>> Yuval Peres: Good afternoon. There is a classic paper that David Wilson wrote 2004 on random adjacent transpositions, and that paper had some open problems that several of us have looked at over the years, and we're delighted to have Hubert Lacoin tell us about the resolution of some of these problems. Please. >> Hubert Lacoin: Thank you very much for the presentation and for the invitation to come and speak here. So what I'm going to talk about the adjacent transposition shuffle and the simple exclusion on one dimensional graph, meaning both the segment and the [indiscernible]. Okay. So here is the [indiscernible] on the segment. So it's a Markov chain on the symmetric group. We start with the identity. And define it in discrete time at each step to get the state of your Markov chain from the state you have in the previous step. What you do is you choose an adjacent transposition uniformly at random, which is the transposition of X, X plus one, where X belongs to the set one up to minus one. And okay. By [indiscernible] the you increase invariant measure for dynamics is the uniform measure on the symmetric [indiscernible] basically you should be close to a uniform [indiscernible]. So okay. [indiscernible] but just for the sake of it, here's a graphic representation that might be clear than the formal definition. So you start from the identity, which I represented, I represent the process graphically on the segment by saying, you know, or by [indiscernible] and okay. What you do at each step is you choose a pair of adjacent location and you inverse the content. This is what you get in step one. And you do it over and over. And this is what you get after six step. And okay. After many times, you should get some [indiscernible] close to somehow if you imagine this is a deck of cards, after many transposition, your deck of cards should be shuffled. So the question is how many steps do you need to shuffle the deck of card? So here there is a small technical problem, that is that after N step, you will have come close with an N transposition so that basically, you're just looking at the signature of the transposition you have obtained. You can tell the [indiscernible] of the number of the steps you've performed and you never you will still keep the memory of the initial condition. So there is two way to bypass this small technicality. So you can easily work with a lazy chin. At each step you will did nothing with probability one half. Or you can work in continuous time, meaning that you put a Poisson clock on each pair of adjacent sides and when the clock ring, what you do is you compose with your [indiscernible] transposition. So if you put Poisson clock with rate one at inch, what you will do is speed up the Markov chain by your factor N minus one you compare to the [indiscernible] one so keep that in mind when comparing results in the literature. Okay. So I will be interested in the [indiscernible] equilibrium in terms of two distances which are [indiscernible] distance and the separation distance so I guess all of you know what they are. So just to [indiscernible] distance, it measures how well you can couple two measures, define them to the same probability space. The separation distance also looks at something that is very often [indiscernible] context of mixing of Markov chains and okay. We denote by DT and DST, distances to [indiscernible] and separation distance respectively. And here, when I say that I started from the identity, but, in fact, I submit it doesn't really matter what condition I start for. This [indiscernible] dependent on the initial condition. Okay. So what I call the mixing time is the first time at which your distance to equilibrium drops below a special epsilon. Epsilon is just a fixed parameter. And what we want to know is the asymptotic behavior of this mixing time when the size of your system N tends to infinity. Okay. So I will recall some results presented in literature concerning this process. So okay. Very often, Markov chain and the symmetric group used as a benchmark to test new techniques in this study of mixing time so that would be a central object in the area. So a first result that have some important historical importance is by Diaconis and Shahshahani. It's a study of the process of discrete time for the complete graph with an [indiscernible]. And what they showed that the mixing time was of order of one half of N again. So okay. Says there is no dependence in epsilon on the right hand side. But just means that the leading [indiscernible] does not depend on epsilon. Okay. I mean around this time the distance quill rub drops abruptly from one to zero. So since this papers, there has been a lot of research concerning also random or semi random transposition shuffle concerning the adjacent transposition shuffle, the first reference I found was by David Aldous in '93. And what he proved is that the mixing time was at least N to N square and at most N square log N. >>: [indiscernible]. >> Hubert Lacoin: Yes, now I switch to continuous time. And many people were convinced that N square again was the good bound but some technicalities prevent [indiscernible] in 2004, David Wilson showed that, okay, asymptotically, right mixing time was somewhere between 1 over pi square and square again and one other 2 pi square and square again. And he conjectured that the lower bound was sharp with some convincing heuristics. Okay. So now I just a tiny bit of motivation. So let's imagine that we want to study very complicated system, like a [indiscernible] particle where [indiscernible] and the velocity and