[Paper Review] On the natural gradient for variational quantum eigensolver

Naoki Yamamoto|arXiv (Cornell University)|Sep 11, 2019

Quantum Computing Algorithms and Architecture12 references44 citations

TL;DR

The paper analyzes how natural gradient optimization, based on quantum geometric metrics, improves VQE parameter updates compared to ordinary gradient and ITE in simple case studies, highlighting when it helps or hinders.

ABSTRACT

The variational quantum eigensolver is a hybrid algorithm composed of quantum state driving and classical parameter optimization, for finding the ground state of a given Hamiltonian. The natural gradient method is an optimization method taking into account the geometric structure of the parameter space. Very recently, Stokes et al. developed the general method for employing the natural gradient for the variational quantum eigensolver. This paper gives some simple case-studies of this optimization method, to see in detail how the natural gradient optimizer makes use of the geometric property to change and improve the ordinary gradient method.

Motivation & Objective

• Motivate the use of natural gradient in VQE by accounting for the geometric structure of parameter space.
• Illustrate how the Fubini-Study quantum metric influences optimization trajectories in VQE.
• Compare natural gradient and imaginary-time evolution based updates to ordinary gradient in simple VQE examples.
• Identify scenarios where natural gradient provides faster convergence and where it may face challenges due to singularities.

Proposed method

• Define the natural gradient update with the quantum Fisher information / Fubini-Study metric F and show it as theta_{k+1} = theta_k - eta_k F(theta)^{-1} ∂f/∂theta.
• Relate natural gradient to imaginary-time evolution (ITE) and show A >= F and F <= A in matrix inequalities to discuss their relationships.
• Compute F explicitly for simple one-qubit and two-qubit hardware-efficient ansatzes and analyze singular points where det(F)=0.
• Use two case studies (single qubit with H = sigma_x and H2 molecule with a 2-qubit Hamiltonian) to compare gradient methods and ITE trajectories.
• Discuss practical considerations such as adding a small diagonal to F to handle near-singular points.

Experimental results

Research questions

• RQ1 How does the quantum-geometry-based natural gradient affect optimization trajectories in VQE compared to ordinary gradient methods?
• RQ2 In which VQE settings (states, Hamiltonians, ansatz structures) does the natural gradient provide faster convergence or encounter problematic singularities?
• RQ3 How does the natural gradient relate to imaginary-time evolution in the VQE context, and when are they equivalent or distinct?
• RQ4 What role do singular points (where det(F)=0) play in the effectiveness of natural gradient in VQE?
• RQ5 Can classical Fisher information-based approaches (Fisher metric on probabilities) be effectively used for VQE optimization?

Key findings

• Natural gradient leverages the geometry of the quantum state space to guide parameter updates faster than ordinary gradient in studied cases.
• In the single-qubit example, natural gradient and ITE converge faster to the ground state by traversing metric-informed trajectories around singular points.
• In the H2 two-qubit hardware-efficient ansatz, ITE equals the natural gradient since the state is real and ⟨∂iφ|φ⟩=0, with singular points tied to separable states and entanglement entropy S(|φ⟩=0 when det(F)=0).
• The energy landscape can include singular regions where F becomes ill-conditioned; adding a small regularization to F can prevent large, destabilizing parameter jumps.
• Natural gradient may be less effective when the target state lies near singular points (e.g., separable states in toy cases), suggesting careful metric analysis of the target state before applying it.