⚛️ quantum physics

A Machine-Verified Proof of a Quantum-Optimization Conjecture

This paper reports a machine-verified resolution of the decade-old Farhi-Goldstone-Gutmann conjecture regarding the QAOA approximation ratio on the ring of disagrees, achieved through a collaborative feedback loop between the Claude Fable 5 language model and the Lean 4 proof assistant that uncovered a hidden dynamical symmetry to construct the proof.

Original authors: Uri Kol, Maor Ben-Shahar, Kfir Sulimany, Dirk Englund

Published 2026-06-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Uri Kol, Maor Ben-Shahar, Kfir Sulimany, Dirk Englund

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: A Robot Mathematician Solves a 10-Year Mystery

Imagine a group of scientists has been stuck on a specific math puzzle for over a decade. It's a question about how well a special kind of quantum computer (called QAOA) can solve a specific type of optimization problem (finding the best way to arrange things). They knew the answer was likely a specific number, but they couldn't prove it was exactly that number.

In this paper, the authors report that they finally solved it. But they didn't do it alone. They used a powerful AI (a large language model named "Claude Fable 5") to write the proof, and then they used a "robot referee" (a software called Lean 4) to check every single step of the AI's work to make sure it was 100% correct.

The Puzzle: The "Ring of Disagrees"

To understand the problem, imagine a ring of people standing in a circle. Each person is holding a flag that can point either Up or Down.

The Goal: The goal is to arrange the flags so that as many neighbors as possible are holding flags in opposite directions (one Up, one Down). This is called the "Ring of Disagrees."
The Machine: The scientists use a quantum computer to try to find the best arrangement. The computer doesn't just guess; it uses a specific recipe (an algorithm) with a certain number of steps, called "depth."
The Conjecture: Back in 2001, three scientists (Farhi, Goldstone, and Gutmann) guessed that if you use a specific number of steps ( $p$ $p$ ), the computer would get a perfect score of exactly $(2p + 1) / (2p + 2)$ .
- Example: If you use 1 step, the best score is $3/4$ . If you use 2 steps, it's $5/6$ .
- They could prove it for 1 step, and later for 2 steps, but for any higher number of steps, it remained an open question.

The Method: The "Draft and Check" Loop

The authors didn't just ask the AI to "solve it." They built a safety system:

The Library: First, they built a massive digital library of math rules in the "Lean" software. Think of this as a dictionary of definitions that the AI and the referee both agree on.
The Gap: They translated the unsolved part of the puzzle into a single, precise sentence that the AI needed to prove.
The AI's Job: The AI (Claude) tried to write a proof. It would write a plan in plain English, try to turn it into code, and then ask the Lean software: "Did I do this right?"
The Referee's Job: Lean is a strict robot. If the AI made even a tiny logical error, Lean would say, "No, that step doesn't follow." The AI would then try again, fixing the error.
The Result: This loop kept going until Lean said, "Yes, this proof is valid."

The human scientists only had to check that the initial translation of the problem was correct. Once that was set, the AI did the heavy lifting, and the robot referee certified the result.

The Breakthrough: Finding a Hidden "Symmetry"

The most exciting part of the paper is how the AI solved it. The proof the AI found was surprising and elegant.

The Old Way: Before this, people tried to solve this by looking at the complex quantum waves of the whole system. It was like trying to untangle a giant knot by pulling on every single thread at once.
The AI's New Way: The AI discovered a hidden "dynamical symmetry." It realized that the complex quantum system could be broken down into many tiny, independent "mini-systems" (like separate gears).
The Analogy: Imagine you have a giant, complicated clock with thousands of gears. Instead of trying to fix the whole clock at once, the AI realized that each gear moves in a simple, predictable pattern that follows the rules of a much simpler field of math called Quantum Signal Processing (QSP).
The "Polynomial" Trick: By viewing the problem through this new lens, the AI turned the difficult quantum problem into a problem about polynomials (mathematical expressions with $x$ 's and numbers). It showed that finding the perfect arrangement is the same as finding a specific polynomial curve that hits certain points.
The Solution: The AI constructed this polynomial explicitly. It proved that such a curve must exist and that it leads exactly to the score the scientists had guessed 20 years ago.

Why This Matters (According to the Paper)

The paper claims this is a major milestone for two reasons:

It Solves a Decade-Old Mystery: It confirms the exact performance limit of this quantum algorithm, which helps scientists know exactly what to expect from quantum computers.
It Proves a New Way of Working: It shows that we can use AI to generate complex mathematical proofs and use formal software to verify them. The authors say this method could be used to solve other hard problems in physics and math in the future, because the "robot referee" ensures the AI doesn't make up facts or hallucinate errors.

