The scale is substantial: 61 qubits and 918 entangling gates. How do we know whether the measured strings still carry the signal of the intended circuit?
The task
Random-circuit sampling asks a processor to execute a fixed, complicated circuit and measure its 61 qubits at the end. Each shot produces one bitstring. At 36 cycles, the released circuit contains 918 CZ gates. The paper used a subset of the 120-qubit Nighthawk r2 processor ibm_phoenix and collected a million bitstrings in 19 seconds of QPU time.
Why a million bitstrings are not enough evidence
Even a poor or uniform sampler can produce a million distinct 61-bit strings: the space has 261 possibilities. A histogram of zeros and ones also cannot show which particular outcomes the ideal circuit favors. Direct cross-entropy benchmarking (XEB) would require ideal probabilities for measured strings. At this depth, calculating those full 61-qubit probabilities for the complete dataset is too costly.
The paper’s quality estimate
The authors instead used two routes: mirror circuits, which test a reversible operation, and circuits divided into three or four smaller patches whose ideal probabilities remain calculable. At 36 cycles, the three estimates are about 0.0022, 0.0018 and 0.0020; the fitted value is about 0.0023. This is a proxy for full-circuit quality, not direct XEB of the million full-width bitstrings.
Where do the classical years come from?
The paper estimates the cost of a specific classical route: many independent tensor contractions for a rejection sampler at the target fidelity. At 36 cycles, its table gives about 1021.82 complex operations per amplitude and roughly 1.2 × 1027 machine FLOPs for a million samples, translated into about 110 Frontier years under the stated assumptions. This is not a lower bound on all classical methods. The authors explicitly leave amplitude reuse, approximate contraction and MPS as open alternatives.
Sources: the preprint and the released code and measurements.
Project overview and article source
Edukaizen · 25 september 2026 / 25 September 2026