⚛️ quantum physics

Statistical Benchmarking of Optimization Methods for Variational Quantum Eigensolver under Quantum Noise

This study systematically evaluates six optimization algorithms for the State-Averaged Orbital-Optimized Variational Quantum Eigensolver applied to the H2 molecule under various quantum noise conditions, revealing that BFGS offers the best balance of accuracy and robustness while providing practical guidance for noise-aware optimizer selection in current quantum hardware.

Original authors: Silvie Illésová, Tomáš Bezděk, Vojtěch Novák, Bruno Senjean, Martin Beseda

Published 2026-03-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Silvie Illésová, Tomáš Bezděk, Vojtěch Novák, Bruno Senjean, Martin Beseda

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find the lowest point in a vast, foggy mountain range. This lowest point represents the most stable, energy-efficient state of a molecule (in this case, a simple hydrogen molecule, H₂). You are using a new kind of map-making tool called a Quantum Computer to help you navigate.

However, there's a catch: the map is glitchy. The quantum computer is "noisy," meaning the data it gives you is fuzzy, like trying to read a sign through a thick fog or while riding a bumpy bus. This is the reality of current quantum hardware, known as the NISQ era (Noisy Intermediate-Scale Quantum).

To find that lowest point, you need a "guide" (an Optimizer) to tell you which direction to step next. The paper you asked about is essentially a massive road test to see which guide works best when the road is bumpy and the weather is bad.

Here is the breakdown of the study using everyday analogies:

1. The Mission: The "State-Averaged" Hike

Usually, hikers just want to find the bottom of the valley (the ground state). But this study is looking for two valleys at once: the main one and a nearby one (the excited state). They call this SA-OO-VQE. It's like trying to find the two deepest spots in a lake simultaneously while the water is choppy.

2. The Contestants: The Six Guides

The researchers tested six different types of guides (optimization algorithms) to see who could navigate the foggy, glitchy terrain best:

BFGS (The Quasi-Newton): Think of this as a smart GPS. It doesn't just look at the slope right in front of you; it remembers the shape of the road you just drove on to predict the best path forward. It's fast and usually very accurate.
SLSQP (The Strict Planner): This guide tries to solve complex math problems to plan every step perfectly. It works great on smooth, paved highways, but in the foggy quantum mountains, it gets confused and often gives up or takes a wrong turn.
Nelder-Mead & Powell (The Explorers): These guides don't use a map or GPS. They are like blind hikers who poke the ground in different directions to see where it's lower. They are slow and take many steps, but they are stubborn and rarely get stuck.
COBYLA (The Efficient Scout): This guide is like a budget traveler. It doesn't need perfect data; it makes do with rough approximations. It's fast and cheap to run, though it might not find the exact bottom of the valley, just a spot very close to it.
iSOMA (The Swarm): Imagine a flock of birds searching for food. Instead of one hiker, you send out a whole flock. They spread out to cover the whole mountain. This is great for finding the absolute best spot in a huge, complex maze, but for a small mountain like H₂, it's overkill and wastes a lot of time and energy.

3. The Obstacles: The "Noise"

The researchers didn't just test these guides on a clear day. They simulated three types of "bad weather" (Quantum Noise):

Sampling Noise: Like trying to hear a whisper in a crowded room. You have to listen many times to get the message right.
Dephasing & Depolarizing: Imagine the compass on your watch starts spinning randomly or the map starts smudging. The data becomes unreliable.
Thermal Relaxation: Imagine the hiker is getting tired and the ground is melting under their feet. The information they are holding starts to fade away completely.

