>> Yuval Peres: All right. Good afternoon everyone. So just yesterday walking, or this morning walking through the building I saw many people in unexpected places at MSR have copies of James’ book on Markov chains. So not just in the theory group. It's extremely influential. But today he'll talk about something else, a variant of DLA Hastings-Levitov aggregation in the small particle limit. James Norris, please. >> James Norris: So thanks Yuval for the chance to talk. And more generally, thanks for the opportunity to spend some time in this beautiful place and very stimulating environment. Before I go any further, so I've not written it on the slide, this is joint work with Amanda Turner, who is a PhD student at Cambridge and is now a lecturer in Lancaster University in the UK. So DLA is a famously hard problem and I'm not really going to talk about DLA, but what I going to talk about is, I suppose, motivated by it. So let's explain in what respect the model I'm going to talk about shares features with DLA and other ways in which it's different. But let's think generally about, well, let’s kind of describe what DLA is and what we're trying to do with DLA. It's a model for the growth of a cluster. So in mathematical terms you could think about clusters being represented by a compact set in some Euclidean space. And it's growing, so we've got a parameterized family of compact sets which are getting bigger. Okay? So K sub T is how big the cluster is at the time T. The mechanism by which the sets grow, this is sort of physics rather than mathematics, we are imagining lots of sticky particles wandering around and every so often a particle bumps into the cluster and sticks there. So the particles the same moving, I mean the simplest, one mechanism by which you might think of these particles moving is that you set them off a long ways from the cluster and they just wander around randomly like Brownian motions and then the first point that they hit, the cluster, they stick there and the cluster grows a little bit. Okay? So the distribution of the point which the cluster will grow at is given by the hitting distribution of Brownian motion from infinity, otherwise known as the harmonic measure from infinity. Okay? So the idea is that we should make the clusters grow so that the growth rate at any point on the boundary is proportional to the amount of the harmonic measure which sits there. And immediately you can see that if you have some boundary which has fingers pointing out then those fingers are more likely to be hit by this Brownian motion than some point deep in a [inaudible] of the boundary of the cluster. So you expect to see that the tips of the fingers will grow preference preferentially and then that's already giving an idea that it, I mean suppose you compare the growth from a nice smooth bowl or disc and from something, on the other hand consider a nice smooth bowl or disc with a little pin prick poking out. Very small. That pinprick will get preferentially grown so that these two domains, which are very close to each other, the evolution might therefore not be close to each other because all that's required is this little prick out of the boundary in order for it to have a completely different evolution. This is not mathematics but this is just some sort of heuristics. But if you try to make sense of a PDE story for the growth by harmonic measure it doesn't work. You try writing down a PDE, it's not well posed. You can't do anything. And indeed, when you look at physical instances where the growth you might reasonably think of as being caused by this sort of mechanism, they look random. So you wouldn't expect it to come out of a PDE. They exhibit sort of fractal and apparently random features. So maybe PDEs aren't the way to go here. I mean, which is interesting for probability because often it's a sort of particle picture of what's happening. You know, you can move over to a PDE picture and the essential aspects of the dynamics are captured by the PDE. But here it seems it’s not going to be that kind of set up. So instead, we might start with a particle model. After all, that's kind of by the sort of physical picture began, and see what happens as the size of the particles get a small. Is there any kind of limit object? Now I'm going to work exclusively in two dimensions which causes already a big disappointment because we would like to be able to understand this in three dimensions. But two dimensions is great because you have all the mechanisms of complex analysis at your disposal. You know, you can play lots of games which you can't do three dimensions. This is the era of two dimensional probability. So we're going to be working two dimensions. So I'm thinking about going from thinking about a continuous evolution for now towards thinking about [inaudible] moves in discrete steps. My time parameters going to be discrete and I'm thinking about a sequence of compact sets which grow, K, 0 is also always going to be the unit disc and I'm going to progressively make it bigger by adding bits to it. I'll assume that the exterior domain for my sets is always simply connected