Papers74°

AlphaProof Nexus 用 Lean 形式化证明解决量子光学悬而未决的 no-go 定理

Why we picked it

DeepMind 和图宾根大学的 AlphaProof Nexus 在 Lean 里搜出了正式证明,回答了 2017 年提出的量子光学问题:高维 GHZ 态用线性光学做不出来。证明由编译器机器验证,结论更可靠。做 AI for Science 或形式化方法的人应该看看这篇。

University of Tübingen 与 Google DeepMind 在 arXiv(2611.26300)发布论文,展示 AI 智能体首次以机器可验证的证明解决一个开放物理问题。该问题源自 2017 年作者提出的量子光学难题:高维 GHZ 态(纠缠光子态)能否仅用线性光学和配对源构建。AlphaProof Nexus 给出的答案是否定的——对于任意 4 个及以上光子、维度不少于光子数的情形,该构建在数学上不可能。

Body

技术路线上,AlphaProof Nexus 在 Lean 定理证明器中搜索形式化证明,将前沿大语言模型与 Lean 编译器的反馈循环结合。由于证明是形式化的,Lean 可机器验证其正确性,无需人工审阅数学推导过程。作者随后人工解读了证明背后的物理含义,并在论文中给出物理解释。

这是形式化证明搜索在纯数学之外产出新结果的最清晰案例之一,表明 LLM 加形式验证的组合可以处理此前被认为需要人类物理学家直觉的问题。

Original · DAIR.AI

Can an AI agent settle an open physics question with a machine-checked proof?

This new paper from the University of Tübingen and Google DeepMind shows that it can.

The question comes from quantum optics. Physicists wanted to know whether high-dimensional GHZ states, a kind of entangled photon state, can be built with only linear optics and pair sources.

The authors first raised it in 2017.

The answer is no. For any even number of photons from 4 upward, with at least as many dimensions as photons, the construction is impossible.

AlphaProof Nexus found the proof. It searches for formal proofs in Lean, combining frontier language models with feedback from the Lean compiler. Because the proof is formal, a computer verifies it.

The authors then worked out what the proof means physically and explained it in the paper.

This is one of the clearest recent examples of formal proof search producing a new result outside pure math.

https://arxiv.org/abs/2611.26300

Chat with Paper: https://academy.dair.ai/papers/a-tensor-algebraic-no-go-theorem-for-high-dimensional-photonic-greenberger-horne-2611.26300

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk, focusing on Europe's AI governance and technology ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include AlphaProof Nexus, Lean, formal verification, Google DeepMind.

Source attribution: curated from DAIR.AI (@dair_ai).