>> Yuval Peres: So welcome. We're delighted to have two lectures today, and the first is by Naomi Feldheim on sine changes of stationary Gaussian processes. Please, Naomi. >> Naomi Feldheim: Thanks. So thanks for the invitation to visit Microsoft, and indeed that's talking about that. It's joint work with my husband, Ohad. And so this is about a simple question that is first defined with our Gaussian stationary processes, then address the question. So definition. I'll begin with the space T to be either R or Z, and that's my time space, and I'm looking at random function F form T to R, which is Gaussian stationary. So Gaussian means all finite marginals have Gaussian distribution, so for any N and for any points T1 till TN in my time space, looking at AF1 till FTN. This has multi-normal distribution, mean zero. Of course, in RN. And secondly, stationary means that my function is invariant with respect to shifts in my time space. So for any S in my time space and, again, any points, if I look at the shifted values by S, this shouldn't change the distribution. So Gaussian processes are very natural for noise, as we know, we like to model things by Gaussianity, and stationary is a very natural property because it describes invariance with time of our noise. So this is a well known family of noises. I'm excited in particular, there is some newer interest in this, in the context of the analytic Gaussian stationary function, so if we add the assumption that almost surely our function is analytic in the case of real time space. In this case the zeros of the function, they form a well-defined point process because they are separated, and they have repulsion. So it is one of the maybe very few processes that we know to construct and study that have repulsion. But today we'll talk about just general Gaussian stationary processes. Okay. So we'll be addressing the following question. Look at the probability, which I denote by HN, probability that our random function stays positive between time zero and N. So no sine changes, or if you're looking at the zeros, it means that there are no zeros at all between zero and N, or maybe it's half the probability that there are no zeros at all. That's why it's called the whole probability, and that's why there's an H here. Random matrix people and people from mathematical physics sometimes call it the gap probability, so it all refers to the same thing. So I guess this is a very natural question to look at, and one would like to understand the symptotics of this is as N grows. So let us look at several simple examples. So example one, let's just take -let's do some examples on Z. So by the way, when my time space is discrete, I'll call it sometimes a sequence, and when my time space is continuous, I might call it just a function. So let us look at sequences. So draw it up here. Example one, 1Z, just FNs which are IID. Then, of course, HN is just half to the N. So there is exponential behavior with N. One might guess that if correlations are weak, then this exponential behavior should persist. So here correlations are as weak as you can get. There are no correlations. So what happens if correlations are zero, say, beyond some point, maybe there is also exponential behavior. So not exactly. Let us look at another simple example. Let's look at the vector at the sequence YN, which is the difference of an IID sequence, so XN minus six, XN plus one minus 6N where XN are IID. So now the whole probability is just the probability that the first difference is positive and then the second difference is positive and so forth, so it's the probability that they are ordered, that the XIs are ordered in a monotonic way, and while this probability is roughly E to the minus N again. It's one over F plus one factorial. So this decays much faster than exponentially. Then after you pose for it and think, well, can this probability decay slower than exponentially? So the answer to this is also yes, it can. So there is very trivial example. It says just take the constant sequence Zed N equals Zed zero. So they're all just equal and you randomize only the first one. Then certainly HN is just a half. It does not decay at all. Okay. One might say that this is maybe degenerate example, but it is possible to construct also nondegenerate examples where the probability does decay but slower than exponentially. So we want to ask what is responsible for the exponential decay. Looking at some sequence, how can I know if it behaves roughly like IID sequence in this since. So question is when does HN behave roughly like an exponential. So we keep this question in mind, and in order to give some characterization of this, I need some more terminology from the world of Gaussian processes and Gaussian stationary processes. So this would be our next part. So some terminology. Okay. I would like to look at what's called the covariance function, which gives the covariance between the value at zero and the value at some point T. Of course, by stationarity, this is the same as the covariance between the value at S and the value at S plus T for any S. So this is a function of one variable for my time space to R. And now since it describes covariance, this is a positive-definite function. In case I work over R over the wheels, I'm going to assume that it is also continuous. I'm going to work only with continuous correlation functions, so assume continuous. And now harmonic analysis says that if we have a positive-definite function which is continuous, then it is the Fourier transform of some positive measure. This is served by Bockner [phonetic] at least in the real case, so there is some measure such that this is the Fourier transform of the measure. So let us explain this. First of all, this is an integral over T star, the dual space. The dual space of Z is minus five pi. The dual space of R is just another copy of R. You might think of minus 5 pi or the circle. That's the same for me. And what is raw? So raw is a positive, so it's a measure on T star. It's nonnegative, and it is finite, and I will also assume some small regularity condition on raw. I will assume that for -- there is some small epsilon such that the momentum polynomial -- the polynomial moment of epsilon converges, so this integral converges. This, by the way, is enough to ensure that our processes is almost truly continuous and I just -- I need it for some regularity conditions. So we have our spectral measure. Now, the covariance function, just looking at it, is a function Fourier variable determines everything about the process. Why is it so? Just because Gaussian processes are always determined by finite -- by their marginals and the marginals, finite marginals are Gaussian, and therefore, they are determined by the covariance matrix. That's all you have to know to answer any probabilistic question. So RT determines everything about the function, and raw is a one, one bijection with it, so if you know raw, you know everything about the process. There is one more property that raw must obey in order for everything to be defined, which is symmetry, so let me add it here. So raw is symmetric around zero. Just because I'm in order for R to be real, raw must be symmetric. And so it turns out that any measure that obeys all of these conditions defines actually covariance function and defines some function abstractly. Okay. So this is quite abstract. Let me tell you a way how to construct the function with a certain spectral measure. I will not use this directly, but I think it