Sebastian Hassinger (00:01.388)
I'm Matthias, good to see you. Thank you very much for joining.

Matthias (00:05.087)
Hi Sebastian, thanks for having me.

Sebastian Hassinger (00:08.249)
I'm I'm really interested in this conversation. I've been wanting to have you on for a while Which is something I say to a lot of guests, but in your case, I'm I'm you're almost a little bit of an enigma to me because you're your professor of or you're trained mathematician Your professor at University of Copenhagen. You're the director of the quantum for life Center which to me I thought I was sort of thinking in my mind that you were

more computational chemist or interested in quantum as it relates to biology, but you also have very strong theoretical foundation to your work. And most recently in the work that we're going to talk about in particular, it's having to do with fault tolerance and sort of a reframing of the question of what is fault tolerance for and how do you achieve it, which I find super fascinating.

So before we get into the details, I would love to hear a little bit about your academic journey, your intellectual journey. How did you get to where you are in your career right now?

Matthias (01:16.417)
Actually, I started physics back in the day at ETH Zurich. And I was fascinated already during my studies about the possibility of quantum information and the possibility that qubits change the way we do information processing. I was reading the original paper by Bennett Brassard and by Eckert.

and I wanted to get into this field.

Sebastian Hassinger (01:49.57)
That's great. yeah, go ahead.

Matthias (01:49.714)
And yeah, what happened then, and I already started doing my master's thesis in the area more and focused on quantum security. And then the studies were very theoretical. So was sort of clear I was going into as a theoretical physicist with very mathematical tools, but combining it.

in order to understand what are the information processing capabilities of quantum bits. So going into computer science.

Sebastian Hassinger (02:25.102)
Right. Right. Right.

Matthias (02:26.677)
combining now all of the three areas, right? Computer science, math, and physics. And what happened in... Yeah, sorry. Now what happened in later years, you know, was possibilities to study an application of this. And this application was in the direction of life sciences and the direction of chemistry. And then I got the chance to...

Sebastian Hassinger (02:32.374)
And you mentioned him and sorry, go ahead. No, go ahead.

Sebastian Hassinger (02:45.475)
Mm.

Matthias (02:55.553)
to lead this quantum for life center where we develop quantum technologies for the benefit in the life sciences and I think that was very exciting but of course I'm still the same. I would like to study what is information processing at the quantum scale.

Sebastian Hassinger (03:07.02)
Yeah.

Yeah, it seems in a way that that kind of broad everything from information representation, the density of information representation to applications in security and complexity theory, fault tolerance, almost the framing of what are the algorithms, what are the applications, what are the practical uses. I wonder if that

very broad set of considerations are maybe in part contributed to your insight into this embedded assumption in quantum fault tolerance that everyone was looking at fault tolerance essentially with classical inputs and classical outputs. You sort of reframe that, I think. Is that accurate around the work we're going to talk about?

Matthias (04:05.685)
Yeah, so when you normally think about the fault tolerance theorem and why do we need fault tolerance? Fault tolerance is sort of like the fault tolerance theorem says that you can simulate a perfect quantum computation on underlying noisy quantum bits. So on quantum bits, when you do an operation, it's noisy. But then there's a theorem or a statement, mathematical statement saying, like, if you just build the software right,

the full 12-1, the error correction layer right, then actually you can enact a perfect computation.

And this I view as the basis for the field. It's really the theoretical foundation for why many people believe we can build a quantum computer and why it's worth doing that. so as quantum computing came further along and it's becoming more sort of reality as we're moving forward in time,

In some sense, for me personally, this theorem sort of like gained importance because it sort of gives us the foundation to stand on. But my own background was more in the area of quantum communication. And in quantum communication, it's not that you want to compute a function. You want to compute the energy of molecule or you want to the factors of a large composite number.

Sebastian Hassinger (05:24.088)
Right.

Matthias (05:45.569)
But what you want to do there is you want to use your quantum computer in a clever way to make a quantum state, to produce a quantum state at the output that you then can send over a noisy channel to someone else. So the quantum computer is actually not delivering a prime number or it's not delivering. The quantum computer is producing a quantum state. That's its purpose. well, how do we, you know,

Sebastian Hassinger (06:05.996)
Mmm.

Sebastian Hassinger (06:09.965)
Mm-hmm.

Matthias (06:14.782)
what does the fault tolerance theorem say about this and what can we do? And the fault tolerance theorem as it stands, has this limited applicability with like classical input and classical output, but the tools behind it you can use more generally. And so one of the first things that we wanted to do, we want to say like, hey, we have all these great results on channel capacities on how much information we can send through

Sebastian Hassinger (06:32.373)
Hmm.

