DifGa: Differentiable Error Mitigation for Multi-Mode Gaussian and Non-Gaussian Noise in Quantum Photonic Circuits

Technical Summary of "DifGa: Differentiable Error Mitigation for Multi-Mode Gaussian and Non-Gaussian Noise in Quantum Photonic Circuits"

Problem Statement
Continuous-variable (CV) quantum photonic circuits, which utilize Gaussian operations and room-temperature hardware, are promising for scalable quantum information processing. However, these systems are fundamentally constrained by optical loss, mode mismatch, and phase drift. While bosonic quantum error-correcting codes (such as GKP) offer fault tolerance, they require non-Gaussian resources, high squeezing, and complex measurements that are currently beyond the capabilities of near-term photonic processors like Xanadu's Borealis.

Conversely, existing error mitigation techniques developed for qubit-based systems (e.g., zero-noise extrapolation) do not translate to CV systems because Gaussian noise acts on phase-space moments rather than discrete Pauli operators. Furthermore, rigorous no-go theorems (Niset et al., Ohliger et al.) establish that perfect Gaussian quantum error correction of Gaussian noise is impossible using only Gaussian operations and vacuum ancillas. Consequently, there is a lack of a hardware-compatible, differentiable framework capable of mitigating noise in CV photonic circuits while operating strictly within the Gaussian regime and targeting experimentally accessible observables.

Methodology
The authors introduce DifGa (Differentiable Gaussian Error Mitigation), a fully differentiable framework designed to suppress noise-induced degradation of quadrature observables ($\langle \hat{x} \rangle$ and $\langle \hat{p} \rangle$) without attempting full state recovery. The methodology relies on the following components:

• Simulation Environment: The framework utilizes the default.gaussian backend of PennyLane, which represents quantum states via their first and second statistical moments (mean vector and covariance matrix), enabling end-to-end automatic differentiation.
• Circuit Architecture: The system employs a multi-mode architecture consisting of a signal mode, an ancilla mode, and an environmental mode (to model loss).
  • Encoding: Input states are prepared using squeezing and displacement operations, followed by an entangling beam splitter.
  • Noise Modeling:
    • Gaussian Loss: Modeled as a beam splitter interaction with a vacuum environment with transmissivity $\eta \in [0.3, 0.95]$.
    • Non-Gaussian Noise: Weak non-Gaussian phase jitter is incorporated via a differentiable Monte-Carlo mixture of random phase rotations with jitter amplitudes $\delta \in [0, 0.7]$.
• Trainable Recovery Layer: A six-parameter trainable Gaussian layer is appended to the circuit, acting locally on the signal and ancilla modes. This layer consists of local phase rotations and complex displacements.
• Optimization: The recovery parameters are optimized via gradient descent to minimize a quadratic loss function defined as the squared difference between the ideal and noisy quadrature expectations. The framework supports both pure Gaussian training and "noise-aware" training using Monte-Carlo averaging for non-Gaussian scenarios.

Key Contributions

1. First End-to-End Differentiable Gaussian Mitigation: The paper presents the first framework for designing Gaussian-only error mitigation protocols that are fully differentiable and compatible with current photonic hardware, avoiding the need for non-Gaussian ancilla states.
2. Gradient Propagation through Mixed States: The authors demonstrate that gradients can be propagated through full mixed-state Gaussian simulations (including Monte-Carlo mixtures for non-Gaussian noise) using PennyLane, enabling classical optimization of noise-aware interferometers.
3. Observable-Level Mitigation: The work shifts the objective from impossible full-state correction to the practical restoration of first-moment quadrature expectations, a regime where Gaussian operations can be effective.
4. Empirical Validation of Ancillary Assistance: The study numerically demonstrates that even shallow Gaussian recovery circuits, when assisted by ancillary modes, can suppress reconstruction errors by several orders of magnitude under moderate loss.

Results

• Gaussian Loss Mitigation: Under pure Gaussian loss, the optimized recovery layer suppresses reconstruction error to near machine precision ($< 10^{-30}$) for moderate transmissivities ($\eta \geq 0.5$). Even at lower transmissivity ($\eta = 0.3$), errors remain extremely small ($\sim 10^{-38}$).
• Multi-Mode vs. Single-Mode: Multi-mode architectures significantly outperform single-mode baselines. While a single-mode baseline without recovery yields an error of $\sim 1.7 \times 10^{-1}$, the multi-mode baseline reduces this to $\sim 1.0 \times 10^{-1}$ passively. With the active trainable recovery layer, the multi-mode system drives the error to numerical zero ($\sim 10^{-41}$), representing a two-order-of-magnitude improvement over the multi-mode baseline.
• Non-Gaussian Robustness: When trained explicitly with non-Gaussian phase jitter, the recovery layer generalizes effectively. For phase jitter amplitudes $\delta \leq 0.2$, the error remains below $10^{-3}$. While performance degrades as $\delta$ increases, noise-aware training reduces errors by more than an order of magnitude compared to Gaussian-trained recovery at large jitter levels ($\delta = 0.7$).
• Scaling and Cost: The computational cost scales linearly with the number of Monte-Carlo samples ($K$). Runtime benchmarks confirm predictable scaling, with moderate sample sizes ($K=8\text{--}16$) providing robust mitigation with runtimes under 10 seconds. Increasing the number of modes ($M$) further improves mitigation performance, driving errors to numerical precision across the tested range.

Significance and Claims
The paper positions DifGa not as a violation of known Gaussian no-go theorems, but as a complementary, experimentally relevant approach. The authors explicitly state that they do not claim to achieve fault-tolerant logical error correction or full state restoration. Instead, the significance lies in providing a software-native, hardware-aligned strategy for noise suppression in near-term photonic processors.

By operating entirely within the Gaussian subspace using physically realizable components (squeezing, displacement, beam splitters) and vacuum ancillas, DifGa offers a practical pathway to enhance the fidelity of measurable observables in noisy photonic circuits. The framework leverages the classical simulability of Gaussian circuits (as noted by Mari and Eisert) as an enabling feature for efficient gradient-based optimization, distinguishing it from non-Gaussian error correction schemes that require resources currently unavailable in integrated photonic platforms.