is -- it helps us understand the object, so if you're given raw, you should take psi N, which is an orthonormal basis of L2 raw of the space, and then take weighted Fourier transform of this. After you have this Fourier transform basis, you can define a function, a random function F, random sequence, which is sum AN phi NT where any N are IID normal random variables. So it is not very difficult to check that this random sound indeed converges and converges to some Gaussian process under the margin of the Gaussian to a stationary process and the spectral measure is indeed raw. So this is a quick way how to construct such functions. Okay. As I see some examples, so first if I take, say, a sequence XN which is IID, what would be the spectral measure? So in order to compute this, I first need to compute R and the covariance, so R of N is just delta 0N. The Fourier transform of this is just an indicator or a constant function indicator of minus 5 pi. So that's my measure in this case. Good. Now, this measure can be viewed also as part of the real line. So now suppose I want a function whose measure is the indicator of minus 5 pi. I regard it as part of the real line. What function would that be? So what I need to do is begin with raw, which is the indicator, and then construct R, which is now the real Fourier transform of this. So this is a sync function. So this would be the covariance, the correlation between two points in distance T of such a function. And now if I really want to understand what it is, I can follow the recipe here of constructing it and construct it as a random sound. So let me tell you the answer, write it here. So the answer is FT being sum AN sine pi T minus N divide by T minus N where AN are again IID [indiscernible]. So what's going on here? You have a sync function. So that's function which looks like this. It vanishes at all other integers, and these least one and zero, and you shift it to any integer, and at every integer you take your random normal amplitude. Then you sum this all up, different things, and you get your random function. So in this form here you can see that FN for a certain integer is exactly AN. So this is indeed an IID sequence. So this function is rather interesting because it is some kind of real stationary interpolation of the IID sequence on the lattice. Right. So this is the function it has spectral measure of the indicator. >>: In your construction of the Gaussian process from the measure? >> Naomi Feldheim: Yeah. >>: I guess is there any easy way to see that F of T is stationary? >> Naomi Feldheim: I would compute the covariance and just it's for, yeah, probably part of the -yeah, and by the way, I did not say, but you probably have to worry about the realness of phi N because this has complex numbers, but it is always possible to do this construction so the phi would be real. Just choose psi N to be odd and even functions, but yeah. Right. So that was our first example. Second example. The second example that we had before, so YN, the sequence which is XN minus XN, XN plus one minus XN. Differences of an IID sequence. So what's the spectral measure here? So first we have to compute the covariance function. So at zero it's two. That's the variance at the point. At distance one or minus one it's -- well, just minus one. And from there on it's just zeros. And now we have to Fourier transform this to get your measure, and this turns out to be two, one minus cosine lambda. So that's like the density, which looks like this. Okay. So this is our second example. Not much to say here. Third example which might -- we might keep in mind, similar to number three there, but write something a bit more general. If we have raw, which is an atomic measure, so sum of sum weights, and then some delta minus lambda N and delta at lambda N. I must keep it symmetric because I want the Fourier transform to be gree [phonetic]. So suppose I have something like this. Then by the construction we have, we Fourier transform we get that our function or sequence is sum with weights of cosines and sines. So it's where AN and B now are both IID norms. So its sum of random waves we've given frequencies, lambda N and some given weights. Now, taking delta measure just at zero will give you the constant function, the examples that we had before. So this is some generalization. Okay. So I think this gives us some understanding between the spectral measure and the process, itself. So at last we are ready to state the result. So on there, so say that the Feldheim theorem. I want to indeed give some condition for our process to have exponentially decay. I'll give separately the conditions for upper bound and lower bound. Even so, they're quite similar. So first let's just give the condition for upper bound. So suppose there is some small alpha and some positive number M, and positive number big M such that if I compare my spectral measure to the Lebesgue measure and it is between those two constants, for any interval inside a small neighborhood of the origin, then there is an upper bound. So, of course, I mean for large enough N, I can borrow this from above my exponential. Lower bound very similar, but I need only the bound from below. So I compare my measure to Lebesgue measure. I want it to be bounded from below for any I in some small neighborhood of the origin, then it is bounded from below. So a few remarks here. First, this comparison to Lebesgue measure, I think the intuitive way to understand it is that Lebesgue measure or the indicator we said is somehow related to an IID sequence. So the indicator of minus pi pi is exactly the IID sequence on the integers, and so an indicator of some smaller interval, the spectral measure of such a thing would be independent on some different lattice, a more space lattice. And so we are comparing it either from both above and below or just from below. Second remark is that indeed for the lower bound we need only the lower bound on the comparison. Here we need both, but maybe it's just part of our method and not truth. Maybe the truth is that you need only comparison from above to give this above. Actually, we can refine this a bit. We can ask for an upper bound near zero in some neighborhood of the origin and the lower bound somewhere else. Just give me it anyplace where this lower bound holds and not necessarily in minus alpha alpha. So the proof still works. So this hints that the lower -- this can be improved. Moreover -- yeah. >>: The lower bound is somewhere else? >> Naomi Feldheim: means that ->>: So lower bound somewhere else The interval or -- >> Naomi Feldheim: Yeah, that this bound is the same and this bound for any I in some other interval may be very far from zero. So -- yeah. I need comparison to Lebesgue measure somewhere from below, but not necessarily in the same interval. Right. And the second thing I want to remark that constants here are very different and it would be interesting to understand if there is really exponential behavior asymptotically. So now let me tell you a bit about the proof in the few minutes I have. And how does this show up. Oh, by the way, before as I am erasing the examples, you can see that in the examples we had, number two, which was faster than exponential, indeed has spectral measure that does not obey this bound from below. It is zero, near zero. So it obeys the upper bound but not the lower bound. It's faster. Okay. So idea. So the main observation is that if you take a spectral measure which is actually the sum of two positive