Common average reference gave negative full residual correlations and positive covariance pseudoinverse scores in independent contact controls from 36 SEEG patients

We compared full residual correlation with covariance pseudoinverse scores after common average reference. We used public SEEG data from 36 patients, with 72 runs and three frequency bands. The reference and the network used the same contacts. Full adjustment for all other channels gave residual correlation -1 when residual variance was positive. Independent contact controls kept the measured variances and removed cross-contact covariance. All 1,102,868 pair evaluations in the 1–80 Hz controls had positive covariance pseudoinverse scores. Small ridge values instead gave scores close to -1. We also checked a valid contrast interpretation of the pseudoinverse scores. The known rank and contrast results are not claimed as new theory. The contribution is a public SEEG benchmark and a reusable audit tool. All 72 raw prefixes passed their hash checks. All 432 rebuilt covariance matrices matched the archive bit for bit. The package preserves code, data, proofs, results, source records, and instructions. The controls do not estimate the fraction of real neural connections or support a clinical claim.

common average reference · SEEG · partial correlation · covariance pseudoinverse · ridge regularization · contact covariance · neural networks · rank deficiency · contrast association · reproducibility