4. The Results: Who Won the Race?

The Champion: BFGS
The Smart GPS (BFGS) won almost every race. Even when the fog was thick and the road was bumpy, it found the lowest energy point with the fewest steps. It was the most accurate and the most efficient. It proved that even in a noisy world, a guide that "remembers" the path works best.
The Runner-Up: COBYLA
If you don't care about finding the perfect bottom, but just want a "good enough" answer quickly and cheaply, COBYLA is your guy. It's the budget option that gets the job done without needing a supercomputer.
The Loser: SLSQP
The Strict Planner (SLSQP) was a disaster. It failed to find the bottom in almost every noisy scenario. It's like a driver who refuses to drive unless the road is perfectly paved; as soon as there's a pothole, they stop the car.
The Over-Engineers: iSOMA
The Swarm (iSOMA) was too slow. For a small problem like a hydrogen molecule, sending a whole flock of birds was unnecessary. It took way too many steps to get the same result the GPS got in seconds.

5. The Big Takeaway

The study concludes that not all guides are created equal, and the "best" one depends on the conditions:

If you want accuracy and speed on current noisy quantum computers, use BFGS.
If you are low on resources and just need a rough answer, use COBYLA.
Avoid SLSQP for noisy quantum problems; it just doesn't work well there.
Don't use the "Swarm" (Global Optimizers) for small, simple problems; it's like using a sledgehammer to crack a nut.

In short: When your quantum computer is glitchy and noisy, you need a guide that is smart enough to handle the errors without getting confused. The paper proves that the "Smart GPS" (BFGS) is currently the best navigator for the quantum wilderness.

1. Problem Statement

Variational Quantum Algorithms (VQAs), particularly the Variational Quantum Eigensolver (VQE), are the leading approach for near-term quantum computing (NISQ era). However, their performance is heavily dependent on the classical numerical optimizer used to minimize the cost function. In real-world quantum hardware, the cost landscape is distorted by various noise sources (sampling noise, decoherence, thermal relaxation), which can cause optimizers to fail, converge to local minima, or require excessive computational resources.

While previous studies have analyzed specific optimizers, there is a lack of systematic, statistically rigorous benchmarking comparing gradient-based, gradient-free, and global optimization strategies across a comprehensive suite of noise models specifically for State-Averaged Orbital-Optimized VQE (SA-OO-VQE). This method is crucial for calculating both ground and excited states, acting as a quantum analog to classical MCSCF methods.

2. Methodology

System and Algorithm

Target System: The hydrogen molecule ( $H_2$ ) at equilibrium bond length (0.74279 Å).
Method: SA-OO-VQE using a Complete Active Space (CAS) approximation with a cc-pVDZ basis set.
Ansatz: A generalized unitary coupled-cluster ansatz with single and double excitations (3 trainable parameters) applied to two orthogonal initial states (Hartree-Fock and first excited singlet).
Cost Function: The ensemble energy of ground and excited states, minimized over orbital parameters ( $\kappa$ ) and circuit parameters ( $\theta$ ).

Optimization Strategies Tested

Six representative optimizers were evaluated:

Gradient-Based: BFGS (Quasi-Newton), SLSQP (Sequential Least Squares Quadratic Programming).
Gradient-Free (Local): Nelder-Mead (NM), Powell's Method (PM), COBYLA (Constrained Optimization BY Linear Approximations).
Global: iSOMA (Improved Self-Organizing Migrating Algorithm).

Noise Models and Estimators

The study utilized Qiskit Aer to simulate various noise conditions:

Sampling Noise: Simulated by varying the number of measurements ( $n_m \in \{256, 512, 1024, 6144\}$ ).
Decoherence Channels:
- Phase Damping: Applied to $R_z$ gates at rates of 1%, 5%, 10%, 20%.
- Depolarizing Noise: Applied to single and two-qubit gates at rates of 1%, 5%, 10%, 20%.
- Thermal Relaxation: Modeled with realistic ( $T_2 \in \{70, 80, 180, 380\}$ $\mu$ s) and extreme/unrealistic ( $T_2 \in \{50, 100, 200, 300\}$ ns) relaxation times.

