Estimates of when photonic devices outgrow classical simulation are hard to compare: even for single-photon boson sampling, quoted thresholds range from 50 to 98 photons, as they rest on different tasks and assumptions. We propose as a common unit the time a classical computer needs to draw one sample from the device's exact output distribution, up to floating-point rounding. Only exactness applies equally to every kind of light, since the cost of approximate sampling depends on the accuracy required. The simulability frontier is the device size at which our fastest exact sampler produces one sample per second on a fixed machine; this device size increases with better hardware and algorithms. As the time per sample grows exponentially with the number of detected photons, doubling the machine speed moves the frontier by about one photon, whereas a better algorithm changes the growth rate. We measure the frontier with photonic-engines, an open-source GPU library of exact samplers for single photons, entangled photons from quantum emitters, vacuum-single-photon superpositions and squeezed light. Each sampler is validated against a simulator sharing no code with it. On one NVIDIA H100 GPU, with Haar-random interferometers, twice as many modes as photonic inputs and photon-number-resolving detectors, the frontier lies at 38.3 single photons and at 31.0 entangled photons from one emitter. For squeezed light, a timing model fitted to measurements places it at 41.5 squeezed photonic inputs of mean photon number 0.5. For our samplers, coherence alone makes simulation harder: sources that emit a photon with probability one half reach 64.2 when vacuum and photon are mixed incoherently, as for lossy single photons, but only 50.0 when they are in coherent superposition. Feed-forward from mid-circuit measurement of up to three modes lowers the single-photon frontier only from 38.3 to 38.1.
If classical algorithms have been successful in reproducing the estimation of expectation values of observables of some quantum circuits using off-the-shelf computing resources, matching the performance of the most advanced quantum devices on sampling problems usually requires extreme cost in terms of memory and computing operations, making them accessible to only a handful of supercomputers around the world. In this work, we demonstrate for the first time a classical simulation outperforming Gaussian boson sampling experiments of one hundred modes on established benchmark tests using a single CPU or GPU. Being embarrassingly parallelizable, a small number of CPUs or GPUs allows us to match previous sampling rates that required more than one hundred GPUs. We believe algorithmic and implementation improvements will generalize our tools to photo-counting, single-photon inputs, and pseudo-photon-number-resolving scenarios beyond one thousand modes. Finally, most of the innovations in our tools remain valid for generic probability distributions over binary variables, rendering it potentially applicable to the simulation of qubit-based sampling problems and creating classical surrogates for classical-quantum algorithms.
The strong efforts towards the realization of scalable quantum computing are motivated by the promise of super-polynomial speed-up in several computational tasks. A first fundamental step in this route is to provide concrete evidence of ‘quantum supremacy'. Such a regime is obtained when a quantum device is capable of solving a specific problem faster than the best classical hardware. A serious attempt to provide a first demonstration of quantum speed-up with affordable resources is represented by boson sampling [1,2], a non-universal quantum-computational protocol exploiting indistinguishable single photons and linear optical devices. In this context, strong theoretical evidence has been reported that the interference pattern generated by a boson sampler is hard to simulate with classical resources, this pattern being related to the calculation of a hard-to-compute quantity known as a matrix permanent. Formal achievement of the regime of quantum supremacy will answer two very relevant questions: Are we able to benefit from the power of quantum mechanics or does there exist a fundamental reason why this achievement is forbidden? Once this regime has been achieved, are we able to certify the correct functioning of the quantum boson samplers or do we face an undecidability problem? Starting from these intriguing and promising theoretical results, an experimental race kicked off with the aim of reaching the quantum supremacy regime with boson sampling by employing a photonic platform [3]. A large amount of effort has been devoted to pushing the limits of current photonic systems [1], in order to significantly increase the size (number of photons and optical modes) that can be handled with a quantum device. In parallel to this quest for improved quantum performance, it is fundamental to benchmark the limits of a classical simulation of boson sampling. More specifically, in the race for quantum supremacy, what is the best possible classical simulation that can be performed? Such an analysis would be necessary to define the computational capacity possessed by such a classical agent to perform the task (see Fig. 1). In the race for quantum supremacy, a user who employs the best classical hardware and software, and a second agent who exploits a quantum device compete to solve the boson sampling task faster. In a recent article, Wu et al. [4] have provided a fundamental step to define and push the limits of a classical simulation of boson sampling. The authors have employed Tianhe-2, one of the world's most powerful super-computers, to perform such a task. More specifically, by using such high-performance computing hardware and by employing the fastest algorithm to date, Wu et al. have shown that calculating the permanent of a |
50×50| matrix can be performed in 100 min. Recent results [5] have shown that this computation corresponds to the classical hardness of drawing a 50-photon sample from a boson sampling distribution. The strong computational capacity of Tianhe-2 also allowed us to address the precision that the known fastest classical algorithms can reach in such a large-scale calculation. This latter aspect is indeed of fundamental importance, since any reliable classical computation must be accompanied by an appropriate error analysis. The study reported in Ref. [4] represents by far the largest calculation to date related to the classical simulation of boson sampling, whose relevance can be found in two different but related aspects. First, it raises the bar for any quantum implementation aimed at reaching quantum advantage, and provides a fundamental benchmark to which any photonic hardware must compare. This result suggests that a larger effort must be dedicated to pushing quantum technologies in this direction. Second, improved classical simulation also has important consequences in the context of boson sampling certification, i.e. the verification that a quantum device is correctly performing the required task. Indeed, first validation procedures in the quantum advantage regime must be assisted by powerful classical computational resources that, while requiring longer processing time than quantum hardware, will necessarily have to perform complex calculations.
