Boson sampling: what did Jiuzhang and Borealis actually prove?

Gaussian boson sampling (GBS) sits behind the loudest “quantum advantage” headlines of the decade — and behind the most complete public record of those headlines being renegotiated. Rosetta Quantum reads that record here as one dataset: four machines across six years, every advantage margin published before 2026 shrunk or overturned by classical algorithm work alone, and a fresh 2026 claim engineered specifically to survive the best-known counterattack. Each section opens with its answer; every number carries its source.

What is Gaussian boson sampling, and what does it prove?

When it works, GBS proves exactly one thing: that a photonic device can draw samples from a probability distribution that no known classical algorithm can reproduce at the same scale and speed. Squeezed states of light enter an interferometer; detectors record which output modes fire and how many photons arrive; the probability of each detection pattern is governed by a matrix function (the hafnian) whose computation is believed to be classically hard. The machine is not running a program — it is being itself, faster than a supercomputer can imitate it. That is a supremacy-style claim, not an application claim; the distinction is the subject of our advantage-vs-supremacy explainer.

Two consequences follow. First, the task is native to the device: nobody needed samples from that distribution before the machine existed. Even coverage sympathetic to the newest result states it plainly — GBS “is not a general-purpose computation” and the machine cannot “break encryption or simulate new drugs” (PostQuantum, 2026). Second, the claim is exactly as strong as the best classical rival on the day you check — and the rival keeps improving after publication day. That structural fact, which Rosetta documented for superconducting hardware in the tensor-networks scoreboard, is the entire history of GBS.

What did Jiuzhang and Borealis actually claim?

Four machines, escalating photon counts, escalating margins — each margin correct against the classical algorithms known on its publication day. The ledger, with a source per row:

Borealis earned one structural first: it was programmable — its 216-mode entanglement topology could be reconfigured — where the first Jiuzhang generations were largely fixed circuits (Madsen et al., Nature 606, 2022).

What did the classical counterattacks do to those numbers?

They shrank or overturned every margin published before 2026 — without touching any hardware. Three episodes carry the story:

January 2022 — nine orders of magnitude, 13 months after the claim. Bulmer et al. improved exact GBS simulation until Jiuzhang 1.0’s “600 million years” became “several months” on the same class of hardware — a billion-fold acceleration from algorithms alone (Science Advances 8, eabl9236, 2022).

August 2023 — the classical explanation that would not die. Martínez-Cifuentes et al. showed that a fully classical “squashed light” model had not been ruled out as an explanation of Jiuzhang 1.0’s published data. The same paper cuts both ways, and Rosetta reports both edges: their Bayesian tests favor the quantum ground truth for Jiuzhang 2.0 and most configurations (Quantum 7, 1076, 2023).

June 2024 — the benchmark falls. Oh, Liu, Alexeev, Fefferman and Jiang built a tensor-network sampler that exploits the experiments’ photon loss and, “using relatively modest computational resources”, outperforms the large-scale experiments on the very benchmarks used as evidence of advantage — the paper states it “disputes the experimental quantum advantage claims” (Nature Physics 20, 1461–1468, 2024; arXiv:2306.03709).

This is the same pattern the tensor-networks post measured for superconducting machines: publication starts the clock, it does not stop it. GBS adds the sharpest single number in the whole series — 10⁹× of claimed margin recovered classically on Jiuzhang 1.0’s task — which is why it is the number on this post’s cover.

Where does the GBS claim stand in September 2026?

On its strongest hardware and its least-tested number. Jiuzhang 4.0 (Nature, May 13, 2026) detects up to 3,050 photons across 8,176 modes fed by 1,024 squeezed sources, at 25.6 μs per sample, with 92% source efficiency — an order-of-magnitude jump in detected photons over anything before it. Its design goal is explicit: the team validated “against all current classical simulation methods, especially the matrix product state algorithms recently designed to exploit photon loss” — that is, against the 2024 counterattack — and estimates that same MPS method would need more than 10⁴² years on El Capitan (arXiv:2508.09092).

Rosetta reads that figure with the series’ own history: (1) it is the team’s estimate against today’s best-known algorithm family, not an externally reproduced measurement; (2) every earlier “years” figure in the table above was also correct at publication against the then-best algorithm. Whether 10⁴² years stands will be decided by the next algorithm, not by the last experiment — and until an external team publishes either a confirmation or a counterattack, the honest status is: claimed, fresh, untested. For the broader photonics map — PsiQuantum’s system promises, Xanadu’s Aurora — see our photonics ledger, which this post does not re-litigate.

Why is a boson sampler so hard to verify?

Because ground truth is unavailable exactly where the claim lives. At low photon numbers the ideal distribution can be computed and checked — Borealis reported fidelities above 99% there (Nature 606, 2022) — but at 219 photons, let alone 3,050, computing the ideal probabilities is the very task claimed to be intractable. Nobody can grade the machine’s answers directly at scale.

Validation therefore leans on proxies: benchmark scores against deliberately wrong models (thermal states, distinguishable photons, squashed light), correlation functions, Bayesian model comparison (PostQuantum, 2026). The 2024 result showed a classical sampler can score better than the experiment on such proxies (arXiv:2306.03709); the Borealis authors themselves wrote that “it remains an open question how to verify that a GBS device is operating correctly” and left open “whether better polynomial-time algorithms for spoofing can be developed” (Nature 606, 2022). Rosetta’s benchmark red-team guide applies here without modification.

Does GBS do anything commercially useful today?

No. The advantage claims are physics claims, not application claims: as of this date there is no GBS application advantage — in graphs, chemistry or machine learning — in the public record, and even favorable coverage of Jiuzhang 4.0 says the machine cannot “break encryption or simulate new drugs” (PostQuantum, 2026). What the GBS program has demonstrably been is a proving ground for photonic engineering — squeezed-light sources, loss control, detection at thousands of photons — components that fault-tolerant photonic roadmaps will need if they are to deliver (our photonics ledger).

What we don’t know

• Whether the >10⁴² years estimate survives the next classical algorithm. History in this exact series says margins fall by orders of magnitude; history is not proof that this one will.
• Whether Jiuzhang 4.0’s samples come from the ideal squeezed-light distribution. Verification at 3,050 photons is statistical by necessity; ground truth is intractable.
• Whether a polynomial-time spoofing algorithm exists for the newest benchmarks — the Borealis authors’ 2022 open question stands unanswered in public.
• Rosetta Quantum has not run or reproduced any of these experiments or simulations in-house; this post reads the public record. Our own measured quantum-vs-classical margins live in the Ledger’s unit, where the scoreboard still shows zero advantaged solves.

Rosetta Quantum publishes verdicts with raw, reproducible data. This is educational content, not a product claim.