Dynamical learning and quantum memory with non-Hermitian many-body systems

Moein N. Ivaki, Austin J. Szuminsky, Achilleas Lazarides, Alexandre Zagoskin, Gerard McCaul, Tapio Ala-Nissila · 2025

Read Dynamical learning and quantum memory with non-Hermitian many-body systems by Moein N. Ivaki, Austin J. Szuminsky, Achilleas Lazarides, Alexandre Zagoskin, Gerard McCaul, Tapio Ala-Nissila (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-07676.txt.

Introduction

Quantum systems that leak energy to their surroundings behave in ways closed systems do not. Moein Ivaki and co-authors study what such leaky systems can learn and remember in this 2025 paper. The systems are called non-Hermitian, and they pass through phases marked by a symmetry that holds or breaks. That shift changes how tangled and how pure their states are, which in turn sets their memory power.

The test case is a web of interacting spins placed on random graphs. The authors find that the first special point, where the system's spectrum turns from real to complex values, matches a sharp drop in learning power. Tuning disorder and spin coupling moves that point, so the learning threshold can be dialed. Inside the learning phase, the system holds information in a usable way.

Catalog record

CREATORMoein N. Ivaki, Austin J. Szuminsky, Achilleas Lazarides, Alexandre Zagoskin, Gerard McCaul, Tapio Ala-Nissila
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE52 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-07676

