Researchers built a continuous‑variable photonic circuit that can learn the correlation structure of multivariate data by matching the circuit’s Pearson correlations to those measured in the data. The model uses two‑mode squeezing to create correlations, adjusts their strength with local phase shifts, and redistributes them across modes with an adjustable interferometer. Training is performed classically; the circuit parameters are tuned so that the simulated correlation matrix matches the empirical one, avoiding the need to generate samples during training.

After training, a simple affine map per mode adjusts the mean and variance of each output variable so that marginal statistics match the targets while keeping the learned correlations unchanged.

In a test on five financial assets, the circuit reproduced all ten pairwise correlations with residuals under 4.1 × 10⁻⁵ and a mean‑squared‑error of 3.7 × 10⁻¹⁰. For twenty assets, the Gaussian circuit correctly recovered the sign in 93.8 % ± 0.9 % of the 190 correlations and achieved 86 % of the pairs with residuals below 0.10.