Towards Quantum Operator-Valued Kernels
Hachem Kadri, Joachim Tomasi, Yuka Hashimoto, Sandrine Anthoine · 2025
Read Towards Quantum Operator-Valued Kernels by Hachem Kadri, Joachim Tomasi, Yuka Hashimoto, Sandrine Anthoine (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-2506-03779.txt.
Introduction
Hachem Kadri and co-authors argue that quantum machine learning should move past simple number-valued kernels. Their 2025 position paper recalls the early excitement around quantum kernels, functions that score similarity using quantum states. Later studies cooled that mood, suggesting little advantage when learning from ordinary classical data.
The authors say the disappointment is partly a framing error. Scalar kernels are too rigid to use what is distinctly quantum, such as entanglement, and the tasks tested were ones classical methods already handle well. They propose operator-valued kernels, which return structured outputs and can express richer dependencies. A small proof of concept shows such kernels exposing structure that scalar versions miss. The paper lays out a roadmap toward harder, structured prediction problems. It asks the field to aim quantum tools where classical ones actually struggle.
Catalog record
| CREATOR | Hachem Kadri, Joachim Tomasi, Yuka Hashimoto, Sandrine Anthoine |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 48 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-03779
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Towards Quantum Operator-Valued Kernels Hachem Kadri, Joachim Tomasi, Yuka Hashimoto, Sandrine Anthoine Abstract Quantum kernels are reproducing kernel functions built using quantum-mechanical principles and are studied with the aim of outperforming their classical counterparts. The enthusiasm for quantum kernel machines has been tempered by recent studies that have suggested that quantum kernels could not offer speed-ups when learning on classical data. However, most of the research in this area has been devoted to scalar-valued kernels in standard classification or regression settings for which classical kernel methods are efficient and effective, leaving very little room for improvement with quantum kernels. This position paper argues that quantum kernel research should focus on more expressive kernel classes. We build upon recent advances in operator-valued kernels, and propose guidelines for investigating quantum kernels. This should help to design a new generation of quantum kernel machines and fully explore their potentials. 1 Introduction Quantum machine learning (QML) is an emerging field of research at the intersection between machine learning and quantum computing with the goal of using quantum computing paradigms and technologies to improve the speed and performance of learning algorithms (see Figure 1 ) [ 12 , 52 ] . Since the seminal works by Havlíček et al. 2019 and Schuld and Killoran 2019 , quantum kernel machines have generated a lot of enthusiasm in this field, especially for exploring the applications of noisy intermediate-scale quantum computers to machine learning [ 67 , 13 , 89 , 44 ] . This enthusiasm has been recently tempered by recent studies that have suggested that quantum kernels could not offer speed-ups when learning on classical data [ 47 , 59 , 53 ] . However, these works have dealt with standard classification and regression tasks for which classical kernel methods are efficient and effective, leaving very little room for improvement with quantum kernels. We strongly believe that we have to focus on more complicated machine learning tasks where classical kernel methods face clear limitations in order to reveal the full potential of quantum kernels. This is the lens through which we will explore quantum kernel machines, relying on recent advances in classical machine learning. We present here our point of view on this ongoing field of research and suggest that quantum kernel research should shift its attention from scalar-valued to operator-valued kernels . Operator-valued kernels generalize standard kernel functions and offer the possibility of tackling various machine learning problems ranging from multitask learning to multiview learning and differential equations modeling [ 68 , 4 , 29 , 57 , 63 , 50 , 82 ] . In this paper, we explore the potential of operator-valued kernels in the context of QML. The operator-valued kernel framework is flexible and gives rise to more expressive quantum kernel feature spaces. We also discuss C^{*} -algebra-valued kernels as a generalization of operator-valued kernels. This could open up a new way to apply mathematical theories to QML for more complicated tasks. Then we consider the application of quantum operator-valued kernels for structured prediction. Structured output learning is the task of learning a mapping between objects of different nature that each can be characterized by complex data structures such as curves, trees and graphs [ 88 , 34 , 56 , 16 ] . Quantum structured prediction should extend the application scope of quantum kernels. Quantum Computing Machine learning QML QSVKs QOVKs QOVKs \supset QSVKs Kernels Methods Quantum Circuits, States Beyond classical expressivity? OVKs enable rich quantum representations Figure 1: Quantum machine learning (QML) is a recent field of research at the intersection between quantum computing (QC) and machine learning (ML). The interaction is two-sided: quantum-enhanced ML (from QC to ML), and ML-based quantum computing (from ML to QC). Most of the interest has concentrated on the use of quantum computing paradigms to improve machine learning algorithms. Quantum scalar-valued kernels (QSVKs) have generated a great deal of interest in the field of QML due to their natural alignment with the kernel trick and their compatibility with hybrid quantum-classical architectures. However, recent findings suggest that their expressive power may be limited in classical data regimes. Quantum operator-valued kernels (QOVKs) offer a more general and expressive framework, potentially unlocking richer hypothesis spaces that are inaccessible to classical or quantum SVKs. QOVKs generalize QSVKs and provide opportunities to go beyond classical expressivity. To support our position, we provide a quantum implementation of operator-valued kernels and present an experimental result on quantum channel estimation demonstrating their superiority over scalar-valued kernels. Notation The so called ‘bra-ket’ notation is used to describe the state of a quantum system. A column vector \psi is represented as ‘ket’ \ket{\psi} . The conjugate transpose (Hermitian transpose) of a ket, a row vector, is denoted by ‘bra’ \bra{\psi}:=\ket{\psi}^{\dagger} , where \dagger denotes the conjugate transpose. The inner product of two vectors \ket{\psi_{1}} and \ket{\psi_{2}} can be written in bra-ket notation as \braket{\psi_{1}|\psi_{2}} , and their tensor product, can be expressed as \ket{\psi_{1}}\ket{\psi_{2}} , i.e., \ket{\psi_{1}}\otimes\ket{\psi_{2}} . 2 Kernels from classical to quantum Machine learning is the branch of artificial intelligence that seeks to develop computer systems which detect patterns in data in order to improve their performance automatically with experience [ 54 ] . The wide-spreading development of acquisition tools together with increasing storage capacities has had us witness an explosion in the amount of available data, as well as the urge to develop methods to handle them properly. In this context, it is crucial to design new large-scale machine learning systems that are able to deal with big data. Quantum machine learning is a relatively recent field of research [ 12 , 24 , 27 ] . This research field is largely driven by the desire to develop artificial intelligence that uses quantum technologies to improve the speed and performance of learning algorithms. There is also interest in investigating the use of machine learning for tackling quantum computing and quantum information problems. The field is evolving rapidly, but many open questions remain and should be addressed to better understand how quantum computers may outperform classical computers on machine learning tasks. This paper identifies potential effective interactions between the fields of quantum computing and kernel machines and lays the ground towards a deeper understanding of what kernel-based learning algorithms look like in a quantum world. Note that we focus here on quantum machine learning with classical data (or classical representation of data) but the ideas presented in this paper could be extended to quantum data. 2.1 Classical kernel methods The research work described in this paper belongs to a large class of learning algorithms, the so-called kernel methods. Since they were introduced by Boser et al. 1992 as a way to construct a nonlinear extension of Support Vector Machines, these methods become very popular. Kernel methods exploit training data through implicit definition of a similarity between data points that can be expressed as a dot product in a feature space, namely reproducing kernel Hilbert space (RKHS). The idea is to transform the data via a feature map \phi into a suitable feature space in which linear learning algorithms could be applied. The inner product between features can be computed using the kernel function; this is the well-known kernel trick, i.e., k(x,y)=\langle\phi(x),\phi(y)\rangle (see