Aglect

Analysis on Graphs
Alexander Grigoryan Lecture Notes University of Bielefeld, WS 2009/10
Contents
1 The 1.1 1.2 1.3 1.4 1.5 Laplace operator on The notion of a graph Cayley graphs . . . . . Random walks . . . . . The Laplace operator . The Dirichlet problem graphs . . . . . . . . . . . . . . . . . . . . . . . . . 5 . 5 . 9 . 13 . 24 . 28 . . . . . . 33 33 34 39 44 51 54
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
2 Spectral properties of the Laplace operator 2.1 Greens formula . . . . . . . . . . . . . . . . 2.2 Eigenvalues of the Laplace operator . . . . . 2.3 Convergence to equilibrium . . . . . . . . . 2.4 More about the eigenvalues . . . . . . . . . 2.5 Products of graphs . . . . . . . . . . . . . . 2.6 Eigenvalues and mixing times in Zn . . . . . m
3 Geometric bounds for the eigenvalues 59 3.1 Cheegers inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 3.2 Estimating 1 from below via diameter . . . . . . . . . . . . . . . . . 67 3.3 Expansion rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 4 Eigenvalues on innite graphs 4.1 Dirichlet Laplace operator . . . . . . . . . 4.2 Cheegers inequality . . . . . . . . . . . . . 4.3 Isoperimetric inequalities . . . . . . . . . . 4.4 Solving the Dirichlet problem by iterations 4.5 Isoperimetric inequalities on Cayley graphs 5 The 5.1 5.2 5.3 5.4 5.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 79 82 84 85 88 93 94 103 106 112 117
heat kernel and the type problem On-diagonal upper estimates of the heat kernel . On-diagonal lower bounds . . . . . . . . . . . . The type problem . . . . . . . . . . . . . . . . . Volume tests for recurrence . . . . . . . . . . . Isoperimetric tests for transience . . . . . . . . .
CONTENTS
Chapter 1 The Laplace operator on graphs

1.1 The notion of a graph
A.Grigoryan
A graph is a couple (V, E) where V is a set of vertices, that is, an arbitrary set, Lecture 1 whose elements are called vertices, and E is a set of edges, that is, E consists of 13.10.09 some couples (x, y) where x, y 2 V . We write x y (x is connected to y, or x is joint to y, or x is adjacent to y, or x is a neighbor of y ) if (x, y) 2 E. Graphs are normally represented graphically as a set of points on a plane, and if x y then one connects the corresponding points on the plane by a line. There are two versions of the denition of edges: 1. The couples (x, y) are ordered, that is, (x, y) and (y, x) are considered as dierent (unless x = y). In this case, the graph is called directed or oriented. 2. The couples (x, y) are unordered, that is, (x, y) = (y, x). In this case, x y is equivalent to y x, and the graph is called undirected or unoriented. Unless otherwise specied, all graphs will be undirected. The edge (x, y) will be normally denoted by xy, and x, y are called the endpoints of this edge. The edge xx with the same endpoints (should it exist) is called a loop. A graph is called simple if it has no loops. A graph is called locally nite if each vertex has a nite number of edges. For each point x, dene its degree deg (x) = # fy 2 V : x yg , that is, deg (x) is the number of the edges with endpoint x. A graph is called nite if the number of vertices is nite. Of course, a nite graph is locally nite. We start with a simple observation. Lemma 1.1 (Double counting of edges) On any simple nite graph (V, E), the following identity holds: X deg (x) = 2#E.
xV
CHAPTER 1. THE LAPLACE OPERATOR ON GRAPHS
Proof. Let n = #V and let us enumerate all vertices as 1, 2, ..., n. Consider the adjacency matrix A = (aij )n i,j=1 of the graph that is dened as follows: aij = 1, i j 0, i 6 j.
This matrix is symmetric and the sum of the entries in the row i (and in the column i) is equal to deg (i) so that n ! n n n X X X X deg (i) = aij = aij .
i=1 i=1 j=1 i,j=1
The entries aij in the last summation are 0 and 1, and aij = 1 if and only if (i, j) is an edge. In this case i 6= j and (j, i) is also an edge. Therefore, each edge ij contributes to this sum the value 1 twice: as (i, j) and as (j, i). Hence,
n X
aij = 2#E,
i,j=1
which nishes the proof. Example. Consider some examples of graphs. 1. A complete graph Kn . The set of vertices is V = f1, 2, ..., ng, and the edges are dened as follows: i j for any two distinct i, j 2 V . That is, any two distinct points in V are connected. Hence, the number of edges in Kn is 1 1X deg (i) = n (n 1) . 2 i=1 2
n
2. A complete bipartite graph Kn,m . The set of vertices is V = f1, .., n, n + 1, ..., n + mg, and the edges are dened as follows: i j if and only if either i < n and j n or i n and j < n. That is, the set of vertices is split into two groups: S1 = f1, ..., ng and S2 = fn + 1, ..., mg, and the vertices are connected if and only if they belong to the dierent groups. The number of edges in Kn,m is equal to nm. 3. A lattice graph Z. The set of vertices V consists of all integers, and the integers x,y are connected if and only if jx yj = 1. 4. A lattice graph Zn . The set of vertices consists of all n-tuples (x1 , ..., xn ) where xi are integers, and (x1 , ..., xn ) (y1 , ..., yn ) if and only if
n X i=1
jxi yi j = 1.
That is, xi is dierent from yi for exactly one value of the index i, and jxi yi j = 1 for this value of i.
1.1. THE NOTION OF A GRAPH
Denition. A weighted graph is a couple (, ) where = (V, E) is a graph and xy is a non-negative function on V V such that 1. xy = yx ; 2. xy > 0 if and only if x y. Alternatively, xy can be considered as a positive function on the set E of edges, that is extended to be 0 on non-edge pairs (x, y). The weight is called simple if xy = 1 for all edges x y. Any weight xy gives rise to a function on vertices as follows: X (x) = xy . (1.1)
y,yx
Then (x) is called the weight of a vertex x. For example, if the weight xy is simple then (x) = deg (x). The following lemma extends Lemma 1.1.
Lemma 1.2 On any simple nite weighted graph (, ), X X (x) = 2 .

xV E
Proof. Rewrite (1.1) in the form (x) = X

yV
xy
where the summation is extended to all y 2 V . This does not change the sum in (1.1) because we add only non-edges (x, y) where xy = 0. Therefore, we obtain X XX X X X (x) = xy = xy = xy = 2 .
xV xV yV x,yV x,y:xy E
Denition. A nite sequence fxk gn of vertices on a graph is called a path if k=0 xk xk+1 for all k = 0, 1, ..., n 1. The number n of edges in the path is referred to as the length of the path. Denition. A graph (V, E) is called connected if, for any two vertices x, y 2 V , there is a path connecting x and y, that is, a path fxk gn such that x0 = x and k=0 xn = y. If (V, E) is connected then dene the graph distance d (x, y) between any two distinct vertices x, y as follows: if x 6= y then d (x, y) is the minimal length of a path that connects x and y, and if x = y then d (x, y) = 0. The connectedness here is needed to ensure that d (x, y) < 1 for any two points. Lemma 1.3 On any connected graph, the graph distance is a metric, so that (V, d) is a metric space. Proof. We need to check the following axioms of a metric.
CHAPTER 1. THE LAPLACE OPERATOR ON GRAPHS 1. Positivity: 0 d (x, y) < 1, and d (x, y) = 0 if and only if x = y. 2. Symmetry: d (x, y) = d (y, x) . 3. The triangle inequality: d (x, y) d (x, z) + d (z, y) .
The rst two properties are obvious for the graph distance. To prove the triangle inequality, choose a shortest path fxk gn connecting x and z, a shortest path k=0 fyk gm connecting y and z, so that k=0 d (x, z) = n and d (z, y) = m. Then the sequence x1 , ..., xn1 , z, y1 , ..., ym is a path connecting x and y, that has the length n + m, which implies that d (x, y) n + m = d (x, z) + d (z, y) .
Lemma 1.4 If (V, E) is a connected locally nite graph then the set of vertices V is either nite or countable. Proof. Fix a reference point x 2 V and consider the set Bn : fy 2 V : d (x, y) ng , that is a ball with respect to the distance d. Let us prove by induction in n that #Bn < 1. Inductive step for n = 0 is trivial because B0 = fxg. Inductive step: assuming that Bn is nite, let us prove that Bn+1 is nite. It suces to prove that Bn+1 n Bn is nite. For any vertex y 2 Bn+1 n Bn , we have d (x, y) = n + 1 so that there is a path fxk gn+1 from x to y of length n + 1. Consider the vertex z = xn . Clearly, the k=0 path fxk gn connects x and z and has the length n, which implies that d (x, z) n k=0 and, hence, z 2 Bn . On the other hand, we have by construction z y. Hence, we have shown that every vertex y 2 Bn+1 n Bn is connected to one of the vertices in Bn . However, the number of vertices in Bn is nite, and each of them has nitely many neighbors. Therefore, the total number of the neighbors of Bn is nite, which implies # (Bn+1 n Bn ) < 1 and #Bn+1 < 1. S Finally, observe that V = Bn because for any y 2 V we have d (x, y) < 1 n=1 so that y belongs to some Bn . Then V is either nite or countable as a countable union of nite sets.
1.2. CAYLEY GRAPHS
9
A.Grigoryan Lecture 2 19.10.09
1.2
Cayley graphs
Here we discuss a large class of graphs that originate from groups. Recall that a group (G, ) is a set G equipped with a binary operation that satises the following properties: 1. for all x, y 2 G, x y is an element of G; 2. associative law: x (y z) = (x y) y; 3. there exists a neutral element e such that x e = e x = x for all x 2 G; 4. there exists the inverse x1 element for any x 2 G, such that x x1 = x1 x = e. If in addition the operation is commutative, that is, x y = y x then the group G is called abelian or commutative. In the case of abelian group, one uses the additive notation. Namely, the group operation is denoted + instead of , the neutral element is denoted by 0 instead of e, and the inverse element is denoted by x rather than x1 . Example. 1. Consider the set Z of all integers with the operation +. Then (Z, +) is an abelian group where the neutral element is the number 0 and the inverse of x is the negative of x. 2. Fix an integer q 2 and consider the set Zq of all residues (Restklassen) modulo q, with the operation +. In other words, the elements of Zq are the equivalence classes of integers modulo q. Namely, one says that two integers x, y are congruent modulo q and writes x = y mod q if x y is divisible by q. This relation is an equivalence relation and gives rise to q equivalence classes, that are called the residues and are denoted by 0, 1, ..., q 1 as integers, so that the integer k belongs to the residue k. The addition in Zq is inherited from Z as follows: x + y = z in Zq , x + y = z mod q in Z. Then (Zq , +) is an abelian group, the neutral element is 0, and the inverse of x is q x (except for x = 0). For example, consider Z2 = f0, 1g. Apart from trivial sums x + 0 = x, we have the following rules in this group: 1 + 1 = 0 and 1 = 1. If Z3 = f0, 1, 2g, we have 1 + 1 = 2, 1 + 2 = 0, 2 + 2 = 1 1 = 2, 2 = 1.
10
Given two groups, say (A, +) , (B, +), one can consider a direct product of them: the group (A B, +) that consists of pairs (a, b) where a 2 A and b 2 B with the operation (a, b) + (a0 , b0 ) = (a + a0 , b + b0 ) The neutral element of A B is (0A , 0B ), and the inverse to (a, b) is (a, b). More generally, given n groups (Ak , +) where k = 1, ..., n, dene their direct product (A1 A2 ... An , +) as the set of all sequences (ak )n where ak 2 Ak , with the operation k=1 (a1 , ..., an ) + (a01 , ..., a0n ) = (a1 + a01 , ..., an + a0n ) . The neutral element is (0A1 , ..., 0An ) and the inverse is (a1 , ..., an ) = (a1 , ..., an ) . If the groups are abelian then their product is also abelian. Example. 1. The group Zn is dened as the direct product Z Z ... Z of n {z } |
n
copies of the group Z. 2. The group (Z2 Z3 , +) consists of pairs (a, b) where a is a residue mod 2 and b is a residue mod 3. For example, we have the following sum in this group: (1, 1) + (1, 2) = (0, 0) , whence it follows that (1, 1) = (1, 2) .
What is the relation of the groups to graphs? Groups give rise to a class of graphs that are called Cayley graphs. Let (G, ) be a group and S be a subset of G / with the property that if x 2 S then x1 2 S and that e 2 S. Such a set S will be called symmetric. A group G and a symmetric set S G determine a graph (V, E) as follows: the set V of vertices coincides with G, and the set E of edges is dened as follows: x y , x1 y 2 S. (1.2)
Note that the relation x y is symmetric in x, y, that is, x y implies y x, because 1 2 S. y 1 x = x1 y
Hence, the graph (V, E) is undirected. The fact that e 2 S implies that x 6 x / / because x1 x = e 2 S. Hence, the graph (V, E) contains no loops.
Denition. The graph (V, E) dened as above is denoted by (G, S) and is called the Cayley graph of the group G with the edge generating set S. There may be many dierent Cayley graphs of the same group since they depend also on the choice of S. It follows from the construction that deg (x) = #S for any x 2 V. In particular, if S is nite then the graph (V, E) is locally nite. In what
1.2. CAYLEY GRAPHS
11
follows, we will consider only locally nite Cayley graphs and always assume that they are equipped with a simple weight. If the group operation is + then (1.2) becomes x y , y x 2 S , x y 2 S. In this case, the symmetry of S means that 0 2 S and if x 2 S then also x 2 S. / Example. 1. G = Z with + and S = f1, 1g . Then x y if x y = 1 or x y = 1, that is, if x and y are neighbors on the real axis: ... ... If S = f1, 2g then x y if jx yj = 1 or jx yj = 2 so that we obtain a dierent graph. 2. Let G = Zn with +. Let S consist of points (x1 , ..., xn ) 2 Zn such that exactly one of xi is equal to 1 and the others are 0; that is ( ) n X S = (x1 , ..., xn ) 2 Zm : jxi j = 1.
i=1
For example, in the case n = 2 we have S = f(1, 0) , (1, 0) , (0, 1) , (0, 1)g . The connection x y means that x y has exactly one coordinate 1, and all others are 0; equivalently, this means that
n X i=1
jxi yi j = 1.
Hence, the Cayley graph of (Zn , S) is exactly the standard lattice graph Zn . In the case n = 2, it looks as follows: j j j j j j j j j j j j j j j j j j j j j j j j j j j j j j
Consider now another edge generating set S on Z2 with two more elements: S = f(1, 0) , (1, 0) , (0, 1) , (0, 1) , (1, 1) , (1, 1)g .
12
The corresponding graph (Z2 , S) is shown here: j j j j j j j j j j j j j j j j j j j j j j j j j j j j j j
3. Let G = Z2 = f0, 1g. The only possibility for S is S = f1g (note that 1 = 1). The graph (Z2 , S) coincides with K2 and is shown here: 4. Let G = Zq where q > 2, and S = f1g. That is, each residue k = 0, 1, ..., q1 has two neighbors: k 1 and k + 1. For example, 0 has the neighbors 1 and q 1. The graph (Zq , S) is called the q-cycle and is denoted by Cq . Here are the graphs C3 and C4 : j , C4 = j j C3 =
5. Consider Zq with the symmetric set S = Zq n f0g. That is, every two distinct elements x, y 2 Zq are connected by an edge. Hence, the resulting Cayley graph is the complete graph Kq . 6. Let G = Zn := Z2 Z2 ... Z2 , that consists of n-tuples (x1 , ..., xn ) of 2 | {z }
n
residues mod 2, that is, each xi is 0 or 1. Let S consist of all elements (x1 , ..., xn ) . such that exactly one xi is equal to 1 and all others are 0. Then the graph (Zn , S) is 2 called the n-dimensional binary cube and is denoted by f0, 1gn , analogously to the geometric n-dimensional cube [0, 1]n . Clearly, f0, 1g1 = K2 and f0, 1g2 = C4 . The graph f0, 1g3 is shown here in two ways: j j j j = f0, 1g3 = j j j j j j j j j j j j j j j j
7. Let G = Zq Z2 . Then G consists of pairs (x, y) where x 2 Zq and y 2 Z2 . Then G can be split into two disjoint subsets G0 = Zq f0g = f(x, 0) : x 2 Zq g G1 = Zq f1g = f(x, 1) : x 2 Zq g ,
1.3. RANDOM WALKS each having q elements. Set S = G1 . Then (x, a) (y, b) , a b = 1 , a 6= b.
13
In other words, (x, a) (y, b) if and only if the points (x, a) and (y, b) belong to dierent subsets G0 , G1 . Hence, the graph (Zq Z2 , S) coincides with the complete bipartite graph Kq,q with the partition G0 , G1 . Denition. A graph (V, E) is called D-regular, if all vertices x 2 V have the same degree D (that is, each vertex has D edges). A graph is called regular if it is D-regular for some D. Of course, there are plenty of examples of non-regular graphs. Clearly, all Cayley graphs are regular. All regular graphs that we have discussed above, were Cayley graphs. However, there regular graphs that are not Cayley graphs (cf. Exercise 8).
1.3
Random walks
Consider a classical problem from Probability theory. Let fxk g be a sequence of k=1 independent random variables, taking values 1 and 1 with probabilities 1/2 each; that is, P (xk = 1) = P (xk = 1) = 1/2. Consider the sum Xn = x1 + ... + xn and ask, what is a likely behavior of Xn for large n? Historically, this type of problem came from the game theory (and the gambling practice): at each integer value of time, a player either wins 1 with probability 1/2 or loses 1 with the same probability. The games at dierent times are independent. Then Xn represents the gain at time n if Xn > 0, and the loss at time n is Xn < 0. Of course, the mean value of Xn , that is, the expectation, is 0 because E (xk ) = 1 (1 1) = 0 and 2 n X E (xk ) = 0. E (Xn ) =
k=1
The games with this property are called fair games or martingales. However, the deviation of Xn from the average value 0 can be still signicant in any particular game. We will adopt a geometric point of view on Xn as follows. Note that Xn 2 Z and Xn is dened inductively as follows: X0 = 0 and Xn+1 Xn is equal to 1 or 1 with equal probabilities 1/2. Hence, we can consider Xn as a position on Z of a walker that jumps at any time n from its current position to a neighboring integer, either right or left, with equal probabilities 1/2, and independently of the previous movements. Such a random process is called a random walk. Note that this random walk is related to the graph structure of Z: namely, a walker moves at each step along the edges of this graph. Hence, Xn can be regarded as a random walk on the graph Z. Similarly, one can dene a random walk on ZN : at any time n = 0, 1, 2, ..., let Xn be the position of a walker in ZN . It starts at time 0 at the origin, and at time
14
n + 1 it moves with equal probability 1/ (2N) by one of the vectors e1 , ..., eN , where e1 , ..., en is the canonical basis in Rn . That is, X0 = 0 and P (Xn+1 Xn = ek ) = 1 . 2N
We always assume that the random walk in question has the Markov property: the choice of the move at any time n is independent of the previous movement. The following picture of the trace of a random walk on Z2 was copied from Wikipedia:
Figure 1.1:
More generally, one can dene a random walk on any locally nite graph (V, E) . Namely, imagine a walker that at any time n = 0, 1, 2... has a random position Xn at one of the vertices of V that is dened as follows: X0 = x0 is a given vertex, and Xn+1 is obtained from Xn by moving with equal probabilities along one of the edges of Xn , that is, 1/ deg (x) , y x, P (Xn+1 = y j Xn = x) = (1.3) 0, y 6 x. The random walk fXn g dened in this way, is called a simple random walk on (V, E). The adjective simple refers to the fact that the walker moves to the neighboring vertices with equal probabilities. A simple random walk is a particular case of a Markov chain. Given a nite or countable set V (that is called a state space), a Markov kernel on V is any function P (x, y) : V V ! [0, +1) with the property that X P (x, y) = 1 8x 2 V. (1.4)
yV
1.3. RANDOM WALKS
15
If V is countable then the summation here is understood as a series. Any Markov kernel denes a Markov chain fXn g as a sequence of random n=0 variables with values in V such that the following identity holds P (Xn+1 = y j Xn = x) = P (x, y) , (1.5)
and that the behavior of the process at any time n onwards is independent of the past. The latter requirement is called the Markov property and it will be considered in details below. Observe that the rule (1.3) dening the random walk on a graph (V, E) can be also written in the form (1.5) where 1 , y x, deg(x) P (x, y) = (1.6) 0, y 6 x. The condition (1.4) is obviously satised because X
yV
P (x, y) =
X
yx
P (x, y) =
X
yx
1 = 1. deg (x)
Hence, the random walk on a graph is a particular case of a Markov chain, with a specic Markov kernel (1.6). Let us discuss the Markov property. The exact meaning of it is given by the following identity: P (X1 = x1 , X2 = x2 , ..., Xn+1 = xn+1 j X0 = x) ) = P (X1 = x1 , X2 = x2 , ..., Xn = xn j X0 = x) P (Xn+1 = xn+1 j Xn = xn(1.7) that postulates the independence of the jump from xn to xn+1 from the previous path. Using (1.5) and (1.7), one obtains by induction that P (X1 = x1 , X2 = x2 , ..., Xn = xn j X0 = x) = P (x, x1 ) P (x1 , x2 ) ...P (xn1 , xn ) . (1.8) Obviously, (1.8) implies back (1.7). In fact, (1.8) can be used to actually construct the Markov chain. Indeed, it is not obvious that there exists a sequence of random variables satisfying (1.5) and (1.7). Proposition 1.5 The Markov chain exists for any Markov kernel. Proof. This is the only statement in this course that requires a substantial use of the foundations of Probability Theory. Indeed, it is about construction of a probability space (, P) and dening a sequence fXn g of random variables n=0 satisfying the required conditions. The set will be taken to be the set V of all sequences fxk g of points of V , that represent the nal outcome of the process. k=1 In order to construct a probability measure P on , we rst construct a probability measure P(n) of the set of nite sequences fxk gn . Note that the set of sequences k=1 of n points of V is nothing other than the product V n = V {z V . Hence, our ... | }
n
strategy is as follows: rst construct a probability measure P(1) on V , then P(n) on
16
V n , and then extend it to a measure P on V . In fact, we will construct a family of probability measures Px indexed by a point x 2 V , so that Px is associated with a Markov chain starting from the point X0 = x. Fix a point x 2 V and observe that P (x, ) determines a probability measure (1) Px on V as follows: for any subset A V , set X P (x, y) . P(1) (A) = x
yA
Clearly, Px is -additive, that is, ! G X (1) Px Ak = Px (Ak ) ,

k=1 k=1
and Px (V ) = 1 by (1.4). (n) Next, dene a probability measure Px on the product V n = V {z V as ... } | follows. Firstly, dene the measure of any point (x1 , ..., xn ) 2 V n by1 P(n) (x1 , ..., xn ) = P (x, x1 ) P (x1 , x2 ) ...P (xn1 , xn ) , x and then extend it to all sets A V n by X P(n) (A) = x
n
(1)
(1.9)
P(n) (x1 , ..., xn ) . x

(n)
(x1 ,...,xn )A
Let us verify that it is indeed a probability measure, that is, Px (V n ) = 1. The inductive basis was proved above, let us make the inductive step from n to n + 1 X (n) Px (x1 , ..., xn+1 )
x1 ,...,xn+1 V
= = =
x1 ,...,xn+1 V
x1 ,...,xn V xn+1 V
P (x, x1 ) P (x1 , x2 ) ...P (xn1 , xn ) P (xn , xn+1 ) X P (x, x1 ) P (x1 , x2 ) ...P (xn1 , xn ) P (xn , xn+1 ) X P (xn , xn+1 )
= 1 1 = 1,
x1 ,...,xn V
P (x, x1 ) P (x1 , x2 ) ...P (xn1 , xn )
xn+1 V
following sense. Fix two positive integers n < m. Then every point (x1 , ..., xn ) 2 V n
Note that measure Px have been
1 (n)
where we have use (1.4) and the inductive hypothesis. n o (n) The sequence of measures Px constructed above is consistence in the
n=1
is not a product measure of n copies of Px P (x; x1 ) P (x; x2 ) :::P (x; xn ) :
(1)
since the latter would
1.3. RANDOM WALKS
17
can be regarded as a subset of V m that consists of all sequence where the rst n terms are exactly x1 , .., xn , and the rest terms are arbitrary. Then we have
(n) (m) Px (x1 , ..., xn ) = Px (x1 , ..., xn ) ,
(1.10)
that is proved as follows: X (m) Px (x1 , ..., xn ) = =
P(m) (x1 , ..., xn , xn+1 , ..., xm ) x P (x, x1 ) P (x1 , x2 ) ...P (xn1 , xn ) P (xn , xn+1 ) ...P (xm1 , xm ) X P (xn , xn+1 ) ...P (xm1 , xm )
xn+1 ,...,xm V
xn+1 ,...,xm V
= P (x, x1 ) P (x1 , x2 ) ...P (xn1 , xn ) = = P(n) x P(n) x