and then a first exploitation of that we are working in the plane is that I can encode the cluster through its complementary domain in terms of conformal map because there's a unique conformal map which takes the exterior domain of the unit disc, so here's my disc K, 0, unit disc and so the D, 0 is complementary to [inaudible] and you get mapped by some conformal map to the complementary domain D, N of the cluster at time N. This three parameter family, which does this but there's only one, which has the property that it fixes infinity and doesn't rotate the plane at infinity at all. So you come up with a unique conformal map associated with these two, with this cluster. You know about the conformal invariants of Brownian motion? If I take a Brownian motion in two dimensions and I look at this image under a conformal map, they're not to a change of timescale. I'm still looking at Brownian motion. So in particular, if I go hitting probabilities, if I start a Brownian motion after infinity here or if you like on a sort of large circle around here uniformly, well, actually I know then that it's going to, by sort of symmetries of Brownian motion I know that hitting distribution here is the uniform distribution. If I map this over here then I'm, again, looking at a time change of Brownian motion. So it's the hitting distribution of the image process on here will just be the harmonic measure from infinity and so this shows us that if I take the uniform distribution here and I knock it over I'm assuming that boundary’s nice here, the boundary of this domain is nice, if I put the measure over here using phi N, then I will get exactly the harmonic measure on the boundary of this cluster. Okay? All right. So this slide describes the, well, actually a big family of models that we were discussing just before for the talk that in some sense this is kind of prior work on this family. Not all of it published by Carson and Makarov, but the name which becomes associated with the family of models, the names Hastings and Levitov; so these are models for aggregation, for kind of growth of clusters in the plane which are encoded using conformal maps. There's one parameter family of these models, parameter Alpha, and I'm actually going to talk exclusively about Alpha recourse zero, which is the easiest one to study in this family, there's, its work in progress to deal with the other Alphas. How far we’ll ever get I'll never know. This recent preprint of Amanda and Fredrik Johansson and Alan Sola where they’ve begun to make a bit of progress with positive IDs of Alpha and just post it on the archive. But for today, apart from this slide, it will be entirely about Alpha equals zero. But let me describe the whole family so you have, so you can see what it is. Okay. So I'm going to consider adding a certain sort of particle to the unit disc and two simple examples you might want to keep in mind are, on the one hand I could think of adding a slit like that of length Delta, or I might think about adding a little ball of diameter Delta. So that's a family of particles because I could think about varying Delta. Okay. Now if I exaggerate this picture a little bit, suppose that K, M is like that. Now I really want to attach a particle here of some certain size, say Delta zero, where the attachment point is chosen according to harmonic measure. So I might want to attach a little ball here. What I'm actually going to do is to attach a particle at some random point on the circle and then I'm going to map it over using phi N. Now the amount of harmonic measure on this little bit of boundary here is quite small. It's hard for Brownian motion to get in here. So when I map it over here to the circle, it maps to a small interval. So in order to have a particle of size Delta here, it has to be a much smaller than Delta over here. Because this is a scale factor of if it’s attached at angle theta, Z to the I theta. That's the scale factor in moving this direction. So I want to add a small particle there and if I was adding a particle of size Delta here, well, this little bit of boundary here has got lots of harmonic measures. Easy to hit. So over here I’d want to attach something which was relatively, let's get this right, relatively large over there. So that actually motivates the formula given here. My Delta ends over on this side. I take a bowl or a slit of size Delta N plus one. This is the pointer, right? Anyone know which button is supposed to activate it? >>: Are you sure this is the pen? Don't you have a silver pen [inaudible]? >> James Norris: Okay. So I take my basic particle size and then I just, if I take Alpha equals two it does just what I was describing, Alpha equals two scales-down Delta so that when the particles map back to K, N it comes out at the right size. So we should aim to get all the particles of the same size over here when Alpha equals two. When Alpha equals zero you don't bother to do that so the particles are the same size over here we’re going to end up adding different size particles over there. And you might want to explore how things varied as we let Alpha go between zero and