Future of Continuous Variable Quantum Computing with DRL
Deep Reinforcement Learning for Non-Gaussian Photonic Quantum State Preparation
Researchers developed a deep reinforcement learning system to produce non-Gaussian states for photonic quantum computing. This machine learning method uses an adaptive, iterative procedure to produce cubic-phase gates with a 96% success rate, unlike other methods that need extreme physical parameters or inefficient post-selection. Training an agent to adjust optical components in real time helps the system handle photon-number-resolving measurements' inherent uncertainty. The paper also proposes directly generating quartic-phase states to avoid complex gate decompositions. Reinforcement learning in quantum phase space provides a scalable path to universal and fault-tolerant continuous variable quantum computing.
Non-Gaussian Gates Problem
Continuous variable quantum computing uses bosonic field (qumode) encoding instead of discrete qubits, which allows quantum optics to scale in free space and on-chip. To be universal, CVQC needs non-Gaussian evolution, at least cubic Hamiltonian evolution.
In optics, establishing these states “classically” is notoriously difficult. Microwave fields in superconducting circuits can achieve large nonlinearities, but third-order optical nonlinearities are too weak for deterministic state preparation. Early “quantum mechanical” attempts to solve this problem included probabilistic photon-number-resolving (PNR) measurements, which had to squeeze 17 dB and detect up to 50 photons.
AI as Quantum Architect
The study team overcame these constraints using deep reinforcement learning. A quantum optical circuit-interacting learning agent employs reinforcement learning to pick the appropriate action based on a reward signal.
A quantum circuit Markov decision process (MDP) model was built. To optimize output state fidelity compared to a target cubic-phase state, the agent adjusted beamsplitter transmittivity, squeezing levels, and displacements. Proximal policy optimization (PPO), chosen for its durability and on-policy character, educated the agent over 5.7 million time steps.
The 96% Success Rate
These numerical trials yielded excellent results. The DRL-driven approach achieved 96% success rate in generating cubic-phase states with γ=0.2. The technique is compatible with current experimental equipment because it used far lower PNR measurement levels and less than 10 dB of squeezing than previous proposals.
After analyzing 1,000 instances, the researchers found noteworthy AI traits:
Self-Correction: The agent often made minor “corrective” displacements to perfect the state before ending with great fidelity.
Environment Resets: This method enhanced loop utilization by instructing the agent to “reset” the circuit and start again if a quantum path failed.
Robustness to Loss: Even with photon loss (99% detector efficiency), the agent modified its approach despite taking longer to train and oscillating in its final displacement steps.
Direct Quartic Gates Cut Complexity
The study introduced quartic-phase gate preparation and cubic-phase states. Creating a quartic-phase gate used to require 29 gates, 15 of which had to be cubic.
The researchers developed a quantum technique to directly create these gates using PNR-based resources by “stamping” the quantum Wigner function to create cubic-polynomial contours. Even while it is probabilistic and requires postselection, this direct technique lays the framework for a near-deterministic machine learning implementation that might drastically reduce quantum computer complexity.
Qumode encoding in CVQC has what benefits? For CVQC, Qumode encoding, which uses bosonic fields instead of native qubits, has many advantages.
Excellent Scalability: Qumode encoding uses quantum optics' scalability in on-chip and free space settings.
By permitting hybrid bosonic qubit encoding, researchers can use Gottesman, Kitaev, and Preskill (GKP) states to encode qubits inside oscillators.
Continuous variable quantum computing systems using qumode encoding can be fault-tolerant and scalable.
A specific platform for quantum field theory simulation is provided by this encoding.
Universality: Qumode encoding and quantum field cubic Hamiltonian evolution make the CVQC system universal.