(nm) (x1 , ..., xn ) Pxn
(x1 , ..., xn ) .
nm V
xn+1 ,...,xm V
In the same way, any subset A V n admits the cylindrical extension A0 to a subset of V m as follows: a sequence (x1 , ..., xm ) belongs to A0 if (x1 , ..., xn ) 2 A. It follows from (1.10) that (m) P(n) (A) = Px (A0 ) . (1.11) x This is a Kolmogorov consistency condition, that allows to extend a sequence of (n) measures Px on V n to a measure on V . Consider rst cylindrical subsets of V , that is, sets of the form where A is a subset of V n , and set A0 = ffxk g : (x1 , ..., xn ) 2 Ag k=1 Px (A0 ) = P(n) (A) . x (1.12)
(1.13)
Due to the consistency condition (1.11), this denition does not depend on the choice of n. Kolmogorovs extension theorem says that the functional Px dened in this way on cylindrical subsets of V , extends uniquely to a probability measure on the minimal -algebra containing all cylindrical sets. Now we dene the probability space as V endowed with the family fPx g of probability measures. The random variable Xn is a function on with values in V that is dened by Xn (fxk g ) = xn . k=1 Then (1.9) can be rewritten in the form Px (X1 = x1 , ..., Xn = xn ) = P (x, x1 ) P (x1 , x2 ) ...P (xn1 , xn ) . (1.14)
The identity (1.14) together with (1.4) are the only properties of Markov chains that we need here and in what follows. Let us use (1.14) to prove that the sequence fXn g is indeed a Markov chain with the Markov kernel P (x, y). We need to verify (1.5) and (1.8). The latter is obviously equivalent to (1.14). To prove the former, write X Px (Xn = y) = Px (X1 = x1 , ..., Xn1 = xn1 , Xn = y)
x1 ,...,xn1 V
x1 ,...,xn1 V
P (x, x1 ) P (x1 , x2 ) ...P (xn1 , y)
18 and
Px (Xn = y, Xn+1 = z) = = whence
x1 ,...,xn1 V
x1 ,...,xn1 V
Px (X1 = x1 , ..., Xn1 = xn1 , Xn = y, Xn+1 = z) P (x, x1 ) P (x1 , x2 ) ...P (xn1 , y) P (y, z)
Px (Xn+1 = z j Xn = y) =
Px (Xn = y, Xn+1 = z) = P (y, z) , Px (Xn = y)
which is equivalent to (1.5). Given a Markov chain fXn g with a Markov kernel P (x, y), note that by (1.14) P (x, y) = Px (X1 = y) so that P (x, ) is the distribution of X1 . Denote by Pn (x, ) the distribution of Xn , that is, Pn (x, y) = Px (Xn = y) . The function Pn (x, y) is called the transition function or the transition probability of the Markov chain. Indeed, it fully describes what happens to random walk at time n. For a xed n, the function Pn (x, y) is also called the n-step transition function. It is easy to deduce a recurrence relation between Pn and Pn+1 . Proposition 1.6 For any Markov chain, we have X Pn+1 (x, y) = Pn (x, z) P (z, y) .
zV
(1.15)
Proof. By (1.14), we have Pn (x, z) = Px (Xn = z) = = X X Px (X1 = x1 , ..., Xn1 = xn1 , Xn = z)
x1 ,...,xn1 V
P (x, x1 ) P (x1 , x2 ) ...P (xn1 , z) .
x1 ,...,xn1 V
Applying the same argument to Pn+1 (x, y), we obtain X P (x, x1 ) P (x1 , x2 ) ...P (xn1 , xn ) P (xn , y) Pn+1 (x, y) =
x1 ,...,xn V
= =
X
zV zV
x1 ,...,xn1 V
Pn (x, z) P (z, y) ,
P (x, x1 ) P (x1 , x2 ) ...P (xn1 , z)A P (z, y)
which nishes the proof.
1.3. RANDOM WALKS Corollary 1.7 For any xed n, Pn (x, y) is also a Markov kernel on V . Proof. We need only to verify that X Pn (x, y) = 1.
yV
19
For n = 1 this is given, the inductive step from n to n + 1 follows from (1.15): XX X Pn+1 (x, y) = Pn (x, z) P (z, y)
yV
= =
XX X
zV zV yV
yV zV
Pn (x, z) P (z, y)
Pn (x, z) = 1.
Corollary 1.8 We have, for all positive integers n, k, X Pn (x, z) Pk (z, y) . Pn+k (x, y) =
zV
(1.16)
Proof. Induction in k. The case k = 1 is covered by (1.15). The inductive step from k to k + 1: X Pn+k (x, w) P (w, y) Pn+(k+1) (x, y) =
wV
wV zV
= =
XX X
zV zV wV
XX
Pn (x, z) Pk (z, w) P (w, y) Pn (x, z) Pk (z, w) P (w, y)
Pn (x, z) Pk+1 (z, y) .
Now we impose one restriction on a Markov chain. Denition. A Markov kernel P (x, y) is called reversible if there exists a positive function (x) on the state space V , such that P (x, y) (x) = P (y, x) (y) . (1.17)
A Markov chain is called reversible if its Markov kernel is reversible. It follows by induction from (1.17) and (1.15) that Pn (x, y) is also a reversible Markov kernel. The condition (1.17) means that the function xy := P (x, y) (x)
20
is symmetric in x, y. For example, this is the case when P (x, y) is symmetric in x, y; then we just take (x) = 1. However, the reversibility condition can be satised for non-symmetric Markov kernel as well. For example, in the case of a simple random walk on a graph (V, E), we have by (1.6) 1, x y . xy = P (x, y) deg (x) = 0, x 6 y which is symmetric. Hence, a simple random walk is a reversible Markov chain. Any reversible Markov chain on V gives rise to a graph structure on V as follows. Dene the set E of edges on V by the condition x y , xy > 0. Then xy can be considered as a weight on (V, E) (cf. Section 1.1). Note that the function (x) can be recovered from xy by the identity X xy = X
yV
P (x, y) (x) = (x) ,
y,yx
which matches (1.1). Let = (V, E) be a graph. Recall that a non-negative function xy on V V is called a weight, if xy = yx and xy > 0 if and only if x y. A couple (V, ) (or (, )) is called a weighted graph. Note that the information about the edges is contained in the weight so that the set E of edges is omitted in the notation (V, ). As we have seen above, any reversible Markov kernel on V determines a weighted graph (V, ). Conversely, a weighted graph (V, ) determines a reversible Markov kernel on V provided the set V is nite or countable and X 0< xy < 1 for all x 2 V. (1.18)
yV
For example, the positivity condition in (1.18) holds if the graph (V, E) that is determined by the weight xy , has no isolated vertices, that is, the vertices without neighbors, and the niteness condition holds if the graph (V, E) is locally nite, so that the summation in (1.18) has nitely many positive terms. The full condition (1.18) is satised for locally nite graphs without isolated vertices. If (1.18) holds, then the weight on vertices X (x) = xy
yV
is nite and positive for all x, and we can set P (x, y) = xy (x) (1.19)
so that the reversibility condition (1.17) is obviously satised. In this context, a reversible Markov chain is also referred to as a random walk on a weighted graph.
1.3. RANDOM WALKS
21
From now on, we stay in the following setting: we have a weighted graph (V, ) satisfying (1.18), the associated reversible Markov kernel P (x, y), and the corresponding random walk (= Markov chain) fXn g. Fix a point x0 2 V and consider the functions vn (x) = Px0 (Xn = x) = Pn (x0 , x) and un (x) = Px (Xn = x0 ) = Pn (x, x0 ) The function vn (x) is the distribution of Xn at time n 1. By Corollary 1.7, we P have xV vn (x) = 1. Function un (x) is somewhat more convenient to be dealt with. The function vn and un are related follows: vn (x) (x0 ) = Pn (x0 , x) (x0 ) = Pn (x, x0 ) (x) = un (x) (x) , where we have used the reversibility of Pn . Hence, we have the identity vn (x) = un (x) (x) . (x0 ) (1.20)
Extend un and vn to n = 0 by setting u0 = v0 = 1{x0 } , where 1A denotes the indicator function of a set A V , that is, the function that has value 1 at any point of A and value 0 outside A. Corollary 1.9 For any reversible Markov chain, we have, for all x 2 V and n = 0, 1, 2, ..., X 1 vn+1 (x) = vn (y) xy (1.21) (y) y (a forward equation) and un+1 (x) = (a backward equation). Proof. If n = 0 then (1.21) becomes P (x0 , x) = 1 (x0 ) xx0 1 X un (y) xy (x) y (1.22)
which is a dening identity for P . For n 1, we obtain, using (1.19) and (1.15), X
y
X 1 vn (y) yx = Pn (x0 , y) P (y, x) (y) y
= Pn+1 (x0 , x) = vn+1 (x) ,
which proves (1.21). Substituting (1.20) into (1.21), we obtain (1.22).
22
In particular, for a simple random walk we have xy = 1 for x y and (x) = deg (x) so that we obtain the following identities: X
yx
vn+1 (x) =
1 vn (y) . deg (y)
un+1 (x) =
1 X un (y) . deg (x) yx
The last identity means that un+1 (x) is the mean-value of un (y) taken at the points y x. Note that in the case of a regular graph, when deg (x) const, we have un vn by (1.20). Example. Let us compute the function un (x) on the lattice graph Z. Since Z is regular, un = vn so that un (x) the distribution of Xn at time n provided X0 = 0.We evaluate inductively un using the initial condition u0 = 1{0} and the recurrence relation: un+1 (x) = 1 (u (x + 1) + u (x 1)) . 2 (1.23)
The computation of un (x) for n = 1, 2, 3, 4 and jxj 4 are shown here (all empty elds are 0): x2Z n=0 n=1 n=2 n=3 n=4
1 16 1 8 1 4 1 4 3 8 3 8
-4
-3
-2
-1
1 2
0 1
1
1 2
1 2 3 8
1 4 1 8 1 4 1 16
... ... n = 10 0.117
...
...
0.205
...
...
0.246
... ...
0.205
... ...
0.117
One can observe (and prove) that un (x) ! 0 as n ! 1. Example. Consider also computation of un (x) on the graph C3 = (Z3 , f1g). The formula (1.23) is still true provided that one understands x as a residue mod 3. We
1.3. RANDOM WALKS have then the following computations for n = 1, ..., 6: x 2 Z3 n=0 n=1 n=2 n=3 n=4 n=5 n=6 0 0
1 2 1 4 3 8 5 16 11 32 21 64
23
1 1 0
1 2 1 4 3 8 5 16 11 32
2 0
1 2 1 4 3 8 5 16 11 32 21 64
Here one can observe that the function un (x) converges to a constant function 1/3, and later we will prove this. Hence, for large n, the probability that Xn visits a given point is nearly 1/3, which should be expected. The following table contains a similar computation of un on C5 = (Z5 , f1g) x 2 Z5 n=0 n=1 n=2 n=3 n=4 ... n = 25 0 0 0
1 4 1 8 1 4
1 0
1 2
2 1 0
1 2
3 0
1 2
4 0 0
1 4 1 8 1 4
0
3 8 1 16
0
3 8 1 16
0
3 8
...
0.199
...
0.202
...
0.198
...
0.202
...
0.199
Here un (x) approaches to 1 but the convergence is much more slower than in the 5 case of C3 . Consider one more example: a complete graph K5 . In this case, function un (x) satises the identity 1X un+1 (x) = un (y) . 4 y6=x The computation shows the following values of un (x): x 2 K5 n=0 n=1 n=2 n=3 n=4 0 0
1 4 3 16 13 64 0.199
1 0
1 4 3 16 13 64 0.199
2 1 0
1 4 3 16 0.203
3 0
1 4 3 16 13 64 0.199
4 0
1 4 3 16 13 64 0.199
24
This time the convergence to the constant 1/5 occurs much faster, than in the previous example, although C5 and K5 has the same number of vertices. The extra edges in K5 allows a quicker mixing than in the case of C5 . As we will see, for nite graphs it is typically the case that the transition function un (x) converges to a constant as n ! 1. For the function vn this means that vn (x) = un (x) (x) ! c (x) as n ! 1 0 (x)
for some constant c. The constant c is determined by the requirement that c (x) is a probability measure on V , that is, from the identity X c (x) = 1.
xV
Hence, c (x) is asymptotically the distribution of Xn as n ! 1. . The function c (x) on V is called the stationary measure or the equilibrium measure of the Markov chain. One of the problems for nite graphs that will be discussed in this course, is the rate of convergence of vn (x) to the equilibrium measure. The point is that Xn can be considered for large n as a random variable with the distribution function c (x) so that we obtain a natural generator of a random variable with a prescribed law. However, in order to be able to use this, one should know for which n the distribution of Xn is close enough to the equilibrium measure. The value of n, for which this is the case, is called the mixing time. For innite graphs the transition functions un (x) and vn (x) typically converge to 0 as n ! 1, and an interesting question is to determine the rate of convergence to 0. For example, we will show that, for a simple random walk in ZN , vn (x) ' nN/2 as n ! 1. The distribution function vn (x) is very sensitive to the geometry of the underlying graph. Another interesting question that arises on innite graphs, is to distinguish the following two alternatives in the behavior of a random walk Xn on a graph: 1. Xn returns innitely often to a given point x0 with probability 1, 2. Xn visits x0 nitely many times and then never comes back, also with probability 1. In the rst case, the random walk is called recurrent, and in the second case transient. By a theorem of Polya, a simple random walk in ZN is recurrent if and only if N 2.
1.4
The Laplace operator

f 0 (x) = lim f (x + h) f (x) h0 h
Let f (x) be a function on R. Recall that
1.4. THE LAPLACE OPERATOR so that f 0 (x) f (x) f (x h) f (x + h) f (x) h h
25
for small h. The operators f (x+h)f (x) and f (x)f (xh) are called the dierence operh h ators and can be considered as numerical approximations of the derivative. What would be the approximation of the second derivative?
f (x+h)f (x) f (x)f (xh) f 0 (x + h) f 0 (x) h h f (x) h h f (x + h) 2f (x) + f (x h) = h2 2 f (x + h) + f (x h) f (x) . = 2 h 2 00
Hence, f 00 is determined by the average value of f at neighboring points x + h and x h, minus f (x) . For functions f (x, y) on R2 , one can develop similarly numerical approximations 2 2 for second order partial derivatives f and f , and then for the Laplace operator x2 y2 f = 2f 2f + . x2 y 2
More generally, for a function f of n variables x1 , ..., xn , the Laplace operator is dened by n X 2f f = . (1.24) x2 k k=1
This operator in the three-dimensional space was discovered by Pierre-Simon Laplace in 1784-85 while investigating the movement of the planets in the solar system, using the gravitational law of Newton. It turns out that the Laplace operator occurs in most of the equations of mathematical physics: wave propagation, heat propagation, diusion processes, electromagnetic phenomena (Maxwells equations), quantum mechanics (the Schrdinger equation). o The two-dimensional Laplace operator (1.24) admits the following approximation: 4 f (x + h, y) + f (x h, y) + f (x, y + h) + f (x, y h) f (x, y) 2 f (x) h 4
that is, f (x, y) is determined by the average value of f at neighboring points (x + h, y) , (x h, y) , (x, y + h) , (x, y h) minus the value at (x, y). This observation motivates us to dene a discrete version of the Laplace operator on any graph as follows. Denition. Let (V, E) be a locally nite graph without isolated points (so that 0 < deg (x) < 1 for all x 2 V ). For any function f : V ! R, dene the function
26 f by
CHAPTER 1. THE LAPLACE OPERATOR ON GRAPHS 1 X f (y) f (x) . deg (x) yx
f (x) =
The operator on functions on V is called the Laplace operator of (V, E). In words, f (x) is the dierence between the arithmetic mean of f (y) at all vertices y x and f (x). Note that the set R of values of f can be replaced by any vector space over R, for example, by C. For example, on the lattice graph Z we have f (x) = while on Z2 f (x, y) = f (x + 1, y) + f (x 1, y) + f (x, y + 1) + f (x, y 1) f (x) . 4 f (x + 1) + f (x 1) f (x) , 2
The notion of the Laplace operator can be extended to weighted graphs as follows. Denition. Let (V, ) be a locally nite weighted graph without isolated points. For any function f : V ! R, dene the function f by f (x) = 1 X f (y) xy f (x) . (x) y (1.25)
The operator acting on functions on V , is called the weighted Laplace operator of (V, ). Note that the summation in (1.25) can be restricted to y x because otherwise xy = 0. Hence, f (x) is the dierence between the weighted average of f (y) at the vertices y x and f (x). The Laplace operator is a particular case of the weighted Laplace operator when the weight is simple, that is, when xy = 1 for all x y. Denote by F the set of all real-valued functions on V . Then F is obviously a linear space with respect to addition of functions and multiplication by a constant. Then can be regarded as an operator in F, that is, : F ! F. Note that is a linear operator on F, that is, (f + g) = f + g for all functions f, g 2 F and 2 R, which obvious from (1.25). Another useful property to mention: const = 0 (a similar property holds for the dierential Laplace operator). Indeed, if f (x) c then 1 X 1 X f (y) xy = c =c (x) y (x) y xy whence the claim follows.
1.4. THE LAPLACE OPERATOR Recall that the corresponding reversible Markov kernel is given by P (x, y) = so that we can write f (x) = X
y
27
xy (x)
P (x, y) f (y) f (x) .
Consider the Markov kernel also as an operator on functions as follows: X P (x, y) f (y) . P f (x) =
y
This operator P is called the Markov operator. Hence, the Laplace operator and the Markov operator P are related by a simple identity = P id, where id is the identical operator in F.
Example. Let us approximate f 00 (x) on R using dierent values h1 and h2 for the steps of x: f 0 (x + h1 ) f 0 (x) 1 f (x + h1 ) f (x) f (x) f (x h2 ) f 00 (x) h1 h1 h1 h2 1 1 1 1 1 = f (x + h1 ) + f (x h2 ) + f (x) h1 h1 h2 h1 h2 # " 1 1 1 1 1 1 = + f (x + h1 ) + f (x h2 ) f (x) : 1 1 h1 h1 h2 h1 h2 h1 + h2
1 1 Hence, we obtain the weighted average of f (x + h1 ) and f (x h2 ) with the weights h1 and h2 , respectively. This average can be realized as a weighted Laplace operator as follows. Consider a sequence of reals fxk gk2Z that is dened by the rules
x0 = 0; xk+1 =
xk + h1 ; k is even xk + h2 ; k is odd.
For example, x1 = h1 , x2 = h1 + h2 , x1 = h2 , x2 = h2 h1 , etc. Set V = fxk gk2V and dene the edge set E on V by xk xk+1 . Now dene the weight xy on edges by xk xk+1 = Then we have (xk ) = xk xk+1 + xk xk1 = 1=h1 + 1=h2 and, for any function f on V , f (xk ) = = 1 f (xk+1 ) xk xk+1 + f (xk1 ) xk xk1 (xk ) 1 1 1 h1 f (xk + h1 ) + h2 f (xk h2 ) ; k is even, 1 1 1=h1 + 1=h2 h2 f (xk + h2 ) + h1 f (xk h1 ) ; k is odd. 1=h1 ; k is even, 1=h2 ; k is odd.
28
1.5
The Dirichlet problem
Broadly speaking, the Dirichlet problem is a boundary value problem of the following type: nd a function u in a domain assuming that u is known in and u is known at the boundary . For example, if is an interval (0, 1) then this problem becomes as follows: nd a function u (x) on [0, 1] such that 00 u (x) = f (x) for all x 2 (0, 1) u (0) = a and u (1) = b where the function f and the reals a, b are given. This problem can be solved by repeated integrations, provided f is continuous. A similar problem for n-dimensional P 2 Laplace operator = n x2 is stated as follows: given a bounded open domain k=1
k
Rn , nd a function u in the closure that satises the conditions u (x) = f (x) for all x 2 , u (x) = g (x) for all x 2 ,
(1.26)
where f and g are given functions. Under certain natural hypotheses, this problem can be solved, and a solution is unique. One of the sources of the Dirichlet problem is Electrical Engineering. If u (x) is the potential of an electrostatic eld in R3 then u satises in the equation u = f where f (x) is the density of a charge inside , while the values of u at the boundary are determined by the exterior conditions. For example, if the surface is a metal then it is equipotential so that u (x) = const on . Another source of the Dirichlet problem is Thermodynamics. If u (x) is a stationary temperature at a point x in a domain then u satises the equation u = f where f (x) is the heat source at the point x. Again the values of u at are determined by the exterior conditions. Let us consider an analogous problem on a graph that, in particular, arises from a dissertation of the problem (1.26) for numerical purposes. Theorem 1.10 Let (V, ) be a connected locally nite weighted graph (V, ), and let be a subset of V . Consider the following Dirichlet problem: u (x) = f (x) for all x 2 , (1.27) u (x) = g (x) for all x 2 c , where u : V ! R is an unknown function while the functions f : ! R and g : c ! R are given. If is nite and c is non-empty then, for all functions f, g as above, the Dirichlet problem (1.27) has a unique solution. Note that, by the second condition in (1.27), the function u is already dened outside , so the problem is to construct an extension of u to that would satisfy the equation u = f in . Dene the vertex boundary of as follows: = fy 2 c : y x for some x 2 g .
1.5. THE DIRICHLET PROBLEM
29
Observe that the Laplace equation u (x) = f (x) for x 2 involves the values u (y) at neighboring vertices y of x, and any neighboring point y belongs to either or to . Hence, the equation u (x) = f (x) uses the prescribed values of u only at the boundary , which means that the second condition in (1.27) can be restricted to as follows: u (x) = g (x) for all x 2 . This condition (as well as the second condition in (1.27) is called the boundary condition. If c is empty then the statement of Theorem 1.10 is not true. For example, in this case any constant function u satises the same equation u = 0 so that there is no uniqueness. The existence also fails in this case, see Exercises. The proof of Theorem 1.10 is based on the following maximum principle. A function u : V ! R is called subharmonic in if u (x) 0 for all x 2 , and superharmonic in if u (x) 0 for all x 2 . A function u is called harmonic in if it is both subharmonic and superharmonic, that is, if it satises the Laplace equation u = 0. For example, the constant function is harmonic on all sets. Lemma 1.11 (A maximum/minimum principle) Let (V, ) be a connected locally nite weighted graph and let be a non-empty nite subset of V such that c is non-empty. Then, for any function u : V ! R, that is subharmonic in , we have max u sup u,
c
and for any function u : V ! R, that is superharmonic in , we have min u inf u. c

Proof. It suces to prove the rst claim. If supc u = +1 then there is nothing to prove. If supc u < 1 then, by replacing u by u + const, we can assume that supc u = 0. Set M = max u
and show that M 0, which will settle the claim. Assume from the contrary that M > 0 and consider the set S := fx 2 V : u (x) = Mg . Clearly, S and S is non-empty. Claim 1. If x 2 S then all neighbors of x also belong to S. Indeed, we have u (x) 0 which can be rewritten in the form X u (x) P (x, y) u (y) .
yx
(1.28)
Since u (y) M for all y 2 V , we have X X P (x, y) u (y) M P (x, y) = M.

