Control settings for multiscale manifold alignment backends

Usage

multiscale_manifold_align_control(
  enabled = FALSE,
  backend = c("spectral", "randomized_dwt"),
  operator_mode = c("compressed_graph_core"),
  rank_per_domain = 128L,
  svd_tol = 1e-06,
  eigen_tol = 1e-08,
  max_levels = 8L,
  eps_rank = 0.001,
  eigen_solver = c("lobpcg", "lanczos"),
  eigen_k_max = 512L,
  knn = 12L,
  sigma = NULL,
  cross_edge_weight = 1,
  mnn_k = 5L,
  max_pairs_per_label = 100L,
  max_unsup_pairs_per_pair = 200L,
  randomized_sketch_size = 256L,
  randomized_oversample = 16L,
  randomized_power_iters = 1L,
  shift_invert_delta = 0.001,
  cg_tol = 1e-06,
  cg_max_iter = 500L,
  seed = 1L,
  verbose = FALSE
)

Arguments

enabled: Logical; enable multiscale mode.
backend: Character backend name. One of `"spectral"` or `"randomized_dwt"`.
operator_mode: Character operator mode. Currently `"compressed_graph_core"`.
rank_per_domain: Integer per-domain truncation rank for compressed operators.
svd_tol: Numeric positive tolerance for truncated SVD decisions.
eigen_tol: Numeric positive tolerance for non-trivial eigenvalues.
max_levels: Integer maximum hierarchy depth.
eps_rank: Numeric relative truncation threshold in `(0, 1]`.
eigen_solver: Character eigensolver selector. One of `"lobpcg"` or `"lanczos"`.
eigen_k_max: Integer cap on eigenvectors for spectral hierarchy.
knn: Integer nearest-neighbor graph degree for within-domain graphs.
sigma: Optional positive scalar graph bandwidth. If `NULL`, auto-tuned per domain.
cross_edge_weight: Numeric non-negative weight for cross-domain correspondence edges.
mnn_k: Integer nearest-neighbor count for unsupervised MNN correspondences.
max_pairs_per_label: Integer cap per label/pair when building label correspondences.
max_unsup_pairs_per_pair: Integer cap per domain pair in unsupervised MNN mode.
randomized_sketch_size: Integer sketch size for randomized backend.
randomized_oversample: Integer oversampling budget for randomized backend.
randomized_power_iters: Integer number of subspace power iterations.
shift_invert_delta: Numeric positive regularization for shift-invert style operators.
cg_tol: Numeric positive tolerance for iterative solves.
cg_max_iter: Integer maximum iterations for iterative solves.
seed: Integer random seed.
verbose: Logical; print backend progress messages.

Value

A list with class `multiscale_manifold_align_control`.