On November 18, 2026, Germany's Paderborn University announced research results comparing three boson sampling methods using its photonic quantum experiment platform, the Paderborn Quantum Sampler (PaQS). The research team matched the amount of light produced by the source and the interferometer conditions, then measured photon-number correlations. Two of the methods showed "nonclassicality," meaning correlations that classical light cannot explain. However, how this changed as the light was made brighter depended on the method. For the method that has traditionally been used for scaling up, the fraction of measurement intervals showing nonclassicality decreased. The study shows that the input state of the light and the type of correlation measured are important when evaluating the performance of photonic quantum computation.
Comparing three methods on the same optical circuit

Boson sampling is a quantum computing approach in which multiple photons are sent into an interferometer and the pattern of photon counts detected at each output is measured repeatedly. Because photons interfere according to quantum mechanics, the photon number at each output varies from trial to trial. By collecting many of these output patterns, samples are obtained from the probability distribution created by the optical circuit. Early boson sampling was limited in scale because it was difficult to prepare many single-photon sources of uniform quality. Gaussian boson sampling (GBS), which uses light states that are comparatively easy to generate, was developed as a result.
PaQS sends time-shifted laser pulses into a waveguide to generate light with adjusted quantum fluctuations. The generated pulses are distributed among multiple optical fibers, and the path lengths are tuned so that the pulses arrive at the same time. They are then fed into a 12-mode interferometer in which the way the light mixes can be changed. In this comparison experiment, 8 inputs and 8 outputs were used, and photon numbers were measured with detectors placed at each output. Even when the light source configuration was switched, the interferometer and the output-side detectors remained the same.
In GBS, the input is "single-mode squeezed vacuum," in which the quantum fluctuation of one component of the light is made smaller than the vacuum-state baseline. Scattershot boson sampling (SBS), on the other hand, generates a pair of light beams at the source and measures one of them with a separate detector. Using that measurement result, the other beam can be treated as a non-Gaussian "Fock state" with a definite photon number. The research team tallied the output data for each detection pattern that indicates the input photon number. However, as the required number of photons increases, the probability of obtaining the target input pattern drops rapidly.
The third method, thermal boson sampling (TBS), uses the same measurement records as SBS but tallies them without using the information indicating the input photon number. The statistics obtained this way correspond to inputting thermal-state light into the interferometer, and can be used as a classical-light comparison. To keep long-term drift in the apparatus from affecting the results, the team switched between the GBS and SBS/TBS configurations every 20 seconds and excluded 100 milliseconds before and after each switch from the analysis. The peer-reviewed paper by Michael Stefszky and colleagues, which summarizes the methods and results, was published in the journal PRX Quantum dated November 17, 2026 (Vol. 7, 043033, DOI: 10.1103/kqyv-lvvy).
Same amount of light, different strength of nonclassicality
The research team examined the correlations of photon numbers detected at different outputs. For example, they measured whether, when the photon number rises at one output, it also rises or instead falls at another. These photon-number fluctuations were expressed as a covariance matrix for each pair of outputs, and the mean photon number was subtracted from the diagonal elements, which correspond to the variance at each output. For classical light, the smallest eigenvalue of this matrix is zero or greater. If it is negative even after accounting for measurement error, that is evidence of nonclassicality that classical light cannot explain. With this method, nonclassicality can be evaluated from output photon numbers without knowing the interferometer settings or the details of the input state in advance.
In the comparison, clear differences appeared in the smallest eigenvalues of the three methods even at the same amount of light. Under the condition where the mean photon number at the source was 0.569 ± 0.011 per input mode, the smallest eigenvalue of SBS swung further to the negative side than that of GBS. For TBS, on the other hand, all observed smallest eigenvalues were positive. The team corrected the measured values for each input mode for optical loss to calculate the mean photon number at generation. The results shown in Figure 4 and Tables III, V, and VI of the paper are as follows.
| Method | Lowest observed smallest eigenvalue | Uncertainty from counting |
|---|---|---|
| GBS | −0.002118 | ±0.000148 |
| SBS | −0.003749 | ±0.000225 |
| TBS | +0.001023 | ±0.000423 |
Measured on the same apparatus. Mean photon number at generation was 0.569 ± 0.011 per input mode, with a squeezing parameter of 0.697. For each method, the value shown is from the 1-second measurement interval, each containing 1 million samples, in which the smallest eigenvalue was lowest.