Most proposed applications of boson samplers, including photonic machine-learning models, do not use the samples themselves but estimates of output probabilities, marginals or overlaps, obtained to additive precision from a finite number of shots. For product inputs such as single photons, these quantities can already be estimated classically with Gurvits's algorithm and its generalisations. We ask whether entangled inputs change this picture. For any
n-photon input that is a matrix-product state of bond dimension
χ, with one photon per site and
L modes per site, we give an unbiased estimator of any output amplitude of an
m-mode interferometer, whose mean-square error after
N draws is at most
∥ψ∥2/N whatever the interferometer, the output pattern and the entanglement of the state. Each draw costs
O(nLχ2+nL3+mn) operations after
O(nLχ3) preprocessing, and an additive error
ε with probability at least
1−δ requires
O(∥ψ∥2ε−2log(1/δ)) draws, with memory of the order of the input size. The main application is to light emitted sequentially by quantum emitters, such as spin-based sources of photonic cluster states, for which
χ is bounded by the dimension of the source's retained quantum memory: one spin gives
χ=2 at any length, and a memory of
k qubits, held in emitters or in photons stored in a delay line, gives
χ≤2k, so the cost stays polynomial as long as
k=O(logn). The same sampler also estimates traced-memory probabilities, overlaps and individual marginal probabilities. All guarantees are additive: the estimator gives neither relative estimates of exponentially small probabilities nor a sampler of the output distribution.
Photon loss makes Gaussian boson sampling easier to simulate classically. A standard approach replaces the lossy input by classical light, and it stays accurate while the mean number of surviving photons grows no faster than the square root of the number
N of squeezed inputs at fixed squeezing and accuracy. We prove that this scale is a limit of replacement by classical light. For explicit interferometers, no state with a nonnegative Glauber--Sudarshan
P representation, Gaussian or not and however correlated, keeps the error in the photon count distribution small beyond this scale. We then surpass this limit by approximating the detection instead of the input. The sampler evolves the state exactly between detections and, after each detection, replaces the conditional state by a Gaussian with the same covariance. For every interferometer its total variation error is
O(Nη3) at polynomial expected cost, where
η is the photon survival probability. The mean surviving photon number can therefore grow as
N2/3 at fixed accuracy. A separate pair construction reaches
N3/4 when every input mode is equally squeezed.
Quantum computing experiments are moving into a new realm of increasing size and complexity, with the short-term goal of demonstrating an advantage over classical computers. Boson sampling is a promising platform for such a goal; however, the number of detected single photons is up to five so far, limiting these small-scale implementations to a proof-of-principle stage. Here, we develop solid-state sources of highly efficient, pure, and indistinguishable single photons and 3D integration of ultralow-loss optical circuits. We perform experiments with 20 pure single photons fed into a 60-mode interferometer. In the output, we detect up to 14 photons and sample over Hilbert spaces with a size up to 3.7×10^{14}, over 10 orders of magnitude larger than all previous experiments, which for the first time enters into a genuine sampling regime where it becomes impossible to exhaust all possible output combinations. The results are validated against distinguishable samplers and uniform samplers with a confidence level of 99.9%.
Sampling from probability distributions of quantum circuits is a fundamentally and practically important task which can be used to demonstrate quantum supremacy using noisy intermediate-scale quantum devices. In the present work, we examine classical simulability of sampling from the output photon-number distribution of linear-optical circuits composed of random beam splitters with equally distributed squeezed vacuum states and single-photon states input. We provide efficient classical algorithms to simulate linear-optical random circuits and show that the algorithms' error is exponentially small up to a depth less than quadratic in the distance between sources using a classical random walk behavior of random linear-optical circuits. Notably, the average-case depth allowing an efficient classical simulation is larger than the worst-case depth limit, which is linear in the distance. Besides, our results together with the hardness of boson sampling give a lower-bound on the depth for constituting global Haar-random unitary circuits.