Laplacian Matrices of...

100
Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program in Applied Mathematics Yale Institute for Network Science MTNS, July 7, 2014

Transcript of Laplacian Matrices of...

Page 1: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Laplacian Matrices of Graphs

Daniel A. Spielman Dept. of Computer Science

Program in Applied Mathematics Yale Institute for Network Science

MTNS, July 7, 2014

Page 2: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program
Page 3: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Outline

Examples of graphs and networks Physical metaphors for graphs and the Laplacian Some applications Algorithms and theorems

Page 4: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Graphs and Networks V: a set of vertices (nodes) E: a set of edges an edge is a pair of vertices

Dan

Donna Allan

Gary Maria

Nikhil

Shang-Hua

A small social network

Page 5: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Graphs and Networks

A big social network

Theorem: There is no nice picture of this network

Page 6: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

from: http://nrc.uchsc.edu/STATES/united-states-map.jpg

Page 7: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

from: http://nrc.uchsc.edu/STATES/united-states-map.jpg

Page 8: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program
Page 9: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Examples of Graphs

Page 10: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Nearest neighbor graphs of points

Page 11: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Nearest neighbor graphs of points

Page 12: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Nearest neighbor graphs from data

Page 13: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Nearest neighbor graphs from data

Page 14: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Cartoon Graphs

Path

Grid

Page 15: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Examples of Graphs

8  

5  

1  

7  3  2  

9  4   6  

10

Page 16: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

8  

5  

1  

7  3  2  

9  4   6  

10  

Dan  

Amy  Allan  

Gary  Maria  

Nikhil  

Shang-­‐Hua  

Examples of Graphs

Page 17: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

8  

5  

1  

7  3  2  

9  4   6  

10  

Dan  

Amy  Allan  

Gary  Maria  

Nikhil  

Shang-­‐Hua  

Examples of Graphs

Page 18: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

How to understand large-scale structure

Use physical metaphors Edges as rubber bands Edges as resistors

Examine processes

Diffusion of gas Spilling paint

Identify structures

Communities Hierarchy

Page 19: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Nail  down  some  verEces,  let  rest  seIle  

View  edges  as  rubber  bands  or  ideal  linear  springs        spring  constant  1  (for  now)  

Graphs as Spring Networks

Page 20: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Nail  down  some  verEces,  let  rest  seIle  

View  edges  as  rubber  bands  or  ideal  linear  springs        spring  constant  1  (for  now)  

Graphs as Spring Networks

When  stretched  to  length    potenEal  energy  is    

`

�2/2

Page 21: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Nail  down  some  verEces,  let  rest  seIle.  

Physics:  posiEon  minimizes  total  potenEal  energy  

subject  to  boundary  constraints  (nails)  

x(a)a

1

2

X

(a,b)2E

(x(a)� x(b))2

Graphs as Spring Networks

Page 22: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Nail  down  some  verEces,  let  rest  seIle  

Energy  minimized  when          free  verEces  are  averages  of  neighbors  

x(a)a

�x(a) =1

da

X

(a,b)2E

�x(b)

Graphs as Spring Networks

is degree of , the number of attached edges da a

Page 23: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

If  nail  down  a  face  of  a  planar  3-­‐connected  graph,  get  a  planar  embedding!  

Tutte’s Theorem ‘63

Page 24: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

3-­‐connected:          cannot  break  graph  by  cuYng  2  edges  

Tutte’s Theorem ‘63

Page 25: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

3-­‐connected:          cannot  break  graph  by  cuYng  2  edges  

Tutte’s Theorem ‘63

Page 26: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Tutte’s Theorem ‘63 3-­‐connected:          cannot  break  graph  by  cuYng  2  edges  

Page 27: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Tutte’s Theorem ‘63 3-­‐connected:          cannot  break  graph  by  cuYng  2  edges  

Page 28: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Graphs as Resistor Networks

View edges as resistors connecting vertices Apply voltages at some vertices. Measure induced voltages and current flow.

