NEWS

---

neurocluster 0.1.0.9000

Performance fixes

• The sparse commute solver now uses eigencore (>= 1.0.3) from CRAN as a required import, replacing optional RSpectra. Large masks no longer need a separate solver installation. Sparse solves use a fixed local RNG seed, check original-Laplacian residuals and orthogonality, and disable implicit dense fallback. The minimum R version is now 4.1, as required by eigencore.
• rena_plus() (cluster4d(method = "rena_plus")) is no longer quadratic in the number of voxels. The ReNA coarsening walked each row of the column-compressed adjacency matrix with Armadillo's row iterator, which costs O(N) per row; it now walks the columns of the transpose, which visits the same neighbours in the same order. Labels are bit-identical to the previous implementation. On the 29,532-voxel neuroim2 mask (20 timepoints, n_clusters = 200) run time drops from 118 s to 0.5 s (#3).
• commute_cluster() (cluster4d(method = "commute")) gains eigen_solver = c("auto", "dense", "sparse"). The dense path formed an N-by-N Laplacian and a full eigen() (O(N^2) memory, O(N^3) time; about 7 GB per dense copy on a 30,000-voxel mask) although only the ncomp smallest eigenpairs are used. Above 3,000 voxels "auto" now computes only those eigenpairs with shift-invert Lanczos; at or below 3,000 voxels results are unchanged. The initial RSpectra implementation finished the 29,532-voxel mask in about 5 s (#3); that timing does not characterize the replacement eigencore backend. Dense and sparse solvers are compared against numerical oracles in the tests.

Continuous integration

• The Native sanitizers workflow's asan job had failed since August 2026 without compiling the package: tools/run-native-sanitizers.sh preloaded the ASan runtime into GNU make, whose own strndup() calls ASan reported as heap-buffer-overflows under strict_string_checks=1. The runtime is now preloaded only into the R process that runs the smoke cases, and the script fails if ASan is not active there. No overflow was found in package code (#2).

Documentation and interface fixes

• ?cluster4d now lists all 13 method values (adding corr_slic, rena_plus and commute) and no longer labels g3s as "recommended"; method order is stated not to be a ranking. The "Algorithm Comparison" table now covers every method, reports measured speed tiers (SNIC is listed as slow on whole-brain masks because its k-NN gradient seeding dominates run time), and states which common arguments each method uses. The quick cluster4d() example now runs during checks (#1).
• cluster4d_rena() and cluster4d_rena_plus() no longer show an inactive spatial_weight = 0.5 default in their usage. spatial_weight is no longer a formal argument; supplying it (now through ...) is still an error. Code that passed a fourth positional argument intending spatial_weight now binds it to max_iterations (cluster4d_rena()) or r (cluster4d_rena_plus()) (#1).