Subtitles section Play video Print subtitles [MUSIC PLAYING] JULIAN KELLY: My name is Julian Kelly. I am a quantum hardware team lead in Google AI Quantum. And I'm going to be talking about extracting coherence information from random circuits via something we're calling speckle purity benchmarking. So I want to start by quickly reviewing cross-entropy benchmarking, which we sometimes call XEB. So the way that this works is we are going to decide on some random circuits. These are interleaved layers of randomly chosen single-qubit gates and fixed two-qubit gates. We then take this circuit, and we send a copy of it to the quantum computer. And then we send a copy of it to a simulator. We take this simulator as the ideal distribution of what the quantum evolution should have done. And we compare the probabilities to what the quantum computer did, the measured probabilities. And we expect the ideal probabilities to be sampling from the Porter-Thomas distribution. We can then assign a fidelity for the sequence by comparing the measured and ideal probabilities to figure out how all the quantum computer has done. And, again, I want to emphasize that we are comparing the measured fidelity against the expected unitary evolution of what this circuit should have done. So here's what experimental cross-entropy benchmarking data looks like. So what we're going to do is we're going to take data both over a number of cycles and also random circuit instances. And, when we look at the raw probabilities, we see that it looks pretty random. So what we're going to do is we then take the unitary that we expect to get, and we can then compute the fidelity. And then, out of this random looking data, we see this nice fidelity decay. So the way that this works is that, by choosing depolarizing single-qubit gates, the errors, more or less, add up. And we get this exponential decay. We can fit this and extract an error per cycle for our XEB sequence. And this is error per cycle, given some expected given unitary evolution. So this is fantastic, but we might be interested in understanding where exactly our errors are coming from. So this is where a technique known as purity benchmarking comes in. And, essentially, what we're going to do is we're going to take this XEB sequence, and then we're going to append it with state tomography. So, in state tomography, we run a collection of experiments to extract the density matrix rho. We can then figure out the purity, which is, more or less, the trace of rho squared. You can think of this as, essentially, the length of a block vector squared, an n-dimensional block vector. So what happens if we then look at the length of this block vector as a function of the number of cycles in the sequence? We can look at the decay of it, and that tells us something about incoherent errors that we're adding to the system. So, basically, the block vector will shrink if we have noise or decoherence. So purity benchmarking allows us to figure out the incoherent error per cycle. And the incoherent error is the decoherence error lower bound. And it's the best error that we could possibly get. So this is great. We love this. We use this all the time. It turns out that predicting incoherent error is harder than you think, and it's very nice to have a measurement in situ in the experiment that you're carrying [? about ?] to measure it directly. So, with that, I'm going to tell a little bit of a story. So, last March Meeting, I was sitting in some of the sessions, and I was inspired to try some new type of gate. So I decided to run back to my hotel room and code it up and remote in and get things running. And I took this data, and my experimentalist intuition started tingling. And I was thinking that, wow, the performance of this gate seems really bad. So, when I was looking at the raw XEB data, I noticed that, for a low number of cycles, the data looked quite speckly, and I was pretty happy with the amount of speckliness that was happening down here. But, when I went to a high number of cycles, I noticed these speckles started to wash out. And I found the degree of speckliness to be a little bit disappointing, not so speckly. So I started thinking about this a bit more. And this is pretty interesting because I'm looking at the data, and I'm assessing the quality somehow, but I don't even know what unitary I've been actually performing. And shouldn't the XEB data just be completely random? So I formed an extremely simple hypothesis. Maybe this speckle contrast is actually telling us something about the purity, right? So, in some sense, what I'm saying is that, if there's low contrast, it's telling us that our system is decohered in some way. So I did something very, very simple, which is I just took a cut like this, and I took the variance of that, and I plotted it versus cycle number. And I see this nice exponential decay. So, after this, I got excited, and I went to talk to some of our awesome theory friends like Sergio Boixo and [? Sasha ?] [? Karakov. ?] And they helped set me straight by going through some of the math. So what we're going to do is we're first going to define a rescaled version of purity that looks like this. And you see that, essentially, this is trace of rho squared. And there's some dimension factors that have to do with the size of the Hilbert space. We define it this way so that purity equals 0 corresponds to a completely decohered state. And