CEOInterviews.AI
Start App
Get an email when Ankur Moitra gives a new interview
One email per new interview. Unsubscribe in one click.
Ankur Moitra
Professor of Mathematics, MIT

Ankur Moitra - Transitions in Quantum Spin Systems - IPAM at UCLA

📅 Jan 13, 2026 Institute for Pure & Applied Mathematics (IPAM) 50 MIN 1293 VIEWS 67 SEGMENTS · 2 SPEAKERS
Recorded 13 January 2026. Ankur Moitra of the Massachusetts Institute of Technology presents "Transitions in Quantum Spin Systems" at IPAM's New Frontiers in Quantum Algorithms for Open Quantum Systems Workshop. Abstract: Classical spin systems exhibit a variety of interesting phase transitions wherein there are sudden and dramatic changes in their macroscopic behavior as the temperature varies. In this talk, I will introduce quantum spin systems and study the relationship between entanglement and temperature. In particular, I will show that above a constant temperature, that depends only on t...
Ankur Moitra 0:00 ↗
Great. So, thanks everyone for coming. This is based on joint work with some really awesome collaborators: Inesh Bakshi, Alan Lou, and Ewen Tang. So, I'm a newcomer to quantum spin systems, but I've spent a lot of time thinking about classical spin systems.
So, in this talk, I'll tell you about some cool results about quantum spin systems, specifically about phase transitions and entanglement. But really, what I want to emphasize the most are the metaphors with classical cases, because I think there's a lot of opportunities to try and build what is this dictionary between what do we know classically and what can we hope to say from a quantum standpoint. So I'm going to try and introduce these things really from a classical perspective first.
So let me start off with introducing some of the basics of statistical physics, because that's really the perspective that I want to take before I get to the quantum case. So statistical physics is, you know, a remarkable feat of 19th century physics where the basic guiding question is how do local interactions between different molecules or particles give rise to very complex macroscopic behavior. And these aren't, you know, exotic things; they're very down-to-earth things.
Like maybe the simplest example of these macroscopic phenomena that we want to understand are things like magnetism. So magnets are familiar to all of us, you know, we've certainly all played with them when we were kids. Maybe we still play with them now. No judgment. But, you know, how does a magnet work? So, in your material, you've got electrons and they're each spinning and they have a choice of how to orient their spin. They can orient it in a clockwise or counterclockwise direction. And of course, when the electron is spinning, it induces a local magnetic field. And the choice of how the spin is oriented affects the choice of how the local magnetic field is oriented as well.
So what is magnetism? This macroscopic phenomenon that's familiar to all of us. All it is is when we have a material where most of the spins line up in the same direction, so that you have a macroscopic magnetic field. And, you know, the way to think about these things is you can imagine in your material you might have electrons that are nearby each other and you can imagine that they communicate with each other through some physical forces and mechanism. But when you have electrons that are very far away from each other in the material, how exactly do they all decide on which way if they don't directly communicate with each other?
So that's what statistical physics is about: it's about this microscopic modeling of the interactions, say, between spins of electrons, and macroscopic phenomena, things like magnetism, that we'd like to understand. Even better, we'd like to probe how these macroscopic phenomena are controlled by things like the strength of the interaction, the temperature at which the materials are living. But, you know, to dive a bit more into the details, you know, for things like ferromagnets, the basic principle is very simple: it's that neighboring spins on the lattice, they prefer to point in the same direction.
Now the physics that, you know, gives rise to that is a little bit complicated. It involves something called an exchange interaction. But for us, this is the starting point of what statistical physics is. You know, what you do is you define a probabilistic model that describes a probability distribution over all configurations that these spins could be in. And that's a truly gigantic set. But you define this distribution in a compact way based on writing down some energy functional. And you study properties of this distribution as a way to understand these macroscopic phenomena. That's what statistical physics is about.
So, you know, the standard example that you see the first time you talk about these things in classical spin systems is the Ising model. So there are a lot of terms here and I'll go through and explain them. Probably many of you will have seen them, but this will be a good opportunity to make sure that we're on the same page in terms of the notation. So first I'm defining a distribution on plus minus one to the n. So I think about there as being n different electrons and the corresponding variable, it's plus one that tells me that the magnetic field is oriented in one direction and if it's minus one it's oriented in the opposite direction.
