Skip to main content
Back to timeline
arXivSource publication:

Foundation tensor networks generate ground states straight from the Hamiltonian, pinning glass-transition parameters across 500,000 disorder realizations

Synopsis

In disordered quantum systems, every disorder realization demands a fresh ground-state solve, so critical-phenomenon studies carry two layers of averaging. Foundation tensor networks (FTNs) use a transformer to generate the projected entangled pair state of the ground state directly from a graph embedding of the Hamiltonian, amortizing optimization over the whole Hamiltonian distribution. With GridBP and replica moment-generating functions, generating, fine-tuning and measuring the overlap parameter for one sample takes under three seconds, and finite-size scaling over 500,000 disorder realizations yields the critical parameters of the quantum Edwards-Anderson model.

Source-provided article image: Foundation Tensor Networks Generalize Across Quantum Lattice Systems
arXiv

Interpretation

Ground states of disordered quantum lattice systems no longer need to be solved one realization at a time: a model trained over a Hamiltonian distribution outputs the ground-state wavefunction for a new disorder realization directly. Neural quantum states could already serve as foundation models across Hamiltonians but still required extensive sampling to estimate observables, while tensor networks contract deterministically yet had to find the ground state separately for each disorder realization. On the transverse-field quantum Ising spin glass (EA-TFQIM) on a two-dimensional square lattice with open boundary conditions, ground states generated by inference alone were as accurate as or more accurate than simple-update imaginary-time evolution in almost all cases, and faster.

The wavefunction comes out as a projected entangled pair state (PEPS), a network of tensor cores arranged to mirror the lattice topology, so observables are computed by tensor contraction rather than sampling from the Born distribution. The generator is a transformer: each site's local couplings, transverse field and position are encoded as tokens, passed through transformer blocks into latent states, and read out by a multilayer perceptron into tensor cores. Training used transfer learning across system sizes with one shared architecture, ran in roughly hours on a single Nvidia A100 GPU, and was checked against exact diagonalization and DMRG results.

Computing the energy expectation value drops to two tensor contractions, and the per-sample overlap parameter drops to a single contraction, making large-scale disorder averaging tractable. The usual routes either build a projected entangled pair operator, which enlarges the bond dimension, or evaluate the Hamiltonian term by term, which scales quadratically with the number of sites; the overlap parameter computed pair by pair would need a quadratic number of contractions. GridBP iterates one-dimensional belief propagation with cached messages, costing the same as two contractions of the bilayer PEPS, and the replica moment-generating function uses a Hessian-vector product to reduce the overlap parameter to one automatic-differentiation evaluation.

Perspective

The demonstrated setting is the transverse-field quantum Ising spin glass on a two-dimensional square lattice with open boundary conditions, independent identically distributed Gaussian couplings with zero mean and unit variance, and a fixed bond dimension with a contraction cutoff; critical parameters come from finite-size scaling fits with bootstrap confidence intervals.

The critical-parameter estimates match and refine literature values that themselves carry significant uncertainty; post-inference energy optimization did not significantly improve results, indicating the PEPS output is already near-optimal at its bond dimension, a judgment limited to the tested bond dimension and system sizes.