Changelog • multivarious

multivarious 0.4.0

Added sparse-aware preprocessing for Matrix inputs. pass() and sparse-safe colscale() transformations preserve sparse storage, while center() and standardize() now use a lazy affine representation instead of materializing dense centered matrices by default.
Added explicit sparse preprocessing policies: sparse = "operator" uses the lazy representation, sparse = "preserve" errors before a densifying transform, and sparse = "allow_dense" opts into dense materialization.
Added matrix-like S4 classes for lazy sparse affine data and its transpose, with methods for %*%, crossprod(), tcrossprod(), t(), subsetting, dimensions, and as.matrix(). Downstream code can now multiply or form cross-products from transformed sparse data without first densifying it.
Updated svd_wrapper() and pca() sparse/lazy paths to use implicit sparse SVD backends where possible, avoiding dense centered intermediates for common sparse PCA workflows.
Updated projector and partial-projection paths to multiply through lazy sparse preprocessing views directly.

transform() on a sparse input with centering or standardization may now return a sparse_affine_matrix object. Use regular matrix algebra methods on that object to stay lazy, or call as.matrix() when explicit dense materialization is desired.

Added regression coverage for sparse preprocessing policies, sparse-safe scaling, lazy centered and standardized sparse transforms, implicit sparse PCA, projector multiplication, and public algebra methods on the lazy sparse affine classes.

CRAN release: 2026-06-05

Removed the optional PRIMME backend from geneig() and cPCAplus() (and dropped PRIMME from Suggests), as the PRIMME package is scheduled for archival on CRAN. The iterative "rspectra" and "subspace" backends and the dense "geigen"/"robust"/"sdiag" backends cover the same generalized eigenproblems.

Added explicit term_scopes and exchangeability overrides to mixed_regress(), with exchangeability metadata now carried through summary(), effect(), and effect-operator printing.
Fixed grouped row-metric whitening/unwhitening for random-effect designs by using the Cholesky orientation implied by the row metric.
Hardened perm_test.effect_operator() by making the supported one-sided alternative explicit, preserving seed metadata, honoring explicit exchangeability schemes, and using a fixed statistic family for each sequential permutation step.
Improved bootstrap.effect_operator() for grouped designs by relabeling duplicated bootstrap clusters as distinct resampled groups, aligning multi-component loadings with an orthogonal Procrustes rotation, and preserving seed metadata.

Added regression tests for grouped whitening, explicit term-scope and exchangeability overrides, effect-operator permutation statistic selection, subject bootstrap relabeling, seed metadata, and PRIMME removal.

CRAN release: 2026-01-21

Changed partial_project() default least_squares from TRUE to FALSE.
Changed project_block.multiblock_projector() default least_squares from TRUE to FALSE (now matches partial_project()).

Fixed reconstruct_new.bi_projector() double-preprocessing bug that caused incorrect reconstruction when applied to held-out data.
Fixed blockwise preprocessing paths that could allocate very large temporary matrices with multiblock data, and hardened standardize() for missing and zero-variance columns.

Rewrote CrossValidation vignette with working examples (fixed broken reconstruct() usage and results extraction).
Cleaned up PermutationTesting vignette: improved structure, replaced dense tables with readable prose.
Cleaned up Regress vignette: broke up long code block into focused subsections.
Cleaned up Extending vignette: removed commented-out code walls, simplified examples.

CRAN release: 2026-01-21

Fixed T/F shorthand to TRUE/FALSE in pca() for CRAN compliance.
Converted \dontrun{} to \donttest{} for executable but slow examples.
Fixed bootstrap.plsc() duplicate argument handling when called with named X/Y arguments.
Fixed regress() PLS method dimension mismatch.
Fixed iris data frame to matrix conversion in examples.

Registered S3 methods: classifier.projector, inverse_projection.projector, perm_ci.pca.
Added missing importFrom directives for coefficients and combn.
Replaced non-ASCII characters with ASCII equivalents in documentation.

CRAN release: 2024-03-28