EntNet spots many-body entanglement from low-order spin correlations alone, using 16 local-unitary invariants that carry across chain lengths and Hamiltonian families
Synopsis
Detecting many-body entanglement no longer requires reconstructing the full density matrix or relying on a predefined witness. EntNet turns two- to four-body Pauli correlations on a one-dimensional spin chain into a representation invariant under local basis rotations, sparsely selects 16 feature channels, and learns their spatial arrangement with a Transformer using rotary positional embeddings. On a finite-temperature XY benchmark it reaches 100% AUROC from expectation values of these low-order observables alone and holds the false-positive rate to 0.003 on 2,000 constructively separable states; the same model transfers without retraining to XXZ, frustrated J1-J2 and random local Hamiltonians and to longer chains.
Interpretation
Entanglement can now be judged directly from low-order spin correlations that experiments can measure, without prior quantum-state tomography. Earlier entanglement criteria generally required density-matrix information or depended on predefined witness constructions and a specific choice of observables. On the finite-temperature XY benchmark EntNet reaches 100% AUROC, and on 2,000 constructively separable states with nontrivial classical correlations it reaches a true-negative rate of 0.997.
The model arranges each lattice site's correlation features into a sequence and compares correlation patterns across positions with self-attention. Local basis rotations change raw Pauli correlations, so the input instead uses local-unitary invariants such as singular values of tensor unfoldings, giving the same features for one physical state in different bases. Training uses 35,000 synthetic spin-chain states, half separable and half certified entangled; sparse gating distills 117 candidate invariant channels into 16, and the model has about 480,000 parameters.
The same frozen 16-channel model applies to other Hamiltonian families and longer chains without retraining. On XXZ, frustrated J1-J2 and random local Hamiltonian benchmarks the AUROC is near 1, and the model also identifies higher-order certified entangled states. The main drop appears in the J1-J2 J2 states, where recall is 0.86 while AUROC stays at 1.0000; adding training states with strong next-nearest-neighbor correlations restores that recall to 1.0000 with the same 16 selected channels.
Perspective
The results cover one-dimensional spin chains, with training and evaluation states synthetic or numerically computed and inputs given as exact correlation expectation values; finite sampling, readout errors and imperfect state preparation are not yet included. The authors propose extending the representation to two- and three-dimensional lattices and to fermionic systems, and reducing the required measurement set to suit programmable quantum simulators such as superconducting-qubit arrays and trapped-ion systems, as well as scattering experiments.
The authors state that EntNet's field-dependent lower boundary at finite temperature characterizes the model's sensitivity to thermally activated correlation signatures rather than an exact separability boundary; in the XX limit at fixed positive temperature, Gibbs states at different finite fields are related by invertible local transformations, so exact separability does not depend on the field. The absence of explicit pair channels at separations 3 and 4 in the 16-channel representation comes from training on 16-site chains, and whether it persists in the thermodynamic limit remains to be established; a smaller feature dimension also does not imply a proportional reduction in Pauli measurements.