GPT-5.6 Sol Pro Resolves 30-Year Convex Optimization Gap

GPT-5.6 Sol Pro Resolves Long-Standing Convex Optimization Gap

GPT-5.6 Sol Pro has proposed a proof that closes a 30-year gap in convex optimization, specifically resolving the quadratic dimension dependence for solving optimization problems over convex, Lipschitz functions on a spherical domain. The result was formally verified using the Lean theorem prover, confirming that the lower bound time complexity is $\Omega(d^2)$ function evaluations, matching the complexity of existing algorithms.

Technical Details of the Proof

The breakthrough focuses on the time complexity required to solve optimization problems over a specific class of functions: convex, Lipschitz functions on a spherical domain. In this field, establishing upper bounds (the runtime of a known algorithm) is generally straightforward, but establishing non-trivial lower bounds—which constrain all possible algorithms—is significantly more difficult.

This proof demonstrates that the lower bound time complexity is equal to the time complexity of a 30-year-old algorithm, requiring $\Omega(d^2)$ function evaluations. This suggests that $d$ is the minimal number of evaluations required when utilizing a gradient oracle, as a gradient can be approximated with $d$ function evaluations.

The Role of Human Expertise and Prompting

While the result was generated by GPT-5.6 Sol Pro in 148 minutes, the process relied heavily on substantial human domain expertise and iterative research. The model did not solve the problem from a simple, one-sentence request; instead, it was guided by a sophisticated prompting strategy:

• Extensive Priming: The model was provided with a ten-page prompt consisting of advanced mathematics designed to prime the model in the right direction.
• Iterative Research: The prompt was informed by a year of prior research conducted by the author using previous model versions (GPT-5.4 and 5.5).
• Methodological Framework: The prompt followed a methodology similar to the one used by OpenAI for the Cyclic Double Cover (CDC) conjecture proof, incorporating various reasonable mathematical approaches and specifications.
• Collaborative Prompting: The author used GPT-5.6 Sol Pro itself to help refine and write the majority of the prompt, providing the model with the CDC prompt as a template and specific problem descriptions.

Formal Verification and Peer Review

To ensure the mathematical validity of the AI-generated proof, the author formally verified the result in Lean. The proof passed the formal verification check, providing a level of certainty that exceeds standard LLM outputs, which are prone to hallucinations. However, community members have noted that the result has not yet undergone traditional peer review, which remains a critical step for academic acceptance.

Community Perspectives on AI in Mathematical Research

The intersection of LLMs and high-level mathematics has sparked significant debate regarding the future of research and the nature of intelligence:

The Shift in Research Skills

Some observers suggest that the role of the mathematician is shifting from implementation and solving "low-hanging fruit" to the conceptualization of new questions.

"In the past, implementation skills were very important, but these days, concepts feel more important... the output varies depending on how much background knowledge you have."

The "Stochastic Parrot" Debate

The ability of LLMs to contribute to unsolved mathematical problems challenges the notion that these models are merely "stochastic parrots" capable only of summarization. Proponents of this view argue that the objective reality of these proofs demonstrates a form of creativity or reasoning that transcends simple pattern matching.

Dystopian Concerns

Conversely, some view the decreasing requirement for human expertise as a bleak trajectory. There is concern that as the "loop engineering" (prompting and orchestration) becomes automated, the demand for human thought will diminish, potentially rendering traditional academic training paths obsolete.