Using Lean 4, an AI system collaborates with a mathematician to formalize the Vlasov equation. The project highlights AI's role in structured proof development. Formalizing mathematical proofs has entered a new era. A recent experiment using the Lean 4 proof assistant showcases how AI, guided by a mathematician, can transform LaTeX documents into verified formalizations. This isn't just about digital conversion. It's about ensuring mathematical statements rely solely on foundational axioms.
The Game of Formalization #
Imagine a game where a mathematician directs an AI to achieve a well-defined goal. The objective? To convert a complete research result into Lean without any 'sorry' placeholders, essentially proving theorems purely from axioms. The entire development compiles, marking a win when the target theorems pass strict verification.
But there's more. To pass a second test, the formalization must produce reusable, self-contained mathematics that can integrate with broader libraries. In this instance, the focus was on the nonlinear Vlasov equation, approached through Dobrushin's mean-field perspective. Existence, uniqueness, and stability estimates were all on the table, along with a mean-field limit and a Lagrangian weak solution principle.
AI's Role in the Process #
In this setup, the human didn’t write the proofs. The mathematician scoped definitions, guided decomposition, and identified library gaps. The AI executed these tasks. This division of labor raises an intriguing question: Are we approaching a future where AI takes over more intricate parts of mathematical research?
Optimal transport machinery emerged as a significant artifact of this process, including properties of the Wasserstein-1 metric and Kantorovich-Rubinstein duality. These form a distinct, self-contained layer comprising about a sixth of the development, with 49 out of 299 declarations, all against Mathlib alone. Notably, the interface consisted of 22 declarations, ensuring no reverse dependencies.
Timeframe and Observations #
It’s worth noting that the core theorems were formalized in a week, while the entire development took about a month. These timelines underscore the efficiency of such AI-human collaborations. Yet, these findings are observations from a single instance. They don't establish general laws but offer a glimpse into the potential future of mathematical formalization.
As AI systems evolve, they might redefine the boundaries of mathematical research. But, can they ever replace the nuanced judgment of human mathematicians? For now, AI acts as an articulate assistant, leaving the final judgment to human expertise.
Get AI news in your inbox
Daily digest of what matters in AI.