SPEAKER_02 0:11

144. University of Southern California. Welcome to Entangled Things, your quantum computing podcast, hosted by Patrick and Cyprian.

SPEAKER_01 0:40

Hey Cyprian, how you doing? Hey, Patrick. I'm doing great. Looking forward for another episode of Entangle Things.

SPEAKER_04 0:47

Oh, well, you're in for a treat because we're we're joined by Todd, who's been here before, but we'll still ask you to introduce yourself to our audience, please, Todd.

SPEAKER_00 0:55

Hi, I'm Todd Brunn. I'm a professor of electrical and computer engineering, physics and computer science at the University of Southern California. I work on quantum computers.

SPEAKER_04 1:09

And that's why we want to talk to you. Also, you're very easy to talk to. So we've been on the show before. You are very deep into error correction. I think just before we started recording, you mentioned that you're going to a conference soon on error correction. It's exciting times for that field right now, isn't it?

SPEAKER_00 1:26

Yeah, well, uh we've arrived at a point where it's not just theory anymore. Uh we've been working on error correction really since very early days of quantum computing, because uh people realized early on you needed to be able to do it if they were ever going to be achievable. Right. But for a long time, it was really theoretical. We proved a lot of sort of results about scaling and things like that. But now people are really building these things. And so that means we have to focus on what you can actually do with the machines we have and the machines we're going to have in the next few years.

SPEAKER_04 2:09

Do you do you think we're ending the NISC or entering the NISC period, noisy intermediate quantum computing? What excites you about the developments? Because I can't when I see a feed, almost every day there's something about you know re fault tolerance and redundancy and error correction. It it's it's permeating the news. Aaron Powell Yeah.

SPEAKER_00 2:30

Well, we're still in the NISC era, but we're entering uh I think the term that people have kind of converged on is the early fault tolerance era. Uh fault tolerance is uh sort of full fault tolerance or scalability is pretty demanding. And so the machines that exist right now can only achieve it up to a certain point. Sort of all the elements are there, but we're still not quite where we need to be to be able to scale to larger and larger sizes. But what we would like to be able to do is use some of the elements of fault tolerance, some error correction or detection, some uh methods for reducing the effects of noise, and apply them to actual programs that can be run that would hopefully do actually useful things. Uh so that's where we're just sort of starting to be in that era. And the goal there would be to be able to incorporate more and more fault-tolerant elements as the capabilities of the machines improve to the point where we we arrive at fully fault-tolerant quantum computers.

SPEAKER_04 3:50

Are you excited about one or two modalities more than others in this? Like, is are any of them running away with the ball?

SPEAKER_00 3:58

I love all my children equally.

SPEAKER_04 4:01

Um We have to say as parents, I know.

SPEAKER_00 4:06

Um It certainly is the case that some of them are further along in some sense than others, but this can change unpredictably. So it's still the case that the superconducting modality is the most advanced, that's the one that has the most investment from industry. They've built the biggest processors so far and have the most impressive results. But other modalities are still very much in the running. Ion traps, which have been around since the very early days. I think you did a podcast on them not long ago. And uh there's some companies.

SPEAKER_03 4:51

Glad to know you're listening.

SPEAKER_00 4:53

Ah some companies working to develop that as well. And then uh a third contender has really advanced rapidly in recent years, and that's neutral atoms. Uh arrays of neutral atoms held in optical lattices and controlled by what they call optical tweezers, so you can actually move the atoms around and reconfigure them. That was sort of nowhere, maybe three, four years ago, or I mean not nowhere, but not really a serious contender. And now it's grown to where they have systems with hundreds of atoms in them that they can manipulate with great control. So who knows? That may overtake everything else.

SPEAKER_04 5:39

Yeah, we we talked to the Harvard team that uh I think it was a year this December last December, a year ago last December, they they created a 48 logical qubit system that made the news. Uh and they were using that. The the light laser tweezers, I think of little mini lightsabers on a chip um used to move things. Isn't it funny that in Star Wars a lightsaber can cut through anything except another laser?

SPEAKER_00 6:04

Yeah. I mean they they can invent how it works to make the story good.

