heap_lattice() is reimplemented. It now performs the Moebius inversion the
documentation always described, reading the empirical characteristic function directly
at the exclusive replica centers of each grain with no binning and with the sampling
floor removed. Its arguments change from span and Mgrid to nboot and seed, and
it gains bootstrap standard errors.
The 1.0.0 reader did not invert. It normalized replica amplitudes, and the unrounded
share it returned was identically its grain-one weight, so the two quantities it
reported separately were one number. On a known-truth simulation it was biased on every
grain. Any weights obtained from it are superseded.
The new reader has a stated regime. It recovers every weight to within about four percent once kappa, the standard deviation of the dequantized base over the coarsest grain, exceeds roughly three, and recovers the unit share at every kappa tested. Below that the grain split is descriptive rather than a measurement, and the documentation says so.
Documentation corrections, from experiments run for the accompanying paper. heap_grid()
now records that its blind mode returned the imposed grid in 0 of 12 test cases where its
verifying mode returned it in 12 of 12, so a blind reading is a candidate rather than a
recovered grid. The README no longer describes heap_detect() as blind spectral
detection; with its shipped settings it abstains in most cells. The package Description
no longer calls the grid reader blind, and no longer calls the base-R baselines faithful,
since the Heitjan-Rubin replica has not been checked against a published implementation.
No change to deheap_kde(), superpose_kde(), adkde(), heap_fraction(),
heap_detect(), or any baseline. Density estimates from 1.0.0 are unchanged.
deheap_kde), superposition
(superpose_kde), and the combined band-capacity-gated estimator (adkde);heap_grid, heap_fraction, heap_lattice);heap_detect);deconv_kde) and the
Heitjan-Rubin multiple-imputation approach (heitjan_mi), and a wrapper for the
Kernelheaping stochastic-EM estimator (sem_kde).
Derived from the spectral-decomposition kernel density estimation of Thornton
(arXiv:2606.15450).