Fixing Quantum Computer Errors After the Run Is Over
Error-corrected quantum computers still let some mistakes through. A Google team proposes cleaning up the answers afterwards, using alarms already recorded.
In one paragraph
Error-corrected quantum computers still let some logical errors slip through. A preprint from Google Quantum AI proposes virtual error cancellation: store each run's record of error alarms, re-decode it later with a slower, more accurate decoder, and give runs where the two decoders disagree a negative weight so leftover errors cancel in the average. In simulations of a distance-7 surface code, leftover errors fell by more than three orders of magnitude. It has not been peer reviewed or tested on hardware.
How does virtual error cancellation work?
Quantum computers protect information by spreading one reliable logical qubit across many fragile physical qubits, arranged in a surface code. Repeated checks record which error alarms fired, a log called the syndrome record, and a fast decoder uses it to correct errors in real time. Because it must decide within microseconds, the fast decoder sometimes guesses wrong.
The authors propose leaving the quantum circuit unchanged and storing every run's syndrome record and result. Afterwards, on an ordinary computer, they learn the noise rates directly from those records, with no separate calibration experiments, and re-decode each record with a slower decoder that finds the most likely error exactly.
Runs where the two decoders agree get a weight slightly above one. Runs where they disagree get a negative weight, so that in the weighted average the remaining logical errors cancel. The authors prove this gives an unbiased answer and test it in simulations of surface codes of distance 3, 5 and 7 under a simplified noise model.
How much did the method reduce quantum errors in simulation?
In simulations of a distance-7 surface code, the method reduced the leftover logical error by more than three orders of magnitude compared with standard decoding. Compared with postselection, the common alternative of throwing away runs where the decoders disagree, it was better by over two orders of magnitude.
The noise rates learned from the stored syndrome records closely matched the true rates, without any extra calibration runs.
The method has a cost: because some runs count negatively, the final average fluctuates more, so more repetitions of the computation are needed for the same precision. The authors measure this extra cost with a ratio they call eta. It is about 4 for standard logical error cancellation and 2 for the best existing method that uses syndrome information, while the new method approaches 1 as the code grows, close to a twofold saving over existing syndrome-aware methods.
Why does post-processing matter for early quantum computers?
Early error-corrected machines will run with fairly small codes, so some logical errors will always remain, and adding qubits to push them down is slow and expensive. A method that improves accuracy using software alone, after the run, could make those machines useful sooner.
The approach also removes a timing constraint. The slow, accurate decoder does not need to keep pace with the hardware, because it works on stored records after the computation has finished.
A fair reading is that the main shift is in how syndrome records are viewed: not only as a by-product of real-time error correction, but as data that can be mined after the fact to correct results.
What are the limits of virtual error cancellation?
Read before you quote it
This is a preprint that has not been peer reviewed, and its limits are clear:
- Simulation only. All results come from simulations under a simplified code-capacity noise model, where errors affect only data qubits, not from quantum hardware or realistic circuit-level noise.
- Pauli noise assumed. Coherent errors and leakage are handled by separate tools or argued rather than tested.
- Cost grows with circuit size. The number of extra repetitions grows with the number of noisy operations in a circuit.
- Expensive components. The accurate offline decoder and the noise learning become computationally expensive for larger codes.
The study does not show the method working on a real quantum computer, and it does not replace quantum error correction: it cleans up the errors that error correction misses, at the cost of extra repetitions.
- It does not show the method working on a real quantum computer.
- It does not remove the need for quantum error correction or make large computations free of extra cost.
- All results come from simulations under a simplified code-capacity noise model, not from quantum hardware or realistic circuit-level noise.
- The analysis assumes Pauli-type noise; coherent errors and leakage are handled only by separate tools or argued informally.
- Cancelling errors costs extra repetitions of the computation, and that cost grows exponentially with the number of noisy operations in a circuit.
- It depends on a slow, accurate offline decoder and on learning every fault rate, which becomes computationally expensive for larger codes.
Takeaways
- saysThe method stores each run's error alarms, re-checks them later with a slower decoder, and gives disputed runs a negative weight so leftover errors cancel.
- saysIn simulations of a distance-seven code, leftover errors fell by more than three orders of magnitude, and by over two compared with discarding disputed runs.
- saysIt's a preprint with simulations under simplified noise, so it still has to be shown on real quantum hardware.
- explainsA syndrome record is the log of which error checks fired during a run; it hints at where errors occurred without disturbing the stored information.
- interpretsA fair reading is that stored error records could become a resource for correcting results after a quantum computation, not just during it.
Ask the paper
In PaperSays you can ask this paper anything, by text or by saying “Hey Aiden”. Three questions listeners asked:
What is virtual error cancellation in quantum computing?
A method that cancels leftover logical errors after a run by reweighting stored results, using error alarm records and two decoders. Hear the full answer in the app →
Why not just throw away the runs where the decoders disagree?
Discarding them leaves bias from errors hidden in the agreed runs; negative weights cancel that bias instead. Ask in the app →
Has virtual error cancellation been tested on a real quantum computer?
No. The preprint reports theory and simulations under a simplified noise model only. Ask in the app →
Briefing reviewed by an editor before publication. The paper belongs to its authors; this page summarises it and links to the original. Paper licence: CC BY 4.0. Corrections: support@papersays.com.