Sebastian Hassinger: 00:30

Welcome to another episode of the New Quantum Era. I'm your host, Sebastian Hassinger. Today's conversation is with Daniel Lidar, a professor at the University of Southern California and another pioneering researcher in the field of quantum computing. I first met Daniel around 2018 when I was creating the researcher access program at IBM Quantum and Daniel proposed a very interesting experiment involving random circuit sampling and dynamical decoupling on IBM's fifty four qubit device, which was not yet launched at that point. As new as I was to the field, I was still struck by Daniel's novel and creative thinking about mitigating noise and probing the limits of what the hardware could do.

Sebastian Hassinger: 01:14

Dynamical decoupling is really fascinating because it's injecting non logical gate pulses in between the logical gate operations, and those pulses have the effect of stabilizing and mitigating the noise that can build up in an idle qubit. So that was sort of an introduction to that topic from Daniel, and I found it really interesting. His career goes back to a very common starting point for many quantum computing experts, the publishing of Shor's algorithm in 1995. And he's been involved in many important explorations and innovations since then. We met and recorded this interview at the APS Summit, with the kind assistance of the American Physical Society.

Sebastian Hassinger: 02:02

Daniel takes me through some of his own origin story and experiences with quantum annealing, quantum error correction and his new efforts to create accurate models of physical qubits powered by GPUs with the startup he co founded Quantum Elements. I really enjoyed catching up with Daniel and I hope you find our conversation as fascinating as I can. Hi, Daniel. Thanks so much for joining us.

Daniel Lidar: 02:59

Hi, Sebastian. It's great to be here.

Sebastian Hassinger: 03:01

So, we've known each other for a while professionally. First, when I was working at IBM and now at AWS. You are, you've been involved in quantum computing for quite some time. Your your roots go back to very early days of the field.

Sebastian Hassinger: 03:19

How did you first sort

Daniel Lidar: 03:20

of get involved? Yeah. Well, I got started towards the, middle of my PhD back in 1995. So it's been thirty years Yeah. Unbelievably.

Daniel Lidar: 03:34

I was working on scattering theory and fractals, and I was getting a little disillusioned with those areas. And then I had the good fortune of running into somebody who I believe was a former podcast guest of yours, Dori Taranov.

Sebastian Hassinger: 03:50

Oh, yeah.

Daniel Lidar: 03:51

Yeah. We were grad students at the same time at HebrewU. And she had just started her PhD in the area. And she told me about this amazing algorithm called Shor's Angle. Right.

Daniel Lidar: 04:03

And I

Sebastian Hassinger: 04:03

She told me the story of her adviser saying, read this paper or do something with it.

Daniel Lidar: 04:08

Well, she certainly got me excited and interested, and I decided to make a big switch and never looked back.

Sebastian Hassinger: 04:15

That's amazing. That's amazing. And so from from '95 is quite a fortuitous time to get involved in in quantum computing, although it it require a lot of patience to stick with it. I I assume sort of was your initial interest in in sort of the the, exploration, the algorithms, that might provide some kind of quantum advantage from from from a theoretical point of view?

Daniel Lidar: 04:41

That's right. Yeah. So back in those days, I wanted to look at whether there were other algorithms that could provide quantum advantage. And my very first paper ever was on simulating Isaac models, classical Isaac models. We actually found a very modest type of quantum advantage in that context.

Daniel Lidar: 05:02

And that led to a postdoc at Berkeley. At the time, it was rare to find any postdocs in quantum computing.

Sebastian Hassinger: 05:09

Right.

Daniel Lidar: 05:09

I found one in the chemistry department at Berkeley, which was a great place to be because Berkeley had Alex Pines who was doing pioneering experiments in NMR, quantum computing. And Birgit Whaley was my postdoc advisor, theoretical chemist by the name of Bill Miller was my second postdoc advisor. And that's the point at which I got interested in in error correction and error suppression and started working on decoherence free subspaces.

Sebastian Hassinger: 05:38

Interesting. And so, that ICEE model, that was that the connection to D Wave? When D Wave came out, with their first machine, their their quantum annealing devices is is a simulation of an Ising model. Right? That's sort of the fundamental, whatever, capability of the device?

