Analytic first-order bootstrap in the design space — dkge_bootstrap

Uses the stored full eigendecomposition to apply first-order perturbations for each bootstrap draw. When the perturbation exceeds the validity region, the method falls back to the exact multiplier bootstrap for that replicate.

Usage

dkge_bootstrap_analytic(
  fit,
  contrasts,
  B = 1000L,
  scheme = c("poisson", "exp", "bayes"),
  ridge = 0,
  align = TRUE,
  allow_reflection = FALSE,
  seed = NULL,
  transport_cache = NULL,
  mapper = "sinkhorn",
  centroids = NULL,
  sizes = NULL,
  medoid = 1L,
  voxel_operator = NULL,
  perturb_tol = 0.2,
  gap_tol = 1e-06,
  ...
)

Arguments

fit: A fitted `dkge` object.
contrasts: Contrast specification accepted by [dkge_contrast()].
B: Number of bootstrap replicates.
scheme: Multiplier distribution (`"poisson"`, `"exp"`, or `"bayes"`).
ridge: Optional ridge added to the reweighted compressed covariance.
align: Logical; when `TRUE` the resampled bases are aligned to the baseline basis via K-Procrustes before contrasts are evaluated.
allow_reflection: Passed to [dkge_procrustes_K()] when aligning bases.
seed: Optional random seed for reproducibility.
transport_cache: Optional cache from [dkge_prepare_transport()].
mapper: Mapper specification used when a cache is not supplied.
centroids: Optional centroids overriding those stored on the fit.
sizes: Optional list of cluster weights passed to [dkge_prepare_transport()].
medoid: Medoid index used during transport.
voxel_operator: Optional matrix mapping medoid values to voxels.
perturb_tol: Maximum absolute coefficient tolerated in the eigenvector perturbation; larger changes trigger a fallback to the full eigensolve.
gap_tol: Minimum eigen-gap tolerated (in absolute value) before triggering a fallback to the full eigensolve.
...: Additional arguments forwarded to [dkge_prepare_transport()] when the transport cache needs to be built.

Value

Same structure as [dkge_bootstrap_qspace()] with additional metadata on the number of fallbacks used.