SPEAKER_04 6:11

That's true. That's true. Well, we can't. We don't have that. We have to actually deal with this thing called physics.

SPEAKER_00 6:16

Sadly, sadly.

SPEAKER_04 6:17

Yeah. Um so what should we be watching out for?

SPEAKER_00 6:21

Well, uh, I think the there are two things that are very exciting right now. One is this early fault tolerant era. So there have been a number of experiments now. For a long time, there were no experiments at all that demonstrated any advantage to doing error correction because the hardware was just too noisy for error correction to benefit you. And then for a while, there was just one, which is from a group at Yale in the superconducting area, where they actually didn't encode things in the in their qubits at all. They encoded things in the microwave cavity that was normally used to control the qubits. They were, they turned that around, so the information was in the cavity and they were controlling it using the qubits. And they just barely showed an advantage. This is against real noise that occurs in the system. And for a while that was it. And now there are a bunch of experiments that have demonstrated advantage, and they keep getting better. Um, but so so we're gonna see more of that. Uh, that's continuing, and that's very exciting. But of course, error correction is not the point. The point is computation. So we really want to use error correction to do computations that we couldn't otherwise do. So there, we aren't quite there yet, but I think we're going to increasingly see that over the next few years.

SPEAKER_01 7:57

So I I saw some announcements uh in in recent time, right, regarding these types of uh demonstrations, right? I think QR had one with the uh uh MIT uh MIT team. That was a very interesting one. They were, if I'm not mistaken, they were claiming like a two to one ratio um from from physical to to logical. What would you say is at the moment, uh like the place where we are? Last time we talked, we were just about to cross the boundary where error correction actually yields like a positive outcome. Um where do you think we are now? Um is are the improvements in error correction mostly coming from better hardware? Are we doing some some other kind of improvements in terms of, let's say, novel approaches to do error correction? What are like the uh let's say the dimensions of the advancements that we're we're we're seeing? I think it would be very interesting for our listeners to kind of like understand some of the driving forces there.

SPEAKER_00 9:13

I think the advances are coming both in the hardware, which is of course absolutely necessary, but also on the theory side in exploring new kinds of codes, new kinds of fault-tolerant techniques. Not all of that has been implemented in hardware yet, but some of it has, and and we're getting closer. And uh I think you need both. In the very early days when there were no experimental systems with more than you know, maybe a couple of qubits, uh, the theory largely was concentrated on what you can do in principle. You know, can you, in principle, do computations of an unlimited size provided the noise is sufficiently low? So people would prove theorems about this, they're threshold theorems. If the noise is below some threshold, then you can scale up your computations to any size you want. And the overhead in terms of error correction for doing that scales nicely with the size of the computation. So that was the early days. And it's good to know that because it meant this wasn't an exercise in futility that that if we could build good enough hardware, then we could actually use these things to do something. Now the machines have crossed that threshold in the sense that when you use error correcting codes, you can show that the data that you're storing in these machines is protected from noise. It's better than not using codes. And they've actually gone further than that. They've shown that encoding it in more powerful codes, what we call higher distance codes, protects it more than lower distance codes. So when you're below the threshold, that's what you would expect. When you're above, though, it actually just makes everything worse because you you increase your overhead, the total amount of noise goes up. So it's that trade-off that you need to be on the right side of. You when you scale up, you increase the noise, but you also increase the power of the code to correct the noise. And if you increase the one more than the other, then you either win or you lose. And uh now we're on the winning side. But the overhead is such that after you do all the encoding, the the number of logical qubits is is still fairly small because these processors are not that big. So they're still small enough that we can simulate everything just with an ordinary classical computer. So we haven't quite gone to the next step, which is where the processors are big enough and the noise level is low enough, and the codes are efficient enough that we can do a computation at the logical level using logical encoded qubits that we couldn't simulate uh with a classical supercomputer, for example. So we're we're close to the edge of that as well. There are some demonstrations that you could say can't be simulated or can't easily be simulated classically. They're mostly problems that no one has any interest in solving except as a demonstration that it can be done. So the next stage is to get to interesting problems, problems people would actually like to be able to solve, uh, perhaps simulations of chemical reactions or nuclear reactions, quantum field theories. And recently there's been some indication that new theoretical ideas and error correction and more efficient encodings may bring Schore's algorithm, which got the whole field running in the beginning, uh, the factoring algorithm that can be used to break public key crypto systems, that that may be breakable in the not too distant future. There was a paper from Google and a second one that I should have looked up before I went on this podcast to remember who who wrote that second paper. But they both made the argument that if you go to more efficient codes, that you may start being able to break the kinds of crypto systems that people are using now in maybe 10 years. And uh that's that's much sooner than many people thought.

