Module lorenzo 

Lorenzo-predictor codec for pointwise (max-abs) bounds, and for RMSE / PSNR targets through rate control of the bound.

Aimed at rough fields with one smooth (or repeating) axis, such as particle trajectories stored time × particle, where interpolating across the rough axis costs more than it predicts.

1. Prediction, in row-major order on f32 reconstructions. Mode 0 is the full N-D Lorenzo predictor (inclusion–exclusion over the axes whose coordinate is > 0); mode k + 1 is the previous point along axis k (the N-D predictor where that coordinate is 0). The mode is picked by real stream size.
2. Quantization, q = round((x − x̃)/2b), checked on the exact f32 reconstruction; failures are escaped raw (as in crate::interp).
3. Coding. Adaptive binary range coding: a zero flag with contexts (symbol class of the previous point along the prediction axis and of the side neighbour, log₂ of the last step along the prediction axis in quanta), a sign bit keyed on the previous point’s sign, and an adaptive Elias-gamma magnitude keyed on the two neighbour classes.

Constants

MAGIC

Stream magic of the Lorenzo codec.

Functions

decode

Decode a Lorenzo stream.

encode_f32

Encode x (row-major, shape dims) with |x − x̂| ≤ b everywhere, trying each prediction mode and keeping the smallest stream.

encode_psnr_f32

encode_rmse_f32 at PSNR = 20 log₁₀(range / RMSE) dB.

encode_rmse_f32

Encode for an RMSE target (rmse absolute): the mode is picked at the bound rmse·√3, then that mode’s bound is rate controlled as in crate::interp::encode_rmse_f32. The target is checked on the exact reconstruction.