Daniel Lidar: 05:58

Yeah. So, actually, the connection I had with D Wave happened, several years later, and it's interesting. You should ask if it's connected to that old Ising model work that ever actually never occurred to

Sebastian Hassinger: 06:09

me. Oh, really? I

Daniel Lidar: 06:12

I ran into the founder of of D Wave when I was in Toronto. I got my first faculty position there after three years at Berkeley. And, Canada was, a place where a lot of really interesting quantum computing work was was happening. The Perimeter Institute was just starting up at that time. They were in their old building that had an amazing coffee machine in a very interesting architecture looking building.

Daniel Lidar: 06:43

Jordi and I, he met accidentally, and he told me about what he was up to setting up this new company, that was interested in in quantum annealing. So that's how that got started.

Sebastian Hassinger: 06:54

Okay. So it's serendipitous, and and the the topic of your prior research never came up. That's interesting. But, I mean, that that is it is an icing model simulation that's sort of the bread and butter of a quantum annealing device. Right?

Daniel Lidar: 07:09

Yeah. A quantum annealing device, as we have them now, those are devices that are basically transverse field Ising model type, simulators. The work that I I was referring to that I did, during my PhD was really about a classical Ising model. Okay. And it is related because you could think about what an annealer does is trying to find the ground state or a good approximation to the ground state of a classical Ising model.

Sebastian Hassinger: 07:36

Right.

Daniel Lidar: 07:37

But the way it gets there is actually very interesting, because it involves a quantum path. You start from a strong field along the x direction Mhmm. And that prepares a a uniform superposition of all spins or all the qubits. And then you slowly turn off that transist field, and at the same time, you turn on the Ising Hamiltonian whose ground state you're interested in. And the path from that uniform superposition to the ground state of the Ising model is fairly arbitrary.

Daniel Lidar: 08:08

There's a lot of inherent robustness in the annealing approach in in in that regard. But that path necessarily takes you through a quantum critical point where the energy gap gets very small. And that determines how fast you can run your calculation. It's important not to excite out of the ground state. So the speed at which you can solve a quantum annealing problem is really set by the minimum quantum gap you encounter along this path from Okay.

Daniel Lidar: 08:39

Uniform field to the finalizing model. But the trajectory itself, involves a lot of interesting quantum dynamics. And if you go a little too fast and you get excitations, then, there are interesting physics to to be observed. For example, Kibble Zurich mechanism is is one popular area of study in that regard. So there's a lot of quantum effects that are at play in trying to find the ground state of, the finalizing model.

Sebastian Hassinger: 09:04

Mhmm. Okay. And, I mean, the you were involved in in some early research papers when they shipped their first device. Is that right?

Daniel Lidar: 09:14

Yes. So this actually fast forward, to 02/2010 or '11 Right. I had, by then, already moved from Toronto to USC. I moved there in in 02/2005 and had not stayed in touch with with D Wave. But serendipitously, I had a visit from somebody by the name of Ned Allen, who was the the chief scientist of Skunk Works at Lockheed Martin Okay.

Daniel Lidar: 09:44

Who

Sebastian Hassinger: 09:44

The name was familiar. Right? The Skunk Works. Okay. Cool.

Daniel Lidar: 09:47

Yeah. And so Ned walked into my office and started telling me about a problem that Lockheed was encountering with the f 35 fighter jet. I told him that I I knew very little about fighter jets, but he persisted. And he told me that the problem was one of essentially software debugging. And they were spending a lot of money on on this.

Daniel Lidar: 10:11

And the problem was getting worse and worse because as is the nature with these very complicated machines, you get millions of lines of code and Right. And and and that just keeps growing, and and the debugging problem gets harder and harder. And he wanted to know whether there was anything in the quantum domain that could help them with software debugging. Mind you again, this was 02/2010.

Sebastian Hassinger: 10:31

Yeah.

Daniel Lidar: 10:32

So I told him I didn't think so, but that if you wanted to work with any kind of quantum hardware manufacturer, at that time, it was basically just one game in town, and that was D Wave. Right. So he he got interested, and we both, took a trip to to Vancouver. And long story short, what happened was that Lockheed decided to invest in buying the very first commercial quantum computer, which was the D Wave that ended up, at USC in in 2011. And from my perspective, the goal was to try to understand the claims that were circulating around that time about whether a D Wave was actually performing some type of quantum computation.

