MATH 262A NOTES, 2/5/13 LATEXED BY JAY CUMMINGS 1. The Second Eigenvalue. Suppose we have two connected graphs G and G0 on the same vertex set V (G). We want to relate the second eigenvalues of the normalized Laplacian of these two graphs. Call these eigenvalues λ and λ0 , respectively. Then we have the following theorem. Theorem 1. Suppose that for each edge {x, y} ∈ E(G) there exist a path Pxy in G0 with |Pxy | ≤ `. Also, for v ∈ V , suppose that dv ≥ ad0 v . Finally, let m be such that for all e ∈ E(G), m ≥ me , where me is the number of paths using e. Given all this, we can conslude that aλ1 . λ0 1 ≥ `m Let’s do some examples before the proof. Example 1. Let G = Kn and recall that λ1 = n/(n − 1). Then a n n 0 λ1≥ ≥ . `m n − 1 `m(max degree) Example 2. Let V = Zp , G = Kn , G0 = Cn . For each pair of points we pick the shortest path, allowing us to set ` = (p − 1)/2, the diameter of G0 . Next, fix an edge. We count m by partitioning by the paths that will use that fixed edge in a shortest path. This is seen to be X p2 − 1 p−1 −i = . m= 2 1 0≤i≤(p−3)/2 So λ0 1 ≥ p p−1 2 · p2 −1 8 ·2 & 8 . p2 But now recall that we actually know the λ1 for the cycle. It is 2 2π 2π 1 2π 2 = = 2 . 1 − cos p p 2 p We got the order of magnitude right, were just off by a constant. 1 2 LATEXED BY JAY CUMMINGS Example 3. Let G0 = Z2 d = Qd be the d-cube, which has vertex set V = {0, 1}d . G is still K2d , ` = d is still the diameter. Given vertices (a1 , . . . , ad ) and (b1 , . . . , bd ). we’ll create a path as follows. Find the first place the two differ, say ai . Then the edges are (x1 , . . . , xt−1 , ∗, xt+1 , . . . , xd ), where ∗ is 0 and then 1. This defines a single edge. Continue in this way. Number of paths through this edge? x = ( – any – , ∗, xt+1 , . . . , xd ) and y = (x1 , . . . , xt−1 , ∗, – any – ). So m = 2d−1 . Therefore 2 2d λ1 ≥ d−1 = 2 . d d2 d Whereas actually λ1 = d2 . So we did give up a bit with this approach. Now let’s prove Theorem 1. Proof. First recall that λ0 1 = RG0 (g), the Rayleigh quotient, for some Rayleigh function g. This g has the property that X g(x)d0 x = 0. Now let’s construct an P upperbound for λ1 by finding a good instance of f . Define f : V (G) → R. We need f (x)dx = 0. To accomplish this, let’s set f (x) = g(x) − c. Then X X X f 2 (x)dx = (g(x) − c)2 dx ≥ a (g(x) − c)2 d0 x x∈V x∈V (G) ≥ a inf c x∈V X (g(x) − c)2 d0 x = a x∈V X g 2 (x)d0 x . x∈V Now, for an edge e = {x, y} ∈ E(G). Define g(e) = |g(x) − g(y)|. {x, y} may not be an edge in G0 , but we do know that there exists a path Px,y if length less P than or equal to ` in G0 . By Cauchy-Schwarz (and telescoping), (f (x) − f (y))2 ≤ |Pxy | e∈Px,y g 2 (e). Next, m X e∈E(G0 X g 2 (e) ≥ X {x,y}∈E(G) e∈Px,y 1 = ` X X g 2 (e) ≥ {x,y}∈E(G) (g(x) − g(y))2 ` (f (x) − f (y)2 , {x,y}∈E(G) the first two inequalities by definition of m and Cauchy-Schwarz, respectively. Then, putting everything together to get a lower bound on λ0 1 , P 2 e∈E(G0 ) g (e) λ0 1 = RG0 (g) = P 2 v∈V g (v)dv P 1 2 a a {x,y}∈E(G0 ) (f (x) − f (y)) m` ≥ = RG (f ) ≥ λ1 . 1P 2 m` m` x∈V (G) f (x)dx a MATH 262A NOTES, 2/5/13 3 Homework • Generalize the last theorem to weighted paths. Theorem 2. Let G be a graph on n vertices. For all x, y ∈ V (G) let Px,y be the path of length less than or equal to ` joining x and y. Define dx dy , qxy = V ol(G) P the weight of Pxy . Assume Q is such that Qe = Pxy ⊇e qxy ≤ Q. Then λ1 ≥ 1/(`Q). Recall that we previously showed the Cheeger ratio to be h ≥ 1/(2q). We will now sketch the proof to the above theorem. In it, we will use the following homework problem. Homework • Show that V ol(G) λ1 = P P x∼y (f (x) {x,y} (f (x) − f (y)2 − f (y))2 dx dy . Proof. We will use the homework assignment. Since X X (f (x) − f (y))2 ≤ |Pxy | f 2 (e) ≤ ` f 2 (e), e∈Pxy e∈Pxy summing over all x, y gives X x,y qxy (f (x)−f (y))2 ≤ ` X x,y qxy X f 2 (e) ≤ ` e∈Pxy X X e∈E(G) qxy f 2 (e) ≤ `Q Pxy ⊇e X f 2 (e). e∈E(G) And moreover, the left-hand side above has X X dx dy qxy (f (x) − f (y))2 = (f (x) − f (y))2 V ol(G) x,y x,y Applying the homework gives the result.