Monthly Archives: March 2013

Mixing Times 6 – Aldous-Broder Algorithm and Cover Times

Posted on by
1

In several previous posts, I’ve discussed the Uniform Spanning Tree. The definition is straightforward: we choose uniformly at random from the set of trees which span a fixed underlying graph. But for a dense underlying graph, there are a very large number spanning trees. Cayley’s formula says that the complete graph K_n has n^{n-2} spanning trees, so to select from this list is impractical.

We seek a better algorithm. In a post about a year ago, I presented the result that the path between two fixed points x and y in the UST is distributed as the path generated by Loop-Erased Random Walk, for which we start at x and delete cycles as they appear. An initial problem might be that this only gives us a single path, which might be enough in some contexts, but in general we will want to specify the whole tree. Wilson’s Algorithm is an unsurprising but useful extension to this equivalence which does just that. You start by constructing the LERW between two vertices, then you add the LERW which connects some other vertex to the path you already have. Then you take a further vertex not currently explored and start LERW there, continuing until you hit the tree that you already have. Iterate this process, which must terminate after at most n steps when there are no vertices which to start from. The tree thus obtained is the UST. The tricky part is proving that the method for selecting which unused vertices to start from has no effect on the distribution of paths between two fixed points.

I want to consider a different algorithm, discovered roughly simultaneously by Aldous and Broder. Start a random walk on the underlying graph at some particular vertex. Every time we traverse an edge which takes us to a vertex we haven’t yet explored, add this edge to the tree. For now I don’t want to give a proof that this algorithm works, but rather to talk about how fast it works, because it ties in nicely with something from the Mixing Times book we’ve been reading recently. It is clear that the algorithm terminates at the first time the random walk has visited every vertex. This is a stopping time, called the cover time of the Markov chain. If we are working with an underlying complete, then we notice that this is annoying, because it means that the cover time will increase like n.log n. That is, it will take an increasingly long time to gather the final few vertices into the tree. Perhaps some combination of Aldous-Broder initially then Wilson’s method for the final o(n) vertices might be preferable?

I want to discuss how to treat this cover time. Often we have information about the hitting times of states from other states \mathbb{E}_x T_y. A relationship between S, the hitting time, defined to be the maximum of the previous display over x and y, and the expected cover time would be useful, especially for a highly symmetric graph like the complete graph where the expected hitting times are all the same.

Matthews’ Method relates these two for an irreducible finite Markov chain on n states. It says:

t_{cov}\leq t_{hit}\left(1+\frac12+\ldots+\frac 1 n\right).

We first remark that this agrees with what we should get for the random walk on the complete graph. There, the hitting time of x from y is a geometric random variable with success probability 1/n, hence expectation is n. The cover time is the standard coupon collector problem, giving expectation n log n, and the sum of reciprocals factor is asymptotically a good approximation.

The intuition is that if we continue until we hit state 1, then reset and continue until we hit state 2, and so on, by the time we hit state n after (n-1) iterations, this is a very poor overestimate of the cover time, because we are actually likely to have hit most states many times. What we want to do really is say that after we’ve hit state 1, we continue until we hit state 2, unless we’ve already done so, in which case we choose a different state to aim for, one which we haven’t already visited. But this becomes complicated because we then need to know the precise conditional probabilities of visiting any site on the way between two other states, which will depend rather strongly on the exact structure of the chain.

Peres et al give a coupling proof in Chapter 11 of their book which I think can be made a bit shorter, at least informally. The key step is that we still consider hitting the sites in order, only now in a random order.

That is, we choose a permutation \sigma\in S_n uniformly at random, and we let T_k be the first time that states \sigma(1),\ldots,\sigma(k) have all been visited. This is a random time that is measurable in the product space, and for each \sigma it is a stopping time.

The key observation is that \mathbb{P}(T_{k+1}=T_k)=1-\frac{1}{k+1}. This holds conditional on any path of the Markov chain because the requirement for the event is that \sigma(k+1) is visited after \{\sigma(1),\ldots,\sigma(k)\}. The statement therefore holds as stated as well as just pathwise. Then, by the SMP, conditional on \{T_{k+1}>T_k\}, we have

T_{k+1}-T_k \leq_{st} t_{hit}.

