Resource-Aware Quantum Big Data Analytics: A Hybrid Quantum–Classical Framework And Scalability Study For Post-Exascale Systems

Abstract

Quantum processors are often proposed as accelerators for data-intensive analytics, yet the cost of loading classical data, repeating circuits, and reading out measurements is rarely included when such proposals are assessed. This paper presents Q-BDA, a five-layer framework that places a quantum subroutine inside a hybrid pipeline and judges it with an end-to-end cost model covering encoding, circuit execution, shot sampling, communication, and classical post-processing. The framework is tested with a fidelity quantum kernel built from a two-repetition ZZ feature map and simulated exactly, without noise, on a single CPU core. Over ten random splits, the quantum-kernel support-vector machine reached 94.5%, 98.7% and 92.5% test accuracy on the Moons, Circles and Breast Cancer data, against 97.2%, 99.5% and 94.2% for a tuned RBF kernel. A kernel build needs one circuit per matrix entry, 291,480 circuits at N = 800, and explicit pairwise simulation was 84 times slower than a statevector shortcut that has no hardware counterpart. When the number of qubits grew from 2 to 12, off-diagonal kernel values fell from 0.39 to 0.0009, test accuracy on Breast Cancer fell from 94% to 73% while the RBF baseline stayed above 94%, and the shots needed to resolve a typical kernel value rose from 1.6 to 1,200. A parametric projection suggests that direct kernel construction for 10⁴ samples would occupy a superconducting-type device for roughly nine days to five months. The results support treating quantum hardware as a workload-selective accelerator, and they provide a reproducible protocol for deciding when it is worth using.