finite measures raw one and two, then the corresponding functions, the corresponding function F can be written as the sum of two independent copies, F1, and F2. F1 has spectral measure raw 1 and F2 has spectral measure raw two. So this is very easy to see this linearity from the definitions, but it is very useful here. How so? Our conditions, so here both conditions our lower bound gives that you can write raw as maybe M times indicator of minus alpha alpha plus something. Now, the indicator corresponds to something that has independence on the lattice. So our function F is sum of F1 plus an independent copy of F2 where F1 -so if we suppose -- so F1 has this spectral measure, so if we suppose alpha, for example, is pi over K for some integer K, then F1KN is IID. Then we want to use this IID sequence in order to bound from below and from above. So let me assume just for simplicity that K equal one, so just F1 on the lattice is IID. Okay. So for the upper bound. So both for function and for sequence. So this was my original whole probability. I can always bound it by some lattice or some sublattice, so this is the sequel then being positive for every N on the lattice I chose, so in this case I choose just the integers one, two to the N. Oh, so that's not F1. Sorry. That's F. Still not F1. So my original function on the lattice. Now, how would I bound this? So as we recall, FN is just F1 plus F2 where F1's are actually just IIDs and F2 is some other Gaussian sequence. So I want to spread it into sum of two events, sum of two probabilities. So the first probability is just F1 being bigger than minus F2N for every N, but given that the sum of F2N is small, smaller than, say, one. The reverse. Sorry. Okay. So the rest I will just bound by the probability that this did not happen. So if this did not happen, the sum of F2Ns is bigger than one. Now I want to show that they are both square. So this is the event I want, but given something that happened to F2s, and this is just probability that this did not happen. I want to show that each is exponential. So for the first parts for -- so that's, say, probability, the first probability you just have -- I will write it. You just have an IID sequence being bigger than some number minus CN where CNs you can think of as given numbers that their average is less than one. That's what you want to compute to bound from above for any N. So of course, this is just multiplication of the probabilities that -- of the standard random variable to be bigger than minus CN, so that's three of CN. And now we use the fees, the local K function and also a monotonic function in order to bound this by E to the N fee of one, something like that. So this is just some Gaussian computations. So that's for the first part. For the second part, this probability, so you notice that what you have here is also a Gaussian random variable. It is the sum of Gaussian random variables, and so you just have to compute its variance and understand this bound. So this is again a Gaussian bound where the hint is that the variance of the sum is roughly the spectral measure at zero. So this is by doing some harmonic analysis. Okay. Lastly, about the lower bound. So lower bound. I rewrite this probability I'm interested in, so this is the probability, and now I want to again use this decomposition into F1 plus F2, in order to -- yeah. >>: So in part two is where -- >> Naomi Feldheim: >>: So this -- -- use the -- part of two you're -- >> Naomi Feldheim: This is -- so, yeah. So there is a sum of two probabilities. I want to show that each is exponential, so this will be ->>: But in part one you just use the lower bound on the -- >> Naomi Feldheim: here, so -- But maybe there is a one over >>: But the upper bound spectral measure is used in part two? >> Naomi Feldheim: Yeah, it's used in part two. It's used in part two. I'm saying that this variance is roughly what happens with the spectral measure at zero and I want to bound this from above, exactly. So this is roughly spectral measure at zero, which is less than some constant big M divide by N. Exactly. Then once you have the variance, it's just a Gaussian computation. Right. Now for the lower bound. So I want to construct some event using my decomposition that will make my function positive on the whole interval. So for instance, I can ask for F1 to be bigger than one on the interval, and simultaneously F2 should be less than a half, say. >>: That's a derivative of a spectral [inaudible]? >> Naomi Feldheim: >>: Right. Yeah. [inaudible]. >> Naomi Feldheim: Yeah. Just M. In case of this condition, actually the spectral measure is absolutely continuous in this small neighborhood, so there is meaning for this. Okay. So this is just the product of two probabilities because F1 and F2 are independent, and now I have to study each of them separately. Now, the first -- the second one has been studied. It's just probability that some Gaussian process is in a small ball as in ran use half for a long time, and so it's called the small ball probability. There are many bounds on that. So we have to do some work. We can apply some bound by telegrand [phonetic] and we're able to bound this by an exponential bound below. Actually, there are bounds from above. As for the first one, what's written here is a very concrete question. We have the concrete function F1 of T, which is actually written here. That's F1 of T because it has spectral measure of an indicator, and I want to understand what's the probability that this is bigger than one for a long time. And so this was done or similar work was done by a few guys a few years ago, so it's Antazena Bachme, Malzo and Torson [phonetic], so actually we have some improvement on their proof, but the main idea is to look at this special series and to construct an event on the series. So for instance, the event will be more or less asking N of the LANs to be bigger than two and show that the rest of the things do not ruin your event and that this event is exponential and yields this. >>: To lower bound -- >> Naomi Feldheim: So there is some construction. >>: -- small [indiscernible] for F2 you could just use the fact that you can borrow a little bit of the measure from -- I mean, from the upper bound. >> Naomi Feldheim: >>: For this, you mean? Yeah. >> Naomi Feldheim: The small probability? Yes. So there is some raw truth that is left and there is the function F2, yeah, if you apply -- you have to apply some small [indiscernible]. But -- so this is more or less the idea. So I'm at end just with some interesting open questions here, so I think this is just the beginning of understanding asymptotics, as you see that we have only bound but they are very far from each other and be interesting to understand the real asymptotics. Also would be interesting to understand what happens if the spectral measure blows up at zero or if the spectral measure is indeed vanishes at zero, or even in some interval around zero. How does this affect the whole probability? And if you are not accustomed to the language of spectral measures, I'll rephrase it. This means that the integral of LX, meaning the correlation function is infinity and this means that the correlation function integrated is zero. So in a sense, we show that if it's finite but not -- but somewhere in between, then the behavior should be exponential, but what exactly would happen if it goes to infinity or to zero. Our guess is that it should depend on the weight that it goes to infinity or zero, but we don't