H.B. Keller Colloquium
Philip B. Stark is Professor of Statistics at University of California, Berkeley where he developed the university's first online course. He has published research on the Big Bang, causal inference, the census, earthquake prediction, election auditing, food web models, the geomagnetic field, geriatric hearing loss, information retrieval, Internet content filters, nonparametrics, the seismic structure of Sun and Earth, spectroscopy, spectrum estimation, and uncertainty quantification for computational models of complex systems. He has consulted for the U.S. Departments of Agriculture, Commerce, Housing and Urban Development, Justice, and Veterans Affairs; the Federal Trade Commission; the California and Colorado Secretaries of State; the California Attorney General; and the Illinois State Attorney. He has testified to Congress and the California legislature, and in litigation concerning employment, environmental protection, equal protection, lending, intellectual property, jury selection, import restrictions, insurance, natural resources, product liability, trade secrets, and advertising. He received his AB from Princeton University and his Ph.D. from UCSD.
Any method of tabulating votes—by computer or by hand—can err or be subverted. If there is a trustworthy paper record of the votes, a risk-limiting audit (RLA) can provide a high probability that if any reported winner did not really win, that error will be corrected before the result is certified. RLAs are recommended as best practice by entities including NASEM and are required or authorized in 15 U.S. states. There are RLA methods for almost every social choice function used in political elections: plurality (single and multi-winner), supermajority, instant-runoff voting (IRV), STAR-Voting, proportional representation, and all scoring rules. The most efficient extant RLA methods reduce checking whether the reported winners really won to multiple instances of the same problem: testing (sequentially) the null hypothesis that the mean of a finite list of bounded, nonnegative numbers is not greater than 1/2. Each test amounts to tracking a gambler's fortune in repeated bets that are fair or subfair if the null is true: a nonnegative supermartingale. By Ville's inequality, unless the mean is greater than 1/2, the chance the auditor multiplies its initial stake by k before going broke is at most 1/k. When reported outcomes are correct, a well-designed RLA typically can confirm who won hundreds of overlapping contests by inspecting only a small fraction of cast ballot cards.
Add this event to your calendar.