Statistical Interference Sampling

• Statistical interference sampling is a technique that infers many-body quantum interference from low-order correlators rather than requiring full distribution reconstruction.
• The method distinguishes between indistinguishable bosons and distinguishable particles by analyzing features like normalized mean, coefficient of variation, and skewness of output correlator distributions.
• Experimental validations in integrated photonics demonstrate the approach’s scalability and efficiency in certifying quantum interference without the need for complete distribution tomography.

Searching arXiv for the primary paper and closely related work on many-body interference, Boson Sampling certification, and interference-aware sampling complexity. Statistical interference sampling denotes an approach to inferring the presence of many-body interference from statistical fingerprints in output samples, rather than from reconstruction of a full many-body distribution. In the setting emphasized by "Experimental statistical signature of many-body quantum interference" (Giordani et al., 2021), the method is a certification or benchmarking protocol for Boson Sampling-like linear-optical devices: one measures output events, computes low-order correlators, summarizes their empirical distribution by a small number of statistics, and then performs a hypothesis test between indistinguishable-boson and distinguishable-particle models. The central claim is that genuine many-body quantum interference can be witnessed from the global shape of a correlation dataset, without reconstructing all high-order correlations or the full output distribution (Giordani et al., 2021).

1. Definition and conceptual scope

In the usage most directly associated with the term, the relevant physical setting is the standard Boson Sampling architecture: $n$ identical photons are injected into an $m$-mode linear interferometer, and the output distribution depends on multiparticle interference among all many-path amplitudes (Giordani et al., 2021). For indistinguishable bosons, the output probabilities involve coherent sums over many permutation amplitudes; for distinguishable particles, those interference terms disappear. In that sense, genuine many-body quantum interference means that measured output statistics cannot be explained by a classical mixture of independent particles or by partially distinguishable alternatives (Giordani et al., 2021).

This framing is closely connected to the broader theory of identical-particle interference. For distinguishable particles, propagation through a linear multiport is described by classical alternatives and independent single-particle probabilities, whereas for identical bosons and fermions the same output event receives contributions from many indistinguishable many-particle paths, which must be added at the amplitude level before squaring (Tichy, 2013). Boson Sampling is the canonical sampling problem arising from this structure, because its output probabilities are governed by matrix permanents rather than determinants, and the sampling task is therefore not reducible to independent-particle Monte Carlo procedures (Tichy, 2013).

Within that context, statistical interference sampling is not itself a sampler. It is a method for deciding, from observed samples, whether the data bear the statistical signature of coherent many-body transport. A plausible implication is that the method addresses a validation problem created by the exponential growth of the full output space: if complete tomography is impractical, one may still test for interference through efficient statistics of the observed sample.

2. Correlator statistics and the $C$-dataset

The protocol studied experimentally in (Giordani et al., 2021) is built from the two-mode correlators

$$C_{ij}=\langle \hat{n}_i \hat{n}_j \rangle-\langle \hat{n}_i\rangle\langle \hat{n}_j\rangle,$$

where $\hat n_i$ is the number operator at output mode $i$. For a given device instance, one collects the full set of pairwise correlators $\{C_{ij}\}$, called the $C$-dataset, and summarizes it by statistics including normalized mean (NM), coefficient of variation (CV), and skewness (S) (Giordani et al., 2021).

In the three-photon experiment reported in that work, the correlators are estimated experimentally from postselected output events $(k,l,m)$ through

$$\langle \hat{n}_i \hat{n}_j \rangle \simeq \frac{1}{N}\sum_{m\ge l\ge k} n^i_{klm}n^j_{klm}\,\mathcal N_{klm},$$

$m$0

where $m$1 is the number of times the output occupation pattern $m$2 is observed, and $m$3 is the sample size (Giordani et al., 2021).

The physical content of the method lies in the claim that the shape of the distribution of pairwise correlations over all output pairs differs across particle statistics. Indistinguishable bosons, distinguishable particles, mean-field-like states, and fermions therefore form different clouds in the low-dimensional feature spaces generated from the $m$4-dataset (Giordani et al., 2021). NM reflects the average correlation scale, CV the relative spread, and S the asymmetry of the correlation distribution; the additional summary statistics considered in the same study probe tails, concentration, and dispersion (Giordani et al., 2021).

This statistical compression should be distinguished from full-distribution inference. The protocol does not attempt to learn the complete Boson Sampling law over all outcomes. Instead, it leverages the fact that low-order correlator statistics remain sensitive to the coherent structure of many-body interference.

3. Experimental realization in integrated photonics

The experimental demonstration in (Giordani et al., 2021) injects up to three photons into a 7-mode integrated photonic circuit fabricated by femtosecond laser writing. The integrated network implements a random 7-dimensional unitary transformation. Photon pairs and triples are generated via a four-fold parametric down-conversion source, with an additional heralding photon acting as trigger, and output detection uses a cascade of in-fiber beam splitters to approximate photon-number-resolving detection (Giordani et al., 2021).

The setup can switch between indistinguishable photons $m$5, obtained by matching temporal and spectral properties, and distinguishable photons $m$6, obtained by removing interference filters and introducing time delays (Giordani et al., 2021). The detection layer produces $m$7 monitored modes plus a trigger channel, and the output distribution for three photons in seven modes contains

$m$8

possible three-photon configurations, including collisions (Giordani et al., 2021).

The test workflow reported in that experiment consists of five steps: measure the output sample; compute the set of pairwise correlators $m$9; extract summary statistics such as NM, CV, and S; place the resulting point in a low-dimensional feature space; and compare that point with theoretical or simulated statistical clouds corresponding to $C$0 and $C$1 (Giordani et al., 2021). For the small 3-photon, 7-mode experiment, the plane $C$2 was particularly effective, even more so than the original $C$3 plane proposed in the benchmark paper (Giordani et al., 2021).

The same study also reports agreement with the reconstructed theoretical distribution using the total variation distance

$C$4

with measured values around $C$5 and $C$6 for the two displayed input states (Giordani et al., 2021). These values are reported as distribution-level consistency checks, but the interference witness itself is based on the correlator-statistics protocol rather than on full-distribution reconstruction.

4. Statistical decision theory and feature-space classification

The classification problem is formulated explicitly as a hypothesis test between $C$7, that the data come from indistinguishable bosons, and $C$8, that the data come from distinguishable particles (Giordani et al., 2021). The decision procedure uses the distance of the experimental point from centroids estimated using either Haar-random unitary ensembles or random circuits with the same structure as the actual device; the latter improves accuracy because it incorporates the actual interferometer architecture (Giordani et al., 2021).

This low-dimensional statistical decision rule is a marked departure from verification schemes based on full output distributions. It replaces exponentially large hypothesis spaces by a comparison of empirical summary statistics against model-generated clouds in feature space (Giordani et al., 2021). A plausible implication is that the method is best understood as a model-based compression of the verification task: it retains discriminative information about particle statistics while discarding most of the combinatorial detail of the output law.

The relation to Boson Sampling is therefore indirect but central. Boson Sampling derives its computational hardness from the coherent many-photon interference of random linear optics, and the certification problem asks whether an experimental device is actually operating in that regime. Statistical interference sampling answers that question by asking whether the observed sample exhibits the characteristic fingerprints of many-body interference, not by attempting to solve the underlying sampling problem itself (Giordani et al., 2021).

This distinction aligns with the broader literature on many-body interference and sampling complexity. Bosonic output probabilities in linear networks arise from coherent sums over up to $C$9 many-particle paths and are governed by permanents, whereas fermionic counterparts are governed by determinants (Tichy, 2013). Statistical interference sampling targets the empirical consequences of that coherent structure at the level of output-correlation statistics.

5. Machine learning and automatic feature discovery

A major extension reported in (Giordani et al., 2021) is the use of machine learning, specifically random forest classifiers, to identify which statistics of the $$C_{ij}=\langle \hat{n}_i \hat{n}_j \rangle-\langle \hat{n}_i\rangle\langle \hat{n}_j\rangle,$$0-dataset are most informative for distinguishing $$C_{ij}=\langle \hat{n}_i \hat{n}_j \rangle-\langle \hat{n}_i\rangle\langle \hat{n}_j\rangle,$$1 from $$C_{ij}=\langle \hat{n}_i \hat{n}_j \rangle-\langle \hat{n}_i\rangle\langle \hat{n}_j\rangle,$$2. The feature set considered includes NM, CV, S, median, median deviation, interquartile range, kurtosis, harmonic mean, ROC-based measures, and the TVD of the normalized $$C_{ij}=\langle \hat{n}_i \hat{n}_j \rangle-\langle \hat{n}_i\rangle\langle \hat{n}_j\rangle,$$3-dataset from uniformity (Giordani et al., 2021).

Using random forest feature importance, that study finds that NM and CV are the most informative features in the reported experimental regime, while skewness also becomes useful, especially as system size grows (Giordani et al., 2021). The stated significance of this result is methodological: instead of assuming a fixed pair of observables in advance, the machine-learning procedure can uncover near-optimal statistical signatures for a given device size and architecture (Giordani et al., 2021).

This suggests a broader interpretation of statistical interference sampling as a feature-selection problem on experimentally accessible observables. The observables are still physically motivated correlator summaries, but the choice of the most discriminating subset need not be fixed a priori. In that sense, the method combines physically interpretable summary statistics with data-driven model discrimination.

The paper does not recast the protocol as a generic black-box classifier. Its framing remains a benchmark for witnessing many-body interference from easy-to-compute output statistics (Giordani et al., 2021). The machine-learning component refines feature selection within that benchmark rather than replacing its physical basis.

6. Resources, scalability, and relation to broader sampling problems

The resource analysis in (Giordani et al., 2021) emphasizes three points: photon-number resolution is a practical bottleneck; sample size matters, but the protocol is shown to work well with about $$C_{ij}=\langle \hat{n}_i \hat{n}_j \rangle-\langle \hat{n}_i\rangle\langle \hat{n}_j\rangle,$$4 events for the present case; and five output modes would already be sufficient for reliable discrimination in this 3-photon, 7-mode experiment. The same paper states that the method is expected to become more effective at larger $$C_{ij}=\langle \hat{n}_i \hat{n}_j \rangle-\langle \hat{n}_i\rangle\langle \hat{n}_j\rangle,$$5 and $$C_{ij}=\langle \hat{n}_i \hat{n}_j \rangle-\langle \hat{n}_i\rangle\langle \hat{n}_j\rangle,$$6 because the statistical clouds separate more strongly, and that the protocol is provably efficient in $$C_{ij}=\langle \hat{n}_i \hat{n}_j \rangle-\langle \hat{n}_i\rangle\langle \hat{n}_j\rangle,$$7 and $$C_{ij}=\langle \hat{n}_i \hat{n}_j \rangle-\langle \hat{n}_i\rangle\langle \hat{n}_j\rangle,$$8 in the sense of the original benchmark idea, based only on easy-to-compute output statistics and not requiring extra hardware or dynamic reconfiguration of the interferometer (Giordani et al., 2021).

These claims should be read against the complexity of the underlying many-body sampling problems. In Boson Sampling with finite mode density, collisions can reduce the classical cost of exact or sequential sampling because repeated output columns simplify permanent computations; the relevant complexity can be summarized through an effective boson number

$$C_{ij}=\langle \hat{n}_i \hat{n}_j \rangle-\langle \hat{n}_i\rangle\langle \hat{n}_j\rangle,$$9

rather than the nominal $\hat n_i$0 (Shchesnovich, 2019). That result concerns classical simulation complexity rather than certification, but it clarifies why interference witnessing and hardness assessment are distinct tasks: an experiment may exhibit many-body interference while still being easier to simulate classically if the interferometer is too small (Shchesnovich, 2019).

A related distinction appears in high-energy physics. "The Quantum Trellis" (Macaluso et al., 2021) addresses a different interference-aware sampling problem, namely sampling parton shower outcomes when amplitudes from many shower histories interfere. There, interference prevents sequential Markov sampling, and the proposed solution combines exact dynamic-programming evaluation of an un-normalized Born-rule density with MCMC over final-state phase space (Macaluso et al., 2021). This is not a Boson Sampling certification protocol, but it illustrates a broader principle also present in statistical interference sampling: interference destroys naive factorization, and practical methods must either certify its presence through statistical signatures or incorporate it explicitly into density evaluation and sampling.

In that broader landscape, statistical interference sampling occupies the niche of efficient empirical validation. Its aim is to determine whether a device exhibits genuine many-body interference from statistical features of the observed output sample, especially in regimes where reconstructing the full exponentially large distribution is infeasible (Giordani et al., 2021). The experimental demonstration on three photons in a 7-mode integrated circuit is therefore best understood as a proof of principle for interference certification by low-order statistical fingerprints rather than by full-distribution tomography.