yx yx
30
Since u (x) = M, all inequalities in the above two lines must be equalities, whence it follows that u (y) = M for all y x. This implies that all such y belong to S. Claim 2. Let S be a non-empty set of vertices of a connected graph (V, E) such that x 2 S implies that all neighbors of x belong to S. Then S = V . Indeed, let x 2 S and y be any other vertex. Then there is a path fxk gn k=0 between x and y, that is, x = x0 x1 x2 ... xn = y. Since x0 2 S and x1 x0 , we obtain x1 2 S. Since x2 x1 , we obtain x2 2 S. By induction, we conclude that all xk 2 S, whence y 2 S. It follows from the two claims that the set (1.28) must coincide with V , which is not possible since u (x) 0 in c . This contradiction shows that M 0. Proof of Theorem 1.10. Let us rst prove the uniqueness. If we have two solutions u1 and u2 of (1.27) then the dierence u = u1 u2 satises the conditions u (x) = 0 for all x 2 , u (x) = 0 for all x 2 c . We need to prove that u 0. Since u is both subharmonic and superharmonic in , Lemma 1.11 yields 0 = inf u min u max u sup u = 0, c
c
whence u 0. Let us now prove the existence of a solution to (1.27) for all f, g. For any x 2 , rewrite the equation u (x) = f (x) in the form X X P (x, y) u (y) u (x) = f (x) P (x, y) g (y) , (1.29)
yx,y yx,yc
for all x 2 . Rewrite the equation (1.29) in the form Lu = h where h is the right hand side of (1.29), which is a given function on . Note that F is a linear space. Since the family 1{x} x of indicator functions form obviously a basis in F , we obtain that dim F = # < 1. Hence, the operator L : F ! F is a linear operator in a nitely dimensional space, and the rst part of the proof shows that Lu = 0 implies u = 0 (indeed, just set f = 0 and g = 0 in (1.29), that is, the operator L is injective. By By Linear Algebra, any injective operator acting in the spaces of equal dimensions, must be bijective (alternatively, one can say that the injectivity of L implies that det L 6= 0 whence it follows that L is invertible and, hence, bijective). Hence, for any h 2 F , there is a solution u = L1 h 2 F , which nishes the proof.
where we have moved to the right hand side the terms with y 2 c and used that u (y) = g (y). Denote by F the set of all real-valued functions u on and observe that the left hand side of (1.29) can be regarded as an operator in this space; denote it by Lu, that is, X Lu (x) = P (x, y) u (y) u (x) ,
yx,y
1.5. THE DIRICHLET PROBLEM
31
The initial function u0 can be chosen arbitrarily to satisfy the boundary condition; for example, take u0 = 0 in and u0 = g in c . In the case f = 0, we obtain the same recurrence relation un+1 = P un as for the distribution of the random walk, although now we have in addition some boundary values. Let us estimate the amount of computation for this method. Assuming that deg (x) is uniformly bounded, computation of P un (x) f (x) for all x 2 requires ' N operations, and this should be multiplied by the number of iterations. As we will see later (see Section 4.4), if is a subset of Zm of a cubic shape then the number of iterations should be ' N 2/m . Hence, the Jacobi method requires ' N 1+2/m operations. For comparison, the row reduction requires ' N 3 operations2 . If m = 1 then the Jacobi method requires also ' N 3 operations, but for higher dimensions m 2 the Jacobi method is more economical.
Example. Let us look at a numerical example in the lattice graph Z for the set = f1; 2; :::; 9g, for the boundary value problem u (x) = 0 in u (0) = 0; u (10) = 1:
How to calculate numerically the solution of the Dirichlet problem? Denote N = # and observe that solving the Dirichlet problem amounts to solving a linear system Lu = h where L is an N N matrix. If N is very large then the usual elimination method (not to say about inversion of matrices) requires too many operations. A more economical Jacobis method uses an approximating sequence fun g that is constructed as follows. Using that = P id, rewrite the equation u = f in the form u = P u f and consider a sequence of functions fun g given by the recurrence relation P un f in un+1 = . g in c
The exact solution is a linear function u (x) = x=10. Using the explicit expression for , write the approximating sequence in the form un (x + 1) + un (x 1) ; x 2 f1; 2; :::; 9g 2 while un (0) = 0 and un (10) = 1 for all n. Set u0 (x) = 0 for x 2 f1; 2; :::; 9g. The computations yield the following: un+1 (x) = x2Z n=0 n=1 n=2 n=3 n=4 ... n = 50 ... n = 81 0 0 0 0 0 0 :::
0:00
1 0 0 0 0 0 :::
0:084
2 0 0 0 0 0 :::
0:17
3 0 0 0 0 0 :::
0:26
4 0 0 0 0 0 :::
0:35
5 0 0 0 0 0 :::
0:45
6 0 0 0 0
1 16
7 0 0 0
1 8 1 8
8 0 0
1 4 1 4 3 8
9 0
1 2 1 2 5 8 5 8
10 1 1 1 1 1 :::
1:00
:::
0:55
:::
0:68
:::
0:77
:::
0:88
:::
0:00
:::
0:097
:::
0:19
:::
0:29
:::
0:39
:::
0:49
:::
0:59
:::
0:69
:::
0:79
:::
0:897
:::
1:00
so that u81 is rather close to the exact solution. Here N = 9 and, indeed, one needs ' N 2 iterations to approach to the solution.
2 For the row reduction method, one needs ' N 2 of row operation, and each row operation requires ' N of elementary operations. Hence, one needs ' N 3 of elementary operation.
32
Chapter 2 Spectral properties of the Laplace operator

Let (V, ) be a locally nite weighted graph without isolated points and be the Lecture 6 02.11.09 weighted Laplace operator on (V, ).
A.Grigoryan
2.1
Greens formula
Let us consider the dierence operator rxy that is dened for any two vertices x, y 2 V and maps F to R as follows: rxy f = f (y) f (x) . The relation between the Laplace operator and the dierence operator is given by X 1 X (rxy f) xy = P (x, y) (rxy f ) f (x) = (x) y y Indeed, the right hand side here is equal to X X X (f (y) f (x)) P (x, y) = f (y) P (x, y) f (x) P (x, y)
y
X
y
f (y) P (x, y) f (x) = f (x) .
The following theorem is one of the main tools when working with the Laplace operator. For any subset of V , denote by c the complement of , that is, c = V n . Theorem 2.1 (Greens formula) Let (V, ) be a locally nite weighted graph without isolated points, and let be a non-empty nite subset of V . Then, for any two functions f, g on V , X
x
f(x)g(x)(x) =
XX 1 X (rxy f) (rxy g) xy + (rxy f ) g(x)xy (2.1) 2 x,y c x y 33
34 CHAPTER 2. SPECTRAL PROPERTIES OF THE LAPLACE OPERATOR The formula (2.1) is analogous to the integration by parts formula Z
b a
f (x) g (x) dx =
00
b a
f 0 (x) g 0 (x) dx + [f 0 (x) g (x)]a
where f and g are smooth enough functions on [a, b]. A similar formula holds also for the dierential Laplace operator in bounded domains of Rn . If V is nite and = V then c is empty so that the last boundary term in (2.1) vanishes, and we obtain X f(x)g(x)(x) = 1 X (rxy f ) (rxy g) xy . 2 x,yV ! (2.2)
xV
Proof. We have X 1 X f (x)g(x)(x) = (f (y) f (x)) xy (x) yV x x XX = (f (y) f (x)) g(x)xy = = XX XX

y x x y x yV
g(x)(x)
(f (y) f (x)) g(x)xy + (f (x) f(y)) g(y)xy +
XX
x
XX
x yc
(f (y) f (x)) g(x)xy (rxy f) g(x)xy ,
yc
where in the last line we have switched notation of the variables x and y in the rst sum. Adding together the last two lines and dividing by 2, we obtain X
x
f (x)g(x)(x) =
XX 1 X (f(y) f(x)) (g(y) g(x)) xy + (rxy f) g(x)xy , 2 x,y x yc
which was to be proved.
2.2
Eigenvalues of the Laplace operator
Let (V, ) be a nite connected weighted graph where N := #V > 1. Let F denote the set of real-valued functions on V . Then F is a vector space over R of dimension N. Hence, the Laplace operator : F ! F is a linear operator in a N-dimensional vector space. We will investigate the spectral properties of this operator. Recall a few facts from Linear Algebra. Let A be a linear operator in a Ndimensional vector space V over R. A vector v 6= 0 is called an eigenvector of A if Av = v for some constant , that is called an eigenvalue of A. In general, one allows complex eigenvalues by considering a complexication of V. The set of all complex eigenvalues of A is called the spectrum of A and is denoted by spec A. All the eigenvalues of A can be found from the characteristic equation det (A id) = 0.
2.2. EIGENVALUES OF THE LAPLACE OPERATOR
35
Here A can be represented as an N N matrix in any basis of V; therefore, det (A id) is a polynomial of of degree N (that is called the characteristic polynomial of A), and it has N complex roots, counted with multiplicities. Hence, in general the spectrum of A consists of N complex eigenvalues. In the case when the underlying vector space is the space F of functions on the graph V , the eigenvectors are also referred to as eigenfunctions. Let us give some examples of explicit calculation of the eigenvalues of the operator L = on nite graphs with simple weight . In this case, we have Lf (x) = f (x) 1 X f (y) . deg (x) yx
Example. 1. Let V consist of two vertices f1, 2g connected by an edge. Then Lf (1) = f (1) f (2) Lf (2) = f (2) f (1) so that the equation Lf = f becomes (1 ) f (1) = f (2) (1 ) f (2) = f (1) whence (1 )2 f (k) = f (k) for both k = 1, 2. Since f 6 0, we obtain the equation (1 )2 = 1 whence we nd two eigenvalues = 0 and 1 = 2. Alternatively, considering a function f as a column-vector f (1) , we can represent the action of L f (2) as a matrix multiplication: Lf (1) f (1) 1 1 = , 1 1 Lf (2) f (2) 1 1 . Its so that the eigenvalues of L coincide with those of the matrix 1 1 characteristic equation is (1 )2 1 = 0, whence we obtain again the same two eigenvalues = 0 and = 2. 2. Let V = f1, 2, 3g with edges 1 2 3 1, that is, (V, E) = C3 = K3 . We have then 1 Lf (1) = f (1) (f (2) + f (3)) 2 and similar identities for Lf (2) and Lf (3) . The action of L can be written as a matrix multiplication: 0 1 0 10 1 Lf (1) 1 1/2 1/2 f (1) @ Lf (2) A = @ 1/2 1 1/2 A @ f (2) A . Lf (3) 1/2 1/2 1 f (3)
Calculating the characteristic polynomial of the above 3 3 matrix and evaluating its roots, one obtains the following eigenvalues of L: = 0 and = 3/2 with multiplicity 2.
36 CHAPTER 2. SPECTRAL PROPERTIES OF THE LAPLACE OPERATOR 3. Let V = f1, 2, 3g with edges 1 2 3. Then Lf (1) = f (1) f (2) 1 Lf (2) = f (2) (f (1) + f (3)) 2 Lf (3) = f (3) f (2) so that the matrix of L is 0 1 1 1 0 @ 1/2 1 1/2 A , 0 1 1
and the eigenvalues are = 0, = 1, and = 2. Coming back to the general theory, assume now that V is an inner product space, that is, an inner product (u, v) is dened for all u, v 2 V, that is a bilinear, symmetric, positive denite function on V V. An operator A is called symmetric (or self-adjoint) with respect to this inner product if (Au, v) = (u, Av) for all u, v 2 V. The following theorem collects important results from Linear Algebra about symmetric operators. Theorem. Let A be a symmetric operator in a N-dimensional inner product space V over R. (a) All eigenvalues of A are real. Hence, we can enumerate all the eigenvalues of A in increasing order as 1 , ..., N where each eigenvalue is counted with multiplicity. (b) (Diagonalization of symmetric operators) There is an orthonormal1 basis fvk gN k=1 in V such that each vk is an eigenvector of A with the eigenvalue k , that is Avk = k vk (equivalently, the matrix of A in the basis fvk g is diag (1 , ..., N )). (c) (The variational principle) Set R (v) = (Av,v) where (, ) is the inner product (v,v) in V (the function R (v) on V n f0g is called the Rayleigh quotient of A). The following identities are true for all k = 1, ..., N: k = R (vk ) =
vv1 ,...,vk1
inf
R (v) =
sup
vvN ,vN 1 ,...,vk+1
R (v)
(where v?u means that u and v are orthogonal, that is, (v, u) = 0). In particular, 1 = inf R (v) and N = sup R (v)
v6=0 v6=0
1
A sequence fvk g of vectors is called orthonormal if (vk ; vj ) = 1; k = j : 0; k 6= j
2.2. EIGENVALUES OF THE LAPLACE OPERATOR
37
One of the consequences of part (b) is that the algebraic multiplicity of any eigenvalue , that is, its multiplicity as a root of the characteristic polynomial, coincides with the geometric multiplicity, that is the maximal number of linearly independent eigenvectors with the same eigenvalue (= dim ker (A id)). An eigenvalue is called simple if its multiplicity is 1. That means, on the one hand, that this eigenvalue occurs in the list 1 , ..., N exactly once, and on the other hand, that the equation Av = v has exactly one non-zero solution v up to multiplication by const . We will apply this Theorem to the Laplace in the vector space F of realvalued functions on V . The eigenvectors of are also called the eigenfunctions. Consider in F the following inner product: for any two functions f, g 2 F, set X (f, g) := f (x) g (x) (x) ,
xV
which can be considered as the integration of f g against measure on V . Obviously, all axioms of an inner product are satised: (f, g) is bilinear, symmetric, and positive denite (the latter means that (f, f) > 0 for all f 6= 0). Lemma 2.2 The operator is symmetric with respect to the above inner product, that is, ( f, g) = (f, g) for all f, g 2 F. Proof. Indeed, by the Green formula (2.2), we have ( f, g) = X f (x) g (x) (x) = 1 X (rxy f ) (rxy g) xy , 2 x,yV
xV
and the last expression is symmetric in f, g so that it is equal also to ( g, f). In the next theorem, we will give a more detailed information about the spectrum of . It will be more convenient to work with the operator L = that is called a positive denite Laplacian. By denition, we have Lf (x) = f (x) 1 X f (y) xy , (x) yV
A graph (V, E) is called bipartite if V admits a partition into two non-empty disjoint subsets V1 , V2 such that x 6 y whenever x and y belong to the same set Vi . In terms of coloring, one can say that V is bipartite if its vertices can be colored by two colors, say black and white, so that the vertices of the same color are not connected by an edge. Example. Here are some examples of bipartite graphs.
and by the Green formula, the Rayleigh quotient of L is P ( f, f ) 1 x,yV (rxy f ) (rxy g) xy (Lf, f ) P = = . R (f ) = 2 (f, f ) (f, f ) 2 xV f (x) (x)
38 CHAPTER 2. SPECTRAL PROPERTIES OF THE LAPLACE OPERATOR 1. Z is bipartite with the following partition: V1 is the set of all odd integers and V2 is the set of all even integers. 2. Zm is bipartite with the following partition: V1 is the set of all points (x1 , ..., xm ) 2 Zm with an odd sum x1 + ... + xm , and V2 is the set of all points from Zm with an even sum x1 + ... + xm . 3. A cycle Cn is bipartite provided n is even. 4. A complete bipartite graph Kn,m is bipartite. 5. A binary cube f0, 1gm is bipartite with the following partition: V1 is the set of all points (x1 , ..., xm ) 2 f0, 1gm with an odd sum x1 + ... + xm , and V2 is the set of all points (x1 , ..., xm ) 2 f0, 1gm with an even sum x1 + ... + xm . Theorem 2.3 For any nite, connected, weighted graph (V, ) with #V > 1, the following is true. (a) Zero is a simple eigenvalue of L. (b) All the eigenvalues of L are contained in [0, 2]. (c) If (V, ) is not bipartite then all the eigenvalues of L are in [0, 2). Proof. (a) Since L1 = 0, the constant function is an eigenfunction with the eigenvalue 0. Assume now that f is an eigenfunction of the eigenvalue 0 and prove that f const, which will imply that 0 is a simple eigenvalue. If Lf = 0 then it follows from (2.2) with g = f that X (f (y) f (x))2 xy = 0.
{x,yV :xy}
In particular, f (x) = f (y) for any two neighboring vertices x, y. The connectedness of the graph means that any two vertices x, y 2 V can be connected to each other by a path fxk gm where x = x0 x1 ... xN = y whence it follows that k=0 f (x0 ) = f (x1 ) = ... = f (xm ) and f (x) = f (y). Since this is true for all couples x, y 2 V , we obtain f const. Alternatively, one can prove this using the maximum principle of Lemma 1.11. Indeed, choose a point x0 2 V and consider the set = V n fx0 g. Since this set is nite and function f is harmonic, we obtain by Lemma 1.11 that inf f inf f sup f sup f. c
c
However, inf c f = supc f = f (x0 ) whence it follows that f f (x0 ) = const . (b) Let be an eigenvalue of L with an eigenfunction f. Using Lf = f and (2.2), we obtain X X f 2 (x) (x) = Lf (x) f (x) (x)
xV xV
1 2
{x,yV :xy}
(f (y) f (x))2 xy .
(2.3)
2.3. CONVERGENCE TO EQUILIBRIUM
39
It follows from (2.3) that 0. Using an elementary inequality (a + b)2 2 (a2 + b2 ), we obtain X X f 2 (x) (x) f (y)2 + f (x)2 xy
xV {x,yV :xy}
= =
X
yV
x,yV
f (y)2 xy +
2
x,yV
f (y) (y) + f (x) (x) .

2
= 2
xV
xV
f (x)2 xy
f (x)2 (x) (2.4)
It follows from (2.4) that 2. (c) We need to prove that = 2 is not an eigenvalue. Assume from the contrary that = 2 is an eigenvalue with an eigenfunction f, and prove that (V, ) is bipartite. Since = 2, all the inequalities in the above calculation (2.4) must become equalities. In particular, we must have for all x y that (f (x) f (y))2 = 2 f (x)2 + f (y)2 , which is equivalent to f (x) + f (y) = 0. If f (x0 ) = 0 for some x0 then it follows that f (x) = 0 for all neighbors of x0 . Since the graph is connected, we obtain that f (x) 0, which is not possible for an eigenfunction. Hence, f (x) 6= 0 for all x 2 . Then V splits into a disjoint union of two sets: V + = fx 2 V : f (x) > 0g and V = fx 2 V : f (x) < 0g . The above argument shows that if x 2 V + then all neighbors of x are in V , and vice versa. Hence, (V, ) is bipartite, which nishes the proof. Hence, we can enumerate all the eigenvalues of L in the increasing order as follows: 0 = 0 < 1 2 ... N1 , Note that the smallest eigenvalue is denoted by 0 rather than by 1 . Also, we have always N1 2 and the inequality is strict if the graph is non-bipartite.
2.3
Convergence to equilibrium
Let P be the Markov operator associated with a weighted graph (V, ). We consider it as a linear operator from F to F. Recall that it is related to the Laplace operator L by the identity P = id L. It follows that all the eigenvalues of P have the form 1 where is an eigenvalue of L, and the eigenfunctions of P and L are the same. Since spec L [0, 2], it follows that spec P [1, 1]. It follows from Theorem 2.3 that 1 is a simple eigenvalue of P , and 1 is not an eigenvalue if the graph is non-bipartite.
40 CHAPTER 2. SPECTRAL PROPERTIES OF THE LAPLACE OPERATOR In the next statement, we consider the powers P n of P for any positive integer n, using composition of operators. Claim. For any f 2 F and any positive integer n, we have the following identity: X P n f (x) = Pn (x, y) f (y) (2.5)
yV
where Pn is the n-step transition function. Proof. For n = 1 (2.5) becomes the denition of the Markov operator. Let us make the inductive step from n to n + 1. We have X P n+1 f = P n (P f) = Pn (x, y) P f (y) = X
y z
Pn (x, y)
= =
X
z
X X
y
X
z
P (y, z) f (z) !
Pn (x, y) P (y, z) f (z)
Pn+1 (x, z) f (z) ,
which nishes the proof. In the last line we have used the identity (1.15) of Proposition 1.6. The next theorem is one of the main results of this Chapter. We use the notation p kf k = (f, f ). Theorem 2.4 Let (V, ) be a nite, connected, weighted graph and P be its Markov kernel. For any function f 2 F, set f= 1 X f (x) (x) . (V ) xV
Then, for any positive integer n, we have n P f f n kfk where
(2.6) (2.7)
= max (j1 1 j , j1 N1 j) .
Consequently, if the graph (V, ) is non-bipartite then n P f f ! 0 as n ! 1, that is, P n f converges to a constant f as n ! 1.
(2.8)
The estimate (2.6) gives the rate of convergence of P n f to the constant f : it is decreasing exponentially in n provided < 1. The constant is called the spectral radius of the Markov operator P . Indeed, since P = id L, the number = 1 is
2.3. CONVERGENCE TO EQUILIBRIUM
41
an eigenvalue of P provided is an eigenvalue of L. Hence, the numbers 1 N1 and 1 1 are respectively the minimal eigenvalue min and the maximal eigenvalue max of P except for 1, and we can write = max (jmin j , jmax j) . Equivalently, is the minimal positive number such that spec P n f1g [, ] , because the eigenvalues of P except for 1 are contained in [min , max ] . Proof of Theorem 2.4. If the graph (V, ) is non-bipartite then by Theorem 2.3 we have 0 < 1 N1 < 2. It follows that both numbers j1 1j and jN1 1j are smaller than 1, whence < 1. Therefore, n ! 0 as n ! 1 and (2.6) implies (2.8). To prove (2.6), choose an orthonormal basis fvk gN1 of the eigenfunctions of L k=0 so that Lvk = k vk and, hence, P vk = vk Lvk = (1 k ) vk . Any function f 2 F can be expanded in the basis vk as follows: f=
N1 X k=0
ak vk
where ak = (f, vk ) . By the Parseval identity, we have kfk = We have Pf = whence, by induction in n,
N1 X k=0 2 N1 X k=0
a2 . k
ak P vk =
N1 X k=0
(1 k ) ak vk ,
P f=
N1 X k=0
(1 k )n ak vk .
On the other hand, recall that v0 c for some constant c. It can be determined from the normalization condition kv0 k = 1, that is, X c2 (x) = 1
xV
whence c = p 1
(V )
. It follows that X 1 a0 = (f, v0 ) = p f (x) (x) (V ) xV
42 CHAPTER 2. SPECTRAL PROPERTIES OF THE LAPLACE OPERATOR and a0 v0 = Hence, we obtain P f f =

n N 1 X k=0
1 X f (x) (x) = f . (V ) xV
(1 k )n ak vk a0 v0
N1 X k=1
= (1 0 )n a0 v0 + =
N 1 X k=1
(1 k )n ak vk a0 v0
(1 k )n ak vk .
By the Parseval identity, we have

