DifGa: differentiable error mitigation for multi-mode Gaussian and non-Gaussian noise in quantum photonic circuits

Abstract We present DifGa , a fully differentiable error-mitigation framework for continuous-variable photonic circuits under Gaussian loss and weak non-Gaussian phase noise. DifGa is an observable-level method: it targets first-moment quadratures ⟨ x ^ 0 ⟩ and ⟨ p ^ 0 ⟩ , rather than full-state recovery or fault-tolerant correction. Using PennyLane’s analytic default.gaussian backend, we model optical loss as beam-splitter coupling to vacuum ( η ∈ [ 0.3 , 0.95 ] ) and phase jitter via differentiable Monte-Carlo mixtures ( δ ∈ [ 0 , 0.7 ] ). A trainable Gaussian recovery layer (local rotations and displacements) is optimized end-to-end by automatic differentiation. In this analytic setting, the method reduces quadrature reconstruction loss by several orders of magnitude under Gaussian loss, reaching numerical residuals for moderate transmissivity; these values should be interpreted as moment-level numerical inversion in simulation, not full physical suppression of all noise channels. Under non-Gaussian jitter, noise-aware training improves robustness relative to Gaussian-only training, and a conservative 10%-of-baseline criterion yields an empirical mitigation threshold of δ ∗ ≈ 0.60 . Runtime scales approximately linearly with the Monte-Carlo sample count. DifGa remains fully within the Gaussian operation set and is consistent with Gaussian no-go theorems.