arXiv:2609.22729
Can Chemically Inspired Parameter Initialization Mitigate Barren Plateaus in Variational Quantum Eigensolvers?
Zhangyu Yang, Jinzhao Sun, Jianpeng Chen, Weitang Li, Zhigang Shuai
Abstract
Recent advances in the variational quantum eigensolver (VQE) have strengthened its promise for molecular electronic-structure calculations on near-term noisy intermediate-scale quantum (NISQ) devices. However, barren plateaus (BPs) remain a major obstacle to scaling VQE, as vanishing gradients can severely obstruct optimization at large scale. While most BP studies focus on randomly initialized circuits and global gradient behavior, practical quantum chemistry VQE typically uses chemically motivated initializations and local optimization around HF- or MP2-derived states. Although chemically informed initializations are widely expected to improve VQE trainability, how they mitigate local barren plateaus remains unclear. We therefore study the occurrence of local barren plateaus in molecular UCCSD-VQE, considering both chemically motivated initialization points and iterates along the corresponding optimization trajectories. In the benchmarks considered here, chemically initialized starting points and variational trajectories show polynomial decay of local gradient variance with system size instead of the exponential BP-like decay observed for random controls, indicating that chemical initialization mitigates barren plateaus in the local optimization regions and highlights the trainability advantage of chemically motivated VQE strategies.
AI Summary
The paper asks whether chemically motivated initialization avoids local barren plateaus in factorized UCCSD-VQE, even when broad random initialization produces exponentially suppressed gradients. It studies Hartree–Fock (HF), MP2, and random parameter centers, plus points along their optimization trajectories, using component-averaged gradient variance over hypercubic parameter patches. A sufficient-condition theorem establishes that an inverse-polynomial Hessian Frobenius norm, polynomially bounded third derivatives, and polynomial parameter count guarantee an inverse-polynomial-radius patch with inverse-polynomial gradient variance. At HF, the diagonal Hessian identity $\mathcal H_{\mu\mu}=2(E_\mu-E_0)$ gives a chemical certificate: one included excitation with $|E_\mu-E_0|\geq1/\operatorname{poly}(q)$ suffices under the regularity assumptions. Noiseless classical simulations cover hydrogen and half-filled Hubbard chains with $q=4,8,12,16,20,24$ spin-orbital qubits and up to 1,818 parameters. Scanning patch radii, the peak local variance and its maximizing radius show power-law-compatible scaling around HF and MP2. Peak-variance exponents range approximately from 2.36 to 3.48 across the benchmarks; hydrogen-chain random controls instead fit exponential decay. Local variance retains power-law-compatible scaling at sampled checkpoints over 25 HF/MP2 optimization iterations, while random-start controls favor exponential scaling over 50 iterations. A derivative-generated tangent-space energy diagnostic provides complementary numerical correlations with local variance. The conclusions concern local, finite-size, noiseless diagnostics, not efficient optimization guarantees. The paper explicitly notes that conditioning, local minima, unfavorable trajectories, and limited ansatz expressivity can still impede convergence. Stronger correlation generally worsens the fitted polynomial suppression, and the tangent-space diagnostic alone cannot certify absence of local barren plateaus.
Error found: The assertion that the normalized excitation gap |E_mu-E_0|/m increases with system size contradicts elementary bounds for the paper's U/t=1 Hubbard benchmarks, indicating a normalization, labeling, or reporting error. The claim to examine full optimization trajectories is unsupported by diagnostics restricted to sampled checkpoints in runs capped at 25 HF/MP2 iterations or 50 random-start iterations, without convergence evidence for all runs.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.