N 1 X n P f f 2 = (1 k )2n a2 k k=1
1kN1
max j1 k j
2n
N1 X k=1
a2 . k
As was already explained before the proof, all the eigenvalues 1 k of P with k 1 are contained in the interval [, ], so that j1 k j . Observing also that
N1 X k=1
a2 kf k2 , k
we obtain which nishes the proof.
n P f f 2 2n kfk2 ,
Corollary 2.5 Let (V, ) be a nite, connected, weighted graph that is non-bipartite, and let fXn g be the associated random walk. Fix a vertex x0 2 V and consider the distribution function of Xn : vn (x) = Px0 (Xn = x) . Then P vn (x) ! (x) as n ! 1, (V ) (2.9)
where (V ) =
xV
(x). Moreover, we have 2 X (x0 ) (x) vn (x) 2n . (V ) (x) xV (2.10)
2.3. CONVERGENCE TO EQUILIBRIUM It follows from (2.10) that, for any x 2 V ,
43
Proof. Since the graph is bipartite, we have 2 (0, 1), so that (2.9) follows from (2.10) (or from (2.11)). To prove (2.10), consider also the backward distribution function un (x) = Px (Xn = x0 ) and recall that, by (1.20), un (x) = Since un (x) = Pn (x, x0 ) = = P 1{x0 } (x) ,
n
s (x) vn (x) n (x) (V ) (x0 )
(2.11)
vn (x) (x0 ) . (x) X

yV
Pn (x, y) 1{x0 } (y)
we obtain by Theorem 2.4 with f = 1{x0 } that un 1{x0 } 2 2n 1{x0 } 2 .
Since 1{x0 } = we obtain that
1 (x0 ) (V )
2 and 1{x0 } = (x0 ) , 2 (x) 2n (x0 )
X vn (x) (x0 )
x
(x)
(x0 ) (V )
whence (2.9) follows. A random walk is called ergodic if (2.9) is satised, which is equivalent to (2.8). We have seen that a random walk on a nite, connected, non-bipartite graph is ergodic. Let us show that if N1 = 2 then this is not the case (as we will see later, for bipartite graphs one has exactly N1 = 2). Indeed, if f is an eigenfunction of L with the eigenvalue 2 then f is the eigenfunction of P with the eigenvalue 1, that is, P f = f. Then we obtain that P n f = (1)n f so that P n f does not converge to any function as n ! 1. A.Grigoryan Hence, in the case of a non-bipartite graph, the rate of convergence of the distri- Lecture 8 (x) bution function vn (x) to the equilibrium measure (V ) is determined by n where 09.11.09 = max (j1 1 j , j1 N1 j) = max (jmax j , jmin j) , where min is the minimal eigenvalue of the Markov operator P , while max is the maximal eigenvalue of P except for 1 (recall that spec P [1, 1]).
44 CHAPTER 2. SPECTRAL PROPERTIES OF THE LAPLACE OPERATOR Given a small number > 0, dene the mixing time T = T () by the condition T = min fn : n g , which gives T ln 1 1. ln (2.12)
With this T and n T , we obtain from (2.11) that
The value of should be chosen so that s (x) (x) << , (x0 ) (V ) which is equivalent to << min
x
s (x) vn (x) (x) . (V ) (x0 )
(x) . (V )
In many examples of large graphs, 1 is close to 0 and N1 is close to 2. Below is a diagram of the interval [0, 2] with marked eigenvalues in such a situation: j j j
1 1 N 1 2
0 =0
In this case, we have ln and ln whence 1 1 = ln 1 j1 1 j 1 1
1 1 = ln 2 N1 , j1 N1 j 1 (2 N1 ) 1 1 1 . T ln max , 1 2 N1 (2.13)
In the next sections, we will estimate the eigenvalues on specic graphs and, consequently, provide some explicit values for the mixing time.
2.4
More about the eigenvalues
The next statement contains an additional information about the spectrum of L in some specic cases. Theorem 2.6 Let (V, ) be a nite, connected, weighted graph with N := #V > 1.
2.4. MORE ABOUT THE EIGENVALUES (a) We have the following inequality 1 + ... + N1 N. Consequently, 1 If in addition (V, ) has no loops then N . N 1
45
(2.14)
(2.15)
1 + ... + N1 = N and N1 N . N 1
(2.16)
(2.17)
(b) If (V, ) is a complete graph with a simple weight , then 1 = ... = N1 = N . N1 (c) If (V, ) is non-complete then 1 1. (d) If (V, ) is a bipartite and is an eigenvalue of L then 2 is also an eigenvalue of L, with the same multiplicity as . In particular, 2 is a simple eigenvalue of L. Hence, we see that a graph with a simple weight is complete if and only if 1 > 1. Also, the graph is bipartite if and only if N1 = 2. Proof. (a) Let fvk gN1 be an orthonormal basis in F that consists of the k=0 eigenfunctions of L, so that Lvk = k vk . In the basis fvk g, the matrix of L is as follows diag (0 , 1 , ...N1 ) . Taking into account that 0 = 0, we obtain trace L = 0 + 1 + ... + N1 = 1 + ... + N1 . (2.18)
Note that the trace trace L does not depend on the choice of a basis. So, choose another basis as follows: enumerate all vertices of V as 0, 1, ..., N 1 and consider the functions 1{k} where k = 0, 1, ..., N 1, that obviously form a basis in F. The components of any function f 2 F in this basis are the values f (k). Rewrite the denition of Lin the form X Lf (i) = f (i) P (i, j) f (xj )
j
= (1 P (i, i)) f (i)
X
j6=i
P (i, j) f (xj ) .
We see that the matrix of L in this basis has the values 1 P (i, i) on the diagonal (the coecients in front of f (xi )) and P (i, j) in the intersection of the column i and the row j away from the diagonal. It follows that trace L =
N1 X i=0
(1 P (i, i)) = N
N1 X i=0
P (i, i) N.
(2.19)
46 CHAPTER 2. SPECTRAL PROPERTIES OF THE LAPLACE OPERATOR Comparison with (2.18) proves (2.14). Since 1 is the minimum of the sequence f1 , ..., N1 g of N 1 numbers, (2.15) follows from (2.14). If the graph has no loops then P (i, i) = 0, and (2.19) implies (2.16). Since N1 is the maximum of the sequence f1 , ..., N 1 g, we obtain (2.17). (b) We need to construct N 1 linearly independent eigenfunctions with the N eigenvalue N1 . As above, set V = f0, 1, ..., N 1g and consider the following N 1 functions fk for k = 1, 2, ...N 1 : 8 i = 0, < 1, 1, i = k, fk (i) = : 0, otherwise. We have Lfk (i) = fk (i) P If i = 0 then fk (0) = 1 and in the sum j6=0 fk (j) there is exactly one term = 1, for j = k, and all others vanish, whence Lfk (0) = fk (0) = 1+ 1 X fk (j) N 1 j6=0 1 X fk (j) . N 1 j6=i
1 N = fk (0) . N 1 N 1 P If i = k then fk (k) = 1 and in the sum j6=k fk (j) there is exactly one term = 1, for j = 0, whence Lfk (k) = fk (k) = 1 1 X fk (j) N 1 j6=k
1 N = fk (k) . N 1 N 1 P If i 6= 0, k then fk (i) = 0, while in the sum j6=k fk (j) there are terms 1, 1 and all others are 0, whence N fk (i) . Lfk (i) = 0 = N 1
N Hence, Lfk = N1 fk . Since the sequence ffk gN1 is linearly independent, we see k=1 N that N1 is the eigenvalue of multiplicity N 1, which nishes the proof. (c) By the variational principle, we have
1 = inf R (f)
f 1
where the condition f?1 comes from the fact that the eigenfunction of 0 is constant. Hence, to prove that 1 1 it suces to construct a function f?1 such that R (f ) 1. Claim 1. Fix z 2 V and consider the indicator function f = 1{z} . Then R (f) 1. We have X (f, f ) = f (x)2 (x) = (z)
xV
2.4. MORE ABOUT THE EIGENVALUES and, by the Green formula, (Lf, f ) = 1 X (f (x) f (y))2 xy 2 x,yV ! X X 1 (f (x) f (y))2 xy = + 2 x=z,y6=z x6=z,y=z X X = (f (z) f (y))2 zy = zy (z),
y6=z y6=z
47
whence R (f) 1 (not that in the case of no loops we obtain the identity R (f) = 1). Clearly, we have also R (cf) 1 for any constant c. Claim 2. Let f, g be two functions on V such that R (f) 1, and their supports2 A = fx 2 V : f (x) 6= 0g and B = fx 2 V : g (x) 6= 0g R (g) 1,
are disjoint and not connected, that is, x 2 A and y 2 B implies that x 6= y and x 6 y. Then R (f + g) 1. It is obvious that f g 0. Let us show that also (Lf) g 0. Indeed, if g (x) = 0 then (Lf ) g (x) = 0. If g (x) 6= 0 then x 2 B. It follows that f (x) = 0 and f (y) = 0 for any y x whence X Lf (x) = f (x) P (x, y) f (y) = 0,
yx
whence (Lf) g (x) = 0. Using the identities f g = (Lf) g = (Lg) f = 0, we obtain (f + g, f + g) = (f, f) + 2 (f, g) + (g, g) = (f, f) + (g, g) and (L (f + g) , f + g) = (Lf, f ) + (Lg, f ) + (Lf, g) + (Lg, g) = (Lf, f ) + (Lg, g) . Since by hypothesis (Lf, f) (f, f ) and (Lg, g) (g, g) it follows that R (f + g) = (Lf, f ) + (Lg, g) 1. (f, f ) + (g, g)
Now we construct a function f ?1 such that R (f ) 1. Since the graph is noncomplete, there are two distinct vertices, say z1 and z2 , such that z1 6 z2 . Consider function f in the form f (x) = c1 1{z1 } + c2 1{z2 }
2
The set fx : f (x) 6= 0g is called the support of f .
48 CHAPTER 2. SPECTRAL PROPERTIES OF THE LAPLACE OPERATOR where the coecients c1 and c2 are chosen so that f ?1 (for example, c1 = 1/ (z1 ) and c2 = 1/ (z2 )). Since R ci 1{zi } 1 and the supports of 1{z1 } and 1{z2 } are not connected, we obtain that also R (f ) 1, which nishes the proof. (d) Since the eigenvalues of the Markov operator are related to the eigenvalues of L by = 1 , the claim is equivalent to the following: if is an eigenvalue of P then is also an eigenvalue of P with the same multiplicity (indeed, = 1 implies = 1 (2 )). Let V + , V be a partition of V such that x y only if x and y belong to same of the subset V + , V . Given an eigenfunction f of P with the eigenvalue , consider f (x) , x2 V+ g (x) = . (2.20) f (x) , x 2 V Let us show that g is an eigenfunction of P with the eigenvalue . For all x 2 V + , we have X X P g (x) = P (x, y) g (y) = P (x, y) g (y)
yV
yV
= P f (x) = f (x) = g (x) , and for x 2 V we obtain in the same way X P g (x) = P (x, y) g (y)
yV +
yV
P (x, y) f (y)
yV
P (x, y) f (y) = P f (x) = f (x) = g (x) .
Hence, is an eigenvalue of P with the eigenfunction g. Let m be the multiplicity of as an eigenvalue of P , and m0 be the multiplicity of . Let us prove that m0 = m. There exist m linearly independent eigenfunctions f1 , ..., fm of the eigenvalue . Using (2.20), we construct m eigenfunctions g1 , ..., gm of the eigenvalue , that are obviously linearly independent, whence we conclude that m0 m. Since () = , applying the same argument to the eigenvalue instead of , we obtain the opposite inequality m m0 , whence m = m0 . Finally, since 0 is a simple eigenvalue of L, it follows that 2 is also a simple eigenvalue of L. It follows from the proof that the eigenfunction g (x) with the eigenvalue 2 is as follows: g (x) = c on V + and g (x) = c on V , for any non-zero constant c. Let us give an example of computation of the eigenvalues of L. Frequently it is more convenient to compute the eigenvalues of the Markov operator P = id L. Denote k = 1k so that fk gN1 is the sequence of all the eigenvalues of P in the k=0 decreasing order, counted with multiplicities. From what we know it follows that 1. k 2 [1, 1] , 2. 0 = 1 is a simple eigenvalue,
2.4. MORE ABOUT THE EIGENVALUES
49
3. N1 > 1 if the graph is non-bipartite, and N1 = 1 is a simple eigenvalue if the graph is bipartite. Example. Let us compute the eigenvalues of the Laplace operator on the cycle graph Cm with simple weight. Recall that Cm = Zm = f0, 1, ..., m 1g and the connections are 0 1 2 ... m 1 0. The Markov operator is given by P f (k) = 1 (f (k + 1) + f (k 1)) 2
We know already that = 1 is always a simple eigenvalue of P , and = 1 is Lecture 9 a (simple) eigenvalue if and only if Zm is bipartite, that is, if m is even. Assume in 10.11.09 what follows that 2 (1, 1) . Consider rst the dierence equation (2.21) on Z, that is, for all k 2 Z, and nd all solutions f as functions on Z. Observe rst that the set of all solutions of (2.21) is a linear space (the sum of two solutions is a solution, and a multiple of a solution is a solution), and the dimension of this space is 2, because function f is uniquely determined by (2.21) and by two initial conditions f (0) = a and f (1) = b. Therefore, to nd all solutions of (2.21), it suces to nd two linearly independent solutions. Let us search specic solution of (2.21) in the form f (k) = rk where the number r is to be found. Substituting into (2.21) and cancelling by rk , we obtain the equation for r: r2 2r + 1 = 0. It has two complex roots p r = i 1 2 = ei , where 2 (0, ) is determined by the condition cos = (and sin = p 1 2 ).
where k is regarded as a residue mod m. The eigenfunction equation P f = f becomes f (k + 1) 2f (k) + f (k 1) = 0. (2.21)
A.Grigoryan
Hence, we obtain two independent complex-valued solutions of (2.21) f1 (k) = eik and f2 (k) = eik . Taking their linear combinations and using the Euler formula, we arrive at the following real-valued independent solutions: f1 (k) = cos k and f2 (k) = sin k. (2.22)
In order to be able to consider a function f (k) on Z as a function on Zm , it must be m-periodic, that is, f (k + m) = f (k) for all k 2 Z.
50 CHAPTER 2. SPECTRAL PROPERTIES OF THE LAPLACE OPERATOR The functions (2.22) are m-periodic provided m is a multiple of 2, that is, = 2l , m
for some integer l. The restriction 2 (0, ) is equivalent to l 2 (0, m/2) . Hence, for each l from this range we obtain an eigenvalue = cos of multiplicity 2 (with eigenfunctions cos k and sin k). Let us summarize this result in the following statement. Lemma 2.7 The eigenvalues of the Markov operator P on a lattice graph Zm are as follows: 1. If m is odd then the eigenvalues are = 1 (simple) and = cos 2l for all m l = 1, ..., m1 (double); 2 2. if m is even then the eigenvalues are = 1 (simple) and = cos 2l for all m l = 1, ..., m 1 (double). 2 Note that the sum of the multiplicities of all the listed above eigenvalues is m so that we have found indeed all the eigenvalues of P . For example, in the case m = 3 we obtain the Markov eigenvalues = 1 and = cos 2 = 1 (double). The eigenvalues of L are as follows: = 0 and = 3/2 3 2 (double). If m = 4 then the Markov eigenvalues are = 1 and = cos 2 = 0 4 (double). The eigenvalues of L are as follows: = 0, = 1 (double), = 2.
Alternatively, one can list all the eigenvalues of P with multiplicities in the following sequence: 2j m1 cos j=0 : m Indeed, if m is odd then j = 0 gives = 1, and for j = 1; :::; m 1 we have 2j 2l j; 1 j m1 ; 2 cos = cos where l = m j; m+1 j m 1; m m 2 because 2j 2j 2l cos = cos 2 = cos : m m m
m1 m1 2 Hence, every value of cos 2l l=1 occurs in the sequence cos 2j j=1 exactly twice. m m In the same way, if m is even, then j = 0 and j = m=2 give the values 1 and 1, respectively, while for j 2 [1; m 1] n fm=2g we have 2j 2l j; 1 j m 1; 2 cos = cos where l = m j; m + 1 j m 1; m m 2 m 1 2 so that each value of cos 2l l=1 is counted exactly twice. m
2.5. PRODUCTS OF GRAPHS
51
2.5
Products of graphs
Denition. Let (X, E1 ) and (Y, E2 ) be two graphs. Their Cartesian product is dened as follows:3 (V, E) = (X, E1 ) (Y, E2 ) where V = X Y is the set of pairs (x, y) where x 2 X and y 2 Y , and the set E of edges is dened by either x0 x and y = y 0 0 0 (x, y) (x , y ) if . (2.23) or y y 0 and x = x0 Clearly, we have #V = (#X) (#Y ) and deg (x, y) = deg (x) + deg (y) for all x 2 X and y 2 Y . For example, Z Z = Z2 and, more generally, Zn Zm = Zn+m . Let us dene a more general notion of the product of weighted graphs. Denition. Let (X, a) and (Y, b) be two locally nite weighted graphs. Fix two numbers p, q > 0 and dene the product graph (V, ) = (X, a) p,q (Y, b) as follows: V = X Y and the weight on V is 8 < pb (y) axx0 , qa (x) byy0 , (x,y),(x0 ,y0 ) = : 0, dened by if y = y 0 , if x = x0 , otherwise.
The numbers p, q are called the parameters of the product. Clearly, the product weight (x,y),(x0 ,y0 ) is symmetric, and the edges of the graph (V, ) are exactly those from (2.23). The weight on the vertices of V is given by X X X (x,y),(x0 ,y0 ) = p axx0 b (y) + q a (x) byy0 = (p + q) a (x) b (y) . (x, y) =
x0 ,y0 x0 y0
Claim. If A and B are the Markov kernels on X and on the product (V, ) is given by 8 p 0 < p+q A (x, x ) , q B (y, y 0 ) , P ((x, y) , (x0 , y 0 )) = : p+q 0, Proof. Indeed, we have in the case y = y 0 P ((x, y) , (x0 , y 0 )) =
3
Y , then the Markov kernel P if y = y 0 , if x = x0 , otherwise.
(2.24)
(x,y),(x0 ,y0 ) paxx0 b (y) p axx0 p = = = A (x, x0 ) , (x, y) (p + q) a (x) b (y) p + q a (x) p+q
(X; E1 ) (Y; E2 ) :
The product graph is also denoted by
52 CHAPTER 2. SPECTRAL PROPERTIES OF THE LAPLACE OPERATOR and the case x = x0 is treated similarly. For the random walk on (V, ), the identity (2.24) means the following: the random walk at (x, y) chooses rst between the directions X and Y with probabilp q ities p+q and p+q , respectively, and then chooses a vertex in the chosen direction accordingly to the Markov kernel there. In particular, if a and b are simple weights, then we obtain 8 < p deg (y) , if x x0 and y = y 0 , q deg (x) , if y y 0 , and x = x0 , (x,y),(x0 ,y0 ) = : 0, otherwise.
If in addition the graphs A and B are regular, that is, deg (x) = const =: deg (A) and deg (y) = const =: deg (B) then the most natural choice of the parameter p and q is as follows 1 1 p= and q = , deg (B) deg (A)
so that the weight is also simple. We obtain the following statement. Lemma 2.8 If (X, a) and (Y, b) are regular graphs with simple weights, then the product (X, a) 1 , 1 (Y, b)
deg(B) deg(A)
is again a regular graph with a simple weight. Note that the degree of the product graph is deg (A) + deg (B). Example. Consider the graphs Zn and Zm with simple weights. Since their degrees are equal to 2n and 2m, respectively, we obtain Zn
1 , 1 2m 2n
Zm = Zn+m .
In other words, the eigenvalues of P are the convex combinations of eigenvalues p q of A and B, with the coecients p+q and p+q . Note that the same relation holds for the eigenvalues of the Laplace operators: since those on (X, a) and (Y, b) are 1 k and 1 l , respectively, we see that the eigenvalues of the Laplace operator on (V, ) are given by 1 pk + q l p (1 k ) + q (1 l ) = , p+q p+q
Theorem 2.9 Let (X, a) and (Y, b) be nite weighted graphs without isolated vertices, and let fk gn1 and f l gm1 be the sequences of the eigenvalues of the Markov k=0 l=0 operators A and B respectively, counted with multiplicities. Then all the eigenvalues of the Markov operator P on the product (V, ) = (X, a) p,q (Y, b) are given by the n o sequence
pk +q l p+q
where k = 0, ..., n 1 and l = 0, ..., m 1.
that is, the same convex combination of 1 k and 1 l .
2.5. PRODUCTS OF GRAPHS
53
Proof. Let f be an eigenfunction of A with the eigenvalue and g be the eigenfunction of B with the eigenvalue . Let us show that the function h (x, y) = f (x) g (y) is the eigenvalue of P with the eigenvalue p+q . We have p+q X P ((x, y) , (x0 , y 0 )) h (x0 , y 0 ) P h (x, y) = = = q X p A (x, x0 ) f (x0 ) g (y) + B (y, y 0 ) f (x) g (y 0 ) p + q x0 p + q y0 p q = Af (x) g (y) + f (x) Bg (y) p+q p+q p q = f (x) g (y) + f (x) g (y) p+q p+q p + q h (x, y) , = p+q which was to be proved. Let ffk g be a basis in the space of functions on X such that Afk = k fk , and fgl g be a basis in the space of functions on Y , such that Bgl = l gl . Then hkl (x, y) = fk (x) gl (y) is a linearly independent sequence of functions on V = XY . Since the number of such functions is nm = #V , we see that fhkl g is a basis in the k +q space of functions on V . Since hkl is the eigenfunction with the eigenvalue pp+q l , k +q we conclude that the sequence pp+q l exhausts all the eigenvalues of P . Corollary 2.10 Let (V, E) be a nite connected regular graph with N > 1 vertices, and set (V n , En ) = (V, E)n . Let be a simple weight on V , and fk gN1 be the k=0 sequence of the eigenvalues of the Markov operator on (V, ), counted with multiplicity. Let n be a simple weight on V n . Then the eigenvalues of the Markov operator on (V n , n ) are given by the sequence k1 + k2 + ... + kn (2.25) n for all ki 2 f0, 1, ..., N 1g , where each eigenvalue is counted with multiplicity. It follows that if fk gN 1 is the sequence of the eigenvalues of the Laplace operk=0 ator on (V, ) then the eigenvalues of Laplace operator on (V n , n ) are given by the sequence k1 + k2 + ... + kn . (2.26) n Proof. Induction in n. If n = 1 then there is nothing to prove. Let us make the inductive step from n to n + 1. Let degree of (V, E) be D, then deg (V n ) = nD. Note that. n+1 V , En+1 = (V n , En ) (V, E) It follows from Lemma 2.8 that n+1 V , n+1 = (V n , n ) 1 ,
1 D nD
X
x0
x0 ,y0
P ((x, y) , (x0 , y)) h (x0 , y) + X
X
y0
P ((x, y) , (x, y 0 )) h (x, y 0 )
(V, ) .
54 CHAPTER 2. SPECTRAL PROPERTIES OF THE LAPLACE OPERATOR By the inductive hypothesis, the eigenvalues of the Laplacian on (V n , n ) are given by the sequence (2.25). Hence, by Theorem 2.9, the eigenvalues on V n+1 , n+1 are given by k1 + k2 + ... + kn 1/D 1/ (nD) + k 1/D + 1/ (nD) n 1/D + 1/ (nD) n k1 + k2 + ... + kn 1 = + k n+1 n n+1 k1 + k2 + ... + kn + k = , n+1 which was to be proved.
2.6
Eigenvalues and mixing times in Zn m
Consider Zn with an odd m so that the graph Zn is not bipartite. By Corollary 2.5, m m the distribution of the random walk on Zn converges to the equilibrium measure m (x) 1 = N , where N = #V = mn , and the rate of convergence is determined by the (V ) spectral radius = max (jmin j , jmax j) (2.27) where min is the minimal eigenvalue of P and max is the second maximal eigenvalue of P . In the case n = 1, all the eigenvalues of P except for 1 are listed in the sequence (without multiplicity): 2l m1 cos . , 1l m 2 This sequence is obviously decreasing in l, and its maximal and minimal values are 2 m 1 2 and cos = cos , cos m m 2 m k1 + k2 + ... + kn n
respectively. For a general n, by Corollary 2.10, the eigenvalue of P have the form (2.28)
where ki are the eigenvalues of P . The minimal value of (2.28) is equal to the minimal value of k , that is min = cos . m The maximal value of (2.28) is, of course, 1 when all ki = 1, and the second maximal value is obtained when one of ki is equal to cos 2 and the rest n 1 m values are 1. Hence, we have max = n 1 + cos 2 1 cos 2 m m =1 . n n (2.29)
For example, if m = 3 then min = cos = 1 and 3 2 max 1 cos 2 3 3 =1 =1 n 2n
2.6. EIGENVALUES AND MIXING TIMES IN ZN M whence 1 1 , 2 1 3 = , = max 1 2 2n n 3, n 4.

