A continuous-variable photonic circuit learns asset correlations: across 190 pairs of twenty stocks, signs and magnitudes come back
Synopsis
A continuous-variable photonic generative model hands the correlation structure of asset returns to an optical circuit: two-mode squeezing creates correlations, local phases set their sign, and a programmable interferometer redistributes them across modes, while training matches only the covariance matrix and needs no sampling. On five assets' daily returns, the optimized circuit reproduces the sign and magnitude of all ten pairwise correlations; on twenty S&P 500 equities, most of the 190 pair signs are recovered and the residuals cluster on the small side.
Interpretation
The task of generating multivariate data moves into a continuous-variable optical circuit: the position quadrature of every mode is read by homodyne detection, one measurement yields one multidimensional sample, and the correlation structure follows directly from circuit parameters. Earlier continuous-variable Born machines also read homodyne samples but trained on those samples with a kernel-based loss, with demonstrations limited to single-mode targets and quantum states of up to three modes. On daily log returns of five assets (AAPL, MSFT, GLD, TLT, BTC-USD), the optimized circuit reproduces the sign and magnitude of all ten pairwise correlations, including the negative correlations of TLT with the two equities and the weak correlations involving GLD.
Training skips sampling and aligns only the covariance matrix of the simulated state with the Pearson correlation matrix computed from data, keeping the pipeline inside Gaussian phase-space methods. The loss is a maximum mean discrepancy between rows of the two correlation matrices, minimized with bound-constrained L-BFGS-B; five modes carry five phases, three squeezing amplitudes and ten beam-splitter angles. All ten perturbed initializations of the five-mode circuit reach a loss below 0.001 within 4600 function evaluations; the twenty-mode circuit with 220 trainable parameters and 190 target entries converges to a comparable plateau across five initializations.
A weak Kerr gate is written as a coherent sum of Gaussian branches, so a non-Gaussian element can still be simulated by symplectic propagation. The Kerr generator is quartic in the quadratures and has no symplectic representation, so direct simulation leaves the Gaussian methods. A finite sum from the Hubbard-Stratonovich identity and Levin collocation is compared against exact Kerr evolution in the Fock backend in the weak-nonlinearity, low-occupation regime, with error growing in strength and photon number.
Perspective
The demonstration runs entirely in classical simulations with Strawberry Fields and Piquasso, on daily log returns of five assets and twenty S&P 500 equities, with the sample correlation matrix over a fixed window as the training target and no forecasting involved.
Training minimizes only the distance between correlation matrices and does not guarantee a match of the full distributions; generated marginals stay close to normal while the target returns have sharper peaks and heavier tails, and BTC-USD, the most volatile asset, retains the largest total variation distance of 0.330 after the affine map. The twenty-equity target correlation matrix is estimated from 48 observations and is itself noisy.
