Analog Quantum Feature Selection with Neutral-Atom Quantum Processors
José J. Orquín-Marqués, Carlos Flores-Garrigós, Alejandro Gomez Cadavid, Anton Simen, Enrique Solano, Narendra N. Hegade, José D. Martín-Guerrero, Yolanda Vives-Gilabert · 2025
Read Analog Quantum Feature Selection with Neutral-Atom Quantum Processors by José J. Orquín-Marqués, Carlos Flores-Garrigós, Alejandro Gomez Cadavid, Anton Simen, Enrique Solano, Narendra N. Hegade, José D. Martín-Guerrero, Yolanda Vives-Gilabert (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-20798.txt.
Introduction
Jose J Orquin-Marques and seven co-authors pick useful input features with an analog quantum method run on neutral atom arrays. Feature choice matters because extra inputs add noise and cost, while a small clean set can predict just as well and read more clearly.
Relevance of each feature is coded as a local energy shift, and overlap between features is coded through atom interactions limited by a blockade radius. Slow evolution toward low energy states yields bit patterns that mark balanced feature subsets. On three public classification tasks the method matches or beats classic ranking and wrapper methods, and for tiny subsets it lifts accuracy while cutting feature counts by three quarters or more.
Catalog record
| CREATOR | José J. Orquín-Marqués, Carlos Flores-Garrigós, Alejandro Gomez Cadavid, Anton Simen, Enrique Solano, Narendra N. Hegade, José D. Martín-Guerrero, Yolanda Vives-Gilabert |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 23 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-20798
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Analog Quantum Feature Selection with Neutral-Atom Quantum Processors José J. Orquín-Marqués, Carlos Flores-Garrigós, Alejandro Gomez Cadavid, Anton Simen, Enrique Solano, Narendra N. Hegade, José D. Martín-Guerrero, Yolanda Vives-Gilabert Abstract We present a quantum-native approach to quantum feature selection (QFS) based on analog quantum simulation with neutral atom arrays, adaptable to a variety of academic and industry applications. In our method, feature relevance—measured via mutual information with the target—is encoded as local detuning amplitudes, while feature redundancy is embedded through distance-dependent van der Waals interactions, constrained by the Rydberg blockade radius. The system is evolved adiabatically toward low-energy configurations, and the resulting measurement bitstrings are used to extract physically consistent subsets of features. The protocol is evaluated through simulation on three benchmark binary classification datasets: Adult Income, Bank Marketing, and Telco Churn. Compared to classical methods such as mutual information ranking and Boruta, combined with XGBoost and Random Forest classifiers, our quantum-computing approach achieves competitive or superior performance. In particular, for compact subsets, say 2–5 features, analog QFS improves mean AUC scores by 1.5–2.3% while reducing the number of features by 75–84%, offering interpretable, low-redundancy solutions. These results prove that programmable Rydberg arrays offer a viable platform for intelligent feature selection with practical relevance in machine learning pipelines, able to turn computational quantum advantage into industrial quantum usefulness. I Introduction Feature selection (FS) is a crucial step in machine learning pipelines, aiming to identify informative and non-redundant variables that enhance predictive performance while reducing overfitting and computational cost Lucas 2014 ; Li et al. 2017 . The task is inherently combinatorial and known to be NP-hard Date et al. 2021 ; Chandrashekar and Sahin 2014 , leading classical approaches to rely on heuristics or greedy algorithms. Quantum computing offers new possibilities for tackling such problems more efficiently, particularly through quantum annealing, variational algorithms, or analog Hamiltonian simulation Wurtz et al. 2024 ; Morgado and Whitlock 2021 ; Nguyen et al. 2023 . In this work, we propose a quantum-computing FS method based on arrays of neutral atoms in Rydberg states, which support tunable long-range interactions and programmable spatial configurations Henriet et al. 2020 ; Keating et al. 2013 ; Lu et al. 2024 . We encode feature relevance via local detunings, based on mutual information with the target, and redundancy through pairwise Rydberg interactions modulated by interatomic distances. These distances are obtained from a multidimensional scaling (MDS) projection of the feature redundancy matrix Demaine et al. 2021 ; Li 2025 . The resulting system undergoes adiabatic evolution, ending in a ground state that reflects the optimal feature subset. Our approach is implemented in simulation using Amazon Braket and evaluated on real-world classification datasets. Compared with classical methods such as Boruta and mutual information ranking Kursa and Rudnicki 2010 ; Vergara and Estévez 2013 , our quantum method shows competitive performance, particularly when selecting compact feature subsets—highlighting its potential in high-dimensional settings where classical search becomes intractable. The rest of the paper is organized as follows: Section II reviews the process of feature selection in classical machine learning. Section III presents