Lie Group Diffusion Models for Hardware-Aware Quantum Circuit Synthesis

Technical Summary: Lie Group Diffusion Models for Hardware-Aware Quantum Circuit Synthesis

Problem Statement
Quantum circuit synthesis aims to compile a target unitary operation $U^\star$ into a sequence of gates executable on a specific quantum processor. This problem possesses a natural hybrid structure: the entangling circuit pattern is discrete and dictated by hardware connectivity constraints, while the local single-qubit gates are continuous variables residing on the compact Lie group $SU(2)$. Existing machine learning approaches, including diffusion models, typically represent circuits in Euclidean space, ignoring the intrinsic curved geometry of the $SU(2)$ manifold and failing to natively incorporate hardware constraints such as qubit connectivity and gate error rates.

Methodology
The authors propose a two-component generative framework that integrates hardware constraints and the geometric structure of quantum gates directly into the synthesis process:

1. Circuit Skeleton Selector: A supervised classifier (a small Multi-Layer Perceptron) that selects a discrete circuit template (skeleton) from a library of nearest-neighbor CZ entangling patterns. The training labels for this selector are derived from a hardware-aware cost function that balances process fidelity against a depolarization cost based on specific gate error rates (e.g., $p_{CZ}$ and $p_{1Q}$).
2. Lie Group Diffusion Model: A skeleton-conditioned denoising diffusion probabilistic model (DDPM) that generates the continuous local single-qubit gates. Unlike standard Euclidean diffusion, this model operates directly on the product manifold $SU(2)^n \simeq S^{3n}$.
   • Representation: Gates are represented as unit quaternions.
   • Forward Process: Noise is added via right-multiplicative Lie algebra perturbations (sampling Gaussian tangent vectors in $\mathfrak{su}(2) \simeq \mathbb{R}^3$ and mapping them back to the group via the exponential map).
   • Reverse Process: The model predicts the noise using a Transformer encoder conditioned on the target Hamiltonian, evolution time, and the selected skeleton. The denoising target is derived from the heat kernel on $SU(2)$, ensuring the learned score respects the manifold's nonlinear geometry.
   • Refinement: The diffusion model acts as a proposal generator. The output circuits undergo a final local optimization step (gradient descent on the gate parameters) to maximize fidelity.

Key Contributions

• Manifold-Aware Diffusion: The work introduces a diffusion process that respects the $SU(2)$ geometry by utilizing the heat kernel and Lie algebra tangent spaces, rather than treating gate parameters as Euclidean vectors.
• Hardware-Aware Synthesis: The framework explicitly incorporates hardware constraints (connectivity and error rates) into the training objective, allowing the model to learn a trade-off between circuit fidelity and implementation cost.
• Hybrid Discrete-Continuous Architecture: By decoupling the discrete skeleton selection from the continuous gate generation, the model addresses the mixed-structure nature of the synthesis problem.
• Controllable Solution Styles: The authors demonstrate that the diffusion prior can be steered to produce circuits with specific physical properties (e.g., large vs. small gate rotation angles) by biasing the training data, without altering the target unitary.

Results
The model was evaluated on three-qubit Hamiltonian simulation targets, including the Transverse Field Ising Model (TFIM), Heisenberg-XXZ Model, and various Pauli-based targets.

• Compilation Success: The full pipeline (Selector + Lie Group Diffusion + Local Refinement) achieved a success rate of 75.8% (fidelity $F \ge 0.99$) across diverse Hamiltonian families. This outperformed both a Haar-random search baseline (65.3%) and a "generated search" baseline using a learned near-Clifford prior (68.3%).
• Ablation Studies:
  • Refinement Necessity: Diffusion without local refinement yielded a very low success rate (8.6%), while refinement starting from Haar-random gates also underperformed. The combination of diffusion and local refinement proved complementary: diffusion provides a geometry-aware global proposal, while refinement performs local high-fidelity optimization.
  • Selector Performance: The learned skeleton selector achieved 75.6% accuracy in matching the "oracle" skeleton (the best possible choice given the hardware cost). The selector successfully avoided unnecessary entanglement for local targets and selected deeper entangling templates only when required.
• Fidelity-Complexity Frontier: Analysis of the fidelity-cost landscape showed that the selector learns to concentrate probability on the "first successful" complexity class (e.g., selecting a 4-CZ skeleton over a 5-CZ skeleton if both meet the fidelity threshold, but the former has lower hardware cost).
• Steering: The model successfully generated circuits for the same target with significantly different total rotation angles, demonstrating control over the solution style.

Significance and Claims
The paper claims that Lie group diffusion provides a "natural generative framework" for hardware-aware quantum circuit synthesis. The significance lies in treating circuit synthesis as a constrained physical design problem where the generative prior respects the underlying geometry of quantum gates. The authors posit that this approach moves beyond random circuit generation to create structured proposal priors that balance fidelity with hardware overhead. They note that while the current work focuses on three-qubit systems as a nontrivial testbed, the modular architecture (separate selector and diffusion components) is designed to be extensible to larger qubit counts and more complex connectivity graphs (e.g., 2D grids or heavy-hex architectures) in future work. The method offers an interpretable machine learning approach where the discrete selection is driven by hardware costs and the continuous generation is grounded in Lie group theory.