2
55
3 , 2n
(2.30)
A.Grigoryan
If m is large then, using the approximation cos 1 2 for small , we obtain Lecture 10 16.11.09 2 22 min 1 , max 1 . 2m2 nm2
1 Using ln 1 , we obtain
ln
1 jmin j
2 , 2m2
ln
1 jmax j
2 2 . nm2
Finally, by (2.12) and (2.27), we obtain the estimate of the mixing time in Zn with m error : 2 ln 1 ln 1 2m nm2 1 = nm2 , , (2.31) T 1 ln max 2 22 2 2 ln
1 assuming for simplicity that n 4. Choosing = N 2 (note that has to be chosen 1 signicantly smaller than the limit value N of the distribution) and using N = mn , we obtain 1 2 ln mn nm2 = 2 n2 m2 ln m. (2.32) T 2 2 For example, in Z10 the mixing time is T 400, which is relatively short given the 5 fact that the number of vertices in this graph is N = 510 106 . In fact, the actual mixing time is even smaller than above value of T since the latter is an upper bound for the mixing time. For comparison, if we choose n = 1, then (2.31) slightly changes to become
T ln
1 2m2 4 = 2 m2 ln m, 2
1 where we have chosen = m2 . For example, in Zm with m = 106 we have T ' 1012 , which is huge in striking contrast with the previous graph Z10 with (almost) the 5 same number of vertices as Z106 !
Let us try to spot those Zn with the minimal mixing time assuming that the number of m vertices N = mn is (approximately) given. The formula (2.32) can be rewritten in terms of m and N = mn as follows: m2 T 2 ln2 N; ln m where we have used n = ln N= ln m. Therefore, to minimize T with a given N , the value of m should be taken minimal, that is, m = 3. In this case, we can use the exact value of given by 1 (2.30), and obtain that, for Zn with large n and " = N 2 = 32n : 3 ln 1 4 4 2n ln 3 2 2 " T = ln 3 n2 = (ln N ) 1:2 (ln N ) : 3 3= (2n) 3 3 ln 3 ln 1 2n
Consider now a bipartite case. Recall that Theorem 2.4 and Corollary 2.5 assert that the distribution function vn (x) = Px0 (Xn = x)
For example, for Z13 with N = 313 > 106 vertices, we obtain a very short mixing time T 250. 3
56 CHAPTER 2. SPECTRAL PROPERTIES OF THE LAPLACE OPERATOR of a random walk on a nite connected weighted graph (V, ) convergences to the (x) equilibrium measure (V ) as n ! 1 provided the graph (V, ) is non-bipartite. For bipartite graphs, the following analogous statement is true (see Exercise 20): Claim. Assume that (V, ) is bipartite. If n is even and n ! 1 then ( 2(x) , x 6 x0 , (V ) e vn (x) ! v (x) := 0, x x0 . Moreover, for all x 2 V , e jvn (x) v (x)j n s (x) , (x0 )
where
= max (j1 1 j , jN2 1j) . Note that here we use N2 rather than N1 as before, because N1 = 2 for bipartite graphs. Recall also that by Theorem 2.6 1 1 and N2 = 2 1 so that in fact = 1 1 . It follows that the mixing time is estimated by log 1 log 1 T 1 1 log
(x) assuming 1 0. Here must be chosen so that << minx (V ) . Let us estimate the mixing time on the binary cube Zn = f0, 1gn that is a 2 bipartite graph. The eigenvalues of the Laplace operator on Z2 = f0, 1g are 0 = 0 and 1 = 2. Then the eigenvalues of the Laplace operator on f0, 1gn are given by (2.26), that is, by k1 + k2 + ... + kn n
where each ki = 0 or 1. Hence, each eigenvalue of f0, 1gn is equal to 2j where n j 2 [0, n] is the number of 1s in the sequence k1 , ..., kn . The multiplicity of the eigenvalue 2j is equal to the number of binary sequences fk1 , ..., k g where 1 occurs n n exactly j times. This number is given by the binomial coecient n . Hence, all the j eigenvalues of the Laplace operator f0, 1gn are 2j where j = 0, 1, ..., n, and on n the multiplicity of this eigenvalue is n . j
Note that the total sum of all multiplicities is
n X n j=0
= 2n ;
that is the number of vertices in f0; 1gn as expected. The trace of the Laplace operator is equal to sum of the eigenvalues =
n X 2j n j=0
2.6. EIGENVALUES AND MIXING TIMES IN ZN M

while by Theorem 2.6 the trace is equal to 2n : Hence, we obtain an identity
n X n j = n2n1 : j j=0
57
Of course, one can prove this identity independently by induction in n. For example, the eigenvalues of f0; 1g are 0; 2 ; 4 ; 2 where 0 and 2 are simple, while 2 and 4 3 3 3 3 have multiplicities 3 = 3 = 3: For comparison, let us try to evaluate the eigenvalues of f0; 1g3 1 2 directly. Numbering the vertices as on the diagram, 7 j j 3 j j j 0 j 4 6 j 2 j j j 1 j j 5 1
3
Using computer packages as Maple, one obtains that the eigenvalues of this matrix are indeed as listed above.
2 The convergence rate of the random walk on f0, 1gn is determined by 1 = n . 1 Assume that n is large and let N = 2n . Taking = N 2 = 22n , we obtain the following estimate of the mixing time:
we obtain that the matrix of the Laplace operator is 0 1 1 0 1 1 3 3 3 1 B 1 1 0 0 3 B 3 B 0 1 1 1 0 3 3 B 1 B 0 1 1 0 3 B 3 B 1 0 0 0 1 B 3 B 0 1 0 0 1 3 3 B @ 0 0 1 0 0 3 0 0 0 1 1 3 3
0 1 3 0 0 1 3 1 1 3 0
0 0 1 3 0 0 1 3 1 1 3
0 0 0 1 3 1 3 0 1 3 1
C C C C C C: C C C C A
ln 1 2n ln 2 (ln N)2 = (ln 2) n2 = 1.4 (ln N)2 . = 1 2/n ln 2
For example, for f0, 1g20 with N = 220 106 vertices we obtain T 280.
58 CHAPTER 2. SPECTRAL PROPERTIES OF THE LAPLACE OPERATOR
Chapter 3 Geometric bounds for the eigenvalues

We are in the general setting of a nite, connected weighted graph (V, ) with N = #V > 1, and our purpose is to obtain the estimate for the smallest positive eigenvalue 1 of the Laplace operator L on (V, ).
3.1
Cheegers inequality
Let (V, ) be a weighted graph with the edges set E. Recall that, for any vertex subset V , its measure () is dened by X (x) . () =
x
Similarly, for any edge subset S E, dene its measure (S) by X , (S) =
S
where := xy for any edge = xy. For any set V , dene its edge boundary by = fxy 2 E : x 2 , y 2 g . / Denition. Given a nite weighted graph (V, ), dene its Cheeger constant by h = h (V, ) = inf () . () (3.1)
V () 1 (V ) 2
In other words, h is the largest constant such that the following inequality is true () h () for any subset of V with measure () 1 (V ). 2 59 (3.2)
60
CHAPTER 3. GEOMETRIC BOUNDS FOR THE EIGENVALUES
Lemma 3.1 We have 1 2h. Proof. Let be a set at which the inmum in (3.1) is attained. Consider the following function 1, x2 f (x) = a, x 2 c / where a is chosen so that f ?1, that is, () = a (c ) whence a= Then 1 R (f ) := (f, f) = and (Lf, f ) = 1X (f (x) f (y))2 xy 2 x,y X = (f (x) f (y))2 xy
x,yc (Lf,f ) (f,f )
() 1. (c )
so that it suces to prove that R (f) 2h. We have
X
xV
f (x)2 (x) = () + a2 (c ) = (1 + a) ()
= (1 + a)2
x,yc
= (1 + a)2 () . Hence, (1 + a)2 () R (f) = (1 + a) h 2h, (1 + a) () which was to be proved. The following lower bound of 1 via h is most useful and is frequently used. Theorem 3.2 (Cheegers inequality) We have h2 1 . 2 Before we prove this theorem, consider some examples. Example. Consider a weighted path graph (V, ) where V = f0, 1, ...N 1g, the edges are 0 1 2 ... N 1, and the weights k1,k = mk , where fmk gN1 is a given sequence of positive numbers. k=1 Then, for 1 k N 2, we have (k) = k1,k + k,k+1 = mk + mk+1 , (3.3)
xy
3.1. CHEEGERS INEQUALITY
61
and the same is true also for k = 0, N 1 if we dene m1 = mN = 0. The Markov kernel is then k,k+1 mk+1 P (k, k + 1) = . = (k) mk + mk+1 Claim. Assume that the sequence fmk gN1 is increasing, that is, mk mk+1 . Then k=1 1 h 2N . Proof. Let be a subset of V with () 1 (V ), and let k 1, k be an 2 edge of the boundary with the largest possible k. We claim that either or c is contained in [0, k 1]. Indeed, if there were vertices from both sets and c outside [0, k 1], that is, in [k, N 1], then there would have been an edge j 1, j 2 with j > k, which contradicts the choice of k. It follows that either () ([0, k 1]) or (c ) ([0, k 1]) . However, since () (c ), we obtain that in the both cases () ([0, k 1]) . We have ([0, k 1]) = =
k1 X j=0 k1 X j=0
(j) =
k1 X j1,j + j,j+1 j=0
(mj + mj+1 ) (3.4)
2kmk where we have used that mj mj+1 mk . Therefore () 2kmk . On the other hand, we have () k1,k = mk , whence it follows that mk 1 1 () , = () 2kmk 2k 2N
1 which proves that h 2N . Consequently, Theorem 3.2 yields
1 . 8N 2
(3.5)
For comparison, in the case of a simple weight, the exact value of 1 is 1 = 1 cos (see Exercises), which for large N is 1 2 5 , 2 N2 2 (N 1) N 1
which is of the same order in N as the estimate (3.5).
62
CHAPTER 3. GEOMETRIC BOUNDS FOR THE EIGENVALUES Now assume that the weights mk satisfy a stronger condition mk+1 cmk ,
for some constant c > 1 and all k = 0, ..., N 2. Then mk ckj mj for all k j, which allows to improve the estimate (3.4) as follows ([0, k 1]) =
k1 X j=0 k1 X j=0
(mj + mj+1 ) jk c mk + cj+1k mk
= mk ck + c1k 1 + c + ...ck1 ck 1 = mk ck + c1k c1 c+1 . mk c1 Therefore, we obtain () c1 () c+1 whence h

c1 c+1
and, by Theorem 3.2, 1 1 2 c1 c+1 2 . (3.6)
Let us estimate the mixing time on the above path graph (V, ). Since it is bipartite, the mixing time is given by ln 1 ln 1 , T 1 1 ln 11 where << mink
(k) . (V )
Observe that
N1 X j=0
(V ) =
(mj + mj+1 ) 2
N1 X j=1
mj
where we put m0 = mN = 0, whence min

k
where
(k) m1 1 PN1 , = (V ) 2M 2 j=1 mj M := PN1

j=1
mj
m1
3.1. CHEEGERS INEQUALITY Setting =

1 M2
63
(note that M N and N can be assumed large) we obtain T 2 ln M . 1
For an arbitrary increasing sequence fmk g, we obtain using (3.5) that T 8N 2 ln M. Consider the weights mk = ck where c > 1. Then we have M=
N1 X j=1
cj1 =
cN1 1 c1
whence ln M N ln c. Using also (3.6), we obtain c+1 T 4N ln c c1
Note that T is linear in N! Consider one more example: mk = kp for some p > 1. Then p p mk+1 1 1 = 1+ 1+ =: c mk k N If N >> p then c 1 +
p N
whence 1 1 2 c1 c+1 2 1 p2 . 8 N2
In this case, M= whence ln M (p + 1) ln N and T 16 (p + 1) 2 N ln N. p2

N1 X j=1
j p N p+1 ,
We precede the proof Theorem 3.2 by two lemmas. Given a function f : V ! R and an edge = xy, let us use the following notation1 : jr f j := jrxy f j = jf (y) f (x)j .
1
Note that r f is undened unless the edge is directed, whereas jr f j makes always sense.
64
Lemma 3.3 (Co-area formula). Given any real-valued function f on V , set for any t2R t = fx 2 V : f(x) > tg. Then the following identity holds: X
E
jr fj =
(t ) dt.
(3.7)
A similar formula holds for dierentiable functions on R: Z Z b 0 jf (x)j = # fx : f (x) = tg dt,

a
and the common value of the both sides is called the full variation of f. Proof. For any edge = xy, there corresponds an interval I R that is dened as follows: I = [f(x), f(y)) where we assume that f (x) f(y) (otherwise, switch the notations x and y). Denoting by jI j the Euclidean length of the interval I , we see that jr fj = jI j . Claim. 2 t , t 2 I . Indeed, the boundary t consists of edges = xy such that x 2 c and y 2 t , t that is, f(x) t and f (y) > t; which is equivalent to t 2 [f (x) , f (y)) = I . Thus, we have X X X (t ) = = = 1I (t) ,
t E:tI E
whence Z
+
(t ) dt =
= = which nishes the proof.
XZ X
E E
+ X E +
1I (t)dt 1I (t)dt X
E
jI j =
jr f j ,
Lemma 3.4 For any non-negative function f on V , such that 1 fx 2 V : f (x) > 0g (V ) , 2 the following is true: X
E
(3.8)
jr f j h
X
xV
f (x) (x) ,
(3.9)
where h is the Cheeger constant of (V, ) .
3.1. CHEEGERS INEQUALITY
65
Note that for the function f = 1 the condition (3.8) means that () 1 (V ), 2 and the inequality (3.9) is equivalent to () h () , because X
xV
f (x) (x) =
and
X
x
(x) = () X
X
E
jr fj =
x,yc
jf (y) f (x)j xy =
xy = () .
x,yc
Hence, the meaning of Lemma 3.4 is that the inequality (3.9) for the indicator functions implies the same inequality for arbitrary functions. Proof. By the co-area formula, we have Z Z X jr f j = (t ) dt (t ) dt.
E 0
By (3.8), the set t = fx 2 V : f (x) > tg has measure 1 (V ) for any t 0. 2 Therefore, by (3.2) (t ) h (t ) whence X
E
jr fj h
(t ) dt
On the other hand, noticing that x 2 t for a non-negative t is equivalent to t 2 [0, f (x)), we obtain Z X Z (t ) dt = (x) dt
0
= = =
xV
X X
xt X xV
(x) 1[0,f (x)) (t) dt Z

0
(x)
1[0,f (x)) (t) dt
(x) f (x) ,
xV
which nishes the proof. Proof of Theorem 3.2. Let f be the eigenfunction of 1 . Consider two sets V + = fx 2 V : f (x) 0g and V = fx 2 V : f (x) < 0g . Without loss of generality, we can assume that (V + ) (V ) (if not then replace f by f). It follows that (V + ) 1 (V ) . Consider the function 2 f, f 0, g = f+ := 0, f < 0.
66
Applying the Green formula (2.2) (Lf, g) = 1 X (rxy f ) (rxy g) xy 2 x,yV 1 X (rxy f ) (rxy g) xy . 2 x,yV
and using so that Lf = 1 f, we obtain 1 X

xV
f(x)g(x)(x) =
Observing that fg = g 2 and2 (rxy f) (rxy g) = (f (y) f (x)) (g (y) g (x)) (g (y) g (x))2 = jrxy gj2 . we obtain P jr gj2 P E 2 1 . xV g (x) (x)
Note that g 6 0 because otherwise f+ 0 and (f, 1) = 0 imply that f 0 whereas f is not identical 0. Hence, to prove (3.3) it suces to verify that X
E
jr gj2
h2 X 2 g (x) (x) . 2 xV
(3.10)
Since
1 (x 2 V : g (x) > 0) V + (V ) , 2 2 we can apply (3.9) to function g , which yields X X r g2 h g 2 (x) (x) .
E xV
(3.11)
Let us estimate from above the left hand side as follows: X X r g 2 = 1 g 2 (x) g 2 (y) xy 2 x,yV E 1X = jg(x) g(y)j1/2 jg(x) + g(y)j1/2 xy xy 2 x,y !1/2 1 X 1 X ( (g(x) g(y))2 xy ) ( (g(x) + g(y))2 xy ) , 2 x,y 2 x,y
2
Here we use the obvious identity a+ a = a2 and inequality + ja+ b+ j ja bj ;
that implies for arbitrary real a; b.
(a+ b+ ) (a b) (a+ b+ ) ;
3.2. ESTIMATING 1 FROM BELOW VIA DIAMETER where we have used the Cauchy-Schwarz inequality X
k
67
ak bk
X
k
!1/2 !1/2 X a2 b2 k k
k
that is true for arbitrary sequences of non-negative reals ak , bk . Next, using the inequality 1 (a + b)2 a2 + b2 , we obtain 2 X r g2
E
X
E
X
E
X jr gj2 g 2 (x) + g2 (y) xy

x,y
jr gj2 jr gj2
X
E
X
x,y
g2 (x)xy
!1/2
!1/2
X
xV
g2 (x) (x)
!1/2
which together with (3.11) yields h X g2 (x) (x)

xV
xV
Dividing by
1/2 g2 (x) (x) and taking square, we obtain (3.10).
X
E
jr gj2
!1/2 X
xV
g 2 (x) (x)
!1/2
3.2
Estimating 1 from below via diameter

1 c D (V )
Theorem 3.5 For any nite connected weighted graph (V, ) of diameter D,
where c = minxy xy (for example, c = 1 for a simple weight). Proof. Let f be the eigenfunction of the eigenvalue 1 . Let us normalize f to have max f = max jf j = 1, and let x0 be a vertex where f (x0 ) = 1. Since (f, 1) = 0, there is also a vertex y0 where f (y0 ) < 0. Let fxk gn be a path connecting x0 and y0 = xn where n D. k=0 Noticing that P 2 1 x,y (f (x) f (y)) xy 1 = R (f) = P 2 2 x f (x) (x) let us estimate the denominator and numerator as follows: X f (x)2 (x) (V ) ,
xV
68 and
X 1X (f (x) f (y))2 xy c (f (xk ) f (xk+1 ))2 2 x,y k=0 !2 n1 X c f (xk ) f (xk+1 ) n k=0 c = (f (x0 ) f (xn ))2 n c n
n1
because f (x0 ) = 1 and f (xn ) 0. Since n D, we obtain 1 c c , n (V ) D (V )
which was to be proved. For comparison, let us recall that 1 on Zn is m 1 = 1 cos 2 2 2 m n nm2
if m is large. Since (Zn ) = 2nmn and diam (Zn ) = nm, the estimate of Theorem m m 3.5 gives 1 1 2 n+1 . 2n m This estimate is good only if n = 1. Example. Let graph (V, E) consist of a graph (V+ , E+ ) (that is a nite connected graph), its copy (V , E ), and a path graph of 2m + 1 vertices zm z(m1) ...z1 z0 z1 ... zm1 zm so that zm is identied with a vertex x+ 2 V+ and zm is identied with the vertex x that is a copy of x+ in V . Let be a simple weight on (V, E). If diam (V+ ) << m << (V+ ) = 2#E+ (which can occur if V+ is a square in Z2 of a side s so that diam (V+ ) = 2s and (V+ ) = 4s (s + 1)), then (V ) 2 (V+ ) and D = diam (V ) 2m 1 . 4m (V+ )
so that the estimate of Theorem 3.5 becomes 1 &
Let us show that for this graph (V, E) a matching upped bound also holds: 1 1 . m (V+ )
3.3. EXPANSION RATE Indeed, consider a function 8 < m, if x 2 V+ , k, if x = zk , f (x) = : m, if x 2 V . X

xV
69
Then by symmetry f ?1, then
(f, f) = and (Lf, f) =
f (x)2 (x) 2m2 (V+ )
1X (f (x) f (y))2 xy 2 x,y

k=m m1 X
= whence
(f (zk ) f (zk+1 ))2 = 2m 1 . m (V+ )
1 R (f)
3.3
Expansion rate
Let (V, ) be a nite connected graph with N > 1 vertices. For any two non-empty subsets X, Y V , set d (X, Y ) = min d (x, y)
xX,yY
Note that d (X, Y ) 0 and d (X, Y ) = 0 if and only if X and Y have non-empty intersection. We will need also another quantity: l (X, Y ) = 1 (X c ) (Y c ) ln , 2 (X) (Y )
that will be applied for disjoint sets X, Y . Then X Y c and Y X c whence it follows that l (X, Y ) 0. Furthermore, l (X, Y ) = 0 if and only if X = Y c . To understand better l (X, Y ), express it in terms of the set Z = V n (X [ Y ) so that 1 (Z) (Z) l (X, Y ) = ln 1 + 1+ . 2 (X) (Y ) Hence, the quantity l (X, Y ) measures the space between X and Y in terms of the measure of the set Z. Let the eigenvalues of the Laplace operator L on (V, ) be 0 = 0 < 1 ... N1 ,
70
where N1 2. We will use the following notation: = N 1 1 =1 N1 + 1 2

N 1 1
+1
(3.12)
Clearly, 2 [0, 1), and = 0 only for complete graphs. It is useful to observe that 1 N1 = 1 =1 1+
1
2 . +1
The relation to the spectral radius = max (j1 1 j , jN1 1j) is as follows: since 1 1 and N1 1 + , it follows that 1 N1 whence . Theorem 3.6 For any two disjoint sets X, Y V , we have d (X, Y ) 1 + l (X, Y ) . ln 1 (3.13) 1 1+
If = 0 then we set by denition l(X,Y ) = 0. ln 1 Before the proof, let us discuss consequences and examples. Example. If D = d (X, Y ) > 1 then the estimate (3.13) implies that 1 1 l (X, Y ) (X c ) (Y c ) 2(D1) exp = =: A D1 (X) (Y ) whence A1 and A1 . In terms of the eigenvalues we obtain 1 N1 A1 . A+1
Since N1 2, this yields an upper bound for 1 1 2 A1 . A+1 (3.14)
Example. Let us show that diam (V ) 1 + (V ) 1 , 1 ln m ln (3.15)
3.3. EXPANSION RATE
71
where m = minxV (x). Indeed, set in (3.13) X = fxg, Y = fyg where x, y are two distinct vertices. Then l (X, Y ) whence d (x, y) 1 + (V )2 (V ) 1 ln ln 2 (x) (y) m 1 (V ) . 1 ln m ln
Taking in the left hand side the supremum in all x, y 2 V , we obtain (3.15). In a particular case of a simple weight, we have m = minx deg (x) and (V ) = 2#E whence 1 2#E diam (V ) 1 + 1 ln . m ln For any subset X V , denote by Ur (X) the r-neighborhood of X, that is, Ur (X) = fy 2 V : d (y, X) rg . Corollary 3.7 For any non-empty set X V and any integer r 1, we have (Ur (X)) (V ) 1+
(X c ) 2r (X)
(3.16)
Proof. Indeed, take Y = V nUr (X) so that Ur (X) = Y c . Then d (X, Y ) = r+1, and (3.13) yields (X c ) (Y c ) 1 1 ln r , 2 ln 1 (X) (Y ) whence (Y c ) (Y ) 2r (X) 1 . (X c )
Since (Y ) = (V ) (Y c ), we obtain
(X c ) (V ) (Y c ) 2r (Y c ) (X) whence (3.16) follows Example. Given a set X, dene the expansion rate of X to be the minimal positive integer R such that 1 (UR (X)) (V ) . 2 Imagine a communication network as a graph where the vertices are the communication centers (like computer servers) and the edges are direct links between the centers. If X is a set of selected centers, then it is reasonable to ask, how many direct links from X are required to reach the majority (at least 50%) of all centers? This is exactly the expansion rate of X, and the networks with short expansion rate provide better connectivity.
72
CHAPTER 3. GEOMETRIC BOUNDS FOR THE EIGENVALUES The inequality (3.16) implies an upper bound for the expansion rate. Indeed, if (X c ) 2r 1, (X) (3.17)
then (3.16) implies that (Ur (X)) 1 (V ). The condition (3.17) is equivalent to 2 1 ln (X) r , 2 ln 1 from where we see that the expansion rate R of X satises 1 ln (X) . R 2 ln 1 Hence, a good communication network should have the number as small as possible (which is similar to the requirement that should be as small as possible for a fast convergence rates to the equilibrium). For many large practical networks, one has the following estimate for the spectral radius: ln1N (recall that 1 and N1 are contained in the interval (1 , 1 + )). Since , it follows that 1 & ln N. Assuming that X consists of a single vertex and that (X ) N , we obtain the (X) following estimate of the expansion rate of a single vertex: R. ln N . ln ln N
c
(X c )
(3.18)
(X c )
For example, if N = 108 , which is a typical gure for the internet graph, then R 6 (although neglecting the constant factors). This very fast expansion rate is called a small world phenomenon, and it is actually observed in large communication networks. The same phenomenon occurs in the coauthor network : two mathematicians are connected by an edge if they have a joint publication. Although the number of recorded mathematicians is quite high ( 105 ), a few links are normally enough to get from one mathematician to a substantial portion of the whole network. Proof of Theorem 3.6. As before, denote by F the space of all real-valued functions on V . Let w0 , w1 , ...., wN1 be an orthonormal basis in F that consists of the eigenfunctions of L, and let their eigenvalues be 0 = 0, 1 , ..., N 1 . Any function u 2 F admits an expansion in the basis fwl g as follows: u=
N1 X l=0
al wl
(3.19)
with some coecients al . We know already that a0 w0 = u = 1 X u (x) (x) (V ) xV
3.3. EXPANSION RATE (see the proof of Theorem 2.4). Denote u =uu=
0 N1 X l=1
73
al wl
so that u = u + u0 and u0 ?u. Let () be a polynomial with real coecient. We have (L) u = =
N1 X l=0 N1 X l=0
al (L) wl al (l ) wl
N1 X l=1
= (0) u +
al (l ) wl .
If v is another function from F with expansion v= then ( (L) u, v) = ( (0) u, v) +

N 1 X l=1 N1 X l=0
bl wl = v +
N1 X l=1
bl wl = v + v0 ,
al bl (l )
N1 X l=1 0
(0) uv (V ) max j (l )j
1lN1 1lN1
jal j jbl j (3.20)
(0) uv (V ) max j (l )j ku k kv0 k . Assume now that supp u X, supp v Y and that D = d (X, Y ) 2
(if D 1 then (3.13) is trivially satised). Let () be a polynomial of of degree D 1. Let us verify that ( (L) u, v) = 0. (3.21) Indeed, the function Lk u is supported in the k-neighborhood of supp u, whence it follows that (L) u is supported in the (D 1)-neighborhood of X. Since UD1 (X) is disjoint with Y , we obtain (3.21). Comparing (3.21) and (3.20), we obtain max j (l )j (0) uv (V ) . ku0 k kv 0 k (3.22)
1lN 1
Let us take now u = 1X and v = 1Y . We have u= (X) , (V ) kuk2 = (X)2 , (V ) kuk2 = (X) ,
74 whence
p ku0 k = kuk2 kuk2 =
(X)
(X) = (V )
(X) (X c ) . (V )
Using similar identities for v and substituting into (3.22), we obtain s (X) (Y ) max j (l )j (0) . 1lN 1 (X c ) (Y c ) Finally, let us specify () as follows: () = D1 1 + N1 . 2 D1
Since max j ()j on the set 2 [1 , N1 ] is attained at = 1 and = N1 and

