Numerical Optimization Strategies for the Variational Hamiltonian Ansatz in Noisy Quantum Environments

Technical Summary: Numerical Optimization Strategies for the Variational Hamiltonian Ansatz in Noisy Quantum Environments

Problem Statement
The paper addresses the critical challenge of selecting appropriate classical optimization algorithms for Variational Quantum Eigensolvers (VQE) operating in the Noisy Intermediate-Scale Quantum (NISQ) era. While variational methods are central to near-term quantum computing, the choice of optimizer is often intuition-based rather than systematic. The authors highlight that finite-shot sampling noise fundamentally distorts the cost landscape, creating false minima, reversing gradient directions, and inducing stochastic drift. These effects are exacerbated in high-dimensional parameter spaces and can lead to violations of the variational principle (where estimated energies appear lower than the true ground state). The study seeks to systematically benchmark optimization strategies to determine which methods remain robust under these noisy conditions, specifically utilizing the truncated Variational Hamiltonian Ansatz (tVHA).

Methodology
The authors conducted a comprehensive comparative study of eight classical optimization algorithms applied to the tVHA framework across four molecular systems: H₂, H₄ (linear chain), LiH (full space), and LiH (active space). The optimizers tested included:

• Gradient-based: BFGS, Gradient Descent (GD), SLSQP.
• Gradient-free/Derivative-free: COBYLA, Nelder-Mead, SPSA.
• Population-based/Metaheuristic: CMA-ES, Particle Swarm Optimization (PSO).

The simulations were performed using a Python-based stack (Qiskit, PySCF) on the Barbora supercomputer. The study evaluated performance under two distinct regimes:

1. Statevector (Noiseless): Exact expectation values to establish baseline convergence.
2. Sampling Noise: Finite-shot simulations (fixed at 6,144 shots per evaluation) to mimic realistic hardware constraints.

The experimental design included 40 independent runs per optimizer-molecule combination, varying between Hartree-Fock (HF) initialization and random parameter initialization. A key methodological component involved "high-shot reevaluation" (using $10^5$ shots) of the best parameters found during noisy optimization to distinguish genuine convergence from statistical artifacts.

Key Contributions and Findings

1. Optimizer Hierarchy Reversal: The study demonstrates a clear inversion in optimizer performance based on the presence of noise.

   • Noiseless: Gradient-based methods (particularly BFGS) converge fastest and achieve machine precision.
   • Noisy: Population-based and stochastic methods (CMA-ES, PSO, SPSA) significantly outperform gradient-based methods. BFGS and SLSQP degrade rapidly or diverge entirely under sampling noise, as their reliance on curvature information becomes unreliable when gradients are obscured by noise. CMA-ES consistently provided the most robust convergence across all system sizes and noise levels.

2. Impact of Initialization: Hartree-Fock initialization provides a significant advantage for small systems (H₂), reducing error variance and mean error. However, this benefit diminishes as system size increases (LiH, H₄). In larger, noisy landscapes, random initialization sometimes outperforms HF initialization for population-based methods, suggesting that broader exploration is more valuable than structured priors in high-dimensional, rugged noise environments.

3. Active Space Truncation: The use of active-space representations (e.g., LiH active space) not only reduces computational cost but also improves noise robustness by an order of magnitude compared to full-space simulations. This simplification allows optimizers that struggle in full space (like COBYLA) to achieve sub-millihartree accuracy.

4. Variational Principle Violations and Noise Floor: The paper confirms that finite-shot sampling frequently results in energy estimates below the true ground state, violating the variational principle. Rather than discarding these as unphysical, the authors argue they indicate a symmetric noise distribution around the optimum. They demonstrate that relying on the "best" individual in a population leads to overfitting noise artifacts. Instead, utilizing the population mean allows for the averaging out of symmetric fluctuations, achieving energy estimation precision beyond the intrinsic sampling limit without the prohibitive cost of increasing shot counts for every evaluation.

5. System Scaling: As molecular complexity increases, the "barren plateau" effect and noise-induced ruggedness become more severe. Deterministic gradient methods lose their functional signal, transitioning optimization from gradient-driven contraction to diffusion dominated by sampling noise.

Significance and Claims
The paper claims to provide a systematic, data-driven guide for optimizer selection in variational quantum chemistry. Its primary contribution is the empirical demonstration that population-based optimizers (specifically CMA-ES) are essential for robust performance in noisy, high-dimensional regimes, whereas gradient-based methods are largely limited to idealized, noiseless, or small-scale scenarios.

Furthermore, the work reframes the "violation" of the variational lower bound not merely as a failure mode but as a statistical feature of the noise floor. By leveraging the steady-state distribution of Evolution Strategies and utilizing population means, the authors show that it is possible to extract high-precision energy estimates despite the presence of sampling noise. The study concludes that for practical NISQ-era applications, the choice of optimizer is as critical as the ansatz design, and that relying on the "best" observed value is a statistical trap; robust estimation requires aggregating population statistics.

The authors maintain a modest stance, noting that their findings isolate finite-shot sampling noise without including coherent gate errors or readout bias, though they assert that sampling noise alone is sufficient to render deterministic methods unusable. Future work is suggested to pair fast gradients with population search and to incorporate realistic hardware error channels.