the [indiscernible] and we can bounce in random directions. And you want to know, well, you start answer something like that, however time do you need to forget about the initial condition. And somehow, you want to simplify this picture, you have to the feature that you really want to keep is the interactions are local, that particularly the only for short times only care about what the neighbors are doing and okay. The simplest way to have a paralysis tem with particles interacting with their neighbors is really the transposition shuffle in the segment. If you want to distinguish between the particle or the [indiscernible] if you want to kind of ignore to have unlabeled particles, to forget about the [indiscernible]. Okay. So now I can present my result concerning the adjacent transpositional shuffle. Specifically what I showed is that the conjecture of David Wilson was right. So that the mixing time was, indeed, asymptotically equivalent of one other 2 pi square and square again and that the separation time was it was just twice as large. So if you look at the distance to equilibrium on the time scale as I said before, it drops abruptly from one to zero around 2 pi square to the minus one. And a natural question on what kind of window you should zoom on now that we see some [indiscernible] transition. So this phenomenon dropped from one to zero, it's called cut off. And the term has been coined by Aldous and [indiscernible], I think. And this window is just called the cut off window and it's still an [indiscernible] problem for the adjacent transposition and the right conjecture is that it should be [indiscernible]. >>: [indiscernible]. >> Hubert Lacoin: I don't get good quantitative bounds. I guess maybe some improvements the best you can hope is really log log N. N square times log log N. I don't know if you can really get there. Okay. So now yeah? >>: [indiscernible]. >> Hubert Lacoin: So this should be [indiscernible]. >>: [indiscernible]. >> Hubert Lacoin: >> So lower bound, the lower bound N square is [indiscernible]. >> Hubert Lacoin: Okay. So now let us look at kind of a simplified process, which is, okay, instead of looking at all the position of all your counts, you decide to paint K of them in black and N minus K in white to consider K count as particles and the other [indiscernible] as empty site. And instead of looking at the bottom line, which is the motion of the particles, and okay. It's not true this the prediction of a Markov chain is always a Markov chain, but here it's the case and it's almost an obvious statement. And if you want to describe the dynamics below, what happens that the particles just make well, [indiscernible] they jump and they are [indiscernible] probability one and if you particle twice to jump on the neighbor which on a site which is already occupied, the jump would be cancelled. So here's how you can define formally the process. So the articles are one and the empty sites zero. The state space is a number of 01 configurations with exactly K1s and an alternative way to define the process is to say that, okay, when you put some Poisson clocks on pair of adjacent sites and when the clock ring you good change the content of two sides. Basically, I mean, if you have two particles and two [indiscernible] you want these interchanging because you don't distinguish between the particles. And for this process as well as the equilibrium measure is a uniform measure. This set of continuation. So now that you define the distance two equilibrium, what we do is [indiscernible] for the separation distance. For this process, where you have no symmetry anymore so the configuration to start with does matter so when you define the distance equilibrium what you do is you take the maximum of all initial condition. [indiscernible] distance and the separation distance. This And then you define the mixing time and separation mixing time in the same manner [indiscernible]. So the result is I obtain is that, okay when you have K particle or when the number of particle and [indiscernible] are similar to [indiscernible] and okay, I [indiscernible] take the number of particles smaller than N over 2, but there is no restriction to that there is individuality between the particle and empty sites. It shows that the mixing time is asymptotically equivalent to 1 over two pi square and square log K. And the separation time is twice as large. So here you need a lower bound in the growth of K if you want to get the lower bound for the separation mixing time. That's a technicality. So the second line is not exactly proven for small K. Okay. And finally, a last resort is when you consider as a process on the [indiscernible] instead of the segment. So in that case, okay, you again, [indiscernible] in total variation. In fact, I use a different method and you can push things further and have the exact cut off window. So these three lines there just use at, okay, you zoom on a window with N square, you will see something. And if you zoom on the something smaller, you won't see anything. Okay. And here, I wrote some limit, but you should think about that as the statement calls for a [indiscernible] we have no [indiscernible]. I guess it exists, but I can't prove it. Okay. So as I told you, the method I use for the segment and for the [indiscernible] are quite different. So, in fact, when I mean that the method are different, I'm talking about the upper bound because the proof of the lower bound is kind of is the same and it's basically [indiscernible] Wilson's paper. And so the advantage of the method, what I will develop, in fact [indiscernible] for the segment, it relies on the [indiscernible] so in particular I use censorship and equality by Peres and Winkler also the [indiscernible] inequality. So the interpretive side, you can use it to divide result for adjacent transposition shuffle and for the post distance separation. On the down side, I mean, it does not export to more general graphs because, in fact, you need to really be on this segment to have a natural order on the space of configuration. Okay. For the circle, I used a comparison results developed by Liggett, which compare the exclusion on the circle with a set of independent random [indiscernible]. And so on the positive side, you can get a sharp result for the cut off window. And in fact, half of the proof would work so if you have your high dimension, say, two dimensional grid, on the down side, you don't get the result for the adjacent transposition shuffle and I don't see how you can get result for the separation distance with this method. Okay. So during the rest of the talk, I will mainly switch to the [indiscernible]. So first, what I will do is to provide the small review of the proof of the lower bound on the mixing time, which is the same for circle and segment and then I will focus on the case of the simple exclusion on the segment with N over 2 particle and okay, give first proof of an upper bound that matches within a constant factor which was already known and explain how you can improve this bound and get in your sharp result by using [indiscernible] arguments. >>: Isn't the lower bound just [indiscernible]. >> Hubert Lacoin: >>: [indiscernible]. >> Hubert Lacoin: >>: Yeah. Yeah. I will do it quickly. Okay. >> Hubert Lacoin: And also, it gives some idea of the heuristics of why it should be true. Okay. So finally I will come back to the circle and give a brief overview of the proof the circle. So okay. Let me call [indiscernible] my assistant. Let's say that I start from an initial condition K. And okay. I call UXT the probability of [indiscernible] at site X at time T. When starting from K. And a simple observation you can make is that as you look at this expectation, in fact, it is the solution of the discrete declaration. So a simple way to see that is to see that, in fact, if you look at the density of particle, the expected density of particle, it's the same for that independent random [indiscernible] and this [indiscernible] is just the generator of a simple [indiscernible]. So a simple corollary of this fact is that if you look at the first [indiscernible] coefficient of [indiscernible] so let's define A1 of [indiscernible] to be the sum from X, from Z, from X equal 1. For zero to N of of X [indiscernible] X by over N what you have is that the expected value of [indiscernible] exponentially fast. And with rate which is just two times one minus X by over N. This is equivalent to pi square over N square. And okay. These guys just okay, the eigenvalue to this function for the generator of the random work. Okay. So one start from a configuration this has all of its particle on the right, okay. So [indiscernible] because you're working with article system and the height function. In fact, what you really have to consider is even okay. I'm using to write it for [indiscernible] X minus one half over N pi. So if you start from [indiscernible] from the left half of the [indiscernible] particle, what you have is that initially, you want a [indiscernible] N. And now if you sample [indiscernible] from equilibrium, basically your E to X will be almost [indiscernible] variables. So this is just some of independent variable of finite variance. So typically, the variance of [indiscernible] of N square. Okay. >> So [indiscernible]. >> Hubert Lacoin: [indiscernible], yeah. So in order for your system to come to equilibrium, what you could think is that, okay, you need the expectation of [indiscernible] to drop at least from this value N to the equilibrium situation so that now the expectation get kind of lost in the [indiscernible]. If you want to turn that into a rigorous proof what you have to control is the variance of A1 of [indiscernible] for all time. So what you can use is that this guy is [indiscernible] and then computing the banding the quadratic variation of this [indiscernible], what you get is [indiscernible] the variance of this guy is bounded by N. So the variance is never larger than the equilibrium variance. And as a consequence of that, if T is smaller than one half of lambda one to the minus one, N, okay, so one half minus delta, what you have is that with hyper ability, A1 [indiscernible] is larger than some constant times N to the one half plus delta. And then from this, you can conclude that the [indiscernible] is close to one because this is not the equilibrium value. So then if you replace the lambda one by asymptotic [indiscernible] what you get is that bound for the mixing time is one over two pi square and N square again. So a reason to believe that this gives you the [indiscernible] mixing time is that, okay, not only is this for the coefficient K exponentially fast but all the others, and the others came with a larger rate, okay. So that basically, I mean, this function should be the one that [indiscernible] to equilibrium [indiscernible] manner. But you cannot turn that into a this argument into a real proof because, okay, this thing provides you N eigen functions or N minus one eigen functions and your Markov chain means the space of dimension [indiscernible] so you can't do the full the full spectral analysis. So in order to find an upper bound, you need to