arXiv:2610.32571
Classical Pre-Training of Large Shallow Parametrized Quantum Circuits through Pauli Propagation and deployment on Quantum Hardware
Saverio Monaco, Jamal Slim, Ran Xue, Dirk Krücker, Kerstin Borras
Abstract
Scaling parametrized quantum circuits to a larger number of qubits is a central challenge in Quantum Machine Learning: gradient estimation via the parameter-shift rule quickly becomes prohibitively expensive, while gradient-free optimizers offer limited convergence guarantees. We apply the Pauli propagation framework to a broad class of Quantum Machine Learning problems in which the loss is expressed as a function of expectation values of the circuit's observables, while keeping the relevant quantity of interest, such as the quantum state or its samples, classically intractable to compute. By propagating the observables backward through the circuit, we obtain an explicit parametric expression for the loss that can be minimized entirely on classical hardware. We also present complementary strategies to make propagation tractable at large scale, including weight-based and degree-based truncations and pruning heuristics. We demonstrate the approach on two benchmarks run on 64-qubit circuits: Variational Quantum Eigensolver on a two-dimensional Ising model, and training of a discrete-type generative model for binary images. In both cases, the classically pre-trained parameters are deployed on the IBM quantum processor ibm_berlin. For generative learning, we additionally show that the circuits considered here require substantially fewer parametrized gates than IQP-style ansätze, reducing transpilation overhead and making deployment on current quantum hardware feasible.
AI Summary
The paper asks whether large, shallow parametrized quantum circuits can be trained classically using observable-based losses, then deployed on quantum hardware. It backward-propagates observables through the circuit to construct explicit trigonometric functions of the parameters, optimized with classical gradients. Pauli-weight and trigonometric-degree cutoffs control expression growth; two exact pruning rules remove branches that cannot contribute to the initial-state expectation value. Demonstrations use 64-qubit circuits on an $8\times8$ grid, with deployment on IBM’s `ibm_berlin`. For VQE on the two-dimensional transverse-field Ising model at $(J,h)=(1,2)$, depth-1 and depth-2 ansätze have 256 and 384 parameters. Training uses 1,000 Adam steps, with exact propagation at depth 1 and weight cutoff $w_{\mathrm{cut}}=12$ at depth 2. Hardware energies remain above the propagated energies; depth 2 converges faster classically but gives slightly worse deployed energy. For binarized MNIST generation, a 256-parameter circuit is trained using an observable-based MMD loss over 100,000 Adam steps. Truncated-propagator evaluation gives test $\mathrm{MMD}^2=(2.05\pm0.18)\times10^{-3}$, versus $(24.11\pm1.16)\times10^{-3}$ and $(2.51\pm0.22)\times10^{-3}$ for IQP baselines with 256 and 2,560 parameters. These scores are not hardware-sample evaluations. Hardware-generated images include recognizable digits and indistinct patterns. Stated limitations include expensive propagation of many generative-loss observables, approximate training, limited shallow-circuit expressivity, and hardware noise. Single-site statistics are reproduced more closely than pairwise correlations. The implemented ansatz is not strictly locally scrambling, so cited average-case truncation guarantees do not directly apply; general-purpose scalability remains open.
Central claim may not hold: The abstract's claim of retaining classically intractable states or samples is unsupported by demonstrating 64-qubit circuits beyond convenient full-statevector simulation. The weight-truncation justification neither directly applies to the demonstrated non-locally-scrambling ansatz nor guarantees accuracy at parameters chosen by surrogate optimization. The truncated depth-2 VQE and generative objectives lack a reported independent accuracy assessment at the trained parameters, despite useful exact depth-1 and hardware demonstrations. The generative comparison uses different expectation-evaluation methods for the proposed model and IQP baselines without quantifying propagation bias, so its uncertainties do not establish that the reported advantage reflects modelling error alone.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.