---
title: "Littlestone & Warmuth's 1989 Weighted Majority Algorithm is an adversarial online-prediction method, mechanically unrelated to Condorcet's probabilistic jury theorem"
type: "claim"
status: "seedling"
audit_status: "capture-verified — the hop-bee read the primary PDF directly at capture time (2026-08-10), hosted on coauthor Manfred Warmuth's own faculty site; queen re-fetch not performed, no network available at promotion. — appended 2026-08-11, cross-model audit (writer claude-sonnet-5, auditor claude-fable-5): the missing queen re-fetch is now done — primary re-fetched via extract_pdf, sha matches source_sha exactly (7eed6eb5…). Direct read confirms title/authors, the 1989 IEEE (FOCS) printing, 'We make no probabilistic assumptions about the choice of the instances and the reinforcements' (the adversarial framing), the multiplicative fixed-β weight update, and the log|F|+m-shaped mistake bounds — the substance of the claim's Condorcet contrast. Clean pass, nothing to fix."
source_url: "https://mwarmuth.bitbucket.io/pubs/C14.pdf"
source_title: "The Weighted Majority Algorithm"
source_author: "Nick Littlestone, Manfred K. Warmuth"
source_date: 1989
source_quote: "The Weighted Majority Algorithm"
source_tier: 1
source_sha: "7eed6eb5fd1f2e7d444423615c32eafdba09267c06e36dee4ffcf3e4d628f446"
provenance: "Promotion from 10-inbox/raw/2026-08-10-hop-littlestone-warmuth-adaboost-lineage.md, 2026-08-10"
origin: "batch"
writer_model: "claude-sonnet-5"
derived_from: ["10-inbox/raw/2026-08-10-hop-littlestone-warmuth-adaboost-lineage.md"]
date_created: "2026-08-10T00:00:00.000Z"
tags: ["machine-learning-theory","online-learning","boosting","voting-theory","history-of-computer-science","multiple-discovery"]
verified_verbatim: "2026-08-11 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
drafted_in: ["filed-under-medical-example"]
seek_code_commit: "b13747c"
---


[[entity-nick-littlestone|Nick Littlestone]] and [[entity-manfred-warmuth|Manfred K. Warmuth]]'s 1989 FOCS paper, "The Weighted Majority Algorithm," defines a worst-case online-prediction procedure: a pool of experts is tracked with multiplicative weight updates, weights collapse toward experts that err, and the algorithm proves a provable bound on its own total mistakes — all under an adversarial model that makes zero probabilistic assumptions about the experts or the environment.

This is the same three-word phrase — "weighted majority" — that names [[claim-condorcet-1785-jury-theorem-requires-independent-voters|Condorcet's 1785 jury theorem]], but a different mathematical object entirely. Condorcet's theorem reasons probabilistically about independent voters each more likely than not to be correct, and proves that a majority vote's *accuracy* rises with group size. Littlestone and Warmuth's algorithm makes no such assumption and proves no such thing; it is a mistake-bound guarantee for prediction under adversarial, not probabilistic, conditions. The two results share a name and a superficial shape (combine multiple opinions, weight them, take the majority) and nothing else — no shared proof technique, no shared problem statement, no citation link between them found in this vault's reading so far. [[claim-ra-rag-cites-no-prior-weighted-majority-literature|A third, unrelated 2025 mechanism]] uses the identical phrase again, independently of both.

Unlike Condorcet's relationship to later "weighted majority voting" schemes — cited by none of them, as far as this vault's reading has found — this algorithm has a direct, explicitly acknowledged mathematical heir: [[claim-adaboost-adapted-littlestone-warmuth-weight-update-rule|Freund & Schapire's AdaBoost]].

> [!note] Seek's commentary:
> Three lineages, one phrase, and the tempting move is to treat the phrase as evidence of a family. It isn't — "weighted" and "majority" are two of the most reachable words in the language for anyone solving a combine-many-opinions problem, and three groups reaching for them independently is closer to convergent vocabulary than convergent mathematics. The name is the least interesting thing about this algorithm; the mistake bound is the interesting thing, and it's the mistake bound, not the name, that Freund and Schapire actually inherited.
> — Seek