Method
NumCosmoXcorKernelIntegrandset_tolerances
Declaration [src]
void
nc_xcor_kernel_integrand_set_tolerances (
NcXcorKernelIntegrand* integrand,
gdouble reltol,
gdouble peak_epsilon
)
Description [src]
Records the criterion integrand was fitted to, in the two parts it actually
has. ncm_function_sample_set_refine() accepts a point when
$$ \Vert f - \tilde f \Vert_2 \le \mathrm{reltol} \Vert f \Vert_2 + a \Vert f \Vert_2^\mathrm{max} $$
so the two govern different regions: the floor is added, not maxed, and is
scaled to the peak, which leaves reltol biting only where the function is
within a few orders of that peak. Keeping them apart matters downstream —
they reach a $C_\ell$ differently, the floor through the product of two
closures. See nc_xcor_compute_full().
A spline-backed integrand is a fit, not the function, and the quadratures
that consume it have no other way to learn how good a fit.
NC_XCOR_METHOD_KERNEL_EXACT integrates it exactly, so this is the whole of
its error budget.