Daniel Lidar: 11:17

And there's no black or white there. It's always a bit of a gray area when there's decoherence. And we knew that there was a lot of decoherence in those devices. But I wanted to understand to what extent we could quantify the quantumness. Mhmm.

Daniel Lidar: 11:34

And also, to see if we could, introduce error correction techniques into this area, which we did, and to look at algorithms that we could run on these devices.

Sebastian Hassinger: 11:45

Sorry to interrupt, but error correction on annealing, I guess I was thinking that error correction was directly associated with gait operations and an annealer is is an analog device. Is there what would the the flavor of an air correction be in an analog, an annealing device? Yeah.

Daniel Lidar: 12:04

That's right. And so that was a bit of a dilemma at the time. How do we even approach error correction in the first place? But fortunately, there were very relevant ideas, from the field of topological quantum computing, actually. So the the basic idea is that if you can create an energy penalty against, errors that, in this case, excites you out of the ground state, then this forms a type of properly, it's really error suppression, not quite error correction.

Daniel Lidar: 12:36

So the first idea was let's suppress, errors that create, excitations. But in addition, there are some simple codes that could be applied as well. Mhmm. Because what you can do is and this is a little technical, but you can take a Hamiltonian, and in this case, transistorizing Hamiltonian, and you can replace every spin operator with the logical operator of some error correcting code. And as soon as you do that, now you have the ability to use that code as an error detecting code Okay.

Sebastian Hassinger: 13:14

At the

Daniel Lidar: 13:15

very least, even if it's not an error correcting code. So strictly speaking, what what we did was was in sense in a sense that replacing the the Bayer Ising model Hamiltonian, transistor fuel Ising Hamiltonian with the encoded version with with respect to some code, particularly, it was actually just a simple repetition code Okay. That was, a good fit for these devices. And then, you get the benefit of, this energy gap protection that I mentioned, energy penalty, against excitation errors, as well as the ability to detect any errors at the end of a run and then, essentially perform post selection on the I

Sebastian Hassinger: 13:59

see. Interesting. That's really fascinating. So did did the D Wave help Lockheed with the debugging problem?

Daniel Lidar: 14:08

Not as far as I know. It was understood pretty early on that this this was not gonna be a a good application here. But Lockheed remained invested in in this area for for quite a long time. And I actually credit the whole birth of commercial quantum computing with the vision that Ned Allen had at the time. Mhmm.

Daniel Lidar: 14:31

Because in in twenty ten, twenty eleven, no commercial quantum computers were were sold anywhere. So this sale that that happened actually triggered a lot of subsequent activity. Interesting. Google, two years later, bought a D Wave machine that they put at NASA, and I was actually somewhat involved in in that story. I had the chance to talk to, the head of NASA at the time, and make the case that this this could be an interesting thing for them to explore.

Daniel Lidar: 15:04

And if you look at the work that that Google did in quantum computing, it actually all started with their work on quantum kneeling.

Sebastian Hassinger: 15:11

Absolutely.

Daniel Lidar: 15:11

And then later, they decided to that they could build a a gate based machine in house. But, yeah, those early days in in the early twenty tens, were the very beginning of of commercial quantum computing. Absolutely.

Sebastian Hassinger: 15:27

Yeah. That's so interesting. There's there's, I think, a general sort of I wouldn't skepticism is too strong a word, but but feeling that that if D Wave's devices had advantage, that, there would be more, proof, more more evidence of that at this stage. That in some sense, I think annealing and and analog operations in general, analog quantum operations in general, are sort of discounted as being too hard to to, control, or they don't map as well to to practical problems as we thought, or but you still think that there is significant value in the operation that you can support on those devices?

Daniel Lidar: 16:14

I do believe that. I think there are multiple paths towards practical applications of of quantum computing. At this point, really, nobody knows for sure

Sebastian Hassinger: 16:24

That's true. What the ultimate winning like the naysayers for quantum annealing have a really good, like, counterpoint. Like, oh, here's the thing. Like, it's all still speculative to a certain degree.

Daniel Lidar: 16:35

Yeah. I there there's a lot of speculation. There's also a lot of of great work that establishes, rigorous results. Yeah. And perhaps the most significant critique of quantum annealing is that there is no analog to the quantum fault tolerance result in in quantum But in terms of hardcore results from from quantum annealing delivering quantum advantage, it it actually is the case that there are some some really interesting breakthroughs Mhmm.