Sebastian Hassinger (06:41.432)
Right.

Matthias (06:42.88)
through a channel, again by the giants like Bennett and Peter Shor and so on, and capacity theorems. Can we actually do this in practice now that it's nearing? Now that the computer is being built, can we do this code? When the machine is noisy. And so this was the first question we asked and it fundamentally kind of led us to this question, well, okay, so here's a quantum computer, my encoding machine.

Sebastian Hassinger (07:01.55)
Mm.

Matthias (07:12.754)
It's supposed to produce this quantum output. And maybe even it takes a quantum input. How do I make sure that I can run this on a noisy machine?

Sebastian Hassinger (07:17.592)
Hmm.

Sebastian Hassinger (07:24.088)
That's so interesting because it's almost reframing the challenge to, it's almost more shifting the focus from von Neumann to Shannon in a sense, right? I mean, it's information being carried on a channel that needs to overcome noise. So then if your input and output are quantum information rather than classical information, what does that do to

the feasibility to the ability to achieve fault tolerance. Does that make it easier?

Matthias (08:02.817)
Well, first of all, in some sense, it's more difficult because the quantum bits, so the quantum bits that come out of your, that you have to prepare, whatever computation, whatever quantum state you want to prepare, say, with your machine, there will always be a last layer of quantum bits, of quantum gates that you perform on them in order for this preparation.

Sebastian Hassinger (08:25.006)
Mm. Mm-hmm. Mm-hmm.

Matthias (08:28.988)
So if you are preparing a quantum state on a noisy machine and you're not preparing it in some kind of logical subspace, but you're actually producing it on the physical qubits, then there will always be a final layer of noise being applied. And the way we circumvent this if the output is classical is simply by assuming that classical bits are stable.

Sebastian Hassinger (08:38.222)
Mmm.

Sebastian Hassinger (08:47.298)
Hmm. Mm-hmm.

Sebastian Hassinger (08:58.104)
Right.

Matthias (08:58.492)
And the quantum bits, their assumption, they have this noise on, so that's the best I can do. And then, but what is interesting about the work that we did is that we can actually show that this is achievable. So regardless of how complicated the quantum state that you want to produce, you can produce it on a noisy machine. But you have to pay this little price at the end, one layer of noise, but not more than that.

Sebastian Hassinger (09:03.884)
Hmm.

Sebastian Hassinger (09:25.42)
Hmm, okay.

Matthias (09:27.424)
So in that sense, that's the statement we can make. And that is the statement that allowed us, for instance, to prove now that we can still communicate at the correct rate through noisy quantum channels from Alice to Bob and so on. So giving back meaning to those capacity results, even at a maximum.

Sebastian Hassinger (09:51.48)
So if I understand it, if the output's classical, essentially you're doing some kind of post-selection or final cleanup of the quantum error. You're creating a fault-tolerant output, in a sense. And in this instance, if you're going quantum output or preparing quantum state and providing that as quantum input, that last

Set of operations is going to be effectively uncorrected So there's noise at the boundary from the quantum preparation state preparation to the quantum Transmission information transmission. Is that is that right? Okay, okay, and and is so is there a is there sort of a budget that you can theoretically establish for for what that What that probability of error is at the at the boundaries of the input and the output?

Matthias (10:32.704)
Absolutely, that's correct.

Matthias (10:49.386)
that's just your gate error of your machine. OK, so if your machine operates, that's just the normal physical error rate of your machine that will limit that. But you don't have to pay more. You could, up here, have assumed that maybe it doesn't work at all anymore. So now you have one, as you said, a boundary of error. And the operations you want to do or the next thing you want to

Sebastian Hassinger (10:53.219)
Okay.

Sebastian Hassinger (10:57.858)
Okay.

Sebastian Hassinger (11:03.649)
Okay.

Mmm.

Matthias (11:16.34)
feed this quantum information into it, has to be robust. The task has to be robust with respect to this that you want to execute has to be robust with respect to this last layer. And this is now an active area of research of trying to figure out which tasks are robust and it's very rich. I think what you can imagine is now that as a separate team, you can think like, now I'm sending this from Alice and Bob.

Sebastian Hassinger (11:34.824)
Mmm. Mm-hmm.

Matthias (11:45.576)
far away, know, you know, continent distance, like we have it now, or but you can also think of like, hey, if I have a small quantum computer on my desk, maybe it has several quantum cores, just like we have CPU cores now, a normal computer, and then also there, the cores need to communicate. And so, and they take a quantum input, and they have to provide a quantum output to the next core, and they, and there is a back and forth communication. And so,

Sebastian Hassinger (11:50.445)
Right.

