Researchers at Google Quantum AI have demonstrated a quantum error-reduction technique on real hardware for the first time, cutting errors in one of the most awkward ingredients of quantum computing by a factor of 40.
The work was carried out on Willow, the company's superconducting processor, and published in the journal Science. Paul Arnold reported the result for Phys.org.
Qubits, the quantum equivalent of digital bits, are fragile and prone to errors, which is what makes building reliable quantum computers so difficult. The established remedy for the particular states needed to run general-purpose quantum algorithms is magic state distillation, which filters flaws out of many noisy copies but is very resource-intensive.
The technique tested here is the alternative: magic state cultivation, which cleans up a noisy magic state directly inside a protected qubit rather than by filtering a crowd of them. Until now it had never been successfully tested on real hardware.
The team encoded a noisy magic state across several physical qubits arranged on a small grid, then used helper qubits to check for errors without destroying the data being checked.
Attempts that failed a check were discarded and restarted.
"Cultivation reduces the error by a factor of 40," the researchers wrote. The refined states reached a fidelity of at least about 99.99 per cent.
About 8 per cent of initial attempts passed every fidelity check.
The team also developed a measurement method it calls kickback tomography, which detects remaining errors using far fewer measurements than traditional techniques require.
The cultivated states were then moved into a larger arrangement compatible with a surface code, which suggests they can be carried into bigger error-correcting codes for storage and later use. The point of the exercise is equipment: making magic states by distillation demands a great deal of extra hardware, and the researchers say the next step is to scale cultivation to larger machines.
The headline figure is best read alongside the acceptance rate, because they are two halves of the same sentence. A fidelity of at least 99.99 per cent was achieved among the attempts that passed every check. About 8 per cent of them did.
This is a finding of general application, and this newspaper proposes to adopt it immediately. Any process whatever can be made arbitrarily accurate, provided its failures are identified and discarded before the accuracy is measured.
Our analysts have accordingly run the method over this newspaper's own output and report a factual error rate of zero, across the 8 per cent of paragraphs that survived the check. The remaining 92 per cent have been restarted.
The vocabulary is the part we admire most. The discarded attempts are not failures; they are simply uncultivated. The surface code receives only what has been cultivated. Nothing in the record is inaccurate, because inaccuracy is the criterion for not being in the record.
The Department of Unnecessary Research notes that distillation, the older method, was criticised for the sheer quantity of hardware it consumed, and that cultivation has reduced the equipment requirement by moving the waste from the machine to the schedule.
On the Willow processor, as elsewhere, the remaining errors have not been corrected. They have been asked to leave.