SPEAKER_01 14:10

Or were prepared for.

SPEAKER_00 14:13

Or were prepared for. I I always thought that was short-sighted. Yeah, well, of course, a lot of this kind of public key encryption is used for very short-term things, and uh and people don't care. You know, you send your credit card number, uh, that credit card number probably won't be valid anymore in 10 years, um, things like that. But uh there is stuff that people don't want to be read even 10 years from now. Right. And uh nuclear silo locations and designs for weapons and I hope not, but maybe they're not supposed to send that kind of stuff over the internet. But you know, mistakes happen. And there's lots of, you know, businesses have secret uh stuff and private information. And then there's cryptocurrencies, which rely on uh public key encryption, basically. So they're vulnerable now.

SPEAKER_04 15:13

I've heard some of the ECCs are particularly vulnerable and there's concern that it they could become catastrophically vulnerable within as as few as two years.

SPEAKER_00 15:23

Uh I don't know if it could be as soon as that. It's not really my area, but uh yeah, I mean it's much sooner than than people had been hoping for. Yeah.

SPEAKER_04 15:37

We we've recently talked to about um Microsoft's uh endeavors, and they're claiming um that their topology inherently has fault tolerant properties. Are you following that at all? Is that is that is that one of your children?

SPEAKER_00 15:56

I mean that's sort of a they're all my children. They're all my um yeah, the uh but uh there are there are a bunch of different approaches people are pursuing that that look very promising and and uh I don't know which of those will will end up winning in the long run either.

SPEAKER_03 16:17

Yeah.

SPEAKER_00 16:17

I'm pursuing my own personal research topics. Oh, nice. But but there are you know it would be nice if one of them or some some of the elements of what I'm working on end up being used in in whatever the final results are.

SPEAKER_04 16:33

We we don't believe there'll be one to rule them all. We I both both Cyprian and I think that there's going to be room for multiple modalities to solve different states of problems in the future, whether it's sensing or computation or whatever.

SPEAKER_00 16:47

That very likely is true. Certainly when you have very different applications like sensing and computation, it would make sense.

SPEAKER_04 16:53

You might use photonics. Photonics wings wins the network with ease.

SPEAKER_00 16:59

Right.

SPEAKER_04 17:00

Just because of the nature of it.

SPEAKER_00 17:01

So something that's still uh a bit of a challenge is interconverting between different modalities. And this is something now that that has been attracting increasing interest is so-called hybrid systems, where you have some qubits of one physical type and some qubits of another physical type, and you try to leverage the advantages of the different kinds and make them work together. But converting from one to another is still technically quite challenging. This is called transduction. And uh uh there are reasons why that's not easy. They they operate on very different timescales, so like atomic and ion qubits may have time scales in the in the megahertz while superconducting or in the gigahertz, so that's a big gap. And uh, you know, coupling things that operate using optical photons to things that operate using microwave photons, there's a huge energy gap between those two. They operate at different temperatures. So not easy, but uh it's a valuable thing to be able to do in principle. So people are increasingly also working on that problem.

SPEAKER_01 18:28

Um one of the things that one of the areas where I've also seen some interesting results lately is improvement uh when it comes to error correction, improvement in the decoding latency. Um where where where do you think we are with from from that point of view? Because I believe that's also extremely important, right, in terms of you could have the the best and the most powerful code ever, right? But if it's gonna uh produce like a terrible slowdown of the operations, um it's it's gonna potentially be almost useless.

