# Techniques: Error-mitigated photonic quantum circuit Born machine

## What the paper contributes (one paragraph)
The paper studies quantum circuit Born machines (QCBMs), a generative quantum machine learning model, in the setting of discrete-variable photonic quantum computing. It proposes a QCBM ansatz built from a universal linear optical interferometer and trained as a variational algorithm against a target distribution. Its main result is that recycling mitigation, an error mitigation technique for uniform photon loss, greatly improves QCBM training in realistic high-loss regimes, turning otherwise untrainable lossy QCBMs into trainable ones. This is demonstrated through Perceval simulations on Gaussian and financial datasets and through a proof-of-concept experiment on the Altair photonic processor. In the supplemental material the authors also prove, via boson sampling statistics, that the modulus squared of a matrix permanent can be estimated to smaller additive error than Gurvits' classical algorithm for the same runtime, and discuss classical hardness and quantum learning supremacy.

## Techniques offered
- **Recycling mitigation (quantum error mitigation for photon loss)** - reuses lossy photonic output statistics normally discarded in post-selection to produce mitigated estimates of the ideal lossless output probabilities, supplying a loss-tolerant error mitigation primitive for near-term linear optical circuits.
  - guarantee: statistical bound; the combined bias and statistical error is provably at most the post-selection statistical error up to a very large number of samples, for a generic interferometer.
  - quote: "This technique makes use of photonic output states that would normally be discarded in post-selection, and provably outperforms post-selection up to large sample sizes."
- **Photonic quantum circuit Born machine (linear-optical QCBM ansatz)** - a generative quantum machine learning model implemented on a parametrised universal interferometer with Fock-state input and threshold detection, supplying a photonic-tailored sampling and probability-estimation model that can be trained variationally.
  - guarantee: none (no proof of training success; classical hardness only conjectured for sufficiently large n, m and limited loss; quantum learning supremacy left open).
  - quote: "QCBMs are a popular choice of quantum generative machine learning models, and we present a QCBM designed for linear optics."
- **Permanent-modulus-squared estimation from boson sampling statistics** - estimates the squared modulus of a matrix permanent from i.i.d. boson sampler runs, supplying a sampling-based estimator that beats Gurvits' classical algorithm in additive error for equal runtime.
  - guarantee: statistical bound; with high probability O(1/epsilon^2) samples give an estimate within additive error epsilon||A||^{2n}, requiring runtime O(n^2 t) for Gurvits to match O(t) boson sampler runs.
  - quote: "We show that computing $|\mathsf{Per}(A)|^2$ from statistics collected from $t$ i.i.d runs of a boson sampler produces, with high probability, a more accurate estimate of $|\mathsf{Per}(A)|^2$ than Gurvits which is run for time $O(t)$."

## Where it could apply
- Target primitive(s): sampling, generative modelling, probability estimation (permanent estimation)
- Target application group(s): quantum_ml, platforms_tooling, finance
- Code family / hardware assumptions: photonic (discrete-variable linear optics, single-photon sources, threshold detectors); no quantum error correction code family; assumes uniform photon loss as the dominant noise

## Caveats
Recycling mitigation's formal performance guarantees from the cited prior work hold for no-collision outputs (at most one photon per mode) and for the lossless permanent-estimation theorems; the paper applies it empirically in the comparable m and n regime without re-proving the bound there, and it does not address photon distinguishability, which needs a different mitigation. The QCBM itself carries no trainability or advantage guarantee. Recycling mitigation is defined and proven in a separate reference (MezherMills2024); this paper supplies its application and integration rather than the underlying proof.