Classical simulation and model concentration in passive linear optics

This paper investigates whether the phenomenon of barren plateaus—where loss functions concentrate exponentially, hindering the training of quantum circuits—exists in passive linear-optical systems. Specifically, it explores the relationship between this concentration and the classical simulability of bosonic systems, testing if there is a fundamental tradeoff similar to that observed in qubit-based architectures.

The authors employ a representation-theoretic framework to analyze the variance of expectation values for particle-number-preserving observables under random passive linear-optical transformations. By decomposing the operator space into irreducible representations (irreps) of the unitary group, they characterize the concentration of these values based on the projection norms (irrep purities) of the input states and observables. They interpret these purities as measures of generalized entanglement and locality, respectively, and compare the resulting concentration profiles against known classical simulation algorithms.

The study establishes that concentration is governed by the misalignment of the projections of the input state and the observable into irreducible representations. While they identify specific Fock-state inputs and observables that evade exponential concentration, they find that these regimes do not provide a clean separation from classical methods. In most cases, the signal remains largely tractable, and any residual part not exponentially suppressed is small enough that it can be approximated by a classical surrogate with polynomially small error. The framework provides a systematic method for identifying potential regimes that might eventually combine the absence of exponential concentration with computational hardness.

Understanding the limits of trainability in passive linear optics is crucial for the development of near-term quantum algorithms. By mapping the landscape of concentration and simulability, this work helps researchers identify which photonic architectures are likely to suffer from vanishing gradients and which might offer a genuine quantum advantage, guiding the design of more robust variational quantum algorithms.