Correct the October 2026 audit findings: Sammon curvature and descent,
coreset sampling with replacement and weighted clustering, stable Stein and
Bures-Wasserstein distances, spherical likelihood bounds and log densities,
Stiefel QR signs, Grassmann optimizer returns, nearest-neighbor self exclusion,
Grassmann metric-learning dimensions, and extreme-weight tangent variances.
Coreset clustering now honors maxiter; coincident samples have an explicit
zero-dispersion sampling rule. Help clarifies that first-order stationarity
does not certify a local minimum. Independent regression fixtures accompany
these corrections; legacy experimental status remains where appropriate.
The September 21, 2026 source revision licenses author-owned package software under GPL-3 and retains third-party attribution. Earlier MIT releases and preserved archives keep their original licenses. Package source comments and roxygen text are now ASCII; formatted help retains accented names.
The September 21 r3 documentation revision identifies the ERP matrices as xDAWN-augmented MEG covariances through a numerical reconstruction of the historical recipe. It records the MNE source's CC0 declaration and retains GPL-2 terms for the shapes-derived gorilla data. Data and algorithms are unchanged.
Correct SPD validation and the affine-invariant metric. Symmetric matrix functions replace general complex-valued matrix logarithms in core SPD paths.
Preserve tiny sphere angles and report nonunique antipodal logarithms and degenerate projected summaries explicitly.
Share mean and median solvers between public summaries and internal callers, with validated weights/controls, descent, and actual termination diagnostics.
Preserve fitted regression geometry, stabilize kernel weights, and correct all-fold cross-validation, singleton subsets, and candidate error tables.
Add explicit geometry specifications and a capability registry. Existing intrinsic/extrinsic aliases remain available with their documented meanings.
Tangent PCA now honors requested rank, preserves the chosen metric, and supports prediction and reconstruction under a trained reference.
K-means uses validated Fréchet-center updates, per-start diagnostics, and fixed-center prediction. The default is now Lloyd; MacQueen remains available.
Landmark wrapping explicitly uses O(p), including reflections, and rejects singular configurations. The default alignment reference is now the first normalized configuration without the former extra covariance-eigenvector rotation; quotient distances retain their interpretation.
Correct additional Grassmann, rotation, positive-simplex and fixed-rank PSD primitives, with explicit cut-locus and projection-uniqueness checks. Intrinsic Stiefel and correlation operations are restricted pending consistent geometry.
Reject invalid correlation matrices, negative simplex masses, rank-deficient frames and implicit fixed-rank truncation. Preserve singleton dimensions and named feature order through fitted summaries and prediction.
Correct FANOVA scaling, permutation bookkeeping, kernel-PCA scores, classical scaling spectra and the mutual-neighbor Isomap graph. Updated help states the narrower assumptions and experimental status of remaining legacy routines.
Cache log-Euclidean matrix transforms within pairwise and cross-distance calls. Add executed data workflows, matched software comparisons, and migration help.
Validation and remaining publication requirements are tracked in the separate write-JSS-Riemann/maintenance/ workspace. Journal submission remains a separate step.
CVXR per recent API changes.riem.phate() is added for sub-manifold learning/visualization.data(ERP).spd.wassbary() is added to compute Wasserstein barycenter of Gaussian distributions.mle.spnorm() for controls with user-defined stopping criteria.mixspnorm().riem.m2skreg() for manifold-to-scalar kernel regression and riem.m2skregCV() for parameter selection using cross validation.NEWS.md file to track changes to the package.