validate — NLL Perturbation Analysis

The validate subcommand performs profile likelihood analysis by perturbing optimized parameters and measuring the NLL response. This helps verify that the optimizer has converged to a true minimum.

Basic Usage

r_analysis validate \
    -n 64 \
    -r "kmeans_BD10000" \
    -ird results/ \
    --steps 5 \
    --scales 1e-3 1e-4

Arguments

Flag	Type	Default	Description
--steps	int	5	Number of perturbation steps in each direction
--scales	float (list)	1e-3 1e-4	Perturbation scale factors
--noise-selection	str	min-value	Noise realization: min-value, min-nll, or integer index

Perturbation Targets

Control which parameters are perturbed:

Flag	Default	Description
--perturb-beta-dust	all	Which \(\beta_d\) indices to perturb
--perturb-temp-dust	all	Which \(T_d\) indices to perturb
--perturb-beta-pl	all	Which \(\beta_s\) indices to perturb

Perturbation spec values:

• all — perturb all indices

• -1 — skip this parameter entirely

• 0:30 — slice notation (indices 0 through 29)

• 0,1,2 — specific indices

• max — only the index with the largest value

• min — only the index with the smallest value

Plot Options

Flag	Type	Default	Description
--plot-type	str (list)	nll-grad	Plot types: nll-grad, nll, grad, grad-maps-{idx}
--aggregate	flag	False	Overlay all runs on the same plot
--no-vmap	flag	False	Use for-loop instead of vmap (less memory)
-t, --titles	str (list)	auto	Custom titles for each plot
-o, --output	str	plots/	Output directory
--output-format	str	png	Format: png, pdf, show
--font-size	int	14	Font size

Plus all common arguments.

Example: Validate specific parameters

r_analysis validate \
    -n 64 -r "kmeans_BD4000_GAL020" -ird results/ \
    --perturb-beta-dust "0:50" \
    --perturb-temp-dust "-1" \
    --perturb-beta-pl "max" \
    --plot-type nll grad \
    --aggregate