Quantum Supremacy Claims That Fell to Classical Simulation

In 2019 Google announced quantum supremacy: its 53-qubit Sycamore processor sampled random circuits in 200 seconds, a task the accompanying Nature paper estimated at 10,000 years on the best available classical supercomputer. The estimate held for roughly two years.

This article traces what happened to that claim and to those that followed, and extracts the structural reasons the pattern recurs. The purpose is not to diminish the hardware achievements, which are real, but to be precise about what was demonstrated, because the gap between the technical claim and its public reception has been consistently large.

What the Sycamore experiment actually claimed

The task was random circuit sampling: execute a randomly chosen circuit and return samples from its output distribution. The verification metric was linear cross-entropy benchmarking, which scores how strongly the observed bitstrings correlate with the ideal distribution's high-probability outcomes.

Two features of the task deserve emphasis. It has no application: nobody wants samples from a random circuit, and the problem was chosen precisely because it is hard to simulate rather than because it is useful. And the fidelity achieved was approximately 0.2 percent, meaning roughly 998 of every 1,000 samples were noise rather than signal, with the metric detecting the residual correlation in the remainder.

The 10,000-year figure assumed a specific classical algorithm, Schrodinger-Feynman simulation, on a specific machine, and assumed the classical simulation would reproduce the distribution exactly.

How the estimate collapsed

Between 2021 and 2022, Pan and Zhang and collaborators reproduced the benchmark using tensor-network contraction on a GPU cluster, first in days and subsequently in hours. IBM had separately argued in 2019 that a disk-based approach on Summit would take days rather than millennia.

The decisive insight was not additional compute. It was that the classical simulation need only match Sycamore's fidelity, not the ideal distribution. Reproducing samples at 0.2 percent fidelity is an exponentially easier problem than exact simulation, and the original estimate had compared an approximate quantum device against an exact classical algorithm. Once the comparison was made like for like, the advantage largely evaporated.

The pattern repeats

In 2023 IBM published a 127-qubit experiment in Nature reporting Ising-model dynamics, framed as utility beyond brute-force classical simulation. Within weeks multiple groups reproduced the results classically: first with tensor networks, then with Pauli propagation and with sparse Clifford-based methods. Several reproductions ran on a single laptop.

The paper's technical claim was narrower than its reception. Beyond brute force is a statement about one family of algorithms, not about classical computation generally, and the reproductions used methods that were not brute force. The claim as written was defensible; the claim as reported was not.

Claim                          Year   Classical response              Time to match
Google Sycamore supremacy      2019   Tensor networks (Pan/Zhang)     ~2 years
IBM 127-qubit utility          2023   Tensor networks + Pauli prop.   weeks
Various boson-sampling         2020+  Classical samplers, spoofing    months to years

Boson sampling and quantum annealing claims have followed similar arcs: publication, coverage, then a classical algorithm matching the result within months to a small number of years.

Why classical methods keep winning

Three structural reasons recur, and they are properties of the situation rather than accidents.

• Noise sets a low bar. A device operating at low fidelity has already discarded most of the quantum information in its output. A classical method may trade accuracy for speed until it reaches the same degraded level and no further, and that trade is exponentially favourable. The noisier the device, the easier it is to match.
• Structure is exploitable. Any circuit computing something useful has regularity, and that regularity is what makes it useful. A circuit with no structure is hard to simulate and also computes nothing. This is a genuine tension in the advantage programme rather than a temporary state of affairs.
• The target is fixed and the pursuit is not. A published claim is a static benchmark. Classical algorithms continue improving against it indefinitely, and the reproduction only has to succeed once.

There is a fourth reason that is less often stated. The classical simulation community is substantially larger than the quantum hardware community and has access to mature tooling, decades of tensor-network methodology from condensed matter physics, and commodity GPU clusters. A claim that stands against that is standing against considerable pressure.

What has not fallen

The position should be stated accurately rather than dismissively. Deep random circuit sampling on newer processors, including Google's Willow-class devices, has not been matched classically. Those circuits are engineered to maximise entanglement and non-Clifford resource simultaneously, defeating stabilizer methods, tensor networks and Pauli propagation at once, and the improved fidelity of the newer devices removes the noise-matching shortcut that defeated Sycamore.

This is a real result. It is also a result about a task that computes nothing of independent interest. The accurate summary of 2026 is that quantum hardware leads on contrived sampling benchmarks and classical methods lead on every benchmark tied to a useful computation. Both halves of that sentence matter.

What follows for practice

The history has a direct operational consequence: before paying for hardware access on the assumption that a problem is classically intractable, verify the assumption. The verification is cheap and the assumption has an unimpressive track record.

The specific methods that answered these claims are not exotic research artifacts. Tensor-network contraction with fidelity accounting and Pauli propagation with truncation bounds are the production engines offered here, described in Tensor networks explained and Pauli propagation, together with the accuracy statements this history demonstrates are necessary.

The design principle is the one the episode teaches. A performance claim without an accuracy statement is not checkable, and claims that are not checkable have repeatedly turned out to be wrong in the direction their authors preferred.