The 2019 Sycamore random circuit sampling experiment and its critics

Short answer: In 2019 a Nature paper from Google's team reported sampling a random quantum circuit on 53 superconducting qubits and claimed this was quantum supremacy for that task. The authors estimated that a classical supercomputer would need about 10,000 years for the equivalent task. Later papers and commentary dispute how far the result demonstrates a quantum advantage. This post describes the claim and the stated criticisms, and takes no side.

The claim

The 2019 Nature abstract says the team used a processor with programmable superconducting qubits to create quantum states on 53 qubits, a computational state space of dimension 2 to the power 53, about 10 to the 16. Measurements from repeated experiments sample the resulting probability distribution, which they verify using classical simulations. They say the Sycamore processor takes about 200 seconds to sample one instance of a circuit a million times, while their benchmarks indicate the equivalent task for a state-of-the-art classical supercomputer would take approximately 10,000 years. They call this an experimental realization of quantum supremacy for this specific computational task.

The paper explains that random circuits suit benchmarking because they have no structure and so allow limited guarantees of computational hardness. The circuits entangle qubits by repeated single-qubit and two-qubit operations.

A philosophical critique

A 2020 arXiv paper titled "What Have Google's Random Quantum Circuit Simulation Experiments Demonstrated" asks whether the results close the debate over quantum supremacy and argues they do not. It gives an overview of the experiments and identifies open questions. The paper separates a general claim that a quantum computer can perform some tasks faster than a classical computer from the narrower Google claim that at least one task can be executed exponentially faster.

A statistical critique

A 2023 blog post by Gil Kalai summarizes concerns from his group's statistical analysis of the 2019 data. The post states that the supremacy claim has largely been refuted by several research groups and lists three concerns: the data do not agree with Google's noise model, a formula for a priori fidelity estimation is surprisingly simple and seems to involve an unexpected independence assumption yet gives very accurate predictions, and the calibration process has surprising statistical properties. These are one critic's views, from a blog and an associated abstract, and the post's claim of refutation is its author's assessment.

Limits

The 2019 benchmark is a specific sampling task with no practical application stated in the three pages used for this post. The 10,000 year figure is the authors' estimate from their benchmarks at the time. The critiques are not settled results either, and none of the three pages includes the Google team's reply.

FAQ

What did the 2019 experiment do?

It sampled the output of a pseudo-random circuit on 53 superconducting qubits, according to the Nature paper.

What was the classical comparison?

The authors estimated about 10,000 years on a state-of-the-art supercomputer, against about 200 seconds on Sycamore for a million samples.

Is the claim accepted?

Not universally. A 2020 arXiv paper argues the results do not close the supremacy debate, and a 2023 blog post says the claim has largely been refuted by several groups.

Why use random circuits?

The authors say random circuits have no structure and so allow limited guarantees of computational hardness.

Does this show a practical use?

The three pages describe a benchmark task and do not claim a practical application.