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Dynamical learning and quantum memory with non-Hermitian many-body systems Moein N. Ivaki, Austin J. Szuminsky, Achilleas Lazarides, Alexandre Zagoskin, Gerard McCaul, Tapio Ala-Nissila Abstract Non-Hermitian (NH) systems provide a fertile platform for quantum technologies, owing in part to their distinct dynamical phases. These systems can be characterized by the preservation or spontaneous breaking of parity-time reversal symmetry, significantly impacting the dynamical behavior of quantum resources such as entanglement and purity; resources which in turn govern the system’s information processing and memory capacity. Here we investigate this relationship using the example of an interacting NH spin system defined on random graphs. We show that the onset of the first exceptional point - marking the real-to-complex spectral transition - also corresponds to an abrupt change in the system’s learning capacity. We further demonstrate that this transition is controllable via local disorder and spin interactions strength, thereby defining a tunable learnability threshold. Within the learning phase, the system exhibits the key features required for memory-dependent reservoir computing. This makes explicit a direct link between spectral structure and computational capacity, further establishing non-Hermiticity, and more broadly engineered dissipation, as a dynamic resource for temporal quantum machine learning. I Introduction The key quantity in all learning processes is memory. This is true of both natural and synthetic agents: The capacity to generalize from experience relies on the ability to store it. Given however that any computation associated with learning must take place on a physical substrate Lloyd 2002 , it is natural to ask what dynamical processes enable it. Whether realized in neural tissue, silicon circuits, or quantum media, computation is ultimately an interpretive structure superimposed on the dynamics that realize it. From this perspective, the pursuit of practical learning architectures becomes a search for dynamical systems capable of encoding, transforming, and recalling information with precision and efficiency. In this hunt, quantum systems present one of the richest targets. The exponential scaling of their underlying state space, long-range nonclassical correlations, and dependence on initial conditions give them the potential to process high-dimensional data far more efficiently than classical methods. It is for this reason that Quantum Machine Learning (QML) seeks to harness these unique features, applied to learning tasks such as optimization, classification, and prediction Biamonte et al. 2017 ; Schuld et al. 2015 ; Dunjko and Briegel 2018 ; Liu et al. 2021 . Among various QML paradigms, quantum reservoir computing (QRC) Fujii and Nakajima 2017 ; Fujii and Nakajima 2020 ; Govia et al. 2021 ; Angelatos et al. 2021 ; Kornjača et al. 2024 ; Ghosh et al. 2019 ; Mujal et al. 2021a ; Hu et al. 2024 ; Senanian et al. 2024 ; Martínez-Peña et al. 2023 is particularly well-suited to near-term devices. Inspired by classical reservoir computing and recurrent neural networks Tanaka et al. 2018 ; Ozturk et al. 2007 ; Lukoševičius and Jaeger 2009 ; Butcher et al. 2013 ; Vrugt 2024 ; Larger et al. 2017 ; Lim et al. 2020 , QRC offloads the burden of processing information onto the untrained dynamics of a physical system, dubbed the reservoir . The rationale for doing so lies in the fact that any input-output relationship can be linearized, provided the input data is embedded into a sufficiently high dimensional feature space Kutz et al. 2016 . By encoding inputs into the dynamics of quantum systems, it is only necessary to train a single linear computational layer, which translates measurements read out from the reservoir into computational outputs. QRC therefore minimizes the need for extensive network training, making it a highly efficient approach compared to conventional variational algorithms Cerezo et al. 2021 ; Larocca et al. 2025 . Such quantum learning scheme is inherently suited to analog quantum devices, as it is agnostic to the precise form of the reservoir dynamics. As such, it bypasses the stringent control requirements of gate-based circuits, offering both hardware efficiency and robustness to noise. Hence, this family of learning approaches aligns well near-terms goals of implementation in noisy intermediate-scale quantum devices Bharti et al. 2022 . QRC has been applied to a range of problems, including quantum state classification, tomography, chaotic system modeling and non-linear processing Tran and Nakajima 2021 ; Palacios et al. 2024 ; Mujal et al. 2023 ; Settino et al. 2024 ; Kobayashi et al. 2024 ; Sornsaeng et al. 2024 . Broadly speaking, there are three properties that enable a reservoir to be used for universal computations. These are fading memory, input separability, and the echo-state property Chen and Nurdin 2019 ; Sannia et al. 2024a ; Grigoryeva and Ortega 2018 ; Nokkala et al. 2021 . There is however an inherent difficulty in satisfying these desiderata in the context of quantum dynamics. In particular, fading memory becomes an impossibility under the restriction of unitary operations and global projective measurement McCaul et al. 2025 . This is an essential challenge in realizing QRC, as in many tasks successful learning requires memory. This dependence is unsurprising however. To paraphrase a famous Albert: learning is what remains when all else is forgotten Einstein 2005 . In this context, dynamical measures of chaoticity and lossiness can be tied directly to the memory capacity and precision of a given temporal learning task Xia et al. 2022 ; Ivaki et al. 2024 ; Martínez-Peña et al. 2021 . For this reason, the characterization and control of dynamical phases of quantum systems is vital to realizing the mathematical features quantum learning algorithms rely on Domingo et al. 2023 ; Fry et al. 2023 ; Harrington et al. 2022 ; Kora et al. 2024 ; Olivera-Atencio et al. 2023 ; Yosifov et al. 2025 . A natural approach to introducing the necessary dynamical map is via partial trace operations, in which one repeatedly writes and erases an input locally in an otherwise unitary system. However, such trace-based mechanisms, as originally proposed in early formulations of QRC Fujii and Nakajima 2017 , offer little flexibility in parameterizing and tuning. This is problematic, as optimizing for generalized learning tasks can only be achieved with parametrized control over the dynamical processes generating it. This gives rise to two key objectives: designing simple, parameterizable dynamical models with tunable memory effects, and developing reliable schemes for their physical implementations. Figure 1: Dynamical learning in a lossy quantum system . (a) An input sequence {\theta_{n}} is locally encoded through x -axis rotations at each step n , and applied to all nodes of a non-Hermitian quantum spin system defined on a random graph. External perturbations manifest as asymmetric spin flip operations at each site l , modeled by f_{l}(\gamma)\,\hat{\sigma}_{l}^{+} and f_{l}(-\gamma)\,\hat{\sigma}_{l}^{-} . In our setup, this function is taken to be linear in the perturbation strength f_{l}(\pm\gamma)\sim h_{l}\pm\gamma . The current input-dependent state of the system is represented by a density matrix \hat{\rho}(\theta) . After nonunitary evolution under the Hamiltonian \hat{\mathcal{H}}_{\rm res} (Eq. ( 1 )) for an optimal duration, diagonal expectation values (i.e., quantum features) are fed into a classical linear regression algorithm to reconstruct past inputs. (b) The model can undergo a spectral transition at a finite critical measurements rate \gamma^{\rm c} , where the spectrum becomes complex. Remarkably, this transition, which is due to a spontaneous symmetry breaking,

Read the complete text