Inside OpenAI's math claimed to accelerate Fourier transforms
openai
| Source: HN | Original article
OpenAI unveils new math claiming to compute Fourier transforms faster than the conventional fast Fourier transform algorithm.
OpenAI has added a new entry to its public “math repository” that claims an algorithm can compute the discrete Fourier transform (DFT) faster than the classic fast Fourier transform (FFT). The paper, posted under problem #130, proposes a complexity of O (n (log n)^{1‑ε}) with ε set to 10⁻¹³, a modest but theoretically significant improvement over the long‑standing O (n log n) bound. A GitHub project (shea256/fourier-transform-below-nlogn) tracks the work as an “explicit power saving for the exact DFT,” linking the advance to recent progress on integer multiplication.
Why it matters is twofold. First, the FFT underpins virtually every digital signal‑processing pipeline—from telecommunications to scientific computing—so any reduction in asymptotic cost could translate into lower energy use and faster runtimes for large‑scale workloads. Second, the claim comes from an AI‑driven research pipeline, echoing OpenAI’s earlier, controversial Navier‑Stokes claim that a model produced a novel solution to a millennium‑prize problem. As we reported on 11 October, that claim sparked debate over the reliability of AI‑generated mathematics. The new Fourier result adds another data point for assessing whether large language models can genuinely extend the frontier of algorithmic theory.
What to watch next are the independent verifications that the community is already mobilising. Researchers on the arg min blog have begun dissecting the proof, while the broader theoretical‑computer‑science field will likely submit formal reviews. If the algorithm survives scrutiny, we may see follow‑on work adapting the technique to related transforms and integrating it into the software stacks that power today’s AI accelerators.
Sources
Back to AIPULSEN