Score specification (S3) for U-stat kernels

A score (kernel) is a symmetric function evaluated on s-sized subsets. Provide either ptr (compiled fast path), or an R eval_batch, or an R eval.

Usage

new_score(
  name,
  order,
  setup = NULL,
  eval = NULL,
  eval_batch = NULL,
  eval_param = NULL,
  eval_param_batch = NULL,
  grad = NULL,
  grad_batch = NULL,
  fd_step = 1e-06,
  fd_scheme = "central",
  ptr = NULL,
  randomized = FALSE,
  symmetric = TRUE
)

Arguments

name: Character scalar.
order: Integer order s, or function(n, data, ctx) -> s.
setup: Optional function(data, ...) -> ctx object (list or externalptr).
eval: R fallback: function(idxs, data, ctx, rng) -> numeric(1).
eval_batch: R fast path: function(subs, data, ctx, rng) -> numeric(nrow(subs)).
eval_param: Optional parameter-aware scalar evaluator: function(idxs, data, theta, ctx, rng) -> numeric(1).
eval_param_batch: Optional parameter-aware batch evaluator: function(subs, data, theta, ctx, rng) -> numeric(nrow(subs)).
grad: Optional analytic scalar gradient: function(idxs, data, theta, ctx, rng) -> numeric(length(theta)).
grad_batch: Optional analytic batch gradient: function(subs, data, theta, ctx, rng) -> matrix(nrow(subs), length(theta)).
fd_step: Finite-difference step used by score_diff_path() when method = "fd" and no explicit step is supplied.
fd_scheme: Finite-difference scheme used by score_diff_path() when method = "fd" and no explicit scheme is supplied. One of c("central", "forward").
ptr: External pointer to compiled score (see vignette); if present, used by engine.
randomized: Logical: does the kernel draw randomness?
symmetric: Logical: permutation symmetry (engine assumes TRUE).

Value

An object of class ustat_score.

Details

If your score depends on a parameter vector (for example Z-estimation), use eval_param/eval_param_batch and optionally grad/grad_batch to enable derivative-aware workflows via score_diff_path().

Examples

sc <- new_score(
  name = "sum2",
  order = 2L,
  eval = function(idxs, data, ctx, rng) sum(data[idxs])
)
is_score(sc)
#> [1] TRUE