Quantinuum Demonstrates Provable Quantum Advantage on 55 Qubits with 137-Billion-to-1 Gap

On September 8, 2026, the Quantinuum team published a paper in *Nature Communications* titled "Unconditional and exponentially large violation of classicality." On their H2 trapped-ion processor, they ran a game called complement sampling: scaling from n = 5 to n = 37, with the largest circuit using 55 qubits and roughly 228 native two-qubit gates. The theoretical gap grows exponentially with scale, reaching 137 billion to one at n = 37 — and this is achieved against a mathematically rigorous classical upper bound, not on the basis of "we assume some classical algorithm is hard." Every quantum supremacy experiment of the past decade — Google Sycamore's random circuit sampling, USTC's Jiuzhang and Zuchongzhi Gaussian boson sampling — has been challenged with "a smarter classical algorithm may exist in the future." This one is not open to that objection.

🎮 A Game Whose Ceiling Is Sealed by Mathematics

The core rule of complement sampling is simple: the referee secretly splits the full set of n-bit strings into two halves, A and B, hands the player one string from A, and asks the player to return a string from B.

• A classical player that receives one string from A can only exclude that single string. It knows nothing about the internal composition of A and B — among the remaining \(2^n - 1\) strings, which belong to A and which to B is a complete black box. As n grows, the optimal classical success rate converges exponentially to 1/2.
• The quantum player receives not a single string from A but the entire set A in quantum superposition. Using a circuit called the "swapper" (built from a Grover diffusion operator and a Bernstein–Vazirani-style structure), it moves the whole superposition of A into B, then measures — ideally always returning a string from B.

The classical upper bound is directly proven by information theory — no computational complexity assumptions required.

Placing this alongside the past decade of quantum supremacy experiments, the difference is fundamental:

| Experiment | Basis of advantage | Relies on unproven complexity assumptions? | |---|---|---| | Google Sycamore random circuit sampling | "Simulating the output distribution is computationally hard" | Yes (complexity conjectures) | | USTC Jiuzhang / Zuchongzhi Gaussian boson sampling | "Approximating matrix permanents is hard" | Yes | | Quantinuum complement sampling (9/8) | Classical bound proven by information theory | No |

The Quantinuum team is careful in the paper: this does not mean quantum computers are 137 billion times faster than classical ones on any commercial task, but that "within this mathematically bounded game, the distance between quantum and classical grows exponentially with scale."

📏 Results Achieved on Hardware

Device: Quantinuum System Model H2 trapped-ion processor — 56 qubits, ytterbium-171 hyperfine qubits, barium ions for sympathetic cooling. All qubits are fully connected, with no need for SWAP gates to move quantum information. Two-qubit gate fidelities are stable at above 99.8%.

Experimental procedure:

• Complement sampling instances measured from n = 5 to n = 37;
• Several thousand distinct circuits run for each n;
• Most values of n ran 200 rounds (split evenly between scoring and input-state quality verification);
• n = 37 ran 1000 rounds;
• The classical upper bound was violated by the quantum strategy at the 1% significance level;
• The advantage grows exponentially with n — fully consistent with theoretical predictions.

The largest circuit used 55 qubits (out of the machine's 56), averaging 228 native two-qubit gates.

📐 How the "137 Billion to One" Number Is Actually Derived

"A theoretical ratio of 137 billion to one at 37 qubits" may sound like marketing, but there is a clear computational chain:

• For the Bernstein–Vazirani-derived instance family, the ratio of ideal quantum to best classical = 2^(n−1);
• At n = 37, \(2^{36} = 68{,}719{,}476{,}736\);
• The paper's "137 billion" figure comes from the specific violation convention used (the ratio's numerator and denominator are normalized differently), but the core fact is exponential growth with n.

In other words, this is not "quantum computers are 137 billion times faster than classical", but "the theoretical boundary of quantum advantage in this mathematically defined game grows exponentially with scale." Once published, this conclusion cannot be erased by "smarter classical algorithms" in the future — because the classical bound was not guessed, it was proven.

⚠️ What It Solves — and What It Doesn't

What it solves:

• The question of whether a quantum advantage's classical bound is mathematically proven;
• Whether the advantage can be efficiently verified classically in polynomial time — complement sampling's output can be efficiently checked on a classical machine;
• Whether it holds as systems scale — experimental data show the gap grows exponentially with n rather than shrinking.

