Analysis: Navier–Stokes Lost in Translation

Core Thesis: The authors argue that using AI to translate mathematical proofs from natural language (NL) into formal systems like Lean does not inherently validate the original NL proof, as the translation process often fails to be "semantically faithful."

📄 Document Metadata

---

🛠️ The Autoformalisation Pipeline

The process of verifying AI-generated mathematical claims typically follows this workflow:

The authors contend that the transition from A $\rightarrow$ B is the critical point of failure. While the mechanical checker can confirm that the Lean code is logically sound, it cannot confirm that the Lean code actually represents the original NL argument.

---

⚠️ The Complexity of Semantic Faithfulness

The central problem is the resolution of ambiguities inherent in mathematical natural language. The authors quantify this difficulty using the Solvability Complexity Index (SCI).

Complexity Comparison

The difficulty of ensuring a semantically faithful translation is astronomically higher than standard computational hurdles:

Key Insight: Because $\text{SCI} = \infty$, providing a guaranteed faithful autoformalisation is informally harder than any computable problem.

---

🔍 Case Study: Navier-Stokes Blow-up

The paper specifically examines OpenAI's announced proof regarding the blow-up of solutions to the Navier-Stokes equations.

The Lean verification proves the NL proof is correct. $\leftarrow$ This is the misconception the authors debunk.

Findings from the Analysis:

• AI translated the NL text into Lean.
• The Lean code passed mechanical verification.
• The Lean proof actually corresponds to the logic used in the NL proof. $\rightarrow$ FAILED

The authors demonstrate that the formalised version in Lean is a mismatch for the natural language argument, meaning the "verification" is an illusion.

---

💻 Technical Summary

To illustrate these failures, the authors provide several examples of "mistranslations." In these cases, the AI creates a valid formal proof, but it is a different proof than the one written in English.

-- Conceptual example of a mismatch
-- NL Proof: "Since X is true, Y follows via Lemma A"
-- AI Translation: "X is true, therefore Y is true (via Lemma B)"
-- Result: Lean says 'True', but the NL logic was bypassed.

Final Conclusion

The paper warns that we cannot rely on AI autoformalisation as a "black box" for truth. If the translation is not semantically faithful, the formal verification is irrelevant to the original claim.

---

Related Links & Tools:

• View PDF: arXiv:2610.08144
• Explore via: alphaXiv, scite.ai, Connected Papers

Figure 1: Representation of the SCI hierarchy showing the position of NL translation relative to the Halting Problem.