arXiv:2610.02178
Learning SYK Hamiltonians
Anurag Anshu, Srinivasan Arunachalam, Sitan Chen, Yeongwoo Hwang
Abstract
We study the problem of learning the dense Sachdev--Ye--Kitaev (SYK) Hamiltonian from copies of its Gibbs state. Existing algorithms for Hamiltonian learning typically rely on geometric locality or bounded interaction degree and therefore do not apply to SYK, where each quartic interaction overlaps with $Θ(n^3)$ others. We show that this obstruction can be overcome by exploiting the random mean-field structure of the model. At any constant temperature, we prove that with high probability over the SYK couplings, the entire Hamiltonian can be learned to inverse-polynomial accuracy using polynomially many samples. Furthermore, when the inverse temperature is restricted to be a sufficiently small constant, we construct a quasipolynomial-time learning algorithm which is qualitatively different from the sample-efficient algorithm.
AI Summary
Can a dense, noncommuting SYK Hamiltonian be reconstructed from copies of its Gibbs state at constant temperature, despite its unbounded interaction degree? This paper studies recovery of the Gaussian couplings of quartic SYK, whose $\Theta(n^4)$ terms each overlap with $\Theta(n^3)$ others. For every fixed known inverse temperature $\beta>0$, it establishes recovery with absolute coupling error $\|\widehat{\vec g}-\vec g\|_2\leq\varepsilon$ using $n^{O(1+\beta^2)}\varepsilon^{-2}\log(n/\zeta)$ Gibbs-state copies. The guarantee holds with probability $1-O_\beta(n^{-1})$ over disorder and conditional measurement success at least $1-\zeta$. This sample-complexity result extends to every fixed even interaction order. The proof uses maximum-entropy reconstruction and an inverse-polynomial lower bound on the log-partition-function Hessian at the random target. It compares BKM covariance with ordinary thermal variance, extracts individual coefficients through Majorana conjugations, and controls local-quench Rényi powers using disorder-averaged nested-commutator moments and Gaussian concentration. A separate algorithm, restricted to sufficiently small constant $\beta$, uses $O_\beta(n^{13}\log(n/\zeta))$ copies and runtime $n^{O(\log n)}+\widetilde O_\beta(n^{14}\log(n/\zeta))$. It achieves the weaker normalized error $\|\widehat{\vec g}-\vec g\|_2/\sqrt{\binom n4}\leq n^{-c}$ for fixed $0<c<1$, with disorder failure $O_\beta(n^{-2(1-c)})$. A logarithmic-degree thermal expansion and symmetry reduce quadratic-bias cancellation to a $2\times2$ calibration problem; Wick/polymer estimates control higher-order errors. Polynomial-time recovery remains unresolved. The all-temperature estimator has no efficient runtime guarantee, and available thermal-estimation methods are not certified along candidate optimization trajectories.
Generative AI built on authors' sketch of the sample complexity argument, and came up with the versions included here
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.