Daniel Lidar: 17:06

As of late. So for example, D Wave themselves published a paper not long ago, just a few days ago actually, on a computational quantum advantage in terms of quantum simulation. They got some pushback. But that's true for all types of quantum advantage results. I have work that will come out pretty soon on demonstrating a computational quantum advantage result, in simulation, but actually in the original quest to do so in the context of optimization.

Daniel Lidar: 17:38

But importantly, it's not in the context of exact optimization, because that seems to be very difficult to

Sebastian Hassinger: 17:44

do Right.

Daniel Lidar: 17:45

Due to the the presence of noise and the finite temperature and so on. So instead, actually, the the advantage result is in the context of approximate optimization

Sebastian Hassinger: 17:53

Okay.

Daniel Lidar: 17:54

Where what we showed was, with the help of, these same kind of error correction methods, error suppression and detection methods, to be precise, that I was talking about earlier, that for some well chosen optimization problems that are natively embeddable on the D Wave chip. So I I wanna stress those are not application level problems. It is actually the case that we were able to, show that we could solve these problems with a time to solution scaling that is better than the best classical algorithms that we were able to benchmark against. And here there's always this catch when you when somebody makes a quantum computational advantage claim. Well, in some cases, can you can prove that there's no better algorithm, classical algorithm out there.

Daniel Lidar: 18:46

But in most cases, certainly in all cases, almost of of practical interest, we don't have a proof of what the best classical algorithm is. So you end up performing some kind of a test, benchmark test against heuristic algorithms or the best known algorithms, classical So that that is the case here. We benchmarked against the best currently available classical algorithms, and we found a better scaling of the type of solution, scaling with problem size for the U Wave device in this case.

Sebastian Hassinger: 19:16

That's really interesting. I suppose the nice side effect of not having definitive proof means that it's a motivator to the classical algorithm designers to come up with better approaches to try to compete, which is good for everybody, I think.

Daniel Lidar: 19:32

That is absolutely true, and I fully expect that Yeah. This race will continue Yeah. Between very clever classical algorithm design. And I'll actually be both happy and unsurprised if if we find out that there's something better we we could have done, because that will keep pushing the envelope for everyone.

Sebastian Hassinger: 19:49

You've you've also, sort of, done work in quantum advantage, proof of quantum advantage with those caveats since best known sort of, benchmarking, on the IBM, devices as well recently. Right?

Daniel Lidar: 20:04

That's right. But actually, in this case, we try to find problems for which we can make rigorous statements about quantum advantage. So there are quantum advantage results that rely on conjectures. So what I mean by that is, for example, quantum supremacy in the style of random circuit sampling. Right.

Daniel Lidar: 20:27

Essentially, it boils down to if a certain computational complexity conjecture is true, something like does p equal NP. It's not exactly that, but it's it's close enough. Then, you can prove that a random circuit sampling, cannot be beaten classically. Right? Or quantum simulation, where you use a quantum computer to perform a simulation of a quantum system, relies on another type of conjecture, which is that simulating quantum mechanics itself is is hard on a classical computer.

Daniel Lidar: 20:59

It's widely believed, but it's still a conjecture.

Sebastian Hassinger: 21:03

Right.

Daniel Lidar: 21:04

On the other hand, there are some problems for which we do actually know what is the best classical algorithm. And typically, those arise in the context of so called black box or oracle type problems. Right. Like Bernstein Vazirani or or Simon's problem. Those are very early problems, from the the very early days of, the founding of of quantum computing.

Daniel Lidar: 21:25

And in those cases, we know rigorously that a quantum computer is going to be the best possible classical algorithm. So that's the context in which we, actually demonstrated quantum advantage, in this case, using IBM devices. First, and in 2023, a paper that used a '27 qubit IBM device Right. And demonstrated, again, a better scaling of the time to solution for the Bernstein Vazirani problem on up to 27 qubits. And that doesn't sound like a lot of qubits.

Daniel Lidar: 21:57

But nevertheless, when you see an exponential scaling versus something that is sub exponential, the exponential scaling being the classical scaling for this problem versus the the sub exponential, quantum scaling on the IBM devices, then it's it's quite clear Right. That you have an advantage. And here too, the key was to use some kind of, air suppression method. Mhmm. Very different from the type that we use on D Wave devices.

