Work

Counting Shots to Certify a Logical Error Rate: The Statistical Budget Behind a Suppression Claim

Year: 2026
Type: article
Abstract: A logical error rate is not measured the way a length is measured. It is inferred from counting failures in a finite number of trials, and the count's statistics dictate how many trials a claim costs. This paper prices certification. To measure a logical error rate at the level of one part in a million to ten per cent relative precision at ninety-five per cent confidence requires about 3.8e8 independent trials, in which about 384 failures are expected to occur; the failures, not the trials, are the information. The same arithmetic at 1.00e-3 needs only 3.8e5 trials, which is why early demonstrations certify easily and mature claims become campaigns. The paper extends the budget to suppression-factor estimation, where two rates must be counted and their uncertainties combine, and to the wall-clock cost of certification campaigns on slow platforms, where a statistically adequate run can take longer than the hardware stays calibrated. Verification is a consumable, and this paper adds it to the machine's budget.
Language: English
Subfield: Software
Open Access status: green
OpenAlex ID: W7233572415
FWCI:
Cites:
Cited by:
Related to: