Canonical Boson Sampling Explained

• Canonical Boson Sampling is a quantum model where n indistinguishable photons are injected into an m-mode interferometer, producing output amplitudes defined by matrix permanents.
• Its methodology leverages Haar-random unitary matrices and collision-free regimes to ensure computational hardness, underpinning the quantum advantage argument.
• CBS serves as the benchmark for various generalized boson-sampling architectures and experimental platforms, including photonic, atomic, and trapped-ion realizations.

Canonical Boson Sampling (CBS) is the Aaronson–Arkhipov boson-sampling model in its fixed-input form: $n$ indistinguishable single photons are injected into specified modes of an $m$-mode passive linear interferometer, the evolution is governed by a unitary matrix, and the output is sampled in the photon-number basis. Its defining feature is that output amplitudes are given by matrix permanents, placing the problem at the intersection of linear optics, many-boson interference, and complexity theory. In the literature, CBS functions both as the canonical hardness benchmark for photonic quantum advantage and as the reference point from which scattershot, Gaussian, lossy, atomic, and other generalized boson-sampling architectures are defined (Liu et al., 2016, Neville et al., 2017).

1. Formal model and probabilistic structure

In the standard CBS setting, one considers an $N$-mode lossless interferometer represented by a unitary matrix $\Lambda \in U(N)$, together with an input Fock state

$$S_B = |s_1,s_2,\ldots,s_N\rangle = \prod_{j=1}^N \frac{(\hat a_j^\dagger)^{s_j}}{\sqrt{s_j!}}|0\rangle, \qquad \sum_{j=1}^N s_j = n,$$

where the canonical case is a fixed pattern of $n$ single photons, typically one photon in each of $n$ distinct input modes. In the Heisenberg picture the mode operators transform as

$$\hat{\mathbb S}\,\hat a_j^\dagger\,\hat{\mathbb S}^\dagger = \sum_{k=1}^N \Lambda_{k,j}\,\hat a_k^\dagger.$$

At the output, photon-number measurements yield a pattern

$$T_B = |t_1,t_2,\ldots,t_N\rangle,\qquad \sum_i t_i = n.$$

The corresponding probability is

$$P_B(T_B; S_B, \Lambda) = \frac{\left|\operatorname{Permanent}(A)\right|^2}{\prod_{i=1}^N s_i! t_i!},$$

where $m$0 is the $m$1 submatrix obtained by repeating columns and rows according to the input and output occupations. In the collision-free sector, with $m$2, this reduces to the familiar permanent-squared law for submatrices of the interferometer (Liu et al., 2016, Shen et al., 2013).

The canonical hardness setting further assumes that the interferometer is Haar-random and that the number of modes is sufficiently large, typically with $m$3 polynomially larger than $m$4 and often $m$5, so that collisions are rare and relevant submatrices are distributed similarly to i.i.d. complex Gaussian matrices. This is the regime in which CBS is usually identified with the original Aaronson–Arkhipov proposal (Bentivegna et al., 2015, Neville et al., 2017).

2. Complexity-theoretic status and supremacy thresholds

CBS is widely regarded as classically intractable because the relevant amplitudes are governed by matrix permanents, and computing the permanent of an arbitrary matrix is $m$6-hard. In the Aaronson–Arkhipov framework, the hardness of approximate sampling from the CBS distribution is tied to average-case conjectures for permanents of random Gaussian matrices and to anti-concentration assumptions; under these assumptions, an efficient classical approximate sampler would imply a collapse of the polynomial hierarchy (Bentivegna et al., 2015, Neville et al., 2017).

This complexity-theoretic picture is robust but not free of controversy when translated into experimental thresholds. One line of work emphasizes that boson sampling is “a promising witness of the supremacy of quantum systems,” and that hardness results have been extended to “scattershot Boson Sampling, approximate and lossy sampling under some reasonable constraints,” while also arguing for “a new threshold that is less error sensitive and experimentally more feasible” than earlier estimates (Latmiral, 2015). A contrasting line of work argues that “near-term quantum supremacy via boson sampling is unlikely,” presenting classical algorithms based on Metropolised independence sampling and reporting that the boson sampling problem was solved for 30 photons with standard computing hardware, while concluding that photon loss implies that demonstrating quantum supremacy by boson sampling “would require a step change in technology” (Neville et al., 2017).

Accordingly, CBS occupies a dual position. As an asymptotic sampling problem, it remains a paradigmatic candidate for post-classical computation. As an experimental target, its feasibility depends sensitively on source quality, optical loss, indistinguishability, and the evolving state of classical simulation algorithms (Latmiral, 2015, Neville et al., 2017).

3. Relation to generalized boson-sampling models

Several prominent generalizations are defined relative to CBS rather than independently of it. Scattershot Boson Sampling replaces the fixed set of occupied inputs by a heralded random subset of inputs generated from many parametric down-conversion sources. If each source emits with probability $m$7, then $m$8 simultaneous heralds occur with probability $m$9, which for $N$0 gives “an exponential improvement in generation rate” over fixed-input boson sampling with $N$1 sources. Conditional on a particular herald pattern, however, the device implements precisely a canonical boson-sampling instance for that input set (Bentivegna et al., 2015).

A second extension uses Gaussian resources. In “Boson Sampling from Gaussian States,” $N$2 two-mode squeezed-vacuum states are used so that one half of each pair is measured and the other enters the interferometer. Conditioning on a herald pattern with exactly $N$3 single photons projects the interferometer input onto the usual CBS Fock state; without adaption, the device samples over all such input patterns, thereby defining a generalized boson-sampling problem that contains CBS as a subset (Lund et al., 2013).

Further extensions preserve the permanent structure while enlarging the sample space. The multi-boson correlation sampling framework samples not only over output ports but also over joint detection times, and in “Sampling of bosonic qubits” single-photon sources carrying polarization qubits are used so that one samples over output ports, detection times, and multi-qubit output states. In that setting the relevant probabilities are still mod-squared permanents, now of matrices that include temporal and polarization overlap factors; CBS is recovered by suppressing time and polarization resolution (Tamma, 2015).

4. Certification, verification, and coarse-grained validation

The verification problem for CBS is structurally difficult because the very distributions of interest are expensive to compute classically. This has produced a large literature on restricted certification tasks rather than full verification. A central observation is that bosonic bunching alone is insufficient: certification schemes based only on bunching can fail against the mean-field sampler, which reproduces bunching tendencies while remaining efficiently classically simulable (Liu et al., 2016).

One proposed remedy is to probe the same device with a specific test input,

$N$4

that is, one photon in every input mode. For the mean-field sampler this yields a multinomial distribution

$N$5

which is independent of the interferometer, whereas genuine boson sampling retains a strongly $N$6-dependent permanent structure. This difference survives moderate partial distinguishability in the numerical analyses reported in the paper (Liu et al., 2016).

A complementary approach uses coarse-grained measurements. “Certification of Boson Sampling Devices with Coarse-Grained Measurements” groups Fock outcomes into “bubbles” defined by an $N$7 distance on occupation vectors and applies a two-sample $N$8 test to the resulting coarse-grained distributions. In the reported simulations, about $N$9 events were sufficient even when the Hilbert-space dimension was $\Lambda \in U(N)$0, with sample size below $\Lambda \in U(N)$1 of $\Lambda \in U(N)$2; the method was used to certify equivalence between boson-sampling devices and to rule out distinguishable-particle and uniform samplers (Wang et al., 2016).

Large photonic experiments have combined such ideas with model comparison. In the 20-photon, 60-mode experiment, the data were validated against distinguishable and uniform samplers using a Bayesian test and the row-norm estimator, with reported confidence level $\Lambda \in U(N)$3 (Wang et al., 2019).

5. Experimental realizations

CBS originated in linear optics, and photonic experiments remain its most direct embodiment. Early scattershot demonstrations coupled six different photon-pair sources to integrated photonic circuits, implementing the same permanent-based sampling structure as CBS while alleviating the fixed-input source bottleneck (Bentivegna et al., 2015). A more advanced photonic realization reported “20 input photons in 60-mode interferometers at $\Lambda \in U(N)$4 state spaces,” with a 60-mode interferometer reconstructed as unitary and Haar-like, strict Aaronson–Arkhipov operation for up to 10 detected photons, and lossy Aaronson–Brod operation for larger input sizes (Wang et al., 2019).

That experiment quantified the bosonic output space by

$\Lambda \in U(N)$5

and, for 20 input photons with 14 detected, reported a Hilbert-space dimension of approximately $\Lambda \in U(N)$6. Two-photon tests on the full 60-mode device yielded average fidelities $\Lambda \in U(N)$7–$\Lambda \in U(N)$8 and average total variation distance $\Lambda \in U(N)$9–$$S_B = |s_1,s_2,\ldots,s_N\rangle = \prod_{j=1}^N \frac{(\hat a_j^\dagger)^{s_j}}{\sqrt{s_j!}}|0\rangle, \qquad \sum_{j=1}^N s_j = n,$$0, while the higher-photon validation rejected distinguishable and uniform samplers at $$S_B = |s_1,s_2,\ldots,s_N\rangle = \prod_{j=1}^N \frac{(\hat a_j^\dagger)^{s_j}}{\sqrt{s_j!}}|0\rangle, \qquad \sum_{j=1}^N s_j = n,$$1 confidence (Wang et al., 2019).

CBS has also been realized outside photonics. “An atomic boson sampler” implements boson sampling with ultracold $$S_B = |s_1,s_2,\ldots,s_N\rangle = \prod_{j=1}^N \frac{(\hat a_j^\dagger)^{s_j}}{\sqrt{s_j!}}|0\rangle, \qquad \sum_{j=1}^N s_j = n,$$2 atoms in a two-dimensional tunnel-coupled optical lattice, where lattice sites are modes and single-particle dynamics generate the unitary $$S_B = |s_1,s_2,\ldots,s_N\rangle = \prod_{j=1}^N \frac{(\hat a_j^\dagger)^{s_j}}{\sqrt{s_j!}}|0\rangle, \qquad \sum_{j=1}^N s_j = n,$$3. The reported system achieved nearly deterministic preparation of up to 180 atoms distributed among $$S_B = |s_1,s_2,\ldots,s_N\rangle = \prod_{j=1}^N \frac{(\hat a_j^\dagger)^{s_j}}{\sqrt{s_j!}}|0\rangle, \qquad \sum_{j=1}^N s_j = n,$$4 sites, indistinguishability $$S_B = |s_1,s_2,\ldots,s_N\rangle = \prod_{j=1}^N \frac{(\hat a_j^\dagger)^{s_j}}{\sqrt{s_j!}}|0\rangle, \qquad \sum_{j=1}^N s_j = n,$$5, loss $$S_B = |s_1,s_2,\ldots,s_N\rangle = \prod_{j=1}^N \frac{(\hat a_j^\dagger)^{s_j}}{\sqrt{s_j!}}|0\rangle, \qquad \sum_{j=1}^N s_j = n,$$6 independent of evolution time, and detection fidelity typically $$S_B = |s_1,s_2,\ldots,s_N\rangle = \prod_{j=1}^N \frac{(\hat a_j^\dagger)^{s_j}}{\sqrt{s_j!}}|0\rangle, \qquad \sum_{j=1}^N s_j = n,$$7. Direct verification of the full distribution was deemed infeasible, so the experiment relied on targeted tests of indistinguishability, unitary calibration, and bosonic bunching (Young et al., 2023).

6. Alternative platforms and theoretical reformulations

CBS is not tied to photons. A trapped-ion proposal encodes bosonic modes in local transverse phonon modes of an ion chain, with Coulomb-induced hopping supplying passive linear mode mixing. In that scheme, the same canonical probability formula

$$S_B = |s_1,s_2,\ldots,s_N\rangle = \prod_{j=1}^N \frac{(\hat a_j^\dagger)^{s_j}}{\sqrt{s_j!}}|0\rangle, \qquad \sum_{j=1}^N s_j = n,$$8

is implemented by deterministic Fock-state preparation, universal mode mixing, and high-efficiency readout, and the authors argue that tens of bosons are feasible with state-of-the-art trapped-ion technology (Shen et al., 2013).

At a more formal level, CBS has also been reformulated as a many-body scattering problem generated by a linear canonical transformation. In this view, the single-particle transformation

$$S_B = |s_1,s_2,\ldots,s_N\rangle = \prod_{j=1}^N \frac{(\hat a_j^\dagger)^{s_j}}{\sqrt{s_j!}}|0\rangle, \qquad \sum_{j=1}^N s_j = n,$$9

induces Fock-space amplitudes

$n$0

while coherent-state, quadrature, generating-function, and semiclassical representations expose different aspects of the same permanent structure. This formulation makes explicit that CBS is the Fock-space manifestation of a linear canonical map and that quantum interference can be interpreted semiclassically as a sum over classical branches (Engl et al., 2015).

Phase-space methods extend this perspective to noisy and lossy settings. “Simulating and assessing boson sampling experiments with phase-space representations” develops complex $n$1-representation and qudit complex $n$2-representation techniques that recover the canonical coincidence formula

$n$3

and then use stochastic phase-space sampling to estimate permanent-squared observables and combined correlation benchmarks with sampling errors below corresponding experimental correlation errors for the same sample budget (Opanchuk et al., 2018).

7. Marginals, coarse observables, and derived computational tasks

Although the full CBS distribution is hard, certain reduced observables are efficiently computable. A recent analytic treatment of one-mode marginals derives the exact single-mode distribution directly from the permanent formula and shows that the 1-mode marginal can be computed in $n$4 time. In that derivation, multiphoton interference reduces to an elementary symmetric polynomial in the single-photon probabilities $n$5, multiplied by a factorial bosonic bunching factor; the method bypasses characteristic-function interpolation and Fourier transforms and leads to scalable bunching metrics compatible with standard threshold detectors (Liu, 1 Mar 2026).

This line of work clarifies that the computational hardness of CBS is concentrated in global many-mode correlations rather than in fixed low-order marginals. It also connects directly to certification: vacuum and low-occupancy one-mode probabilities carry robust signatures of bunching and can separate indistinguishable-boson behavior from distinguishable-particle models in regimes where full-distribution verification is out of reach (Liu, 1 Mar 2026).

CBS has also been used as a primitive for classical decision and function problems. One proposal defines a coarse-grained distribution over bins of Fock outcomes and introduces a most-probable-bin function $n$6 that maps an input seed to the label of the most probable output bin. Decision and function problems are then built from arrays of such outputs. The notable feature is that, for a boson-sampling device, the number of samples required to identify the most probable bin is independent of the Hilbert-space dimension, whereas brute-force classical evaluation remains tied to exponentially many permanent computations. The authors present this as a possible route toward cryptographic applications of boson sampling (Nikolopoulos et al., 2016).

Canonical Boson Sampling therefore remains simultaneously a precise mathematical model, an experimental benchmark, and a source of broader methodological developments. Its enduring significance lies in the fact that a minimal physical architecture—fixed Fock inputs, passive linear evolution, and photon-number measurement—generates distributions whose exact structure is controlled by permanents, whose approximate sampling is believed to be classically intractable, and whose reduced observables, variants, and certification tasks continue to define a substantial research program across photonics, atomic physics, trapped ions, and complexity theory.