Rice coding
Rice coding is Golomb coding with a power-of-two divisor, storing an unsigned integer as a unary quotient followed by a fixed-width remainder.
Bit cost: With parameter k, the divisor is 2^k and the codeword length for x is floor(x/2^k) + 1 + k bits. A poor k makes the unary part expensive; smaller inputs alone do not guarantee a smaller complete record. [2]
FLAC: Signed prediction residuals are folded with ZigZag before coding. Partitions have their own 4-bit or 5-bit parameter field. Its all-ones value escapes to fixed-width signed residuals, with a further 5-bit width field; width zero represents an all-zero partition. [1]
Zero-run arithmetic: For one partition of N zero residuals, k = 0 costs N bits after the parameter field; the zero-width escape costs 5. Escape wins when N > 5, ties at 5 and loses below 5. This compares residual modes with the same surrounding structure, not complete FLAC files; a constant subframe is another choice. [2]
Interpretation: Decoding needs k and the residual count. A valid codeword is not an integrity check, and the Rice payload omits prediction state and container metadata. [1]
Relations
Citations
- RFC 9639: FLAC, sections 4.4 and 9.2.7 (Rice codes and escaped partitions)
- Hesela: exact Golomb/Rice bit-cost examples and partition accounting
Further reading
Sources & caveats
- The partition threshold is exact format arithmetic, not a measured audio compression ratio or an encoder-policy recommendation.
- The worked model accounts for method, partition order and parameter fields separately; frame headers, warm-up samples, padding and checksums are excluded.
- Tests and generated examples: https://hesela.dev/research/golomb-rice.test.mjs and https://hesela.dev/research/golomb-rice.json.