Build a kernel-matrix accessor (internal) — .make_kernel

Wraps a fully-materialised kernel matrix in a list of accessor closures. The SMO solver and other iterative consumers read kernel values through this interface so that future phases can replace the underlying representation (e.g., the F6 LRU cache backed by Rcpp) without touching the consumers.

Usage

.make_kernel_accessor(Omega)

Arguments

Omega: Kernel matrix (N x N), already with any required jitter and symmetrisation applied by the caller.

Value

A list with components:

get_column(p): Returns column p of Omega (length-N numeric vector).
get_diag(): Returns diag(Omega) (length-N numeric vector). Cached at construction.
get_entry(p, q): Returns the scalar Omega[p, q].
get_matvec(v): Returns the matrix-vector product as.numeric(Omega %*% v) (length-N). Used for one-shot gradient refreshes; preserves BLAS efficiency.
n: Integer, equal to nrow(Omega).

Details

This F2 implementation is a thin wrapper over the materialised matrix: every closure delegates directly to base-R indexing or BLAS. The diagonal is computed once at construction (diag() is called a single time) and reused on every get_diag() call.

Symmetry assumption. The wrapped matrix is assumed symmetric. Both \(\Omega\) (a kernel matrix for symmetric K) and \(\Omega_s = \tfrac{1}{2}(\Omega + a\,\Omega^*)\) (used by the symmetric models when K satisfies Assumption 3) are symmetric throughout this package, so callers may use get_column(p)[k] in place of a row read Omega[p, k]. The SMO solver relies on this in its WSS3 step.