A fast and frugal Gaussian Boson Sampling emulator

Tom Dodd, Javier Martínez-Cifuentes, Oliver Thomson Brown, Nicolás Quesada, Raúl García-Patrón · 2025

Read A fast and frugal Gaussian Boson Sampling emulator by Tom Dodd, Javier Martínez-Cifuentes, Oliver Thomson Brown, Nicolás Quesada, Raúl García-Patrón (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-14923.txt.

Introduction

Tom Dodd, Javier Martínez-Cifuentes, and three co-authors show that a modest classical computer can now mimic a flagship quantum sampling experiment. Gaussian boson sampling sends photons through an optical network and records the patterns, a task long used to claim quantum advantage. Matching the sampling rate of the hundred-mode experiments was thought to need supercomputers.

Their emulator beats those experiments on standard benchmark tests while running on a single processor or graphics card. The method parallelizes cleanly, so a handful of machines match rates that once took more than a hundred graphics cards. Results like this keep raising the bar that quantum hardware must clear to claim a real edge.

Catalog record

CREATORTom Dodd, Javier Martínez-Cifuentes, Oliver Thomson Brown, Nicolás Quesada, Raúl García-Patrón
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE96 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-14923

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A fast and frugal Gaussian Boson Sampling emulator Tom Dodd, Javier Martínez-Cifuentes, Oliver Thomson Brown, Nicolás Quesada, Raúl García-Patrón Abstract 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 last six years have seen an increasing number of implementations of quantum algorithms claiming to have reached quantum computational advantage over their classical counterparts; these range from random circuit sampling on quantum superconducting devices [ 2 , 14 ] , Gaussian or single-photon boson sampling on photonic hardware [ 41 , 39 , 24 , 8 , 23 ] and atomic ensembles [ 37 ] , respectively, or the estimation of dynamical observables in trapped ions [ 17 ] and superconducting circuits [ 21 ] . On hypothetical error-free quantum hardware, all the candidate proposals are believed to be out of reach for classical computers, challenging today’s most advanced supercomputers with sizes beyond 50 qubits or hundreds of modes and photons. For sampling problems, this is often backed by complexity-theoretic arguments [ 5 , 1 , 10 , 15 , 4 ] . Figure 1 : (1) An instance of GBS is characterized by its ideal optical circuit, or equivalently the covariance matrix of its associated Gaussian state; (2) An experiment implements a version affected by losses and other imperfections; (3) One can extract the ground truth of the experiment from tomographic data or applying a good model of imperfections to the noiseless GBS instance; (4) 10 million samples are generated by quantum hardware in approximately 10 seconds; (5) The samples are benchmarked using cumulant statistical analysis and XEB; (6) The first stage of our algorithm computes a list of \binom{144}{5} correlators that are later transformed into cumulants from the ground truth covariance matrix, as detailed in section I.1 ; (7) The second stage generates samples from the corresponding cumulants as detailed in section I.2 ; (8) Using 8 GPUs, we can generate 10 million samples in 115 minutes, later validated with the same benchmarking tests. In reality, current quantum devices are prone to errors and imperfections that limit their performance. Unless fault-tolerance based on quantum error-correction is implemented, the quality of the computations decays rapidly. The pervasive presence of errors is the key enabler behind all classical algorithm implementations that claim to have successfully disproved previous claims of quantum advantage, such as Refs. [ 31 , 30 ] for random quantum circuits, Ref. [ 28 ] for Gaussian boson sampling (GBS), and Ref. [ 20 , 33 , 3 ] for the estimation of observables of complex spin-system dynamics. If classical algorithms have been successful in reproducing the estimation of expectation values of observables of some recent quantum experiments, even using off-the-shelf computing resources [ 20 , 33 , 3 ] , the simulation of sampling tasks, both qubit- and bosonic-based, are known to pose significantly harder challenges to classical simulations. The implementations of these algorithms required an extreme cost in terms of memory and computing operations, making them accessible to only a handful of supercomputers around the world and unreasonable to reproduce on a regular basis. To date, there has been no demonstration of a classical algorithm that matches the performance on standard benchmarking tests of the most advanced quantum devices on a sampling problem using off-the-shelf computational resources. In this work we demonstrate for the first time a classical simulation of imperfect GBS matching and sometimes outperforming the actual quantum experiments in a series of standardized tests, with the need of only a single CPU or GPU, while Ref. [ 28 ] required more than one hundred GPUs. Additionally, our implementation for 144 modes requires 2 GB to encode the quantum state before the measurement, three orders of magnitude less than in Ref. [ 28 ] . Being embarrassingly parallelizable, we were able to generate 10 million samples in 115 minutes using 8 contemporary GPUs on one single cluster node. These significant improvements in memory use, running time and energy costs makes mainstream what was before accessible only to a handful of facilities in the world. Our main technical contribution is the design of an algorithm to sample from the joint probabilities of M bit samples by exploiting a decomposition in terms of cumulants of increasing order and approximated marginals of the original distribution. Truncating the expression to order K allows us to capture all relevant statistics of the hardware experiment while making its running time scale as O(M^{K}) . The algorithm benefiting from a polynomial scaling of order K , should remain a contender against future hardware improvements and larger system sizes, such as the recent Jiuzhang 4.0 experiments [ 23 ] . We believe the algorithm is easily generalizable to photo-counting and single-photon scenarios. Finally, most of the innovations in our tools remain valid for generic probability distributions over binary variables, rendering it applicable to the simulation of qubit-based sampling problems and creating classical surrogates for hybrid classical-quantum algorithms. I Emulator description Error-free GBS is believed to be a hard sampling problem where an instance over M modes is defined by a set of squeezed state sources located at a given set of input ports and a linear-optical interferometer, characterized by an M\times M unitary matrix U selected randomly over the Haar measure. Except for the Borealis experiment [ 24 ] , where photo-counting was implemented at the output, all other GBS demonstrations so far have implemented threshold detection [ 41 , 39 , 8 ] . In this work we focus on the second type of GBS, consisting of binary outcomes encoding the absence or presence of photons. Historically, the ground truth of the GBS experiment used in its validation tests is the description of the Gaussian state at the output of the real interferometer taking into account the presence of losses. This approach is different from the one used in, for example, random circuit sampling, where one computes benchmarks such as the cross linear entropy with respect to the ideal noiseless instance. Whether it is more appropriate to use a lossy ground truth for photonic quantum advantage remains an open research problem [ 25 ] . With our goal being to design an emulator of noisy GBS, it is natural and non-controversial for us to consider the lossy ground truth as the benchmark

Read the complete text