Pillar C · State as of 2026-09-11

Tensor networks: the classical rival that keeps winning?

As of September 11, 2026, tensor networks — classical algorithms that compress quantum states — have matched or beaten four flagship “beyond-classical” claims after publication: Google’s Sycamore sampling (claimed Oct 2019, spoofed Mar 2021, arXiv:2103.03074), IBM’s 127-qubit utility experiment (claimed Jun 14, 2023, out-simulated in under two weeks, arXiv:2306.14887), D-Wave’s annealing supremacy (claimed Mar 12, 2025, contested the same month, Science adx2728) and IBM’s doped-Clifford run (claimed Jul 2026, reproduced as “compatible” in about three weeks, arXiv:2608.13110). One 2025 claim, Google’s Quantum Echoes, still stands. Rosetta Quantum treats tensor networks as the default challenger every new claim must outlive.

→ Leer en español

State as of: 2026-09-11

Status as of: 2026-09-11. Every number below carries its source next to it. This is a claim explainer in the Rosetta Quantum series: neutral, dated, and built to be checked.

What is a tensor network?

A tensor network is a classical data structure that stores a large quantum state as a web of small tensors instead of one exponentially large vector, keeping only as much entanglement as a tunable knob — the bond dimension — allows (Orús, Annals of Physics 349, 117, 2014). Matrix product states (MPS) chain the tensors in one dimension; projected entangled pair states (PEPS) tile them in two or three; belief propagation (BP) is a fast approximate way to contract the web (Tindall et al., PRX Quantum 5, 010308, 2024). When the real state carries limited or structured entanglement — the usual situation on today’s noisy hardware — the compressed classical copy can be accurate enough to reproduce, and sometimes beat, the machine’s own published numbers.

That one fact explains why the same family of methods keeps appearing on the other side of quantum announcements. It is not a rival computer. It is the referee’s ruler — and Rosetta Quantum tracks it as such.

Which beyond-classical claims have tensor networks already matched?

Four flagship claims since 2019 — and the response window has shrunk from months to weeks.

Claim	Claimed (source)	Tensor-network response (source)	Gap	Status at 2026-09-11
Sycamore random circuit sampling	Google, Nature 574, 505 (Oct 2019): 200 s vs an estimated 10,000 years classical	Spoof: Pan & Zhang, arXiv:2103.03074 (Mar 2021); sampling solved: PRL 129, 090502 (2022), about 15 h on 512 GPUs; then arXiv:2406.18889 (2024): 1,432 GPUs, faster than Sycamore itself with higher XEB	≈17 months to first spoof	matched and surpassed
IBM 127-qubit utility experiment	Kim et al., Nature 618, 500 (Jun 14, 2023)	Tindall et al., arXiv:2306.14887 (Jun 26, 2023) → PRX Quantum 5, 010308 (2024): BP tensor network more accurate than the processor; also Begušić, Gray & Chan, Sci. Adv. 10, eadk4321 (2024)	under two weeks	matched and surpassed in accuracy
D-Wave annealing supremacy	King et al., Science eado6285 (Mar 12, 2025)	Tindall et al., arXiv:2503.05693 (Mar 2025) → Science adx2728 (2026): lattice-matched tensor networks up to 900 qubits, bond dimension 8–32, workstation-class hardware; also Mauron & Carleo, arXiv:2503.08247	same month	contested — D-Wave’s Comment, arXiv:2504.06283 argues the hardest corners were not simulated
IBM doped-Clifford sampling	IBM, arXiv:2607.25941 (Jul 2026): 70 qubits, 70 entangling layers, 468 T gates	Manabe, Gu & Pan, arXiv:2608.13110 (Aug 13, 2026): temporal-boundary tensor-network contraction, 2,051 batches in 37.3 min on 256 H100 GPUs, log-XEB 0.35034 [0.29763, 0.40305] — compatible with IBM’s fidelity bound	≈3 weeks	compatible with the claimed bound; diagnosis, not refutation (our explainer)
Google Quantum Echoes (OTOC)	Google, Nature (Oct 2025): reported more than 10,000× faster than the best known classical algorithm	No classical match published. Google’s own analysis, arXiv:2604.15427 (Apr 2026), argues BP tensor networks cannot feasibly simulate it	open	standing as of 2026-09-11

Two readings of the same table are honest at once. Reading one: every claim that stood on circuits with limited entanglement was eventually matched, usually fast. Reading two: nothing here says the standing claims must fall — the D-Wave episode in particular is genuinely contested in both directions, in public, with both sides linked above. Rosetta scores that as the system working.

