Generalization Bounds for Quantum Learning via Rényi Divergences

Naqueeb Ahmad Warsi, Ayanava Dasgupta, Masahito Hayashi · 2025

Read Generalization Bounds for Quantum Learning via Rényi Divergences by Naqueeb Ahmad Warsi, Ayanava Dasgupta, Masahito Hayashi (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-2505-11025.txt.

Introduction

Naqueeb Ahmad Warsi, Ayanava Dasgupta, and Masahito Hayashi derive new generalization bounds for quantum learning algorithms, in a paper from twenty twenty-five. Generalization error tells how well a learner performs beyond its training data. The authors build on a recent frame and a new meaning for expected true loss. Their bounds are stated through Rényi divergences.

A variational approach works out the quantum Rényi divergences. Next to the standard Petz divergence, they add a modified sandwich divergence. In analysis and in numbers, bounds from the modified divergence beat Petz-based ones. The authors also give probabilistic generalization bounds by two routes. One uses the modified sandwich divergence with a classical Rényi divergence. The other uses a smooth max Rényi divergence. The work sharpens the theory of when quantum learners can be trusted.

Catalog record

CREATORNaqueeb Ahmad Warsi, Ayanava Dasgupta, Masahito Hayashi
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE173 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-11025

Browse the full library · About the Tower of Babel

Full text

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

Generalization Bounds for Quantum Learning via Rényi Divergences Naqueeb Ahmad Warsi, Ayanava Dasgupta, Masahito Hayashi Abstract This work advances the theoretical understanding of quantum learning by establishing a new family of upper bounds on the expected generalization error of quantum learning algorithms, leveraging the framework introduced by Caro et al. (2024) and a new definition for the expected true loss. Our primary contribution is the derivation of these bounds in terms of quantum and classical Rényi divergences, utilizing a variational approach for evaluating quantum Rényi divergences, specifically the Petz and a newly introduced modified sandwich quantum Rényi divergence. Analytically and numerically, we demonstrate the superior performance of the bounds derived using the modified sandwich quantum Rényi divergence compared to those based on the Petz divergence. Furthermore, we provide probabilistic generalization error bounds using two distinct techniques: one based on the modified sandwich quantum Rényi divergence and classical Rényi divergence, and another employing smooth max Rényi divergence. 0 0 footnotetext: Indian Statistical Institute, Kolkata 700108, India. Email: [email protected], [email protected] 0 0 footnotetext: School of Data Science, The Chinese University of Hong Kong, Shenzhen, Longgang District, Shenzhen, 518172, China International Quantum Academy, Futian District, Shenzhen 518048, China Graduate School of Mathematics, Nagoya University, Nagoya, 464-8602, Japan. Email: [email protected] I Introduction Quantum learning theory has emerged as a burgeoning field at the intersection of quantum computation and machine learning, promising enhanced capabilities for data analysis and pattern recognition. Recent years have witnessed significant advancements in developing theoretical frameworks and algorithms for quantum learning, as highlighted in works such as [ 1 , 2 , 3 , 4 , 5 , 6 , 7 ] . Among these, the work by Caro et al. [ 8 ] has provided a crucial foundation by introducing a comprehensive quantum learning framework. This framework uniquely addresses the challenges inherent in learning from quantum data, particularly when considering the interplay between classical training data and quantum data states. A key contribution of Caro et al. [ 8 ] lies in its analysis of quantum learning algorithms’ performance through the lens of quantum information theory. By carefully considering the process of training on classical data in conjunction with quantum systems, and the subsequent generation of both classical and quantum hypotheses, Caro et al. [ 8 ] laid the groundwork for a rigorous evaluation of generalization in quantum learning scenarios. Their framework specifically tackles the issue of evaluating the learned hypothesis against a loss operator, a process complicated by the inherent measurements and post-processing that can perturb the quantum data. To overcome this, Caro et al. [ 8 ] introduced the concept of a test-train bipartition of the quantum data, allowing for the analysis of potential correlations and entanglement across these partitions