Contributed Session: Quantum Machine Learning
Paper ID: 13
Authors: Hrvoje Kukina and Susanne Yelin
Title: Parameterized Quantum Circuits for Deep Reinforcement Learning
Abstract: Deep reinforcement learning (DRL) is a standard approach to sequential decision-making under uncertainty, where value functions are often approximated with deep neural networks such as Deep Q-Networks (DQNs). Parameterized quantum circuits (PQCs) have emerged as a flexible modeling primitive for near-term quantum hardware, enabling hybrid quantum–classical predictors in which a PQC produces bounded expectation-value features that are processed by classical layers. A practical limitation of many quantum learning approaches, however, is the need for quantum hardware not only during training but also at deployment time. In this work, we study PQC-driven value-based agents in the CartPole control benchmark. Using a common DQN-style training loop, we compare classical DQN baselines, a PQC-based Q-network, and a hybrid workflow that leverages classical shadows for evaluation-only deployment. To probe near-term practicality, we further examine robustness by injecting additive Gaussian perturbations into the quantum feature vector across multiple noise amplitudes.
Paper ID: 36
Authors: Gianluca Scanu, Luca Barletta and Stefano Rini
Title: JGRA: Jacobian Geometry Robustness Assessment in NISQ Noise-Aware Quantum Neural Networks
Abstract: The NISQ era places stringent constraints on quantum computation, where noise and decoherence fundamentally limit performance. In classical deep learning, model robustness and resilience to perturbations are well studied: deep neural networks (DNNs) maintain high performance despite pruning, noise injection, and structural perturbations due to inherent redundancy in their representations. A central challenge in quantum machine learning is to transfer this notion of robustness to quantum neural networks (QNNs) under realistic NISQ noise. While classical deep learning exhibits robustness through structural redundancy, analogous principles for QNNs remain underdeveloped. We propose JGRA: a framework for assessing robustness in noise-aware QNNs via Jacobian geometry, capturing model sensitivity to parameter perturbations induced by noise. Our method includes entropy-matched noise calibration, noise-aware training, and noise-conditioned Jacobian extraction, yielding geometric descriptors that link clean-regime structure to noisy inference behaviour. We also empirically demonstrate that these descriptors encode predictive information about robustness under unseen noise.
Paper ID: 57
Authors: Mirko Legnini and Julian Berberich
Title: Noise Resilience and Robust Convergence Guarantees for the Variational Quantum Eigensolver
Abstract: Variational Quantum Algorithms (VQAs) are a class of hybrid quantum-classical algorithms that leverage on classical optimization tools to find the optimal parameters for a parameterized quantum circuit. One relevant application of VQAs is the Variational Quantum Eigensolver (VQE), which aims at steering the output of the quantum circuit to the ground state of a certain Hamiltonian. Over the last few years, significant research effort has been devolved in characterizing the optimization landscape of such problems. In particular, relevant problems include the classification of singular points and convergence guarantees. In our paper (available at https://doi.org/10.48550/arXiv.2601.16758), we investigate convergence guarantees when an ideal VQA is affected by noise, both coherent and incoherent. We provide upper bounds for the distance between the perturbed and the ideal optimal parameters. Furthermore, we derive convergence guarantees for perturbed VQAs. Last, we address the specific case of depolarizing noise and coherent control errors. Our results are supported by numerical simulations implemented via Pennylane.
