Quantum Computing News

Deep Reinforcement Learning for Non-Gaussian Photonic Quantum State Preparation

To address the historical challenges of producing non-Gaussian states for photonic quantum computing, researchers have created a deep reinforcement learning framework. This machine learning method achieves a 96% success rate by using an adaptive, iterative procedure, whereas existing approaches for producing cubic-phase gates frequently need extreme physical parameters or inefficient post-selection. The system efficiently manages the intrinsic unpredictability of photon-number-resolving measurements by training an agent to modify optical components in real-time. Additionally, the work presents a unique approach that may be able to avoid the requirement for intricate gate decompositions by directly generating quartic-phase states. The authors offer a scalable route toward universal and fault-tolerant continuous variable quantum computing by using reinforcement learning to traverse quantum phase space.

You can also read Schur-Weyl Duality Allows a Quantum System Work Extraction

The Non-Gaussian Gates Challenge

Unlike traditional quantum computers that use discrete qubits, continuous variable quantum computing uses bosonic field (qumode) encoding, which provides exceptional scalability through quantum optics in both free space and on-chip configurations. However, access to non-Gaussian evolution, at least cubic Hamiltonian evolution, is necessary for CVQC to be genuinely universal.

In the optical realm, creating these states “classically” is infamously hard. Large nonlinearities can be achieved by microwave fields in superconducting circuits, but deterministic state preparation cannot use third-order optical nonlinearities since they are far too weak. Photon-number-resolving (PNR) measurements were used in earlier “quantum mechanical” attempts to tackle this problem, but they were mostly probabilistic and needed stringent requirements, such squeezing levels of 17 dB and detecting up to 50 photons.

You can also read Why the PNR Photon number resolving detector matters in 2026

AI as the Quantum Architect

The study team used a deep reinforcement learning framework to get over these restrictions. Through interaction with an environment, in this case a quantum optical circuit, a learning agent uses reinforcement learning to choose the best course of action based on a reward signal.

The group created a Markov decision process (MDP) model of the quantum circuit. To optimize the output state’s fidelity in comparison to a target cubic-phase state, the agent had to adjust circuit characteristics, including beamsplitter transmittivity, squeezing levels, and displacements. The agent was trained over 5.7 million time steps using the proximal policy optimization (PPO) approach, which was selected because of its resilience and on-policy nature.

The 96% Success Rate

These numerical trials produced impressive findings. With a value of γ=0.2, the DRL-driven procedure generated cubic-phase states with a 96% success rate. The technique is compatible with current experimental equipment since it was accomplished with significantly lower PNR measurement values and less than 10 dB of squeezing than earlier ideas.

The researchers noticed intriguing emergent characteristics from the AI when evaluating 1,000 episodes:

• Self-Correction: The agent would frequently make small “corrective” displacements to perfect the state before concluding after achieving high fidelity.
• Environment Resets: This technique maximized the utilization of the loop by teaching the agent to “reset” the circuit and begin from the beginning state if it realized that a certain quantum path was unlikely to succeed.
• Robustness to Loss: Although it took more training time and displayed oscillatory behaviors in its final displacement steps, the agent adjusted its approach even when simulated with photon loss (99% detector efficiency).

You can also read IonQ ID Quantique Achieves ISO 14001 for Sustainable Quantum

Overcoming Complexity: Quartic Gates Directly

In addition to cubic-phase states, the study pioneered quartic-phase gate preparation. In the past, achieving a quartic-phase gate necessitated a laborious breakdown into 29 distinct gates, 15 of which had to be cubic.

By “stamping” the quantum Wigner function to produce the required cubic-polynomial contours, the researchers presented a novel quantum technique that allows the direct creation of these gates utilizing the same PNR-based resources. This direct approach lays the groundwork for a future, near-deterministic machine learning implementation that might greatly lower the complexity of quantum computers, even if it is now probabilistic and necessitates postselection.

Feature	Cubic-Phase Gate Generation	Quartic-Phase Gate Generation
Generation Method	Iterative process using a quantum optical circuit with an added in-loop displacement.	Direct generation using a fundamental quantum optical algorithm.
Success Rate	Near-deterministic (96%) success rate achieved through training.	Currently probabilistic, requiring postselection of specific photon number detection patterns (n1 = n2).
Traditional Complexity	Serves as a standard resource for universal continuous variable quantum computing.	Traditionally required a decomposition into 29 separate gates, 15 of which were cubic.
Machine Learning Status	Driven by a deep reinforcement learning (DRL) agent using Proximal Policy Optimization (PPO).	Full numerical ML simulations are a work in progress due to higher computational requirements.
Squeezing Levels	Requires squeezing no higher than 10 dB.	Squeezing is held constant at a higher level, specifically 12 dB (r=1.38).
Hilbert Space Truncation	Simulates effectively with a truncation of 31 photons.	Requires a significantly larger Hilbert space of at least 60 photons.
Physical Mechanism	Leverages interference in quantum phase space to “shape” the Wigner function to high cubicity.	“Stamps” the Wigner function with a displaced Fock state at nearly opposite azimuths in phase space.