a purity of 1 corresponds to a completely pure state. We then assume a depolarizing channel model, which is to say that the density matrix is some polarization parameter p times a completely pure state psi plus 1 minus p times the completely decohered uniform distribution 1/D. Then, if you plug these into each other, we can see that these parameters directly relate. The polarization parameter is directly related to the square root of purity, which kind of makes sense because the polarization is telling us how much we're in a pure state versus how much we're in a decohered state. OK, so now let's talk about what it looks like if we're actually going to try and measure this density matrix rho. If we're measuring probabilities from the pure state distribution, we'd expect to get a Porter-Thomas distribution like we talked about before. And what we notice is that there is some variance to this distribution that is greater than 0. However, if we're measuring probabilities from the uniform distribution, which corresponds to a decohered state, we say that there's no variance at all. And, again, this is an ECDF or an integrated histogram. So this corresponds to a delta function with no variance. And it turns out, if you then go back and actually do the math, you can directly relate the purity to the variance of the probability distributions times some dimension factors. So the point is that we can directly relate purity to variance experimentally from measured probabilities. So is this purity? Yes, it turns out. So let's talk about what this looks like in an experiment. So here we took some very dense XEB data, both the number of cycles this way and number of circuit instances that way. And so what I'm going to do is I'm going to draw a cut like this. I'm going to plot the ECDF, the integrated histogram distribution. And we see that it obeys is very nice Porter-Thomas line. And then, if I look over here when we expect the state to have very much decohered, we see that it does look like this decohered uniform distribution like that. So then, if we plug all this data into the handy-dandy formula that we showed on the previous slide, we can then compare it to the purity as measured by tomography. And we see that these curves line up almost exactly. And, indeed, they provide an error lower bound. They're outperforming the XEB fidelity because there can be some, for example, calibration error. So speckle purity agrees with tomography. We get it for free from this raw XEB data. And we are noticing this by basically observing a general signature of quantum behavior. That's what these contrasts mean. So here's a simple analogy. So, if you take some kind of [? neato ?] coherent laser, and you send it through a crazy crystal medium, you can see this speckle pattern showing up on, for example, a wall. But, if you take a kind of boring flashlight, and you point it at something, you just get a very blurred out pattern. So this makes sense, which is to say that, if you have a very decohered state, and you try and perform a complex quantum operation to it, it's really not going to tell you anything. Nothing is going to happen. OK, so let's talk about how to actually understand the cost of these different ways of extracting purity in terms of experimental time. So, typically, when we're doing an XEB experiment, we take N random circuits times N repetitions per circuit. And then we may optionally add some number of additional tomography configurations to figure out how much data we need to take. So, for example, in 2 qubits, we may take 20 random circuits, 1,000 repetitions per circuit. And then, for 2 qubits, we have to add an additional 9 tomography repetitions. And that means we're taking 200,000 data points to extract a single data point here, which is pretty expensive. If you look at the order of scaling, the number of circuits in XEB we'd typically pick to be about constant. The number of repetitions it turns that scales exponentially due to the size of the Hilbert space. But then we also have this additional exponential factor doing full state tomography. So what we're taking away from this is that full state tomography really is overkill for extracting the purity. We can get it just from the speckles with the same [? information, ?] [? only ?] we're doing it exponentially cheaper. We're actually getting rid of this 3 to the N factor. And I do want to point out that there's still this exponential factor that remains due to having to figure out the full distribution of probabilities. So I want to now actually show data for scaling the number of qubits up to larger system sizes. So we can, for example, extend the XEB protocol to many qubits in a pattern that looks something like this. And we can measure for 3 qubits, 5, 7, or all the way up to 10 qubits to directly extract a purity. And we see that this still works. We see that we get nice numbers. We get a nice decays out of this. And we can also then compare to the XEB directly. I want to point out that, going all the way up to here, we use kind of standard numbers for the number of sequences, the number of stats, but then, at this point, we started to feel the exponential number of samples that we needed. We had to crank the repetitions up to 20,000, but even that is really not that much for extracting purity for a 10 qubit system. OK, so now the last thing that we can do is we can actually do some pretty cool error budgeting.