Paper ID: 96
Authors: Iwo Wojtakajtis, Maria Płatek, Rafał Balicki, Karina Leśkiewicz, Michał Szczęsny and Tomasz Kajdanowicz
Title: Hardware-Efficient Ansatz Design and Noise-Aware Analysis of a Variational Quantum Classifier for IQM Spark
Abstract: While hardware-efficient ansätze have been introduced in quantum machine learning (QML), a critical gap remains regarding the systematization and documentation of design intuitions, coupled with a lack of concrete empirical testing and evaluation on physical quantum processing units (QPUs). Consequently, there is a need to demonstrate concrete methods for evaluating ansätze design. Standard, hardware-agnostic models still dominate the literature, often incurring massive transpilation overheads that severely degrade circuit fidelity. In this empirical study, we demonstrate that to effectively mitigate these overheads, ansatz design should be strictly tailored to the native gate set of the target device. We present a comprehensive comparison on the binary classification benchmark, evaluating a simulator-oriented ansatz against a hardware-aligned variant tailored to the native entanglers of the IQM Spark. Moving beyond simple training accuracy, our framework evaluates five complementary dimensions: (1) compiled resource costs (2) estimated fidelity proxies; (3) theoretical expressibility via Kullback-Leibler divergence to the Haar distribution; (4) optimization robustness via five-fold cross-validation under a phenomenological expectation-value noise model; and (5) end-to-end classification performance directly on the physical QPU using Accuracy and F1 Metrics. Our evaluation reveals that the hardware-aligned ansatz reduces compiled circuit depth by 60%, requiring only 19 native entanglers versus 26 in the simulator-oriented baseline. This translates to superior end-to-end classification performance on the physical IQM Spark QPU, where the tailored model achieves 82.9% accuracy (F1: 0.783), outperforming the standard approach's 75.7% (F1: 0.640). The physical performance advantage is underpinned by higher estimated fidelity proxies (80.35% vs 66.31%) KL-divergence analysis reveals a depth-depended tradeoff expressibility - at 2, the hardware-aligned ansatz achieves higher expressibility (0.096967 vs. 0.068663), while at depth 4 the simulator oriented variant recovers an advantage (0.000893 0.002184), but the difference is irrelevant. In both cases the differences do not translate into measurable classification performance gaps on physical QPU. Our results establish hardware-native co-design as a strategic necessity rather than a theoretical compromise. While hardware-efficient models remain competitive in simulation, they demonstrate clear superiority on physical hardware. We release out five-dimensional evaluation framework as a reproducible methodology for benchmarking QML ansätze on near-term QPUs.
Paper ID: 85
Authors: Friso Meeusen, Ivana Nikoloska and Rianne Lous
Title: Quantum Graph Fourier Neural Operator
Abstract: Ordinary and partial differential equations (ODEs and PDEs) serve as the mathematical description of how numerous physical processes evolve, e.g., in fluid dynamics or quantum mechanics. Traditionally, these equations are solved numerically, but there has been an increasing interest in using machine learning algorithms to replace or speed-up traditional methods. One type of algorithms, referred to as neural operators, learn a mapping of function spaces between the input function and the corresponding solution of the PDE found numerically. The rise of quantum computing and quantum machine learning, enables opportunities to develop quantum neural operators for more efficient data-processing. In this work, we propose Quantum Graph Fourier Neural Operators (QGFNO) that represent the underlying PDE domain as a graph. We train the QGFNO using different and PDE systems with increasing complexity, both on noiseless simulators as well as on emulators of neutral atom and photonic quantum computing devices. We present initial results on the general learning capability of the quantum circuit and the relation to its hyperparameters, types of noise, and noise levels. These results will show further insights regarding the extent to which quantum neural operators provide an advantage in solving PDEs, if any, over their classical counter-part as well as providing a first example of this more advanced quantum machine learning application running on hardware.
Paper ID: 46
Authors: Jiarui Hu, Wenzhi Chen, Shipra Singh and Lijun Chen
Title: An Efficient Quantum Algorithm for Sparse Linear Bandits via Smoothed Exploration
Abstract: Stochastic linear bandits are an important model for sequential decision-making, yet classical algorithms are constrained by a O(d * sqrt(T)) regret lower bound. While recent quantum bandit algorithms achieve exponential improvements in the time horizon T, they usually do not exploit sparsity in the system, resulting in polynomial scaling with the ambient dimension d. We propose Q-Lasso-OFU, an efficient quantum algorithm that achieves a regret that scales linearly with sparsity s while depending weakly on the ambient dimension d. Our algorithm leverages a smoothed exploration and a variance-adaptive weighted Lasso for efficient sparse recovery. We show that it achieves a regret bound of O(s sqrt(log d) log T). This result presents the first provable quantum speedup that simultaneously overcomes the sqrt(T) bound of the classical algorithms and mitigates the curse of dimensionality in linear bandits. Numerical experiments are provided to complement the theoretical analysis.