Spectral Graph Theory, Linear Solvers, and...

Post on 29-Jun-2020

7 views 1 download

transcript

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Spectral Graph Theory,Linear Solvers, and

Applications

Gary Miller

Carnegie Mellon Universityjoiny work with Yiannis Koutis and David Tolliver

Theory and Practice of Computational LearningJune 9, 2009

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Linear Systems

3x 2y −z = 32x −5y 4z = 7−x 1/2y 2z = 2

Fundamental ConstraintSystem

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Matrix Form

3 2 −12 −5 4

−1 1/2 2

xyz

=

372

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Solving the General Case

• Dense Case: O(n3) or O(n2.81) Strassen.

• Sparse Case: Still O(n2.81).

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Solving the General Case

• Dense Case: O(n3) or O(n2.81) Strassen.• Sparse Case: Still O(n2.81).

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

An Easy Case

• Upper and Lower Triangular Systems

• O(m) time where m = number of nonzerosentries.

• Goal: Find more easy cases.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

An Easy Case

• Upper and Lower Triangular Systems• O(m) time where m = number of nonzeros

entries.

• Goal: Find more easy cases.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

An Easy Case

• Upper and Lower Triangular Systems• O(m) time where m = number of nonzeros

entries.• Goal: Find more easy cases.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Symmetric Matrices

• Assume A is symmetric, A = AT .

• Assume A is positive definite, xTAx > 0 forx 6= 0

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Symmetric Matrices

• Assume A is symmetric, A = AT .• Assume A is positive definite, xTAx > 0 forx 6= 0

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Direct MethodsGaussian Elimination Matrices

• Goal: algorithms that minimize work andspace.

• Trick: View nonzero entries as an undirectedgraph and view pivoting as a graphoperation.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Direct MethodsGaussian Elimination Matrices

• Goal: algorithms that minimize work andspace.

• Trick: View nonzero entries as an undirectedgraph and view pivoting as a graphoperation.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Pivoting

Viewing pivoting as a graph operation.

• Let G = (V, E) be a graph and v a vertex.

• PIV OT (v) :• Make a clique out of neighbors of v.• Remove v.

• Fill: New edges formed.• Work: All edges “touched”.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Pivoting

Viewing pivoting as a graph operation.

• Let G = (V, E) be a graph and v a vertex.• PIV OT (v) :

• Make a clique out of neighbors of v.• Remove v.

• Fill: New edges formed.• Work: All edges “touched”.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Pivoting

Viewing pivoting as a graph operation.

• Let G = (V, E) be a graph and v a vertex.• PIV OT (v) :

• Make a clique out of neighbors of v.• Remove v.

• Fill: New edges formed.

• Work: All edges “touched”.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Pivoting

Viewing pivoting as a graph operation.

• Let G = (V, E) be a graph and v a vertex.• PIV OT (v) :

• Make a clique out of neighbors of v.• Remove v.

• Fill: New edges formed.• Work: All edges “touched”.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Good Pivot Strategies

1970s and 1980s• Planar systems: O(n3/2) work and O(n log n)

fill/space.

• 3D Systems: O(n2) work and O(n3/2)fill/space.

• O(n3/2) space is a problem for ML sizeproblems.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Good Pivot Strategies

1970s and 1980s• Planar systems: O(n3/2) work and O(n log n)

fill/space.• 3D Systems: O(n2) work and O(n3/2)

fill/space.

• O(n3/2) space is a problem for ML sizeproblems.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Good Pivot Strategies

1970s and 1980s• Planar systems: O(n3/2) work and O(n log n)

fill/space.• 3D Systems: O(n2) work and O(n3/2)

fill/space.• O(n3/2) space is a problem for ML size

problems.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Pure Iterative Methods

Solving Ax = b.• Basic method: x(i+1) = (I − A)x(i) + b

• Convergence/Rate is determined by ||I −A||.• Accelerated Methods: Chebyshev Iteration,

Conjugate Gradient.• CG is still too slow.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Pure Iterative Methods

Solving Ax = b.• Basic method: x(i+1) = (I − A)x(i) + b

• Convergence/Rate is determined by ||I −A||.

• Accelerated Methods: Chebyshev Iteration,Conjugate Gradient.

• CG is still too slow.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Pure Iterative Methods

Solving Ax = b.• Basic method: x(i+1) = (I − A)x(i) + b

• Convergence/Rate is determined by ||I −A||.• Accelerated Methods: Chebyshev Iteration,

Conjugate Gradient.

