Spectral and Electrical Graph Theory. Daniel A. Spielman Dept. of Computer Science Program in Applied Mathematics Yale Unviersity

Transcription

1 Spectral and Electrical Graph Theory Daniel A. Spielman Dept. of Computer Science Program in Applied Mathematics Yale Unviersity

2 Outline Spectral Graph Theory: Understand graphs through eigenvectors and eigenvalues of associated matrices. Electrical Graph Theory: Understand graphs through metaphor of resistor networks. Heuristics Algorithms Theorems Intuition

3 Spectral Graph Theory Graph G =(V,E) Matrix A rows and cols indexed by V Eigenvalues Eigenvectors Av = λv v : V IR

4 Spectral Graph Theory Graph G =(V,E) 1! 2! 3! 4! Matrix A rows and cols indexed by V A(i, j) =1if(i, j) E Eigenvalues Eigenvectors Av = λv v : V IR ! 2! 3! 4!

5 Example: Graph Drawing by the Laplacian

6 Example: Graph Drawing by the Laplacian

7 Example: Graph Drawing by the Laplacian L(i, j) = 1 if (i, j) E deg(i) if i = j 0 otherwise 8 9

8 Example: Graph Drawing by the Laplacian L(i, j) = 1 if (i, j) E deg(i) if i = j 0 otherwise 8 9 Eigenvalues 0, 1.53, 1.53, 3, 3.76, 3.76, 5, 5.7, 5.7 Let x, y IR V span eigenspace of eigenvalue 1.53

9 Example: Graph Drawing by the Laplacian Plot vertex i at (x(i),y(i)) Draw edges as straight lines

10 Laplacian: natural quadratic form on graphs where D is diagonal matrix of degrees 1! 2! 3! 4!

11 Laplacian: fast facts zero is an eigenvalue Connected if and only if λ 2 > 0 Fiedler ( 73) called algebraic connectivity of a graph The further from 0, the more connected.

12 Drawing a graph in the line (Hall 70) map minimize trivial solution: Solution So, require x 1, x =1 Atkins, Boman, Hendrickson 97: Gives correct drawing for graphs like

13 Courant-Fischer definition of eigvals/vecs

14 Courant-Fischer definition of eigvals/vecs (here )

15 Courant-Fischer definition of eigvals/vecs (here )

16 Drawing a graph in the plane (Hall 70) map minimize

17 Drawing a graph in the plane (Hall 70) map minimize trivial solution: So, require x 1, x 2 1

18 Drawing a graph in the plane (Hall 70) map minimize trivial solution: So, require x 1, x 2 1 diagonal solution: So, require x 1 x 2 Solution up to rotation

19 A Graph

20 Drawing of the graph using v 2, v 3 Plot vertex i at

21

22 The Airfoil Graph, original coordinates

23 The Airfoil Graph, spectral coordinates

24 The Airfoil Graph, spectral coordinates

25 Spectral drawing of Streets in Rome

26 Spectral drawing of Erdos graph: edge between co-authors of papers

27 Dodecahedron Best embedded by first three eigenvectors

28 Intuition: Graphs as Spring Networks edges - > ideal linear springs weights - > spring constants (k) Physics: when stretched to length x, force is kx poten9al energy is kx 2 /2 Nail down some ver9ces, let rest se;le

29 Intuition: Graphs as Spring Networks Nail down some ver9ces, let rest se;le i Physics: minimizes total poten9al energy (i,j) E x(i) (x(i) x(j)) 2 = x T Lx subject to boundary constraints (nails)

30 Intuition: Graphs as Spring Networks Nail down some ver9ces, let rest se;le x(i) i Physics: energy minimized when non- fixed ver9ces are averages of neighbors x(i) = 1 d i (i,j) E x(j)

31 Tutte s Theorem 63 If nail down a face of a planar 3- connected graph, get a planar embedding!

32 Spectral graph drawing: Tutte justification Condition for eigenvector Gives λ small says x(i) = 1 d i λ x(i) (i,j) E x(j) for all near average of neighbors i

33 Spectral graph drawing: Tutte justification Condition for eigenvector Gives λ small says x(i) = 1 d i λ x(i) (i,j) E x(j) for all i near average of neighbors For planar graphs: λ 2 8d/n [S-Teng 96] λ 3 O(d/n) [Kelner-Lee-Price-Teng 09]