Sebastian Hassinger (12:01.079)
Right.

Sebastian Hassinger (12:06.391)
Right.

Matthias (12:14.963)
This result is really to be seen or this drive towards studying quantum computation with quantum input and output is really, you should really think of it as for a network or multi-core processor in the small and in the large scale.

Sebastian Hassinger (12:29.974)
Right. Right. Right. Interesting. Because I was thinking, of course, was thinking in terms of networking, as you say, long distance communication, arbitrary distance. But once you think about distributed computing within a system, then that reframes it to, why would you force QPUs to speak in classical information terms to one another? You should probably enable them to speak.

Matthias (12:56.959)
He's sexy.

Sebastian Hassinger (12:58.88)
in quantum information terms because you don't want to actually, you want to avoid the overhead. Yeah, of course. Yeah, it's just a classical computer with some quantum bits in it. That's really interesting. So then again, so let's just put this in some sort of a practical example. You've got two QPUs, there's some modality and you have

Matthias (13:02.924)
Otherwise, you lose their speed up, right? You lose the speed up if you...

Sebastian Hassinger (13:27.264)
some number of qubits in each of those QPUs and you want to treat them essentially as a distributed system, you could have, you can imagine a photonic connection between those two QPUs with some kind of, as you say, boundary of error as the final step between the connection. But as long as you can correct for that boundary error, that input and output error,

then you can effectively have noiseless or fault tolerant error corrected communication between those two cube cues and then essentially treat it as one computational system. Is that right?

Matthias (14:11.231)
Yeah, that's sort of the idea. So what we want to build, of course, I mean, of course, they might still be limited because they're in two different places. But we want to enable, we want to build the to make, we want to build the tools, the mathematical tools, the coding tools to make it possible to treat those as two-fold tolerant machines interacting.

Sebastian Hassinger (14:12.952)
That's amazing.

Sebastian Hassinger (14:37.57)
Mm-hmm. Okay.

Matthias (14:39.465)
with a quantum input and a quantum output. And that is so that other designers that design an algorithm and design distributed algorithms, for instance, that they can run them on top of this layer.

Sebastian Hassinger (14:57.75)
And there's a lot of work going on around interconnects, often having to do with transduction from microwave qubits or frequencies to telecom frequencies or, you know, photonic links between ion traps, between neutral atom traps or neutral atom chambers. How does this theory relate to those engineering challenges? Does it change anything in the way those

challenges get approached or is it articulating what they're trying to do in practical terms?

Matthias (15:38.559)
I...

Matthias (15:42.792)
I think in a sense you can see the motivation behind studying distributed computing just like classically is that there will be situations in which it's easier to build more cores of a limited size rather than build one large core. And so in some sense what we want to do is we want to give the flexibility.

And we make the tools to allow the flexibility of the algorithm designer and of the hardware designer to make that possible. We don't want to limit them. And of course, there's a number of ways. So our first results here were kind of like on the abstract, in a more abstract realm, where we took constructions, for instance, multiply concatenated codes that

Sebastian Hassinger (16:14.35)
Right.

Matthias (16:38.759)
you might say are maybe not the ones that will be immediately implemented or directly relevant to current hardware. But what has happened is that with the advent of quantum LDPC codes that have a very good encoding rate, one thing that, for instance, we were able to do is build these encoders and decoders more efficiently. So one of our latest works is about how can we make this more efficient? How can we make this more space efficient?

Sebastian Hassinger (17:01.582)
Hmm. Hmm.

Matthias (17:08.575)
And we're going in this direction, but of course there will be the question, how do we translate this? What does this actually tangibly mean in an implementation? And I think that's work I'm very excited about and hopeful comes.

Sebastian Hassinger (17:23.66)
Right, right, right, right. Yeah, mean, LDPC codes are quite promising, but they also rely, the LD stands for long distance. So being able to have the theoretical framework for fault tolerant connections between cores makes it easier to establish those long distance connections that allow LDPC to work, I assume.

Matthias (17:55.688)
Yeah, I mean...

That's true that the low density parity check codes, some of their connections can be long distance. Now, for the codes, when I was speaking of actually one core and using one of those codes in order to enable the computation within one core, in order to make somehow the overhead of default tolerant construction to limit the amount of overhead needed there.

So it's not directly related to the distributed architecture that we've been talking about before of these cores interacting, but rather of how do I make one of the constructions in one of the cores as efficient as possible.

