The trainability of photonic quantum circuits

This paper investigates the trainability of variational quantum algorithms (VQAs) implemented on photonic hardware. Specifically, it addresses why some photonic circuits remain trainable at scale while others suffer from vanishing gradients or prohibitive sampling costs, and it seeks to define a rigorous framework for identifying which observables allow for efficient training.

The authors introduce a trainability framework based on the ratio of sample variance to circuit variance. In this context, circuit variance measures the fluctuations of the loss function across the parameter space (related to the barren plateau phenomenon), while sample variance measures the statistical noise inherent in estimating the observable from finite measurements. By calculating these variances for photon-number observables—specifically photon-number monomials and polynomials—the authors determine the scaling of the number of samples required to resolve gradients as the system size increases.

The study identifies two distinct regimes for photonic circuits. Fixed-order photon-number polynomials (PNPs) exhibit a polynomially scaling variance ratio, meaning they can be trained efficiently as the system size grows. Conversely, high-order polynomials and observables based on output probabilities typically require exponentially many samples, rendering them untrainable. The authors also demonstrate that within the trainable regime, certain observables—such as those implemented via neural network structures—provide a polynomial speed-up over classical simulation methods, establishing photonic platforms as viable candidates for near-term quantum machine learning.