have any concrete work on that. Okay. So that's all. Thank you very much. [applause] >> Ohad Feldheim Ohad Feldheim: >> Yuval Peres: Any other questions? >>: So seems like the obstruction imposed by example number two does not exist if you assume that the correlations are positive. Do you have any idea how to apply the fact that correlations are nonnegative and ->> Naomi Feldheim: Yes, that's a good point. So if correlations are positive, indeed, it's very easy to show that there is existence of the limit. So if all correlations are positive, so there's existence of the limit log of this probability divide by N, because if all correlations are positive, you can show that log of this probability are -- there's some sub [indiscernible] and ->>: [indiscernible]. >> Naomi Feldheim: Yeah, and there is -- so exactly. So existence of limit is a one line proof, but still it would be interesting also to say what is this limit and it does not apply to correlations like the sine [[indiscernible] correlations. >>: That limit is positive, can you infer anything about the process? >> Naomi Feldheim: >>: Again? Suppose it's positive. If limits is -Can you infer anything about the process? >> Naomi Feldheim: I guess that for many processes it would be positive, so ->>: Yeah. It rules out certain processes. >> Naomi Feldheim: Oh. >>: If you say anything causes about the process that can be ruled out by this [inaudible]. >> Naomi Feldheim: Yes, so again, my guess would be that the information you get is precisely that you do not have either blowup or zero in -- at zero. So I guess that it would be exactly in correspondence with saying that there was finiteness of the spectral measure at zero or finite integration of the correlations on the whole line, I guess, but I think we do not know something. >>: My compliments to [indiscernible] to say something [indiscernible] Direction. >> Naomi Feldheim: happens. >> Yuval Peres: Yes. Okay. >> Naomi Feldheim: Yes. I agree. Yeah. That Thank Naomi, once again. Thank you. [applause] >> Yuval Peres: So for the second part Ohad Feldheim will tell us about 3-coloring the discrete torus. >> Ohad Feldheim: Yes. So thanks Yuval, for the invitation and the hospitality and, well, it is going to be completely different now with the presentation and not a blackboard talk, and the topic is completely different, so I hope we'll make this context switch easily. And I'll tell you today about 3-coloring discrete torus and the there is also a subtitle for people interested in statistical mechanics which said the rigidity of zero temperature 3-states anti-ferromagnetic Potts model, and I will explain it as we progress. And this is all joint work with the Ron Peled, also from Tel Aviv University. So the object we're going to deal with is a proper 3-coloring of the discrete lattice ZD. So we take the lattice with nearest neighbor adjacency, so for example, in two dimensions every vertex is adjacent to four other neighbors, and we would like to take uniformly chosen 3-coloring, prophecy coloring of this lattice, but of course, since it's infinite, we can't. So we must limit ourselves to some boundary conditions. There are two classical ways of doing so. So one way's what's called 0-boundary conditions, so we take some bounded set, and this set, we choose it to have an even boundary. So what do I mean? The lattice in every dimension is a bipartite graph, so we want all the vertices on the boundary to be of the same bipartition class, let's say from now on the even, even by position class. We color all of those vertices by the color zero, and then we pick uniformly chosen 3-coloring. And the periodic boundary condition is when we will actually just coloring the torus, so we will have N to denote the side of the torus in each direction. And what we do is we look only on the set from zero to N minus one in each coordinate and we wrap around, so this one is also adjacent to this zero. So this is a discrete torus. We're looking with 3-colorings of this. The benefit of this boundary condition is that it didn't force many vertices to be colored by particular color. So in a way it might be more natural boundary condition to reflect our thought of the infinite, of the infinite lattice. So we will be interested in uniformly chosen purposely colorings of those objects, and we're only going to discuss high dimensions. So imagine that the lattice is of dimension, say, 10,000, and what my pictures depict is just a two-dimensional section of it. Otherwise, things would not be -- I mean, they would not be representing it correctly. And, well, we could say that there is a natural combinatorial motivation to look on this object because while the lattice is natural graph and the 3-colorings are natural here because there are only two 2-colorings of those sets, so the coloring is the minimum number of colors where this would make sense. But there's also some motivation from statistical mechanics, so there is a phenomenon called anti-ferromagnetism, and in the narrow sense it reflect the fact that on a solid matter the electrons of adjacent items or molecules tend to take opposing spins, but more generally it is a situation wherein a solid matter in lattice, a adjacent sites tend to have a different state. This is modeled by the q-states anti-ferromagnetic Potts model, which is the model of q-coloring of the lattice where every configuration is chosen with the proportion which is inverse exponential of the number of forbidden edges. Those are edges that connect the color to the same -- two vertices colored by the same color. And when we -- so there is a parameter here bitter which says how much we dislike those edges, and when bitter is taken to infinity, which corresponds to temperature being taken to zero, then we are left with just q-colorings of the discrete torus. So this is how those two problems relate and, well, we'll get back to this idea mainly in our open questions because we would very much like to generalize what I'll tell you about today on the 3-colorings of discrete torus to the real Potts model, and with bitter, which is not infinity. So what could be asked about this model? So I think very natural question is what is the relative frequencies of the color in a typical coloring. So for example, is it one set, one set, one set. So I pick a coloring at random. Will I see one of set of the vertices colored by zero, one set colored by one, one set colored by two. And we can ask more sophisticated questions, such as does the typical coloring follow some pattern. So if we see a coloring, does it look like a random one. And in particular, there's a question which physicists think interesting, and this is there is a long-range correlation or what is the nature of long-range correlations, so what does this mean? Imagine that our torus is huge. So I said that D is large, but imagine that N is much larger now, and with we've examined it in a small region, and we want to know what can we tell from its behavior in this region very far away from them. And if all centralized observables on those two regions, their correlation in the case exponentially with their distance, we would say that there is no long-range correlation. Now, after studying this object, one can make the following educated guess. So maybe a coloring looks roughly like this. So what is this coloring? All the even sites here except for one are colored by zero, and the odd lattice points are colored uniformly by -- uniformly independently by zero, by one and two. So I guess that