Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Random Walks on Graphs: A Survey
Robert Lazar
Department of MathematicsIowa State University
April 26, 2016
Our Mission
I Goals:� Clear understanding of Random Walks including basic
definitions� Find a few bounds for random walk properties� Learn a few concrete examples
I Tool: Random Walks on Graphs: A Survey (1993), L.Lovasz
Robert Lazar (ISU) 2 of 28
Outline
I Random Walk as a Markov Chain� Stationary Distribution� Time-Reversibility� Expected Return Time
I Eigenvalue Connection
I Main Parameters� Bound on Access Time
I Universal Traverse Sequence� Existence for d-regular Graph
Robert Lazar (ISU) 3 of 28
Random Walk
Definition
Let G be a graph on n vertices and v0 be a vertex selected bysome probability distribution P0. A random walk on the graph Gwith starting vertex v0 is a sequence of vertices v0 = X0,X1 · · · ,where Xi+1 is uniformly selected from N(Xi ).
Definition
A Markov chain is a sequence of random variables X0,X1, · · ·with state space V , where the probability moving to variable Xi+1
only depends on variable Xi . A Markov chain is said to beirreducible if any state in the state space has a non-zeroprobability of traveling to another state. Let the transition ofprobabilities for the Markov Chain be M.
Robert Lazar (ISU) 4 of 28
Random Walks Continued
Example
Let G be K3 with a leaf. Then
M =
0 1
313
13
12 0 1
2 012
12 0 0
1 0 0 0
M is row stochastic as it is the sum of all of its neighbors. We canalso define M in the following way:
Robert Lazar (ISU) 5 of 28
Random Walks Continued
Definition
The t-th step of the walk can be represented as Pt = (Mᵀ)tP0,where M = DAG is the matrix of transition probabilities of thisgraph G ,
mij =
{1
d(i) , if ij ∈ E ,
0, otherwise,
D is the diagonal matrix with dii = 1d(i) and P0 is some initial
distribution.
When G is d-regular we have
M = DA =1
dA.
Robert Lazar (ISU) 6 of 28
Random Walks Continued
Example
Continuing our example above with G being a K3 with a leaf.
AG =
0 1 1 11 0 1 01 1 0 01 0 0 0
,D =
13 0 0 00 1
2 0 00 0 1
2 00 0 0 1
.Then,
M = DAG =
0 1
313
13
12 0 1
2 012
12 0 0
1 0 0 0
Robert Lazar (ISU) 7 of 28
Random Walks Continued
Definition
The probability distribution P0 is stationary for the graph G ifP1 = P0.
In fact, if P0 is stationary, Pt = (Mᵀ)tP0 = P0.
Observation
For any graph G the distribution,
π(v) =d(v)
2m,
is stationary.
P0 is the matrix where∑n
i=1 d(vi ) = 2m and each row is12m (d(v1), d(v2), · · · , d(vn)). Clearly P0 is row stochastic and it isstationary as the entries in each row of M are 0 or 1
d(i) .
Robert Lazar (ISU) 8 of 28
Random Walks Continued
Is this stationary distribution unique and is there a way to derive it?
Remark
If the Graph G is connected then the stationary distribution isunique.
Theorem
Perron-Frobenius: For a non-negative irreducible square matrix,there exists a unique distribution π such that Pπ = π and∑π(i) = 1. As well as a real eigenvalue λ which has maximum
absolute value among all eigenvalues and multiplicity 1.
Since G is connected then a random walk corresponds to anirreducible Markov Chain. So by the Perron-Frobenius Theorem,the stationary distribution is unique.
Robert Lazar (ISU) 9 of 28
Random Walks Continued
Definition
A random walk is time-reversible if π(i)pij = π(j)pji . In otherwords, walking from i to j is the same as from j to i .
Example
Let M be from the example above with P0 = I ,
M =
0 1
313
13
12 0 1
2 012
12 0 0
1 0 0 0
.Since M is not symmetric this walk is not time-reversible.
Robert Lazar (ISU) 10 of 28
Random Walks Continued
Remark
If we require a random walk to be time-reversible then we obtainthe distribution π(i) = d(i)
2m .
Since pij = 1d(i) we have
π(i)
d(i)=π(j)
d(j)= k .