be smarter and to use [indiscernible]. So right now, I need a picture from the screen. Okay. So the first thing we do is basically, we kind of change the Markov chain. We study to an isomorphic one. So we choose to study the height function to the particular system, which is defined by each time you see a particle, you want to you make your I function climb by one and each time you [indiscernible] you go down by one. And because your N over 2 particle, you end up at zero. And if you look at the transition for the dynamics of the high function are, it corresponds to [indiscernible] of corners in your high function. So like when you look at maximum, you can look at minimum and look at minimum and look at maximum. And okay, we equip this state of function with the natural order, which is that we see that some high function is larger than another if it's [indiscernible] above the other. So you call [indiscernible] corresponding [indiscernible]. So I will continue on the board now. And so the first remark to make is that this order is somehow conserved by the dynamics. We haven't shown [indiscernible] meaning that we can find a grand coupling. So a coupling of [indiscernible] we can couple the Markov chain starting from [indiscernible] in a manner that if two condition are ordered, then they remain ordered all through the time. Okay. And so if you try to construct a coupling in the most naive way, what you would do is to say okay, I would put a Poisson clock on each coordinate, and when it rings if I see a corner, I will switch it. But this would not provide you a [indiscernible] coupling. What you're going to do instead is say for each site, I will not take one clock process, but two clock processes. For each X, the two clock processes which I 2x up and 2x down and if 2x up rings, then okay. If 2x up rings and X is a local minimum, I turn it to a local maximum and if 2x down rings, and X is the local maximum, I turn it into a local minimum. And I will [indiscernible] use two different process for up flips and down flips and you should do things that way and this is kind of straightforward proof. So, in fact, using this idea, you can get an upper bound for the mixing time that matches your lower bound up to a constant factor. So [indiscernible] so I can [indiscernible]. Okay. So what you want is a bound to turn variation distance equilibrium and you know that this is smaller than the maximum over pairs of initial condition of the distance between [indiscernible] these are the distribution of [indiscernible]. This is this a triangular inequality. And then you want to bound this guy. So this is the [indiscernible] distance is obtained as picking the best coupling between these two probability and see by how much by what probability is different so in particular, you should take a given coupling, it will give you an upper bound and [indiscernible] distance so if we take our [indiscernible], this is smaller than the probability that under other [indiscernible]. And now because we have a dynamic that conserves nor [indiscernible], we know that the two paths that will be the latest one to couple are the highest one and the lowest one. So now you write them as a [indiscernible] and this is because, well, I mean, basically, once these two have coalesced, all of the other conditions are sandwiched between them so they will. Yeah. And okay. So this is. >>: Wouldn't you upper bound [indiscernible]. >> Hubert Lacoin: Yeah. Okay. So now what you get is that this is the probability that A1 of [indiscernible]. The difference between the two coefficients is positive so like different definition, this is [indiscernible]. And okay. When this guy is positive, it's larger than 1 over N so you can say that this is smaller than N times the expectation of two A1. Okay. Just symmetry here. You can say that this is smaller than N to the three times exponential minus lambda 1T just by using the exponential detail of this guy. And then [indiscernible] taking T equal lambda 1 to the minus 1 times four times log N. We get that this is smaller than N to the minus one. [indiscernible]. So here we've lost a constant 8 in the game and the reason that basically for the lower bound, we wanted A1 to be smaller than N to the three half and, okay, the other notation was N to the one half where high function becomes N to the three half and here we require it to be smaller than 1 over N so this explain. And okay. So basically, what we need is some other idea in order to make this to improve this upper bound. Now we use a picture again. So the reason why it's difficult to control the mixing time of this object is because it's very high dimensional. So within thing we can try to do is to say that we are going to [indiscernible] lower dimensional process, which I call the skeleton by just instead of carrying about the whole height function, we only care about the K minus one points that are equally spaced on the segment. Okay. So this is what I will care about. And what we want to say is that basically, this guy's the equilibrium, then shortly afterwards, the whole system is at the equilibrium. And the two that we want to use to say that is the [indiscernible] equality. Basically what we will do is after some time, we will freeze the dynamics at this point so that after the equilibrium equilibrium, what we need is only to mix smaller systems. So I only need this picture. So I will continue on the board. And okay. So the idea is to define, we want to show that if at time one half, plus delta and [indiscernible] the skeleton the equilibrium. Okay, then things become okay. And the