When does photon entanglement still allow classical estimates?

The input is a chain, not a list of independent photons

An optical interferometer mixes modes occupied by indistinguishable photons. For a product input, classical methods can estimate some output quantities to additive accuracy without reproducing the device’s sampling process. Entanglement seems to obstruct those methods because the amplitudes combine many input configurations. The authors show that this obstacle can be handled for a broad structured family: photonic matrix-product states, or MPSs.

An MPS represents a sequence of sites linked by a finite-dimensional internal index. Here each site contains one photon, which may occupy one of L modes; the index connecting neighboring sites has dimension χ, called the bond dimension. It tracks entanglement across cuts in that chosen site order, not every possible notion of multipartite entanglement. For light emitted one photon at a time, the source’s retained quantum memory bounds χ. A single two-level spin gives χ no greater than 2, regardless of how many photons are emitted. A memory of k qubits gives χ ≤ 2^k.

The central contribution is an unbiased Monte Carlo estimator for any chosen output amplitude of an m-mode interferometer. Rather than enumerate the MPS’s potentially exponentially many configurations, it samples a continuous Gaussian-weighted distribution one site at a time. At each step it contracts the already-sampled part of the chain with a compact summary of the remaining sites. This makes the sampling exact in the mathematical model and allows the estimator to handle entangled inputs without expanding them into separate product states.

For N independent draws, the amplitude estimate has mean-square error at most ||ψ||²/N, where ||ψ|| is the input-state norm. Each draw costs O(nLχ² + nL³ + mn) operations after O(nLχ³) preprocessing; here n is the photon count. The authors also give an additive-error guarantee: accuracy ε with failure probability at most δ uses O(||ψ||² ε⁻² log(1/δ)) draws. Thus the work is polynomial in the bond dimension and inverse target error, and for sequential sources remains polynomial in photon count when the retained memory grows no faster than logarithmically with n, assuming the other problem parameters are polynomially bounded.

The method also extends to estimates of probabilities when the emitter memory is discarded, overlaps between states, and selected output-mode marginals. These are useful distinctions: a marginal can combine many complete output patterns into a coarser quantity, while an overlap compares states without listing every output. The paper reports numerical comparisons with brute-force calculations and formal proofs in Lean for several central identities and sampling laws. Its operation-count analysis, numerical results and complexity discussion are not all covered by those formal proofs.

What the guarantee does not say

The result concerns estimates to additive precision, not relative precision. For outcomes whose probabilities are exponentially small, an additive error can be far larger than the probability itself; the algorithm therefore does not make such rare outcomes accurately measurable. Nor does estimating amplitudes, overlaps or marginals provide a classical sampler for the full boson-sampling output distribution. The authors explicitly leave the classical hardness of sampling for these MPS inputs open.

There are statistical and physical limits as well. The estimator has a finite second moment, which supports the stated error bound, but its fourth moment can be infinite, so an observed sample variance is not necessarily a reliable diagnostic. The main model assumes a specified fixed-photon-number input with one photon per site and a known state representation. The paper discusses extensions to vacuum emissions, certain known losses and locally bounded photon-number support, but unresolved loss and realistic source imperfections are not generally captured by the basic model. The work is a theoretical preprint, not evidence that a photonic device has demonstrated a computational advantage or that the estimates are experimentally practical in every regime.