1V

0V

Page 29: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Graphs as Resistor Networks

View edges as resistors connecting vertices Apply voltages at some vertices. Measure induced voltages and current flow. Induced voltages minimize

X

(a,b)2E

(v(a)� v(b))2

Subject to fixed voltages (by battery)

Page 30: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Graphs as Resistor Networks

View edges as resistors connecting vertices Apply voltages at some vertices. Measure induced voltages and current flow.

1V

0V

0.5V

0.5V

0.625V 0.375V

Page 31: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Graphs as Resistor Networks

View edges as resistors connecting vertices Apply voltages at some vertices. Measure induced voltages and current flow. Current flow measures strength of connection between endpoints. More short disjoint paths lead to higher flow.

Page 32: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Learning on Graphs [Zhu-­‐Ghahramani-­‐Lafferty  ’03]

Infer values of a function at all vertices from known values at a few vertices.

Minimize

Subject to known values

X

(a,b)2E

(x(a)� x(b))2

0

1

Page 33: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Learning on Graphs [Zhu-­‐Ghahramani-­‐Lafferty  ’03]

Infer values of a function at all vertices from known values at a few vertices.

Minimize

Subject to known values

X

(a,b)2E

(x(a)� x(b))2

0

1 0.5

0.5

0.625 0.375

Page 34: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

The Laplacian quadratic form X

(a,b)2E

(x(a)� x(b))2

Page 35: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

x

TLx =

X

(a,b)2E

(x(a)� x(b))2

The Laplacian matrix of a graph

There is a symmetric matrix L so that

Page 36: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Weighted Graphs

Edge assigned a non-negative real weight measuring strength of connection spring constant 1/resistance

wa,b 2 R

x

TLx =

X

(a,b)2E

wa,b(x(a)� x(b))2

(a, b)

Page 37: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

The Laplacian matrix of a graph

x

TLx =

X

(a,b)2E

wa,b(x(a)� x(b))2

1! 2!

3!4!

Page 38: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

The Laplacian matrix of a graph

To minimize subject to boundary constraints, set derivative to zero. Solve equation of form

Lx = b

x

TLx =

X

(a,b)2E

wa,b(x(a)� x(b))2

Page 39: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Can solve Laplacian equations absurdly quickly

Lx = b

Asymptotically best algorithm (today) takes time

O(mplog n log(1/✏))

Where m is number of non-zeros and n is dimension (Cohen-Kyng-Pachocki-Peng-Rao ’14)

Page 40: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Can solve Laplacian equations absurdly quickly

Lx = b

Asymptotically best algorithm (today) takes time

O(mplog n log(1/✏))

Where m is number of non-zeros and n is dimension (Cohen-Kyng-Pachocki-Peng-Rao ’14)

Good code: LAMG (lean algebraic multigrid) – Livne-Brandt CMG (combinatorial multigrid) – Koutis

Page 41: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Measuring boundaries of sets

Boundary: edges leaving a set

S S

Page 42: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Measuring boundaries of sets

Boundary: edges leaving a set

S S

Page 43: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Measuring boundaries of sets

Boundary: edges leaving a set

S

0 0

0 0

0 0

1

1 0

1 1

1 1 1

0

0 0

1

S

Characteristic Vector of S:

x(a) =

(1 a in S

0 a not in S

Page 44: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Measuring boundaries of sets

Boundary: edges leaving a set

S

0 0

0 0

0 0

1

1 0

1 1

1 1 1

0

0 0

1

S

Characteristic Vector of S:

x(a) =

(1 a in S

0 a not in S

x

TLx =

X

(a,b)2E

(x(a)� x(b))

2= |boundary(S)|

Page 45: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Eigenvalues and Eigenvectors A n-by-n symmetric matrix has n real eigenvalues and eigenvectors such that

�1 � �2 · · · � �n

v1, ..., vn

Lvi = �ivi

Page 46: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Spectral Graph Theory A n-by-n symmetric matrix has n real eigenvalues and eigenvectors such that

�1 � �2 · · · � �n

v1, ..., vn

These eigenvalues and eigenvectors tell us a lot about a graph!