In short: The authors used a human-AI team to solve a 20-year-old quantum puzzle. The AI found a clever shortcut using a hidden symmetry, and a computer program verified every step, turning a guess into a mathematically certain fact.

Technical Summary: A Machine-Verified Proof of a Quantum-Optimization Conjecture

Problem Statement
The paper addresses a long-standing open problem in quantum optimization regarding the performance of the Quantum Approximate Optimization Algorithm (QAOA). Specifically, it targets the Farhi, Goldstone, and Gutmann (FGG) conjecture [1], which posits that for the "ring of disagrees" problem (equivalent to the MaxCut problem on a cycle graph or a 1D antiferromagnetic Ising model with periodic boundary conditions), a depth- $p$ QAOA circuit achieves an exact optimal approximation ratio of:
$r^*_p = \frac{2p + 1}{2p + 2}$
This conjecture holds for even system size $n$ in the regime $2p + 2 \leq n$ . While the upper bound $r^*_p \leq (2p+1)/(2p+2)$ was established via an exact decomposition of the residual energy into independent momentum modes [21], the existence of a set of $2p$ variational angles that simultaneously saturates this bound for all modes remained unproven. The core difficulty lies in solving a system of $2p$ transcendental equations derived from the SO(3) rotations of the QAOA layers, where a solution is not guaranteed by simple parameter counting.

Methodology
The authors employed a "generate-then-certify" paradigm, combining human-guided autoformalization with a large language model (LLM) and a formal proof assistant.

Formalization: The team built upon an existing Lean 4 library for quantum information (developed in parallel work on machine-verified quantum cryptography [29]). They extended this library to formalize QAOA components, the Ising cost Hamiltonian, the mixing operator, and the exact decomposition of the problem into decoupled momentum modes using the Jordan-Wigner transformation and Anderson pseudospin formalism.
Gap Isolation: The open problem was reduced to a single formal statement in Lean: the existence of angles $\gamma, \beta$ such that the per-mode residual energy vanishes for all $p$ momentum modes simultaneously.
AI-Driven Proof Construction: The LLM, Claude Fable 5, was tasked with constructing a proof for this statement. The process utilized an agentic feedback loop (Fig. 1):
- The model generated natural language proof plans.
- It performed internal numerical checks using Python to validate candidate constructions before formalization.
- It attempted to formalize the proof in Lean 4.
- The Lean compiler and kernel provided diagnostics (type errors, open goals), which the model used to revise its strategy.
Verification: The final output is a Lean 4 proof that compiles cleanly using standard classical axioms from Mathlib [3]. Human verification was limited to ensuring the formal statement faithfully encoded the original conjecture; the logical derivation was entirely machine-certified.

Key Contributions and Results
The paper presents the first machine-verified proof of the FGG conjecture. The proof, generated by the LLM and verified by Lean, reveals a hidden dynamical symmetry in the problem that transforms a difficult existence question into an explicit construction:

Symmetry Reduction: The proof identifies that the per-mode dynamics, though originally described by SO(3) rotations, can be lifted to the double cover SU(2). This reduces the problem to a single-qubit geometric quantity.
Quantum Signal Processing (QSP) Connection: The authors demonstrate that the QAOA propagator for each mode corresponds to a Quantum Signal Processing (QSP) sequence. This maps the steering condition (driving the state to the cost axis) to the requirement that a specific off-diagonal entry of an SU(2)-valued Laurent polynomial vanishes at specific nodes.
Explicit Construction: By leveraging the parity constraints of the QSP polynomial (class-L pairs), the model constructs a unique polynomial $T(w)$ that vanishes at the required $p$ momentum nodes. The normalization of this polynomial at the unsteerable mode ( $k=\pi$ ) naturally yields the constant $1/(2p+2)$ , confirming the conjectured bound.
Angle Recovery: Using Haah's factorization theorem [39], the proof explicitly maps the constructed polynomial back to the QAOA angles, demonstrating that a valid set of parameters exists.

Significance
The paper claims significance on two primary fronts:

Resolution of a Decade-Old Conjecture: It settles the exact expressive power of depth- $p$ QAOA on the ring of disagrees, providing a rigorous performance guarantee for a canonical benchmark in combinatorial optimization.
Validation of AI-Assisted Formal Verification: The work demonstrates that LLMs can successfully navigate complex mathematical reasoning when coupled with formal verification tools. The methodology proves that models can discover non-trivial mathematical structures (such as the connection between QAOA and QSP) and that the "hallucination" risk in long mathematical arguments can be mitigated by a deterministic verification loop. The authors note that this approach is applicable to other open conjectures in quantum information and broader scientific domains where problems can be formally stated.

The paper concludes that the proof is striking not only for its result but for its structure, which unifies free-fermion symmetries, SU(2) dynamics, and QSP theory—a route not previously explored in the literature on this specific conjecture.