SPEAKER_00 19:04

Yeah, so that's an excellent point. And in fact, that is one of the practical issues. When we talk about how good codes are, we're usually talking about their intrinsic properties, like their rate, how many physical qubits per logical qubit, how many errors they can correct, their distance, um sometimes details of what you would need to do to measure what we call the error syndrome that tells you what error had happened. But the decoding problem is also very important because in principle there are lots and lots of great codes, but they're useless to us because decoding them is a computationally hard problem. And so you couldn't use them in practice. So we need codes that are good enough, but that also have efficient decoding algorithms. And this latency issue is the reason. Right now, all the demonstrations of quantum error correction, pretty much, they encode things, they do their computation, whatever it is. Sometimes it's just storing it and protecting it from noise. Sometimes they actually try to do some encoded computation. And then they measure everything and they they do the decoding in post-processing, right? They don't do it in real time. But for a practical quantum computer that can run big computations, you need to be able to measure what errors are happening right now, figure out what the appropriate correction is, and apply that back onto the code while the computation is ongoing. Right. And so the time it takes to run the decoding algorithm is key in that. And also just the communication time. Do you have to take that information outside of your quantum computer to some classical computer in the outside world, run a program there to figure it out, and then send the information back to the quantum processor about what correction it should do. If that takes too long, then you'll fall behind the errors. And even if your code in principle should be able to correct all the errors, in practice you can you can't.

SPEAKER_03 21:32

You'll be overwhelmed.

SPEAKER_00 21:34

Yeah, exactly. So that's also something that people are working on. And that's that's also an area where hardware improvements are key. Right now for the first few generations of the quantum processors that like IBM and Google and so forth are building, they didn't have the capability to measure and feedback. during the computation really. They couldn't do that at all. Now they can, so they've made that advance, but it's fairly slow. And uh we need to get to the point where they can do that and do it quickly enough that they can uh correct errors as they happen during a computation. And it may be that the best code in practice isn't the most powerful code, but the code where we can do that. And also as far as decoders, there are there are ideal decoders that sort of find the optimal correction. We may be better off with a fast but dirty decoder that works most of the time and uh just build in a little redundancy in other ways to to compensate for its imperfection. But there's actually a a great uh precedent in classical computing the development of turbocodes and and low density parity check codes which happened I think mainly in the 1990s once classical computers were fast enough that they could run decoding algorithms uh quickly enough people started using these codes because there were approximate decoding algorithms that theoretically you couldn't guarantee that they would work but it turned out that in practice they do.

SPEAKER_01 23:36

It was just good enough.

SPEAKER_00 23:38

It was good enough and they could use them on this big class of codes that that were very effective. And so there was a big leap uh there used to be a joke in coding theory um claude Shannon the the the creator of information theory proved that you could do error correction and uh achieve communication rates he's the one who who discovered the formula for the capacity of a channel that is still the one that everyone uses and he proved it using what are called random codes um which means that if you just randomly generate codes, they're mostly pretty good. In principle, but he was proving what you can do in principle. But in practice those codes are useless because the decoding problem is too hard. So the joke in the field was almost all codes are good codes except for the ones we know and and of course it wasn't so much the ones we know as the ones we know how to decode efficiently. Right. So in the quantum case uh so that was all mainly for for uh for communication and uh computation just by making the hardware more and more reliable they managed to avoid having to do really onerous coding and decoding in the classical case but in the quantum case we need it. So we're we're trying to steal ideas from communication in the classical case and apply them to computation in the quantum case and having fast efficient decoders is absolutely key to that.

SPEAKER_04 25:32

So I'm I I want to take the opportunity and if this is not a valid question that I understand that. Is there something that you wish everyone knew about error correction that you have to beat out of your students or or or get them to stop thinking about it a certain way or start thinking it a certain way is there one bad one cardinal bad habit or misconception that you'd like to debunk? I try not to beat my students it's frowned on but uh I was in the military so it was it was required.

SPEAKER_00 26:06

Oh I see yeah no I I mean we we could do the whole you know the beatings will continue until morale improves.

SPEAKER_04 26:15

It works for me. There was a staple at West Point.