second time, I will show that you can prove what I will give you ideas of how you prove this. So what we define is some [indiscernible] dynamics, prime T so here I assume that I stop from the very top initial condition. And it's not true that it has to be the worst >> [indiscernible]. >> Hubert Lacoin: >>: What is the coefficient [indiscernible]. >> Hubert Lacoin: >>: Sorry? One half plus delta, okay. [indiscernible]. >> Hubert Lacoin: 1 over pi square, yeah. >>: [indiscernible]. >> Hubert Lacoin: I think of okay. I first select delta and then I take K called one over delta. Thank you. So I defined [indiscernible] prime of T, which is a modified dynamic, which is just [indiscernible] NT zero and for which the skeleton is freezed for T larger than T zero. So [indiscernible] that means I just okay. If I call XI the coordinate of my skeleton, which I am just I times N over K, what I will do is I will sign in the clocks to the XI for I equals one to T minus one. And the reason the result, which was yeah. Which is published in [indiscernible] in 2001 and maybe appeared in the paper in 2011, by [indiscernible], which is called the censorship in equality and a way to read it in our particular setup is to say that if you look at the distance to equilibrium for your original dynamics, then it's smaller than the distance to equilibrium for your modified dynamics. And this is [indiscernible] you start from the maximal initial condition. So here I'm kind of [indiscernible] something how it prove as general result, but if I have time to explain how things work on the [indiscernible] you'll see that the main interest for this method is to use the adjacent transposition shuffle where you can start from any condition you want. So now, if you look for times after times zero, your dynamics just becomes [indiscernible] to the skeleton. It's [indiscernible] K independent dynamics on systems which are of size L, which is N over K. So each of these systems, they are not N over [indiscernible]. It's not L over two particles, but what I said before still apply. So the know the systems are mixing the time [indiscernible] log N. In fact, with this time, you should take the [indiscernible] L square over N. By large [indiscernible], I mean something that does not depend on K. There will be really close to equilibrium so that the product of the consistent will also be close to equilibrium. And this guy is just N square log N by T square. So what this tells you is that for T equal T zero plus delta N square log N, I'm much after this mixing time and therefore, condition teal my skeleton, my system is at equilibrium so if I know that I am [indiscernible] my skeleton was at equilibrium, the full system will be at equilibrium. Okay. So now what remains to prove is that the skeleton is indeed at equilibrium. So here I told you that basically you want to get a shot down, you have to take K large. I will do the proof of K equal one and I hope this will be sufficiently convincing for to show you the proof is also right for K as large as you want. So okay. So another consequence of my [indiscernible] is that if you look at any time the identity of your distribution with respect to equilibrium, it's an increasing function in the sense that if this function at a given configuration, you can be an incomplete function and you should take in that function which is above another, this function will be larger. And again, this relies on the fact that you start from the maximum configuration and this thing is also stable by projection. So what I will do now is to say that, I will look in the distribution of the middle point. So I will look at [indiscernible] at coordinate N over two and I call this guy bar. I call new bar is the low of eta bar under zero T zero and I call the new bar equilibrium low of eta bar. And what I have is that D over new bar over D of new bar is an increasing function. So it's defined on the subset of Z. And okay. So what I know is that if I look at the expectation of eta bar and the measure new bar, it's equal to U N over 2 T zero okay. Which is just a solution of these equations starting from the and this is because of the time I've chosen, you can [indiscernible] to resolve the [indiscernible] equation. It's small O of square root of N. So what you have that is, okay, the expectation is smaller than the equilibrium [indiscernible] so this kind of indicates that the two measures should be close, but you can't turn that into proof. It's easy to found counter examples. But you have additional information that this [indiscernible] function and you can use that to complete. So I will do that now. What view in general is that the [indiscernible] variation distance between new and new bar, it's smaller than some constant times the expectation of [indiscernible]. Okay. If you prove that, you're done. So okay. I will start from this term and then I will show that it's larger than this one. So if I look at this guy, okay, I can always subtract the expectation and the equilibrium, which is equal to zero. I can write it as the well, I write it as an integral. And it's the integral of the product of two positive function. One is identity minus one. Okay. So when you have a two increasing functions, what you can use is you will have this in equality. If you use this right away, you won't get anything. You will just get that this thing is larger than zero. >>: [indiscernible]. >> Hubert Lacoin: Oh, yeah. So what we will do first is split this thing under two events. So as the density is increasing, what you know is that the total variation distance, it's equal