Reducing diagonal load increased strong negative partial correlation density in 80 to less than 150 Hz CAR matrices from 19 SEEG records

We tested diagonal loading of common average reference (CAR) covariance in 19 public SEEG records from OpenNeuro ds003876 version 1.0.2. We used two minutes per record and a band from 80 to less than 150 Hz, with 118 to 122 Hz excluded. A strong negative edge had a partial correlation at or below -0.9. Reducing the load from 0.01 to 0.000001 of mean variance raised the equal-patient mean edge fraction from 0.001141 percent to 94.735204 percent. The increase was 94.734 percentage points. Its 95 percent patient bootstrap interval was 84.216 to 100.000 percentage points. All patient changes were positive. One record had a much smaller response. A proof shows that every loaded full-order partial correlation approaches -1 under two assumptions. The CAR constraint must be the only zero mode, and the diagonal load must approach zero from above. Independent contacts with equal variance have the same limit. The package provides a finite-load bound and a diagnostic for the known CAR term in the precision diagonal. These scores cannot, by themselves, establish inhibitory brain connections. Clinical value was not tested. The package preserves compact inputs, edge scores, code, source versions, hashes, retrieval recipes, and checks. Fresh raw ranges from 19 of the 19 records exactly reproduced their input covariance arrays. Two new analysis processes gave exact numerical outputs. The cohort is shared with earlier AutoSci work. This study adds a different calculation and does not add independent patients.

common average reference · partial correlation · diagonal loading · stereoelectroencephalography · precision matrix · zero mode · OpenNeuro