SPEAKER_00 26:20

Ah all right that's where I went I I guess um maybe I'm I'm lucky that I I didn't go to West Point I I'm very delicate. Let's uh I mean there are some misconceptions but I think the the most unintuitive thing about quantum coding is that you can figure out what the errors are and correct them without just measuring everything. Because classically of course it doesn't matter when you measure things it doesn't hurt they don't change yeah but quantum mechanically they they do change and so you want to you want to design your codes so that you can measure just what you want to know which is what errors happen without sort of going over the line and measuring the data that you're storing. And the fact that that's possible at all is not obvious. So the the early developers of quantum error correcting codes were very insightful when they realized that that was possible.

SPEAKER_04 27:35

Yeah the analogy I think of is like I'm tracking an animal I'm not allowed to look at like the basilisk or the hot or the Medusa right but I can look at their tracks.

SPEAKER_00 27:44

Yeah.

SPEAKER_04 27:45

And so I'm watching their tracks and I can see the way they've come without turning the stone. I don't know if that's a good analogy or not.

SPEAKER_00 27:53

No it's an excellent analogy yeah so you know uh the key to hunting a basilisk well first don't do it at all but second um if you must bring a big big gun and blindfold you know if you if you have a big enough uh blast shotgun yeah yeah but uh yeah you you you need to look but not too closely and uh and that's the same thing with quantum error correcting codes though fortunately you don't turn to stone as far as I know so what are the tracks then in that analogy that I've used?

SPEAKER_04 28:30

Is is it is it their are you looking I I guess I don't are you looking at temperatures are you looking at I I guess it depends on the modality of course.

SPEAKER_00 28:40

Yeah but all the codes work in a somewhat similar way I mean and and actually classical codes work this way too though classically you know this isn't an issue so people didn't really worry about it. But in a code you're spreading the information that you're trying to protect redundantly over a larger number of bits or qubits in the quantum case. And classically you would just go in and you'd look at all the bits and you'd say oh this isn't a valid codeword so some error must have happened. So I will figure out what codeword it most likely you'd look at parity and things like that. Right. Well that's the thing the parodies are used to define the code so so that's how you know whether it's a valid codeword or not you look at these parodies the you know whether in different subgroups of your bits whether you have an even or an odd number of ones versus zeros. So uh but classically you can calculate that just by looking at each of the bits and counting the number of turning to stone and and you don't turn to stone which is which is lucky um given the amount of time everyone stares at their phones, we'd all you know the world would be filled with statues seems like it that might have happened. Yeah well a good point um but quantum mechanically one of the key insights was you could measure the parodies of bits without actually knowing the values of the individual bits. So for instance if I have two bits uh then we'd say the parity is is even or the parity is zero. If both bits are the same, if they're both zero or both one or if I have three bits if I have if two of them are ones or none of them are ones then those would all be even parity. So classically you just you look at the bits and you say how many of them are ones and you calculate the parity. But quantum mechanically if you do that then you collapse the wave function you've collapsed the wave function you've gone too far you've looked at the state and your computation turns to stone. But uh but you don't the the remarkable thing is you don't have to do that. You can actually measure the parodies without measuring the values of the individual bits which is not at all obvious. So I would call that a very unintuitive thing as well but it's related to the the same one I was saying before. So that's what you have to do you have to look but not too closely and that's how you do it. You measure these parodies but you don't measure the individual qubits.

SPEAKER_04 31:30

That's fascinating and I and very appreciated because these are the things that you work very hard and it's those epiphanies like I Cyprian shared very early on in our conversations that he was trying to understand quantum and it evaded him which meant he was on the right track. It was only when he let go of understanding it in a cognitive sense that it all made sense. And you actually use the math you also dove into the math in order to understand it.

SPEAKER_01 32:00

When you act when you resort to math right in the that's right yeah you're in you're either doing on the right track or you're in serious trouble when you resort to math so what one of the things that I wanted to also ask you Todd is uh of course besides the inherent stability of the of the qubits right in the in the hardware like what else can the hardware itself do to help error correction?

