Study compares 10 quantum processors from Quantinuum, IBM and IQM in error correction operations

In the IBM processors analyzed, measurements in the middle of a calculation increase the effective error by around ten times regarding the same task without them. In Quantinuum’s, this extra cost is comparable to that of a two-qubit gate, the operation that connects two qubits to each other. This is reflected in a quantum computing study that has been published on arXiv by JA Montañez-Barrera and Kristel Michielsen, from the Jülich Supercomputing Center, which compares 10 processors from Quantinuum, IBM and IQM.

These measurements are essential for correcting errors, because they allow errors to be detected and the result to be used to guide subsequent operations, but they can also introduce new errors or leave qubits waiting. That is why the authors propose a pretest that estimates whether a chip is ready, without launching the entire experiment, which requires a lot of calculation time.

The test uses a version of the QAOA algorithm, a method that seeks good solutions to mathematical problems, posed so that solving them requires connections and measurements similar to those of certain correction codes. With each added layer, it is measured how quickly the quality of the solutions worsens: 1 is equivalent to the optimal solution and 0.5 is random. The circuits reach 2,950 measurements intermediate.

With surface code, which protects information through checks spread across a network of qubits, Quantinuum’s Helios-1 scored highest with 25 qubits of data at all depths analyzed, with differences of two to four standard deviations per point and 50 executions in each one. With chains of 30 qubits the trend was similar, but the uncertainty does not allow the improvement to be confirmed.

Up to 91 qubits of data and the limit of all eight zones

In broader testing, Helios-1 maintained above-chance results with 81 qubits of data in surface code, 91 in triangular color code, and 48 in bivariate-bicycle codes, a family being studied to reduce resource requirements. These essays They measure whether the processor can withstand demanding operating patterns and they do not demonstrate that errors at that size were corrected. In 91, only seven of its 98 physical qubits were free as auxiliaries, so the measurements were divided into batches.

The cost of measurements on Helios-1 grew with longer chains, something that the study links to its eight operating zones: not all measurements can fit at the same time and some qubits wait. By running more operations in parallel, the test saw a clear improvement.

To validate it, the authors compared it with protected memory experiments in 11 zones of the IBM Phoenix processor. The highest-rated areas tended to have fewer logical errors, with an average deviation of 1.6 places in the classification versus between 2.1 and 2.7 for the calibration measures examined, although only in small patches of surface code and on a single processor.

The researchers also mention that more tests will be needed before this method replaces full experiments. Furthermore, the work is a preprint and has not yet undergone peer review.