DiffSum - A Simple Post-Election Risk-Limiting Audit

DiffSum is approximately as efficient as Bravo— compare (3) with the estimate 2 ln(1/α)/m for Bravo [2] (here α is the risk limit). Moreover,DiffSum does not need an initial estimate of the vote shares, and Bravo is inefficient when this estimate is inaccurate. Error rate: The error rate bounds given in Step 1 are based on extensive simulations for δ = 0 to 4, d = 3 to 7, n = 10, and c = d+δ. We measured the error rate over 10,000 simulated elections in each case. Each simulation estimated the error rate when the election was a tie, a worst-case scenario; with more realistic margins the error rate drops dramatically, so that in practice even c = d should give very reliable audits.