Both GBS and SBS showed negative values well beyond the uncertainty from counting. In particular, for SBS, the smallest eigenvalue was lower than for GBS under the same generated light level, indicating that nonclassicality appeared more strongly in the second-order photon-number correlations. However, this comparison shows a difference in an indicator based on second-order correlations; it is not a comparison of computation speed or of quantum computing capability itself.
Brighter light makes GBS nonclassicality harder to detect
What is especially interesting in this study is that GBS and SBS showed different trends when the amount of light was increased. For GBS, at a mean generated photon number of 0.569 ± 0.011 per input mode, more than 80% of the 1-second measurement intervals had a negative smallest eigenvalue. However, when the mean photon number was increased to 2.152 ± 0.090, that fraction fell below 2%. This fraction uses as its denominator the total number of 1-second GBS intervals obtained at each light level, and the intervals showing negative values include some whose difference from zero is not clear once counting error is considered.
SBS, on the other hand, showed the opposite trend. Under the same two light conditions, the lowest smallest-eigenvalue for each condition dropped from −0.003749 ± 0.000225 to −0.027048 ± 0.000708. In other words, the more the light was increased, the further the indicator of nonclassicality based on second-order correlations swung to the negative side. These values correspond, respectively, to the squeezing parameters of 0.697 and 1.176 shown in Table V of the paper.
In actual GBS, the phase of the input light drifts over time, so the smallest eigenvalue fluctuates between measurement intervals even at the same amount of light. However, the research team considered that phase fluctuation and optical loss in the circuit alone might not fully explain the trend when the light was increased, and so ran numerical simulations. They modeled the 8×8 optical transformation corresponding to the experiment with 100% transmission, and generated 1 million samples for each of nine settings in which the input phase was fixed within each trial. The upper limit on the number of photons handled in the calculation was 15. The simulations also confirmed a trend in which the second-order-correlation indicator of GBS nonclassicality weakens beyond a certain point as the light increases.
That said, detector performance also bears on the interpretation of the experimental results. The detectors used here can distinguish photons with high accuracy only up to 3 photons per pulse at each output. In a separate numerical check, the team found that counting events with 4 or more photons as 3 can bias the smallest eigenvalue toward the negative side more than the true value for GBS and TBS at high light levels. This effect may also have contributed to the part of Figure 4(b) in which the smallest eigenvalue of high-light GBS swings negative again. The detection limit on photon number not only makes higher-order correlations harder to capture but can also bias measured values further negative than the classical-light boundary.
Technology needed to measure higher-order correlations
The research team points out that when the light in GBS is increased, the quantum entanglement contained in the output may appear in correlations of order higher than second. The nonclassicality determination used here relied on photon-number statistics up to second order and did not evaluate higher-order correlations. Even if the second-order indicator weakens at high light levels, quantum entanglement might still be captured using other statistics. The authors state that measuring higher-order correlations with sufficient precision requires detectors that can resolve more than 4 photons.
Paderborn University also explains that as the number of photons increases and the optical circuit becomes more complex, simulation by classical computers becomes harder. However, this is an explanation that assumes large-scale photonic quantum computing systems. The nonclassicality measured in this paper is one necessary condition for demonstrating quantum computational advantage; it is not a comparison between PaQS's processing time and the runtime of the best classical algorithm. Evaluating actual computation speed would require a comparison under conditions that produce equivalent output.
The research team has published the measurement data for the three methods and a Python loading example on Zenodo. This allows third parties to recompute the nonclassicality determination from the same measurement records. As of November 20, 2026, no replication by an independent research team had been confirmed. If other apparatuses can reproduce the weakening of the second-order-correlation indicator at high light levels, and if higher-order correlations can be measured at the same time, it will be possible to investigate in more detail which statistics reveal quantum entanglement as the amount of light increases.
