Lean formally verifies Riemann Hypothesis proof

A glowing, intricate lattice structure emerging from a dark, abstract surface, symbolising a mathematical proof verified through formal logic, with light tracing paths through interconnected nodes.

On 20 November 2026, mathematician Levent Alpöge released a complete formal proof of the Riemann Hypothesis as a Lean library of approximately 17.4 million lines of code. The proof, developed with assistance from the AI system Claude, also extends to the Generalized Riemann Hypothesis (GRH) for Dirichlet L-functions. Within 48 hours, the Lean and Mathlib communities, alongside independent researchers, completed a comprehensive verification process, confirming the proof’s correctness beyond reasonable doubt.

The verification relied on multiple independent tools and criteria. Crucially, the proof compiles without errors using Lean 4.33.1 with a pinned version of Mathlib (commit 0df444a3). No sorry, admit, or unsafe constructs were found. The five main theorems—including the classical Riemann Hypothesis and its generalisations—depend only on Lean’s three standard axioms: propext, Classical.choice, and Quot.sound.

Proof verification across multiple checkers

Independent proof checkers have accepted the formal proof in its exported form. The con-leche and con-ron checkers, both with formal soundness proofs relative to set theory, confirmed that the proof justifies the target statements. The Rust-based nanoda kernel also accepted the full export. The lean4lean verifier replayed all 175,187 compiled modules successfully, with only five requiring increased step limits due to computational complexity.

An independent sandboxed rebuild of the entire project—without network access and using read-only source—reproduced the same compiled output byte for byte. This reproducibility eliminates concerns about hidden dependencies or build-time manipulation. The exported proof, weighing over 61 GB, has been validated by multiple parties across different hardware and software environments.

Statement correctness and arithmetic implications

The proof targets Mathlib’s own definition of the Riemann Hypothesis, ensuring alignment with established formal mathematics. Two independent tools, con-leche --compare and con-ron --compare, confirmed that the theorem proved in Alpöge’s library exactly matches the challenge statement derived solely from Mathlib.

Further validation came from combining the proof with an independent formalisation of Robin’s and Lagarias’s criteria—arithmetical equivalents of the Riemann Hypothesis. When linked with the kimihiro64/Robin1984 library (commit acab1a3), the combined system proves both the Robin and Lagarias inequalities without additional axioms. These results, verified by con-leche, con-ron, and nanoda, confirm the proof’s implications in elementary number theory, independent of zeta function formalism.

What it means

For mathematicians, this marks the first machine-verified solution to one of the most profound open problems in mathematics. For the formal verification and AI-assisted research communities, it demonstrates the maturity of proof assistants like Lean in handling frontier mathematical reasoning. The collaboration with Claude suggests a new model for human-AI mathematical discovery.

For businesses and institutions in the UK, UAE, US, and Kazakhstan investing in formal methods or cryptographic research, this event underscores the strategic value of proof automation. The Riemann Hypothesis underpins conjectural security assumptions in some number-theoretic cryptography. While no immediate break is implied, the ability to formally verify such deep results strengthens trust in mathematical foundations used in security-critical systems.

Organisations relying on long-term cryptographic assumptions should monitor developments in formal verification, as future proofs may invalidate or reinforce current standards. Investing in formal methods capability now allows proactive adaptation to such shifts.

The full proof is publicly available for replication. Instructions include building from GitHub, verifying SHA-256 hashes, checking axioms, and replaying with external checkers like lean4checker. No license was specified by the author, so reuse remains legally ambiguous until clarification.

Sources

Leave a Reply

Your email address will not be published. Required fields are marked *