Note that by the definition of t_{hit}, this bound on the hitting time T_{k+1} is unaffected by concerns about where the chain actually is at T_k (since it is not necessarily at \sigma(k)).

So, removing the conditioning, we have:

\mathbb{E}\Big[T_{k+1}-T_k\Big]\leq\frac{1}{k+1}t_{hit},

and so the telescoping sum gives us Matthews’ result.

One example is the cover time of random walk on the n x n torus, which turns out to be

O(n^2(\log n)^2).

If anyone remembers that Microsoft screensaver from many years ago which started with a black screen and a snake leaving a trail of white pixels as it negotiated the screen, this will be familiar. The last few black bits take a frustratingly long while to disappear. Obviously that isn’t quite a random walk, but it perhaps diminishes the surprise that it should take this long to find the cover time.

There are a couple of interesting things I wanted to say about electrical networks for Markov chains and analytic methods for mixing times, but the moment may have passed, so this is probably the last post about Mixing Times. Plans are in motion for a similar reading group next term, possible on Random Matrices.

Related articles

Minimum Spanning Trees

Posted on by
2

In my last post, I discussed the Uniform Spanning Tree. To summarise very briefly, given a connected graph on n vertices, a tree is a subgraph, that is a subset of the edges, which is connected, but which contains no cycles. It turns out this requires the tree to have n-1 edges.

We are interested in natural mechanisms for generating randomly chosen spanning trees of a given graph. One way we can always do this is to choose uniformly at random from the set of possible trees. This UST is in some sense canonical, but it is worth knowing about some other measures on trees that might be of interest.

A family of natural problems in operations research concerns an arbitrary complex network, with some weight or cost associated to each connection. The question is how to perform some operation on the network so as to minimise the resulting cost. Perhaps the most famous such problem is that of the Travelling Salesman. The story is that a salesman needs to visit n locations and wants to do the trip as efficiently as possible. This might be thought of as some sort of financial or time cost, but proably the easiest way to set it up is to imagine he is trying to minimise the distance he has to travel. It is not hard to see why this problem might genuinely arise in plenty of real-world situations, where a organisation or agent is trying to be as efficient as possible.

It might be the case that it is not possible to travel between every pair of locations, but we needn’t assume that for now. So if he knows the distance between any pair of cities, he wants to know which of the possible routes gives the shortest overall distance. The problem is that there are n! routes, and this grows roughly like n^n, which is faster than exponential, so for as few as 20 cities it has turned into a comparison which is too large to compute.

There are various algorithms which reduce the number of routes that must be checked, and some approximation methods. But if you want the exact answer, it is not currently possible to calculate this in polynomial time.

Minimal Spanning Trees and Uniqueness

For the travelling salesman, we were looking for the minimal cost spanning path. In the case of the complete graph, this is the same as the minimal cost non-repeating path of length n-1. Such paths are a subset of the set of spanning trees on the underlying graph. So what if we look instead for the minimal cost spanning tree? This exists as after all, there are only finitely many spanning trees.

So far, this has been deterministic, but we were looking for a random spanning tree. We can achieve this by choosing the weights at random. Anything other than assigning the weights as an IID sequence seems likely to be complicated, but there isn’t a canonical choice of the distribution of the weights. Our first question will be whether the distribution of the weights affects the distribution of the induced MST. In fact it will turn out that so long as the distribution is continuous, it has no effect on the distribution of the MST. The continuous condition might seem odd, but it is present only to ensure that the weights almost certainly end up generating a unique MST.

It turns out that there is a straightforward greedy algorithm to find the MST once the weights are known. We will examine some consequences of this algorithm in the random setting. First we check uniqueness. The condition required for uniqueness is that the weights be distinct. Note that this is slightly weaker than the statement that all of sums of (n-1)-tuples be distinct, which immediately implies a unique MST.

We now prove this condition. Suppose we have distinct weights, and an associated MST. If the underlying graph is a tree, then the result is clear. Otherwise, add some extra edge e, with weight w(e). By the definition of a tree, this generates exctly one cycle. Consider the other edges, say e_1,\ldots,e_k in this cycle. If any of w(e_i)>w(e) then we can replace e_i with e to get a spanning tree with smaller weight, a contradiction of the claim that we started with an MST. So by distinctness of weights, we conclude that w(e)>w(e_i) for all i.

