The challenge of representing complex quantum states in multiparticle systems has limited progress in fields such as materials science and quantum chemistry. Felix Frohnert, Emiel Koridon, and Stefano Polla at Leiden University have developed a new unsupervised method to discover highly compressed representations of these states. Their approach uses neural networks to identify the smallest “latent space” needed to accurately describe the ground state of a system, effectively matching the number of essential properties. This breakthrough avoids a fundamental problem in quantum simulation and allows scientists to directly minimize energy within this simplified space, promising more efficient and accurate modeling of complex quantum materials.
By applying an autoencoder neural network to Fermi-Hubbard model data, the researchers identified a minimal latent space that accurately represents the ground state manifold, reducing the computational costs associated with storing and manipulating these complex states. This approach effectively captures the essential correlations within the ground state, enables accurate reconstruction of the wave function from a compressed latent representation, and paves the way for scalable simulations of strongly correlated fermion systems, which are central to understanding matter with emergent quantum phenomena.
The reconstruction quality of this method reaches a threshold at latent dimension L-1, which matches the inherent degrees of freedom of the system. The trained decoder acts as a differentiable variational analyzer, allowing direct energy minimization within the latent space. Importantly, this approach avoids the challenge of ensuring physical plausibility, as the learned manifold implicitly limits the optimization to physically valid quantum states. The exponential growth of Hilbert space with system size poses fundamental challenges in quantum many-body physics, making direct simulation difficult for all but the smallest systems, but approximate methods provide a viable path.
The ground state of the Hubbard model exists on a low-dimensional manifold
This study demonstrated that the ground state of the Hubbard model, the standard model in condensed matter physics, lies on a surprisingly low-dimensional manifold. Specifically, the intrinsic dimension of the ground state manifold is approximately L-1. where L is the number of lattice sites in the system. This means that the ground state can be accurately represented with far fewer parameters than previously expected. Unsupervised autoencoders, a type of neural network, have proven highly effective at learning these compressed representations and encoding important information into low-dimensional latent spaces. The number of dimensions of L-1 is important. Below this dimension, the quality of the reconstruction decreases significantly, but above it the latent space becomes unstable. This compressed latent space serves as a variational analysis for energy optimization, allowing researchers to efficiently find approximate ground states by minimizing the energy in the latent space and decoding back to the original system.
The researchers also successfully applied this compression approach to two-body reduced density matrices, which contain more detailed information about ground state correlations. The intrinsic dimension remains L-1, suggesting that additional latent dimensions are not required for additional correlation information. Systems with degeneracy in the ground state (e.g., systems with an odd number of sites) exhibit more complex latent space structures and slightly worse reconstruction quality, highlighting the need for autoencoders that take the underlying symmetries and degeneracy into account. This research has several important implications, including the possibility of efficient quantum simulation, the discovery of new quantum phases, improved variational algorithms, and a deeper understanding of quantum complexity.
Learned latent space captures quantum ground state
This work demonstrates an unsupervised machine learning framework that can discover compressed representations of complex quantum many-body ground states. By applying an autoencoder neural network to the Fermi-Hubbard model, the researchers identified the smallest latent space that accurately reconstructs the state of the system. These spatial dimensions correspond to the inherent degrees of freedom of the system. This approach avoids limitations associated with the representation of quantum states by implicitly restricting optimization to physically plausible configurations within the learned latent space. Analysis of the learned representation reveals a clear hierarchy of feature importance, determined by examining the influence of different physical observations on the latent coordinates.
Results consistently show that density and on-site interaction terms dominate the latent representation across a range of system sizes, while nearest-neighbor correlation plays a relatively minor role. This suggests that the autoencoder prioritizes local quantities that are directly related to the system's possibilities and interactions when constructing a compact description. Systems with an odd number of sites, tested as a means of exploring the limits of the compression approach, exhibited slightly higher reconstruction losses compared to systems with an even number of sites, indicating room for future improvements.