[1 ,N 1 ]
max j ()j =
N1 1 2
it follows from (3.30) that N1 1 2 D1 N1 + 1 2 D1 s (X) (Y ) . (X c ) (Y c )
Rewriting this inequality in the form D1 1 exp (l (X, Y )) and taking ln, we obtain (3.13). The next result is a generalization of Theorem 3.6. Set k = N1 k . N1 + k
Theorem 3.8 Let k 2 f1, 2, ..., N 1g and X1 , ..., Xk+1 be non-empty disjoint subsets of V , and set D = min d (Xi , Xj ) .
i6=j
Then D 1+
1 max l (Xi , Xj ) . ln 1k i6=j
(3.23)
Example. Dene the k-diameter of the graph by diamk (V ) =

{x1 ,...,xk+1 } i6=j
max
min d (xi , xj ) .
3.3. EXPANSION RATE
75
In particular, diam1 is the usual diameter of the graph. Applying Theorem 3.8 to sets Xi = fxi g and maximizing over all xi , we obtain diamk (V ) 1 + ln (V ) m , ln 1k (3.24)
where m = inf xV (x) . We precede the proof of Theorem 3.8 by a lemma. Lemma 3.9 In any sequence of n + 2 vectors in n-dimensional Euclidean space, there are two vectors with non-negative inner product. Note that n + 2 is the smallest number for which the statement of Lemma 3.9 is true. Indeed, if e1 , e2 , ..., en denote an orthonormal basis in the given space, let us set v := e1 e2 ... en . Then any two of the following n + 1 vectors e1 + v, e2 + v, ...., en + v, v have a negative inner product, provided > 0 is small enough. Proof. Induction in n. The inductive basis for n = 1 is obvious. The inductive step from n 1 to n is shown on Fig. 3.1. Indeed, assume that there are n + 2 vectors v1 , v2 , ..., vn+2 such that (vi , vj ) < 0 for all distinct i, j. Denote by E the 0 orthogonal complement of vn+2 , and by vi the orthogonal projection of vi onto E.
v3 v2 v1
v1
E
v 3 v 2
vn+2
Figure 3.1:
0 The dierence vi vi is orthogonal to E and, hence, colinear to vn+2 so that 0 vi vi = i vn+2
(3.25)
for some constant i . Multiplying this identity by vn+2 we obtain

0 (vi , vn+2 ) (vi , vn+2 ) = i (vn+2 , vn+2 ) .
76
0 Since (vi , vn+2 ) = 0 and (vi , vn+2 ) < 0, it follows that
i =
(vi , vn+2 ) > 0. (vn+2 , vn+2 )
Rewriting the identity (3.25) in the form

0 vi = vi i vn+2 ,
0 0 By the inductive hypothesis, out of n + 1 vectors v1 , ..., vn+1 in (n 1)-dimensional Euclidean space E, there are two vectors with non-negative inner product, say, 0 0 vi , vj 0. It follows from (3.26) that also (vi , vj ) 0, which nishes the proof. Proof of Theorem 3.8. We use the same notation as in the proof of Theorem 3.6. If D 1 then (3.23) is trivially satised. Assume in the sequel that D 2. Let ui be a function on V with supp ui Xi . Let () be a polynomial with real coecients of degree D 1, which is non-negative for 2 f1 , ..., k1 g . Let us prove the following estimate
we obtain, for all distinct i, j = 1, ..., n + 1 0 0 (vi , vj ) = vi , vj + i j (vn+2 , vn+2 ) .
(3.26)
klN1
max j (l )j (0) min

i6=j
ui uj (V ) . ku0i kku0j k
(3.27)
Expand a function ui in the basis fwl g as follows: ui = It follows that ( (L) ui , uj ) = (0) ui uj (V ) + (0) ui uj (V ) + max
klN1 k1 X l=1 k1 X (i) (j) (l ) al al N1 X l=0 (i) al wl
= ui +
k1 X l=1
(i) al wl
N1 X l=k
al wl .
(i)
N1 X l=k
(l ) al al
(i) (j)
(l ) al al
(i) (j)
Since also ( (L) ui , uj ) = 0, it follows that max j (l )j ku0i k u0j
l=1 j (l )j ku0i k u0j . k1 X l=1
klN1
(0) ui uj (V ) +
(l ) al al .
(i) (j)
(3.28)
In order to be able to obtain (3.27), we would like to have

k1 X l=1
(l ) al al 0.
(i) (j)
(3.29)
3.3. EXPANSION RATE
77
dimensional Euclidean space. By Lemma 3.9, there are two vectors, say a(i) and a(j) such that a(i) , a(j) 0, which exactly means (3.29). For these i, j, we obtain from (3.28) ui uj (V ) max j (l )j (0) 0 0 klN1 kui k uj whence (3.27) follows. In particular, taking ui = 1Xi and using that ui = and ku0i k = we obtain
klN1
for any two vectors a = (a1 , ..., ak1 ) and b = (b1 , ..., bk1 ). Also, consider the vectors (i) (i) (i) a = a1 , ..., ak1 for i = 1, ..., k + 1. Hence, we have k + 1 vectors in (k 1)-
In general, (3.29) cannot be guaranteed for any couple i, j but we claim that there exists a couple i, j of distinct indices such that (3.29) holds (this is the reason why in (3.27) we have min in all i, j). To prove that, consider the inner product in Rk1 given by3 k1 X (a, b) = (i ) ai bi ,
i=1
(Xi ) (V )
(Xi ) (Xic ) , (V ) s (Xi ) (Xj ) c . (Xic ) Xj
max j (l )j (0) min

i6=j
(3.30)
Consider the following polynomial of degree D 1 D1 k + N 1 () = , 2
which is clearly non-negative for k . Since max j ()j on the set 2 [k , N1 ] is attained at = k and = N 1 and
[k ,N 1 ]
max
j ()j =
N1 k 2
D1 s
it follows from (3.30) that N1 k 2 D1 N1 + k 2 D1 min

i6=j
whence (3.23) follows.
(Xi ) (Xj ) c , (Xic ) Xj
3 By hypothesis, we have (i ) 0. If (i ) vanishes for some i then use only those i for which (i ) > 0 and consider inner product in a space of smaller dimension.
78
Chapter 4 Eigenvalues on innite graphs

Let (V, ) be a locally nite connected weighted graph with innite (countable) set Lecture 14 21.12.09 of vertices.
A.Grigoryan
4.1
Dirichlet Laplace operator
Given a nite subset V , denote by F the set of functions ! R. Then F is a linear space of dimension N = #. Dene the operator L on F as follows: rst extend f to the whole V by setting f = 0 outside and then set L f = (Lf) j . In other words, for any x 2 , we have L f (x) = f (x) where f (y) is set to be 0 whenever y 2 . / X
yx
P (x, y) f (y)
Denition. The operator L is called the Dirichlet Laplace operator in . Example. Recall that the Laplace operator in the lattice graph Z2 is dened by 1X f (y) . Lf (x) = f (x) 4 yx
1 L f (a) = f (a) f (b) 4 1 L f (b) = f (b) (f (a) + f (c)) 4 1 L f (c) = f (c) f (b) . 4 Consequently, the matrix of L is 0 1 1 1/4 0 @ 1/4 1 1/4 A 0 1/4 1 79
Let be the subset of Z2 that consists of three vertices a = (0, 0), b = (1, 0), c = (2, 0) , so that a b c. Then we obtain for L the following formulas:
80
CHAPTER 4. EIGENVALUES ON INFINITE GRAPHS
p and the eigenvalues are 1, 1 1 2. 4 For comparison, consider as a nite graph itself. Then deg (a) = deg (c) = 1 and deg (b) = 2 so that the Laplace operator on as a nite graph is dened by Lf (a) = f (a) f (b) 1 Lf (b) = f (b) (f (a) + f (c)) 2 Lf (c) = f (c) f (b) . The matrix of L is 0 1 1 1 0 @ 1/2 1 1/2 A 0 1 1
the eigenvalues are 0, 1, 2. As we see, the Dirichlet Laplace operator of as a subset of Z2 and the Laplace operator of as a graph are dierent operators with dierent spectra. Returning to the general setting, introduce in F the inner product (f, g) = X
x
f (x) g (x) (x) .
Lemma 4.1 (Greens formula) For any two functions f, g 2 F , we have (L f, g) = where 1 = U1 () . Proof. Extend both functions f and g to all V as above. Applying the Green formula of Theorem 2.1 in 1 and using that g = 0 outside , we obtain (L f, g) = = X Lf (x) g (x) (x) (4.2) 1 X (rxy f) (rxy g) xy , 2 x,y
1
(4.1)
X 1 X (rxy f ) (rxy g) xy (rxy f ) g(x)xy 2 x,y c x1 ,y1 1 X 1 = (rxy f ) (rxy g) xy . 2 x,y

1
x1
We have used the fact that the last sum in (4.2) vanishes. Indeed, the summation can be restricted to neighboring x, y. Therefore, if y 2 c then necessarily x 2 c 1 and g (x) = 0. Since the right hand side of (4.1) is symmetric in f, g, we obtain the following consequence. Corollary 4.2 L is a symmetric operator in F .
4.1. DIRICHLET LAPLACE OPERATOR
81
Hence, the spectrum of L is real. Denote the eigenvalues of L in increasing order by 1 () 2 () ... N (), where N = #. As for every symmetric operator, the smallest eigenvalue 1 () admits the variational denition: 1 () =
f F \{0}
inf
R (f) ,
(4.3)
Here the second equality is true by Lemma 4.1. Note that the ranges x 2 and x, y 2 1 of summations in (4.4) can be extended to x 2 V and x, y 2 V respectively, because the contributions of each additional term is 0. Indeed, for the denominator it is obvious because f (x) = 0 outside . For the numerator, if x 2 1 and y x / then y 2 so that f (x) = f (y) = 0 and rxy f = 0. / Theorem 4.3 Let be a nite non-empty subset of V . Then the following is true. (a) 0 < 1 () 1. (b) 1 () + N () 2.Consequently, spec L [1 () , 2 1 ()] (0, 2) . (c) 1 () decreases when increases. Proof. (a) Let f be the eigenfunction of 1 (). Then we have P 2 1 (L f, f ) x,y1 jrxy f j xy 2 1 () = = P , 2 (f, f ) x f (x)(x) (4.5)
where the Rayleigh quotient R (f ) is dened by P 2 1 (L f, f) x,y1 (rxy f) xy 2 P = . R (f) = 2 (f, f ) x f (x)(x)
(4.4)
(4.6)
which implies 1 () 0. Let us show that 1 () > 0. Assume from the contrary that 1 () = 0. It follows from (4.6) that rxy f = 0 for all neighboring vertices x, y 2 1 , that is, for such vertices we have f (x) = f (y). Fix a vertex x 2 . Since is nite and V is innite, the complement c is non-empty, let y 2 c . Since (V, ) is connected, there is a path connecting x and y, say fxi gn where x0 = x and i=0 xn = y. Then let k be the minimal index such that xk 2 c (such k exists because x0 2 while xn 2 c ). Since xk1 2 and xk1 xk , it follows that xk 2 1 . Hence, all the vertices in the path x0 x1 ... xk1 xk belong to 1 whence we conclude by the above argument that f (x0 ) = f (x1 ) = ... = f (xk ) . Since f (xk ) = 0 it follows that f (x) = f (x0 ) = 0. Hence, f 0 in , which contradicts the denition of an eigenfunction. This proves that 1 () > 0. To prove that 1 () 1, we use the trace of the operator L . On the one hand, trace (L ) = 1 () + ... + N () N1 () . (4.7)
82
On the other hand, since L f(x) = f (x) X

y
P (x, y)f(y),
the matrix of the operator L in the basis 1{x} x has the following values on the diagonal: 1 P (x, x) . It follows that X trace (L ) = (1 P (x, x)) N. (4.8)
x
Since
Applying (4.3) to the function jfj , we obtain P 2 1 x,yV (rxy jfj) xy 2 P . 1 () R (jfj) = 2 xV f (x)(x)
Comparing (4.7) and (4.8), we obtain 1 () 1. (b) Let f be an eigenfunction with the eigenvalueN (). Then we have similarly to (4.6) P 2 1 x,yV (rxy f) xy 2 . N () = R (f ) = P 2 xV f (x)(x)
it follows that
(rxy f )2 + (rxy jfj)2 = (f (x) f (y))2 + (jf(x)j jf (y)j)2 2 f 2 (x) + f(y)2 (f 2 (x) + f(y)2 ) xy P 1 () + N () 2 xV f (x)(x) P P 2 xV yV f 2 (x)xy P = f 2 (x)(x) P xV 2 f (x) (x) 2 = 2. = P xV 2 xV f (x)(x)
x,yV
This together with part (a) implies (4.5). (c) If increases then the space F also increases. Clearly, the inmum of the functional R (f) over a larger set must be smaller. Hence, by (4.3), 1 () decreases.
4.2
Cheegers inequality
= fxy 2 E : x 2 , y 2 c g .
Recall that, for any subset of V , the edge boundary is dened by
Also, for any subset S of the edge set E, its measure is dened by X (S) = .
S
4.2. CHEEGERS INEQUALITY Denition. For any nite subset V , dene its Cheeger constant by h () = inf (U) , U (U)
83
where the inmum is taken over all non-empty subsets U of . In other words, h () is the largest constant such that the following inequality is true (U) h () (U) (4.9) for any non-empty subset U of . Theorem 4.4 (Cheegers inequality) We have 1 1 () h ()2 . 2 The proof is similar to the case of nite graphs. We start with the following lemma. Lemma 4.5 For any non-negative function f 2 F , the following is true: X X jr f j h () f (x) (x) .
E xV
(4.10)
Proof. By the co-area formula of Lemma 3.3, we have Z X jr f j (Ut ) dt,

E 0
where Ut = fx 2 V : f (x) > tg . Since Ut for non-negative t, we obtain by (3.2) (Ut ) h () (Ut ) whence X
E
jr f j h ()
(Ut ) dt
On the other hand, as in the proof of Lemma 3.4, we have Z X (Ut ) dt = f (x) (x) ,
0 xV
which implies (4.10). Proof of Theorem 4.4. in the form
Hence, to prove (4.10), it suces to verify that X

2
Let f be the eigenfunction of 1 (). Rewrite (4.6) P 2 E jr fj 1 () = P . 2 xV f (x) (x) (4.11)
h ()2 X 2 jr f j f (x) (x) . 2 xV E
84
The same computation as in the proof of Theorem 3.2 shows that X r f 2

E
Applying (4.10) to function f 2 , we obtain X X r f 2 h () f 2 (x) (x) .

E xV
(4.12)
X
E
jr fj2
X
xV
f 2 (x) (x)
!1/2
Combining this with (4.12) yields h () X

xV
f 2 (x) (x)
X
E
jr fj2
!1/2 X
xV
f 2 (x) (x)
!1/2
Dividing by
xV
1/2 f 2 (x) (x) and taking square, we obtain (4.11).
4.3
Isoperimetric inequalities
Denition. We say that the graph (V, ) satises the isoperimetric inequality with function (s) if, for any nite non-empty subset V , () ( ()) . (4.13)
We always assume that the function (s) is non-negative and is dened for all s inf (x)
xV
(4.14)
so that the value () is in the domain of for all non-empty subsets V . Example. If is a simple weight then (V, ) always satises the isoperimetric inequality with function (s) 1. Indeed, any nite subset has at least one edge connecting with c (because of the connectedness of (V, E)). For the lattice graph Z the sharp isoperimetric function is (s) 2. As we will show later on, Zm satises the isoperimetric inequality with function m1 (s) = cm s m with some constant cm > 0. The relation between isoperimetric inequalities and the Dirichlet eigenvalues is given by the following theorem. Theorem 4.6 Assume that (V, ) satises the isoperimetric inequality with function (s) such that (s) /s is decreasing in s. Then, for any nite non-empty subset V, 1 () ( ()) (4.15) 2 where (s) = 1 (s) . 2 s
4.4. SOLVING THE DIRICHLET PROBLEM BY ITERATIONS
85
Denition. We say that (V, ) satises the Faber-Krahn inequality with function (s) if (4.15) holds with this function for any nite non-empty subset V . We always assume that (s) is non-negative and has the domain (4.14). Hence, the isoperimetric inequality with function (s) implies the Faber-Krahn 2 1 (s) inequality with function (s) = 2 s . Example. Any graph with a simple weight satises the Faber-Krahn inequality 1 with function (s) = 2s2 . The lattice graph Zm satises the Faber-Krahn inequality with function (s) = 0 2/m cm s . Proof of Theorem 4.6. We have, for any non-empty subset U , (U) ( (U )) = It follows that h () whence by Theorem 4.4 1 1 1 () h ()2 2 2 ( ()) () 2 = ( ()) . ( ()) ( (U )) (U) (U) . (U ) () ( ()) , ()
4.4
Solving the Dirichlet problem by iterations

L u = f (4.16)
Let be a nite non-empty subset of V . Consider the Dirichlet problem
where f 2 F is a given function and u 2 F is to be found. This problem was considered in Section 1.5 and it was shown in Theorem 1.10 that it has a unique solution. Now we can prove it much easier. Indeed, the operator L has the spectrum in (0, 2). In particular, 0 is not in the spectrum, which means that this operator is invertible, which is equivalent to the solvability of (4.16). In what follows we will use the notion of the norm of a linear operator. Let V be a nite dimensional linear space with inner product (, ). Recall that the norm p of any vector v 2 V is dened then by kvk = (v, v). Given a linear operator A : V ! V, dene its norm by kAk = sup kAvk . vV\{0} kvk
In other words, kAk is the smallest value of a constant C that satises the inequality kAvk C kvk for all v 2 V. It is easy to verify the following properties of the norm (although we will not use them):
86
CHAPTER 4. EIGENVALUES ON INFINITE GRAPHS 1. kA + Bk kAk + kBk 2. kAk = jj kAk 3. kABk kAk kBk For symmetric operators, there is the following relation to the spectrum:
Lemma 4.7 If the operator A is symmetric then kAk = max jj .

spec A
(4.17)
Proof. We have kAk2 = = The expression R (v) = (A2 v, v) (v, v) (Av, Av) vV\{0} (v, v) sup (A2 v, v) . vV\{0} (v, v) sup
is the Rayleigh quotient of the operator A2 , and sup R (v) is the maximal eigenvalue of A2 . The eigenvalues of A2 have the form 2 where is an eigenvalue of A. Hence, the maximal eigenvalue of A2 is max (2 ) = (max jj)2 where runs over all eigenvalues of A. It follows that kAk2 = (max jj)2 , whence (4.17) follows. Let us investigate the rate of convergence in Jacobis method of solving (4.16). Set P = id L and rewrite (4.16) is the form u = P u + f. (4.18)
Let us construct successive approximations un 2 F to the solution as follows: u0 = 0 and un+1 = P un + f. (4.19)
Theorem 4.8 The following inequality is true for all positive integers n: kun uk n kuk where = 1 1 (). Consequently, un ! u as n ! 1.
4.4. SOLVING THE DIRICHLET PROBLEM BY ITERATIONS Proof. Indeed, subtracting (4.18) from (4.19), we obtain whence un+1 un = P (un u) , kun+1 uk kP k kun uk .
87
(4.20)
The eigenvalues of P are k = 1 k () . The sequence fk g is decreasing, and by Theorem 4.3 we have that 0 1 < 1 (follows from 0 < 1 () 1) 1 + N 0 (follows from 1 () + N () 2), that is, N 1 . Hence, we have spec P [N , 1 ] [1 , 1 ] which implies max jk j = 1 . Denoting 1 by , we obtain from (4.20) that whence it follows by induction that kun+1 un k kun uk ,
which was to be proved. Since < 1, we obtain that indeed kun uk ! 0 as n ! 1, that is, un ! u. Given > 0, dene the rate of convergence of fun g to u by that is, T = min fn : n g , T
kun uk n ku0 uk = n kuk ,
ln 1 1. ln In other words, after T iterations in Jacobis method, we obtain that The value of should be chosen to have << 1 so that kun uk is a small fraction of kuk and un can be considered as a good approximation for u. It is standard to set = 1 . With this convention, we obtain e T 1 1 = ln 1
ln
1 11 ()
kun uk kuk .
Since ln 11 () 1 () (indeed, it follows from e 1 ) we obtain the upper 1 bound for the convergence rate 1 . (4.21) T 1 () Example. Let be a non-empty nite subset of Zm and N = #. Since Zm satises the Faber-Krahn inequality with function (s) = cm s2/m , it follows that for some positive constant c = c (m). Therefore, we obtain T CN 2/m . 1 () cN 2/m ,
88
4.5
Isoperimetric inequalities on Cayley graphs
Let G be an innite group G and S be a nite symmetric generating set of G. Let (V, E) be the Cayley graph (G, S) and be a simple weight on (V, E). Recall that the degree of every vertex in (V, E) is jSj so that (x) = jSj (here jAj is the cardinality of a set A). Note also that jSj 2. Let e be the neutral element of G, and dene the balls centered at e by Br = fx 2 V : d (x, e) rg for any r 0. Theorem 4.9 Let V (r) be a continuous non-negative strictly increasing function on [0, +1) such that V (s) ! 1 as s ! 1. Assume that, for all non-negative integers r, (Br ) V (r) . (4.23) Then the Cayley graph (G, S) satises the isoperimetric inequality with function (s) = c0 where c0 =
1 4|S|
1 +1 V 1 (2|S|)
(4.22)
V 1
s , s jSj , (2s)
(4.24)
Remark. Since V (0) (B0 ) = (e) = jSj, the function V 1 (s) is certainly dened on [jSj , +1). Hence, the function V 1 (2s) is dened on [jSj , +1) and is strictly positive on this interval. It follows that the domain of (s) contains [jSj , +1) and, consequently, all the values () for all non-empty subsets V . Example. In Zm we have (Br ) ' rm for r 1 so that we can take V (r) = crm . Then V 1 (s) = c0 s1/m and we conclude that Zm satises the isoperimetric inequality with function m1 s (s) = c00 1/m = c00 s m , s 00 where c is a positive constant depending on m. Example. There is a large class of groups with the exponential volume growth, that is, the groups where (4.23) holds with V (r) = exp (c0 r). For such groups Theorem s 4.9 guarantees the isoperimetric inequality with function (s) = c ln s . Combining Theorems 4.6 and 4.9, we obtain the following: Corollary 4.10 Under the conditions of Theorem 4.9, the Cayley graph (G, S) satises the Faber-Krahn inequality with the function 1 (s) = c 1 (2s) V with some positive constant c > 0. 2
4.5. ISOPERIMETRIC INEQUALITIES ON CAYLEY GRAPHS
89
In Zm we obtain as above (s) = cs2/m . On the groups with exponential volume growth, we have (s) = c (ln (2s))2 . Proof of Theorem 4.9. Let us introduce some notation that will be used only in this proof. For any function f on V with nite support, set X kfk := jf (x)j .
xV
Also, consider jr f j as a function of 2 E, and set krfk := X

E
jr fj =
1 X jf (x) f (y)j . 2 x,yV : xy
We will obtain certain lower bounds for krf k which will then allow to obtain a lower estimate for (). Indeed, for f = 1 we have 1, 2 jr fj = . 0, 2 / It follows that krf k = jj = () , where we have used that = 1 for any edge . For any z 2 G, consider the function fz on G that is dened by fz (x) = f (xz) . That is, fz is a spatial shift of f by the right multiplication by z. Claim 1. If s 2 S then kf fs k 2 krf k . Recall that the edges x y are determined by the relation x1 y 2 S that is equivalent to y = xs for some s 2 S. Hence, for any s 2 S, we have x xs and X X kf fs k = jf (x) f (xs)j jf (x) f (y)j = 2 krf k . (4.26)
xV x,yV : xy
(4.25)
Claim 2. If z 2 Bn then kf fz k 2n krfk. Any z 2 Bn can be represented in the form z = s1 s2 ...sk where si 2 S and k n. Then we have X kf fz k = jf (x) f (xz)j X
xV xV
jf (x) f (xs1 )j
+... X jf (xs1 ...sk1 ) f (xs1 ...sk )j . +

xV
X
xV
jf (xs1 ) f (xs1 s2 )j
90
The rst sum on the right hand side is bounded by 2 krfk as in (4.26). In the second sum make change y = xs1 so that it becomes X jf (y) f (ys2 )j ,
yV
which is bounded by 2 krfk as in (4.26). In the same way, we have, for any i = 1, ..., k, X X jf (xs1 ...si1 ) f (xs1 ...si )j = jf (y) f (ysi )j 2 krfk .
xV yV
Hence, it follows that kf fz k 2k krfk 2n krf k . Claim 3. For any positive integer n and any function f on V with nite support, dene a new function An f (averaging over n-balls) on V by X 1 f (y) . An f (x) = jBn j
{y:d(x,y)n}
Then the following inequality is true: kf An fk 2n krf k . (4.27)
The condition d (x, y) n means that there is a path between x and y of at most n edges. Since moving along an edge means the right multiplication by some s 2 S, we obtain that y = xs1 ...sk for some k n and s1 , ..., sk 2 S. Setting z = s1 ...sk we obtain that d (x, y) n is equivalent to y = xz for some z 2 Bn . Hence, we have X kf An fk = jf (x) An f (x)j = = = = X 1 X f (xz) f (x) jBn j zB xV n X 1 X (f (x) f (xz)) jBn j xV zBn X 1 X jf (x) f (xz)j jBn j zB xV n 1 XX jf (x) f (xz)j jBn j zB xV n 1 X kf fz k jBn j zB
n
xV
2n krfk ,
4.5. ISOPERIMETRIC INEQUALITIES ON CAYLEY GRAPHS
91
where in the last line we have used Claim 2. Claim 4. Let be a non-empty nite subset of V , and n be a positive integer such that1 jBn j 2 jj. Then we have 1 () . (4.28) () 4 jSj n Set f = 1 . Then we have, for any x 2 V , X 1 An f (x) = f (y) jBn j {y:d(x,y)n} X 1 f (y) jBn j yV = It follows that kf An f k X
x
1 1 jj . jBn j 2
jf (x) An f (x)j
1 jj , 2
while krf k = () (cf. (4.25). Comparing the above two lines and using (4.27) we obtain 1 () jj . 4n Noticing that () = jSj jj (because the degree of any vertex is jSj), we obtain (4.28). Claim 5. For any non-empty nite set V , we have () ( ()) where is dened by (4.24). Choose n to be minimal positive integer with the property that This implies (Bn ) 2 () which is equivalent to jBn j 2 jj so that (4.28) holds. The minimality of n implies that either n = 1or n > 1 and V (n 1) < 2 () . In the both cases, we obtain that Since contains at least one vertex and the measure of this vertex is jSj, we have V 1 (2 ()) V 1 (2 jSj) =: C > 0, whence 1 + 1 V 1 (2 ()) . n C Substituting into (4.28), we obtain () which was to be proved.
1
V (n) 2 () .
n 1 + V 1 (2 ()) .
4 jSj
1 1
() 1 = c0 ( ()) , + 1 V (2 ())
Note that such n always exists because jBn j ! 1 as n ! 1.
92
Chapter 5 The heat kernel and the type problem

