A projected bound shortened 70 to 110 Hz power guards in 400 SEEG pulses under stated input error balls

We tested 400 public SEEG pulse epochs from five patients. The feature was causal 70 to 110 Hz power in 62-sample windows. Input errors could occur only in 25 samples around each pulse, under a stated Euclidean error ball. All methods used the same nominal voltage, error budget, and 5% earlier-power target. The projected bound shortened every guard against the voltage triangle bound. The median paired reduction was 7.8125 ms. The projected and sharp power guards matched in 345 of 400 pulses. They differed by at most one sample in 391 pulses. The sharp solve gave a further median gain of zero and a maximum primary gain of 4.3945 ms. We supply a proof, attaining errors, exact epochs, source hashes, code, tables, and four figures. All 30 fresh raw files passed their hashes. Fresh extraction and a full numerical replay reproduced the saved data and results. Guards cover a finite grid and use later observed voltage. They are retrospective processing limits. This study does not validate the artifact budget or measure neural recovery, decoder performance, or device performance.

SEEG · causal filtering · power error bounds · trust-region optimization · brain-computer interfaces · signal processing