Heavy tails and pruning in programmable photonic circuits for universal unitaries

Area of Science:

• Quantum computing
• Photonic integrated circuits
• Deep learning hardware

Background:

• Programmable photonic circuits are promising for universal unitaries in quantum computing and deep learning.
• Increasing circuit scale amplifies noise, degrading fidelity of quantum operators and deep learning weight matrices.

Purpose of the Study:

• To demonstrate the stochastic nature of large-scale programmable photonic circuits.
• To develop high-fidelity universal unitaries by pruning superfluous rotations.
• To reduce the complexity and improve the efficiency of photonic hardware for quantum and deep learning applications.

Main Methods:

• Analysis of large-scale programmable photonic circuits to identify heavy-tailed distributions of rotation operators.
• Application of network pruning techniques, inspired by the Pareto principle, to photonic circuit architecture.
• Extraction of a universal architecture for pruning random unitary matrices in the Clements design.

Main Results:

• Demonstrated a nontrivial stochastic nature in large-scale programmable photonic circuits, characterized by heavy-tailed distributions.
• Revealed power law and Pareto principle in conventional programmable photonic circuits due to hub phase shifters.
• Showcased that removing superfluous rotations ('the bad') improves fidelity and energy efficiency in photonic hardware.

Conclusions:

• Designed pruning of superfluous rotations enables high-fidelity universal unitaries in large-scale photonic circuits.
• Network pruning is applicable to photonic hardware design, leveraging revealed power law and Pareto principle.
• This work lowers the barrier for achieving high fidelity in quantum computing and photonic deep learning accelerators.