arXiv:2610.10697
The surprising simplicity of forecasting quantum chaotic many-body dynamics
Francesco Perciavalle, Agostino Gallo, Francesco Plastina, Gianluigi Greco, Nicola Lo Gullo, Carlo Adornetto
Abstract
Accurately simulating chaotic quantum many-body dynamics remains a major computational challenge for classical methods, due to the rapid buildup and spatial spreading of entanglement during the evolution. This raises the question of whether machine learning can provide an effective alternative for predicting quantum dynamics. We address this question by formulating quantum dynamics as a time-series forecasting problem, using a few-qubit PXP chain, realizable with Rydberg-atom arrays, as a benchmark. By varying the initial state, the system spans dynamical regimes ranging from ergodic behavior to quantum many-body scarring, providing a controlled setting for testing forecasting models across qualitatively different dynamics. We compare two contrasting architectures: an expressive nonlinear Transformer and DLinear, a simple linear forecasting model. The Transformer accurately predicts dynamics in the more ergodic regime, but its performance progressively deteriorates as the initial state approaches the scarred limit. In contrast, DLinear remains accurate across the entire family of initial states, with its main deviations consisting of small high-frequency oscillations that have little effect on the overall prediction error. Remarkably, these results show that observables generated by complex quantum many-body dynamics can be forecast with high accuracy through a simple linear mapping from past to future observations. This reveals that the complexity of the underlying quantum evolution need not translate into an equally complex forecasting problem.
AI Summary
The paper asks whether forecasting observables of chaotic quantum many-body dynamics requires an expressive nonlinear model. It compares DLinear, a linear time-series forecasting architecture, with an autoregressive Transformer on a PXP chain exhibiting both thermalizing and quantum-scarred dynamics. The benchmark uses exact evolution of an $L=10$ open chain for 500 initial states sampled from a two-parameter family interpolating between polarized and Néel configurations, including entangled superpositions. Each trajectory contains 55 features: local magnetizations and all connected two-site $zz$ correlations. Models receive 500 samples covering $t\in[0,10)$ and predict 300 samples over $t\in[10,16)$, with $\Delta t=0.02$. There are 400 training/validation trajectories and 100 held-out test trajectories. Both models use a loss penalizing value errors and discrepancies in temporal increments and directions; DLinear predicts the entire horizon directly, whereas the Transformer generates it recursively. DLinear yields low, comparatively uniform mean absolute errors across the initial-state family. Its raw directional accuracy is approximately $50\%$, attributed to small high-frequency prediction artifacts; smoothing raises it above $90\%$ across much of parameter space. Transformer errors increase toward the scarred regime, with local absolute errors approaching $0.7$ in the selected worst-case trajectory. Its error grows more strongly over the forecast horizon. The stated limitations are the single finite chain, restricted initial-state family, selected observables, and fixed forecasting horizon. Generalization to other Hamiltonians, larger systems, circuits, and open-system dynamics remains unestablished.
Error found: Because DLinear uses direct multi-step prediction and the Transformer uses teacher-forced autoregressive prediction, their comparison does not isolate linear versus nonlinear expressiveness as the reason for the performance difference. The paper gives incompatible Transformer protocols—sliding-window next-step forecasting and target-sequence encoder–decoder forecasting—leaving the implementation underlying the reported comparison ambiguous.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- Francesco Perciavalle and Agostino Gallo contributed equally to this work. 11 pages, 5 figures