Google's method for checking quantum computer answers, for R&D teams
ShiniFrontiers · Source: Google Quantum AI
What is it, and who gains?
What it is: A quantum computer is a new kind of computer built on the physics of atoms. It works with qubits, the quantum version of an ordinary computer's bits, which lets it take on problems ordinary computers struggle with, such as how molecules behave. Today's machines are noisy: tiny disturbances cause small errors as they run, so their answers come out slightly wrong. The long-term fix, called error correction, is machines that repair their own errors while they work, and those are still being developed. Until they arrive, every answer is corrected afterwards by calculation, called error mitigation, and the authors expect it to stay useful even after. This paper, from Google's quantum research team, is a better method of error mitigation.
Who it's for: Any company paying for quantum computer work in chemistry or materials. A quantum computer's answer is only as good as the error mitigation behind it, and this method tells you whether to trust it. Machine time is bought in shots, single runs of the program, and in the authors' simulations this method needed 10 to 100 times fewer of them.
The catch: It is not yet proven on real hardware. Every new result in the paper comes from computer simulations of small, 16-qubit circuits; a circuit is the program a quantum computer runs. Only an earlier, simpler version has run on Google's own machine.
A. Knowing whether to trust the answer — for heads of R&D
Problem: The most common method, zero-noise extrapolation, runs the program at several deliberately higher noise levels and traces the trend back to zero noise. It assumes the whole answer fades at one steady rate. When it doesn't, the answer can be up to 30% wrong, more shots don't fix it, and nothing warns you.
What it does: Picture a ration godown holding rice, which keeps, and onions, which rot fast. Working back to the original stock with one average loss rate gets it wrong. Google's method works out the mix first, then corrects rice and onions each at their own rate. It keeps refining until the total stops changing, and when it stops changing, you can trust the answer.
B. Paying for fewer shots — for whoever budgets the work
Problem: The exact method, probabilistic error cancellation, is right on average for any program, but too expensive for large problems: the shots it needs multiply as the problem grows.
What it does: In the godown, that method is like checking for every item that could possibly be stored. Google's method checks only for what is actually there. In the authors' simulations, that took 10 to 100 times fewer shots for the same accuracy.
See the difference
| The situation | Today | With Google's method |
|---|---|---|
| Knowing whether to trust the answerGoogle's molecule experiment, on its own quantum computer | Error 0.2: five times the 0.04 allowed | Error 0.018: inside the allowed error (earlier version) |
| Knowing whether to trust the answerCan you trust the corrected answer? | Up to 30% off, with no warning | Refined until it stops changing |
| Paying for fewer shotsShots for the same accuracy | The exact method: the most shots | 10 to 100 times fewer shots |
Rows 2 and 3 are from the authors' simulations. The godown is our example.
- What a company gains
Paid machine time, and decisions built on wrong answers.
- What it would take
Nothing to buy yet. It is a published method, and the authors share their code on request.
- How ready
- Early research
Shown in simulation, and an earlier version works on Google's quantum computer. Tests on real hardware are next.
Join free to unlock: how it works, what to do now, what it could change for the business, and what is not yet proven.
Join free to read it all →