two. Obviously, Alpha equals two because your best hope for something which looks like a DLA, it's believed that actually the whole family of models is of interest and possibly Alpha equals zero is some sort of connection with Eden model which is another famous graph modeling probability. >>: [inaudible]? >> James Norris: Alpha equals one. Sorry. Yes. Alpha equals zero. You can see it's going to be much simpler because I'm not doing anything here and that's the one we could do something with. All right. And little bit of light relief. Why you might be interested in this sort of thing. Here are some pictures of instances, either physical or computational, where your cluster’s grown. And they show interesting features that don't look like bowls. They look random and they look kind of fractal. Mathematics has not really had much success yet in explaining what's going on here. But I think we ought to, right? It's, there’s some math to be done here. So these are caused by some sort of electrical effect like lightning, Lichtenberg features, famous source of paperweights. There’s a different sort of growth model coming from biology. You can see sort of fingers coming out and maybe this is growing more like a, roughly like a bowl. It's a simulation of DLA, the colors in these pictures encoding times at which particles are added. And this is a simulation which Amanda did of our model HL zero. It's not too different to this simulation of DLA. Again, the colors are decoding different epochs, but which one is adding particles? This slide just reviews the model which is going to be the subject for the rest of the talk. So if you’ve got any, if you want me to explain then I'm happy to take questions. We can stop with unit disc and fix a parameter Delta. Delta in the end is going to go to zero. We’re going to be thinking about scaling limits as the size of the particle goes to zero. The particle added in each stage will either be [inaudible] slits or discs. The picture here was, this picture was the slit model. Then there will be some unique conformal map which kind of encodes adding that particle to the unit disc, subjects to these normalizations, and the way the model is constructed is the following: the probability comes from a sequence of independent uniform random variables on unit, [inaudible] on the circle. Feature’s going to be the angle at which the particle is attached on the circle. Probably the best formula to look at to understand what's going on is this one. P is my basic particle, could be a little disc attached at point 1 on the unit circle, then you rotate the disc by random angle, and then you map it over to the cluster using the conformal map phi N. So you end up with a new bit added to the cluster. As the cluster evolves you add more and more particles. The challenge is to understand what the limiting shape of these clusters, described the stochastic structure of the clusters which form in this way. Oh, yeah. So maybe one thing to say is this: the parallel’s here with SLE, right? So if I look at the inverse map, then like I kind of get the composition in the right order for a stochastic flow. Oh. Maybe I did inverse on that. Plus this is not too visible, but, so the inverse maps of these conformal maps are stochastic flows. I can put a point into the flow and see it evolve. And this will follow some mock off evolution. And I mean, you know, in SLE you want to study either the trace or the whole. What you actually get access to it through is the level of the flow and much of the hard work is to translate stuff you can prove for this into properties of this. Okay? And that's exactly paralleled in the math we do here. But we can understand this map pretty well and then we work hard at it and are able to understand the inverse. So this is the theorem on the shape of the cluster. See what it says. This stuff at the beginning isn't so important. You should look at these three bullets. We’re really interested, so this describes how the cluster has evolved after a certain number of iterations of adding particles. Delta, we are considering the limit is Delta goes to zero. The size of the particles become small. When the number of particles, when you’ve added a number of particles of order, one over Delta squared, you've grown a part, grown approximately a ball of radius, well, some microscopic size. So the interesting values of N in this theorem are really when N is like one over Delta squared. Because if you add, if you scale down the size of the particles you're adding, then you need to scale up the number of particles that you add in order to see anything appreciable happen and the way that you need to do the scaling is this. So what do the three things say? Well, the first one says when we add a particle then in fact it ends up close to a scale up by the current size of the cluster to the point which it was actually attached on the unit circle. So we attached a point in the unit circle. We mapped it over to the cluster. It turns out really just some kind of moved out and scaled up. I mean this is despite the fact that the cluster has an extremely complicated boundary and in