A common average reference constraint explained strong negative partial correlation scores under small diagonal loading in 72 SEEG samples

We tested whether common average reference can make full-order SEEG partial correlation scores strongly negative. We used 72 fixed 20-second samples from 36 patients in OpenNeuro ds004100 version 1.1.3. For the 2–40 Hz band, the mean patient fraction of contact pairs with scores at most −0.9 was 88.7% at dimensionless diagonal loading 10⁻⁶. Its 95% patient bootstrap interval was 85.5% to 91.8%. The fraction was zero at loading 10⁻². An independent-contact covariance model gave 99.9978% at loading 10⁻⁶ after the same reference. A proof shows the minus-one limit when there is one null mode with a strictly positive vector and positive contrast eigenvalues. The bound gave a mean patient fraction of 84.6% for certified scores at most −0.9 at loading 10⁻⁶. All 72 inputs were rebuilt from fresh fixed source byte ranges and matched the saved files. Direct solves passed 1,080 checks. The package preserves an audit tool, proof, covariance data, source identities, code, results, and reproduction instructions. The result explains a processing constraint. It does not establish neural inhibition or a best reference or loading value.

common average reference · partial correlation · precision matrix · diagonal loading · stereo-EEG · epilepsy · reference constraint