Abstract
Quantum neural networks (QNNs) and parameterized quantum circuits (PQCs) are essential for near-term quantum machine learning but face scalability issues due to over-parameterization and hardware constraints. We introduce LiePrune, a mathematically grounded one-shot structured pruning framework. By representing gates in a dual Lie group–Lie algebra and quantum geometric space, LiePrune enables aggressive compression. Experiments demonstrate over 10× compression with minimal performance loss across classification and generative tasks, supported by theoretical guarantees on redundancy detection and functional approximation.
Introduction: The Scalability Hurdle in QNNs
Quantum Neural Networks (QNNs) are notoriously difficult to scale. They often suffer from "over-parameterization," where circuits contain more gates than necessary, leading to "barren plateaus"—regions where gradients vanish and training stalls. Furthermore, these dense circuits frequently exceed the limited "qubit budgets" and gate depth capacities of Noisy Intermediate-Scale Quantum (NISQ) hardware.
While classical deep learning has successfully employed pruning to remove unnecessary neurons, quantum circuits present unique challenges. Quantum gates must remain "unitary" (mathematically reversible) to be physically implementable. Naive pruning often violates this structure. Additionally, quantum-specific properties, such as the curvature of the Hilbert space, are typically ignored by classical algorithms, leaving significant optimization potential untapped.
Methodology: The LiePrune Framework
LiePrune addresses these challenges by treating quantum gates as elements of a Lie Group. The core innovation is a Dual-Space Identifier that evaluates gates through two lenses:
- Lie Algebra Space: Every unitary gate is represented as the exponential of a Lie algebra element (its "generator"). LiePrune analyzes these generators to identify mathematically similar gates.
- Quantum Geometric Space: The framework utilizes the Fubini–Study distance—a measure of how much a gate alters a quantum state—and the Quantum Fisher Information to determine a gate's "sensitivity" or task-specific importance.
The Workflow
- Partitioning: Gates are grouped into subsets based on their Lie subgroups (e.g., gates acting on the same qubits in a similar manner).
- Redundancy Graph: Within these groups, the algorithm constructs a graph where edges represent "geometrically close" (redundant) gates.
- One-Shot Merging: Instead of simple deletion, LiePrune selects a "core" gate and merges redundant neighbors into it via weighted Lie algebra addition. This "one-shot" process occurs in a single pass, eliminating the need for expensive iterative re-training.
Experimental Evaluation
LiePrune was evaluated across four distinct benchmarks:
- MNIST (4-vs-9): Binary classification of handwritten digits (8 qubits).
- FashionMNIST (Sandal-vs-Boot): Image classification of clothing (10 qubits).
- Bars-and-Stripes (BAS): Quantum generative modeling (8 qubits).
- LiH VQE: Variational Quantum Eigensolver for the ground-state energy of Lithium Hydride (12 qubits).
Key Results and Insights
LiePrune achieved significant compression across all tasks. On MNIST and Bars-and-Stripes, the framework reached 8× compression (reducing parameters from 288 to 36), while FashionMNIST saw 10× compression.
As shown in the results, while "raw" pruning (without fine-tuning) initially causes accuracy drops, a brief period of post-pruning fine-tuning (+FT) recovers nearly all performance. Notably, in the Bars-and-Stripes task, the pruned model slightly outperformed the original (0.751 vs 0.743), suggesting that pruning may act as a beneficial form of regularization.
The VQE Challenge
Pruning proved more difficult for quantum chemistry (LiH VQE). Although 12× compression was achieved, the energy error remained high even after fine-tuning. This indicates that chemical simulations are far more sensitive to gate removal than classification tasks, requiring higher precision.
Discussion and Future Directions
LiePrune demonstrates that algebraic "compression" is superior to simple deletion. By merging generators, the underlying physical structure of the circuit is preserved. Furthermore, the algorithm's linear scalability makes it suitable for real-time compression on edge quantum devices.
However, limitations remain. The sensitivity of VQE tasks suggests a need for "chemistry-aware" or symmetry-preserving pruning strategies. Future work will also focus on testing LiePrune on actual NISQ hardware to evaluate its performance against real-world gate errors and decoherence, as well as exploring redundancies across different Lie subgroups to further increase compression ratios.
Summary updated