Provably Robust Training of Quantum Circuit Classifiers Against Parameter Noise

Lucas Tecot, Di Luo, Cho-Jui Hsieh · 2025

Read Provably Robust Training of Quantum Circuit Classifiers Against Parameter Noise by Lucas Tecot, Di Luo, Cho-Jui Hsieh (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-2505-18478.txt.

Introduction

Lucas Tecot, Di Luo, and Cho-Jui Hsieh write a 2025 paper on quantum computing. Quantum circuit classifiers sort data using circuits whose settings can be tuned by training. Noise in those settings is a chief barrier to trust in near-term quantum machines.

The team gives a training theory and method with proof that the result resists noise in the settings. The method links to evolution strategies, a classic search approach from computing. It needs only small changes to common training tools and fits many circuit types. Tests on sorting quantum phases show the approach works in practice. The authors frame it as a step toward quantum tools that stay sound on real, noisy hardware.

Catalog record

CREATORLucas Tecot, Di Luo, Cho-Jui Hsieh
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE41 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-18478

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Provably Robust Training of Quantum Circuit Classifiers Against Parameter Noise Lucas Tecot, Di Luo, Cho-Jui Hsieh Abstract Advancements in quantum computing have spurred significant interest in harnessing its potential for speedups over classical systems. However, noise remains a major obstacle to achieving reliable quantum algorithms. In this work, we present a provably noise-resilient training theory and algorithm to enhance the robustness of parameterized quantum circuit classifiers. Our method, with a natural connection to Evolutionary Strategies, guarantees resilience to parameter noise with minimal adjustments to commonly used optimization algorithms. Our approach is function-agnostic and adaptable to various quantum circuits, successfully demonstrated in quantum phase classification tasks. By developing provably guaranteed optimization theory with quantum circuits, our work opens new avenues for practical, robust applications of near-term quantum computers. 1 Introduction In the past few decades, the field of quantum computing has grown dramatically [ 15 , 18 , 21 , 19 , 22 ] . Enticed by the potential of speed-ups over classical computers, great efforts have been devoted to not only building quantum computers [ 6 ] , but also exploring how to achieve these speed-ups when the proper devices exists [ 4 ] with a wide variety of approaches, from carefully-crafted algorithms [ 14 ] to data-learned optimized approaches [ 11 ] . However, all existing quantum algorithms suffer from the existence of noise on the current quantum hardware. Until the theory and practice of error-corrected quantum computing makes significant progress, all quantum algorithms need to find ways to be robust to the noise that exists in noisy intermediate-scale quantum (NISQ) computers [ 23 ] . Furthermore, there are many types of noise, from noise inherent to quantum systems that can’t be fully isolated from their environment, to simple instrumentation noise that occurs whenever preparing a continuously-parameterized operation [ 24 , 8 ] . As a result, a number of works have explored different ways to fight noise in the NISQ era [ 32 , 30 , 31 ] . In parallel, significant effort has been made to improve the noise robustness of machine learning models. One strong method to achieve this is randomized smoothing, which certifies a "smoothed" classifier’s robustness to input perturbations by analyzing the distribution of outputs under noise [ 12 , 35 , 20 , 28 , 26 , 25 ] . In this work, we connect tools from classical machine learning to the quantum setting and offer another tool to fight against noise for algorithms deployed on NISQ devices. Specifically we tackle parameter gate noise, in which the prepared parameters of a continuously-parameterized quantum gate may be different than it is intended to be. Our method is a certified robustness method, in which we can guarantee a certain level of noise will not affect the end result of the quantum process. We highlight our contributions as follows: 1. Our work provides a provably guaranteed framework and theory for training parameterized quantum circuit classifier under parameter noise. While most other methods consider adversarial attacks on inputs or mid-circuit noise inherent to quantum systems, we explicitly tackle the use case that accounts for instrumentation error in devices. As long as the desired noise-model and algorithm can be expressed as a parameterized circuit, our method can be used to find an optimal robustness certificate for changes in those parameters. 