ssMRCD 2.0.3

• Improved input checks and rescaling for weightsin ssMRCD().
• ssMRCD() now uses the names of a list input X as group names.
• In ssMRCD(), the tuning results now use correct names and order of groups. Bugs fixed that led to an error.
• Documentation updated.
• scores(): data is now only centered by the local means before computing scores, score distances and orthogonal distances. Previously, the data was additionally scaled by the (non-robust) root mean square of each variable per group, so results differ from version 2.0.2.
• scale.ssMRCD() now has the arguments of the generic scale() and only scales the data stored in the ssMRCD object; use scale = FALSE for centering only.
• plot.locOuts(): calling plot() with default arguments failed. Invalid values for type, scale and observation now give informative error messages.
• plot.ssMRCD(): ellipse plots now work for data without column names, the title reflects tolerance, and manual_rescale can take any positive value.
• ssMRCD(): best6pack now correctly returns, for each group, the index of the initial h-set of the best starting combination (previously a single, unrelated value).
• ssMRCD(): fixed a rare case where poorly conditioned initial h-sets could be used if only one well-conditioned initial h-set was available in a group, and a failure if only one combination of starting values remained.
• ssMRCD(): tuning with tuning$method = "local contamination" no longer resets the global random seed. Use set.seed() before calling ssMRCD() for reproducible results. Results therefore differ from version 2.0.2.
• geo_weights(): groups with a single observation and factors with unused levels no longer fail or produce NA. W and centersN are now named by group.
• ssMRCD(): if the global MCD target matrix is singular (e.g. variables constant within groups), the MRCD is now used as target matrix as intended. Previously, a warning was given and the estimation failed.
• locOuts(): numeric group labels that are not consecutive (e.g. 3, 7, 20) no longer fail.
• time_weights(): off_diag with N - 1 values no longer fails.
• groups_gridbased(): observations on the upper boundary of the grid are now assigned to the last cell instead of failing.
• residuals.ssMRCD(): performance improvement for type = "trimmed_mean". Invalid values of type now give an error.
• locOuts(): observations without neighbors within dist now give a warning (they cannot be flagged) and an error is given if no observation has neighbors. The message that k is ignored when both k and dist are given is now shown.
• ssMRCD(): tuning with tuning$method = "residuals" no longer fails if lambda contains duplicated values. Tuning with "local contamination" gives an informative error if tuning$cont is too small.
• ssMRCD(): inverse covariance matrices (MRCDicov) are now computed via Cholesky decomposition.
• Internal clean-up (unused computations removed, simplified code).
• plot.msPCA() now returns the plot itself for a single type and a correctly named list of plots for several types.
• msPCA(): tuning of eta and gamma now gives correct results if the grids are not sorted increasingly.
• scores(): performance increased.
• msPCA(): the automatic increase of rho no longer stops at the first infeasible value when msPCA() is called inside tryCatch(). A clear error is given if no feasible solution is found. A starting rho above the maximal rho no longer fails.
• msPCA(): a large thresholding tolerance (eps[1]) no longer produces NaN loadings (at least the largest loading per group is kept).
• msPCA(): the progress bar during tuning of eta/gamma is now controlled by show_progress (default FALSE) and only shown in interactive sessions. Diagnostic output for infeasible ADMM steps is now part of the message and can be suppressed with suppressMessages().converged is now also correct if convergence is reached in the last iteration. Correlation matrices are now also recognized if their diagonal is 1 only up to numerical precision.
• msPCA plots and summary now consistently use the group and variable names of the fitted object.
• plot(type = "score_distances") for msPCA objects now uses the first k components as documented (previously only the last component was used, and k was ignored).
• cellMGGMM(): Sigmai and Sigma_reg are now returned on the original scale. groups now contains the full group vector (previously a single value).
• cellMGGMM(): removed console output from compiled code.
• cellMGGMM(): progress output now uses message() and is only shown with silent = FALSE. Non-convergence and numerical inaccuracies are reported as warnings. No plot is drawn.
• cellMGGMM(): informative errors for groups with fewer than 4 observations, fewer than 2 variables, and variables with too few distinct values or zero MAD within a group.
• cellMGGMM(): Updated reference.

ssMRCD 2.0.2

• Version 2.0.2 not released on CRAN. Supplementary code for paper “Outlier-Robust Multi-Group Gaussian Mixture Modeling with Flexible Group Reassignment” (Puchhammer, Wilms and Filzmoser, 2026, https://doi.org/10.48550/arXiv.2504.02547).
• Updated references: sparse robust PCA for multi-source data is now published (Puchhammer, Wilms and Filzmoser, 2026, doi:10.1007/s11222-025-10771-y).
• cellMGGMM(): setting maxcond = Inf now disables regularization (rho = 0).
• cellMGGMM() no longer returns the original data matrix X.
• Documentation fixes.