Skip to contents

Slow delete-one jackknife for rank-weighted L-stat rules (fixed ranks)

Source: R/jackknife-rank-slow.R
jackknife_slow_rank.Rd

A reference implementation that recomputes aggregates per deletion using the provided neighbor ranks. This does not recompute ranks after deletion (useful as a correctness oracle and for testing).

Usage

jackknife_slow_rank(ranks, a, phi, normalize = TRUE)

Arguments

ranks

Integer matrix (T x (s+1)): for each target t, ranks[t, k] gives the positive 1-based index of the k-th neighbor, for k = 1, ..., s+1. (Must exclude t itself if targets are a subset of reference.)

a

Finite numeric length-s vector of L-stat coefficients \((a_1,\dots,a_s)\).

phi

Either a finite numeric matrix (n_ref x q) with row j equal to \(\phi(j)\), or a function phi(j, t) returning a finite length-q numeric vector.

normalize

Scalar logical; if TRUE (default) divide aggregates by sum(a).

Value

A list with:

theta0

numeric length-q baseline aggregate

theta_loo

numeric matrix (n_ref x q) of delete-one aggregates

info

list with fields s, T, q, normalize

Details

Like jackknife_fast_rank(), this helper targets compact-support finite-rank L-stat rules. It is not the internal engine used by the user-level exact DNN wrappers, which instead update the full rank matrix under deletion and reevaluate the DNN rank-weight rule.

Works with:

  • phi as a numeric matrix (n_ref x q): row j equals \(\phi(j)\).

  • phi as a function function(j, t) that returns a length-q numeric, allowing target-dependent contributions.

Examples

set.seed(1)
n <- 20
p <- 3
s <- 4
X <- matrix(rnorm(n * p), n, p)
# toy ranks: nearest-neighbor indices precomputed elsewhere (T x (s+1))
# here we fake something deterministic just for the example
ranks <- cbind(matrix(rep(1:s, each = n), n, s), s + 1L)
y <- rnorm(n)
a <- rep(1, s)
out <- jackknife_slow_rank(ranks, a, cbind(y))
str(out)
#> List of 3
#>  $ theta0   : num 0.77
#>  $ theta_loo: num [1:20, 1] -0.0162 0.594 0.4118 0.5772 0.77 ...
#>  $ info     :List of 4
#>   ..$ s        : int 4
#>   ..$ T        : int 20
#>   ..$ q        : int 1
#>   ..$ normalize: logi TRUE