Recall that Pn (x, y) is the n-step transition function of the random walk on a weighted graph (V, ), that can be determined inductively as follows: P1 (x, y) = P (x, y) and X Pn (x, y) = Pn1 (x, z) P (z, y)
zV
(cf. Proposition 1.6). Furthermore, Pn (x, y) is reversible with respect to measure (x), that is Pn (x, y) (x) = Pn (y, x) (y) . Denition. The function pn (x, y) := Pn (x, y) (y)
is called the heat kernel of (V, ) (or the transition density of the random walk). The reversibility condition clearly implies that pn (x, y) is symmetric in x, y that is pn (x, y) = pn (y, x) . Other useful properties of the heat kernel that come from the corresponding properties of the transition function, are as follows: X Pn f (x) = pn (x, y) f (y) (y) ,
yV
X
yV
pn (x, y) (y) 1, X
zV
pn+m (x, y) =
pn (x, z) pm (z, y) (z) .
Lemma 5.1 We have for all x, y 2 V and n, m 2 N: pn+m (x, y) (p2n (x, x) p2m (y, y))1/2 93 (5.1)
94
CHAPTER 5. THE HEAT KERNEL AND THE TYPE PROBLEM
Proof. It follows from the symmetry and semigroup identity that !1/2 !1/2 X X 2 2 pn (x, z) (z) pm (z, y) (z) pn+m (x, y) = X
zV zV
pn (x, z) pn (z, x) (z)
zV !1/2
= (p2n (x, x) p2m (y, y))1/2 .
X
zV
pm (z, y) pm (y, z) (z)
!1/2
One of the most interesting problems on innite graphs is the rate of convergence to 0 of Pn (x, y) as n ! 1, which is equivalent to the rate of convergence to 0 of the heat kernel pn (x, y). The question amounts to obtaining upper and lower estimates of pn (x, y) for large n, that will be considered in this Chapter. Example. Let (V, ) be Z with a simple weight. Since Z is shift invariant, pn (x, y) is also shift invariant, that is, pn (x, y) = pn (x z, y z) for any integer z. Hence, it suces to understand pn (x, 0). It follows from Exercise 16 that ( n 1 , x n mod 2 2n+1 x+n 2 pn (x, 0) = (5.2) 0, otherwise, where n
m
is the binomial coecient, and 1 pn (0, 0) p as n ! 1, n even. 2n
(for odd n we have pn (0, 0) = 0). It follows that pn (0, 0) ! 0 as n ! 1. Since pn (x, x) = pn (0, 0) for any x 2 Z, it follows from (5.1) that also pn (x, y) ! 0 as n ! 1 for all x, y 2 Z.
5.1
On-diagonal upper estimates of the heat kernel
In this section (V, ) is an innite locally nite connected weighted graph that satises in addition the conditions 1 xy M for all x y, deg (x) D for all x 2 V, (5.3)
for some constants M and D. The rst condition is trivially satised for a simple weight, the second condition is always satised on Cayley graphs. Lemma 5.2 The conditions (5.3) imply that, for any non-empty nite set A V , (U1 (A)) C0 (A) , where C0 = C0 (D, M) . (5.4)
5.1. ON-DIAGONAL UPPER ESTIMATES OF THE HEAT KERNEL Proof. Since (x) = P
yx
95
xy , it follows from (5.3) that 1 (x) MD.
Therefore, for any nite set A, we have jAj (A) MD jAj . Recall that the the r-neighborhood of A is dened by Ur (A) = fy 2 V : d (x, y) r for some x 2 Ag , and the balls of radius r are dened by Br (x) = fy 2 V : d (x, y) rg . It follows that Ur (A) = S Br (x) .
xA
whence it follows that
The ball B1 (x) consists of the vertex x and of the vertices y x so that jB1 (x)j D + 1. Hence, X jU1 (A)j jB1 (x)j (D + 1) jAj ,
xA
(U1 (A)) MD (D + 1) (A) . The next theorem is the main result of this section. Theorem 5.3 If (V, ) satises (5.3) and the Faber-Krahn inequality with function (s) = cs1/ , for some , c > 0, then the following estimate is true pn (x, y) Cn . for all x, y 2 V , n 1 and some C = C (, c, C0 ) . Example. If the weight is simple then we always have the Faber-Krahn inequality 1 with function (s) = 2s2 , that is, with = 1/2. Assuming that the degree is uniformly bounded, we obtain by Theorem 5.3 that pn (x, y) Cn1/2 . In particular, pn (x, y) ! 0 as n ! 1. Note that this estimate is sharp in Z (up to a constant multiple). Example. If (V, ) is a Cayley graphs satisfying the volume growth condition (Br ) crm (5.6) (5.5)
96
then we have the Faber-Krahn inequality with the function (s) = c0 s2/m whence Since (5.6) is satised in Zm , we see that the estimate (5.7) holds in Zm . As we will see later on, the power nm/2 is sharp here. Proof. Let F be the set of all functions on V with nite support, where supp f = fx 2 V : f (x) 6= 0g . Clearly, F is a linear space of innite dimension. Observe that f 2 F implies that Lf and P f belong to F, because supp (P f) U1 (supp f) . Consider in F the inner product (f, g) = X pn (x, y) Cnm/2 . (5.7)
f (x) g (x) (x) .
(5.8)
xV
The approach to the proof is as follows. For a xed z 2 V , denote fn (x) = pn (x, z). We will show that fn+1 = P fn . Set X pn (x, z)2 (x) = p2n (z, z) . bn := (fn , fn ) =
xV
We will show that fbn g is a decreasing sequence and will estimate the dierence bn bn+1 = (fn , fn ) (P fn , P fn ) , which will imply an upper bound for bn and, hence, for p2n (z, z) . Then Lemma 5.1 will allow to estimate pn (x, y) for all x, y 2 V. The technical implementation of this approach is quite long and will be split into a series of Claims. The denition (5.8) of (f, g) remains valid provided only one of the functions f, g has a nite support, and the other is an arbitrary function on V . In particular, for any f 2 F, we write X (f, 1) = f (x) (x) .
xV
Claim 0. If f 2 F then (P f, 1) = (f, 1) . Indeed, using the Green formula of Theorem 2.1 in = U1 (supp f), we obtain (f, 1) (P f, 1) = (Lf, 1) X Lf (x) 1 (x) (x) =
x
XX 1 X (rxy f) (rxy 1) xy (rxy f ) xy . 2 x,y x yc
5.1. ON-DIAGONAL UPPER ESTIMATES OF THE HEAT KERNEL
97
The rst sum is 0 because rxy 1 = 0. In the second sum, y 2 and x y imply / that x 2 supp f whence rxy f = 0 so that the second sum is also 0, which proves / the claim. Consider now the following functional Q (f, g) = (f, g) (P f, P g) , that is dened for all f, g 2 F. Also, we write Q (f) = Q (f, f ) . Claim 1. If is a nite non-empty subset of V , f 2 F and U1 (supp f ) then Q (f) 1 () (f, f ) . In particular, Q (f) 0 for any f 2 F. Clearly, supp (P f) so that P f = P f where P = id L . Let 1 = 1 1 () so that spec P [1 , 1 ] and kP k 1 . Then Q (f ) = = (f, f) (P f, P f ) kf k2 2 kfk2 1 (1 1 ) (1 + 1 ) kfk2 1 () kf k2 .
Claim 2. Let g, h 2 F and a real constant c be such that g 0 on V , h = c on supp g, h c outside supp g. Then Q (h, g) 0. We have Q (h, g) = (h, g) P 2 h, g = h P 2 h, g . = P2 (x, y) (x) . xy The measure (x) on vertices, that is determined by the weight , coincides with xy (x) because X X = P2 (x, y) (x) = (x) . (x) = xy
yV yV
Note that P 2 is a Markov kernel, reversible with respect to measure (x). As was explained in Section 1.3, P 2 induces a weight on V that is dened by xy
Let L = id P 2 be the Laplace operator of (V, ). Setting = supp g, we obtain by the Green formula on (V, ): X (L h, g) = L h(x)g(x)(x)
x
XX 1 X (rxy h) (rxy g) (rxy h) g(x) . xy xy 2 x,y x yc
98
The rst sum is equal to 0 because h (x) = h (y) = c in . The second sum is nonpositive because for all x 2 and y 2 c , we have g (x) 0 and h (x) = c h (y) whence rxy h 0. It follows that Q (h, g) = h P 2 h, g = (L h, g) 0. Claim 3. If f 2 F and c is a positive constant then Q (f c)+ Q (f) . Set g = (f c)+ and h = f g so that f = g+h. Since Q is a bilinear functional, we have Q (f) = Q (g + h) = Q (g) + Q (h) + 2Q (h, g) . Note that g 0. Let us verify that h satises the conditions of Claim 2. Indeed, if x 2 supp g, that is, g (x) > 0 then f (x) > c, g (x) = f (x)c and, hence, h (x) = c. If x 2 supp g, that is, g (x) = 0 then f (x) c and, hence, h (x) = f (x) g (x) c.By / Claim 2, we conclude that Q (h, g) 0. By Claim 1, we have Q (h) 0 whence it follows that Q (f ) Q (g) . Claim 4. Let f be a non-negative function from F. For any s 0 dene the set s by s = U1 supp (f s)+ . Then1 In particular, for Q (f ) 1 (s ) ((f, f) 2s (f, 1)) . s= we obtain 1 (f, f) , 4 (f, 1) (5.9)
1 Q (f ) 1 (s ) (f, f) . 2
Set g = (f s)+ . By Claims 1 and 3, we have Q (f) Q (g) 1 (s ) (g, g) . On the other hand, we have Indeed, if f s then g = f s and g 2 f 2 2sf. (5.10)
g 2 = f 2 2sf + s2 f 2 2sf, and if f < s then g = 0 and f 2 2sf = (f 2s) f 0.Summing up (5.10) against measure (x), we obtain (g, g) (f, f) 2s (f, 1) , whence (5.9) follows. Claim 5. Let ffn g be a sequence of non-negative functions on V such that n=0 f0 2 F, (f0 , 1) = 1, and fn+1 = P fn . Set bn = (fn , fn ) .
1
Note that, for s = 0; (5.9) coincides with Claim 1.
5.1. ON-DIAGONAL UPPER ESTIMATES OF THE HEAT KERNEL Then
99
where c0 = 1 c (4C0 )1/ . 2 By induction, we obtain that fn 2 F and (fn , 1) = 1 (by Claim 0). Note that bn bn+1 = (fn , fn ) (P fn , P fn ) = Q (fn ) . Estimating Q (fn ) by Claim 4 and choosing s= we obtain where On the other hand, we have 1 (fn , fn ) 1 = bn , 4 (fn , 1) 4 (5.12)
bn bn+1 c0 b1+1/ , n
(5.11)
1 bn bn+1 1 (s ) bn , 2 s = U1 supp (fn s)+ .
1X 1 1 supp (fn s)+ = (x 2 V : fn (x) > s) fn (x) (x) = (fn , 1) = . s xV s s By Lemma 5.2, we obtain that (s ) Hence, by the Faber-Krahn inequality, 1 (s ) c (s )1/ c (4C0 )1/ b1/ , n (5.13) 4C0 C0 = . s bn
which together with (5.12) yields (5.11). A.Grigoryan Claim 6. If fbn gn=0 is a sequence of positive real numbers satisfying (5.11) then Lecture 18 18.01.10 bn C 0 n where C 0 = (/c0 ) . We use an elementary inequality y x (x y) , x+1 (5.14)
that is true for all > 0 and x > y > 0. Indeed, by the mean-value theorem, we have y x y x = = 1 xy yx
1 where 2 (y, x), whence (5.14) follows. Applying (5.14) with = , we obtain 1/ bn+1
b1/ n
bn bn+1 bn
1+1/
c0 bn bn
1+1/
1+1/
c0 .
1/
Summing up this inequality from 0 to n, we conclude that bn C 0 n .
c0 n
and bn
100
1 Now we can nish the proof as follows. Fix a vertex z 2 V and set f0 = (z) 1{z} . Then f0 2 F and (f0 , 1) = 1. Dene the sequence ffn g inductively by fn+1 = P fn and show that, in fact,
fn (x) = pn (x, z) for any n 1. We have f1 (x) = P f0 (x) = and fn+1 (x) = = X
yV yV
X
yV
P (x, y) f0 (y) =
P (x, z) = p1 (x, z) (z)
P (x, y) fn (y) p1 (x, y) pn (y, z) (y)
= pn+1 (x, z) . The sequence ffn g satises the hypotheses of Claim 5. Setting bn = (fn , fn ) = p2n (z, z) , we obtain by Claims 5,6 that p2n (z, z) C 0 n , for all z 2 V . Using Lemma 5.1 and (5.15), we obtain that pk+l (x, y) (p2k (x, x) p2l (y, y))1/2 C 0 (kl)/2 , (5.16) (5.15)
for all x, y 2 V and positive integers k, l. Given an integer n 2, represent it in the form n = k + l where l = k for even n and l = k + 1 for odd n. In the both cases, we have n1 n lk , 2 4 whence by (5.16) pn (x, y) C 00 n .
(x,y) Finally, for n = 1 we obtain p1 (x, y) = P(y) 1 because P (x, y) 1 and (y) 1 by (5.3). As we have seen in the last part of the proof, the estimate (5.5) is equivalent to the on-diagonal estimate pn (x, x) Cn .
For that reason, (5.5) is also frequently referred to as an on-diagonal estimate of the heat kernel. The point is that this estimate does not take into account the distance between points x, y, which could improve the estimate. Indeed, if d (x, y) > n then obviously pn (x, y) = 0. Theorem 5.3 can be extended to a general Faber-Krahn function (s) as follows.
5.1. ON-DIAGONAL UPPER ESTIMATES OF THE HEAT KERNEL
101
Theorem 5.4 If (V, ) satises (5.3) and the Faber-Krahn inequality with a positive decreasing function (s) on (0, +1) then, for all positive integers n and all x, y 2 V , pn (x, y) C 1 (n/8) ,
where C is a constant and the function is dened by Z s dv (s) = . 1 v(v) Example. In the case (s) = cs1/ we obtain Z 1 s 1/1 (s) = v dv = c0 s1/ 1 c0 s1/ c 1
and 1 (t) c00 t , which gives the previous theorem.
Example. Let (s) = ln2 c as on the Cayley graphs with exponential volume (2s) growth. Then Z Z dv 1 2s 2 du 1 3 1 s 2 ln (2v) ln u = ln (2s) (s) = c 1 v c 2 u 3c whence 1 (t) 1 exp c0 t1/3 and 2 One can show that, for a large family of Cayley graphs, there is a similar lower bound, with dierent values of constants C, c00 . Proof. The proof goes the same lines as that of Theorem 5.3. The only place where we have used the Faber-Krahn inequality was the estimate (5.13) in Claim 5, which now becomes 4C0 . 1 (s ) ( (s )) bn Putting together with (5.12) and setting C = 4C0 , we obtain C 1 bn . bn bn+1 2 bn Using (5.18), we estimate bn from above as follows. Consider the function (b) = 1 C , b b pn (x, y) C exp c00 n1/3 . (5.17)
(5.18)
which is monotone decreasing in b 2 (0, +1). It follows that Z

bn bn+1
(b) db (bn bn+1 ) (bn ) =
bn bn+1 1 , C 2 bn bn
102
which implies by the change v =
where in the last inequality we have used (5.18). Summing up this inequality from 0 to n, we obtain Z b0 db n C , 2 bn b b
C b
that
Z Recall that
C/bn C/b0
dv n . (v) v 2 X
x 2 f0 (x) (x) .
(5.19)
b0 = (f0 , f0 ) = Since X
x
f0 (x) (x) = (f0 , 1) = 1, X

x
it follows that
2 f0 (x) 2 (x) 1.
Since (x) 1, we obtain that b0 = because C = 4C0 > 4. Hence, whence
X
x
2 f0 (x) (x) 1 < C,
C b0
> 1, and (5.19) implies that = Z

C/bn 1
C bn
dv n , (v) v 2
and
n C 1 bn 2 bn 1 C . (n/2) we obtain bn = p2n (z, z) and
As in the previous proof, choosing f0 =
1 1 , (z) {z}
p2n (z, z)
C . (n/2)
For any integer n 2, using (5.16) and choosing k, l as in the previous proof, we obtain C C pn (x, y) 1 . (5.20) 1/2 (n/8) ( 1 (k/2) 1 (l/2)) For n = 1 we have p1 (x, y) 1. By increasing if necessary the value of C, we can have 1C 1 so that (5.20) is satised also for n = 1. (1/8)
5.2. ON-DIAGONAL LOWER BOUNDS
103
5.2
On-diagonal lower bounds
Theorem 5.5 For any even positive integer n and for any non-empty nite set V , the following estimate holds: sup pn (x, x)
xV
(1 1 ())n . ()
(5.21)
In particular, if 1 () 1/2 then sup pn (x, x)

xV
exp (21 ()n) . ()
(5.22)
Since the set is arbitrary, one can rewrite (5.22) in the form sup pn (x, x) sup
xV V
exp (21 ()n) , ()
where sup is taken over all non-empty nite subsets . In this form, it appears as an inequality between two functions of n. Example. We claim that in Zm pn (x, x) cnm/2 (5.23)
for all even positive integers n and x 2 Zm . Indeed, x r 2 N and take = Br . One can show that C 1 (Br ) 2 r (see Exercise 39). It follows that, for large enough r, C exp r2 n exp (21 (Br )n) c0 . sup pn (x, x) (Br ) rm x p Choosing r n and noticing that pn (x, x) does not actually depend on x (see Lemma 5.6 below), we obtain (5.23). Combining together the on-diagonal upper and lower bounds for the heat kernel in Zm , we obtain that, for all x 2 Zm and all even n, pn (x, x) ' nm/2 . Note that, for odd n, pn (x, x) = 0. Example. Using this method, one can prove the following lower bound on Cayley graphs of polycyclic groups with exponential volume growth: pn (x, x) c exp Cn1/3
for even n, which matches the upper bound (5.17). In this case, constructing an appropriate set is more complicated and requires the use of the structure of the group.
104
Instead of giving the denition of a polycyclic group, let us give an example. Consider a semi-direct product Z2 o Z that, as a set, consists of couples (x, a) where x 2 Z2 and a 2 Z, and the group operation is dened by (x, a) (y, b) = (x + M a y, a + b) where M is a 2 2 matrix with integer coecients and with det M = 1 (then also M 1 has integer coecients). If M = id then we obtain just Z3 . For a less trivial M, for example, for M = 2 1 , we obtain a group with exponential volume growth. 1 1 Proof. The estimate (5.22) follows from (5.21) using the inequality 1 exp (2) that is true for 0 1/2. Indeed, it is obviously true for = 0 and = 1 and, 2 hence, is true for 2 [0, 1/2] because the function 1 is linear and exp (2) is convex. Let us prove (5.21). We use again the Markov operator P = id L, that acts on functions f 2 F as follows: X P (x, y) f (y) . P f (x) =
yV
Recall that the powers P n of the Markov operators are given by X Pn (x, y) f (y) , P n f (x) =
yV
Fix a non-empty nite set V and consider the operator Q = P = id L in F , that is, X Qf (x) = P (x, y) f (y) .
y
where Pn (x, y) is the transition function that can be dened inductively by P1 (x, y) = P (x, y) and X Pn (x, y) = Pn1 (x, z) P (z, y) . (5.24)
zV
where Q1 (x, y) = P (x, y) and
The distinction between Q and P is that the range of the summation of the former is restricted to y 2 . For any positive integer n, consider the powers Qn . By induction, one obtains X Qn f (x) = Qn (x, y) f (y) , (5.25)
y
Qn (x, y) =
X
z
Qn1 (x, z) P (z, y) .
(5.26)
The function Qn (x, y) can be regarded as the transition function for a random walk with the killing condition outside . The comparison of (5.24) and (5.26) shows that Qn (x, y) Pn (x, y) .
5.2. ON-DIAGONAL LOWER BOUNDS
105
Consider trace Qn . the one hand, (5.25) means that the matrix of this operOn ator in the basis 1{x} x has on the diagonal the values Qn (x, x) so that trace Qn = X
x
Qn (x, x)
X
x
Pn (x, x) .
On the other hand, the operator Q has the eigenvalues 1 k (), k = 1, ..., N where N = jj. The operator Qn has the eigenvalues (1 k ())n whence trace Q =
n N X k=1
(1 k ())n (1 1 ())n .
We have used that all the terms in the above sum are non-negative, which is true because n is even. Comparing the two expressions for the trace, we obtain X (1 1 ())n Pn (x, x) = sup pn (x, x)(),
x
X
x
pn (x, x)(x)
whence (5.21) follows. The following lemma will enable us to obtain the lower bounds for pn (x, x) on Cayley graphs. Lemma 5.6 On any Cayley graph (G, S), the value of pn (x, x) does not depend on x, that is, pn (x, x) = pn (y, y) for all x, y. Proof. Let us show that the heat kernel is invariant under the left multiplication, that is, pn (x, y) = pn (zx, zy) (5.27) for all x, y, z 2 G, which will imply for y = x and z = x1 that pn (x, x) = pn (e, e) . Note that x y is equivalent to x1 y 2 S. Since (zx)1 (zy) = x1 z 1 zy = x1 y we see that x y if and only if zx zy. For n = 1 we have p1 (x, y) = xy P (x, y) = . (y) (x) (y)
If x, y are not neighbors then p1 (x, y) = 0. If x y then it follows that p1 (x, y) = 1 1 . = deg (x) deg (y) jSj2
106
Since the right hand side is the same for all couples x y, we obtain (5.27) for n = 1. Let us make the inductive step from n 1 to n: X pn (zx, zy) = pn1 (zx, w) p1 (w, zy) (w) = X
uG uG wG
pn1 (zx, zu) p1 (zu, zy) (u) pn1 (x, u) p1 (u, y) (u)
= pn (x, y) .
Corollary 5.7 If under the conditions of Theorem 5.5, (V, ) is a Cayley graph and 1 () 1/2 then we have pn (x, x) for all x 2 V and even n. Proof. By Lemma 5.6, we have, for any x 2 V , pn (x, x) = sup pn (x, x)
xV
exp (21 ()n) ()
(5.28)
so that (5.28) follows from (5.22).
5.3
The type problem
and transient otherwise, that is, if there is x 2 V such that Px (Xn = x innitely often) < 1.
Denition. We say that the random walk fXn g on (V, ) is recurrent if, for any x 2V, Px Xn = x innitely often2 = 1,
The type problem is the problem of deciding whether the random walk is recurrent or transient. Theorem 5.8 The random walk is transient if and only if
X n=1
pn (x, x) < 1
(5.29)
for some/all x 2 V .
5.3. THE TYPE PROBLEM
107
Corollary 5.9 (Polyas theorem) In Zm the random walk is transient if and only if m > 2. Proof. Indeed, in Zm we have X
n
pn (x, x) '
X 1 nm/2 n
and the latter series converges if and only if m > 2. We start the proof of Theorem 5.8 with the following lemma. Lemma 5.10 If the condition
X n=1
pn (x, y) < 1
(5.30)
holds for some x, y 2 V then it holds for all x, y 2 V. In particular, if (5.29) holds for some x 2 V then it holds for all x 2 V and, moreover, (5.30) holds for all x, y 2 V . Proof. Let us show that if (5.30) holds for some x, y 2 V then the vertex x can be replaced by any of its neighbors, and (5.30) will be still true. Since the graph (V, ) is connected, in a nite number of steps the initial point x can be then replaced by any other point. By the symmetry, the same applies to y so that in the end both x and y can take arbitrary values. Fix a vertex x0 x and prove that
X n=1
pn (x0 , y) < 1.
We have Pn+1 (x, y) = whence pn (x0 , y) = It follows that

