Variational processing of multimode squeezed light
Integrated multimode quantum optics is a promising platform for scalable continuous-variable quantum technologies leveraging multimode squeezing in both the spatial and spectral domains. However, on-chip measurement, routing and processing the relevant ``supermodes'' over which the squeezing resource is distributed still scales quadratically with the number of modes $N$, causing rapid increase in photonic circuit size and number of required measurements. Here, we introduce a variational scheme, relying on self-configuring photonic networks (SCN) that learns and extracts the most-squeezed supermodes sequentially, reducing both the circuit size and the experimental overhead. Using homodyne measurement as a cost function, a sparse SCN discovers the $l\ll N$ most significant supermodes using $O(lN)$ physical elements and optimization steps. We analyze and numerically simulate these architectures for both real-space and frequency-domain implementations, showing a fidelity close to unity between the learned circuit and the supermode decomposition, even in the presence of optical losses and detection noise. In the frequency domain, we show that circuit size can be further reduced by using inverse-designed surrogate networks, which emulate the layers learned thus far. Using two different frequency encoding schemes -- uniformly- and non-uniformly-spaced frequency bins -- we reduce an entire network (learning all $N$ supermodes) to $O(N)$ and even $O(1)$ modulated cavities. Our results point toward chip-scale, resource-efficient quantum processing units and demultiplexers for continuous variable processing in multimode quantum optics, with applications ranging from quantum communication, metrology, and computation.