So I'm defining a distribution on states of my system. Now I would never hope to write down that distribution just in a table because then I'm writing down 2 to the n probability values and there's no interesting way to work with those objects. So instead what you do is you define a distribution implicitly through defining a Hamiltonian. So this Hamiltonian H is going to be a function that takes my configuration little x that lives in plus minus one to the n and it spits out the energy of that configuration. And for an Ising model all it is is it's the sum of two terms. One of them is a quadratic interaction term and the other term is linear in the x's.
And the way to think about what each of these terms do is to think about them in isolation. So if I had no interactions, then I would end up defining a distribution that's a product distribution. So it's independent across all of the coordinates. But my corresponding external field h_i would tell me, you know, how much the external magnetic field favors one direction over the other. Right? So that's what this first term means. This is this linear term in my energy function. And then of course what's critically important for us are the interactions between spins.
So that's captured by this interaction matrix J. That's an n by n matrix. And the way to think about it is that if you look in some, the i row and the j column, the value in the matrix tells you how much i and j prefer to point in the same or opposite directions. And you get things like ferromagnetism and antiferromagnetism based on some choice of, you know, whether these should be positive or negative signs.
So now we have a Hamiltonian, you know, given some configuration on the state space that spits out an energy, and the remaining thing that we need to do in order to define a distribution is we set an inverse temperature beta. So if you think about beta, which is our inverse temperature, is going off to infinity, then what's going to happen is that this distribution is concentrated only on the minimizers of H of x, so the ground states. But now as beta becomes very small, you know, the inverse temperature is becoming very small, so we're heating up our system, then we tend to favor distributions that have higher entropy even though they might end up with a larger energy.
And the entire thing, finally, we normalize it so that it is a distribution, and that's the partition function. So this is what an Ising model is. You know, there's a lot of beautiful mathematical questions that you can study about this that have very deep physical significance, and you can generalize it in a lot of natural ways, right? You know, what happens, for example, if I want higher-order interactions? It's not just pairs of spins that are interacting; maybe I've got triples of spins, quadruples of spins. That really is just a minor change on this model, where instead of having a degree 2 polynomial sitting upstairs for my Hamiltonian, I might have a higher degree, like a degree k polynomial, right? So any questions about this? Please stop me anytime if you guys have questions.
All right, so, you know, these are oftentimes called graphical models, and one of the reasons for that is because by looking at an associated graph to your Ising model, you can learn a lot of important properties about the distribution you're defining. So if I take my Ising model and I form a graph where I've got nodes, you know, each one represents a coordinate in my random variable, then I can create an edge set on this graph whenever there's a nonzero interaction. And by assumption, we're always going to assume that this interaction matrix J is symmetric. So we're really going to get undirected edges here.
And the critical defining property in the classical case, the thing that, you know, we exploit all the time when we talk about sampling or learning algorithms, is what's called the Markov property. So the way that this works is that if you're given two spins, I want to figure out whether the spin at site i is pointing up or down. I want to figure out whether the spin at site j is pointing up or down. In general, they won't be independent, but they'll have a lot of nice conditional independence properties. So if you tell me some other spins in my lattice, then it might turn out that based on what you've told me, i and j are now independent. Right?
And so the critical point is that this Markov property, you know, comes from the separators of the graph. So if you take your Ising model and you write down this graph based on which pairs of spins actually have a nonzero interaction, and you give me any pair of sites i and j, and you condition on another set of sites U, as long as they're a graph-theoretic separator, there's no path that starts at i, ends at j, and avoids U, you'll have the property that you are guaranteed that the distribution you've defined has this conditional independence property.
So this is the part that makes a lot of sampling and learning possible, is really exploiting this property. And, you know, if we're trying to build this dictionary between what happens in classical spin systems, what happens in quantum spin systems, I should warn you about the things that, you know, are going to come later when we talk about quantum spin systems. A lot of the things that we hold near and dear classically are just going to go completely out the window. So there are ways to talk about the Markov property; unfortunately, it fails very badly exactly for entanglement.
