Autonomously Extending the SYK Spectral-Edge Theorem with AI

We are releasing a new paper in mathematical physics. The research, proof development, writing, verification, and review were all carried out by AI, demonstrating sustained end-to-end research. There’s a technical summary below and the full AI-generated manuscript linked at the bottom, but before that, here’s our perspective on this.

Even if AI capabilities were frozen at where they are today, the impact on science would already be significant. The coding capabilities alone are already changing how a lot of science is done, including in areas that previously never relied heavily on programming.

Our perspective, however, is that AI will advance beyond that, and we will soon be in a world where higher-order AI discoveries are the norm (by which we mean AI results that are themselves built on top of AI discoveries). This result that we’re sharing is itself potentially such an example, given that the original result cited in this paper also directly mentions AI usage.

The proliferation of higher-order AI results leads to a new set of challenges, beyond just verifying correctness. As we build systems that are continuously working at the frontier of existing knowledge, how do we ensure that progress (and compute) is being directed towards solving problems that meaningfully advance technology and science? The landscape of worthwhile questions evolves as new results add to existing knowledge. This is of course an inextricable part of science itself, and the best researchers are able to ask questions that meaningfully progress their field of work. The successful AI research systems will be ones that reflect this dynamic process of knowledge-creation, creating a compounding loop where each completed project expands the system’s tools and ability to pursue new problems.

Systems for this new paradigm of science will transform technology development and make it economical for many more to answer questions typically reserved for technical teams with high budgets. Initio is building this in the computational sciences.

Summary

The work studies the Sachdev–Ye–Kitaev model (SYK). SYK is a simplified model of a quantum system with strong interactions and has become an important testing ground for ideas in quantum chaos, many-body physics, and quantum gravity.

One difficult question is how to determine the model’s spectral edge, i.e. the highest and lowest energies the system can reach. A recent theorem solved this problem for the standard four-body version of SYK.

Our work extends this result to every fixed even interaction order. It also determines how the spectral edge behaves as the interaction order becomes large. The ideation, research, proof development, verification, writing, and review were performed entirely by AI.

Overview

The SYK model describes ($N$) interacting fermions. Its interactions are random, and the parameter ($q$) determines how many fermions participate in each interaction.

Each particular set of random interactions produces a matrix called a Hamiltonian. The Hamiltonian’s eigenvalues represent the possible energies of the system. The largest and smallest eigenvalues form the spectral edges.

The central question we address is:

As the number of fermions becomes very large, what are the highest and lowest possible energies?

At finite size, these energies depend on the random interactions. Our theorem shows that after the right rescaling, this randomness disappears and the spectral edges converge to fixed values in the large-system limit. For every fixed even $q \geq 4$, the upper edge converges to a deterministic value $E_q$, while the lower edge converges to $-E_q$.

This gives the limiting extreme energies for every even interaction order:

$$q=4,6,8,10,\ldots$$

The lower spectral edge is the ground-state energy, one of the central quantities used to understand a quantum system at zero temperature. The overall distribution of eigenvalues does not tell us where the highest and lowest energies lie (understanding the bulk of the spectrum does not determine its extremes).

A recent theorem by Yukun He established the exact spectral edge for the quartic SYK model, the case $q = 4$.

This new work extends that theorem to every fixed even interaction order. Doing so required rebuilding the argument with the interaction order treated as a general parameter. Several parts of the proof change with $q$, including the Majorana algebra, normalization factors, combinatorics, locality estimates, cavity expansion, zero-temperature analysis, and concentration bounds.

The result shows that the mathematical structure behind the quartic theorem continues to hold throughout the full fixed-$q$ family.

Our paper also determines how the edge behaves as the interaction order grows:

$$E_q \sim \frac{\sqrt{2}}{q}.$$

It proves the first correction to this expression as well. This confirms, rigorously, a large-$q$ prediction from the physics literature.

Full paper available here.