⚛️ quantum physics

Bridging Frustration and Non-Hermiticity via COMPASS: An Adaptive Biorthogonal Neural Quantum State Framework

This paper introduces COMPASS, an adaptive biorthogonal neural quantum state framework that enables stable simulations of non-Hermitian many-body systems by combining autoregressive sampling with symmetry-aware ansatz selection, revealing how frustration protects against spectral instability and generates novel diabolic ring topologies.

Original authors: Lavoisier Wah, Flore K. Kunst, Mohamed Hibat-Allah

Published 2026-07-16

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Lavoisier Wah, Flore K. Kunst, Mohamed Hibat-Allah

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a massive, three-dimensional puzzle, but the pieces are made of a strange, wobbly material that changes shape every time you touch them. This is the world of non-Hermitian quantum physics. In our everyday world, things usually follow strict, predictable rules (like a ball always falling down). But in the quantum realm, particles can exist in multiple states at once, and when they interact with their environment—like losing energy or gaining it from a source—the math describing them becomes "non-Hermitian." This means their energy levels can become complex numbers, involving both real values (like normal numbers) and imaginary ones (mathematical concepts that don't exist on a standard number line).

For decades, scientists have struggled to simulate these tricky systems because the usual tools break down. The standard way to find the lowest energy state of a system (the "ground state") relies on the assumption that the system is perfectly balanced and stable. But when a system is non-Hermitian, it's like trying to balance a spinning top on a moving skateboard; the usual methods get confused, often leading to wrong answers or getting stuck in loops. Furthermore, these systems often involve "frustration," where the particles can't all be happy at the same time because their neighbors are fighting over who gets to be in which state. Understanding how these unstable, frustrated systems behave is crucial because they might hold the keys to new types of lasers, better sensors, and exotic materials that don't exist in nature yet.

Enter a new team of researchers who have built a digital "smart search engine" called COMPASS to navigate this chaotic landscape. Think of COMPASS as a highly adaptable robot explorer. Instead of using a rigid, pre-made map (a fixed computer model), this robot can grow and shrink its own brain as it explores. It uses a special trick called "biorthogonal" math, which is like having two different pairs of eyes—one looking at the system from the past and one from the future—to make sure it sees the whole picture clearly. The robot doesn't just guess; it uses a clever two-step dance. First, it hunts for the lowest energy, and then it double-checks its work to ensure it hasn't accidentally found a fake solution.

The researchers discovered something surprising about how to build this robot's brain. They found that for some systems, using a "complex" brain (one that handles imaginary numbers) actually makes the robot hallucinate, breaking the rules of physics and giving fake results. In these cases, a "real" brain (one that only deals with normal numbers) is the only way to find the truth. However, for other, more chaotic systems, the complex brain is absolutely necessary to see what's really happening. It turns out that the "right" tool depends entirely on the specific rules of the puzzle you are solving.

Using this smart, adaptable approach, the team simulated systems with up to 200 particles in a line and 100 particles in a grid, which is a massive leap forward in size compared to what was possible before. They applied this to a famous model of "frustrated magnetism" (where particles are stuck in a tug-of-war). They found that the "frustration" itself acts like a shield, protecting the system from becoming unstable and chaotic up to a certain point. But when they made the frustration itself "complex" (adding that wobbly imaginary ingredient), something magical happened: the system didn't just get messy; it formed a brand-new, ring-shaped pattern of energy crossings that has never been seen before. This "diabolic ring" is a topological feature where the identity of the ground state swaps around in a loop, a behavior that simply cannot happen in our normal, stable world.

In short, this paper doesn't just give us a better calculator; it teaches us that to understand the wildest, most unstable corners of the quantum world, we need to be flexible. We can't just use the same old tools; we have to choose our mathematical tools carefully, sometimes simplifying them to keep them honest, and sometimes embracing their complexity to see the hidden, swirling patterns of reality that lie just beyond the edge of stability.

Technical Summary: Bridging Frustration and Non-Hermiticity via COMPASS

Problem Statement
Neural Quantum States (NQSs) have emerged as a powerful variational ansatz for strongly correlated many-body systems. However, extending these methods to non-Hermitian (NH) systems presents significant challenges. Traditional Variational Monte Carlo (VMC) fails in NH settings because left and right eigenstates are biorthogonal rather than orthogonal. Existing NH approaches often rely on fixed network architectures, Markov Chain Monte Carlo (MCMC) sampling (which introduces autocorrelation and burn-in issues), and unconstrained complex-valued ansätze. These limitations can lead to optimization instabilities, hidden biases, and spurious results, particularly near exceptional points (EPs) or in systems with specific symmetries like Parity-Time (PT) symmetry. Furthermore, the intersection of NH physics and frustrated magnetism remains largely unexplored, with previous studies focusing primarily on integrable or stoquastic models without frustration.