So, you know, one of the really fun things about talking about classical spin systems is you can draw very pretty pictures, and there's a lot of beautiful, deep mathematics. You know, it's led to fields, metals, a lot of interesting things have come out of it. And, you know, one way to visualize what these models look like is to take a typical sample from an Ising model, let's say a ferromagnetic one where neighboring spins prefer to be pointing in the same direction, and then to think about varying the temperature.
And so what I've drawn here is, you know, imagine that you have a white dot for a spin that's pointing in one direction. You've got a black dot for a spin that's pointing another direction. Now, if I take a typical sample from the Ising model, you know, at high enough temperature, this is what it looks like. It just looks like noise. It looks like a speckled pattern of, you know, black and white squares. But, you know, as you slowly decrease temperature, something very beautiful happens, which is you end up with almost all of the spins pointing in the same direction, right?
And this is what's called a phase transition, is when you have a dramatic change in the macroscopic properties of your system, you know, that are driven by these microscopic interactions. So people know lots of beautiful things, like in fact these transitions happen very sharply. So there's in general a very critical temperature at which you change very quickly from this picture to this picture. And there's a lot of deep works on trying to understand where those phase transitions happen. What are the algorithmic implications for them? What kinds of structural properties you can prove about a typical sample from this distribution. Right?
So, you know, now we should start talking about quantum spin systems. So the same way classical spin systems were a way for us to describe what a system looks like at thermal equilibrium, well, now I want to understand what does a quantum system look like at thermal equilibrium. So usually, you know, if I give this talk, which I've only given it a couple times, you know, I don't know much quantum, but the audience knows even less quantum than I do. So I have to introduce quantum. Here we seem to be in the reverse situation. So let me tell you how I think about it. Okay?
And that's just for me. You know, you guys already know all this stuff. But at least it'll give me an excuse to, you know, make sure that we've got the same notation in mind. So we all know what a density matrix is. So it's, you know, of an n-qubit system. It's a 2 to the n by 2 to the n matrix. Of course we require that it's positive semidefinite, and critically, it better be trace one, right? And at least the way that I like to think about these is sort of as a non-commutative generalization of classical distributions.
Like what if I took a density matrix, you know, it's 2 to the n by 2 to the n, it's positive semidefinite, it's trace one, but I add this further restriction that not only does it satisfy those properties, but also it's diagonal. So in that case, if it's positive semidefinite, if it's diagonal, if its trace is one, then I have no choice but all of the values along the diagonal are all non-negative and they better also sum to one. So this is our distribution sitting right there.
Remember I told you for classical distributions you never want to write them down explicitly, you know, as a 2 to the n table of all of the probabilities of every configuration. But that's exactly what the density matrix would do in the special diagonal case, because they'd show up exactly as these diagonal entries. Right?
So now we also have to think about what it means to generalize, you know, this two-local Hamiltonian that only cares about pairs of interacting spins or triples of interacting spins. And the way we think about this is we think about, you know, terms H in our that are going to compose our Hamiltonian, which we can build up in simple ways. Even though they're 2 to the n by 2 to the n matrices, we can describe them much more succinctly. Like what if I told you a 4x4 matrix A and then I tensor it with a bunch of copies of the identity?
Well, then we should think about this like, you know, a term in the Hamiltonian that only cares about the first two qubits because it's an identity on the rest of the system. And more generally, we can think about these, you know, k-local Hamiltonians as being ones where you tell me a 2 to the k by 2 to the k matrix that tells me what is the interaction between the corresponding k-tuple of qubits, but then you also tell me what its support is. So which k qubits it actually interacts on, and the rest of it is tensored with the identity. Right?
So I like to think of this obviously as a generalization of the types of local interactions that we use to describe what is the energy of our system in the classical setting. So now we're good because we can talk about what the quantum Gibbs state is. We've already seen it many times, but, you know, I'm going to think about the terms in the Hamiltonian as each of them being k-local. So they're the tensor product of k Paulis. So that already induces some normalization on the H's.
But then I'm going to allow you to add up these terms in the Hamiltonian with weights lambda_a, but I'm going to impose some constraint that maybe the absolute value of the lambda_a is at most one, because really what I'll care about is, you know, how this inverse temperature beta affects properties of my quantum Gibbs state. And I'm just choosing some normalization so that we have some well-defined, you know, way to measure what this beta is.
