Delta
Delta Transforms
The XorDelta and FloatDelta transforms perform cross-tensor residual encoding. They subtract a reference plane (from a dependency tensor) from the current plane to produce a highly compressible residual.
XorDelta
XorDelta computes a byte-level XOR residual (residual[i] = current[i] ^ reference[i]).
Because XOR is involutive, the forward and inverse operations are identical. XorDelta is guaranteed to be byte-exact and lossless for all inputs.
FloatDelta
!WARNING
FloatDeltais marked as DONE_WITH_CONCERNS because it violates PTWM's byte-exact round-trip invariant.
FloatDelta performs an element-wise float subtraction (current - reference).
IEEE 754 floating-point arithmetic does not guarantee that (a - b) + b == a bit-exactly when a - b introduces a rounding error. Therefore, FloatDelta may be slightly lossy (by ~1 ULP). It exists for opportunistic delta encoding where near-identical tensors yield a residual so small that the coding gain outweighs the slight approximation.
References
- Generic Delta Encoding. Wikipedia