As π is row stochastic and∑
d(i) = 2m,
1 =∑
π(i) =∑
kd(i) = k∑
d(i) = k2m.
Therefore, we obtain k = 12m .
Robert Lazar (ISU) 11 of 28
Random Walks Continued
Hence,
π(i) =k
pij=
d(i)
2m
as obtained above. If G is d-regular, then∑
d(i) = dn and
π(i) =1
n,
the uniform distribution.
Robert Lazar (ISU) 12 of 28
Random Walks Continued
Remark
1) The expected return rate to an edge is 2m.
2) The expected return rate to a vertex is 2md(i) , if G is d-regular
the expected return rate is n.
Proof.
Let E be the expected value that we return to an edge/vertex withprobability 1
k . Then E = 1k + (1 + E )(1− 1
k ). Solving for E givesus E = k. For 1) we use k = 1
2m as each edge has equally likely
chance of being used. For 2) use k = d(i)2m .
Now, π will represent the unique stationary distribution. We willsee how this relates to eigenvalues.
Robert Lazar (ISU) 13 of 28
Eigenvalue Connection
Theorem
For a connected non-bipartite graph G , the stationary distributionis also a limiting distribution.
Proof.
Recall that M = DAG does not always have to be symmetric.Consider
N = D1/2AGD1/2 = D−1/2MD1/2.
D is invertible as graph is connected. Eigenvalues of N are thesame as the eigenvalues of M. Since the matrix is real symmetric,we can order its eigenvalues as λ1 ≥ · · · ≥ λn with correspondingorthonormal eigenvectors v1, · · · , vn.
Robert Lazar (ISU) 14 of 28
Eigenvalue Connection Continued
Proof.
By spectral decomposition,
N = V ᵀλV =n∑
i=1
λivᵀi vi .
Let w = (√
d(1), · · · ,√d(n))ᵀ,
Nw = D1/2AGD1/2w = D1/2AG1 = D1/2(d(1), · · · , d(n))ᵀ = 1w .
Therefore, w is an eigenvector of N with eigenvalue 1. ByPerron-Frobenius Theorem, 1 = λ1 > · · · ≥ λn ≥ −1. Hence,v1 = w√
2m, the normalized w eigenvector.
(D1/2v1vᵀ1D−1/2) = π.
Robert Lazar (ISU) 15 of 28
Eigenvalue Connection Continued
Proof.
Then
M = D1/2ND−1/2
= (D1/2v1vᵀ1D−1/2 +
n∑i=2
λiD1/2vᵀi viD
−1/2
= π +n∑
i=2
λiD1/2vᵀi viD
−1/2.
Hence,
Mt = π +n∑
i=2
λtiD1/2vᵀi viD
−1/2.
Robert Lazar (ISU) 16 of 28
Eigenvalue Connection Continued
Proof.
If G is bipartite then its spectrum is symmetric about the origin(Lovasz, Eigenvalues of graphs 2007). If G is bipartite
AG =
[0 BC 0
],
so the sum of the eigenvalues is 0. If G is non-bipartite, then bythe Perron-Frobenius theorem, λn > −1 hence
limt→∞
Mt = π +n∑
i=2
λtiD1/2vᵀi viD
−1/2 = π + 0 = π.
Robert Lazar (ISU) 17 of 28
Main Parameters
Definition
The access time Hij is the expected number of steps before nodej is visited, starting from node i .
Definition
The cover time is the expected number of steps to reach everynode. If no starting node is specified we take the worst case.
Definition
The mixing rate is a measure of how fast the random walkconverges to its limiting distribution.
Robert Lazar (ISU) 18 of 28
Access Time
Example
Determine the access time for two points of a path on n nodeslabeled 0, · · · , n − 1. Notice that H(k − 1, k) is one less than theexpected return time to vertex k. As calculated before, theexpected return time is
∑d(i)
d(k) . Since we have a path of length k ,∑d(i) = 2(k − 1) + 2 = 2k and d(k) = 1. Hence, the expected
return time is 2k so H(k − 1, k) = 2k − 1. Now,
H(i , k) =k∑
j=i+1
H(i , j) =k∑
j=i+1
2(j)− 1 = k2 − i2.