Daniel Lidar: 22:25

In this case, I'm talking about dynamically coupling, which is an air suppression method. Yeah. We we can talk about what that is. But let me just mention that the then we had a follow-up result, which is still being peer reviewed. And and that is a quantum advantage at the 27 qubit level.

Daniel Lidar: 22:43

So going up to 127 qubits, because you always want to look at time to solution as a function of problem size. Right. And this was for Simon's problem. Simon's problem is actually a very interesting problem because it was the precursor to Shor's algorithm. Essentially, it's about period finding.

Sebastian Hassinger: 22:59

Right.

Daniel Lidar: 23:00

So the result there Is

Sebastian Hassinger: 23:02

phase estimation as well, or is that something that Shor added in?

Daniel Lidar: 23:06

That is something that is, more special to

Sebastian Hassinger: 23:10

Shor's Right. Okay.

Daniel Lidar: 23:11

Yeah. But, there is, essentially, you could say that, yes, there is a type of Fourier transform that that happens, in Simon's algorithm as well. Yeah. That's right.

Sebastian Hassinger: 23:24

Yeah. Interesting. So okay. So you in that case, you used dynamical decoupling as well, on a 27 qubit?

Daniel Lidar: 23:31

Yeah. Yeah. So what happened was that when we tried to run these algorithms, Bernsav Rosarani and and Simon's algorithm later, without doing any kind of error suppression, the results were not good. Right. There was essentially no quantum advantage Right.

Daniel Lidar: 23:45

To be seen. And I just wanna make clear that, when I say quantum advantage, what I'm talking about is the scaling of the time to solution as a function of problem size. So we're not talking about a hundred x speed up Right. Or something like that. Right?

Daniel Lidar: 23:59

It it's it's crucial that you look at as a function of problem size. It's this is really the the heart of the claim of a quantum computational speed up. It is an advantage that grows Right. As polynomial or maybe as an exponential in the best case, as a function of the number of variables in the problem, which is the number of qubits in in this case. So that's what talking about here.

Daniel Lidar: 24:21

Time's a solution is a function of the number of qubits. Okay. In that context, it turned out that there was no advantage to be seen, if you ran the problem in its naked form. Right. But then we try to do air suppression using dynamical decoupling, and and dynamic decoupling is is a fascinating field.

Daniel Lidar: 24:40

Has its origins and, actually, in the Han and Co experiment from 1950, it's really the basis for NMR magnetic resonance imaging as we know it. So there's a tremendous amount of knowledge that the NMR community has developed in this context. But the quantum information community has also contributed greatly to the development of these dynamically coupling sequences that that you can think of as essentially suppressing decoherence, they don't error correct, they merely suppress. Right. But in the context of running an algorithm on this devices, this can be extremely powerful.

Daniel Lidar: 25:17

And that's what actually made the difference when we were able to successfully demonstrate quantum advantage.

Sebastian Hassinger: 25:25

And these are these are microwave pulses that you inject in between gate operations on a qubit. Right? That's that's essentially what the dynamical decoupling is?

Daniel Lidar: 25:34

Yeah. Dynamical decoupling is what you said. So the way that you use these pulses is in its simplest form. If you look at a quantum circuit, typically, there are what we like to call idle gaps. And so every qubit is, at some point, just sitting around waiting for for something to happen while, the other qubits are being acted upon by by Gates.

Daniel Lidar: 25:59

And while a qubit is sitting idle, it decoherence. Right. So rather than letting it decohere, the idea is, let's try to prevent it from decoherence or slow down its decoherence as much as possible. And so this is where dynamic decoupling comes in. What you can do is you can you can fire pulses, at this qubit or a group of qubits, that are not doing anything.

Sebastian Hassinger: 26:21

Mhmm.

Daniel Lidar: 26:22

And simple way to think about the way this works, by no means the the actual, more elaborate way in which it was eventually done, is that dynamically coupling is like doing time reversal. Essentially, you're you're flipping the sign of the interaction between that qubit and its environment. I see. And you constantly flip it back and forth. And so it's as if somebody's walking forward and backward.

Daniel Lidar: 26:45

Right. Right. And on average, they're going nowhere. And that's that's what this cubit is doing.

