Diagonal shrinkage gave lower next minute beta loss than an equal contact power correlation target in 19 SEEG records

We tested two covariance targets on public SEEG from 19 patients in OpenNeuro ds003876 version 1.0.2. Adjacent bipolar channels shared source contacts. One target was diagonal. The other target used minus one half correlation between channels with a shared contact in opposite roles. This pattern follows from independent zero mean contacts with equal power. Both models chose one of eight weights within the first minute. Diagonal shrinkage gave lower Gaussian beta loss in the next minute in all 19 records. Its mean advantage was 0.528144 nats per complex channel and frequency bin. A patient bootstrap gave a descriptive 95 percent interval of 0.430156 to 0.631308. Reverse time order and a contact quality screen retained the direction. Disjoint pairs gave identical models and scores. This exploratory result applies to these records, the 13 to 29.5 Hz bins, the fixed targets, and the tested grid. It does not establish neural connectivity or clinical accuracy. The package contains code, contact Fourier data, score grids, source metadata, retrieval recipes, and verification reports.

SEEG · spectral covariance · covariance shrinkage · beta band · bipolar channels · prediction loss · negative result