The Effect of Noise on the Performance of Variational Algorithms for Quantum Chemistry

Waheeda Saib, Petros Wallden, Ismail AkhalwayaQCE 2021 · DOI

TL;DR

This paper investigates how noise affects the performance of hardware-efficient ansatze within Variational Quantum Eigensolver (VQE) algorithms for quantum chemistry. They find that noise significantly alters the ranking of optimal ansatze, making ideal-world performance a poor predictor of real-world utility, and that the "expressibility" metric is not a reliable indicator of VQE performance.

What Problem Does This Paper Address?

Quantum computers hold immense promise for simulating molecular systems, a task crucial for drug discovery and materials science. Variational Quantum Algorithms (VQAs), like the Variational Quantum Eigensolver (VQE), are a leading candidate for near-term quantum computers due to their hybrid quantum-classical nature. A key component of VQE is the "ansatz," a parameterized quantum circuit that explores the solution space. Hardware-efficient ansatze are popular because they are designed to be easily implemented on current noisy quantum hardware. However, the performance of these ansatze can be heavily influenced by noise, and it's unclear which ansatz performs best under realistic noisy conditions. This paper aims to bridge this gap by systematically evaluating how different noise models impact the performance and ranking of various hardware-efficient ansatze for a quantum chemistry problem.

Methods

The researchers employed the Variational Quantum Eigensolver (VQE) algorithm to approximate the ground state energy of the hydrogen molecule (H₂). They focused on "hardware-efficient ansatze," which are circuits composed of parameterized single-qubit gates and two-qubit entangling gates (CX and CZ) tailored to the specific hardware.

Key methodological aspects include:

  • Ansatz Selection: Twelve different hardware-efficient ansatze were designed with a circuit depth of one and a width of four qubits, varying the types and arrangements of CX and CZ gates.
  • Noise Simulation: Simulations were performed using both ideal (noise-free) and noisy quantum conditions. The noisy simulations utilized a specific noise model derived from the IBM Quantum Experience (specifically, the IBMQX2 device noise model), incorporating gate errors, measurement errors, and decoherence.
  • Performance Evaluation:
    • VQE Performance: The primary metric was the energy difference between the simulated ground state energy and the known true ground state energy of H₂. Ansätze were ranked based on this energy difference.
  • Expressibility: The "expressibility" of each ansatz was calculated. Expressibility quantifies how well a parameterized quantum circuit can represent the Hilbert space of quantum states. It was measured by comparing the distribution of quantum state fidelities produced by the ansatz to a uniform distribution of Haar random states.
  • Correlation Analysis: The paper investigated the correlation between an ansatz's expressibility and its VQE performance using Pearson correlation coefficients and scatter plots.
  • Noise Model Impact: The effect of different IBM Quantum device noise models on the ranking of ansatze was analyzed.

Datasets

Not discussed in the paper. The study uses a simulated hydrogen molecule (H₂) as the benchmark problem, for which the ground state energy is known.

Key Insights and Ideas

  • Noise is a Game Changer for Ansatz Ranking: The most striking finding is that the ranking of optimal ansatze changes dramatically when moving from ideal (noise-free) to noisy quantum simulations. An ansatz that performs best in ideal conditions may perform poorly in a noisy environment, and vice-versa. This highlights the critical need to evaluate ansatze under realistic noise conditions.
  • Expressibility is a Poor Predictor of VQE Performance: The "expressibility" metric, intended to gauge a circuit's ability to represent quantum states, showed a weak (and sometimes negative) correlation with VQE performance for both ideal and noisy simulations. This suggests that high expressibility does not necessarily translate to better accuracy in quantum chemistry VQE tasks.
  • Non-uniform Sampling Can Boost Expressibility: The study demonstrated that expressibility can be significantly improved by using non-uniform parameter sampling during its calculation, rather than uniform sampling. This further weakens expressibility as a standalone metric, as it can be manipulated.
  • Hardware-Specific Noise Matters: The ranking of ansatze varied not only between ideal and noisy conditions but also across different noise models of IBM quantum devices. This implies that the choice of ansatz is not universal and must be tailored to the specific quantum hardware and its associated noise characteristics.
  • CX vs. CZ Gates: The paper observed that for noisy simulations, circuits with mixed CX and CZ gates often outperformed those with only CX gates. This is attributed to CX gates not being "native" on many IBM Quantum devices, leading to higher effective noise when implemented through sequences of native gates.

Key Results

  • VQE Performance: In ideal simulations, circuits 9 and 3 showed the smallest energy differences (0.0161 and 0.0167 hartree, respectively). In noisy simulations, circuits 2 and 4 performed best with energy differences of 0.2089 and 0.2109 hartree. Circuit 10 was consistently among the worst performers in both ideal and noisy scenarios.
  • Expressibility Ranking: In ideal simulations, circuits 12 and 11 were the most expressible (0.020 and 0.027), while circuits 10 and 9 were the least expressible (0.691 and 0.648). In noisy simulations, this order shifted significantly, with circuits 12 and 11 becoming much less expressible (0.946 and 1.459), and circuits 2 and 1 becoming the most expressible.
  • Correlation: The Pearson correlation coefficient between expressibility and VQE energy difference was -0.195 for ideal simulations and a mere 0.012 for noisy simulations, indicating a very weak or negligible linear relationship.
  • Noise Model Impact: Figure 5 visually demonstrates that the ranking of ansatze varies considerably across different IBM Quantum noise models, with some models leading to drastically different optimal circuit choices.

Limitations and Future Work

  • Limited Circuit Depth and Width: The study focused on ansatze with a circuit depth of one and a width of four qubits. Future work could explore the impact of noise on deeper and wider circuits.
  • Single Chemistry Problem: The evaluation was performed on a single, relatively simple chemistry problem (H₂). Extending the analysis to more complex molecules and chemical systems would be valuable.
  • Specific Noise Model: While using a real IBM device noise model is a strength, future work could explore a broader range of noise models and their effects.
  • Expressibility Metric: The authors suggest that alternative quantities and numerical methods should be considered for evaluating ansatz suitability, as expressibility proved to be an inadequate measure.
  • Machine Learning for Ansatz Selection: The authors propose using machine learning techniques to train classifiers that can predict suitable ansatze based on problem characteristics and hardware noise models.
  • Broader Ansatz Types: Future research could include a more extensive analysis of different ansatz types, beyond hardware-efficient ansatze, and explore various gate sets.

Summary updated

BibTeX
@inproceedings{saib2021effect,
  author = {Waheeda Saib and Petros Wallden and Ismail Akhalwaya},
  title = {{The Effect of Noise on the Performance of Variational Algorithms for Quantum Chemistry}},
  booktitle = {QCE},
  year = {2021},
  doi = {10.1109/qce52317.2021.00020},
}
Figure