Struct LogBiasTable 

pub struct LogBiasTable { /* private fields */ }

Precomputed observed − expected log-transition table for a fixed model pair. The effective-length correction evaluates log_bias for a context at every transcript position; log_bias evaluates BOTH models (each re-encoding the context and sweeping all CONTEXT_LENGTH positions), so a context costs two encodes + two table sweeps. Folding the pair into a single difference table diff[pos·ROWS+idx] = obs − exp (built once per quant run, the models being fixed during correction) collapses that to one encode + one sweep — ~1.4× on the per-position factor build, the dominant cost of the seqBias sweep.

eval equals log_bias(obs, exp, ctx, rc) up to floating-point reassociation: it sums Σ(obs−exp) rather than (Σobs) − (Σexp), a difference of ~1e-15 per context (machine epsilon), far below quant-output resolution.

Implementations

impl LogBiasTable

pub fn new(observed: &SBModel, expected: &SBModel) -> Self

Build the difference table from a normalized observed/expected pair.

pub fn eval(&self, context: &[u8], rev_comp: bool) -> f64

log_bias for a context (one encode, one table sweep).