But the quantum Gibbs state is sort of the obvious thing. It was e to the minus beta H, and that's what it is now. We take our 2 to the n by 2 to the n Hamiltonian, we multiply by this inverse temperature, and now we take the matrix exponential. This is already a little bit different because before our Hamiltonian was a function that takes in the state of the system, but now it's this giant object that we can never hope to actually explicitly write down, right?
And of course, you know, by construction, you know, this quantum Gibbs state I've written down here is a density matrix. It's definitely PSD and trace one. And now I can get to at least one of the main questions that I'll be interested in for this talk, is the same way that we cared about, you know, how temperature affects macroscopic phenomena like magnetization that people have studied for a long time in statistical physics. What about how temperature affects, you know, the types of resources we care about in quantum spin systems? So how does it affect things like entanglement?
All right. So last little bit of, you know, quantum background for me, not for you, is, you know, how are we thinking about entanglement? So, you know, measuring entanglement in a general density matrix is a bit challenging to define, but, you know, for our purposes, I'll just be concerned more with the crude question of what kinds of states are entangled versus which kinds of states are unentangled.
So we can talk about a very special class of density matrices, product states, where we just take a bunch of rho_i's that are 2x2 density matrices and we tensor them together. And of course the result that we get is itself a density matrix. But, you know, the important thing is that these types of density matrices that we get in this way shouldn't really be thought of as quantum objects at all because they don't have any entanglement.
So one more twist on this definition is that maybe we won't just be interested in product states, but we'll be interested in mixtures therein. So I'll say that a state is separable if I can write it as a convex combination of product states where I can write down some of these rho_i's, right? And the important thing is that, you know, for our purposes, separable states are ones that only have classical correlations.
So if I talk about measuring these density matrices, I'm not going to get anything that's inherently quantum about it because I can realize these, you know, instead as basically being like product distributions or mixture of product distributions.
So let me say something trivial before we get to the main results. You know, there's always something silly you can say about the relationship between the temperature and entanglement, right? Because you can always send beta to zero. You know, remember beta is our inverse temperature. So what we're doing is we're just setting it to, you know, we're setting the temperature to infinity. In that case, it's very obvious what happens to the entanglement because if I were literally to set beta equals zero, all I'm going to get is I'm going to get some scalar multiple of the identity.
And this is something that is trivially separable because I know exactly how to write it as a product state. It already is written in exactly that form. So in some sense we already know a very crude statement that, you know, if the temperature were infinite, obviously there's no entanglement in my quantum spin system. But, you know, the point I want to make is that this is a non-physical temperature.
So if you think about quantifiably how small you need beta to be or how large you need the temperature to be, the point is that as my system grows, the number of qubits I care about, you know, I actually need the temperature to be growing too. In fact, you know, I need beta here to be like 1 over 2 to the n to say anything non-trivial about this. So this is not like, you know, it's not even like taking your spin system and throwing it into the sun. It's like doing something even worse to it. Right?
So likewise in classical spin systems, what you really care about are physically meaningful temperatures. What I care about are proving phase transitions about macroscopic phenomena where I can identify where the action is happening in a way that's, you know, scale-invariant of the system. Maybe it only depends on things like the strength of the interaction and the degree for each qubit. How many other qubits is it interacting with?
So that'll be the name of the game here, is I still want to understand the relationship between temperature and entanglement, but I want to do it in physically reasonable settings, which this trivial version of, you know, what I'm saying is definitely not right. So what happens at physically realizable temperatures? So let me tell you about our main results. So what we showed was that, you know, what's called the sudden death of entanglement. So this was the first mathematical proof of it.
So it turns out that if you have a quantum Gibbs state, remember that the terms in the Hamiltonian are tensor products of, you know, k Paulis, and we've normalized so that the coefficients we add them up with have absolute value at most one, then in that case, if we define two different parameters, d and k, so d is the degree, so you care about for each of the qubits how many different terms in the Hamiltonian hit it, right? That's the degree of interactions.