X
X
z
P (x, z) Pn (z, y) P (x, x0 ) Pn (x0 , y) ,
Pn+1 (x, y) pn+1 (x, y) Pn (x0 , y) = . 0 ) (y) (y) P (x, x P (x, x0 )

0
which was to be proved. Proof of Theorem 5.8: the suciency of (5.29). Fix a vertex x0 2 V and denote by An the event fXn = x0 g so that, for any x 2 V , Px (An ) = Px (Xn = x0 ) = Pn (x, x0 ) = pn (x, x0 ) (x0 ) . P By Lemma 5.10, the condition (5.29) implies n pn (x, x0 ) < 1 whence X Px (An ) < 1.
n
X 1 pn (x , y) pn+1 (x, y) < 1, P (x, x0 ) n=1 n=1
(5.31)
108 We have
where Bm =
Px (Xn = x0 innitely often) = Px (8m 9n m Xn = x0 ) T S = Px An m nm T Bm = Px S

m nm
Therefore,
The sequence fBm g is decreasing in m, which implies that T Bm = lim Px (Bm ) = 0. Px

m m
An . It follows from (5.31) that X Px (An ) ! 0 as m ! 1. Px (Bm )

nm
Px (Xn = x0 innitely often) = 0,
(5.32)
and the random walk is transient3 . Note that the condition (5.32) that we have proved, is in fact stronger than the denition of the transience as the latter is Px0 (Xn = x0 innitely often) < 1 for some x0 2 V . We will take advantage of (5.32) later on. The proof of the necessity of condition (5.29) in Theorem 5.8 will be preceded by some lemmas. Lemma 5.11 (Strong maximum principle) Let u be a subharmonic function on V , that is, such that Lu 0 on V . If, for some point x 2 V , u (x) = sup u, then u const . In other words, a subharmonic function on V cannot attain its supremum unless it is a constant. Proof. See Exercise 41. Denition. Fix a nite non-empty set K V and consider the function vK (x) = Px (9n 0 Xn 2 K) . The function vK (x) is called the hitting (or visiting) probability of K. also the function hK (x) = Px (Xn = x0 innitely often) , Consider
that is called the recurring probability of K. Clearly, we have v 1 on K and 0 hK (x) vK (x) 1 for all x 2 V . In the next two lemmas, the set K will be xed so that we write v (x) and h (x) instead of vK (x) and hK (x), respectively.
3 Alternatively, one can use a lemma of Borel-Cantelli that says the following: if (5.31) is satised then the probability that the events An occur innitely often, is equal to 0 (which is exactly (5.32)). In fact, the above argument contains the proof of that lemma.
109
Lemma 5.12 We have Lv (x) = 0 if x 2 K (that is, v is harmonic outside K), / and Lv (x) 0 for any x 2 K. Proof. If x 2 K then we have by the Markov property / v (x) = Px (9n 0 Xn 2 K) = Px (9n 1 Xn 2 K) X = P (x, y) Py (9n 1 Xn1 2 K) = = X
y
X
y
P (x, y) Py (9n 0 Xn 2 K) P (x, y) v (y)
so that v (x) = P v (x) and Lv (x) = 0. If x 2 K then Lv (x) = v (x) P v (x) = 1 P v (x) 0 because P v (x) P 1 (x) = 1. Lemma 5.13 The sequence of functions fP n vg is decreasing in n and
n
lim P n v (x) = h (x)
(5.33)
for any x 2 V . Proof. Since Lv 0, we obtain P n v P n+1 v = P n (v P v) = P n (Lv) 0 so that fP n vg is decreasing. Hence, the limit in (5.33) exists. Consider the events Bm = f9n m Xn 2 Kg . Obviously, the sequence fBm g is decreasing and the event T
m
Bm = f8m 9n m Xn 2 Kg
is identical to the event that Xn 2 K innitely often. Hence, we have h (x) = Px We claim that T
m
Bm
= lim Px (Bm ) .
m
(5.34)
Px (Bm ) = P m v(x).
(5.35)
110
Indeed, for m = 0 this is the denition of v (x). Here is the inductive step from m 1 to m using the Markov property: X Px (9n m Xn 2 K) = P (x, y) Py (9n m Xn1 2 K) = = X
y y m y
P (x, y) Py (9n m 1 Xn 2 K) P (x, y) P m1 v (y)
= P v (x) . Combining (5.34) with (5.35), we obtain (5.33). Proof of Theorem 5.8: the necessity of (5.29). Assume that the random walk is transient and show that (5.29) is true. Let x0 2 V be a point where Px0 (Xn = x0 innitely often) < 1. Consider the hitting and recurring probabilities v (x) and h (x) with respect to the set K = fx0 g. The above condition means that h (x0 ) < 1. It follows that v 6 1 because otherwise P n v 1 for all n and by Lemma 5.13 h 1. As we know, Lv (x) = 0 for x 6= x0 and Lv (x0 ) 0. Claim 1. Lv (x0 ) > 0. Assume from the contrary that Lv (x0 ) = 0, that is, Lv (x) = 0 for all x 2 V . Since v takes its maximal value 1 at some point (namely, at x0 ), we obtain by the strong maximum principle that v 1, which contradicts the assumption of the transience. Denote f = Lv so that f (x) = 0 for x 6= x0 and f (x0 ) > 0. Claim 2. We have for all x 2 V
X n=0
P n f (x) v (x) .
(5.36)
whence it follows that ! m1 X P n f = (id P m ) f = f P m f f. L

n=0
Fix a positive integer m and observe that (id P ) id +P + P 2 + ... + P m1 = id P m
Set vm =
Obviously, vm has a nite support and Lvm f. For comparison, we have Lv = f and v 0 everywhere. We claim that vm v in V . Indeed, let = supp vm so that
m1 X n=0
P n f.
111
outside m the inequality vm v is trivially satised. In we have L (v vm ) 0. By the minimum principle of Lemma 1.11, we have min (v vm ) = inf (v vm ) . c

Since the right hand side is 0, it follows that v vm 0 in , which was claimed. Hence, we have m1 X P n f v,
n=0
whence (5.36) follows by letting m ! 1. Using that supp f = fx0 g, rewrite (5.36) in the form
X n=0
pn (x, x0 ) f (x0 ) (x0 ) v (x)

X n=0
whence it follows that
pn (x, x0 ) < 1.
Setting here x = x0 we nish the proof. Corollary 5.14 Let K be a non-empty nite subset of V . If the random walk is recurrent then vK hK 1. If the random walk is transient then vK 6 1 and hK 0. Hence, we obtain a 0-1 law for the recurring probability: either hk 1 or hK 0. Proof. Let x0 be a vertex from K. Obviously, we have v{x0 } (x) vK (x) . Therefore, if the random walk is recurrence and, hence, v{x0 } 1 then also vK (x) 1. Since hK = lim P m vK , (5.37)
m
for all x0 , x 2 V . It follows from the proof of Theorem 5.8 that h{x0 } (x) = 0 (cf. (5.32)). If fXn g visits K innitely often then fXn g visits innitely often at least one of the vertices in K. Hence, we have X h{x0 } . hK
x0 K
it follows that hK 1. Let the random walk be transient. Then by Theorem 5.8 and Lemma 5.10, we have X pn (x0 , x) < 1
n=1
Since h{x0 } 0, we conclude that hK 0. Finally, (5.37) implies that vK 6 1.
112
5.4
Volume tests for recurrence
In this section, let us x an integer-valued function (x) on V with the following two properties: For any non-negative integer r, the set Br = fx 2 V : (x) rg is nite and non-empty. If x y then jrxy j 1. For example, (x) can be the distance function to any nite non-empty subset of V . Theorem 5.15 (The Nash-Williams test) If
X r=0
1 = 1, (Br )
(5.38)
then the random walk on (V, ) is recurrent. Note that Br is non-empty because otherwise the graph (V, ) would be disconnected. An alternative way of stating this theorem is the following. Assume that V is a disjoint union of a sequence fAk g of non-empty nite subsets with the following k=0 property: if x 2 Ak and y 2 Am with jk mj 2 then x and y are not neighbors. Denote by Ek the set of edges between Ak and Ak+1 and assume that
X k=0
1 = 1. (Ek )
(5.39)
Then the random walk on (V, ) is recurrent. Indeed, dening (x) = k if x 2 Ak S we obtain that Br = r Ak and Br = Er . Hence, (5.39) is equivalent to (5.38). k=0 Let us give two simple examples when (5.39) is satised: 1. if (Ek ) Ck for all large enough k; 2. if Ekj C for a sequence kj ! 1 (in this case, (Ek ) for k 6= kj may take arbitrarily big values). Proof of Theorem 5.15. Consider the hitting probability of B0 : v (x) = vB0 (x) = Px (9n 0 : Xn 2 B0 ) . Recall that 0 v 1, v = 1 on B0 , and Lv = 0 outside B0 (cf. Lemma 5.12). Our purpose is to show that v 1, which will imply the recurrence by Corollary 5.14.
5.4. VOLUME TESTS FOR RECURRENCE
113
We will compare v (x) to the sequence of functions fuk g that is constructed k=1 as follows. Dene uk (x) as the solution to the following Dirichlet problem in k = Bk n B0 : Luk = 0 in k (5.40) uk = f in c k
where f = 1B0 . In other words, uk = 1 on B0 and uk = 0 outside Bk , while uk is harmonic in k . By Theorem 1.10, the problem (5.40) has a unique solution. By the maximum/minimum principle of Lemma 1.11, we have 0 uk 1. c Since uk+1 = uk on B0 and uk+1 0 = uk in Bk , we obtain by the maximum principle that uk+1 uk in k . Therefore, the sequence fuk g increases and converges to a function u as k ! 1. The function u has the following properties: 0 u 1, u = 1 on B0 , and Lu = 0 outside B0 (note that Luk ! Lu as k ! 1). Comparing v with uk in k and using the maximum principle, we obtain that v uk , whence it follows that v u . Hence, in order to prove that v 1, it suces to prove that u 1, which will be done in the rest of the proof. By Exercise 23, the solution of the Dirichlet problem (5.40) has the minimal value of the Dirichlet integral 1 X D (u) = (rxy u)2 xy , 2
x,yU1 (k )
among all functions u that satisfy the boundary condition u = f in c . Since u 1 k c in B0 , u 0 in Bk , and U1 (Bk ) Bk+1 , we have D (u) = 1 X (rxy u)2 xy . 2 x,yB
k+1
Choose a function u with the above boundary condition in the form u (x) = ( (x)) , where (s) is a function on Z such that (s) = 1 for s 0 and (s) = 0 for s k + 1. Set S0 = B0 and Sr = fx 2 V : (x) = rg for positive integers r. Clearly, Br is a disjoint union of S0 , S1 , ..., Sr . Observe also that if x y then x, y belong either to the same Sr (and in this case rxy u = 0) or the one to Sr and the other to Sr+1 , because j (x) (y)j 1. Having this in mind, we obtain D (u) =
k X k X k X r=0
r=0 xSr ,ySr+1
X X
(rxy u)2 xy ( (r) (r + 1))2 xy
r=0 xSr ,ySr+1
( (r) (r + 1))2 (Br ) .
114 Denote
m (r) := (Br ) and dene (r) for r = 0, ..., k from the following conditions: (0) = 1 and (r) (r + 1) = ck , r = 0, ..., k m (r) (5.41)
where the constant ck is to be found. Indeed, we have still the condition (k + 1) = 0 to be satised. Summing up (5.41), we obtain (0) (k + 1) = ck so that (k + 1) = 0 is equivalent to ck = Hence, assuming (5.42), we obtain D (u) =
k X r=0 k X r=0
1 m (r)
k X
r=0
1 m (r)
!1
(5.42)
X 1 c2 k = ck . m (r) = c2 k m (r) m (r)2 r=0

k
By the Dirichlet principle, we have D (uk ) D (u) whence D (uk ) ck . On the other hand, by the Green formula X Luk (x) uk (x) (x) = X 1 X (rxy uk )2 xy 2 x,yB xB
k+1
(5.43)
Bk+1
c k+1 yBk+1
(rxy uk ) uk (x)xy .
c c The last sum vanishes because if y 2 Bk+1 and x y then x 2 Bk and uk (x) = 0. The range of summation in the rst sum can be reduced to Bk because uk = 0 outside Bk , and then further to B0 because Luk = 0 in Bk n B0 . Finally, since uk 1 in B0 , we obtain the identity
X
B0
Luk (x) (x) =
1 X (rxy uk )2 xy = D (uk ) . 2 x,yB

k+1
It follows from (5.43) that X

B0
Luk (x) (x) ck .
Since u takes the maximal value 1 at any point of B0 , we have at any point x 2 B0 that P uk (x) 1 and Luk (x) = uk (x) P uk (x) 0.
5.4. VOLUME TESTS FOR RECURRENCE Hence, at any point x 2 B0 , we have 0 Luk (x) (x) ck . By (5.38) and (5.42), we have ck ! 0 as k ! 1, whence it follows that Luk (x) ! 0 for all x 2 B0 .
115
Hence, Lu (x) = 0 for all x 2 B0 . Since Lu (x) = 0 also for all x 2 B0 , we see / that u is harmonic on the whole graph V . Since u takes its supremum value 1 at any point of B0 , we conclude by the strong maximum principle that u 1, which nishes the proof. The following theorem provides a convenient volume test for the recurrence. Theorem 5.16 If
X r=0
r =1 (Br )
(5.44)
then the random walk is recurrent. In particular, this is the case when
2 (Brk ) Crk
(5.45)
for a sequence rk ! 1. The condition (5.45) holds in Zm with m 2 for the function (x) = d (x, 0). Hence, we obtain one more proof of the recurrence of Zm for m 2 (cf. Corollary 5.9). We need the following lemma for the proof of Theorem 5.16. Lemma 5.17 Let f r gn be a sequence of positive reals and let r=0 vr = Then
r X i=0
i .
(5.46)
n n X 1 1X r . 4 r=0 vr r=0 r
n1 2
whence
Proof. Assume rst that the sequence f r g is monotone increasing. If 0 k then 2k+1 X i (k + 1) k v2k+1
i=k+1
1 k+1 1 2k + 1 . k v2k+1 2 v2k+1

n 2
Similarly, if 0 k
then v2k
2k X
i=k+1
i kk
116 and
1 k 1 2k = . k v2k 2 v2k
n1 [X] [n] n n 2 2 X 1 X r 2k + 1 X 2k 4 + = , k v2k+1 v2k vr r=0 k=0 k=0 k=0
It follows that
which was claimed. Now consider the general case when the sequence f r g is not necessarily increasing. Let fer gn be an increasing permutation of f r gn and set r=0 r=0 v er =
r X i=0
e Note that vr vr because vr is the sum of r smallest terms of the sequence fi g e whereas vr is the sum of some r terms of the same sequence. Applying the rst part of the proof to the sequence fei g, we obtain
n n n n X 1 X 1 1X r 1X r = , e 4 r=0 er v 4 r=0 vr r=0 r r=0 r
i . e
which nishes the proof. Proof of Theorem 5.16. Set for any r 1
Sr = fx 2 V : (x) = rg = Br n Br1 and S0 = B0 . Then we have X (Br ) = xy =

xBr ,y Br /
xSr ,ySr+1
xy
xSr ,yV
xy =
xSr
(x) = (Sr ) .
Denoting vr = (Br ) and r = (Sr ) and observing that the sequences fvr g and f r g satisfy (5.46), we obtain by Lemma 5.17 and (5.44) that
X r=0
X 1 1X r 1 = 1. (Br ) 4 r=0 vr r=0 r

Hence, (5.38) is satised, and we conclude by Theorem 5.15 that the random walk on (V, ) is recurrent. We are left to show that (5.45) implies (5.44). Given positive integers a < b, we have b b a X X X b2 a2 b (b + 1) a (a + 1) , r= r r= 2 2 2 r=a+1 r=1 r=1 whence it follows that
b X r 1 b2 a2 . v vb 2 r=a+1 r
5.5. ISOPERIMETRIC TESTS FOR TRANSIENCE
117
By choosing a subsequence of frk g, we can assume that rk 2rk1 . Then we have, using (5.45),
rk X X r X r v v r=0 r +1 r k r=r
k1
2 2 X 1 rk rk1 vrk 2 k 2 2 1 X rk rk1 2 2C k rk 2 rk1 1 X 1 2 = 2C k rk 1 X3 = 1, 2C k 4
which was to be proved.
5.5
Isoperimetric tests for transience
Theorem 5.18 Let the graph (V, ) satisfy the hypothesis (5.3). (a) If (V, ) satises the isoperimetric inequality with function (s) such that Z ds <1 (5.47) 2 (s) then the random walk on (V, ) is transient. (b) If (V, ) satises the Faber-Krahn inequality with function (s) such that Z ds <1 (5.48) 2 (s) s then the random walk on (V, ) is transient. For example, in Zm we have (s) = cs Z ds s2
m1 m
and (5.47) becomes
m1 m
< 1,
1 2 (s) . 2 s2
which is satises provided 2 m1 > 1, that is, m > 2. m Proof. Suces to prove (b) because (a) follows (b) with (s) = Theorem 5.4, the Faber-Krahn inequality implies that pn (x, x) where (r) = Z C 1 (cn)
r 1
By
ds . s(s)
118
CHAPTER 5. THE HEAT KERNEL AND THE TYPE PROBLEM P

1 n 1 (cn)
Hence, it suces to prove that
< 1, which is equivalent to
dt < 1. (t)
Changing s = 1 (t), we arrive at Z Since 0 (s) =

1 , s(s)
0 (s) ds < 1. s
the latter condition is equivalent to (5.48).
Corollary 5.19 Let (G, S) be a Cayley graph and let Br = Br (e) where e is the neutral element of G. The random walk on (G, S) is recurrent if and only if
X r = 1. jBr j r=1
(5.49)
then the random walk is transient. Let V (r) be a smooth strictly increasing function on [0, +1) such that (Br ) = V (r) for all non-negative integers r. Then (5.50) implies Z rdr < 1. V (r) 0 (s) = c s . V 1 (2s)
Proof. Note that on (G, S) we have for any set A G that (A) = jAj jSj . Hence, (5.49) implies (5.44) and the random walk is recurrent by Theorem 5.16. Let us prove that if X r <1 (5.50) jBr j r=1
By Theorem 4.9, the graph (G, S) satises the isoperimetric inequality with function
Let us show that (s) satises (5.47) which will imply the transience by Theorem 5.18. Indeed, we have c
2
ds = 2 (s)
[V 1 (2s)] ds 1 [change s = V (r) ] 2 s 2 1 Z 2 r dV(r) = 2 2 V 1 (2) V (r) Z 1 2 r d . 2 V(r) 0 R

0
Integrating by parts, we obtain 2 Z

R 0
1 r d V(r)
2
2r2 = V(r)
+4
R 0
rdr 4 V(r)
R 0
rdr . V(r)
5.5. ISOPERIMETRIC TESTS FOR TRANSIENCE Letting R ! 1, we obtain Z Z rdr 1 2 2 r d 4 < 1, V(r) V(r) 0 0
119
whence (5.47) follows. It is known from Group Theory that for Cayley graphs the following two alternatives take places: 1. either (Br ) ' rm for some positive integer m (the power volume growth), 2. or, for any C, N, we have (Br ) CrN for large enough r (the superpolynomial volume growth). It follows from Corollary 5.19 that, in the rst case, the random walk is recurrent if and only if m 2, while in the second case the random walk is always transient.

Aglect

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Aglect

Uploaded by

Copyright:

Available Formats

Analysis on Graphs

Alexander Grigoryan Lecture Notes University of Bielefeld, WS 2009/10

Chapter 1 The Laplace operator on graphs

CHAPTER 1. THE LAPLACE OPERATOR ON GRAPHS

1.1. THE NOTION OF A GRAPH

Lemma 1.2 On any simple nite weighted graph (, ), X X (x) = 2 .

Proof. Rewrite (1.1) in the form (x) = X

1.2. CAYLEY GRAPHS

CHAPTER 1. THE LAPLACE OPERATOR ON GRAPHS

Note that the relation x y is symmetric in x, y, that is, x y implies y x, because 1 2 S. y 1 x = x1 y

1.2. CAYLEY GRAPHS

CHAPTER 1. THE LAPLACE OPERATOR ON GRAPHS

The corresponding graph (Z2 , S) is shown here: j j j j j j j j j j j j j j j j j j j j j j j j j j j j j j

CHAPTER 1. THE LAPLACE OPERATOR ON GRAPHS

1.3. RANDOM WALKS

strategy is as follows: rst construct a probability measure P(1) on V , then P(n) on

CHAPTER 1. THE LAPLACE OPERATOR ON GRAPHS

Clearly, Px is -additive, that is, ! G X (1) Px Ak = Px (Ak ) ,

P(n) (x1 , ..., xn ) . x

P (x, x1 ) P (x1 , x2 ) ...P (xn1 , xn )

is not a product measure of n copies of Px P (x; x1 ) P (x; x2 ) :::P (x; xn ) :

since the latter would

1.3. RANDOM WALKS

that is proved as follows: X (m) Px (x1 , ..., xn ) = =

= P (x, x1 ) P (x1 , x2 ) ...P (xn1 , xn ) = = P(n) x P(n) x

P (x, x1 ) P (x1 , x2 ) ...P (xn1 , y)

CHAPTER 1. THE LAPLACE OPERATOR ON GRAPHS

Px (Xn = y, Xn+1 = z) = = whence

Px (Xn = y, Xn+1 = z) = P (y, z) , Px (Xn = y)

Proof. By (1.14), we have Pn (x, z) = Px (Xn = z) = = X X Px (X1 = x1 , ..., Xn1 = xn1 , Xn = z)

P (x, x1 ) P (x1 , x2 ) ...P (xn1 , z) .

P (x, x1 ) P (x1 , x2 ) ...P (xn1 , z)A P (z, y)

which nishes the proof.

Pn (x, z) Pk (z, w) P (w, y) Pn (x, z) Pk (z, w) P (w, y)

Pn (x, z) Pk+1 (z, y) .

CHAPTER 1. THE LAPLACE OPERATOR ON GRAPHS

P (x, y) (x) = (x) ,

A.Grigoryan Lecture 4 26.10.09

1.3. RANDOM WALKS

X 1 vn (y) yx = Pn (x0 , y) P (y, x) (y) y

= Pn+1 (x0 , x) = vn+1 (x) ,

which proves (1.21). Substituting (1.20) into (1.21), we obtain (1.22).

CHAPTER 1. THE LAPLACE OPERATOR ON GRAPHS

1 vn (y) . deg (y)

1 X un (y) . deg (x) yx

... ... n = 10 0.117

CHAPTER 1. THE LAPLACE OPERATOR ON GRAPHS

The Laplace operator

Let f (x) be a function on R. Recall that

1.4. THE LAPLACE OPERATOR so that f 0 (x) f (x) f (x h) f (x + h) f (x) h h

CHAPTER 1. THE LAPLACE OPERATOR ON GRAPHS 1 X f (y) f (x) . deg (x) yx

A.Grigoryan Lecture 5 27.10.09

P (x, y) f (y) f (x) .

CHAPTER 1. THE LAPLACE OPERATOR ON GRAPHS

The Dirichlet problem

1.5. THE DIRICHLET PROBLEM

and for any function u : V ! R, that is superharmonic in , we have min u inf u. c

Since u (y) M for all y 2 V , we have X X P (x, y) u (y) M P (x, y) = M.

CHAPTER 1. THE LAPLACE OPERATOR ON GRAPHS

1.5. THE DIRICHLET PROBLEM

CHAPTER 1. THE LAPLACE OPERATOR ON GRAPHS

Chapter 2 Spectral properties of the Laplace operator

f (y) P (x, y) f (x) = f (x) .