Can a photonic mesh learn to undo squeezing and mode mixing?

Can a circuit correct a complicated optical transformation by watching its own output? Zhong and colleagues propose a way to train a photonic mesh to undo an unknown operation that mixes light modes and squeezes their fields. Instead of measuring the entire device, calculating an inverse on a computer, then programming it, their “symplectic Reck mesh” adjusts itself using local optical measurements.

A light mode can be described by two linked quadratures, analogous to the real and imaginary parts of its field. A Gaussian optical operation transforms these quadratures through a matrix. For a lossless device, that matrix must be symplectic: it preserves the relationships required by quantum mechanics. Ordinary programmable optical meshes can rearrange modes using phase shifters and beam splitters, but those passive components cannot represent every symplectic transformation. Squeezers add the missing capability: they reduce fluctuations in one quadrature while increasing them in the other.

The proposed mesh has a triangular sequence of two-mode gates, followed by one single-mode gate per mode. Put after the unknown device, it aims to make the combined response behave like the identity—leaving each input mode unchanged. The training proceeds one pair at a time. Small coherent light pulses probe a selected input mode; measurements of the output quadratures reveal how that input affects nearby outputs. The circuit then sets a two-mode gate to cancel one unwanted response block, moving through the mesh in a fixed order. Once the off-diagonal responses are removed, individual gates correct what remains on each mode. An output displacement can be measured and canceled separately.

The authors give two ways to choose each gate. A feedback optimizer can adjust its controls until the measured unwanted response is small, without needing a detailed calibration model. Alternatively, a closed-form update computes a gate from measured response blocks; that approach is direct but requires a model to translate the desired gate into hardware settings. Because each step measures the already-modified circuit, later updates can partly account for fixed errors introduced earlier.

A key mathematical result explains when one step becomes difficult. The authors define a local condition number, κ, from the area spanned by two measured response vectors relative to a conserved “pivot.” When κ equals one, passive optics can perform the cancellation; for κ greater than one, the gate requires squeezing, and larger κ can amplify measurement errors. A zero pivot can block a step for some responses. The full triangular sequence succeeds for generic lossless transformations—the exceptions occupy a measure-zero set—but being close to an exception can still make training fragile. Extra passive gates can avoid those failures in a modified, more component-heavy architecture.

The paper tests the idea in simulations, not on a photonic device. With matched total measurement budgets, the closed-form method had a median normalized residual of 0.0046 at 15 modes and 0.047 at 30 modes. Its per-setting shot count was about 1.07 million and 1.03 million, respectively; the feedback methods used 1,000 shots per setting. The residual increased with mesh size. In a separate simulated test with fixed gate-calibration errors, in-situ training became more accurate than tomography followed by offline compilation once calibration errors outweighed measurement noise. At small errors, tomography was slightly more accurate because its estimate was globally projected onto a symplectic matrix. These are results for the paper’s simulated targets, noise assumptions and training budgets, not universal performance guarantees.

Practical limits remain. The analysis assumes a lossless unknown operation. Uniform loss can still permit mode separation in simulations, but unequal loss leaves residual mixing; no mesh can restore quantum information lost to the environment. Near-failing nulling steps and larger meshes also demand more squeezing and can amplify errors. The authors suggest additional passive components to ease some squeezing requirements, at the cost of a larger circuit. Whether the proposed local measurements and gates can be implemented with sufficient precision in a real continuous-variable platform remains an open experimental question.