New non-Euclidean neural quantum states from additional types of hyperbolic recurrent neural networks

This paper introduces new non-Euclidean neural quantum state (NQS) architectures—specifically Poincaré and Lorentz variants of RNNs and GRUs—and demonstrates through Variational Monte Carlo simulations that these hyperbolic models consistently outperform their Euclidean counterparts in representing hierarchical quantum many-body systems.

The Gist

Imagine you are trying to map out a massive, sprawling city.

If you use a standard flat paper map (which is what scientists usually do), you’ll quickly run into trouble. As the city grows and the streets branch out into complex neighborhoods and suburbs, the flat map starts to distort. You can’t fit all the detail in without stretching the streets or overlapping them.

This paper is about a new way to "map" the incredibly complex behavior of quantum particles by using a different kind of "paper"—one that isn't flat, but curved like a saddle or a coral reef.

Here is the breakdown of the research:

1. The Problem: The "Flat Map" Limitation

In quantum physics, scientists try to predict how a group of particles (like spins in a magnet) will behave. This is called finding the "ground state." Because these particles interact in complex, branching patterns, the math becomes a nightmare.

Currently, scientists use Neural Networks (AI) to guess these patterns. Most of these AI models are "Euclidean," meaning they assume the world is flat. But quantum particles don't live in a flat world; they live in a world of complex, hierarchical relationships. Using a flat AI to map a hierarchical quantum system is like trying to wrap a piece of flat gift wrap around a basketball—it’s going to wrinkle and fail.

2. The Solution: "Hyperbolic" AI

The author introduces a new type of AI called Non-Euclidean Neural Quantum States (NQS).

Instead of a flat map, this AI uses Hyperbolic Geometry. Think of hyperbolic space like a piece of kale or a coral reef. It has "extra room" as you move outward. Because the space expands exponentially, it can fit complex, tree-like structures (hierarchies) much more naturally than a flat map can.

The paper introduces three new "flavors" of this curved AI:

• Poincaré models: Think of this like a circular disk where the edges are infinitely far away.
• Lorentz models: Think of this like a vast, open, curved landscape that doesn't have a "boundary" or an edge.

3. The Experiment: The Ultimate Stress Test

The researcher tested these new "curved" AIs against the old "flat" AIs using two famous physics puzzles: the Heisenberg J1J2 and J1J2J3 models. These are essentially mathematical playgrounds where particles interact with their neighbors in increasingly messy and "frustrated" ways.

The researcher scaled the test up to 100 particles, which is a massive amount of complexity.

4. The Results: The "Small but Mighty" Winner

The results were a huge success for the curved models. Here is what they found:

• Curved is better than Flat: In almost every single test, the hyperbolic (curved) AIs beat the Euclidean (flat) AIs. The flat AIs simply couldn't "wrap" around the complexity of the particles.
• The "Lorentz RNN" Surprise: This was the most exciting part. The researcher tested two types of AI: a heavy, complex one (the GRU) and a lighter, simpler one (the RNN). Usually, the heavy one wins because it has more "brain power" (parameters).
  • However, the Lorentz RNN—a lightweight, simple model—often beat the heavy, complex models.
  • Analogy: It’s like finding out that a nimble, lightweight mountain bike can actually navigate a rugged forest faster and more effectively than a heavy, high-tech SUV. Even though the SUV has more "features," the bike’s shape is just better suited for the terrain.

Summary

In short, this paper proves that if you want to understand the "shape" of quantum physics, you shouldn't use a flat ruler. You should use a curved tool. By using Hyperbolic Geometry, we can create AI that "fits" the natural complexity of the universe, allowing us to solve much harder problems with much smarter, more efficient models.

Technical Summary

Technical Summary: New Non-Euclidean Neural Quantum States from Additional Types of Hyperbolic Recurrent Neural Networks

Author: H. L. Dao
Date: April 28, 2026
Field: Quantum Physics / Machine Learning (Neural Quantum States)

---

1. Problem Statement

The paper addresses the challenge of approximating the ground state wavefunctions of quantum many-body systems (specifically frustrated Heisenberg spin models). Traditional Neural Quantum States (NQS) rely on Euclidean neural network architectures (ANN, CNN, RNN, Transformers). However, these Euclidean constructions often struggle to efficiently represent quantum states that possess hierarchical or tree-like entanglement structures.