Methodology: The COMPASS Framework
The authors introduce COMPASS (Complementary Optimization Method for Progressive Adaptive State Search), a framework based on biorthogonal adaptive recurrent neural quantum states. The method integrates three core components:

Adaptive Autoregressive Architecture: The framework employs two independent Recurrent Neural Networks (RNNs) with Gated Recurrent Unit (GRU) cells to represent the right ( $|\Psi_R\rangle$ ) and left ( $\langle\Psi_L|$ ) eigenstates. Unlike static networks, COMPASS utilizes a curriculum learning strategy where the network capacity (hidden dimension) progressively increases during training. Parameters from a converged smaller network are transferred to a larger one, initializing new weights with small random values. This approach mitigates optimization pathologies in large variational spaces and improves scaling.
Biorthogonal Representation: The method explicitly parameterizes independent left and right wavefunctions, satisfying the biorthonormal condition $\langle\Psi_L^i | \Psi_R^j \rangle = \delta_{i,j}$ . The wavefunction amplitudes are factorized using an autoregressive chain rule, allowing for exact sampling without MCMC. This eliminates burn-in periods and autocorrelation effects, providing statistically independent samples and more stable gradient estimates.
Complementary Optimization (Warm-Start Strategy): The optimization alternates between two complementary loss functions:
- Energy Minimization: A biorthogonal energy-based loss ( $L_e$ ) targets the ground-state energy.
- Variance Minimization: A variance-based loss ( $L_v$ ) ensures the state is an eigenstate (zero variance condition).
  The total loss is $L = \lambda_t L_e + (1-\lambda_t) L_v$ . The process begins with pure energy minimization ( $\lambda_t=1$ ) to rapidly approach the ground-state manifold. Once convergence is detected, the weight $\lambda_t$ is adiabatically reduced to prioritize variance minimization, refining the eigenstate quality. This "warm-start" protocol ensures stable convergence to the correct biorthogonal ground-state pair.

Key Contributions and Results

Ansatz Selection and Symmetry Breaking: The study demonstrates that the choice of ansatz is physically decisive in NH systems.
- For PT-symmetric Hamiltonians in the unbroken phase, unconstrained complex ansätze (cRNN) can spontaneously break PT symmetry during optimization, leading to spurious imaginary energies even when the physical system should remain real. Real-valued ansätze (pRNN) in an appropriate basis naturally constrain the optimization to the correct physical manifold, yielding accurate results.
- Conversely, for generic NH Hamiltonians without symmetry protection and with complex spectra, complex ansätze are essential to accurately capture complex ground-state properties. Real ansätze fail to represent the necessary phase information, leading to systematic underestimation of complex observables.
Scalability and Efficiency: By combining adaptive architectures with exact autoregressive sampling, COMPASS achieves significantly faster convergence and lower variance compared to static architectures and MCMC-based methods. The authors report simulations of 1D NH systems up to $N=200$ spins and 2D systems up to $N=100$ spins, representing the largest system sizes reported to date for these dimensions.
Frustration and Non-Hermiticity in the $J_1-J_2$ Chain: The framework was applied to two NH extensions of the frustrated spin-1/2 $J_1-J_2$ $J_{1} - J_{2}$ Heisenberg chain:
1. Staggered Imaginary Field ( $H_1$ ): This model introduces a PT-symmetric staggered field. The results reveal that the frustration gap acts as a quantitative shield against PT-symmetry breaking. The critical threshold for symmetry breaking ( $\gamma_c$ ) scales with the spectral gap ( $\Delta$ ). At the Majumdar-Ghosh (MG) point (maximum frustration), the system exhibits the strongest protection.
2. Complex Next-Nearest-Neighbor (NNN) Coupling ( $H_2$ ): Here, non-Hermiticity is introduced directly into the frustration parameter ( $J_2 \to J_2 + i\gamma$ ). This model exhibits a novel spectral topology: a diabolic ring. Unlike exceptional rings where eigenvectors coalesce, this ring consists of closed curves of real-energy level crossings (diabolic points) where states with distinct imaginary parts cross in their real parts. This structure is controlled by the phase of the complex coupling, not just its modulus, and has no Hermitian analog. The ring encloses a region where the spectrum remains "accidentally real," connecting the MG degeneracy to NH spectral stability.

Significance and Claims
The paper claims to establish a principled framework for the direct study of 1D and 2D NH many-body systems without relying on Hermitian embeddings, adiabatic continuations, or symmetry simplifications. The authors emphasize that physically informed ansatz selection is a foundational requirement for reliable NH simulations, rather than merely a technical detail; maximal expressivity (using complex ansätze everywhere) can be detrimental.

The work positions NQSs as a robust tool for exploring the complex landscape of NH quantum matter, including frustrated systems. By bridging adaptive architectures, biorthogonal optimization, and symmetry-aware modeling, COMPASS enables the discovery of emergent spectral topologies, such as the diabolic ring, which were previously inaccessible to conventional numerical approaches. The authors note that their findings regarding the necessity of complex ansätze for generic NH systems align with recent independent studies on dissipative quantum systems.

Technical Summary: Bridging Frustration and Non-Hermiticity via COMPASS

More like this