34 Small eigenvalues are not enough Plot vertex i at (v 3 (i),v 4 (i))

35 Graph Partitioning

36 Spectral Graph Partitioning [Donath-Hoffman 72, Barnes 82, Hagen-Kahng 92] S = {i : v 2 (i) t} for some t

37 Measuring Partition Quality: Conductance S Φ(S) = # edges leaving S sum of degrees in S For deg(s) deg(v )/2

38 Spectral Image Segmentation (Shi-Malik 00)

39 Spectral Image Segmentation (Shi-Malik 00)

40 Spectral Image Segmentation (Shi-Malik 00)

41 Spectral Image Segmentation (Shi-Malik 00)

42 Spectral Image Segmentation (Shi-Malik 00) edge weight

43 The second eigenvector

44 Second eigenvector cut

45 Third Eigenvector

46 Fourth Eigenvector

47 Cheeger s Inequality [Cheeger 70] [Alon-Milman 85, Jerrum-Sinclair 89, Diaconis-Stroock 91] For Normalized Laplacian: L = D 1/2 LD 1/2 λ 2 /2 min S Φ(S) 2λ 2 And, is a spectral cut for which Φ(S) 2λ 2

48 McSherry s Analysis of Spectral Partitioning p p q S T Divide vertices into S and T Place edges at random with Pr [S-S edge] = p Pr [T-T edge] = p Pr [S-T edge] = q q<p

49 McSherry s Analysis of Spectral Partitioning p q p S T E [ A ]= p q } S q p } T S T

50 McSherry s Analysis of Spectral Partitioning E [ A ]= p q } S q p } T S T v 2 (E [ L ]) is positive const on S, negative const on T View A as perturbation of E [ A ] and L as perturbation of E [ L ]

51 McSherry s Analysis of Spectral Partitioning v 2 (E [ L ]) is negative const on S, positive const on T View A as perturbation of E [ A ] and L as perturbation of E [ L ] Random Matrix Theory [Füredi-Komlós 81, Vu 07] With high probability L E [ L ] small Perturbation Theory for Eigenvectors implies v 2 (L) v 2 (E [ L ])

52 Spectral graph coloring from high eigenvectors Embedding of dodecahedron by 19 th and 20 th eigvecs.

53 Spectral graph coloring from high eigenvectors p p p p p Coloring 3-colorable random graphs [Alon-Kahale 97]

54 Independent Sets S is independent if are no edges between vertices in S

55 Independent Sets S is independent if are no edges between vertices in S Hoffman s Bound: if every vertex has degree d S n 1 d λ n

56 Networks of Resistors Ohm s laws gives i = v/r In general, i = L G v with w (u,v) =1/r (u,v) Minimize dissipated energy v T L G v 1V 0V

57 Networks of Resistors Ohm s laws gives i = v/r In general, i = L G v with w a,b =1/r a,b Minimize dissipated energy By solving Laplacian v T L G v 0.5V 0.5V 1V 0V 0.375V 0.625V

58 Electrical Graph Theory Considers flows in graphs Allows comparisons of graphs, and embedding of one graph within another. Relative Spectral Graph Theory

59 Effective Resistance Resistance of entire network, measured between a and b. Ohm s law: r = v/i R eff (a, b) = 1/(current flow at one volt) 1V a b 0.53V 0.27V 0V 0.2V 0.33V

60 Effective Resistance Resistance of entire network, measured between a and b. Ohm s law: r = v/i R eff (a, b) = 1/(current flow at one volt) = voltage difference to flow 1 unit

61 Effective Resistance R eff (a, b) = voltage difference to flow 1 unit Vector of one unit flow has 1 at a, -1 at b, 0 elsewhere i a,b = e a e b Voltages required by this flow are given by v a,b = L 1 G i a,b

62 Effective Resistance R eff (a, b) = voltage difference of unit flow Voltages required by unit flow are given by Voltage difference is v a,b = L 1 G i a,b v a,b (a) v a,b (b) =(e a e b ) T v a,b =(e a e b ) T L + G (e a e b )

63 Effective Resistance Distance Effective resistance is a distance Lower when are more short paths Equivalent to commute time distance: expected time for a random walk from a to reach b and then return to a. See Doyle and Snell, Random Walks and Electrical Networks

64 Relative Spectral Graph Theory For two connected graphs G and H with the same vertex set, consider L G L 1 H work orthogonal to nullspace or use pseudoinverse Allows one to compare G and H

65 Relative Spectral Graph Theory For two connected graphs G and H, consider L G L 1 H = I n 1 if and only if G = H

66 Relative Spectral Graph Theory For two connected graphs G and H, consider L G L 1 H I n 1 if and only if G H

67 Relative Spectral Graph Theory For two connected graphs G and H, consider 1 1+ eigs(l GL 1 H ) 1+ if and only if for all x IR V 1 1+ xt L G x x T L H x 1+