Sebastian Hassinger (18:45.358)
Okay, interesting. Interesting. And I was really curious. You debated Gil Kalai in 2025. He's one of the more visible skeptics about quantum computing. He believes that, or he has doubt of whether quantum computers can exist at all in practice. And I think his core argument sort of revolves around

error rates that we can't get error rates low enough at meaningful scale, that there's some incompatibility between quantum error rates or quantum coherence and the scale that you require to do meaningful computation. Is that sort of your response? Is it sort of rude in what you just said, that you're trying to find

tools to achieve that scale without having to build a single core that has a million qubits or whatever the arbitrarily large number is.

Matthias (19:53.139)
Yeah, that's right. mean, default tolerant tools will enable you to build better logical qubits on top of noisy hardware. And of course, if the physical qubits are limited, if they're limited to some noise rate, then we still believe by having more of them, we'll be able to build better and better logical qubits.

And yeah, I think that's Gil's point that this shouldn't be possible. yeah, I mean, I was debating him. can watch the debate on YouTube. Yeah, I was on the other side. I don't see any obstacles to us building a quantum computer. And I'm very excited about the project going forward.

Sebastian Hassinger (20:27.342)
Yes, I know you're on the other side.

Sebastian Hassinger (20:42.476)
Yeah, absolutely. I can imagine taking the other side of that argument in general principle, and you're making real contributions to the theoretical toolkit that we need in order to overcome those skeptical objections. Are there any concerns or questions that Gil has raised that you still

you know, have, I think our open questions or our concerns that you're thinking about in your work.

Matthias (21:17.756)
I think he makes his questions, what are the assumptions? And kind of like maybe pushes us a little bit. And actually, it was sort of like refreshing because you can start to think about what is really the essence and say, well, I don't really see an obstacle. So I see there's assumptions going into the error correcting architecture. For instance, there are certain assumptions of the independence of errors that seem to be

Sebastian Hassinger (21:35.928)
Hmm.

Matthias (21:47.515)
adhered by the hardware being built at present. so the assumptions that going into this fault tolerance theorem, and now we can always say like, OK, is it that precise noise model is a little bit different? I believe the techniques in order to establish fault tolerance are rather general and robust.

Sebastian Hassinger (22:05.646)
Mm.

Matthias (22:16.55)
And the full tolerance theorem gives us this basis of saying, well, yes, based on this theorem, we can build the architecture. And now, of course, if you want to do this in practice, there will always be the question, how do I do this most efficiently, most effectively? Which code do I choose in practice? Which noise model do I have? Et cetera. What are the obstacles of going? But I don't see the physical.

Sebastian Hassinger (22:37.87)
Mm. Mm-hmm.

Matthias (22:47.72)
physical barrier there.

Sebastian Hassinger (22:49.102)
Right. It's interesting. You sometimes describe your work as developing quantum software for a future quantum computer. You've just been talking about sort of the theoretical grounding that will make that computer more useful for applications. you run the Quantum for Life Center. Do you have sort of a set of

of applications, practical uses for this future quantum computer that you're reverse engineering. The software I would need in order to provide value for quantum for life, for life sciences, would have to have this set of requirements and therefore I need a system of this scale and therefore I need the tools in order to marshal that level of compute resources.

part of your process to a certain degree.

Matthias (23:50.982)
Yeah, to certain degree. within, we have run a large project on trying to understand the binding affinity of a drug molecule, of a ligand to a target protein. And I've been working together with biologists and chemists in order to take this large protein and zoom in on the relevant quantum region and understand how we can carry out this quantum simulation on a quantum computer.

and then feed it back into a large classical machine learning assisted workflow so that we can compute the properties of the entire system with a quantum input. with a small quantum computation. And then the question is like how big does it have to be in order to...

Sebastian Hassinger (24:47.18)
Right.

Matthias (24:47.997)
to impact that and I think this is super exciting. the software stack we've built is running very well at the moment. We've been running it with a classical backend, but as soon as the hardware comes along, can slot in the hardware, the quantum hardware, and run this larger computation. And therefore make impact in the biological realm.

Sebastian Hassinger (25:09.422)
That's really cool.

Sebastian Hassinger (25:15.852)
Right, right. That's really fascinating. And I'm also curious in that life sciences setting, often quantum sensing applications are being discussed sort of in a better medical sensors or more higher resolution, almost atomic or molecular resolution scans. Are there instances that you think that there may be

sensing applications in a life science setting that provide that quantum input into the quantum computer that would take advantage of exactly the kind of theory that you're building for that quantum input output fault tolerance.

