Simulated quantum reservoirs showed their strongest short-term forecasting at a narrow point where information scrambling was beginning, according to a preprint study. At a 17-step forecast horizon, precision peaked near the upper edge of the scrambling-onset region in every network size studied from N = 6 to N = 20.
At that onset, precision grew with network size with exponent α = 2.64 ± 0.12, compared with 0.35 ± 0.07 in frozen dynamics and 0.90 ± 0.09 in scrambled dynamics. Since an exponent above 1 means faster-than-linear growth with size, the first result is what the study calls superextensive scaling. The advantage was strongest at horizons of 10 to 20 steps, remained at 42 steps, and disappeared by 84 steps, when scaling was sublinear in every region.
A narrow operating window
The work asks whether learning performance can outpace operating costs as a reservoir grows, and whether an edge-of-chaos condition is necessary or information scrambling can be enough. The reservoirs were weakly disordered two-dimensional transverse-field Ising spin networks, controlled by the dimensionless transverse-field ratio h_x/J.
Researchers classified the dynamics using late-time saturation of an out-of-time-order correlator, or OTOC. They labelled C < 1 as frozen and C > 2 as scrambled, with the onset between those ranges. The classification organized a sweep across the single field ratio, but the authors do not treat the scrambling diagnostic as proof of chaos.
For forecasting, the team generated a Mackey-Glass series of 800 time steps, discarded the first 30, reserved the final 100, and used a 70:30 training-to-test split. Forecast horizons ranged from 1 to 85 steps. The closed-system protocol reset the reservoir after every input, while histories of raw features supplied temporal context classically rather than keeping it in the quantum state.
Forecasting used 10 network-disorder and input-function realizations up to N = 18, and 5 realizations at N = 20. That means the high-size estimate rests on fewer realizations than the smaller network results.
More than a forecasting peak
A separate diagnostic tracked how broadly the quantum state occupied computational-basis states. Its participation ratio peaked near the upper edge of the onset window, and the maximum grew exponentially with N, rising by more than two orders of magnitude between N = 6 and N = 20.
Expressive capacity, estimated from the rank of a feature Gram matrix, told a slightly different story. It grew sublinearly in the frozen region, with α = 0.75, but superextensively at onset and in the scrambled region, with α = 1.42 and α = 1.72. The representative field ratios were 0.25 for the frozen case, an onset window from 1 to 2, and 6.00 for the scrambled case.
Memory changed as scrambling increased
Memory shifted as scrambling increased. Linear memory was largest in the frozen region, while progressively higher nonlinear orders peaked at stronger fields; the quartic contribution peaked in the scrambled regime. The result suggests a redistribution from simple linear memory toward higher-order nonlinear components, rather than a single memory peak shared by every order.
Total memory also depended on how forgetting was supplied. The open-system simulations compared collective and local relaxation at a common relaxation rate. At h_x = 1.85 J and γ = J/2, total memory scaled with α = 1.47 ± 0.15 for collective relaxation and α = 0.78 ± 0.09 for local relaxation. Collective relaxation outperformed local relaxation for all tested γ when N > 4, while the two channels crossed near N = 6.
These memory calculations used Monte Carlo wavefunction quantum trajectories, averaging 100 trajectories per realization. There were 10 realizations at N = 4, 6 and 8, and 2 at N = 10 and 12. The authors interpret collective relaxation as the forgetting mechanism associated with the superextensive memory result, but that interpretation has not been tested on a device.
A result bounded by simulation
The evidence remains finite and model-specific. Forecasting was carried out on network sizes up to N = 20, memory analysis used N = 4 to 12, and the forecasting benchmark was one Mackey-Glass series. The study does not establish that the same scaling will hold for other tasks or physical hardware.
The document is a preprint identified as arXiv:2608.25511v1, dated 26 August 2026. It says supporting data and code are available from the corresponding author on reasonable request. Funding came from an AQSN microgrant from the Hon Hai (Foxconn) Research Institute and support from the Pawsey Supercomputing Research Centre through the National Computational Merit Allocation Scheme 2026, with additional NCRIS support for the simulation hub.
Paper data and sources
Original title: Superextensive learning in quantum reservoirs at the onset of information scrambling
Authors: Jonas Freiheit, Francesco Campaioli
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text