Conversely, suppose we remove some edge e which IS in the MST. We end up with exactly two connected components. Consider all the edges in the underlying graph between the two components, and suppose that one of these f satisfies w(f)<w(e). Then if we add in edge f, which is by construction not in the original MST, we end up with a smaller total weight than we started with, a further contradiction.

We can summarise this in a neat form. Given an edge e between x and y, consider the set of all edges in the underlying graph with weight LESS THAN w(e). Then if x and y are in different components, the edge e must be in the MST. Since we have an explicit description of which edges are present, it follows that the MST is unique. The problem is that working out the component structure of the graph with higher weights removed is computationally rather intensive. We want a slightly faster algorithm.

Kruskal’s Algorithm

Several rather similar algorithms were developed roughly simultaneously. Prim’s algorithm is a slight generalisation of what we will discuss. Anyway, for now we consider Kruskal’s algorithm which has the advantage that it can be described without really needing to draw a diagram.

We start by ordering the weights. Without loss of generality, we might as well relabel the edges so that

w(e_1)< w(e_2)<\ldots< w(e_{|E|}).

Now, by the condition derived in the argument for uniqueness, we must have e_1 and e_2 in any MST. Now consider e_3. Unless doing so would create a cycle, add e_3. Then, unless doing so would create a cycle, add e_4. Continue. It is clear that the result of this procedure is acyclic. To check it is actually a spanning tree, we show that it is also connected. Suppose not, and two of the components are A and B. Let e be the edge between A and B with minimal weight. According to the algorithm, we should have included e in our MST because at no point would adding it possibly have created a cycle. So we have proved that this greedy algorithm does indeed give the (unique) MST.

A useful consequence of this is that we know the two edges with overall minimum weight are definitely in the MST. In the search for a random measure on spanning trees, what is most important is that we didn’t use the actual values of the weights in this construction, only the order. In other words, we might as well have assumed the weights were a random permutation from S_{|E|}. This now answers our original question about how the random weight MST depends in distribution on the underlying edge weight distribution. So long as with probability one the weights are distinct (which holds if the distribution is continuous), then the distribution of the resulting spanning tree is constant.

It’s not too hard to show this isn’t the same as UST: n=4 suffices as a counterexample. But the difference in asymptotic behaviour of properties such as the diameter is of interest, and will be explored in the next post.

Related articles

Uniform Spanning Trees

Posted on by
7

For applications to random graphs, the local binomial structure and independence means that the Galton-Watson branching process is a useful structure to consider embedding in the graph. In several previous posts, I have shown how we can set up the so-called exploration process which visits the sites in a component as if the component were actually a tree. The typical degree is O(1), and so in particular small components will be trees with high probability in the limit. In the giant component for a supercritical graph, this is not the case, but it doesn’t matter, as we ignore vertices we have already explored in our exploration process. We can consider the excess edges separately by ‘sprinkling’ them back in once we have the tree-like backbone of all the components. Again, independence is crucial here.

I am now thinking about a new model. We take an Erdos-Renyi process as before, with edges arriving at some fixed rate, but whenever a cycle appears, we immediately delete all the edges that make up the cycle. Thus at all times the system consists of a collection (or forest) of trees on the n vertices. So initially this process will look exactly like the normal E-R process, but as soon as the components start getting large, we start getting excess edges which destroy the cycles and make everything small again. The question to ask is: if we run the process for long enough, roughly how large are all the components? It seems unlikely that the splitting mechanism is so weak that we will get true giant components forming, ie O(n) sizes, so we might guess that, in common with some other split-merge models of this type, we end up with components of size n^{2/3}, as in the critical window for the E-R process.

In any case, the scaling limit process is likely to have components whose sizes grow with n, so we will have a class of trees larger than those we have considered previously, which have typically been O(1). So it’s worth thinking about some ways to generate random trees on a fixed number of vertices.

Conditioned Galton-Watson

Our favourite method of creating trees is inductive. We take a root and connect the root to a number of offspring given by a fixed distribution, and each of these some offspring given by an independent sample from the same distribution and so on. The natural formulation gives no control over the size of the tree. This is a random variable whose distribution depends on the offspring distribution, and which in some circumstances be computed explicitly, for example when the offspring distribution is geometric. In other cases, it is easier to make recourse to generating functions or to a random walk analogue as described in the exploration process discussion.

