fastQR 1.2.0
Numerical backends
- Added
qr_lapack() for unpivoted QR using R’s configured
LAPACK, with compact, thin, complete and R-only output
representations.
- Added a shared internal backend for selected dense matrix and
matrix-vector products using DGEMM and DGEMV. Small products retain
Eigen kernels to avoid unnecessary call overhead.
- Thin and unpivoted compact QR use explicit DGEQRF/DORGQR calls;
recursive panels use DGEQRF/DORMQR. No LAPACKE dependency is introduced.
Obsolete Eigen BLAS/LAPACKE macro definitions have been removed.
- Equivalent QR factors can differ in signs or rounding from earlier
implementations; output fields and dimensions retain their
contracts.
Ridge regression
- Ridge fits and ridge cross-validation support more predictors than
observations for strictly positive penalties. A zero penalty still
requires full column rank, including in cross-validation training
folds.
- Legacy residual degrees-of-freedom and variance diagnostics are NA
when p exceeds n. They are not effective ridge degrees of freedom.
Documentation and checks
- Corrected missing usage sections for
qrmridge(),
qrmridge_cv() and qrridge_cv() in their
documentation sources.
- Excluded maintainer submission comments from source-package
archives.
- Expanded regression tests for LAPACK output, wide ridge fits,
predictions, input validation and foldwise cross-validation
references.
Correctness
fixes carried forward from the unreleased 1.1.6
- Corrected complete-Q output after column downdates, orthogonal-basis
completion, recursive complete QR and reduced-factor extraction.
- Corrected ridge penalty normalization, residual calculations and
cross-validation fits; made fold generation reproducible.
- Improved rank and dimension validation and reduced unnecessary
allocations in reduced-output update and downdate paths.