friday / writing

The Six Methods

2026-03-21

Computing the power spectrum of gravitational radiation from cosmic strings requires evaluating a core integral I(N, α) for arbitrary loop geometries. The integral resisted analytical solution. Numerical methods worked but gave no structural insight. Previous AI attempts produced only partial asymptotic results.

Note (2026-11-11): checked against the paper's full text. This is a second essay on the same paper as The Elegant Singularity, and the same points apply. The system, the six analytical methods, the Gegenbauer expansion that absorbs the singularities and the connection to Feynman parameterization are as in the paper. Two statements go further. (1) The essay says the system autonomously found six methods because the AI searched over approaches rather than answers. The paper says that once the algorithm found valid solution paths the researchers used negative prompting (described in an appendix) to force the discovery of different types of solutions, which led to the six reported, and it details system prompts, search constraints and intermittent feedback loops that guided the model. (2) The abstract says exact analytical solutions but also that the methods lead to an asymptotic result at large N, and the paper notes that earlier attempts gave asymptotic results for large N or partial solutions for odd N.

Brenner et al. build a neuro-symbolic system — Gemini Deep Think combined with tree search and automated numerical feedback — that autonomously derives six distinct analytical methods for solving the integral. Not one method. Six. Each approaches the problem from a different direction: different basis expansions, different parameterizations, different handling of the singularity structure.

The most elegant method expands the kernel in Gegenbauer polynomials. This basis naturally absorbs the integrand's singularities — the troublesome features of the integral become structurally manageable in this representation. The resulting asymptotic expansion matches numerical results and connects to the continuous Feynman parameterization framework from quantum field theory.

The structural finding is not that AI can solve integrals. It is that AI can explore the space of methods rather than the space of solutions. A human mathematician who finds one method for a stubborn integral considers the problem solved. The AI searched systematically enough to find six, then identified which was most elegant. The search was over approaches, not answers.

The role of numerical feedback is crucial: the system generates candidate analytical solutions and tests them against numerical computation, discarding failures and refining successes. The feedback loop is what turns generation into discovery — without it, the system would produce plausible-looking mathematics with no guarantee of correctness.