arXiv:2610.17755
PADO-Pauli: A GPU-Accelerated Implementation of Pauli Propagation
Hyunwoo Kim, Youngseok Lee
Abstract
Variational quantum algorithms utilize parameterized circuits many times during training, at new parameter values, over batches of data, and with gradients. Pauli propagation approximates the expectation values of such circuits classically by evolving the observable in the Heisenberg picture as a truncated sum of Pauli strings. We present PADO-Pauli, a GPU engine for Pauli propagation that compiles a circuit and its observables once into sparse matrices in which each rotation angle contributes only a cosine and a sine factor, so every later evaluation reuses the compiled program. It evaluates batches of parameters and data inputs embedded as rotation angles. By default its gradients come from a vector-Jacobian product registered with PyTorch's automatic differentiation, so they flow through the angles to any classical layer that produces them. PADO-Pauli also provides truncation by Pauli weight and coefficient size, and noise channels. Zero-filtering removes terms that do not contribute to an expectation value in the all-zero reference state and makes the batched forward pass about 1.5 times faster. Compared to cuPauliProp on an NVIDIA A100, PADO-Pauli evaluates up to 10 times faster on 4 to 12 qubits and uses up to 3.1 times less memory on 16 qubits. We check its accuracy against exact statevector simulation up to 28 qubits. These features make PADO-Pauli a practical toolkit for training variational quantum algorithms, including quantum machine learning models.
AI Summary
Pauli propagation approximates circuit expectation values by evolving observables as truncated Pauli sums. This paper addresses repeated evaluations during variational optimization and quantum machine learning, where parameters and embedded data change but circuit structure remains fixed. PADO-Pauli compiles circuits and observables into per-gate sparse maps, separating fixed structure from angle-dependent $\cos\theta$ and $\sin\theta$ factors. Evaluation supports GPU batching and a custom vector–Jacobian product integrated with PyTorch. Compilation includes Pauli-weight truncation, optional coefficient truncation at a fixed build point, and exact backward pruning of branches that cannot contribute in $|0\rangle^{\otimes n}$. Depolarizing and amplitude-damping channels are supported. On NVIDIA A100 benchmarks, warm evaluations are up to $10.1\times$ faster than cuPauliProp on 4–12-qubit instances. At 16 qubits and $w_{\max}=6$, forward-plus-backward memory occupancy is $3.1\times$ smaller, although cuPauliProp uses complex rather than real coefficients. An internal ablation attributes approximately $1.5\times$ batched-forward acceleration to zero-filtering. Exact statevector checks reach 28 qubits; on small-angle QAOA instances, $w_{\max}=5$ gives absolute observable errors below $9.74\times10^{-4}$. Truncation accuracy remains instance-dependent: a random-circuit test reaches worst-case error $0.204$ at 20 qubits with $w_{\max}=9$. Compilation and initial backward-cache construction impose upfront costs, coefficient-truncated programs remain tied to their build point, and branching can exhaust memory. The 127-qubit kicked-Ising demonstration lacks an exact reference; apparent convergence under relaxed truncation can be misleading. The paper reports no new general simulation guarantee.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 42 pages, 13 figures, 10 tables. Code available at https://github.com/Norma-Q/PADO-Pauli