Lvi = �ivi

Page 47: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Spectral Graph Drawing [Hall ‘70]

31 2

4

56 7

8 9

Arbitrary Drawing

Page 48: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Spectral Graph Drawing Plot vertex at draw edges as straight lines

[Hall ‘70]

(v2(a), v3(a))a

31 2

4

56 7

8 9

12

4

5

6

9

3

8

7Arbitrary Drawing

Spectral Drawing

Page 49: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

A Graph

Page 50: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Drawing of the graph using v2, v3

Plot vertex at a (v2(a), v3(a))

Page 51: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program
Page 52: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

The Airfoil Graph, original coordinates

Page 53: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

The Airfoil Graph, spectral coordinates

Page 54: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

The Airfoil Graph, spectral coordinates

Page 55: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Spectral drawing of Streets in Rome

Page 56: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

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

Page 57: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Dodecahedron

Best embedded by first three eigenvectors

Page 58: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Graph Partitioning

Page 59: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Spectral Graph Partitioning

for some

[Donath-Hoffman ‘72, Barnes ‘82, Hagen-Kahng ‘92]

S = {a : v2(a) � t} t

Page 60: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Spectral Graph Partitioning

Performance guarantees come from Cheeger’s Inequality

[Donath-Hoffman ‘72, Barnes ‘82, Hagen-Kahng ‘92]

Page 61: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Spectral Image Segmentation (Shi-Malik ‘00)

200 400 600 800 1000 1200

100

200

300

400

500

600

700

800

900

Page 62: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Spectral Image Segmentation (Shi-Malik ‘00)

2 4 6 8 10 12 14 16

2

4

6

8

10

12

Page 63: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Spectral Image Segmentation (Shi-Malik ‘00)

2 4 6 8 10 12 14 16

2

4

6

8

10

12

Page 64: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Spectral Image Segmentation (Shi-Malik ‘00)

2 4 6 8 10 12 14 16

2

4

6

8

10

12

Page 65: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Spectral Image Segmentation (Shi-Malik ‘00)

2 4 6 8 10 12 14 16

2

4

6

8

10

12

edge weight

Page 66: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

The second eigenvector

50 100 150 200 250 300

50

100

150

200

Page 67: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Second eigenvector cut

50 100 150 200 250 300

50

100

150

200

Page 68: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Third Eigenvector

50 100 150 200 250 300

50

100

150

200

50 100 150 200 250 300

50

100

150

200

Page 69: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Fourth Eigenvector

50 100 150 200 250 300

50

100

150

200

50 100 150 200 250 300

50

100

150

200

Page 70: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Laplacian  Linear  Systems  Lx = b

Algorithms

Solve  in  Eme                        where            =  number  of  non-­‐zeros  entries  of  L                  

O(m log

c m)

m

(S-­‐Teng  ’04,  ‘14)

Page 71: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Laplacian  Linear  Systems  Lx = b

Algorithms

Solve  in  Eme                        where            =  number  of  non-­‐zeros  entries  of  L                      Emes                                  for      -­‐approximate  soluEon.  

O(m log

c m)

m

log(1/�) ✏

��x� L�1b��L �

��L�1b��L

(S-­‐Teng  ’04,  ‘14)

Page 72: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Laplacian  Linear  Systems  Lx = b

Algorithms

Solve  in  Eme                        where            =  number  of  non-­‐zeros  entries  of  L                      Emes                                  for      -­‐approximate  soluEon.  

