Error-bounded lossy compression
Error-bounded lossy compression permits approximate reconstruction while constraining a specified error measure under stated input and implementation conditions.
Error measure: A pointwise absolute bound limits every |x - reconstructed x|. A pointwise relative bound divides each error by that value's magnitude, whereas a block-relative bound uses a block norm. These denominators answer different questions; zero values need an explicit policy. [1]
Scope: Error control concerns numerical reconstruction, not bitwise identity or the accuracy of a later simulation. The zfp error analysis separates conversion, transform and bit-plane truncation errors; a local reconstruction bound is not a proof of stability under repeated application updates. [1]
Example: Native zfp compression of [1, 2^-30, -2^-40, 0] at precision 16 loses both small nonzero values: maximum absolute error is 2^-30, but maximum pointwise relative error is 100%. Zero is excluded from the relative denominator and checked separately. [2]
Relations
- related
- zfp, Lossless compression
Limitations
- A requested tolerance is meaningful only with its units, norm, supported value domain and codec configuration; a size limit is a different constraint. [1, 2]
Citations
- Diffenderfer et al. (2019): Error Analysis of ZFP Compression for Floating-Point Data, sections 4-6, arXiv v3
- Hesela: native zfp 1.0.1 error budgets and padded-block byte accounting
Sources & caveats
- General error-measure synthesis, illustrated by zfp rather than a guarantee shared by all lossy codecs.
- The native inputs are constructed controls, not sampled scientific measurements. Positive-tolerance checks passed for the tested arrays only.