Sebastian Hassinger: 26:49

So do you have to make sure that you return it to its original sign before the next operation? So there is some kind of structure to the dynamic that you have playing with. It's gotta be structured.

Daniel Lidar: 26:59

There's a a ton of structure to these sequences, and and and, yes, that's right. So, ideally, you you would like a dynamically coupling sequence to complete an identity operation. Okay. But that identity operation is a nontrivial identity because it it forces the the qubit environment to cancel out up to some order. This is not complete error correction, let's be very clear about that.

Daniel Lidar: 27:20

No entropy is being removed from the circuit. In fact, you're injecting some entropy because no such thing as perfect operation.

Sebastian Hassinger: 27:29

It feels like that would increase the overall noise of the circuit, but you're actually acting to suppress the noise.

Daniel Lidar: 27:34

Yeah. So if you try to do dynamically coupling naively, it's not going to work. And by naively, what I mean is if you use pulses that aren't themselves self correcting. So we now know that there exist pulse sequences that have an inherent, I think it's fair to say fault tolerance built into them. They self correct over the course of a full cycle.

Daniel Lidar: 27:59

So if there is some coherent error buildup, because there's a little bit of over rotation or or the wrong axis being rotated around, as you apply a sequence like that, a cleverly designed dynamically coupling sequence is going to take care of that and is going to eventually get you to the identity operation at the end of the cycle, despite the fact that there were errors along the way.

Sebastian Hassinger: 28:23

Right. Right. And, I mean, more recent, and I think actually presenting this week at APS, work that includes logical dynamical decoupling or LDD, I think you're referring to as. Right?

Daniel Lidar: 28:35

Yeah. That's right. So you can take this up to the next level instead of just looking for gaps where the qubit is sitting around doing nothing and and addressing it with dynamically coupling this way. You can combine ideas from dynamically coupling with ideas from quantum error correction. And logical dynamic decoupling or LDD is the idea of using the logical operators of a quantum error correcting or quantum error detecting code.

Daniel Lidar: 29:04

And those operations, those logical operators, can have the same effect as as what I described before. But there's there's one major advantage in using logical operations of of of a code has done on decoupling operations. And that is that now you can naturally combine that on decoupling with an error correcting or detecting code. Mhmm. They now literally fit hand in glove, and they, reinforce each other.

Daniel Lidar: 29:28

Mhmm. So in particular, logical dynamically coupling can actually suppress the logical errors that a code is not able to deal with.

Sebastian Hassinger: 29:39

Does that effectively sort of increase the d value for that that code then?

Daniel Lidar: 29:44

By the d value, you mean the the distance The distance. Yeah. Of the code.

Sebastian Hassinger: 29:47

Yeah.

Daniel Lidar: 29:48

Yes. The way that the construction in our new LDD paper works is essentially that's what happens. You can change the distance of the code. You can even use an an error detecting code, essentially as an error correcting code, in this context. And the reason it works is because when you apply these logical dynamically coupling pulses, again, they remove all logical errors, or not completely removed, they suppress all logical errors that the code, by design, was not able to cope with.

Daniel Lidar: 30:18

In addition, they also suppress some of the physical errors that the code was designed to to deal with. And what's left over is a subset of all the errors that the code was originally supposed to to correct. And because you've now removed some of the errors, the the the code is is now essentially a higher distance cut. It has to do less work.

Sebastian Hassinger: 30:39

Right. Right. That's spectacular. Is this theoretical work, or have you actually carried this out experimentally at this point?

Daniel Lidar: 30:47

We have carried it out experimentally. There's there's a theorem in in this new work, that formalizes everything I just said. But, yeah, we use this approach to, for the first time on a Transmon qubit processor, demonstrate high fidelity logical Bell state generation. So beyond generation, we generate logical Bell states in this case, the context is the four-two-two code. It's just really just four physical qubits, a code that gives you two logical qubits out of four physical qubits.

Daniel Lidar: 31:17

And those two logical qubits, we prepare a Bell state, a logical Bell state. And then we show that we can, using dynamically coupling in this form, in combination with the properties of the four two two error detecting code, keep those logical Bell states alive for much longer than, is possible if if you don't do any of this. But also much longer than is possible if you only use the 422 code itself Mhmm. As as an error detecting code with with post selection.