O(m log

c m)

m

log(1/�) ✏

��x� L�1b��L �

��L�1b��L

kxkL =px

TLx

(S-­‐Teng  ’04,  ‘14)

Page 73: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Algorithms

Two  main  ingredients:      Graph  Sparsifiers  

   Low  Stretch  Spanning  Trees                      

(S-­‐Teng  ’04,  ‘14)

Page 74: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Approximating Graphs

for all

A graph H is an -approximation of G if ✏

1

1 + � xTLHx

xTLGx 1 + �

x

Page 75: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Approximating Graphs

for all x

A very strong notion of approximation Preserves all electrical and spectral properties Preserves boundaries of every set

A graph H is an -approximation of G if ✏

1

1 + � xTLHx

xTLGx 1 + �

Page 76: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Sparsifying Graphs

for all 1

1 + � xTLHx

xTLGx 1 + �

x

(Batson-S-Srivastava ’09) Every graph G has an -approximation H with edges

A graph H is an -approximation of G if

n(2 + ✏)2/✏2

Page 77: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

for all 1

1 + � xTLHx

xTLGx 1 + �

x

(Batson-S-Srivastava ’09) Every graph G has an -approximation H with edges (S-Srivastava ‘08) Can quickly find one with edges by sampling according to effective resistances

A graph H is an -approximation of G if

n(2 + ✏)2/✏2

Sparsifying Graphs ✏

Cn log n/✏2

Page 78: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

If                                  ,  kAk < 1 (I �A)�1 = I +A+A2 +A3 + · · ·

=Y

i�0

(I +A2i)

Solving Equations (S-­‐Peng  ‘14)  

Page 79: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

If                                  ,  kAk < 1 (I �A)�1 = I +A+A2 +A3 + · · ·

=Y

i�0

(I +A2i)

If                                  ,  kAk < 1 A2i ! 0

⇡kY

i=0

(I +A2i)

Sparsify    

Solving Equations (S-­‐Peng  ‘14)  

I +A2i

Page 80: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Spanning trees of graphs

Connected  No  cycles  Have  a  unique  path  between  every  two  verEces  

Page 81: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Connected  No  cycles  Have  a  unique  path  between  every  two  verEces  

Spanning trees of graphs

Page 82: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

The stretch of a spanning tree

stT (G) =X

(a,b)2E

path-lengthT (a, b)

(Alon-­‐Karp-­‐Peleg-­‐West  ’91)

Page 83: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

path-­‐len  5  

stT (G) =X

(a,b)2E

path-lengthT (a, b)

The stretch of a spanning tree

Page 84: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

path-­‐len  3  

stT (G) =X

(a,b)2E

path-lengthT (a, b)

The stretch of a spanning tree

Page 85: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

path-­‐len  1  

stT (G) =X

(a,b)2E

path-lengthT (a, b)

The stretch of a spanning tree

Page 86: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

stT (G) =X

(a,b)2E

path-lengthT (a, b)

For  every  G  there  is  a  T  with  stT (G) O(m log n log log n)

The stretch of a spanning tree

(Elkin-­‐Emek-­‐S-­‐Teng  ‘04,        Abraham-­‐Bartal-­‐Neiman  ’08,  Abraham-­‐Neiman  ‘12)  

Page 87: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

stT (G) =X

(a,b)2E

path-lengthT (a, b)

For  every  G  there  is  a  T  with  stT (G) O(m log n log log n)

And,  we  can  compute  it  in  Eme  O(m log n log log n)

(Elkin-­‐Emek-­‐S-­‐Teng  ‘04,        Abraham-­‐Bartal-­‐Neiman  ’08,  Abraham-­‐Neiman  ‘12)  

The stretch of a spanning tree

Page 88: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program
Page 89: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Kelner-­‐Orecchia-­‐Sidford-­‐Zhu  ‘13  

Want  the  flow  saEsfying  demands  that  minimizes    

Kirchoff:  is  the  flow  saEsfying  demands  such  that  

for  all  cycles  C  in  the  graph  

X

(a,b)2C

fa,bra,b = 0,

X

(a,b)2E

f2a,bra,b

Page 90: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Kelner-­‐Orecchia-­‐Sidford-­‐Zhu  ‘13  

Push  flow  around  a  cycle  unEl  it  saEsfies  KCL.        This  decreases  the  energy.    To  decide  which  cycles,    choose  a  low-­‐stretch  tree.  Each  non-­‐tree  edge    determines  a  cycle.  

