Estimate variance using Lanczos method — lanczos

Estimate squared standard error (variance) for a parameter or derived quantity without forming the Hessian matrix directly

Usage

lanczos_variance(
  Hq,
  q,
  x = attr(Hq, "env")$x0,
  k = c(25, 30),
  min_spectral_ratio = 1e-10,
  orthogonalize = FALSE
)

Hq: function that calculates the product H %*% q given probe q and parameters x
q: vector to use when calculating variance, either an indicator for a single parameter, or a gradient evaluated at the MLE for a derived quantity
x: parameter vector used when calculating the Hessian matrix
k: can be a vector
min_spectral_ratio: is the ratio of minimum to maximum ratio, where values < minimum are truncated to minimum (min_spectral_ratio=0 disables truncation)
orthogonalize: Whether to do two-pass Gram-Schmidt re-normalization (much slower)