Sebastian Hassinger: 31:48

So this this is a four qubit device. Was this in a a university lab, or is it who whose device was it?

Daniel Lidar: 31:55

This was actually on the IBM over the cloud Oh, okay. Devices on the latest and greatest machines from from IBM. Four qubit codes, in this case, just because, we wanted to start with something simple.

Sebastian Hassinger: 32:09

I see.

Daniel Lidar: 32:09

I see. And then, of course, there are a lot more physical qubits on these devices. So so in fact, we're able to leverage the parallelism that you get from running many copies of this.

Sebastian Hassinger: 32:19

I see. I see. Would that be I mean, would would that be sort of a the path forward, would that be to to increase the the physical qubits you're involved in involving in the logical code? Or would it be to do more of these sort of parallel smaller units?

Daniel Lidar: 32:35

Oh, definitely the former. The former. So the next thing we'd like to do is, demonstrate that we can make a high fidelity logical entangled states, with, much larger codes.

Sebastian Hassinger: 32:46

And the the the IBM chips still have the the heavy hex layout, so it's it's two or three connections, I believe, per qubit. Is there any would there be any advantage in in more of a lattice work layout, like the Google chip or the IQM chip or Nakamura san's chip from RIKEN, they're all, you know, they each have four connections, all nearest neighbor connections. Would that affect the the performance of the logical code?

Daniel Lidar: 33:15

Yes. Higher connectivity as a general rule is always better because it reduces the the number of swap operations that you have to perform. Every time you have to perform a swap, you're losing a little bit. So, yes, I'm I'm a fan of highly connected architectures.

Sebastian Hassinger: 33:32

And in turn talking of architectures, what about this sort of it feels like we're at this point now where plain vanilla transmods to a certain degree are are falling away, and you're getting cat qubits, dual rail qubits, you know, the IBM's new plan chip that's in the works is an implementation of a QLDPC architected chip. Do you think that there'll be advantages, or can you adapt your approach to those sort of more esoteric kind of designs of and layouts of chips?

Daniel Lidar: 34:09

Yeah. Absolutely. These results are all fairly agnostic Mhmm. Ultimately about the specifics of the qubits to the architecture, because they're based on on general design principles. Mhmm.

Daniel Lidar: 34:22

The four two two code, for example,

Sebastian Hassinger: 34:23

is Right.

Daniel Lidar: 34:23

The code that you can, put on on pretty much, any modality and, any type of of architecture. So, So, yeah, I'm really looking forward to working with these, fancier, even better qubits.

Sebastian Hassinger: 34:34

I know. There's so many coming down the pipe. It's very exciting. And I I guess the the last last topic I wanna touch on, underlying all of this seems to be and you mentioned sort of getting fascinated with, sources of noise and characterization of noise. And, of course, that's that's the motivator for getting better at error correction.

Sebastian Hassinger: 34:55

You know, what you've done a lot of work even in in simulation of noise. Is there something sort of emerging in our understanding of how noise causes problems in in chips that's that's helping us improve sort of the overall performance? I guess this is I mean, yes, obviously, error correction, but I I guess, I mean, in a more fundamental way. Do we understand quantum noise better than we did a decade ago?

Daniel Lidar: 35:23

Yeah. I mean, that's a deep question with a lot we could go into there. So and the short answer to your question is, of course, yes. But the point is really that error models are, getting better, and in in certain important ways. So a lot of the the results about, for example, the accuracy threshold of fault tolerance are based on rather simple error models.

Daniel Lidar: 35:53

Typically, the error model is a type of Monte Carlo simulation, where you take, let's say, depolarizing channel or any kind of noise channel. And, essentially, you imagine you have a circuit, and you you sprinkle errors at random in the circuit. And and then you can simulate the effect of that noise, and you calculate some kind of a fidelity quantity. And you can extrapolate that to to different size from using different sizes, can extrapolate and you can obtain a a threshold type result. The problem is that real life noise isn't at all like that.

Daniel Lidar: 36:29

It's not necessarily Markovian. It has memory characteristics. In other words, it can be correlated in time, and it can also be correlated in space. And more sophisticated understanding of these more realistic noise models is absolutely essential in order to to get realistic results about accuracy thresholds and the performance of error correcting codes more generally. And I think that is an extremely rich and important area that the community has tried to move towards, and there are some people who have been working on it for a while, that is accounting for non Markovian type of effects, correlated errors.

