Google Quantum AI Pushes Willow to All 105 Qubits in Contextuality Tests
Google Quantum AI measured contextuality on Willow, winning the magic square game 98.3% of the time and sustaining quantum advantage across 105 qubits.
Takeaways
- Google Quantum AI demonstrated quantum contextuality at scale on Willow, published on the cover of Science.
- Qubits won the Mermin-Peres magic square game 98.30% of the time, beating the 88.89% classical limit.
- Kochen-Specker-Bell inequality measurement reached chi equal to 5.590, above the noncontextual bound of 4.
- GHZ parity game exceeded classical win rates up to 71 qubits; multipartite entanglement confirmed across 45 qubits.
- Hidden Linear Function solved on grids up to 105 qubits with success probability above 0.5.
- Code and data released via Zenodo and ReCirq; no unconditional quantum advantage claimed.
Google Quantum AI has tested quantum contextuality across four experiments on Willow, including circuits that occupied all 105 superconducting qubits. The results, reported in Science and summarized in Google’s overview, range from a direct violation of a noncontextuality inequality to shallow-circuit tasks that compare quantum hardware with restricted classical strategies.
Quantum systems can exhibit entanglement among subsystems and contextuality across compatible measurement settings. Contextuality rules out models that assign every observable a pre-existing value independent of which compatible observables are measured with it. The property matters for computing because it is tied to resources used by quantum algorithms, while its algebraic constraints can expose hardware errors that aggregate fidelity scores conceal.
| Experiment | What it tests | Scale |
|---|---|---|
| Mermin-Peres magic square | Coordination beyond the classical win-rate ceiling | Nine input pairs |
| Kochen-Specker-Bell inequality | Direct violation of a noncontextual bound | Two data qubits plus ancillas |
| GHZ parity game | Multipartite contextual correlations | Up to 71 qubits |
| 2D Hidden Linear Function | Quantum and constrained classical circuit depth | 9 to 105 qubits |
Nine squares, one impossible assignment
In the Mermin-Peres magic square game, two players fill entries in a 3 × 3 grid with +1 or −1. Every row must multiply to +1, while every column must multiply to −1. A fixed grid cannot satisfy both conditions because multiplying all row products gives +1 and multiplying all column products gives −1.
The cooperative version assigns Alice a random row and Bob a random column. Each player returns three values with the required product, and their answers must match at the intersecting cell. They cannot communicate after receiving their assignments. Any classical strategy based on fixed values can win at most eight of the nine possible input pairs, producing an 88.89% ceiling.
A quantum strategy uses shared Bell pairs and compatible measurements selected for each row or column. An ideal implementation wins every round because the outcomes need no globally consistent, context-independent assignment. On Willow, Google ran each input pair 100,000 times and reported an average win rate of 0.9830 ± 0.0001. Device noise accounted for most of the gap from the ideal result.
A direct score for contextuality
The Kochen-Specker-Bell test converts the same logical conflict into a numerical inequality. Its score, χ, combines expectation values for row and column products using the signs required by the magic square. Every noncontextual hidden-variable model satisfies χ ≤ 4, while quantum mechanics permits a maximum of 6.
Executing the test required three compatible observables to be measured sequentially on the same two data qubits without discarding or resetting them between measurements. Ancilla qubits recorded each result through quantum nondemolition readout. Across 180 measurement contexts and 1,000 shots per context, Willow reached χ = 5.590 ± 0.008. The authors attribute the remaining distance from 6 to gate and readout errors.
Seventy-one qubits play the parity game
The team extended the test with a Greenberger-Horne-Zeilinger parity game involving N players. Each player receives an input bit and returns an output without communicating; the combined outputs must satisfy a parity rule determined by the inputs. The best classical win rate approaches 50% as the player count grows, while an ideal GHZ strategy wins every valid round.
Willow’s win rate remained above the classical bound at every tested size through 71 qubits, although noise steadily reduced the margin. A separate analysis certified genuine multipartite entanglement across 45 qubits, confirming correlations that could not be decomposed into smaller entangled groups.
A shallow circuit fills Willow
The largest experiment implemented the two-dimensional Hidden Linear Function problem used by Bravyi, Gosset, and König in their 2018 theorem. That work studies circuit depth, the number of sequential gate layers required to solve a problem when gates are local and have bounded fan-in.
The quantum circuit applies Hadamard gates, S gates selected from a random symmetric binary matrix, four layers of controlled-Z gates arranged on the processor grid, another round of Hadamards, and measurement. Its gates belong to the Clifford set, which makes the circuit suitable for efficient verification with the Stim stabilizer simulator.
For every grid size from 9 to 105 qubits, the team generated 1,000 random instances and collected 100 shots per instance. Success stayed near 100% on small grids and fell to roughly 60% at 105 qubits, remaining above the 0.5 threshold used in the paper’s conjectured finite-size comparison. Under the paper’s retry-adjusted depth metric, the hardware run scored about six layers, compared with four for an ideal noiseless implementation.
The 105-qubit caveat
The 2018 theorem provides an asymptotic logarithmic lower bound on the depth of eligible classical circuits. With the constants used in the analysis, the rigorous separation becomes quantitative only at scales around one million qubits. The 105-qubit experiment therefore depends on a conjectured finite-size baseline.
- Classical model: The comparison covers bounded-depth, geometrically local circuits. General classical algorithms fall outside the theorem.
- Simulation cost: Stim can efficiently simulate these Clifford circuits on conventional hardware. The claimed separation concerns circuit depth under specific architectural constraints.
- Computational claim: The experiment does not establish unconditional quantum advantage on a practical workload.
Algebraic tests expose hardware drift
Aggregate random-circuit benchmarks compress many failure modes into a single fidelity estimate, which can obscure small coherent calibration errors such as systematic over-rotations and under-rotations. The contextuality experiments require exact parity and commutation relations across many operations, causing those errors to appear as specific reductions in χ or game win rates.
In stabilizer-based fault-tolerant models, Clifford operations remain efficiently classically simulable, while suitable non-Clifford magic states enable universal computation. Contextuality provides one way to characterize the additional computational resource supplied by those states. Google has separately reported magic-state cultivation on its hardware, although the contextuality study did not run a fault-tolerant algorithm.
Code is public; replication remains
The team released plotting data on Zenodo and circuit-generation and analysis code through ReCirq. Those artifacts allow other researchers to regenerate instances, rerun the analysis, and adapt the tests to different processors. No independent hardware group has reproduced the results, leaving the proposed use of contextuality as a cross-platform metric without external validation.
The paper identifies molecular simulation in chemistry and materials science as possible downstream applications, but it reports no domain simulation. The immediate output is a reproducible set of structured hardware tests spanning direct contextuality measurements, multiplayer games, and chip-wide shallow circuits.