to new of A minus new of A where A is the [indiscernible] so it's you can write it in this manner, okay? This is true for some value of small A. So I'll choose to spread this in two parts. And then I'll use this correlation equality on for both sides so I have to be careful because, okay, these things aren't probabilities anymore. But what you obtain [indiscernible] is that you have eta bar minus its larger than okay. The expectation of eta bar conditioned to A times U of A minus U of A. Plus the expectation of eta bar conditioned to the compliment of A which applied by a new of the compliment of A minus U of the compliment of A. And then this is equal to okay. Your total variation distance times the [indiscernible] of eta bar conditioned to A minus, the expectation of eta bar. Okay. So what you need to do to conclude is prove that this guy is larger than some constant times square. So you can discuss the value of small A. So if small A is larger than zero, this guy is larger than the [indiscernible] of eta bar conditioned to eta bar larger than zero. Okay. And this is the square root of N. If it's smaller than zero, and you don't know anything about this guy, but you would know the same thing for this guy. Okay. And this [indiscernible] to conclude. So if you want to do it for the whole skeleton, basically, so you will take, instead of taking just the expectation, you will take the sum of expectations for all of the coordinates and basically instead of using this simple correlation in equality, you use the TKG in equality for your K coordinates and it should read the right way, basically. That works. So this is what I had to say about the simple [indiscernible] on the segment maybe I can tell you a bit more about how you can make this proof work for our adjacent transposition. Okay. So for adjacent transposition, the first thing is when you have a permutation, you want to map it into so you map to the particular configuration and to height function. You want to map the permutation into discrete surfaces, okay which I'll just functions from one from zero N square to R. Okay. And which are defined as [indiscernible]. The height at coordinate XY is the [indiscernible] from that equal one to X. And okay. XY over N and this shift us just to make this center at the equilibrium. And okay. If you have this height function, I mean, you can easily define another on the set of permutation. And again, this [indiscernible] is by the dynamics so that you have [indiscernible] and you can use the censorship and the equality. Then the censors you will use is different in that case. You will do censoring two times. So the first thing you will do is you will start so again you have a parameter of K, which is one over delta, and what you start doing is you run the dynamics during your time delta N square log N, okay. And what you do is that during this time, yeah, you censor the K minus one transposition that corresponds to the XI and basically this has the effect of shuffling the your [indiscernible] each segment and it makes what the count number, say, from one to N over K indistinguishable. So basically, to describe fully about the permutation, you need N height function with this first step, you reduce to K height function, okay? Then you N censor the dynamics and you run it for time T zero. And after time T zero, basically what you know is that if you look just at this coordinate, XI is just IN over K. So if you just look at the skeleton, you can put to it equilibrium just using what I told you there. Okay? So if you looked at your T line functions, you can couple them with an equilibrium like that. Okay. So like you can couple with [indiscernible] so that all the [indiscernible] coincide. And once you've done that, basically, you're done because you can say for the remaining time, for a time delta N square again, what I will do is I will censor the [indiscernible] corresponding to the X size again and this will allow me to kind of squeeze the kind of [indiscernible] on the board. Okay. So okay. Basically [indiscernible] anything, but this is the idea of the proof. So if I have some time left now, I can talk about the technique which is used for the second. >>: I'd like to hear it. >> Hubert Lacoin: So for the [indiscernible], basically, the first eigenvalue of the operator is four times as large. The time you need for instant for the solution and for your [indiscernible] to become comparable for equilibrium situation is now one over 8 pi square and square again. And the first thing you would like to show is what you know, by this time, I assume that they have N over two particles. By this time, what I have is that starting from any [indiscernible] configuration, the expected value of the present of the particle somewhere is it's one half [indiscernible] big O of N to the minus one half. What I would like to show first is that, okay, if you look at a segment and you count the number of particles you have into it, you should just take the you should compare that to the expectation. You will be up at most by square root of N. So let's say that I call STX, which is a sum from Z equal one to X of eta. I will just forget about the initial condition. Minus its expectation. For all [indiscernible] and zero, the probability that there is an X such that STX is larger than square root of N times S, which is smaller than two times exponential minus CS square. I can find the consistency [indiscernible]. So basically, my deviations are sub Gaussians. >>: [indiscernible]. >> Hubert Lacoin: So basically, what you want is to replace this line by what? Okay. Corollary