Create and explore imaginative 3D worlds with our new product Dreamfield (opens in a new tab)

Determine the cause of cuPauliProp gradient discrepancies

Identify the cause of the discrepancy between cuPauliProp's expectation-value agreement and its gradient deviations from the central-difference reference on the 9- to 12-qubit benchmark instances and on the benchmark with batched embedding inputs.

Background

The paper compares PADO-Pauli and NVIDIA cuPauliProp against central-difference gradient references on untruncated benchmark circuits and on a benchmark involving batched embedding inputs. Although cuPauliProp's expectation values agree with the reference values, its gradients show substantially larger deviations at 9 to 12 qubits and on the batched-embedding benchmark. The paper reports the discrepancy but does not determine its source, leaving the causal explanation unresolved.

References

As driven by our harness, cuPauliProp's gradient matches the reference to $1.6\times10{-7}$ at $4$ and $6$ qubits but deviates by up to $1.8\times10{-2}$ at $9$ to $12$ qubits and by $3.0$--$3.5\times10{-2}$ on the instance with batched embedding inputs below, while its expectation values match throughout; we did not trace the cause, and its backward timings are quoted as measured.

— PADO-Pauli: A GPU-Accelerated Implementation of Pauli Propagation  (2610.17755 - Kim et al., 13 Oct 2026) in Section 5.2, paragraph “Gradient accuracy”