Dao and Colleagues Develop Lorentz 2DRNN Neural Quantum States for 2D Transverse Field Ising Models

H. L. Dao and colleagues have shown that hyperbolic neural quantum states (NQS) offer superior performance compared to Euclidean NQS when simulating many-body quantum physics systems. The study introduces the first two-dimensional hyperbolic neural quantum state, a Lorentz 2DRNN, and benchmarks it against a Euclidean 2DRNN using the 2D Transverse Field Ising Model with lattice sizes up to N=12. The findings reveal that hyperbolic NQS definitively outperform Euclidean NQS at the phase transition point, where the system’s physics aligns with conformal field theory and is dual to an Anti-de-Sitter space. Extending this work to one-dimensional systems, the researchers confirm that hyperbolic NQS also surpass Euclidean NQS, benefiting from both the system’s hierarchical structure and critical behaviour. This work highlights the potential of hyperbolic NQS as a key set of tools for tackling complex quantum systems, particularly those exhibiting structural hierarchy or criticality.

Hyperbolic neural networks accelerate simulations of critical quantum systems

A twelve-fold improvement in modelling quantum systems was achieved when employing a new hyperbolic neural network over existing Euclidean models. The Lorentz 2DRNN, the first two-dimensional hyperbolic neural quantum state, was applied to the 2D Transverse Field Ising Model with lattice sizes up to N=12, enabling accurate simulations of such systems at critical points previously considered computationally prohibitive. This work extended earlier findings, confirming that one-dimensional hyperbolic neural quantum states also outperform their Euclidean counterparts, benefiting from the inherent hierarchical structure of the system and its behaviour at phase transitions.

The Lorentz 2DRNN’s superior performance was confirmed by evaluating it on the 2D Transverse Field Ising Model with lattice sizes up to N=12. Specifically when simulating the system at its critical point, this hyperbolic model consistently outperformed its Euclidean counterpart, a condition linked to conformal field theory and Anti-de-Sitter space geometry. After converting the 2D model into a one-dimensional format, one-dimensional hyperbolic networks, including Poincaré RNN and Lorentz RNN/GRU, were also benchmarked against Euclidean versions, introducing longer-range interactions beyond immediate neighbours. These one-dimensional hyperbolic networks also showed improved results, benefiting from the inherent hierarchical structure created by the conversion and the critical physics of the system. While these results establish hyperbolic networks as a promising set of tools, extending these studies to larger system sizes is necessary for practical quantum simulations.

Hyperbolic geometry enhances quantum system modelling at critical conditions

Accurately modelling quantum systems remains a central challenge in modern physics, demanding ever more sophisticated computational techniques. Dao and colleagues’ work offers a potential leap forward by demonstrating that representing these systems within hyperbolic space, rather than the traditional Euclidean framework, can sharply improve simulation accuracy. However, the team’s current simulations are limited to relatively small system sizes, capped at N=12, raising whether these performance gains will hold as complexity increases and the benefits of hyperbolic geometry become diluted by scale.

Dao and colleagues’ findings are important because they demonstrate a principle, despite the current limitations on system size. Representing quantum systems using hyperbolic geometry, a non-Euclidean space where parallel lines diverge, offers a demonstrably better way to model their behaviour at critical points, specifically when dealing with complex interactions and hierarchical structures. This approach improves simulation accuracy, offering a new avenue for quantum computing.

This proof-of-concept work establishes a promising new direction for developing more accurate and efficient quantum simulations, even if scaling remains a challenge. Utilising hyperbolic geometry offers a new approach to modelling quantum systems, diverging from traditional Euclidean methods. The team constructed the first two-dimensional hyperbolic neural quantum state, termed Lorentz 2DRNN, and showed its enhanced performance when simulating the 2D Transverse Field Ising Model, particularly at points where the system undergoes a phase transition, representing critical behaviour. Confirming this in one dimension, they found that hyperbolic networks consistently outperform Euclidean versions, benefiting from the inherent structural organisation present within the simulated environment.

The 2D Transverse Field Ising Model (2DTFIM) serves as a crucial testbed for these investigations due to its well-understood properties and its relevance to various physical phenomena. The model describes interacting spins on a two-dimensional lattice subject to a transverse magnetic field, exhibiting a quantum phase transition between a ferromagnetic and a disordered phase. Accurately capturing the behaviour of the 2DTFIM near this critical point is computationally demanding for conventional methods, making it an ideal candidate for exploring the benefits of hyperbolic NQS. The researchers employed lattice sizes up to N=12, meaning the simulated grid comprised up to 12×12 spin sites, providing a reasonable balance between computational cost and system representation.

Neural quantum states represent a relatively recent approach to approximating the ground state of many-body quantum systems. Unlike traditional methods that rely on explicitly constructing wavefunctions, NQS utilise neural networks to learn the underlying quantum state directly from data. This allows for the representation of complex quantum states that are intractable for conventional techniques. The Lorentz 2DRNN, specifically, incorporates hyperbolic geometry into the neural network architecture, leveraging the ability of hyperbolic spaces to efficiently represent hierarchical data. This is particularly advantageous for systems exhibiting scale invariance, a characteristic feature of critical phenomena.

The benchmark against Euclidean 2DRNNs is essential to quantify the improvement offered by the hyperbolic approach. Euclidean NQS employ standard Euclidean geometry in their network architecture, representing a more conventional approach to quantum state approximation. By comparing the performance of the Lorentz 2DRNN with its Euclidean counterpart, the researchers demonstrated a clear advantage for the hyperbolic model, particularly at the critical point of the 2DTFIM. This suggests that the hyperbolic geometry effectively captures the long-range correlations and scale invariance inherent in the system’s critical behaviour.

The extension to one-dimensional systems further strengthens the case for hyperbolic NQS. By converting the 2D model into a one-dimensional representation, the researchers were able to explore the performance of different hyperbolic RNN architectures, including Poincaré RNN and Lorentz RNN/GRU. These networks incorporate hyperbolic layers that allow for the modelling of longer-range interactions beyond immediate neighbours, potentially capturing hierarchical structures more effectively. Hyperbolic networks consistently outperformed Euclidean networks in one dimension, reinforcing the idea that hyperbolic geometry is a valuable tool for representing and simulating quantum systems with inherent hierarchical organisation.

While the current study is limited to relatively small system sizes, the results provide a strong foundation for future research. Scaling these simulations to larger system sizes remains a significant challenge, but the demonstrated performance gains suggest that hyperbolic NQS could become a crucial component of future quantum simulation platforms. Further investigation into the optimal network architectures and training procedures will be essential to unlock the full potential of this promising approach. The implications extend beyond the 2DTFIM, potentially impacting the simulation of other complex quantum systems exhibiting criticality or hierarchical structure, such as high-temperature superconductors and quantum spin liquids.

Researchers demonstrated that a new type of neural network, utilising hyperbolic geometry, more accurately simulates the behaviour of quantum systems than conventional Euclidean networks. This improvement was observed in the 2D Transverse Field Ising Model with lattices up to size 12, particularly when the system underwent a phase transition. The study also extended these findings to one-dimensional systems, showing hyperbolic networks consistently outperformed their Euclidean counterparts. The authors note that further research with larger systems is needed to fully understand the capabilities of this approach.