is that for T equal the value that I want, I can replace ST I can replace this guy by one half and this won't affect this. Sol the proof of the statement, let's say that, okay, here I'm going to have to work to prove it [indiscernible] for all X and this is important. But let's say that we just want to prove the statement for the middle point. What we want to have is [indiscernible] transform of yeah, of STN over 2. And I would like to show that, okay, for all [indiscernible] in one, this guy is smaller than exponential N and [indiscernible] square times some [indiscernible]. You have a bound of this type, you can prove this for N over 2 and then basically, by dichotomy, you can prove also for N over 4 and 3N over 4 and you will get better bound because your intervals are smaller and then if you [indiscernible] you will be able to sum everything. And okay. To prove this, in fact, you just have to look in the literature. So this you can see that it's a function of the particle configuration, okay. And the particle configuration, you can [indiscernible]. The position of the N particles, okay, you don't know which particle is at the position at the N over 2 particles. You don't know which particle at this place, but as a function of the symmetric, this is not a problem. And there is, in fact, two papers by Liggett in the '70s which tells you that if you have a function which is positive definite, so in the sense that if you take any two dimensional margin of this function, it's positive definite and you look at its expectation for a simple exclusion process, it's smaller than expectation for the simple or for [indiscernible]. Okay. And if you want to evaluate this guy, okay, at a given time, it's just a sum of independent [indiscernible] with different parameter and, therefore, I mean, you have subtracted the means so the [indiscernible] transform is just [indiscernible] and you're done. So now we'll come back to the screen for last time. Okay. And so now you want to say that once you've these situations of the square root of N around the mean density of particles, this is enough to [indiscernible] within the time N square. So what you start doing is you draw the height function of your system as before. So let's say that you start from zero [indiscernible] decide to draw your height function. Let's say that now you start from deterministic configuration that the situation square root of N. So it can be as bad as you want. Can just go out for square root of N step and then go down. And you want to couple it with equilibrium, but I mean, now your height function, they're not attached anymore. They can wander as high and as low as they want. So what you will do is, okay, let's say that you take a particle configuration at equilibrium, define you can from it, you can define the height function with zero translation and what you do is you will take two [indiscernible] of this height function, you will place it one below the configuration and one above and then, okay, you can run dynamic on this height function that conserves the other and what you will do is just to make the blue guy and the red guy coalesce. Okay. And we'll take this up. You remember the picture. So the screen up, please. So I call side one of T the blue guy, which is somebody starting from equilibrium and that [indiscernible]. And side two of T is the red guy, which was on the bottom. And what I'm interested is the volume between the two, which is the sum from all the [indiscernible] of this guy. And, okay, it's random because the amount of which I had to shift the height function to make things order is random. But it's an order N to the three halves. Okay. And what I want is to find a coupling on the dynamics as the high function is that this guy fluctuates as much as it can so that it squeezes rapidly to zero. Okay. But you should look at this guy it's [indiscernible] so it has no drift. You can only count on the fluctuations to go down to zero. So the way you will define your coupling is to say that, okay, as soon as corners are distant, this is a [indiscernible] independently so you maximize a situation and when they are together they [indiscernible] together so that you can [indiscernible] the other. So any fixed time you're at the equilibrium at the beginning they have no contact with one another. Okay. Maybe I have to say something before that. So you should look at [indiscernible] with this coupling. It only makes plus one and minus one jumps and it's a [indiscernible] so if you make a time change, you can well, it's just a simple random work. It starts from N to the three halves to the time it needs to reach zero is of order N 3. So what you would need to show naively is that when your time changes the weight of N so that you really need time N square. So initially, this is true. Like your two guys have no contact and typically in the past, you have N variable corners that can flip independently. But as soon as the area starts to be small, this is not true anymore. So what you have to do to their [indiscernible] is to [indiscernible] in time to show that, well, okay, the time by which you have two few corners is well, you have only you have many corners as soon as the array is large enough is basically the idea, and then you do this, yeah, multi scale analysis in order to say that basically, I mean, when you have abandoned time change which depends on AT and which is good enough to bring you to zero [indiscernible] times N square. Okay. So I guess I will stop here and thank you for your attention. >> Yuval Peres: Any comments or questions? Thank you.