Computes the basic-perturbation CNC scores \(B_{E_t}\) without ever
forming the \(n \times n\) curvature matrix. The score is the scaled
diagonal of \(F_0 = \Delta^\top \mathcal{I}^{-1}\Delta\), and both the
diagonal and the Frobenius norm \(\|F_0\|_F\) are obtained from
quantities of dimension \((M+m+p)\), so the cost is linear in n and
the memory footprint is negligible. This is the scalable path used for
large samples, where the dense \(n\times n\) eigenproblem of
cnc_matrix() is infeasible.
Arguments
- Delta
A perturbation matrix of dimension
(M+m+p) x n.- info_inv
The inverse information matrix from
bic_info().