Role of quantum hardware noise in PQFM-driven performance gains

Determine the exact role of intrinsic quantum hardware noise during execution of the Heisenberg-ansatz Projected Quantum Feature Map (PQFM) circuits on IBM Quantum Heron hardware (e.g., ibm_torino) in shaping the transformed feature vectors and in producing the observed out-of-sample AUC improvements for estimating fill probabilities of corporate bond RFQ responses, relative to noiseless quantum simulation and classical features.

Background

The paper reports substantial uplifts in out-of-sample AUC when machine learning models are trained on PQFM features generated by noisy quantum hardware, particularly with the “longer” Heisenberg circuit on the IBM Heron r1 device (ibm_torino), compared with both classical features and noiseless quantum-simulated PQFM features.

Despite reproducibility across multiple Heron devices and variations of the circuit (including preserving decoherence periods with idle operations), the authors were unable to reproduce these uplifts via classical simulations with artificially induced noise, and note that the precise impact of hardware noise on the PQFM outputs and downstream model performance remains unknown.