Limitations of Quantum Advantage in Unsupervised Machine Learning

Apoorva D. Patel · 2025

Read Limitations of Quantum Advantage in Unsupervised Machine Learning by Apoorva D. Patel (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-2511-10709.txt.

Introduction

Apoorva D. Patel examines when quantum machine learning can truly beat classical methods at unsupervised learning, the task of finding the probability distribution that best describes data and using it to predict observables. Classical models usually fit Boltzmann distributions of Hamiltonians with many tunable parameters, while quantum versions replace those distributions with density matrices.

Patel argues an advantage appears only when a method uses features of density matrices that classical probability laws simply do not have. Whether such features matter depends on both the input data and the observables being targeted. Through explicit examples, the paper brings out the constraints that limit possible quantum advantage, with effects for data analysis and for sensing uses.

Catalog record

CREATORApoorva D. Patel
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE17 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-10709

Browse the full library · About the Tower of Babel

Full text

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

Limitations of Quantum Advantage in Unsupervised Machine Learning Apoorva D. Patel Abstract Machine learning models are used for pattern recognition analysis of big data, without direct human intervention. The task of unsupervised learning is to find the probability distribution that would best describe the available data, and then use it to make predictions for observables of interest. Classical models generally fit the data to Boltzmann distribution of Hamiltonians with a large number of tunable parameters. Quantum extensions of these models replace classical probability distributions with quantum density matrices. An advantage can be obtained only when features of density matrices that are absent in classical probability distributions are exploited. Such situations depend on the input data as well as the targeted observables. Explicit examples are discussed that bring out the constraints limiting possible quantum advantage. The problem-dependent extent of quantum advantage has implications for both data analysis and sensing applications. I BACKGROUND Over the last decade, the subject of machine learning has grown explosively in many directions. One reason for that is the availability of big data in many fields. Various types of sensors and detectors have become affordable and convenient. They collect huge amount of data (with space and time labels), which need to be analysed to make decisions. Often there is no time and space to store the data; interesting features must be extracted quickly and the remaining data discarded. The challenge is then to carry out these tasks efficiently. Examples of such situations span astronomy, satellite imagery, surveillance, weather and climate, internet traffic, model system tests and validation (for design purposes), experimental observations, collider physics and genetic information. The analysis tackling these problems typically undergoes three steps: recognise interesting patterns, classify them and then make appropriate recommendations. The tedious job of learning from the big data is given to a computer, together with some rules, so that the results can be extracted without much human intervention. The popular learning algorithms based on a large number of tunable parameters include [ 1 , 2 , 3 ] : \bullet Supervised learning, where training data with known pattern labels is provided to the computer program. \bullet Unsupervised learning, where similar examples and analogies are provided to the computer program. \bullet Reinforcement learning, where reward functions evaluating the extracted results are provided to the computer program. These algorithms generally use neural network methods, heuristically and without proof, and obtain probabilistic results. Moreover, the analysis uses fast linear algebra techniques. When the number of tunable parameters is small, say tens or hundreds, the results do not offer much predictive power. But when the number of tunable parameters becomes large, going up to millions and billions, there is a vast improvement in predictive power. Affordable availability of computational resources that can handle such large number of tunable parameters, with multiple GPU systems, is the other reason for the rapid advances in this field. Almost parallel developments in the field of quantum computation has added another aspect to the adventure; may be one can achieve more by a suitable combination of classical and quantum strategies [ 4 ] . The aim of this presentation is to quantify the extra advantage that can be obtained by combination of the two strategies, in the specific case of unsupervised learning. II UNSUPERVISED LEARNING The task of unsupervised learning is to ascertain the probability distribution of the data generated in a specific scenario, which can then be used to make predictions. Initially one guesses the probability distribution based on analogies, and then systematically evolves it to improve the approximation. This strategy needs two ingredients: a criterion to measure the difference between two probability distributions, and an algorithm to reduce the difference by varying parameters characterising the guessed distribution. A good criterion to measure the difference between two probability distributions is the Kullback-Leibler (KL) divergence, i.e. the relative entropy of the two distributions: D_{KL}(p\|q)=\sum_{x}p(x)\log(p(x)/q(x))~, (1) where x denotes the data labels. It is the amount of information lost when the distribution p(x) is used to approximate the distribution q(x) . It is asymmetric in p(x) and q(x) , but it is always non-negative and vanishes only when p(x) and q(x) coincide [ 5 ] . So a parametrised p(x) is fit to the data q(x) , by minimising D_{KL}(p\|q) . When D_{KL}(p\|q) is expanded around its minimum, which is zero, the leading quadratic form gives the Fisher information metric, whose eigendirections obey the corresponding Cramér-Rao bounds for the estimation of parameters. II-A Boltzmann Machines Of course, one cannot make much progress without a practical parametrisation of p(x) . The crucial assumption that is made is that the target probability distribution q(x) arose from some physical process, and so it can be reasonably approximated by the thermal Boltzmann distribution, p(x)=e^{-\beta H(x)}/(\sum_{states~x}e^{-\beta H(x)})~, (2) in terms of a local Hamiltonian H(x) containing many variational parameters. A neural network parametrisation of these distributions is called a Boltzmann machine, with the state variables x represented by network nodes. In this parametrisation, when some degrees of freedom are not observed, they get summed over. Such a sum produces reduced probability distributions; the unobserved degrees of freedom are mapped to hidden variables of a neural network, while the rest remain as the observable (or visible) degrees of freedom. The favoured realisations of Boltzmann machines are stochastic Ising spin-glass models with an external field. They are versatile enough to describe a variety of physical behaviour (e.g. regular, periodic, glassy, chaotic), as well as are easy to simulate. They are defined by the Hamiltonian H=-\sum_{ij}w_{ij}s_{i}s_{j}-\sum_{i}b_{i}s_{i}~, (3) where w_{ij} is the interaction strength between the spins s_{i} and s_{j} , and b_{i} is the external bias field. In simulations starting from a high temperature, the spin network evolves from a random configuration towards its thermal equilibrium state (i.e. simulated annealing). The Hamiltonian defined above has too many parameters, so based on observed interaction structure between neurons, it is simplified to construct restricted Boltzmann machines where interactions exist only between the observable and the hidden variables but not among the observable ones or among the hidden ones. Such a formulation is further extended to define deep Boltzmann machines with multiple layers of hidden variables, where interactions exist between successive layers but not within any particular layer. Fig. 1 illustrates such a network. Fig. 1: An Example of a deep Boltzmann machine with one visible and two hidden layers. The lines denote the interactions between nodes. III QUANTUM GENERALISATION In the quantum setting, the classical probability distribution generalises to the quantum density matrix \rho(x) , which provides the complete description of any quantum state. The expectation values of observables undergo the replacement: \langle O\rangle=\sum_{x}p(x)O(x)\longrightarrow Tr[\rho(x)O(x)] . Since \rho(x) is positive semi-definite, it can always be expressed as e^{-\beta H(x)} in terms of a suitable quantum Hamiltonian. With this change, the reduced probability distributions generalise to reduced density matrices obtained by partial trace over hidden variables. The relative entropy of quantum distributions generalises to S(\rho\|\sigma)=Tr[\rho(x)(\log\rho(x)-\log\sigma(x))]~, (4) which remains non-negative and vanishes only when \rho(x) and \sigma(x) coincide [ 5 ] . Also, the relation between S(\rho\|\sigma) and the quantum Fisher information metric (and the Cramér-Rao bounds) has the same structure as in the classical case. The novel quantum feature is that

Read the complete text