Build a sparse cost graph from neighbour indices/distances (CSR+CSC) — uot_build_sparse_cost_knn

Helper for CPU pipelines: given neighbor indices and distances (e.g. from `RANN::nn2`), compute a sparse bipartite cost graph between source nodes and a fixed template.

Usage

uot_build_sparse_cost_knn_cpp(
  nn_idx,
  nn_dists,
  n_cols,
  lambda_anat = 1,
  F = NULL,
  G = NULL,
  lambda_feat = 0,
  max_dist,
  extra_i = NULL,
  extra_j = NULL,
  extra_dists = NULL,
  Y_template = NULL,
  prior_map = NULL,
  lambda_prior = 0,
  prior_sigma = NA_real_
)

Arguments

nn_idx: Integer matrix (n x k) of 1-based neighbor indices into the template support.
nn_dists: Numeric matrix (n x k) of Euclidean distances aligned with `nn_idx`.
n_cols: Number of template nodes (m).
lambda_anat: Weight for anatomical distance-squared term.
F: Optional source features (n x D).
G: Optional template features (m x D).
lambda_feat: Weight for feature distance-squared term.
max_dist: Optional maximum distance threshold; edges with `nn_dists > max_dist` are dropped.
extra_i: Optional 1-based row indices for additional edges.
extra_j: Optional 1-based column indices for additional edges.
extra_dists: Optional distances for additional edges (aligned to `extra_i`, `extra_j`).
Y_template: Optional template coordinates/features (m x d). When provided together with `prior_map` and `lambda_prior > 0`, adds a soft diagonal/identity bias term based on squared distances in template space.
prior_map: Optional integer vector (length n) giving the preferred template index (1..m) for each source node. Use NA for “no prior” on a row.
lambda_prior: Nonnegative weight for the soft prior term.
prior_sigma: Optional positive scale (in the units of `Y_template`). When finite, the prior penalty uses `||y_j - y_pref||^2 / prior_sigma^2`.

Value

A list with CSR fields `row_ptr`, `col_idx`, `cost` and CSC fields `col_ptr`, `row_idx`, `cost_csc`, plus dimensions.