# MinHash and deduplication denominators This is a finite arithmetic illustration, not a hashing-library benchmark, storage measurement or new MinHash theorem. There are no external dependencies. Run with Node.js: ```sh node --test model.test.mjs node --input-type=module -e 'import { examples } from "./model.mjs"; console.log(JSON.stringify(examples(), null, 2))' ``` The tests enumerate 225 ordered nonempty subset pairs of four identities under all 24 permutations. Each pair obeys the classical collision identity by exact integer cross-multiplication. All 256 pairs, including empties, also satisfy coordinatewise-minimum union. Enumeration is not a set of independent sampled coordinates and does not validate practical hash functions or uncertainty bounds. An empty/empty Jaccard comparison returns null rather than imposing a convention. The payload examples define existing identities A={a,b}. Incoming recipes are a,c or a,a,a,c. One size unit is one byte; a has either 1 or 9 bytes and b,c have 1 byte each. Identity a always denotes the same bytes within a case. Count every incoming occurrence whose identity already exists in A. The numerator/denominator pairs are 1/2, 9/10 and 3/4, while set Jaccard remains 1/3. These are logical payload reuse fractions, not measured physical space savings. No compression, metadata, replication, deletion, collision, false-match, index-coverage or I/O cost is modeled. Two shard pairs with the same local similarities can have different union similarities. This demonstrates why scores alone cannot recover union overlap. Compatible signatures retain information that the scalar scores discard. Primary evidence: - Broder et al., Min-Wise Independent Permutations (JCSS 2000), pp. 632-633: https://www.eecs.harvard.edu/~michaelm/postscripts/jcss2000.pdf - Broder, On the resemblance and containment of documents (SEQUENCES 1997), section 2: https://skeptric.com/resources/broder97resemblance.pdf - datasketch implementation, reviewed 2026-11-20, not executed here: https://ekzhu.com/datasketch/_modules/datasketch/minhash.html Source of truth: content/research/minhash in the repository. publish.mjs copies the three public files without changing their bytes; --check validates copies. The article imports examples() directly, so its table uses the tested values. ## Reviewed hash-family counterexample Choosing uniformly from the 20 permutations h(x)=(a*x+b) mod 5, with a in {1,2,3,4} and b in {0,1,2,3,4}, maps each distinct ordered input pair uniformly to the 20 distinct ordered output pairs. The test enumerates all those input/output cases. This is pairwise independence **for permutations** (Broder et al., equation 3), not independent outputs allowing collisions. Yet sets {0,1} and {0,2} have matching minima for 7 of these 20 functions, rather than the Jaccard value 1/3. The same mathematical construction therefore separates pairwise independence from the exact min-wise requirement in equation 4. It is a bounded arithmetic counterexample, not a measured bias in datasketch. The model returns all orderings so the count is independently inspectable.