our quantum feature selection (QFS) method via neutral atom encoding. Section IV describes the data embedding process based on feature importance. Section V details the evaluation methodology, and Section VI reports the results. Finally, Section VII provides the conclusions and perspectives for future work. II Feature Selection in Machine Learning Feature selection is a fundamental task in machine learning and data preprocessing, aiming to identify a subset of the most informative variables from a high-dimensional dataset. By retaining only the most relevant features, one can improve model interpretability, reduce overfitting, enhance generalization, and significantly decrease computational burden Guyon and Elisseeff 2003 ; Li et al. 2017 . Mutual information (MI) is one of the most popular metrics for feature selection due to its ability to capture nonlinear dependencies between a feature and the target variable. For a feature X and class label Y , MI is usually represented as I(X;Y) . Features with higher mutual information values are more relevant for classification or regression tasks Chandrashekar and Sahin 2014 . However, selecting features based solely on individual relevance often leads to redundancy, where multiple features carry overlapping information. To address this, modern methods consider both relevance to the target and redundancy among features. The trade-off can be formalized as \max_{z\in\{0,1\}^{N}}\!\left[\sum_{i}I(x_{i};y)z_{i}-\lambda\sum_{i<j}I(x_{i};x_{j})z_{i}z_{j}\right], (1) where z_{i} indicates whether feature i is selected and \lambda balances relevance and redundancy. This formulation naturally maps to a Quadratic Unconstrained Binary Optimization (QUBO) problem, solvable by classical heuristics or quantum-enhanced approaches Lucas 2014 ; Mücke et al. 2023 . Recent studies have explored quantum annealing, variational and evolutionary quantum algorithms, and analog quantum simulators to solve QUBO-based feature selection tasks more efficiently Date et al. 2021 ; Albino et al. 2023 ; Hamed et al. 2011 ; Vivek et al. 2024 . In this work, we exploit the analog simulation capabilities of neutral atom systems to implement a quantum-native feature selection scheme that encodes both mutual information and redundancy into the spatial arrangement and control parameters of a Rydberg atomic array. III Neutral Atom Encoding for Quantum Feature Selection Neutral atoms excited to Rydberg states have emerged as a promising hardware platform for quantum computation and simulation, particularly in analog quantum optimization. These systems combine excellent scalability, tunable long-range interactions, and flexible spatial programmability D’Arcangelo et al. 2023 ; Phuttitarn et al. 2024 ; Lu et al. 2024 ; Su et al. 2024 . Individual atoms are confined by optical tweezers and coherently manipulated via laser pulses that drive transitions between the ground state \ket{g} and an excited Rydberg state \ket{r} , forming an effective qubit. Each atom behaves as a two-level system with \ket{g}\!\equiv\!\ket{0} and \ket{r}\!\equiv\!\ket{1} . The time-dependent Hamiltonian for an array of N atoms is \displaystyle H_{\text{Ryd}}(t) \displaystyle=\frac{\Omega(t)}{2}\sum_{i}\!\left(e^{i\phi(t)}\ket{g_{i}}\!\bra{r_{i}}+\text{h.c.}\right)-\Delta_{g}(t)\sum_{i}\hat{n}_{i} (2) \displaystyle-\Delta_{l}(t)\sum_{i}p_{i}\hat{n}_{i}+\sum_{i<j}V_{ij}\hat{n}_{i}\hat{n}_{j}, where \Omega(t) and \phi(t) are the Rabi frequency and phase of the driving laser, \Delta_{g}(t) and \Delta_{l}(t) implement global and local detunings, p_{i} encodes feature-dependent weights, and V_{ij} are van der Waals interactions given by V_{ij}=\frac{C_{6}}{|\vec{x}_{i}-\vec{x}_{j}|^{6}} , with C_{6} determined by the principal quantum number. The strong distance dependence of V_{ij} leads to the Rydberg blockade : if two atoms are closer than a critical distance R_{b} , their joint excitation is suppressed since V_{ij} exceeds the laser bandwidth. The blockade radius is R_{b}=\left(\frac{C_{6}}{\sqrt{\Omega^{2}+\Delta^{2}}}\right)^{1/6} . This mechanism enforces spatial exclusion, making Rydberg arrays natural candidates for constraint-based optimization problems such as Maximum Independent Set (MIS) Dalyac and Henriet 2022 ; Goswami et al. 2024 ; Gonzalez-Cuadra et al. 2024 . Here, each atom represents a feature, and spatial arrangement encodes redundancy constraints. Compared with superconducting qubits or trapped ions, neutral atoms offer scalability to hundreds of sites, flexible geometry, tunable interactions, and efficient fluorescence-based readout Johnson and Rolston 2010 ; Saffman et al. 2010 ; Farouk et al. 2024 ; Llenas and Lamata 2024 . III.1 Mapping Feature Selection to the Rydberg Hamiltonian The FS objective combining relevance and redundancy is defined as Q(x;\alpha)=-\alpha\sum_{i}I_{i}x_{i}+(1-\alpha)\sum_{i<j}R_{ij}x_{i}x_{j}, (3) where I_{i}=I(x_{i};y) and R_{ij}=I(x_{i};x_{j}) . The final Hamiltonian becomes H_{\text{Ryd}}(T)=-\Delta_{l}^{\text{max}}\sum_{i}p_{i}\hat{n}_{i}+\sum_{i<j}V_{ij}\hat{n}_{i}\hat{n}_{j}, (4) linking feature relevance to detunings and redundancy to interaction strengths. III.2 Adiabatic Evolution Protocol Starting from all atoms in \ket{g} , we interpolate adiabatically toward the problem Hamiltonian. Schedules