Of course, there is nothing to stop us conditioning on the total size of the population. This is equivalent to conditioning on the hitting time of -1 for the corresponding random walk, and Donsker’s theorem gives several consequences of a convergence relation towards a rescaled Brownian excursion. Note that there is no a priori labelling for the resulting tree. This will have to be supplied later, with breadth-first and depth-first the most natural choices, which might cause annoyance if you actually want to use it. In particular, it is not obvious, and probably not true unless you are careful, that the distribution is invariant under permuting the labels (having initially assumed 1 is the root etc) which is not ideal if you are embedding into the complete graph.

However, we would like to have some more direct constructions of random trees on n vertices. We now consider perhaps the two best known such methods. These are of particular interest as they are applicable to finding random spanning trees embedded in any graph, rather than just the complete graph.

Uniform Spanning Tree

Given a connected graph, consider the set of all subgraphs which are trees and span the vertex set of the original graph. An element of this set is called a spanning tree. A uniform spanning tree is chosen uniformly at random from the set of spanning trees on the complex graph on n vertices. A famous result of Arthur Cayley says that the number of such spanning trees is n^{n-2}. There are various neat proofs, many of which consider a mild generalisation which gives us a more natural framework for using induction. This might be a suitable subject for a subsequent post.

While there is no objective answer to the question of what is the right model for random trees on n vertices, this is what you get from the Erdos-Renyi process. Formally, conditional on the sizes of the (tree) components, the structures of the tree components are given by UST.

To see why this is the case, observe that when we condition that a component has m vertices and is a tree, we are demanding that it be connected and have m-1 edges. Since the probability of a particular configuration appearing in G(n,p) is a function only of the number of edges in the configuration, it follows that the probability of each spanning tree on the m vertices in question is equal.

Interesting things happen when you do this dynamically. That is, if we have two USTs of sizes m and n at some time t, and condition that the next edge to be added in the process joins them, then the resulting component is not a UST on m+n vertices. To see why, consider the probability of a ‘star’, that is a tree with a single distinguished vertex to which every other vertex is joined. Then the probability that the UST on m vertices is a star is \frac{m}{m^{m-2}}=m^{-(m-3)}. By contrast, it is not possible to obtain a star on m+n vertices by joining a tree on m vertices and a tree on n vertices with an additional edge.

However, I think the UST property is preserved by the cycle deletion mechanism mentioned at the very start of this post. My working has been very much of the back of the envelope variety, but I am fairly convinced that once you have taken a UST and conditioned on the sizes of the smaller trees which result from cycle deletion. My argument is that you might as well fix the cycle to be deleted, then condition on how many vertices are in each of the trees coming off this cycle. Now the choice of each of these trees is clearly uniform among spanning trees on the correct number of vertices.

However, it is my current belief that the combination of these two mechanisms does not give UST-like trees even after conditioning on the sizes at fixed time.

Related articles

Beyond Erdos-Renyi: more realistic models of networks

Posted on by
4

The claim is often made that the study of random graphs such as the Erdos-Renyi model is worthwhile because it gives us information about complex systems which exist in the real world. The internet or social networks provide the example du jour at the moment, but it’s equally plausible to think about traffic flows, electrical systems or interacting biological processes too.

If this were entirely true, it would be great for two reasons. Firstly, in my opinion at least, it is a beautiful subject in its own right, and to have a concrete applicable reason to continue studying it would make it even better. (Not to mention the dreaded competition for funding…) Secondly, Erdos-Renyi is so simple. After all, it involves little more than adding some simple topology to a collection of IID Bernoulli random variables, and so it would surely be possible to draw some significant conclusions about how complicated real-world objects interact without too much mathematical effort.

Unfortunately, but unsurprising, this simplicity is a drawback as far as applications go. It is fairly clear that most real-world systems cannot offer any property even approaching the niceness of the independent, same probability edges condition. But rather than consign E-R to the ‘pretty but useless’ category of mathematical structures, we should think carefully about exactly why it fails to be a good model for real-world networks, and see whether there are any small adjustments that could be made to improve it.