While previous work introduced the Poincaré hyperbolic GRU to exploit non-Euclidean geometry, it was limited to a single architecture and a single model of hyperbolic space. The central problem is whether expanding the library of hyperbolic architectures (RNNs and different hyperbolic models) can provide more efficient, accurate, and parameter-light approximations for complex, highly frustrated quantum systems.

2. Methodology

The author extends the NQS framework by constructing and implementing three new types of hyperbolic recurrent neural networks:

• Poincaré RNN (based on the Poincaré disk model).
• Lorentz RNN (based on the Lorentz hyperboloid model).
• Lorentz GRU (Gated Recurrent Unit).

Mathematical Framework:

• Hyperbolic Operations: The networks replace Euclidean addition, matrix multiplication, and pointwise multiplication with their hyperbolic analogs (Möbius addition, Möbius matrix multiplication, etc.) using exponential and logarithmic maps to transition between the manifold and the tangent space.
• Wavefunction Construction: The paper utilizes an autoregressive approach to build complex NQS wavefunctions. The probability amplitude $P(\vec{\sigma})$ is generated by the recurrent layer, while the phase $\phi(\vec{\sigma})$ is generated via a separate dense layer with a Softsign activation.
• Variational Monte Carlo (VMC): The models are trained by minimizing the expectation value of the Hamiltonian (ground state energy) using the local energy formula.

Experimental Setup:

• Models: 1D Heisenberg $J_1J_2$ and $J_1J_2J_3$ models (up to 100 spins).
• Benchmarks: Six NQS types: Euclidean RNN/GRU, Poincaré RNN/GRU, and Lorentz RNN/GRU.
• Constraints: To prevent "numerical explosions" inherent in hyperbolic geometry (where states drift to the edge of the disk or infinity), the author implements spatial constraint hyperparameters: a maximal radius $R_{max}$ for Poincaré models and a maximal spatial norm $L_{max}$ for Lorentz models.

3. Key Contributions

1. Architectural Expansion: The first introduction of Lorentz-based RNN and GRU architectures into the NQS field.
2. Comparative Benchmarking: A large-scale VMC study comparing Euclidean vs. Hyperbolic geometries and Poincaré vs. Lorentz sub-geometries across different levels of quantum frustration.
3. Parameter Efficiency Analysis: Demonstrating that simpler hyperbolic architectures can outperform more complex Euclidean ones with significantly fewer parameters.
4. Stability Protocols: Development of spatial constraint mechanisms ($R_{max}$ and $L_{max}$) to stabilize training in non-Euclidean manifolds.

4. Results

The experiments yielded several critical findings:

• Hyperbolic Superiority: In almost all cases, hyperbolic NQS (both Poincaré and Lorentz) outperformed their Euclidean counterparts of the same architecture.
• Lorentz vs. Poincaré:
  • For RNNs, the Lorentz RNN was the definitive winner, consistently outperforming both Poincaré RNN and Euclidean RNN.
  • For GRUs, the performance was more nuanced; Lorentz and Poincaré GRUs "took turns" being the top performer depending on the specific coupling constants ($J_2, J_3$), suggesting that the optimal geometry may depend on the underlying physics of the system.
• The "Lorentz RNN" Surprise: Most notably, the Lorentz RNN—despite having up to three times fewer parameters than the Euclidean GRU—outperformed the Euclidean GRU in 8 out of 8 experimental settings. This suggests that the hyperbolic geometry of the Lorentz model is a more powerful inductive bias for these quantum systems than the sophisticated gating mechanism of a Euclidean GRU.
• Complexity vs. Geometry: While GRUs generally provide better approximations than RNNs when using the same geometry, the pure hyperbolic geometry of a simple Lorentz RNN is often more effective than the complex gating of a Euclidean GRU.

5. Significance

This work proves that geometry is a critical component in the success of NQS. By moving from Euclidean to hyperbolic space, researchers can capture the hierarchical nature of quantum entanglement more effectively. The discovery that a low-parameter Lorentz RNN can outperform a high-parameter Euclidean GRU provides a roadmap for developing more efficient, scalable neural network models for quantum many-body physics, potentially reducing the computational cost of simulating large-scale quantum systems.