version 1.1.0

Disassortative mating

• am_simulate() and am_covariance_structure() now support negative (disassortative) cross-mate correlations. The equilibrium covariance becomes diagonal minus low rank in that case, tracked by an attr(U, "sign") that rb_dplr() honors automatically.
• fixed am_covariance_structure() returning NaN at r = 0
• fixed h2_eq() returning NaN at r = 0, where the answer is just h2_0, since no assortment leaves heritability where it started. The closed form divided by 2r when the bracket it multiplied also vanished there, and the same cancellation made small r ill conditioned: at r = 1e-12 it was wrong in the fifth digit. It is now defined through vg_eq(), as equilibrium genetic variance over equilibrium total variance, which divides by nothing that vanishes. rbahadur simulate --r 0 previously reported equilibrium h2 NaN.
• rb_dplr() gains a sign argument, and its infeasibility error now names the offending locus and suggests concrete remedies
• documented that negative r leaves the Bahadur feasible region sooner than positive r, and that the magnitude of r that samples reliably depends on sample size as well as min_MAF; see ?am_covariance_structure for measured envelopes

Genotype output

• am_simulate() gains path, format, and batch_size for streaming genotypes to disk in batches, removing the full-matrix allocation
• new write_genotypes() and read_genotypes() supporting individual-major int8, variant-major int8, and PLINK bed
• README documents reading the int8 and bed output into Python
• genotype matrices are now checked before integer conversion, so fractional or non-finite dosages are rejected rather than silently truncated
• genotype metadata is now parsed strictly: unknown layouts, incompatible dtypes, invalid dimensions, duplicate fields, and unsupported versions are reported as corruption instead of being guessed; int8 readers and rbahadur info also detect dosage bytes outside 0, 1, and 2
• PLINK output from am_mosaic() now carries the reference panel’s chromosome, physical position, genetic map, marker ID, and allele metadata when present

Local LD

• new am_mosaic(), which combines the genome-wide linkage disequilibrium induced by assortative mating with the local linkage disequilibrium induced by limited recombination. Causal variants are drawn with rb_dplr() and the intervening markers are filled by copying contiguous blocks from a reference panel, following Algorithm S4 of the supplementary note. Unlike the published vignette, block boundaries are drawn from a genetic map rather than uniformly, so breakpoints concentrate where recombination occurs.
• am_mosaic() accepts the same path, format, and batch_size arguments as am_simulate(). The variant-major layouts stream over markers and the individual-major layout streams over people, so no full genotype matrix is held either way.
• new kg_reference() returning a bundled 1000 Genomes panel (520 haplotypes across 2500 common SNVs in a 1 Mb window of chromosome 22, GRCh38, with genetic map positions), so examples and tests exercise real LD offline
• new vcf_to_panel() to build a panel from a phased VCF and a PLINK genetic map, and download_1kg_panel() to fetch a region of 1000 Genomes directly
• vcf_to_panel() also accepts unphased calls, with a prominent warning that their supplied allele order is treated as phase and can create artificial LD
• VCF records are now width-checked independently, symbolic alleles are not mistaken for SNVs, marker positions must be ordered, and genetic-map interpolation is restricted to the retained chromosome
• streamed reference downloads are written to temporary files and published to the cache only after the requested interval has demonstrably completed; map archives and extracted maps are likewise published atomically
• new vignette walking through the method, verifying that the assortative mating structure is preserved exactly and that local LD matches the panel, and documenting the tradeoff between block length and the infinitesimal limit underlying vg_eq()

Performance

Except for the corrected one- and two-variable rb_dplr() edge cases, output on previously tested simulation dimensions is unchanged: it is bit-identical under the same seed across all formats and batch sizes and leaves the random number stream in the same state.

• am_mosaic() expands markers 3 to 6 times faster and in roughly a third of the memory. Measured on a 2,000 haplotype by 100,000 marker panel with n = 1,000 and 400 causal variants: in memory 19.3s to 5.7s, individual-major 18.7s to 7.3s, variant-major 31.0s to 5.9s, and bed 36.3s to 6.0s, with peak resident memory falling from 4.0 GB to 1.3 GB.
• reference panels are now held as one byte per allele rather than being expanded to integer on entry, which was four times the memory and, on a genome-scale panel, more time than the simulation it was preparing for
• the variant-major and bed layouts no longer rescan every individual at every marker to find who has crossed a block boundary; boundaries are compiled once into a marker-ordered schedule
• the individual-major layout copies blocks out of a transposed panel, so a haplotype is a sequential read rather than a stride over the panel
• rb_dplr() draws its uniforms one locus at a time rather than allocating an n by m matrix of them up front, the same equivalence am_simulate()’s streaming path already relied on
• PLINK bed packing recodes by lookup and packs a block of variants per call instead of one call per variant, which also speeds up write_genotypes()

Command line interface

• new rbahadur executable, shipped in exec/. rbahadur simulate streams a simulation to disk without opening R, and rbahadur info inspects an existing run and verifies the data file against its metadata. Exit status separates usage errors (1) from runtime failures (2).
• rbahadur_cli_path() returns the location of that script so it can be put on the search path; rbahadur_main() exposes the same interface from R
• --csv writes portable _pheno.csv and _variants.csv sidecars alongside the R-only .rds, for pipelines that continue outside R
• unknown or misspelled command-line options and extra positional arguments are now rejected as usage errors instead of being silently ignored
• duplicate options are rejected, and --seed 0 is accepted as a valid R seed

Other

• added a testthat suite
• utils added to Imports
• simulations with fewer than 50 causal variants remain supported but now warn that the equilibrium variance and heritability formulas are large-locus targets; the warning can be disabled with options(rBahadur.warn_small_m = FALSE)
• declared and preflight-checked the external commands used by download_1kg_panel()
• simulation, Bernoulli, equilibrium, reference-panel, metadata, and PLINK inputs now receive explicit domain and integer validation; the broken one- and two-variable recursion edges in rb_dplr() and rb_unstr() are handled directly

version 1.0.0

• citation updated after publication
• typos fixed in documentation

version 0.9.2

• am_simulate() can use user-specified allele frequencies
• am_simulate() can now return optionally return haplotypes