Google Quantum AI's Virtual Error Cancellation targets leftover logical errors

Started by BinaryMonk91, Oct 18, 2026, 03:47 AM

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Topic: Google Quantum AI's Virtual Error Cancellation targets leftover logical errors   Views(Read 79 times)

BinaryMonk91

Google Quantum AI researchers have described a new technique called Virtual Error Cancellation, or VEC, and it is a clever bit of thinking. It is a classical step that runs after a fault tolerant circuit has finished. Nothing changes about how the circuit runs on the hardware, and no separate calibration runs are needed. The aim is to strip out the small logical errors that slip past error correction

The trick is to decode every shot twice. One decoder is fast enough for real time use, such as minimum weight perfect matching, which has to meet microsecond deadlines. The other is slower but more accurate, such as a tensor network maximum likelihood decoder that can run offline. When the two agree, the shot gets a positive weight slightly above one. When they disagree, the shot is kept rather than thrown away and given a negative weight tied to the type of error

Averaged over many shots, those weights make the leftover logical error cancel out. The noise settings needed for the weights are worked out from syndrome data the experiment has already recorded, using a method called Walsh-Hadamard linear inversion. That means no extra hardware time is spent learning the noise. Quantum Computing Report points out that simply discarding disagreeing shots introduces its own bias, which VEC avoids

The numbers are striking. In simulations of a distance seven rotated surface code at a physical error rate of 0.005, VEC pushed the residual logical error rate below 7.95 x 10^-10. That is more than 1,600 times lower than plain decoding at about 1.32 x 10^-6, and more than 200 times better than throwing away the disagreeing shots. The sampling overhead was only about 1.3, compared with roughly 4 for standard logical probabilistic error cancellation

It is worth being clear that these are simulation results from a preprint, not a hardware demonstration. Still, the idea fits neatly into the way real machines will work, with a fast decoder keeping up during the run and slower analysis afterwards. If it holds up on a real chip, it could make early fault tolerant machines useful sooner than expected

What do you make of using negative weights instead of postselection? And how soon do you think we will see this tested on real hardware?

sudo train me a model

SpikeDudley05

Over 1,600 times lower than plain decoding sounds almost too good, so I went looking for the catch. The comparison is against matching alone, which is a fairly basic baseline. Even so, beating postselection by more than 200 times is impressive. Both comparisons are fair ways to frame the gain

Numbers like that always deserve a second look. This one seems to hold up

Memory Jamie

An overhead of 1.3 is the figure that jumps out at me. Probabilistic error cancellation has always been held back by how many extra shots it needs. Getting that close to one changes the maths for anyone with a limited budget of machine time. Academic groups in particular will appreciate that

Cloud access is expensive. Every saved shot is money back in the grant

Gareth23

Could this work for algorithms that need mid circuit decisions? Feed forward depends on the real time decoder being right at that moment

Post processing cannot go back and change a branch that was already taken. That seems like a natural limit worth exploring in the paper

Plenty of near term algorithms do not need feed forward, so there is still a lot of room for it. Sampling and estimation tasks would suit it well

Ava12

Pairing a fast decoder with a slow one makes a lot of sense. Nobody can run maximum likelihood decoding in real time at scale yet, but plenty of people can run it overnight. This gives that slower work a clear job. The real time decoder keeps the machine moving, and the careful one cleans up the results afterwards. It is a nice division of labour

Dean_52

Ideas like this are why software keeps pace with hardware in this field. Everyone focuses on qubit counts, but clever classical processing can deliver gains just as large. The quantum and classical sides need each other. VEC is a good example of that partnership

Hardware headlines get the clicks, but software quietly does half the work. Good to see it getting some attention
Press F to pay respects

DarkSideRichard47

Simulation results at p = 0.005 are encouraging, but real hardware noise is messier. Correlated errors, leakage and drift are hard to model fully. My hope is that Google tries this on Willow or its successor before long

Then we will know how much of the gain survives contact with a real chip. Leakage in particular is something simulations often leave out. Still, the method looks flexible enough to adapt

Sorted Echo

Distance seven is a sensible test size. It is large enough to matter and small enough to simulate with tensor networks. Scaling up to distance eleven or beyond will be the next challenge for the offline decoder. Those decoders get expensive quickly as codes grow

TomTiz

Keeping the disagreeing shots is the smart part here. Throwing them away always felt wasteful, and the bias it introduces is a known headache. Giving them negative weights turns a problem into useful information

It is a neat piece of lateral thinking from the Google team. Clever ideas are often the simple ones in hindsight
Always open to a good discussion

Related Topics (1)