IFISC researchers reveal how quantum network boundaries shape machine learning performance

•  Researchers from IFISC (CSIC-UIB) show that the learning ability of quantum neural networks can depend entirely on how their boundaries are connected.
• Published in Optica Quantum, the study reveals how the Liouvillian skin effect enables control of information processing through dissipation. 

A new study, led by IFISC researchers, demonstrates that quantum neural networks can exhibit the Liouvillian skin effect, a phenomenon where boundary conditions drastically affect the behavior of quantum systems. This discovery opens new pathways for controlling quantum machine learning performance, highlighting the role of dissipation and network architecture in emerging quantum technologies. 

Can the boundary of a quantum system determine how well it learns? A study published in Optica Quantum answers this question by showing that the performance of quantum reservoir computers, networks used to process temporal information, can be fully governed by their boundary conditions. The research, conducted by researchers of the Institute for Cross-Disciplinary Physics and Complex Systems (IFISC, CSIC-UIB), reveals that even in irregular quantum architectures, adding or removing a network link can radically alter the network’s computational capabilities. 

The study focuses on a phenomenon known as the Liouvillian skin effect. In this context, certain modes of a quantum system become localized near its boundaries due to asymmetrical dissipation, significantly influencing system dynamics. “We show that this effect, traditionally studied in a fundamental context, also emerges in quantum neural networks and can describe learning performance” explains Antonio Sannia, lead author of the study. 

The team used open quantum systems where dissipation induces chiral particle hopping across network nodes. In this setup, the information to be learned, such as patterns in time series, is encoded not in the system's Hamiltonian but in its dissipative structure. Crucially, when the boundaries are closed, the network fails to retain any useful information. However, learning is restored when the boundaries are slightly opened, and performance increases dramatically. 

“We found that the ability of a quantum reservoir to perform tasks like memory retention or nonlinear classification depends entirely on how a single network's link is connected”, adds Sannia. 

The findings suggest that skin effects could become essential tools in the development of robust and flexible quantum technologies, especially in the realm of neuromorphic and reservoir computing.

Antonio Sannia, Gian Luca Giorgi, Stefano Longhi, and Roberta Zambrini, "Liouvillian skin effect in quantum neural networks," Optica Quantum 3, 189-194 (2025) https://doi.org/10.1364/OPTICAQ.541744

• El diari de la UIB