Light passing through a birefringent crystal changes its polarization depending on wavelength. This profile is like a mosaic assembled from rare flashes, where each shard hints at the shape of its neighbors. The problem is that under photon starvation, when only a handful of photons are detected, the accuracy of estimates hits a fundamental limit: shot noise dominates and the signal-to-noise ratio plummets. In such situations, the standard classical shadows method, which works pixel by pixel, yields scattered points with no connection between them. The new work proposes a different path: functional classical shadows (FCS), which reconstruct the profile as a single function using the entire dataset.

The idea was born from quantum tomography: there, classical shadows estimate properties of a quantum state from a small number of measurements, but they require identical copies of the state. In polarimetry, however, each measurement yields a state with a different value of a classical parameter, namely wavelength. The authors formalize the problem as supervised learning of a function ρ(x) that maps wavelength to a polarization state. For each event, a random unitary transformation is applied, followed by a measurement in a fixed basis. From these data, a classical snapshot is computed: a mathematical object that need not be a physical state but allows estimating the loss. Instead of pixel-by-pixel optimization, they introduce a parametrization of the hypothesis η(x) as a pure state with angles θ(x) and φ(x), and minimize a global loss function averaged over the entire training set.

The experiment confirmed the approach: in the wavelength range 808–812 nm, the phase and polarization angle profiles were recovered. The standard method gives scattered points, while FCS reproduces a smooth linear dependence by using correlations between neighboring wavelengths. Theoretically, they prove that the loss estimate is unbiased and the variance is bounded, which guarantees the method's efficiency for a number of samples T = O(maxₓ||η₀(x)||²_shadow,x / ε² · log(|F|/δ)). The key advantage is using the entire dataset to reconstruct a single function, which provides better regularization than directly fitting individual pixels.

The method's prospects are broad: it can be extended to mixed-state hypotheses using trace distance and optimal Helstrom projectors, which would allow application to more realistic noisy scenarios. It is also interesting to adapt FCS to phase imaging and other types of optical measurements where the function ρ(x) describes spatial or temporal dependencies. Combining it with active learning methods for optimal selection of measurement points is promising. Overall, the work transfers the classical shadows paradigm from quantum computing to metrology with natural parameter variability, opening the way for applying quantum-inspired methods in classical polarimetry and phase imaging where every photon counts.