Wierichs, David ORCID: 0000-0002-0983-7136 (2023). Estimation of gradients and analysis of gradient-based optimizers for variational quantum algorithms. PhD thesis, Universität zu Köln.
PDF
Dissertation_Wierichs_final.pdf - Accepted Version Bereitstellung unter der CC-Lizenz: Creative Commons Attribution. Download (5MB) |
Abstract
On its path towards computational advantage, quantum computing hardware still is at an early prototypical stage, not yet allowing the use of error correction codes and algorithms that are provably more performant than their classical competitors. Even so, it might be the case that the current noisy quantum devices can be used for relevant computations that are out of reach for current classical computers, if only for a few specific applications, and without any performance guarantees. The question is thus whether the quantum devices in the near future can be promoted to more than a mere bridge technology, which gave rise to the field of research on noisy intermediate-scale quantum (NISQ) algorithms. A substantial part of these research efforts, and also the main topic of this thesis, concerns variational quantum algorithms (VQAs), which aim at using NISQ devices in a hybrid setup together with classical computers. In such a setup, a computational problem is encoded in terms of an observable, typically a Hamiltonian, such that determining the minimal energy (or the corresponding ground state) would yield the solution. With the purpose of generating candidates for the ground state, a classical computer selects a quantum circuit from a family of circuits and the quantum device executes the circuit to prepare the corresponding state. The family of circuits is typically defined in form of a parametrized quantum circuit (PQC) and a particular circuit can be chosen by fixing its parameters. In return the classical computer receives measurement outcomes of selected observables on the candidate states. The algorithm proceeds by optimizing over the parameter space in order to find for (a useful approximation of) the target state. There are many variants and proposed use cases for VQAs which has led to a modular structure of the algorithms. In this thesis, the first chapter focuses on derivative estimators for objective functions that are based on PQCs, which is a subroutine commonly used in the optimization within VQAs. It begins with a review of (componentwise) differentiation methods for PQC-based objective functions, followed by a detailed comparison of these methods using the example of a ubiquitous class of PQCs. This comparison confirms and complements recent results on the topic and has implications for gradient estimation in practice. The second chapter continues on the theme of gradient estimation for optimization, and covers so-called parameter-shift rules, a particular class of derivative estimators. It starts with a brief review of the literature, followed by an extension of said estimators to a specific class of gates for quantum chemistry calculations, which was published as part of a larger manuscript. The main part of the chapter is a publication about the generalization of the parameter-shift rule to a larger class of quantum gates.It presents derivative estimators for various gates, some of which have been shown in the literature to be optimal, together with a thorough cost analysis for both classical simulators and quantum hardware. The third and final chapter gives a short outline of VQAs and provides the context for a second publication, which analyses different algorithms for the optimization task in VQAs. It compares the quantum natural gradient optimizer (QNG) to two established gradient-based methods from classical non-convex optimization. This is done with numerical experiments, in which QNG shows favourable convergence properties and enhanced robustness against symmetry-breaking PQCs for highly structured problems. This work uses the variational quantum eigensolver (VQE), a popular example of a VQA, on spin chain Hamiltonians as its benchmark problem and the experiments are based on extensive classical simulations. A promising direction for future work is to investigate individual building blocks of the large VQA construct separately and to develop metrics for these blocks that allow to predict their usefulness in practice. This may reduce the complexity of single research efforts and lead to insights that are as modular as VQAs themselves.
Item Type: | Thesis (PhD thesis) | ||||||||
Translated title: |
|
||||||||
Translated abstract: |
|
||||||||
Creators: |
|
||||||||
URN: | urn:nbn:de:hbz:38-714998 | ||||||||
Date: | 2023 | ||||||||
Language: | English | ||||||||
Faculty: | Faculty of Mathematics and Natural Sciences | ||||||||
Divisions: | Faculty of Mathematics and Natural Sciences > Department of Physics > Institute for Theoretical Physics | ||||||||
Subjects: | Physics | ||||||||
Uncontrolled Keywords: |
|
||||||||
Date of oral exam: | 11 January 2023 | ||||||||
Referee: |
|
||||||||
Funders: | ML4Q Cluster of Excellence | ||||||||
Refereed: | Yes | ||||||||
URI: | http://kups.ub.uni-koeln.de/id/eprint/71499 |
Downloads
Downloads per month over past year
Export
Actions (login required)
View Item |