coloring makes sense. So indeed, we've limited ourselves very much in respect to the even sublattice, but now on the old sublattice we have complete freedom. So if all the vertices were colored -- all the even vertices were colored by zero, we would have a two to the number of vertices over two configurations. Now, of course, it pays to have some vertices on the even sublattice not colored by zero because, well, for example, the first one that we had gives us another degree of freedom equal to half the number of vertices and all that we lose is that it's forced its neighbors, but we can expect that as the dimension becomes higher, the percentage of those vertices will decay dramatically. So is this really how it looks? So the conjecture is that on two dimension, not at all. So my pictures here, if you thought of them as two dimensions, it's not the case. But for all dimensions higher than two -- and higher than two because there are some senses of fractional dimensions and in those senses really you can see the two should be the critical one, the critical dimension. So for all dimensions higher than two there is some rigidity. So there is some behavior like this that on the one sublattice the vertices tend to take the same color. Okay. So this conjecture has been established for 0-boundary conditions in high dimensions. So what Peled shows in 2010 that the typical 3-coloring with 0-boundary conditions has nearly all the even vertices taking the color 0. So formally we say that if d is large enough and we have uniformly chosen 3-coloring with 0-boundary conditions has the expected number of even vertices not colored by zero out of all even vertices, decays roughly exponentially with the dimension. Yes. And whether this log factor is really necessary is open. So this result didn't work for periodic boundary conditions, and the reasons will become clearer as I advance along this talk. And of course, it's opening low dimensions, and as I said, it's widely open and when we have no clue how to advance towards it, it's not clear that combinatorics is the right approach to get to low dimensions. Could be that one should do completely different things. Now, towards showing this on bounded tori, we have the following remarkable result by Galvin and Engbers. So what they show in our context is that if you fix N and you take high enough dimension, so this is not what we wanted. We wanted to take large D but to take N to infinity. Then the result holds. Typical 3-coloring with periodic boundary conditions is nearly constant on either even/odd sublattice. So of course, this is the counterpart of the result I've shown on the lattice slide because, well, when we're coloring with periodic boundary conditions, we have no bias towards the even sublattice or the odd sublattice or towards one of the colors. So this is what we would expect. So why is this so remarkable? Mainly because it works also for q-colorings, and there are very few results that work for general q-colorings, and as you'll see when I'll explain a bit about the proof of a 3-coloring, you will see that we really use something special about them. It's really not easy to generalize it. But it is not our purpose. So what I will talk about today is the parallel phenomenon for periodic boundary conditions. So what we show is that a typical 3-coloring with periodic boundary conditions is nearly constant in either the even or the odd sublattice, so this was our goal. And formally, so what is the formal counterpart if d is large enough and we uniformly choose a periodic boundary conditions 3-coloring, and now we're looking on all colors and both sublattices and we look on where there is a color that is so dominant that there is as few as possible vertices not colored by this color. So the number of vertices not colored by this color of all the vertices on this sublattice will decay exponentially with the dimension, roughly. So this is more or less natural. Natural generalization. Okay. Observe that when I said that there is even and odd sublattices, I have implicitly said that N must be even. This is not a weakness of the proof. It's how things really are. Odd tori do not resemble the entire lattice in the sense that they are not bipartite, and so conjecture like this that I've presented here cannot hold for them, and something else happens. We have a very good idea about what happens in odd tori. Nothing of this is written, but maybe it will help for other results. We're not sure it's interesting enough on its own, on its own accord. And the proof that I'll tell you about today introduces topological techniques to the problem, so it's not that we're going to do again what was done for the periodic boundary conditions. Actually, we're going to use that result, but it's going to be not so simple to transfer it to the torus, and this is what I will explain, how we do this transformation. So what I'll tell you about now is what is the real challenge. I mean, well, where is the effort in doing this. So in order to tell you anything about proofs in 3-colorings, I absolutely must tell you about homomorphism height functions. Those are the objects that allowed us to understand so much better 3-colorings that has the numbers of close. So what are homomorphism height functions? Those are simply homomorphisms form a graph G, in our case the ZD2Z, and this means just a mapping from this graph ZD2Z that satisfies that adjacent vertices have their images deferred by exactly one. And you should think about it like a topographical map on our set. So this reflects the height and the requirement of difference one represents some kind of continuity. And like we did for 3-colorings, we can pick a uniformly chosen such height function with certain boundary conditions. So for 0-boundary conditions everything is just the same. Pick a set with even boundary, color its boundary, so pick the height of its boundary vertices to be zero, and then pick the rest uniformly along such -- among such a height functions. For periodic boundary conditions we must add at least one vertex whose height we fix, because if we will not do so, we will have a degree of freedom of lifting the entire height function by any constant. So we pick one such vertex and fix it and we call this usually pointed periodic boundary condition uniformly chosen height function. So why do I tell you about this object? So actually, on ZD, this object in 3-colorings are just the same object. This is the same. So what do I mean? There is a bijection between those two objects, and it goes as follows: So if I give you a height function and I take a modulo three of it, you will get a proper coloring. Well, of course, if two vertices their height differs by one, then modulus three, their height will also differ by one, so they will not be -- will not get the same color. To see that this is a bijection, we have to explain how to calculate the height of a particular vertex, so how do we do it? So we can think of the coloring as encoding height differences on this lattice. So from zero to one we go up by one, and from one to two we go up by one, from two to zero we go up by one, and in the other direction we go down. And now to calculate the height of a vertex, what I can do is to integrate those height differences along a path. So for example, if I want to know what was the height of this vertex, I am saying, okay, from zero to one I go up by one, so it's one. From one to two I go up by one, so it's two. From two to zero I go up by one, so it's three. And now