Skip to main content
Back to timeline
arXivSource publication:

A quantum sampler learns a whole phase diagram classically: HyperQCBM matches 1,500 independent updates after about 550 local ones on a twelve-site chain

Synopsis

Quantum circuit Born samplers previously had to be fitted separately at each Hamiltonian point, with nothing reused across a phase diagram. HyperQCBM maps Hamiltonian parameters to circuit angles through a hypernetwork, trains jointly across several ground-state distributions, then fine-tunes with local reference data. At 40 boundary points of a twelve-site Haldane chain, adaptive pretraining plus fine-tuning reached a final mean total variation distance of 0.216 versus 0.237 for independent training, matching the independent fits' final value after about 550 local updates instead of 1,500.

Source-provided article image: Classically Trained Quantum Generative Models with Shared Training and Fine Tuning for Quantum Matter
arXiv

Interpretation

A quantum circuit sampler can now cover a whole family of ground-state measurement distributions instead of being fitted from scratch at each Hamiltonian point. The prior route trained circuit angles independently at a single parameter point, sharing no information between points. Across 40 boundary coordinates of the twelve-site Haldane chain, adaptive pretraining plus fine-tuning gave a final mean total variation distance of 0.216, against 0.237 for independent training and 0.234 for angle-only fine-tuning, and the advantage held in each of four initializations.

The sharing is carried by a hypernetwork that turns normalized Hamiltonian parameters into a scale and shift of shared base angles. The hypernetwork and base angles are trained jointly, with each update averaging losses over several reference points, so what is learned is the map from parameters to angles. On a 1,024-site chain and a 4 by 256 lattice, the fine-tuned circuits had lower squared maximum mean discrepancy from the reference distributions than independently trained circuits at the evaluated point in each geometry, about one third of the independent error on the chain and about one seventh on the lattice.

The model coefficients needed for training are estimated by Monte Carlo, without constructing the full output distribution. The earlier phase-averaging estimator applied only to commuting IQP generators and did not carry over to noncommuting Pauli rotation circuits. The estimator is unbiased for ordered Pauli rotation circuits and its root mean squared sampling error is bounded independently of qubit count, but the number of amplitudes each sample evolves is set by the generators, and restricting active sites brings that number down to polynomial size.

Perspective

The demonstrated scope is limited to spin model families such as the transverse-field Ising model and the Haldane chain, with exact diagonalization supplying small-system references and matrix product states capped at bond dimension 128 supplying references for the 1,024-site chain and the 4 by 256 lattice; the large-system comparison is made at one evaluation point per geometry, while the twelve-site fine-tuning comparison covers 40 boundary coordinates and four initializations.

The large-system references are finite-bond matrix product state approximations that reached the sweep limit at the evaluation point, so held-out samples measure agreement with that approximate ground state; on the twelve-site chain, independent training actually gave lower per-coefficient error at order eight, showing that matching overall correlation strength does not guarantee the magnitude or sign of individual coefficients.