Step 06Added in 2026 · recovery quality
One dataset said lasso wins. Do a hundred?
The 2021 report compared estimators on a single noisy dataset. Here every estimator is refitted on many realisations of the same sources, so each recovery score comes with an interval and each comparison becomes a paired test with an effect size.
For a true time course the score is the largest absolute correlation with any retrieved time course,
and the same for spatial maps. When the retrieved source with the matching index is the best match, this equals the matched score summed in 2021. The table under Figure 6.1 reports how often that happens. Penalties stay at their 2021 values throughout.
Using the 2021 parameters. Results match the report.
Noise seed 30034. Realisation 1 is the dataset from step 1, which is the 2021 dataset.
Penalties fixed at their 2021 values
Ridge λ = 0.5, both lassos ρ = 0.625, PCR ρ = 0.001. Maps are scored against the layout X was built with (the aligned convention).
Figure 6.1
Per-source recovery, with 95% intervals
Figure 6.2
Do the 2021 comparisons hold beyond one dataset?
Optional. Needs your own Anthropic or OpenAI key, and sends only the instructions and figure summary shown below.
Sensitivity
How much noise can each method take?
The assignment fixed the noise variances at 0.25 (temporal) and 0.015 (spatial). The grid below varies both. Least squares and ridge degrade smoothly, the maps with spatial noise and the time courses with temporal noise. The lasso is different. Its penalty was tuned at the 2021 noise level, and when temporal noise grows, standardising X shrinks the signal under the fixed threshold until the maps are wiped out.
30 cells. Every cell reuses the same seeded draws (seed 30034), scaled to its variances.
Estimator
Figure 6.3
Lasso (as submitted): recovery of the spatial maps across noise levels
Optional. Needs your own Anthropic or OpenAI key, and sends only the instructions and figure summary shown below.
Next · How it was done
Methods
Model, assumptions, evaluation design, decision records, model card and the AI use statement.