[Paper Review] The Optimization Landscape of Hybrid Quantum-Classical Algorithms: from Quantum Control to NISQ Applications

Xiaozhen Ge, Re-Bing Wu+1 more|arXiv (Cornell University)|Jan 19, 2022

Quantum Computing Algorithms and Architecture4 citations

TL;DR

This paper analyzes the optimization landscape of hybrid quantum-classical algorithms, revealing a morphological transition from trap-free landscapes in small quantum systems to barren plateaus in large-scale NISQ devices. It identifies that expressivity and trainability are in tension, with ansatz structure and initialization strategies critical for avoiding vanishing gradients and enabling practical quantum advantage.

ABSTRACT

This review investigates the landscapes of prevalent hybrid quantum-classical optimization algorithms in many rapidly developing quantum technologies, where the objective function is either computed by a natural quantum system or a quantum ansatz that is engineered, but the optimizer is classical. In any particular case, the nature of the underlying control landscape is fundamentally important for systematic optimization of the objective. In early studies on the optimal control of few-body dynamics, the optimizer could take full control of the quantum systems to be manipulated whose Hilbert space dimension is relatively small. Stepping into the noisy intermediate-scale quantum (NISQ) era, the experimentally growing computational power of the ansatz expressed as quantum hardware may bring quantum advantage over classical computers, but the classical optimizer is often limited by the available control resources. Across these different scales, we will show that the landscape's geometry experiences morphological changes from favorable trap-free landscapes to easily trapping rugged landscapes, and eventually to barren-plateau landscapes on which the optimizer can hardly move. This unified view provides the basis for understanding classes of systems that may be readily controlled out to those with special consideration, including the difficulties and potential advantages of NISQ technologies, as well as seeking possible ways to escape traps or plateaus, in particular circumstances.

Motivation & Objective

• To understand the geometric structure of optimization landscapes in hybrid quantum-classical algorithms across different system scales.
• To explain why classical optimization struggles with large-scale NISQ devices due to barren plateaus and rugged landscapes.
• To identify conditions under which optimization remains efficient, particularly in the context of variational quantum algorithms.
• To explore strategies—such as structured ansatzes and intelligent initialization—that mitigate barren plateaus and improve trainability.
• To clarify the trade-off between quantum ansatz expressivity and classical optimizer efficiency in achieving quantum advantage.

Proposed method

• Analyzes the control landscape geometry of quantum systems under varying control resources, from full control in few-body systems to limited control in NISQ devices.
• Applies theoretical frameworks from quantum optimal control and variational quantum algorithms to map the evolution of landscape morphology.
• Investigates gradient scaling behavior in parameterized quantum circuits (PQCs), particularly the variance of partial derivatives as a proxy for barren plateau formation.
• Evaluates structured ansatzes such as quantum convolutional networks, tree tensor networks, and step-controlled circuits to assess their resistance to barren plateaus.
• Proposes initialization schemes that avoid random 2-designs and reduce parameter space dimensionality to preserve gradient information.
• Considers layerwise training and pretraining via classical neural networks to improve convergence and avoid early optimization failure.

Experimental results

Research questions

• RQ1 How does the optimization landscape of hybrid quantum-classical algorithms evolve as the quantum system size increases relative to available classical control resources?
• RQ2 What causes barren plateaus in variational quantum algorithms, and under what conditions can they be avoided or mitigated?
• RQ3 To what extent do specific ansatz structures—such as convolutional or tree-structured circuits—reduce the risk of exponentially vanishing gradients?
• RQ4 Can intelligent initialization strategies prevent the optimizer from getting trapped in suboptimal regions, especially in high-dimensional parameter spaces?
• RQ5 What is the fundamental trade-off between quantum circuit expressivity and classical optimizer trainability in NISQ-era applications?

Key findings

• In small quantum systems with abundant control resources, the optimization landscape is trap-free, enabling efficient convergence via greedy algorithms.
• With limited control resources, false local minima emerge, leading to optimization failure even with stochastic methods.
• Barren plateaus arise in large-scale NISQ devices due to exponentially vanishing gradients, particularly when the ansatz is a 2-design or highly expressive.
• Structured ansatzes such as quantum convolutional neural networks and tree tensor networks exhibit polynomially decaying gradients, avoiding barren plateaus.
• Initialization strategies that avoid 2-designs or reduce parameter space dimensionality significantly improve optimization convergence and reduce training iterations.
• Error mitigation techniques are essential for combating noise-induced barren plateaus, as demonstrated in VQE implementations on Sycamore hardware for hydrogen chain binding energy calculations.