Daniel Lidar: 37:12

But there's a lot that remains to be done in order to improve our confidence in the claims that are being made around particular threshold values. So I'm always a bit skeptical when I hear a number like, we have a threshold of, let's say, you know, 1%. Did you properly account for all non Markovian effects in, in arriving at this number, for example? And and and typically, answer is probably not. So I've actually started a company that is trying to take this problem seriously.

Daniel Lidar: 37:48

It's called quantum elements. We're building open system quantum simulation software that is trying to be as as accurate and realistic as as possible. One of the goals is to to try to provide more accurate estimates of real noise to the community and to to customers who are are building quantum hardware.

Sebastian Hassinger: 38:12

And the the models that you're building, are those, I mean, are you creating sort of a a mathematical model for for a a better model of how the noise, behaves, or are you training a machine learning model to to mimic that the the real world noise? So we're trying

Daniel Lidar: 38:30

to build better models. We're trying to also develop, techniques that will help with, speeding up the simulation problem here. Mhmm. Because this is a classical simulation problem. This is why we need quantum computers

Sebastian Hassinger: 38:43

Right. In the

Daniel Lidar: 38:43

first place. So there's a bit of a, a vicious cycle and bootstrapping going on here. Yeah. So the emphasis is on on building, more accurate models, faster solution methods for these types of models. And and about AI, well, a the role of AI here is to essentially plug into a feedback loop Mhmm.

Daniel Lidar: 39:07

Between the error models, simulations, and actual devices, learn the noise from the devices Right. In order to improve the error models and so on.

Sebastian Hassinger: 39:19

Right. Right. And you mentioned before also, before we started recording, that that quantum elements is now playing a role in in, the bring up process for I mean, I can easily imagine quantum elements being useful in the design and simulation of of the qubits you're building and, you know, the testing and the r and d process. But but in terms of operationally bringing qubits up, how does that that noise model help?

Daniel Lidar: 39:46

Yeah. So we're very interested in this bring up problem. The bring up problem is essentially I like to think of it as somebody delivers a piano to your house, and you are not an experienced piano player and certainly not a piano tuner. So you start to play your piano, and it sounds a little off, and so now you need to call in a tuner. So that's the bringer problem.

Daniel Lidar: 40:10

Well, there's a lot of, role here for a an open system simulator, the type that I I talked about before, connecting, to that hardware in order to learn the noise on on that particular hardware and help with not just air correction that actually comes later, but also with, calibration

Sebastian Hassinger: 40:32

Mhmm.

Daniel Lidar: 40:33

And, getting the processor ready to be used at a level it which it operates as close to optimal as possible with highest possible fidelities and so on. Right.

Sebastian Hassinger: 40:44

That makes total sense now. Think that, I mean, there's there's a whole team in a a typical quantum hardware r and d team that's just calibrating the machine to try to make it, you know, its operational state as optimal as possible.

Daniel Lidar: 40:59

That's right. This is exactly the the pain point that we are trying to address. And I remember as a graduate student walking into an optics lab and seeing all these lasers on on an optical table and being very impressed with how much work was being done to to align all the lasers. Was told then that it could easily be a PhD to just get all that alignment going before you get an experiment. And so there's been a lot of automation in in that area, and it's time to bring that automation to bear also on superconducting qubits and and spin qubits.

Daniel Lidar: 41:37

And those are the the two particular modalities that we're interested in.

Sebastian Hassinger: 41:40

Fantastic. Well, Daniel, thank you very much. It's been super interesting. And I always enjoy talking with you, and this has been no exception. So thank you.

Sebastian Hassinger: 41:49

Thank you.

Daniel Lidar: 41:50

It was a lot of fun.

Sebastian Hassinger: 41:58

That's it for another episode. Thanks for listening as always. Thank you for Doctor. Daniel Lidar for his time. Thanks to APS for their support.

Sebastian Hassinger: 42:08

The podcast is a production of the New Quantum Era hosted by me, Sebastian Hassinger, with theme music by Omar Costa Hamido. You can find past episodes on www.newquantumera.com or on blueskynewquantumera.com. If you enjoy the podcast, please subscribe and tell your quantum curious friends to give it a listen.