This is something I’ve been meaning to read up about for ages and ages. What follows is based heavily on the Albert and Barabasi 2002 review paper. I suspect that many of the open problems and intuitive calculations have since been finished and formalised, but for an overview I hope that doesn’t matter hugely. I’ve also leafed through the relevant chapters of Remco Van der Hofstad’s notes, but am setting the details and the exercises aside for the holidays when I have a bit more time!

Problems with Erdos-Renyi

Recall that G(n,p) takes n vertices, and adds edges between any pair of vertices independently with probability p.

One property shared by most real-world networks is the scale-free phenomenon, which says that the degree distribution has a power law tail. The Albert-Barabasi papers gives a comprehensive survey of data verifying this claim. By contrast, G(n,p) has degree distribution which is approximately Poisson as n grows. This is concentrated near the average degree with a thin exponential tail, so does not satisfy this requirement. I was and still am a bit confused by the term ‘scale-free’. The idea is certainly that the local structure is independent of the size of the system, which seems to be true for the degree distributions in sparse ER, that is where p = O(1/n). But I think the correct heuristic is that it doesn’t matter how far zoomed in you are – the macroscopic structure looks similar for n vertices as for n^2 vertices. This certainly fails to be true for ER, where no vertex has O(n) neighbours, whereas with a power law tail, this does hold.

The main consequence of this is that there are a few vertices with very high degree. These are often called ‘hubs’ and parallels are drawn to the internet, where key websites and servers connect lots of traffic and pages from different areas. The idea is that the hubs are almost certainly well-connected to each other, and this offers a step towards a small-world phenomenon, where the shortest path between any two vertices is very small relative to the size of the system. This notion was introduced to mainstream culture by Stanley Milgram’s ‘Six degrees of separation’ experiment in the 60s, where it became clear that subjects were able to deliver a package to a complete stranger on the other side of the USA, using only personal contacts, in about six stages. The graph theoretic notion for this is the diameter, defined as the maximal graph distance between two points. Here, the graph distance means the length of the shortest path between the points. This definition, with the max-min formalism looks rather complicated, but isn’t really. The diameter of an Erdos-Renyi graph for fixed p, increases like log n, which is small relative to n, and so this property holds.

A quick glance at your list of Facebook friends will confirm that the independent edges condition in an Erdos-Renyi random graph is not a plausible model for social networks. How many friends do you have? Let’s say about 1000, more to make the calculation easier than because you’re necessarily very popular. How many does your friend Tom have? Let’s say 1000 again. As was in the news a few months ago, there are now over a billion people on Facebook. Let’s say exactly a billion (that is 10^9 for these purposes). So both you and Tom are friends with 1/10^6 of the total membership of the network. So how large would you expect the overlap of your friendships to be, if they were all chosen independently at random? Well, the probability that you are both friends with Alice is 10^-12, and so the expected number of your mutual friends is 10^-12 x 10^9 = 10^-3 which is substantially less than 1. Yet I imagine if you substituted names suitably, you and Tom might well have over 50 mutual friends if you were, say, in the same year at school or niversity and haven’t yet purged your list.

We want a statistic that records this idea quantitatively. There are various candidates for such a clustering coefficient. The underlying notion that we might expect there to be greater connectivity between neighbours of some fixed point v than in the graph as a whole gives an intuition for a possible definition. Compare the proportion of triangles in the graph to the cube of the proportion of edges. When this ratio is large, then there is a lot of clustering. In the E-R case, we would expect these to be equal, as the probability of forming a triangle is equal to the cube of the probability of the presence of each of the three independent edges that make up the triangle.

So we have three properties of real networks that we would like to incorporate into a model: small diameter, power-law degree distribution, and high clustering. To avoid this turning into a book, I’m going to write a paragraph about each of the possibilities discussed by Albert and Barabasi.

Generalised Random Graph

The degree distribution will typically emerge as a consequence of the construction of a given model. The general idea here is to condition on the degree distribution having the form we want, and see what this does to the structure. Of course, the choice of how to do this conditioning is absolutely key. It certainly isn’t obvious what it means to ‘condition G(n,p) to have power-law distribution’, since the very idea of a power-law vs exponential tail requires the number of vertices to be large.

