DNN weights (exact full-U), distance-based ranking — weights_dnn

Exact DNN weights are derived from the full rank permutation for each target. They are not constructed by keeping only the first s observations. For the exact finite-population rule, the support in rank order is j <= n - s + k; in particular, for the paper-faithful k = 1 DNN this becomes j <= n - s + 1. Fixed top-k kernels are implemented for any 1 <= k <= s. The fixed-k extension of the theory is essentially direct, but the currently written manuscript derivations cover the k = 1 case.

Usage

weights_dnn_exact(
  X,
  X0 = NULL,
  x0 = NULL,
  distance = "l2",
  k,
  s,
  ctx = NULL,
  return = c("W")
)

Arguments

X: numeric matrix (n x p)
X0: numeric matrix (m x p); mutually exclusive with x0
x0: numeric length-p single target (convenience for m=1)
distance: distance spec or registered name (e.g., "l2")
k: integer; neighbors used in each s-subset (1 <= k <= s)
s: integer; U-statistic order / subset size
ctx: optional context
return: character vector of extras to include; any of c("W","ord","ranks","dx")

Value

a weights_result with W and metadata

Examples

set.seed(1)
x_mat <- matrix(rnorm(20), 10, 2)
x0_mat <- matrix(rnorm(6), 3, 2)
out <- weights_dnn_exact(x_mat, x0_mat, distance = "l2", k = 1L, s = 3L, return = "W")
dim(out$W)
#> [1] 10  3