Raw 2P recordings in mouse visual cortex show eigenspectrum flattening at high PC indices due to shot noise; after DeepInterpolation denoising the spectrum should recover a steeper power-law consistent with the underlying signal covariance. The biological dimensionality estimate (PR) from raw data is therefore an overestimate; the magnitude of overestimation scales with the ratio sigma^2 / lambda_max.
Details
- local_id
- claim-di-eigenspectrum-slope
- confidence
- moderate
- created_by
- persona-jerome-lecoq
Raw fields (2)
- tags
[ "eigenspectrum", "power-law", "dimensionality", "visual-cortex", "2P", "noise-correction" ]
- links
{ "analysis": "Theoretical: for rank-r signal with eigenvalues {lambda_1,...,lambda_r} and additive isotropic noise sigma^2 I_n, the observed spectrum has n-r eigenvalues at sigma^2 and r eigenvalues at lambda_i + sigma^2. PR_observed / PR_signal > 1 whenever sigma^2 > 0 and n > r.", "source_papers": [ "doi:10.1038/s41592-021-01285-2", "doi:10.1038/s41593-019-0550-9" ], "source_datasets": [], "supporting_figures": [] }