from zero to two I go down by one, so it's two again. And from two to zero I go up by one and it's supposed to be three, and indeed, this vertex is called by height -is the height [indiscernible]. So how can I convince you this is really a mapping? I should show you that on different passes you get the same height. So this is the same as showing that the longest cycle you don't accumulate any height. To show this, it's enough to show that on the cycles which generate all the cycles, you don't get any height. In all ZD, in all -- no matter what is the dimension, the cycles are generated from small four cycles. And now to see that you can't gain height on a four cycle, you can, for example, say that it is even, the height that you gain, because you go over four vertices, and the resulting height might be the same modulo three as the starting height, so it would be the same modulo six, and so in four steps moving from one value modulo six to another value modulo six, so it must be the same value. Okay. Now, this bijection of -- sorry -- does not extend to the torus. Well, actually, this is an example, right? So this coloring matches only this height function. And this seven, say, and this zero, they do not fulfill the requirement to be a height function. And this is main obstruction. Why so? Because we know very much about height functions. So in that paper of Peled from 2010, he shows that no matter if it's on the lattice or on the torus, height functions have what we want. They tend to be a nearly constant on either even or odd sublattice. And for 3-colorings, the way that the result is obtained is just saying, okay, these are the same objects, so they also satisfy this property. And the challenge here is that because we don't have the bijection, and we wish to transfer this to -- the result to 3-colorings, then we would have to do something else. We will have to somehow go around this problem that we don't have the bijection. Okay. So in order to try and understand 3-colorings on the torus, now that we don't have bijection with height functions, let's see what we can put them in bijection with. So we take the coloring on the torus, and one thing that we can do with it is to pull it back to the plane. All right? So the plane is ZD, the lattice ZD is universal color of this torus, and so every function here could be pulled here. And what we do is just put copies of the same coloring all over the plane. Now, on the plane we have bijection, so we can apply it and get the height function. And we can ask ourselves, what kind of height functions will we get? So they will not be periodic like the coloring, not necessarily, but at least because the coloring is periodic, then the height differences would be periodic. So if here we have some height difference, then in the next copy we'll have the same height difference. So functions whose slope, whose gradient is periodic, we call them quasi-periodic functions, and they are defined by the heights on one copy, on one quasi-period, and by how much height do you accumulate when you wrap around the torus in each standard basis direction. Okay? So here, for example, you have a slope six when you go along E1, and you have slope zero when you go along E2. So these are the slopes that they write here on this site. And it is not so hard to convince ourselves that, first of all, we'll always get slopes which are zero modulo six. The argument is similar to what I explained before. A vertex in its translation must be the same modulo three in height and must be the same modulo two, and that every such quasi-periodic height function defines a periodic 3-coloring. So we have this bijection. Now, if we were lucky, let's say, and our coloring turn out to be a quasi-periodic function with slope zero, then it's excellent. We can pull it back to the -- we can push it back to the torus and we will have exactly the function which corresponds to this coloring. So in fact, 3-coloring, flat 3-colorings correspond homomorphism height functions. So what is the plan halfly [phonetic]? We have 3-colorings. We want to -- we know that they're the same as periodic 3-colorings on ZD, which are the same as quasi-periodic functions on ZD, and we know that the periodic functions on ZD are the same as the homomorphism height function on the torus, and for them we have the result that we want. So the entire challenge, the entire purpose of this work is to show that those two things are roughly the same, that is, that if we pick a quasi-periodic function on the ZD, usually it will be periodic. And how are we going to do this? So first abstractly we know that flat 3-colorings are in one-to-one correspondence with height functions on the torus. Now we have those slope 3-colorings. What you would like to do is we'd like to flatten them. We are going to map them to flat 3-colorings and then use this mapping to get height functions on the torus. Now, we would like to show that, first of all, our mapping is few to one. So not -- there are not many sources to a particular height function; and second, that the resulting height functions are very unusual, and because they are very unusual and we know that this is a small subset and this is few to one, then small times few is not so much, and we are maintain with the notion that most height functions are flat 3-coloring -- most 3-colorings are flat and we can grow results from height functions to flat 3-colorings. This is the entire purpose of our work. Now, let's be more precise about what we do. So let's look only on quasi-periodic functions with a particular slope M. So we're restricting ourselves to one slope. >>: Is it a vector or -- >> Ohad Feldheim: It's a vector. It's a vector. It has in every direction different. What we're going to do is we're going to map all of those functions with this particular slope to functions with zero slope, and this we're going to do in a one-to-one way. Yes. So given the slope, the mapping is one to one, and this is the sense in which I said it's few to one. One image can have -- if it has different sources, they must have different slopes. And now after doing this, we will find in the image a very long level set. So think of level set as in -so this is an approximation of a continuous function, so think of it as a level set in a continuous function, and if time will permit, we'll also give a definition later on. And the fact that long level sets are rare was exactly the instrument that was used in this Peled 2010 paper to show that things are flat. He showed that long level sets are extremely uncommon. This was the main instrument. And this would tell us that those images are very rare. And so we would know that if the image of those psi M is just a small subset of each. Psi M is just a small sum. So we would be left only with the periodic function, or mainly with the periodic function. So this is the grand scheme, okay? So what we have done so far is that our challenge is taking quasi-periodic functions and mapping them one-to-one to periodic functions. And this is the call of this work. So how can we do this? >>: Kind of it cannot be for every -every N, but could get -- like decay with slope, right? I mean, that, right? I mean, you showed that you know, you get some small subset for presumably if you summed over M, you >> Ohad Feldheim: No, so actually we'll show that all the images are very, very rare. So the entire set of images for all Ms together, its percentage out of colorings will be just a -- so it is something like some constant to