable
Lossy compression is appropriate for derived visualizations, previews, and media, where perceptual quality is what matters. It is not appropriate for the authoritative raw record, where any silent loss would corrupt reproducibility. The Kronos published record keeps authoritative data lossless; lossy methods appear only in generated media and visual previews, never in the deposited numeric datasets.