SPEAKER_00 32:26

Are there any kind of capabilities that are maybe modality specific that could help or it's just the the quest to get more stable qubits and the rest of it happens essentially at a layer that's that's that's above is there anything else that can be uh can help the the problem of error correction there yeah there there are a lot of things that affect it so the intrinsic error rate of the individual qubits and and also of the the operations the gates that we do on them that's number one of course and uh they all have to have a very low rate of noise but other things are important too if you want to measure these parities without measuring the individual qubits it's very hard to do that if the qubits are physically far apart from each other. So these qubits are they they actually have locations in space for for uh superconducting their little devices etched onto a chip for ion traps their individual ions at a particular location in the trap for neutral atoms their atoms at a particular location in the lattice and if the the parity that you're measuring is of qubits that are far apart from each other it's very difficult to measure that without measuring the individual qubits so you can in some cases get around that by physically moving them to be close to each other. So ion traps can do that and uh and and neutral atoms can do that. But superconducting qubits they're they're etched on the chip the qubits are where they are and and then there's the question of how long range can you connect things together. So um with with solid state implementations like superconducting and semiconducting qubits they generally interact with the the qubits around them physically so making things interact that are far apart is more challenging. So you have to find ways around that issue. So that affects the kinds of codes you can use because different codes have different demands for which subsets of the qubits you need to measure parities of and how local they are, how easily you can lay them out on say a 2D surface or even in three dimensions. And so so that affects which codes you can use how easy it is to to do these measurements that you need to do to figure out what the errors are. So from that point of view um ion traps and and neutral atoms are are more flexible because you can physically move the qubits they're slow however right so the operations are relatively slow compared to superconducting qubits where the operations are really fast you know 10 nanoseconds or something like that.

SPEAKER_04 35:45

So there's a trade-offs yeah there's a lot of trade-offs and that's why we still don't really have a single uh contender so you make error correcting very fun and much less painful than when I had at school um we we've we've we've been talking for a while i think I suspect we could talk for much longer but but as we start to reel in on time is there anything else you want to let people know anything you're doing that you want to highlight um we really do always appreciate your time well I appreciate that you're making this uh subject accessible to a larger number of people and and so people can appreciate the technical difficulties of what we're doing but also the promise.

SPEAKER_00 36:37

I'll just say I mean I've been working on error correction and fault tolerance I'm one of many people working in this area where we're all trying to come up with ideas to make this problem easier and make the ultimate goal of useful quantum computers closer. But there are a lot of other elements to building a quantum computer and of course for the applications we we know some things quantum computers are good for but we also would like to find new ones new new algorithms new applications and so that's a very active area as well that probably should be more active than it is um so it it's become a big enterprise in the very early days so I first learned about quantum computing in the in the early 90s and it wasn't a field at all. It was just an idea an idea a few people were kicking around and now I have no idea how many people are working in this field certainly thousands maybe tens of thousands it's extraordinary to have seen that growth and you kind of can sometimes feel a little like a cog in the machine but that's modern technology and it shows that we're a lot closer to quantum computing being a real technology not just an idea.

SPEAKER_04 38:10

I read recently that in Europe in just before the printing press was invented there were only 30,000 books in all of Europe. And so I feel like we're the people making the printing press you know in in some way. I mean we're reporting on it more than you're making it but uh we do appreciate you know what you're doing.

SPEAKER_00 38:29

Thank you very much and I appreciate that because if you weren't doing that then then uh people would wonder you know where where where this money is going and why why it's an important thing to invest in. So I appreciate that. And yeah I mean you go from handwritten books where one book costs the same as your house to to now where we're in the information age and there's more information available to everyone than they could read in a hundred lifetimes. Right it's extraordinary to be part of it.

SPEAKER_04 39:07

Well we're hoping you'll be back on on continue to be a regular here. We really do appreciate your time and and thanks for joining us. Thank you so much.

SPEAKER_02 39:15

Thank you it's been a real pleasure all right bye everybody see you next time everyone bye cybercrime is one of the biggest threats to businesses of all sizes and industries with almost half a million open cyber positions the problem is compounded by the lack of available talent in the marketplace. At Pulsar Security our elite team of highly credentialed experts collaborate with you to assess your current defenses and develop solutions tailored to your specific needs. With services ranging from cybersecurity education to advanced penetration testing and red teaming you can start reducing your risks today. Visit pulsarsecurity dot com and let's secure your digital future together