principle it could be attached anywhere. The particles don't get attached just anywhere. They just get attached kind of on a scaled up version of the unit circle. The particles don't get distorted hugely. Usually they're all close to each of the C, N plus I theta N. The second thing says there aren't any big holes in the cluster. So any point which is within the kind of approximate shape occupied by the cluster, there is some bit of the cluster K, N which is close to it. So we fill out a disc. And the second one is simply saying, this accounts for where all the particles go up to a certain maximum value of N, which can be taken to be pretty large. Remember it’s Delta minus two which is really of interest. So this is much larger, but this is just saying that none of the other particles penetrate in two, the particles are all attached in the boundary layer. It's a growing disc and you never get a particle which is attached kind of further and then it should be beyond an epsilon. So the limit shape is a ball, it fills out the ball, and we have some good control about where the particles are attached and there’s a final cluster. So that's one level of description. Disappointing in a way because it's a deterministic limit and we are hoping to see some sort of random effects as Delta goes to zero. So the rest of the talk is about discovering some random effects in the cluster which you might anticipate by looking at the picture here. So we're trying to describe this by theorem. What is there to say? Well, you can see bits of it are picked up by the theorem I just stated that it fills out the disc; you can see that as time goes by it decretes in layers, but what we are able to do is to be able to say something about the structure of these fingers. So that's really what the rest of the talk is devoted to, understanding the structure of the fingers. And to get access to that we have to think about this thing we call the harmonic measure flow. It's helpful understanding these things to take logs. So the boundary of the cluster becomes this, instead of E to the C, N we're looking at the points C, N plus I theta. And that’s the boundary of the cluster there. So if that’s the boundary of the cluster K, N and then we add another particle, so that's the boundary of the cluster K, N plus one. And when we take the, this is realized by some intertwining of the conformal map with the exponential function of the map from this line since we, the harmonic measure parameterizes the boundary here by theta going from not to two pi. Okay. So we can map this by phi, let's call it little phi N, distinct it from the big phi N which was exponentiated. So this is parameterized by theta from not to 2 pi. On the other hand, so is the other boundary, they boundary of K, N plus one is also parameterized by theta. And the boundary of K, N is a subset of the new boundary. So there is a map on the interval from not to 2 pi which takes the, see each point of the boundary here parameterized by harmonic measure, so suppose that's the parameters set there gets mapped to some parameter value in the new boundary of K N plus one only there has to be a jump. The map is essentially the identity, a way for some region here around the particle, but we have to kind of force in a new bit of boundary here. So you end up with this map, G N plus one or G, what did I call it? G theta N plus one. So when one gets a flow of the harmonic measure as you add more and more particles, and it’s by looking at this flow that we can understand the behavior of the fingers. So the way it goes is that we take, we always take a limit of these flows in some weak sense and understand the limit measure. In order to do that you have to have some space in which the flows live. To do weak limits in probability you have to understand the space in which the flows live in. So this is the space which the flows live in. So there's a picture here of a non-decreasing right continuous function which sort of a repeat itself periodically. So if I consider this set of all such functions D without, here without the periodic condition, then I'm going to say two functions like that are close if when you rotate the access, so instead of thinking it as being a non-decreasing right continuous function, I could just draw axes like that cross and it becomes a contraction. Okay? And if I consider two such functions, I could look at the uniform of distance between those two contractions. You’ve got to tilt your head by 45 degrees. So the distance of this function from the identity is just the maximum value of that it never gets away from that diagonal, for example. But instead of looking at the uniform norm we have to look at some usual way of just [inaudible] a little bit so it's a locally uniform norm. Okay. So the harmonic measure flow is a flow in the sense it has a kind of normal property that you'd expect to flows. You know, if you act by a one-time interval by the flow and then you flow on more by the next time interval then you get what you’d expect by taking the flow over the larger time interval. However this, the flow property turns out not to be robust under the