minus N to the power D minus one. So this is very, very, very, very rare. So even though we have a polynomial number of slopes, it's insignificant in comparison to how rare those sets are. Okay. So how do we flatten a function? Well, fortunately, in one dimension it's really easy. What you do is you say I take this function, I know that it gains a certain slope. In this case it gains slope six. Let's find the first time that it gains -- that it gained a three height. This is this green line. First time that we gain three. Now, let's reflect the slope of the function from this point and until we finish one period, okay? So if we gain that to here half the slope and we were supposed to gain another half here, so instead we will lose this half and we will come up with something flat, and if we do this periodically, we will get a flat function. So we reflect immediately after height M1 over 2, but what can we do if we have several Ms? So it doesn't generalize very easily. And to see how it generalizes, we must know something about how multi-valid functions look, how do they -continuous functions look on the torus. And topology will give us this answer. So this is how a typical height function looks from very far away, let's say, on the torus. I mean, this is how such functions look. And this is a continuous on the loop, of course. So the black lines here, they are level lines, okay? And you can see that the height when they heat this -- the X-axis, and they have a very interesting structure. So first of all, there are two kinds of them. One kind is those that we call trivial. One way to think about those is that they are the level lines that when we project them to the torus, they split it in two parts. Okay? So, and the other way to think about you start inside one of them and you wrap around, you will definitely end it, a translation of it. Okay? And going to be important in the proof. them is that if make an entire inside a copy of this notion is And there are nontrivial ones that when you start from here and you start traveling around, you might cross them just one. You'll never meet them, a copy, I mean, you will not meet another copy until you get to the other side. So those nontrivial level contours, at least in the continuous setting we would like to think of them as some kind of co-dimension one manifolds. They are very important in understanding how to do this reflection, so they have the following structure. They are some sort of interlaced. So you see this is one copy of such a level line. This is another translation of it. If I translate this level line by one and by N in this direction, I'll see here a copy. If I translate by another one, I'll see another copy. And the same goes for another level line. So if this is the next level line of height six, then there will be here another level line which is just of the same height difference, and they have this ordered structure which we -- I'm not sure that I will get to describe it more accurately, but it is the parallel of what happens one dimension. So you can think of it that in this direction things look very similar to how the one dimension and case looked. So what is going to be my tactics? I'm going to, first of all, mark for myself only increasing level lines. So first I go to height three, then go to height six. I know those three and six. I already reached height six. Then I reach height nine and height 12, et cetera. And only the nontrivial ones. So I don't mark this and I don't mark this. Now I will have some way of finding the first level line which is nontrivial. Let's say it's this one for the moment. And I can now find also its translation. So its translation is this level line. And I can go back and look where did I gain half the height, so I go from this 15 and half the height between 15 and three is nine, so it's this line. And I'm going to reflect everything that goes on between those two lines. So it's going to be this region. And so I will get a flat function. But how will I be able to recover it? So even after the flattening, if I'll take these three, go to the right to the next translation, then go back until I see a height difference of six, which I can calculate because I know the slope, I will know that between this line and this line is the domain that I'm going to have to flip again to get back to my original function. Okay. So this is roughly the idea. This idea is very topological. So if I want to do this for continuous functions, I know to do it only for polynomials or for very nice functions. And I will have to take those notions of what is nice function, what is a -- what are what are those co-dimension one manifolds and somehow formalize it in the discrete terms. This is going to be the main challenge in taking this topological concept and creating a discrete algorithm. So first of all, let's discuss a bit about what are those level lines. So far they were very abstract. Let's make them a bit more concrete. So originally in the work of Peled, level lines were defined in the following way. So he defined, first of all, a sublevel set of a vertex, say this is vertex V and the sublevel set of a certain height, height K, is just taking the connected component of this vertex in the graph without the vertices of this height. So I'm traveling without crossing vertices of height one. This was the notion of a sublevel set, and then the level set would be the contours bounding this set. This was the original definition. So apparently this is not good enough for our purposes, so we had to improve upon it a bit. And the reason is that this notion of a level lines that I described to you of those co-dimension one manifolds, it turns out that the corresponding term in the discrete settings are sets such that they're connected and the complements are connected. If a set is connected and its complement is connected, then we call the boundary separating them a co-dimension one discrete manifold. So in order to turn this set which didn't have this property into a set whose boundary satisfies this, what we do, instead of defining a level set for one vertex, we define a level set for two vertices. So we want to look at the level set between this vertex and this vertex. This zero and this three of height two. And what we do, we add to this. So we look at the connected comparance [phonetic] of these three, and we add to the connected component of zero everything that is not in this connected component. So we pick this blue regions and add them into the set, and so now we will have a set which is connected and also its complement is connected. So this is the definition of what is a sublevel set separating two vertices in our context. And the boundary of this is what we call a level set, and now, that may be the most interesting topological object that we have in our proof, the important is COM is the following: If we have a set of such level sets that they are co-connected, that is, that they are connected and their complement is connected, and that we know that their translates do not intersect, so we can prove this also for our level sets, then they will satisfy one of three of three properties. So it goes like this. We look at such a level set and we look at all of its translations in the world. This is the set, the set -- do I have it here? No, I didn't write its name, but all the U plus T where T is some translation from the set in time ZD. have some set and all of its translations. So we And now this set will -- all of those translations will satisfy one