What it doesn't solve:

• Trusted state-preparation assumptions: the referee and player share the same processor. Future experiments should place the referee and player on two physically separated quantum machines connected by a genuine quantum channel, to fully remove reliance on honest input-state preparation;
• Hardware noise: larger n means more accumulated two-qubit gate errors. The measured score at n = 37 is slightly below the theoretical bound — the quantum machine still beats the classical bound by a huge margin, but the measured gap increasingly deviates from theory as n grows;
• Quantum error correction: the experiment used no quantum error-correcting codes. Scaling further will require either higher gate fidelities or error correction;
• Any practical utility: complement sampling doesn't compute anything anyone wants. It is a non-classicality test, closer to a Bell inequality violation than to factoring large integers or simulating molecules.

The paper's conclusion is restrained: the game "can serve as a scalable benchmark for quantum hardware," and its violations grow exponentially with input size.

🧭 Placing It in the Week's Quantum Landscape

| Event | Date | Subfield | Difference from Quantinuum complement sampling | |---|---|---|---| | QuEra Nature spin readout | 9/1 | Neutral-atom quantum computing | Hardware: spin readout fidelity | | Diraq + Equinix silicon spin qubits in data centers | 9/1 | Silicon quantum commercialization | Hardware → commercial deployment | | IBM modular cryogenic coupler | 8/30 | Superconducting modularity | Hardware: scaling path | | Pasqal Matriq Xenomi quantum economy | 8/30 | Simulated quantum economy | Algorithms + economic modeling | | Stanford optical cavity arrays | Sept | Optical qubit readout | Hardware: cavity array scaling | | Quantinuum complement sampling | 9/8 | Benchmarking | Benchmark: provable advantage |

In the second half of 2026, the quantum field is showing a clear division of labor: the hardware camp tackles "more qubits + fewer errors," while the benchmarking camp tackles "how to mathematically prove quantum is genuinely useful." Quantinuum's work gives the benchmarking camp a yardstick that doesn't depend on betting whether classical algorithms will catch up.

🔁 The Significance: Quantum Benchmarks Move from "Guessing" to "Proving"

Over the past decade, the most common criticism of quantum supremacy claims has been that "the advantage rests on a complexity conjecture we believe but haven't proven" — every classical-algorithm improvement forces the superiority boundary to be re-argued. Complement sampling turns this from an empirical matter into a theorem:

• The classical bound comes from information theory, independent of computational complexity conjectures;
• The output is efficiently verifiable on a classical machine — no need to rerun the entire quantum process classically;
• The advantage grows exponentially with scale, aligned with — not opposed to — hardware noise; larger systems show clearer advantages, contrary to the old intuition that "bigger systems are harder to demonstrate";
• The team's next step is two physically separated quantum machines + a quantum channel, pushing from single-machine self-attestation to network verification.

Lines worth tracking next:

• Separated-machine verification: when will Quantinuum deploy referee and player on two physically separated H-series machines connected by a genuine quantum channel — the next step in the "quantum network" dimension;
• Benchmark extensions: can complement sampling be reworked into something closer to practical tasks, e.g., hardcoding molecular simulation or constraint-satisfaction problems into the A/B split;
• Porting to other platforms: can superconducting, photonic, and neutral-atom routes achieve comparable advantages on complement sampling — a first step toward an industry-wide benchmark;
• QOBLIB in parallel: IBM announced on September 1 that QOBLIB (Quantum Optimization Benchmarking Library) added 1200+ instances, 500+ solved to optimality — the optimization and supremacy tracks are simultaneously establishing comparable public standards.

📚 References

1. Nature Communications: "Unconditional and exponentially large violation of classicality," Benedetti et al., DOI: 10.1038/s41467-026-77413-3, 2026-09-08 2. arXiv preprint: 2511.11008 (2025-11-14) 3. Quantum Brief: "Quantinuum's 55-qubit experiment beats the best classical score by 137 billion to one," 2026-09-09 4. The Quantum Insider: "Game On — Quantinuum Demonstrates Exponential Edge Over Classical Strategies," 2026-09-08 5. Phys.org: "A new game demonstrates quantum advantage with provable classical limits," 2026-09 6. Quantinuum Blog: "Unlocking Quantum Advantage with Complement Sampling," 2026-09-08 7. IBM QOBLIB (Quantum Optimization Benchmarking Library) September 1 expansion update