sorts of topology which is possible to put on this flow space. There's a weaker property, a weaker flow property which is robust and that is written here. So these increasing functions have right and left limits. And the right way to think about the flow property to make it robust is say when you compose the left limits of two time intervals then you get to the left limit which is something that is less than or equal to the left limits of the longer time interval and then something analogous with inequality the other way around for the right limits. One property that a flow might have is that it be continuous in descent, that if you look at how far things move over short time intervals, it's not very far. So I want to consider continuous, in this sense weak flows, and on that space there's a way of defining a metric which turns the space of these flows into complete separable metric space. So there you have an object where you can start thinking about weak convergence. So we’ve looked at that picture before. That's how you define the distance between the two flows. Now there's a famous object in probability which is that coalescing Brownian flow which I now want to introduce because it turns out to be the limit object for the harmonic measure flows. So if I start with, think about Brownian motions on a circle and take a collection of starting points and I round Brownian motions forwards from these planes, and I have a rule that if I go through the starting points in turn, and if I hit Brownian motion already there, I just join up with it. Then it's easy to see that you get something, a distribution on coalescing Brownian parts which is independent of the order which is you set things up, and you can have more than one points in here so eventually you could have Brownian motions going from every rational point forwards, and then there's a good way to kind of complete this so that you end up with, you can say for every point on this line here where it goes to on this line. And these are pictures that you end up with is that actually every, they’re only finite images. Every point here actually is coalesced into one by a certain later time, and so that's what the function looks like. If you draw a graph of the function then it looks like a staircase. Not all the same size, but graph of this function is a staircase like that. >>: [inaudible]? >> James Norris: Oh. So this is, there's a theta here and it gets mapped to F of theta over there, all of those going to that one. Okay. So this is theta and that's F theta, like that. So, in fact you can think of the coalescing, or think of Arratias flow as being a probability measure on this continuous weak flow space. It doesn’t live just as, sometimes it’s necessary to complete spaces in order to support measures. It doesn't live on the space of perfect flows, but it does live nicely on the space of these continuous weak flows. So this is basically a result of Arratia. It's just reformulated in the language of these continuous weak flows. There's a unique [inaudible] property measure on the space with the property that if I, so this means I can’t drop a point into the flow and see where it goes, but it does a Brownian motion because from this [inaudible] characterization it would imply that this is a Brownian motion. And if I did this as a martingale, it says that if I drop two different points into the flow, then as soon as they hit each other they perform independent Brownian motions, again by [inaudible] characterization up until the time they first hit each other, TEE, and thereafter they perform a single Brownian motion because as soon as we get after this time we have to compensate the thing by the usual Brownian drift. Okay? So this is a neat way to give a martingale characterization of the Arratia flow. So more propaganda. This is a good way to think about the Arratia flow. You can invert all these continuous weak flows; just look at the inverse map. The time reversal map, which just takes this inverse, turns out to give an isometry of this space and it also preserves the law of the coalescencing Brownian flow. All right. So the harmonic measure flow is going to converge to the coalescing Brownian flow. And that's because the harmonic measure flow turns out to be one of these, what we've called disturbance flows. So I need to describe what a disturbance flow looks like in general. Suppose I take some basic disturbance. So I call it disturbance because it's not quite the identity. It's a little disturbance of the identity. So it looks like the identity for most values of theta, which we not in 2 pi, but then this is some little region where it's not quite the identity. Okay? And then suppose that I then randomize this theta here, I can sort of move this up and down the line to vary the value of theta. So there’s a particular function G, so this might be a picture of G theta, I specify G not and then G theta is obtained from it by translation. So I fix the disturbance G, G not, and then I form a flow just by iterating with random values of theta. So these are random functions. I choose these thetas uniformly random. How will a