# Estimating duplicate payload: choose the sampling unit first A tiny occurrence sample can miss duplication even when every occurrence has the same chance of selection. That does not make all sampling estimators invalid: the selection design, estimator and target quantity must be specified together. ## Exact constructed example Consider six equal-length blocks: A,A,A,A,B,C. Original payload is 6 byte-units; unique payload is 3. The reduction factor is 2 and the saved fraction is 50%. There are 15 equally likely two-position subsets. Six contain two copies of A, giving a within-sample saved fraction of 1/2. The other nine contain two distinct identities, giving zero. The average is therefore (6/15)*(1/2) = 1/5, or 20%. Extrapolating that within-sample percentage underestimates the actual 50% saving. This is an exact counting result, not a claim about the bias of every possible estimator built from occurrence samples. For a population of N equal-length occurrences, identity j appearing f_j times, and a uniform k-position subset without replacement, the expected number of distinct sampled identities is: E[D_sample] = sum_j (1 - choose(N-f_j, k) / choose(N, k)) E[naive saved fraction] = 1 - E[D_sample] / k Use choose(a,k)=0 when a