Statistical Analysis Framework

To ensure robust conclusions, the authors employed a rigorous statistical pipeline:

PERMANOVA (Permutational Multivariate Analysis of Variance): To test for significant differences in the centroids (location) of optimization results across different noise settings.
PERMDISP: To test for heterogeneity in multivariate dispersion (spread) among groups.
Mardia's Test: To check for multivariate normality (skewness and kurtosis).
Friedman Test & Wilcoxon Signed-Rank Tests: For non-parametric ranking of optimizers and pairwise comparisons with Holm correction for multiple testing.
Bootstrapping: Used to generate 95% prediction ellipses for visualizing result variability.

3. Key Contributions

Comprehensive Benchmarking: The first systematic comparison of six distinct optimization classes (gradient-based, derivative-free, global) applied to SA-OO-VQE under a wide spectrum of realistic and extreme quantum noise models.
Statistical Rigor: Moving beyond simple mean-error reporting, the study utilizes advanced non-parametric statistical tests (PERMANOVA, PERMDISP, Friedman) to validate that observed performance differences are statistically significant and not artifacts of noise distribution.
Noise-Aware Optimization Guidelines: The paper provides specific, data-driven recommendations for selecting optimizers based on the dominant noise channel and the trade-off between accuracy and computational cost (number of function evaluations).
Identification of Failure Modes: Detailed characterization of how specific noise types (e.g., depolarizing vs. dephasing) uniquely impact different algorithmic families, revealing that SLSQP is fundamentally unsuited for noisy VQE landscapes.

4. Key Results

Performance Under Ideal Conditions

BFGS was the most efficient, converging in ~37 evaluations with high accuracy.
SLSQP also converged quickly but showed instability in noisy regimes.
Global Optimizer (iSOMA) performed poorly, requiring >1000 evaluations and failing to reach the global optimum, indicating it is unnecessary for low-dimensional problems like $H_2$ .

Performance Under Sampling Noise

BFGS remained the top performer, achieving accuracy within $10^{-3}$ of the ideal value with hundreds of evaluations.
COBYLA offered a strong alternative for low-cost applications, requiring fewer evaluations than BFGS but with slightly lower accuracy.
SLSQP failed to converge in all sampling noise scenarios, regardless of measurement count.

Performance Under Decoherence (Phase & Depolarizing)

Dephasing Noise: BFGS, NM, and PM performed well. Interestingly, for Nelder-Mead, higher noise levels paradoxically reduced the number of evaluations required, suggesting noise can sometimes simplify the landscape for specific algorithms.
Depolarizing Noise: This had a more severe impact than dephasing. BFGS maintained the best balance of accuracy and stability. COBYLA and SLSQP suffered significant accuracy degradation.
Thermal Relaxation:
- Under realistic conditions, BFGS was consistently the best. SLSQP was the worst performer.
- Under extreme (unrealistic) relaxation times, all optimizers converged to similar poor energy values (the "noise floor"), rendering the choice of optimizer irrelevant as the quantum information was lost.

Statistical Rankings

Top Tier: BFGS, Nelder-Mead (NM), and Powell (PM) formed a statistically indistinguishable group of high performers.
Mid Tier: COBYLA and iSOMA showed intermediate performance.
Bottom Tier: SLSQP was statistically significantly worse than all other methods across all categories.

5. Significance and Conclusion

This study provides critical practical guidance for researchers and practitioners in the NISQ era:

Optimizer Selection: For small-to-medium molecular systems, BFGS is the recommended default due to its robustness and efficiency. If computational cost (function evaluations) is the primary constraint, COBYLA is a viable alternative. SLSQP should be avoided for noisy VQE applications.
Noise Limitations: The research highlights that while algorithmic improvements matter, there is a "noise floor." Under extreme decoherence, no optimization strategy can recover the correct physics, emphasizing the need for hardware improvements or error mitigation techniques alongside better optimizers.
Future Directions: The authors suggest that for larger, more complex systems with higher dimensionality and potential "barren plateaus," global or population-based methods (like iSOMA or CMA-ES) may become necessary, and future work should focus on noise-resilient optimizers and the integration of error mitigation.

In summary, the paper establishes that noise-aware optimization strategies are essential for reliable quantum chemistry computations, with BFGS emerging as the most robust choice for current hardware limitations.