And then you care about k, which is for each of the terms in the Hamiltonian, how many different qubits does it touch. Then it turns out that if those are constants in your spin system, which usually they are, then there is some universal constant that, you know, when the temperature is above that threshold, it's not just that the entanglement is very small, but that it's literally zero. There's not a shred of entanglement in this, right?
So in some sense this result was a little surprising to us. In fact, we proved the theorem without knowing we proved it, which is kind of an interesting quirk. The thing we were really interested in was an algorithm for preparing quantum Gibbs states. So what we wanted was an efficient algorithm that maybe sidestepped our inability to prove things about Lindbladians or Davies generators, and we wanted some other algorithm that can prepare quantum Gibbs states.
And I'll tell you what that algorithm is, but somehow what was kind of a cute accident was that as a byproduct of showing that algorithm worked, we accidentally proved this theorem, you know, about the death of entanglement. So somehow you get this very interesting physical phenomenon as a result of taking an algorithmic perspective on things.
All right. So let me pontificate about some stuff. Maybe this will be controversial. I don't know. But I'll just say some things before I tell you about the proof. In fact, the proof is not too hard. So I can give you the main ideas of the proof just in whatever time remains for this talk. But I want to take a step back, you know, at least as an outsider to, you know, quantum spin systems, and at least tell you some thoughts that are going around in my head.
So if you think about, you know, quantum computation, certainly one of the main goals is to give quantum advantage, but for useful problems. So we all know about, you know, Shor's algorithm. So we know that if we had, you know, general-purpose quantum computers, we could do things we believe are not solvable in polynomial time. We can break RSA, and, you know, the world would forever change. Of course, we don't have general-purpose quantum computers yet. Maybe we will in some time. You know, probably doesn't seem like it's anytime soon.
But one of the goals that people put up in the interim was quantum supremacy, and there it's a very different change on, you know, what you're aiming for, because what you want is you want any problem, no matter how artificial, that quantum computers that we currently have seem to solve and classical computers seem not to be able to solve. So there are some caveats that come with that, which is that a lot of times you end up with, you know, problems that aren't so useful. They don't help you do interesting things because you build some quantum device and it sort of computes whatever it does in nature.
And then even worse, you sort of don't get much. There's no great ways to have a system of beliefs about whether there aren't better classical algorithms for doing it. There's been a lot of ebb and flow on what people believe. But, you know, certainly a lot of us here in this room believe that one path to useful quantum advantage is quantum Gibbs samplers, right? So we could hope that maybe, you know, quantum Gibbs states describe the natural world around us.
Maybe they help us probe in the sciences to understand quantum phenomena in chemistry or condensed matter physics. And we could hope that if we have, you know, even near-term quantum computers, that we discover new science as a byproduct of it, right? But, you know, the point I want to make, at least thinking about it from a complexity theory standpoint, is that there's a limit to how optimistic you can be about this.
Let's say our goal was to do quantum advantage for some useful problem, you know, for science, for quantum chemistry, condensed matter physics. And even in our, you know, most optimistic of A, the trouble is that there are some things we probably shouldn't expect to do with quantum computers because we know even classically these things are computationally hard.
So the thing is that even for classical spin systems, you know, it's not just that there are phase transitions in things like magnetization, but there are phase transitions in everything you could imagine. There are phase transitions in temporal mixing. There are phase transitions in strong spatial mixing like in UNTO. There are even computational phase transitions where the problem of sampling from the corresponding Gibbs distribution or approximating the partition function goes from computationally easy, solvable in polynomial time, to being NP-hard.

2 more exchanges in this transcript

Sign in free to read the rest of this interview. No card required.

Sign in to read the full transcript

Cite this transcript

APA, MLA, BibTeX
APA

Moitra, A. (2026, January 13). Ankur Moitra - Transitions in Quantum Spin Systems - IPAM at UCLA [Interview transcript]. Institute for Pure & Applied Mathematics (IPAM). CEOInterviews.AI. https://ceointerviews.ai/interview/3143456/

MLA

Ankur Moitra. "Ankur Moitra - Transitions in Quantum Spin Systems - IPAM at UCLA." Institute for Pure & Applied Mathematics (IPAM), 13 Jan. 2026. Transcript, CEOInterviews.AI, https://ceointerviews.ai/interview/3143456/.

BibTeX
@misc{moitra2026_3143456,
  author       = {Ankur Moitra},
  title        = {Ankur Moitra - Transitions in Quantum Spin Systems - IPAM at UCLA},
  howpublished = {Interview transcript, Institute for Pure \& Applied Mathematics (IPAM). CEOInterviews.AI},
  year         = {2026},
  month        = {jan},
  url          = {https://ceointerviews.ai/interview/3143456/},
  note         = {Speaker-attributed transcript with timestamps}
}