Matthias (26:00.316)
Yeah, I believe so. So my colleague here, Eugene Poulsik, who's also part of the Quantum for Life Center, he actually senses magnetic field with atomic clouds and magnetic field of cells of some mammals. And actually, he is really moving forward at spectacular speed with this project. And of course, that is exactly

Sebastian Hassinger (26:09.496)
Hmm.

Sebastian Hassinger (26:20.334)
Hmm.

Matthias (26:29.657)
in his atomic cloud, one day we might take the state, the quantum state of the atoms there and process it within a full-tolerant quantum computer, and then it's a quantum input to this quantum computer, precisely. And yeah, that's a very important point. And I'm looking forward to that happening.

Sebastian Hassinger (26:56.332)
Yeah. Well, and again, I mean, there's another interesting analogy to the beginning of the information age in the 40s, 50s. You know, there were classical non-universal computing devices that were used for signal processing primarily. So I wonder if there's going to be, you know, sort of the quantum sensing and then some amount of quantum signal processing that's not, not universal.

But then that almost prepares the quantum state as an input into the universal quantum computer that's fault tolerant in the method that you're describing.

Sebastian Hassinger (27:40.334)
It's interesting. And then the really exciting thing that you said before, the idea that you don't need a quantum computer to compute the entire molecular ground state or excited state or whatever of the entire biological system. You actually just are zooming in on particular quantum surfaces, let's call them. They're not, but whatever, you know what I mean, areas.

and getting a more exact simulation of what's happening in a very precise interaction and then using that to seed a larger scale machine learning classical workload that's much more familiar to us. But with that more accurate starting point that potentially gets us to a better answer more quickly, guys, I presume, right?

Matthias (28:33.533)
Yeah, so for us in the first place it's better. It's about the quality of the simulation.

And of course there's challenges. One challenge is how do I take the small exact quantum computation and embed it into this larger classical in order to obtain an energy for the entire system. And that in its own is a fantastic problem that my colleagues from biology and chemistry have worked on some years and now quantum computing comes in with a new computing modality.

to impact that workflow. And one other aspect I think is very exciting that maybe puts spotlight on what is it that we might be using a quantum computer for. In this case, it's training data for machine learning model about how interfaces of ligand and protein work.

And also that in order to find one useful result, you might actually have to do many small quantum computations. So many small quantum computations will give this input to this model. And then you compute one final answer. So I think this is a very, very interesting new. certainly, for me, it gave me a new perspective on what quantum computing might be capable of in the future.

Sebastian Hassinger (29:47.96)
Hmm. Mm-hmm.

Sebastian Hassinger (30:09.794)
That's amazing. do you have, to put you on the spot, do you have sort of a sense either of a timeline or a scale of fault-tolerant qubits that you sort of think will be required for that sort of uniquely quantum-enabled result? Quantum advantage, whatever you want to call it.

Matthias (30:33.826)
I think the field is moving so fast and it's not going to be many years that we'll see impact. The field is really accelerating and you see this that now the software development, so quantum error correction, fault tolerance, algorithms development is going hand in hand with what happens on the hardware side. And this is really inspiring to see that developments in the hardware spur development on the

Sebastian Hassinger (30:37.24)
Yeah.

Matthias (31:03.292)
software side and the other way around. And I think, you know, this will lead to more co-design of hardware and software. And I think this is really a pivotal moment now that the whole field is going through this change in transition. You know, a few years back, I would have told you that the really important thing is bridging the gap between the hardware and the software. Now it's not anymore about bridging the gap. It's about working together.

and moving along and putting these two areas together. They're really touching and getting them together and getting them to work in the most efficient way. Obviously the hardware is not super large scale yet, but it will be getting there very fast. And the better we co-design, the faster it will be.

Sebastian Hassinger (31:56.398)
Right, right. Yeah, that's what I find so fascinating. And I think, you have such a an advantageous perspective because as you said, you're starting from a pure sort of quantum information starting point, you thought through communication and security kind of applications from the Bennett Brissard and Shore kind of perspective. You're trying to create this sort of

It's like theory, but it's practical theory because it's, as you said, it's trying to bridge the end application with the hardware design in a way that gets us to a much more efficient path to actually realizing really useful applications in life science or in other areas. So, Matias, thank you so much for joining. This has been really fascinating and I'm looking forward to seeing what you continue to work on, what you do in the future.

Whenever you get that first sort of interesting quantum input into a classical machine learning model, I want to have you back so we can talk through what you've achieved there.

Matthias (33:09.084)
Thanks you, thank you, thanks for having me.

Sebastian Hassinger (33:11.924)
Excellent.