68 Relative Spectral Graph Theory In particular, for x T L G x = 1 1+ xt L G x x T L H x S (a,b) E x(a) = 1 a S 0 a S 0 (x(a) x(b)) 2 = E(S, V S)

69 Relative Spectral Graph Theory 1 1+ xt L G x x T L H x 1+ For all S V 1 1+ E G(S, V S) E H (S, V S) 1+

70 Expanders Approximate Complete Graphs Expanders: d-regular graphs on n vertices high conductance random walks mix quickly weak expanders: eigenvalues bounded from 0 strong expanders: all eigenvalues near d

71 Expanders Approximate Complete Graphs For G the complete graph on n vertices. all non- zero eigenvalues of L G are n. For, x 1 x =1 x T L G x = n

72 Expanders Approximate Complete Graphs For G the complete graph on n vertices. all non- zero eigenvalues of L G are n. For, x 1 x =1 x T L G x = n For H a d- regular strong expander, all non- zero eigenvalues of L H are close to d. For, x 1 x =1 x T L H x [λ 2, λ n ] d

73 Expanders Approximate Complete Graphs For G the complete graph on n vertices. all non- zero eigenvalues of L G are n. For, x 1 x =1 x T L G x = n For H a d- regular strong expander, all non- zero eigenvalues of L H are close to d. For, x 1 n d H x =1 x T L H x d is a good approximation of G

74 Sparse approximations of every graph 1 1+ xt L G x x T L H x 1+ For every G, there is an H with (2 + ) 2 n/ 2 edges [Batson-S-Srivastava] Can find an H with in nearly-linear time. O(n log n/ 2 ) edges [S-Srivastava]

75 Sparsification by Random Sampling [S- Srivastava] Include edge (u, v) with probability p u,v w u,v R eff (u, v) If include edge, give weight w u,v /p u,v Analyze by Rudelson s concentration of random sums of rank-1 matrices

76 Approximating a graph by a tree Alon, Karp, Peleg, West 91: measure the stretch

77 Approximating a graph by a tree Alon, Karp, Peleg, West 91: measure the stretch T

78 Approximating a graph by a tree Alon, Karp, Peleg, West 91: measure the stretch i 3 j T

79 Approximating a graph by a tree Alon, Karp, Peleg, West 91: measure the stretch T

80 Low-Stretch Spanning Trees For every G there is a T with stretch T (G) m 1+o(1) (Alon- Karp- Peleg- West 91) where m = E stretch T (G) O(m log m log 2 log m) (Elkin- Emek- S- Teng 04, Abraham- Bartal- Neiman 08) Conjecture: stretch T (G) m log 2 m

81 Algebraic characterization of stretch [S- Woo 09] stretch T (G) = Trace[L G L 1 T ]

82 Algebraic characterization of stretch [S- Woo 09] stretch T (G) = Trace[L G L 1 T ] Resistances in series sum In trees, resistance is distance. 1 1 v : 0 1 a 2 2 T b

83 Algebraic characterization of stretch [S- Woo 09] stretch T (G) = Trace[L G L 1 x T L G x = (x(a) x(b)) 2 (a,b) E = ((e a e b ) T x) 2 (a,b) E = x T (e a e b )(e a e b ) T x (a,b) E = x T ( (e a e b )(e a e b ) T )x (a,b) E T ]

84 Algebraic characterization of stretch [S- Woo 09] stretch T (G) = Trace[L G L 1 T ] Trace[L G L 1 T ]= (a,b) E = (a,b) E = (a,b) E Trace[(e a e b )(e a e b ) T L 1 T ] Trace[(e a e b ) T L 1 T (e a e b )] (e a e b ) T L 1 T (e a e b )

85 Algebraic characterization of stretch [S- Woo 09] (a,b) E stretch T (G) = Trace[L G L 1 T ] (e a e b ) T L 1 T (e a e b )= (a,b) E = (a,b) E R eff (a, b) stretch T (a, b)

86 Notable Things I ve left out Behavior under graph transformations Graph Isomorphism Random Walks and Diffusion PageRank and Hits Matrix-Tree Theorem Special Graphs (Cayley, Strongly-Regular, etc.) Diameter bounds Colin de Verdière invariant Discretizations of Manifolds

87 The next two talks Tomorrow: Solving equations in Laplacians in nearly-linear time. Preconditioning Sparsification Low-Stretch Spanning Trees Local graph partitioning

88 The next two talks Thursday: Existence of sparse approximations. A theorem in linear algebra and some of its connections.