2. Our method is simple, easy to deploy, and naturally connected to Evolutionary Strategies (ES), an optimization algorithm already commonly used by the variational quantum algorithm (VQA) community [ 16 , 3 ] . Due to its ability to avoid computing costly quantum gradients [ 2 ] , a practitioner can achieve robustness certificates using our method with minimal effort. Furthermore, our method can be applied on top of any error-mitigation method, allowing to be easily combined with other methods to further boost the robustness they provide. 3. Our approach is successfully demonstrated on quantum phase classification tasks. Our results present clear robustness-variance trade-off and a robustness-variance correlation, which provide insights and understanding on the sensitivity of the parameters in quantum parameterized circuits. In summary, we develop a provably noise-resilient training theory and algorithm for parameterized quantum circuits. With its flexibility and ease of adaptation, our approach achieves noise robustness, paving the way for practical applications on near-term quantum computers. 2 Method 2.1 Noise-resilient Theorem Definition 2.1 (PQC Classifier) A classifier C is called a parameterized quantum circuit (PQC) classifier if it is constructed in the following way: \displaystyle C(\theta,x)_{i}=\bra{\phi_{0}}V^{\dagger}(x)U^{\dagger}(\theta)A_{i}U(\theta)V(x)\ket{\phi_{0}} (1) \displaystyle U(\theta)=U_{L}(\theta_{L})\cdots U_{2}(\theta_{2})U_{1}(\theta_{1}) (2) \displaystyle U_{l}(\theta_{l})=\prod_{m}e^{-i\theta_{m}H_{m}}W_{m} (3) where C(\theta,x)_{i} is the quantum classifier’s probability assigned to class i , A_{i} are easily measurable observables, V(x) is a data-dependent unitary, W_{m} is an unparameterized unitary, H_{m} is a Hermitian operator, and \theta_{l} is the l -th element of \theta . PQC classifiers have been considered in various quantum machine learning setups [ 11 , 27 , 30 ] . While we follow the above particular form of PQC in this work, our discussion is general as long as the classifier comes from a quantum circuit with unitaries depending on x and \theta . In this work, we focus on noise effect on the given PQC classifier parameters. If the PQC classifier parameters are fully robust to noise, any possible noisy perturbation on the parameters should not change the correct classification of the task. More precisely, we formalize this notion as follows. Definition 2.2 (Noise-resilient PQC Classifier) A PQC classifier is parameter noise-resilient if the following is true \displaystyle\forall(x,y)\in D, (4) \displaystyle\left(\argmax_{i}[C(\theta,x)_{i}]=y\right)\implies (5) \displaystyle\qquad\left(\argmax_{i}[C(\theta+\delta,x)_{i}]=y\right) (6) for all perturbations \delta sampled from a given domain. Our goal is to develop robust training theory and algorithms for PQC classifier so that it is provably parameter noise-resilient. To tackle this problem, we integrate approach from classical machine learning with parameterized quantum circuits. More specifically, we develop randomized smoothing certified robustness theory [ 28 ] under the setting of PQC classifier. To start with, we introduce the concept of smoothed PQC classifier. Definition 2.3 (Smoothed PQC Classifier) Let C be a PQC classifier with possible prediction classes \gamma=\{1,\dots,N\} . A smoothed PQC classifier G_{\sigma} is defined as: \displaystyle G_{\sigma}(\theta,x)=\argmax_{z\in\gamma}\mathbb{P}\left(\argmax_{i\in\gamma}\left[C(\theta+\epsilon,x)_{i}\right]=z\right), (7) where \epsilon\sim\mathcal{N}(0,\Sigma) and \Sigma is a diagonal matrix with vector \sigma^{2} as the diagonal. Theorem 2.1 (Noise-resilient Condition) Let C be a PQC classifier and G_{\sigma} to be the corresponding smoothed PQC classifier. If G_{\sigma}(\theta,x)=c_{a} , then G_{\sigma}(\theta+\delta,x)=c_{a} for any \delta vectors that satisfy \displaystyle\|\delta\oslash\sigma\|_{2} \displaystyle<\frac{1}{2}\big(\Phi^{-1}(p_{A})-\Phi^{-1}(p_{B})\big) (8) where \oslash is the Hadamard (element-wise) division, \|\cdot\|_{2} is the L_{2} norm, \Phi^{-1} is the inverse of the standard Gaussian CDF, and \displaystyle p_{A} \displaystyle=\mathbb{P}\left(\argmax_{i\in\gamma}[C(\theta+\epsilon,x)_{i}]=c_{a}\right) (9) \displaystyle p_{B} \displaystyle=\max_{c\neq c_{a}}\mathbb{P}\left(\argmax_{i\in\gamma}[C(\theta+\epsilon,x)_{i}]=c\right) (10) with \epsilon\sim\mathcal{N}(0,\Sigma) and \Sigma is a diagonal matrix with vector \sigma^{2} as the diagonal. Using the above theorem, by smoothing a given PQC (i.e. re-evaluating the model multiple times, with \epsilon sampled from a multivariate Gaussian each time), we can guarantee with high probability that any noise \delta caused by the environment will not change the prediction results, as long as it satisfies the given bound. All we need to do is estimate bounds on p_{A} and p_{B} using concentration inequalities and then we can directly apply the theorem (See Section B.3 for more details). Since standard PQC classifiers require multiple evaluations due to the

Read the complete text