Title: QADR: A Scalable and Trainable Quantum Machine Learning Framework for Distributed Entanglement Reduction

Abstract

The training of Variational Quantum Circuits (VQCs) within the Noisy Intermediate-Scale Quantum (NISQ) era is hindered by significant computational bottlenecks. Specifically, classical simulation of statevectors requires memory that grows exponentially with the number of qubits, scaling as $\mathcal{O}(2^n)$. Furthermore, global cost functions are prone to barren plateaus, a phenomenon where the variance of gradients diminishes exponentially, following the rate $\mathcal{O}(1/2^n)$. To address these challenges, this study presents the Quantum Algorithm for Distributed Reduction of Entanglements (QADR). This hybrid quantum-classical machine learning approach partitions a global $n$-qubit VQC into localized sub-circuits. These sub-circuits function primarily within the causal light cones surrounding individual target qubits.

By adopting this distributed architecture, QADR significantly lowers the memory requirements for classical simulation, reducing the scaling from $\mathcal{O}(2^n)$ to $\mathcal{O}(n \cdot 2^{2d+1})$, where $d$ represents the radius of the light cone. This method also inherently alleviates the issue of global barren plateaus. We evaluated the performance of QADR against standard global VQCs, Support Vector Machines (SVM), and two bespoke classical neural networks with matched parameters (CANN and PMNN). The benchmarks utilized the MNIST dataset and a high-dimensional diagnostic task involving NASA IMS wind turbine drivetrain data. The results highlight QADR’s robust scalability; it operated successfully with $n_{\text{features}}=2000$, a scale at which standard global VQCs failed due to memory exhaustion. Moreover, QADR achieved performance levels that matched or surpassed those of the optimized classical models.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC