pub fn jirak_order(x: &[f64], config: JirakConfig) -> Result<JirakResult>Expand description
Jirak’s (2012, Ann. Statist. 40(2), eq. 2.14) refitted order estimator
q̂⁽⁵⁾: fit Yule–Walker AR(k) for every dimension k ≤ d_n, studentize
each coefficient vector with its exact null standard error
se_i = sqrt(γ̂*_ii(k) · σ̂²(k) / n) (the Statistic::JirakBand
studentization at dimension k), apply the common Gumbel threshold
calibrated on the maximal dimension d_n, and return the largest
per-dimension last-crossing.
The refits recover power that the single full-dimension fit
(Statistic::JirakBand inside fh_order) can lose: a coefficient
that is small in the AR(d_n) representation may be large in a smaller
one, and under H₀ the extra comparisons are already paid for by the
d_n-normalized threshold. Costs O(d_n⁴) — negligible for the
poly-log default (d_n = ⌈2 log n⌉).