Robert Lazar (ISU) 19 of 28
Cover Time
Example
Let G be the complete graph on n vertices with τi representing thefirst time i vertices have been visited. Then, τi+1 − τi representsthe amount of time for a new vertex to be visited. Now, theprobability we can visit a new vertex is n−i
n−1 since there are n − ivertices we have not visited and n − 1 vertices that are not ourinitial vertex. As above,
E (τi+1 − τi ) =1
n−in−1
=n − 1
n − i.
Therefore, the cover time is
E (τn) =n−1∑i=1
E (τi+1−τi ) =n−1∑i=1
n − 1
n − i= (n−1)
n−1∑i=1
1
i︸ ︷︷ ︸HarmonicSeries
≈ n log n
Robert Lazar (ISU) 20 of 28
Cover Time Continued
In 1993 Feige was interested in the Coupon Collector Problem. Ifyou want to collect each of n different coupons, and you get arandom coupon every day in the mail how long do you have towait?
Theorem
(Feige 1993). The cover time of a regular graph on n nodes is atmost 2n2.
Remark
The access time is at most 2n2 since we have to cover the wholegraph in that amount of time.
Robert Lazar (ISU) 21 of 28
Universal Traverse Sequence
Definition
Let G be a connected d-regular graph and v0 some starting vertex.Label the edges incident to each node as {1, · · · , d}. A traversesequence for the current label and starting vertex v0 is a sequence(h1, · · · , ht) ⊆ {1, · · · , d}t such that if we start a random walk atv0 then at the the i-th step we leave the edge labeled hi , we havevisited every node. A universal traverse sequence for a fixed nand d is a sequence which is a traverse sequence for every d-regulargraph on n vertices, every labeling of it, and every starting point.
Robert Lazar (ISU) 22 of 28
Universal Traverse Sequence Continued
Example
K3. (1,2,2) is Universal while (1,1,1) is not. Let 0 represent thelabeling (1,2) and 1 represent the labeling (2,1). Then the 8labelings are (0,0,0),(0,0,1),(0,1,0),(1,0,0),(0,1,1),(1,1,0),(1,0,1)and (1,1,1).
Robert Lazar (ISU) 23 of 28
Existence of Universal Traverse Sequence
Theorem
For every d ≥ 2 and n ≥ 3, there exists a universal traversesequence of length O(d2n3 log(n)).
Proof.
I Construct a random sequence
I Calculate probability that sequence is not traverse
I Use Bound on Cover Time
I Use Markov’s Inequality
I Conclude that there is at least one universal traverse sequence
Robert Lazar (ISU) 24 of 28
Existence of Universal Traverse Sequence Cont.
Proof.
Let t = 8dn3 log n and H = (h1, · · · , ht) is a random walk. Theprobability that H is not traverse is bounded by the cover time,2n2. Let X be the event that we have not seen all nodes, byMarkov’s Inequality
P(X ≥ 4n2) ≤ E (X )
4n2<
2n2
4n2=
1
2.
We can think of 4n2 as a random walk, in fact we will have t4n2
more random walks that also should not cover the graph, eachhaving an independent probability of 1
2 . The probability that wehave not seen all vertices after t-steps is less than
2−t
4n2 = 2−2dndlog2 ne ≤ n−2dn.
Robert Lazar (ISU) 25 of 28
Existence of Universal Traverse Sequence Cont.
Proof.
For a fixed vertex in a d-regular graph on n-vertices, there are(n − 1
d
)≤ nd
choices for its neighbors. So for each vertex we have at most nd
different edges hence there are at most ndn d-regular graphs withdifferent edge labelings. Combining our results, we have ndn
possible labelings, n possible starting vertices and probabilityn−(nd+2) that we have not seen all vertices after t steps.So the probability that H is not a traverse sequence for one ofthese graphs is
nndnn−(nd+2) = n−1 < 1,
as n ≥ 3. So there exists some universal traverse sequence.
Robert Lazar (ISU) 26 of 28
Further Work
I The access time can be rewritten in terms of the spectraallowing for a deeper connection with the eigenvalues of thegraphs adjacency matrix
I Calculating the expected number of steps before two RandomWalks starting at different vertices collide
I Applications from electrical networks and statics can beapplied to obtain results about Random Walks
I Applications of the mixing rate
Robert Lazar (ISU) 27 of 28
Thank You
Robert Lazar (ISU) 28 of 28