point move under this flow? Well, for most of the time it doesn't move very much because you're away from the disturbance. Every now and again the disturbance lands close to where you are and you get moved a little bit. And it's symmetric. So you’re just as likely to go up as you are to go down. Okay? So just think about motion of one point. What's the scaling limit going to be? Well by symmetry it's going to be Brownian motion. Now what about putting two points into this flow? How are they going to move? Well, it's going to be different times that you move the two points, provided their separated. Okay? So they'll move as independent Brownian motions until they’re close and then when they’re close together they'll tend to get hit, they'll get moved in the same way by the disturbance and so no surprise, really not very difficult to show at least for kind of finitely many points, that the limit object is coalescing Brownian motions. Okay? But in order to get those that we wanted it was necessary to consider this limit at the level of the flow is not just the level of finitely many points. So back to the slide. So you specify a disturbance and get a disturbance flow, you can do a sort of diffusive rescaling of the flow, which is described here, I don’t expect you to absorb it all, but if you look at the flow on various scales, just as when you prove convergence of random walks to Brownian motion, you can work in sort of, score a whole topology to understand that as the right topology. This is sort of scorer hot metric on the flow space and in that metric the rescaled disturbance flows converge weakly to the coalescing Brownian flow. Okay? So the limit of the disturbance flows, this is kind of a nice way to think, to realize the coalescing Brownian flow. In fact, I guess I should explain what this criteria, so one of the expresses is the size of the disturbance is becoming small. The size of the disturbance is smaller than the scale at which we are trying to look at the flow. Now the harmonic measure flow, the flow of the harmonic measure on the boundary of the cluster turns out to be exactly one of these disturbance flows. And so it, that flow converges to the coalescing Brownian flow. That, in fact, we knew a long time ago. But then by combining that with this sort of precise location of the clusters, which is given by the other theorem, we are able to transfer that information about the harmonic measure flow to the shapes of the fingers in the cluster. So this slide is just saying that the harmonic measure flow is converging to the coalescing Brownian flow. Now I want to talk about the fingers. If I take any point in the cluster, there's a nation of ancestry because the particle attached to another particle and you can trace back in time and watch where your ancestors were. Okay? And they won’t go sort of straight [inaudible] into the unit circle. They'll move around a little bit. So the finger of a point is its ancestral line in space. And I've drawn this picture in there. It’s a logarithmic scale. This also an escape route associated with a point. If I take some point and attach a piece of thread to it and then that thread leading out of the cluster through the gaps between the particles and I pull that thread tight, then I get a unique path from that point out to infinity. Outside the cluster. So for any given point, there's a finger going in and there's an escape route going out. >>: [inaudible] line segments and circle [inaudible]? >> James Norris: This is misleading because these have all been distorted by the conformed lab. But essentially, yes. The escape routes will have line segments and also bits which are round particles. Yeah. So because we know essentially where all the particles are by the earlier theorem, we can establish the limiting shapes of the fingers and the gaps by referring the particles sent to their thetas which come in from the harmonic measure. So in the end, for example show that if I take a finite collection of the sort of space and time starting points, so those will be points in this diagram, and consider for each point the finger going in and the escape route going out, so there are two paths associated to each starting point, one going forwards in time, one going backwards in time. So that gives a probability measure on I guess, paths which, a finite collection of paths which coalesce both when you go forwards in time and when they go backwards in time because the fingers tend to coalesce. But also the escape routes tend to coalesce. Right? >>: So the fingers goes one way, the gaps goes the other way? >> James Norris: Yeah. Let me draw a picture. Theta from not to two pi to a collection of points to start from, we trace back the finger here and this finger and this finger, they join up with that one. Okay? And this is a good time to have a, you trace the escape routes for this. The cluster is all over the place, right? So your escape route is very much constrained by the cluster. There are no big holes in the cluster. Here's the escape route for this one. Like that. So this escape route’s coming out here. There are more escape, it’s more