pub fn finite_diff_gradient<S: Scalar>(
f: &dyn Fn(&[S]) -> S,
x: &[S],
g: &mut [S],
)Expand description
Compute a finite-difference gradient using the canonical central FD step
cbrt(S::EPSILON) * (1 + |x|), which scales with |x| to remain valid for
large-magnitude parameters. The additive scaling prevents silent zero-gradient
failures that occur when an unscaled step falls below the precision floor
of x + h == x (manifesting for |x| > ~5e7 at f64 precision).