The first idea for achieving this gives the vertices ‘stubs’, which join up in pairs to form edges. We decide on the distribution of stubs according to this power law, then pair them up uniformly at random. Obviously, there is a possibility of getting some loops, but this is not going to happen so often as to be a genuine problem in the limit. This construction is similarly open to the branching process exploration ideas well covered for the E-R random graph, though we have to be careful to size-bias the degree distributions when necessary. There is still an underlying independence in the location of edges though, so it is reasonably clear that the amount of clustering may be closer to E-R than to the real examples cited.

The other possibility suggested is to retain the independent edge property, but give the vertices weights, and let the probability of an edge between two vertices be some sensible function of the weights. In the end it turns out to make little difference whether the weights are chosen deterministically or randomly, but by taking the weights i.i.d. with infinite mean, we can generate a so-called generalised random graph where the degree distribution has a power law.

Watts-Strogatz

In the WS model, the idea is to interpolate between a graph with maximal clustering and a random graph. A d-regular graph, say on a ring, where every vertex is connected to its d nearest neighbours has high clustering, but large diameter, as for example it takes roughly n/2d steps to get to the other side of the ring. Whereas in the standard E-R model we add edges with some fixed probability p, here we replace edges with some fixed probability p. That is, we take an edge in the regular graph and with some small probability we remove it and instead add an edge between two vertices chosen uniformly at random. The theoretical motivation is that removing a few edges doesn’t destroy the high clustering evident in the regular graph, but even a sparse random graph has small diameter, so adding a few ‘long-range’ edges should be enough to decrease the diameter significantly.

It obviously needs to be checked that a substantial drop in diameter occurs before a substantial decrease in clustering, and there is a calculation and diagram to support this intuitive idea in the paper. The one drawback of this model is that it fails to provide the power-law degree distributions we want. After all, an E-R graph has a concentrated degree distribution, and a d-regular graph has all degrees the same, so we would expect some interpolation between the two to have a concentrated distribution as well. Nonetheless, this model accords well with an idea of how complex networks might form, particularly if there is some underlying geometry. It is reasonable to assume that an initial setup for a network would be that people are connected to those closest to them, and then slowly acquire distant contacts as time progresses.

Preferential Attachment – Barabasi-Albert model

Most of our intuition for networks can be extended to an intuition for the formation of networks. The idea of prescribing a degree distribution is neat, but it doesn’t give any account to the mechanism of formation. Complexity emerges over time, and a good model should be able to describe why this happens. The Barabasi-Albert model takes this as its starting point, with the aim of producing a highly clustered system dynamically. Recall that we can describe G(n,p) as a process by coupling, then increasing p from 0 to 1, and seeing edges emerge. The independence assumption can be lifted through the coupling, and so which edge appears next is independent of the current state of the system.

This is what we need to relax. Recall the motivating idea of ‘hubs’, where a small collection of vertices have very high connectivity across the whole system, as observed in several real situations. A consequence of this is that new edge is more likely to be attached to a hub, than to a pair of poorly connected vertex elsewhere. But it turns out that this idea of preferential attachment isn’t enough by itself. Because as a network forms, it is not just the connectivity that increases, but also the size of the system itself. So in fact it makes sense to add vertices rather than edges, and join the new vertices to existing vertices in proportion to the degrees of the existing vertices. This combination of growth and preferential attachment is key to the scale-free graphs that this Barabasi-Albert model generates. Relaxing either mechanism returns us to the case of exponential tails. However, there are methods in the literature for generating such graphs without the need for a dynamic model, but they are harder to understand and describe. None I have seen so far has a high clustering coefficient.

Hubs are effectively a way to reduce the diameter. Recall the description of Milgram’s experiment where he encouraged randomly chosen people to send a package to Harvard. For the purposes of this model, an undergraduate from Wyoming or a husband from Alabama moving in with his wife in Boston are clear hubs, as for very many people near their previous home, they represent a good connection to Harvard. So it is unsurprising that BA, which reinforces hubs, has a sub-logarithmic small diameter.

Conclusions

I’m not entirely what conclusions I should draw from my reading. Probably the main one is that I should read more as there is plenty of interesting stuff going on in this area. Intuitively, it seems unlikely that there is going to be a single model which unites the descriptions of all relevant real-world networks. As ever, it is pleasant to find structures that are both mathematically interesting in their own right and relevant to applied problems. So it is reassuring to observe how similar many of the models discussed above are to the standard random graph.

Related articles