of three states. Either they are all disjoint, or all of the complements are disjoint, or they are all ordered, totally ordered by inclusion. This idea that they are all totally ordered by inclusion is the counterpart of this topological observation that we have those translations of a level set containing one another. So those are nontrivial level sets. So since the description was a bit hard, let's look at some pictures. So here are -- three level sets and their translation. So one is those blue level sets. You can see that they are the sets that separate this minus one and this zero. And indeed, this reflects like the peak of a mountain. It is a trivial level set. And you can see that when you translate it -- sorry -- when you translate it, you get disjoint copies. The counterpart of those are those yellow sets where it's like we're trying to get from a vertex outside the peak of a mountain to the mountain, itself. So if those were valleys -- maybe this is a better description -- those are valleys and those are mountains, and they satisfy that their translations are the complements of their translation of disjoint. So the complement of the set here is just this mountain peak, and the translation of it is this one, is this next mountain peak. And finally, we have those red level sets, and for those red level sets you can see that they are really ordered. So this is one level set, and the next translation is this one, and all of the translations will be ordered. Okay. So I think this is more or less enough about those objects. I can just tell you that what we should do now, if we would posit, is to show that if we have a slope, we must have some nontrivial level set, and then we should explain how to recover -recover our reflections, how to define it, and those things are very difficult technically. I mean, if I would really like to give you a concrete proof, unfortunately, this turned out to be -- so only the technical elements of showing the topological counterparts took more than 20 pictures. Okay. So let me conclude with some open problems. So odd tori, we know very well what's going on. It's not really -- I mean, I wouldn't bet my right arm on it, but it seems very plausible what is the structure that we're going to see, and I think it would be interesting in the context of measures where you have a slope, you are forced to have some particular slope. But the real challenge starts showing this for four colors and more. So as soon as you get to four colors and more, you no longer have this nice bijection with height functions, but you still can define different phases and their interfaces. So it doesn't look completely hopeless. There are some preliminary ideas, but I think it would still be remarkable if someone would know how to solve it and there is -- there are some people who believe that if you'll solve it for four colors, you probably know how to solve it for a number of colors. I haven't made up my mind if I share this view. Another challenge is nonzero temperature. So this says we allow some edges which are not respecting the coloring. Now, we don't have the real correspondence with height functions, but some structure of height function still remains, so we still have some abstract cover space where the height function is defined. It still is related to height functions in some way, and there are very interesting topological phenomenon that happens here, but it still seems a bit far. And of course, low dimension, which I don't know. I wouldn't recommend trying to tackling it using those methods. I mean, it would be interesting to push the dimension down, but if you are trying to reach to three dimensions, you'll have to have a very brilliant idea. I would like just to comment a bit about some small relation between this thing and small -- and low dimensions. So I discard my -- our results here for tori with all the sides of the tori of length N. But they would work also when there are non-even sides of the torus. And in particular, they would work for tori of the form, so N to the two times two to the D. So those are tori that have many dimensions, but many of them are very compact. So here you would still see this phenomena, all -- everything I said here would hold, while if the torus was just N times two to the D, then it is not. It doesn't hold. So at least this is one way to see that there is something special about dimensions two and more. But of course, this is not really the low dimension case that we're interested in. Thank you. [applause] >> Yuval Peres: Questions or comments for Ohad? >>: Yeah. Do you have result for, say, bitter depending on N? >> Ohad Feldheim: For bitter depending on N? See, so what you are asking me is if I can -- so we were interested always in result that are independent of N, right, right. >>: [inaudible]. >> Ohad Feldheim: But you're asking if Galvin -- so I don't know. I would start at looking at the result of Galvin Engbers, because they are less sensitive to those things probably, but I don't know. >>: But do you expect that -- I mean, it seems that there's a fundamental problem to move from real 3-colorings to almost 3-colorings. >> Ohad Feldheim: Yes, but also, I mean, this line of salt also made people think that it's hard to prove it on the torus, and we were able to do it using some topological tricks. So it's possible that using more sophisticated topological tricks you can still overcome the fact that you would have some bad edges. Actually, when you look on configurations with bad edges, you can classify them according also to certain topological structures, and you can -- or even then say that certain topological structures are not so likely to happen. >>: But I mean, like, it really becomes some kind of a analytic property if you have only few bad edges. I mean, topology, you know, presumably, you know, doesn't it, you know, detect so well whether, like the density of the bad edge. I mean, you can -- even if there is one bad edge, you'll have some serious topological flaw [inaudible]. >> Ohad Feldheim: So you might and you might not. Right? So when you have a bad edge, you have now this kind of a four cycle, and if you have a four cycle of this site, of this type, this is very bad. Right? Because it means that you have -- that you are increasing in height when you go around it and you are no longer correspond any height function. You don't have to have cycle like this. So for example, 0011 is okay for us. I mean, it wouldn't cause any trouble. There would be trouble around, okay? But -- and when you look on those bad cycles, then those co-dimension two point -co-dimension two manifolds they wrap around, they create some kind of structure, and it might be possible to say something about the structure, and if the structure you can show that it's really small, so you will never have big loops when you wrap around something bad happens, it might be possible still to obtain a combinatorial result. So I'm telling things that are really half-baked, but at least you can see how I think on approaching topological -- I mean, approaching topologically those faulty 3-colorings. >>: Okay. [inaudible]. Can you just briefly explain the heuristics, why the change of behavior between the dimension two and three, and not, say, three and four? What's the ->> Ohad Feldheim: Briefly explain. So I think not briefly enough for this situation, but I can explain it not briefly, if you'll have time later on. >> Yuval Peres: Okay. So we'll continue in private. Let's thank both our speakers.