and enabling the derivation of upper bounds on the expected generalization error in terms of both classical and quantum information-theoretic quantities. This prior work thus provides a vital starting point for further investigations into the theoretical limits and practical capabilities of quantum learning algorithms. Building upon these foundational concepts from learning theory, the study of quantum learning aims to understand the capabilities and limitations of learning algorithms that utilize quantum resources. In classical learning theory, a central challenge lies in bridging the gap between the expected empirical loss, directly estimated from training data, and the theoretically critical expected true loss on unseen data. This difference is defined as the generalization error [ 9 , 10 , 11 , 12 , 13 ] , a key figure of merit for assessing how well a learning algorithm’s output generalizes to new, independent data. A substantial body of work in classical learning theory has been dedicated to analyzing this generalization error [ 14 , 15 , 16 , 17 , 18 , 19 , 20 , 21 , 20 , 22 , 23 , 24 , 25 ] . Notably, an information-theoretic approach, as employed in [ 15 , 16 , 17 , 21 , 25 , 20 , 22 , 23 ] , often utilizes the combination of variational lower bounds on divergence [ 26 ] and concentration inequalities like Hoeffding’s Lemma [ 27 , Lemma 2.2 ] (or sub-Gaussianity assumptions [ 27 , Section 2.3 ] for unbounded loss functions). The generalization error, inherently dependent on the stochastic nature of the training data and the learned hypothesis, is typically studied through two primary ways: its expectation and its probabilistic behavior. Investigations into the expected generalization error in classical learning have successfully employed the aforementioned information-theoretic tools. For instance, the references [ 15 , 16 ] established bounds on the expected generalization error based on the mutual information between the training data and the resulting hypothesis. Furthermore, Modak et al. [ 21 ] derived upper bounds on the expected generalization error by leveraging the variational form of Rényi divergence [ 28 ] . Complementarily, a significant amount of research in classical learning theory has also focused on the probabilistic generalization error. Works such as [ 24 , 25 ] have proven upper bounds on the generalization error that hold with a certain probability. Specifically, Esposito et al. [ 25 ] demonstrated that the generalization error is probabilistically upper-bounded by Rényi divergence (up to additive and multiplicative constants) with high probability, achieving this through a change of measure technique rooted in Hölder’s inequality (refer to Fact 2 ). It is worth noting that many existing works in both classical [ 14 , 15 , 16 , 17 , 18 , 19 , 20 , 21 , 25 , 20 , 22 , 23 ] and quantum [ 8 ] learning theory often assume the loss function (or loss observable in the quantum case) to be sub-Gaussian. However, as observed in the classical setting due to Hoeffding’s inequality (Fact 7 ), this assumption is less restrictive for bounded loss functions. In this work, we establish a similar result for quantum learning by first proving a quantum analogue of Hoeffding’s inequality, which subsequently allows us to demonstrate that a bounded loss operator is trivially sub-Gaussian in the quantum context. The main contribution of this paper is to extend classical results on generalization error, specifically those obtained in [ 21 , Theorem 3 ] and [ 25 , Corollary 2 ] , to the quantum setting within the framework of [ 8 ] . We achieve this by proving upper bounds on the expected generalization error for quantum learning algorithms using variational forms involving Rényi-divergence-type quantities. Given that our evaluation is based on measurements, we obtain variational bounds based on measured Rényi divergence. However, these bounds involve optimization over the choice of measurement, making their calculation challenging. To circumvent this, we focus on two specific types of quantum Rényi divergences: Petz quantum Rényi divergence [ 29 ] and sandwich quantum Rényi divergence [ 30 , 31 ] . While sandwich quantum Rényi divergence is generally smaller than Petz quantum Rényi divergence, it only provides an upper bound on measured Rényi divergence for \alpha\geq 1/2 , whereas Petz quantum Rényi divergence is applicable for \alpha\in(0,1)\cup(1,\infty) (See Fact 28 ). To address this limitation, we introduce a modification of the sandwich quantum Rényi divergence using the reverse sandwich quantum Rényi divergence with parameter \alpha<1/2 . By

Read the complete text