Why do tensor networks keep catching today’s quantum hardware?

Because noise compresses. A noisy processor cannot build or sustain the deep entanglement that would make its state incompressible, so a truncated classical description — a tensor network with modest bond dimension — often captures everything the device actually did. Tindall et al. attribute their PRX Quantum result to the fact that the Eagle experiment’s wavefunction carried mostly “tree-like correlations” that belief propagation represents almost exactly (PRX Quantum 5, 010308, 2024). The D-Wave response works the same way: build the network in the exact geometry of the annealer’s lattice and the required bond dimension stays small — 8 to 32 for systems up to 900 qubits (Science adx2728, 2026).

Tool	What it is	Where it breaks	Source
MPS (matrix product state)	1D chain of tensors; entanglement capped by bond dimension	volume-law entanglement in 2D and beyond	Orús 2014
PEPS (2D/3D networks)	lattice-shaped tensor web matching the hardware geometry	exact contraction is #P-hard — approximation is mandatory	Schuch et al., PRL 98, 140506 (2007)
Belief propagation contraction	fast approximate contraction; exact on trees	loopy, highly entangled states; corrections needed	Tindall et al., PRX Quantum 5, 010308 (2024)
Temporal-boundary contraction	contracts a 1D brickwork circuit along time	width grows with depth — helped by structure in the circuit	Manabe, Gu & Pan, arXiv:2608.13110 (2026)

A second, structural reason: the targets are public. Every counterattack in the table was possible because the claim shipped with enough circuit and instance detail to attack — which is exactly what radical reproducibility asks for. Google, IBM and D-Wave all published enough to be checked, and D-Wave replied to its challengers in public (arXiv:2504.06283). That deserves credit, and it is the opposite of misconduct: this is what verification looks like when it works.

Where do tensor networks hit their real limits?

The limits are as real as the wins, and they are theorems and measurements, not vibes. Exact contraction of a 2D tensor network is #P-hard (Schuch et al., PRL 98, 140506, 2007) — everything in the table above runs on approximations that silently fail when entanglement grows past what the bond dimension can hold. Google’s own team argues precisely this for its Quantum Echoes experiment: the output states are largely incompressible, the required bond dimension explodes, and the cost of BP tensor networks scales exponentially in system size (Bermejo et al., arXiv:2604.15427, Apr 2026). We flag plainly: that analysis is authored by Google Quantum AI — an interested party defending its own claim — and it will stand or fall when independent groups weigh in. It will be judged the way every row above was judged: on the arXiv, in public.

The forward-looking limit matters more. Error-corrected processors, if the published roadmaps deliver, will run deeper circuits whose entanglement will outgrow any fixed truncation — that regime is where tensor networks will stop being the whole referee and verification itself will get harder. That is a future-tense sentence about announced plans, not a measured fact.

How does Rosetta Quantum read a new advantage claim, given this history?

Rosetta Quantum applies four questions, all derived from the table above. First: start the clock — record the claim date and watch for the tensor-network response; the documented record is ≈17 months (Sycamore), under two weeks (IBM utility), same month (D-Wave), ≈3 weeks (doped Clifford). Second: check what classical baseline the claim itself used, because a claim that never ran a serious tensor network is a weak-baseline benchmark waiting to be corrected. Third: prefer claims that publish circuits and instances — unattackable claims are unverifiable claims. Fourth: read “compatible” carefully — the doped-Clifford result confirms IBM’s fidelity bound while also showing a classical machine reproducing the sampling task; whether that survives as an advantage is an open debate we already cover. From this date, Rosetta will log the claim-to-response gap for each new beyond-classical claim in its ledger, and will publish that series as it accumulates.

What we don’t know

Whether Quantum Echoes will stay standing. The only published infeasibility analysis is vendor-authored (arXiv:2604.15427); independent classical attacks have not been published as of 2026-09-11. Absence of a counterattack is evidence, not proof.
Whether the corners D-Wave’s Comment defends — the largest 3D lattices, the fastest-correlating ensembles, higher-order observables (arXiv:2504.06283) — will be reached classically. Open in both directions.
Where exactly the compressibility frontier sits. There is no clean, agreed-upon boundary between “tensor networks will catch it” and “they will not”; each episode is settled empirically.
Our own limit: Rosetta Quantum has not re-run any of these simulations in-house. This post reads the public record. If we reproduce one of these episodes on our own hardware, it will be published with its own seal and raw artifacts.