A Givens-exchange ansatz for molecular variational eigensolvers

Technical Summary: A Givens-exchange ansatz for molecular variational eigensolvers

Problem Statement
Accurate calculation of molecular ground-state energies is essential for computational drug discovery, particularly for systems involving strong correlation or large active spaces where classical electronic-structure approximations become expensive or unreliable. The Variational Quantum Eigensolver (VQE) addresses this by optimizing a parameterized quantum circuit to minimize the expectation value of a qubit Hamiltonian. However, the practical efficacy of VQE is constrained by ansatz design:

• Chemically motivated ansatzes (e.g., UCCSD) encode physical excitation structures but often result in circuits too deep for near-term hardware.
• Hardware-efficient ansatzes use shallow layers but may lack the expressivity to capture molecular correlations or suffer from poor trainability.
• Quantum Architecture Search (QAS) methods (e.g., Reinforcement Learning-based approaches like HyRLQAS) automate circuit construction to find compact, high-performing architectures. While effective, these methods introduce significant computational overhead through repeated circuit construction and policy updates, and the resulting architectures can be difficult to interpret chemically.

The paper addresses the need for an ansatz that balances accuracy, reproducibility, and interpretability without relying on the computational cost of unrestricted architecture search.

Methodology
The authors propose a fixed-topology Givens-exchange ansatz that avoids architecture search entirely. The methodology is grounded in the geometric interpretation of the complex Stiefel manifold, where the ansatz represents an ordered unitary trajectory preserving state normalization.

• Circuit Structure:
  • Initialization: The circuit starts from a computational-basis state $|b_0\rangle$ selected to have the lowest diagonal Hamiltonian expectation value.
  • Layers: The ansatz consists of $L=2$ blocks of all-pair Givens exchange macros interleaved with local $R_Y$ rotation layers.
  • Givens Exchange Macro ($G_{ij}$): A two-qubit operation acting on qubits $i$ and $j$ that rotates the single-excitation subspace spanned by $|01\rangle$ and $|10\rangle$ while leaving $|00\rangle$ and $|11\rangle$ invariant. In a Jordan–Wigner (JW) occupation basis, this corresponds to a particle-conserving single-excitation operation.
  • Sequence: The full state is prepared as $|\psi(\theta)\rangle = R_L G_{L-1} \dots G_0 R_0 |b_0\rangle$, where $R$ layers contain local $R_Y$ rotations and $G$ layers contain ordered sequences of Givens exchanges for all qubit pairs.
• Optimization Protocol:
  • Parameters (rotation angles $\alpha$ and exchange angles $\beta$) are optimized using the Adam optimizer to minimize the Hamiltonian expectation value.
  • Initialization: Angles are initialized from a zero-mean normal distribution (scale 0.05).
  • Evaluation: Exact diagonalization is used only after optimization terminates to compute ground-state energies, errors, and fidelities. It is not part of the variational loop.
  • Reproducibility: Results are reported as six-seed means (seeds: 17, 11111, 22222, 33333, 44444, 55555) to demonstrate typical performance rather than cherry-picking the best seed.

Key Contributions

1. Definition of a Fixed Topology: The authors define a mathematically explicit Givens-exchange ansatz with a fixed, ordered sequence of rotation and exchange layers, removing the need for architecture search.
2. Stiefel-Manifold Interpretation: The work frames the ansatz as a structured trajectory on the Stiefel manifold, providing a geometric justification for using ordered exchange macros as chemically meaningful variational building blocks.
3. Rigorous Benchmarking Protocol:
   • The study distinguishes between coefficient-verified Hamiltonians (LiH-6, H2O-8) and public-specification candidates (BeH2-6) or stress cases (LiH-4 parity-mapped).
   • It separates the primary Jordan–Wigner comparisons from parity-mapped/tapered cases to avoid conflating different variational landscapes.
   • It reports results as six-seed means with standard deviations, offering a transparent view of optimization sensitivity.
4. Resource Accounting Clarity: The paper explicitly separates pre-compilation macro counts (Givens exchanges) from hardware-native gate counts (CNOTs), acknowledging that Givens macros require compilation to native gates.

Results
The method was evaluated on LiH-6, BeH2-6, and H2O-8 Hamiltonians using the Jordan–Wigner mapping without tapering.

• Chemical Accuracy: Across all six seeds, the method achieved chemical accuracy (error < 1.60 mHa) for LiH-6, H2O-8, and the BeH2-6 candidate in every run.
• Error Magnitudes:
  • LiH-6: Mean error of $0.000000124$ Ha ($0.000124$ mHa).
  • H2O-8: Mean error of $0.000128558$ Ha ($0.128558$ mHa).
  • BeH2-6: Mean error of $0.000002152$ Ha ($0.002152$ mHa).
• Comparison with QAS Baselines:
  • On LiH-6, the six-seed mean error ($0.000124$ mHa) was lower than the published point errors of HyRLQAS ($0.590$ mHa), TensorRL-QAS, CRLQAS, and other reinforcement learning-based baselines.
  • On H2O-8, the mean error ($0.128558$ mHa) was also lower than the published point errors of the compared QAS methods.
• Stress Tests:
  • The method showed sensitivity to initialization and mapping changes. For the LiH-4 parity-mapped stress case, 5 out of 6 seeds passed chemical accuracy, but the distribution included high-error local solutions.
  • For LiH-6 at a stretched geometry (3.4 Å), performance degraded significantly, with only 1/6 seeds passing for JW mapping and 0/6 for parity mapping.
• Ablation: A depth ablation showed that $L=2$ (two exchange blocks) was necessary to achieve chemical accuracy across all seeds; $L=1$ was insufficient for BeH2 and H2O.

Significance and Claims
The paper claims that a fixed-topology Givens-exchange ansatz serves as an accurate, reproducible, and search-free reference template for molecular VQEs.

• Accuracy vs. Search Cost: The method achieves state-of-the-art accuracy (lower mean errors than several QAS methods) without the computational overhead of learning a gate-placement policy.
• Trade-off: The authors acknowledge that their ansatz uses a larger "pre-compilation macro budget" (more two-qubit operations) than some compact, search-optimized circuits. They frame this not as a hardware-resource victory, but as a trade-off: paying a higher macro-operation cost to gain a transparent, systematically defined prior that avoids the instability and cost of architecture search.
• Scope: The method is most effective for near-equilibrium Hamiltonians with direct occupation-like interpretations (e.g., Jordan–Wigner). Its performance degrades on stretched bonds or parity-mapped effective Hamiltonians, highlighting the importance of the ansatz's geometric alignment with the specific encoding.
• Role: The primary contribution is providing a baseline against which search-based methods and future compiled variants can be evaluated, demonstrating that simple, fixed molecular templates can outperform complex learned architectures in terms of mean accuracy and reproducibility.