constrained than it looks by the picture because I’m only considering finitely many points here, but here are the escape routes coalescing. Is that all of them? >>: [inaudible]? [inaudible] obtained in some version as a scaling image of [inaudible]? >> James Norris: So this is a picture which I'm, this is a sketch of something which is read off the cluster. You sort of take logs of the cluster, okay? But then this converges, in a weak sense, to the backwards and forward lines in the coalescing Brownian flow. That's the theorem. Okay? So we do understand kind of the stochastic structure. So we simulated the limit law, didn't succeed in making it very much like, so this is a simulation of backwards and forwards lines in the coalescing Brownian flow, but exponentiated around, so the light blue ones go out and the dark blue ones go in. So if you could see, at each point there is a light blue line going outwards and there's a dark blue line going inwards and they all coalesce, right? So that's supposed to look like this one, which is what the Hastings-Levitov simulation. >>: [inaudible] fingers [inaudible]? >> James Norris: That's correct. You have to imagine the gaps. >>: [inaudible] match the pictures [inaudible]? >> James Norris: Sure. Okay. But, okay. So you can just look at the dark blue lines in there, right? And then that's supposed to look like this one considered in monochrome. Okay. So I'll stop there. >>: Questions or comments? >>: Do you know how they simulated that? >> James Norris: I know roughly. I'm not the one who did it. So this is among the simulation. >>: But, I mean>> James Norris: Okay. So you basically work out where each pixel goes. The slit map, conformal map is something explicit, right? >>: Oh, okay. >> James Norris: Okay? You can just work it out. You're not slit map is for the other upper half plane. You just move around and you get the slit map for the disc. So that's something you can ask your computer to do and you can also get your computer to generate your random thetas and then you compose and you see where each point goes. I didn't get a picture. >>: I think it would be interesting to offer a little bit of historic perspective because this picture was not made in this high-resolution [inaudible]. It was created first in the late 80s actually. And, so 10 years before Hastings and Levitov. The only trace that left in the literature [inaudible] worked on this. He was actually the only trace of the literatures by a guy named Richard Rockburg[phonetic] gave a talk at [inaudible]conference. The title was actually Stochastic Movement, so that’s supposed to be, no actually it’s different. And in many ways the generation of just composing, you can fix conformal map [inaudible]. Now, of course, the computers were very lean, slow and [inaudible]. >> James Norris: So if I could add a sort of personal [inaudible] to that. I was just thinking what happens if you take SLE and you drive it by a poisson random measure. And then you, of course if you increase the intensity of the poisson random measure, you'd expect to get back to the big measure. If you derive SLE, but if you derive SLE by the big measure you just get an expanding disc. Nothing interesting. So do you get anything interesting when you drive it by a poisson random measure? Well, actually you do. Even when the particle size goes to zero, even when the kind of effective each little atom in the measure goes to zero, there's still something stochastic left if you do proper rescaling. Of course, we know this don't we? If we look at the lower large numbers and you do a proper rescaling, these Brownian fluctuations, they're not on the scale of this cluster. There are on a small scale here. So, okay. I reversed from the SLE to this, but then fell on this older idea. [inaudible] Makarov, I guess. >>: I guess I have one more question. So do you think one could take this and construct a model closer to SLE by building this but still adding some artificial randomness on the size of the fingers? So, in other words, if you look at an individual finger do expect it to look locally like a DLA finger or just with different structure? I mean, I don't expect it to look, construct a real DLA, but maybe something that will be a little more like it by building this and then kind of just run>> James Norris: Yeah. So there's another work by the three, Turner; Johansson, Sola, and Turner, where they investigate a rotationally inhomogeneous version of this story. So they add particles of different sizes at different angles. And that develops in a rather interesting way. I think, so you could have some sort of, you could cite in advance some sort of random way or>>: That's what I was>> James Norris: One might even try to think how could you drive that randomness by, I mean, the nice thing to do perhaps would be to sort of use stochastic fluctuations, which are kind of present in this to drive the, in a sense putting Alpha positive as doing exactly this, right? So that's the hard way to do it. So you're saying well, do it another way and make it look like DLA. Yes. That's a really good idea. >>: [inaudible]. >>: Okay. So thanks.