X

(a,b)2C

fa,bra,b = 0,

Page 91: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Kelner-­‐Orecchia-­‐Sidford-­‐Zhu  ‘13  

Choose  off-­‐tree  edge  with  prob  

Push  flow  around  that  cycle  to      minimize  energy  and      saEsfy  Kirchoff.  

With  right  data  structure,            each  operaEon  takes  logarithmic  Eme.  

Solve  in  Eme  O(m log

2 m log logm log ✏�1)

pa,b ⇠ path-lengthT (a, b)

Page 92: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

1.0  

1.0  

1.0  

1.0   1.0  

1.0   1.0  

1.0  

0   0  

Kelner-­‐Orecchia-­‐Sidford-­‐Zhu  ‘13  

3.0  3.0  

Start  with  a  simple  flow  

Page 93: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

1.0  

1.0  

1.0  

1.0   1.0  

1.0   1.0  

1.0  

0   0  

Kelner-­‐Orecchia-­‐Sidford-­‐Zhu  ‘13  

3.0  3.0  

Minimize  energy  on  a  cycle  

Page 94: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

0.8  

0.8  

0.8  

1.2   1.2  

1.0   1.0  

1.0  

0   0  

Kelner-­‐Orecchia-­‐Sidford-­‐Zhu  ‘13  

3.0  3.0  

Minimize  energy  on  a  cycle                and  saEsfy  Kirchoff  

Page 95: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

0.8  

0.8  

0.8  

1.2   1.13  

1.0   1.07  

1.0  

0   0.07  

Kelner-­‐Orecchia-­‐Sidford-­‐Zhu  ‘13  

3.0  3.0  

Minimize  energy  on  another  cycle  

Page 96: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

0.8  

0.8  

0.8  

1.13   1.13  

1.07   1.07  

1.0  

0.07   0.07  

Kelner-­‐Orecchia-­‐Sidford-­‐Zhu  ‘13  

3.0  3.0  

Minimize  energy  on  another  cycle  

Page 97: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

0.8  

0.8  

0.8  

1.13   1.13  

1.07   1.07  

0.71  

0.36   0.36  

Kelner-­‐Orecchia-­‐Sidford-­‐Zhu  ‘13  

3.0  3.0  

Minimize  energy  on  another  cycle  

Page 98: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

0.83  

0.83  

0.83  

1.24   1.24  

0.93   0.93  

0.62  

0.31   0.31  

Kelner-­‐Orecchia-­‐Sidford-­‐Zhu  ‘13  

3.0  3.0  

Number  of  iteraEons  is  proporEonal  to  the  stretch  

Page 99: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

Fast Laplacian Solvers

Asymptotically best algorithm (today) takes time

O(mplog n log(1/✏))

(Cohen-Kyng-Pachocki-Peng-Rao ’14)

Good code: LAMG (lean algebraic multigrid) – Livne-Brandt CMG (combinatorial multigrid) – Koutis

A powerful computational primitive!

Page 100: Laplacian Matrices of Graphsfwn06.housing.rug.nl/.../wp-content/uploads/2014/09/DanielSpielman… · Laplacian Matrices of Graphs Daniel A. Spielman Dept. of Computer Science Program

My  web  page  on:  Laplacian linear equations, sparsification, local graph clustering, low-stretch spanning trees, and so on.

My class notes from “Graphs and Networks” and “Spectral Graph Theory”

Lx = b, by Nisheeth Vishnoi

To learn more