• CG is still too slow.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Pure Iterative Methods

Solving Ax = b.• Basic method: x(i+1) = (I − A)x(i) + b

• Convergence/Rate is determined by ||I −A||.• Accelerated Methods: Chebyshev Iteration,

Conjugate Gradient.• CG is still too slow.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Preconditioned Iterative Methods

Solving B−1Ax = B−1b = b′.• Basic method: x(i+1) = (I −B−1A)x(i) + b′

• Computing the term z = B−1Ax(i)

• y = Ax(i) Forward Multiply• Bz = y Solve the preconditioner system

• Goal: Minimize the number of iteration whileminimizing the cost of the solve.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Preconditioned Iterative Methods

Solving B−1Ax = B−1b = b′.• Basic method: x(i+1) = (I −B−1A)x(i) + b′

• Computing the term z = B−1Ax(i)

• y = Ax(i) Forward Multiply• Bz = y Solve the preconditioner system

• Goal: Minimize the number of iteration whileminimizing the cost of the solve.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Preconditioned Iterative Methods

Solving B−1Ax = B−1b = b′.• Basic method: x(i+1) = (I −B−1A)x(i) + b′

• Computing the term z = B−1Ax(i)

• y = Ax(i) Forward Multiply• Bz = y Solve the preconditioner system

• Goal: Minimize the number of iteration whileminimizing the cost of the solve.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Classic Preconditioners

• Jacobi: B = Diagonal(A).

• Gauss-Seidel: B = UpperTriangular(A).• SSOR: B = (L + 1

ωD) 1ωD(L + 1

ωD)

• Still too slow and unreliable.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Classic Preconditioners

• Jacobi: B = Diagonal(A).• Gauss-Seidel: B = UpperTriangular(A).

• SSOR: B = (L + 1ωD) 1

ωD(L + 1ωD)

• Still too slow and unreliable.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Classic Preconditioners

• Jacobi: B = Diagonal(A).• Gauss-Seidel: B = UpperTriangular(A).• SSOR: B = (L + 1

ωD) 1ωD(L + 1

ωD)

• Still too slow and unreliable.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Classic Preconditioners

• Jacobi: B = Diagonal(A).• Gauss-Seidel: B = UpperTriangular(A).• SSOR: B = (L + 1

ωD) 1ωD(L + 1

ωD)

• Still too slow and unreliable.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Graph LaplaciansApplications

Graph Laplacian• G = (V,E, w) weighted undirected graph, wij > 0.

• Weighted incidence matrix:

Aij ={

wij if eij ∈ E0 otherwise

• Degree of vi: di =∑

j wij

D =

d1 0. . .

0 dn

• Laplacian: L = D −A

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Graph LaplaciansApplications

Graph Laplacian• G = (V,E, w) weighted undirected graph, wij > 0.• Weighted incidence matrix:

Aij ={

wij if eij ∈ E0 otherwise

• Degree of vi: di =∑

j wij

D =

d1 0. . .

0 dn

• Laplacian: L = D −A

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Graph LaplaciansApplications

Graph Laplacian• G = (V,E, w) weighted undirected graph, wij > 0.• Weighted incidence matrix:

Aij ={

wij if eij ∈ E0 otherwise

• Degree of vi: di =∑

j wij

D =

d1 0. . .

0 dn

• Laplacian: L = D −A

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Graph LaplaciansApplications

Graph Laplacian• G = (V,E, w) weighted undirected graph, wij > 0.• Weighted incidence matrix:

Aij ={

wij if eij ∈ E0 otherwise

• Degree of vi: di =∑

j wij

D =

d1 0. . .

0 dn

• Laplacian: L = D −A

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Graph LaplaciansApplications

Graph Laplacian• G = (V,E, w) weighted undirected graph, wij > 0.• Weighted incidence matrix:

Aij ={

wij if eij ∈ E0 otherwise

• Degree of vi: di =∑

j wij

D =

d1 0. . .

0 dn

• Laplacian: L = D −A

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Graph LaplaciansApplications

Classic Applications of the Laplacian

• View each edge a conductor with conductance wij .

• Let V be a column vector of voltages• If LV = c the c residual current needed to maintain the

given voltages.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Graph LaplaciansApplications

Classic Applications of the Laplacian

• View each edge a conductor with conductance wij .• Let V be a column vector of voltages

• If LV = c the c residual current needed to maintain thegiven voltages.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Graph LaplaciansApplications

Classic Applications of the Laplacian

• View each edge a conductor with conductance wij .• Let V be a column vector of voltages• If LV = c the c residual current needed to maintain the

given voltages.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Graph LaplaciansApplications

Graph Laplacian’s and the HeatEquations

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Graph LaplaciansApplications

Graph Laplacian’s and RandomWalks

Transition Matrix: D−1LG

Fundamental Eigenvectors: O(n + m) (Spielman Teng)Trick: Inverse Powering only requires O(log n) iterations.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Graph LaplaciansApplications

Laplacian’s and Spring Mass Systems

• G = (V, E, w) weighted graph and wij isviewed a spring constant.

• M is a diagonal matrix of mass constants• Fact: Modes of vibration of Spring-Mass

system G, M are:Eigen-pairs of LGx = λMx.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Graph LaplaciansApplications

Laplacian’s and Spring Mass Systems

• G = (V, E, w) weighted graph and wij isviewed a spring constant.

• M is a diagonal matrix of mass constants

• Fact: Modes of vibration of Spring-Masssystem G, M are:Eigen-pairs of LGx = λMx.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Graph LaplaciansApplications

Laplacian’s and Spring Mass Systems

• G = (V, E, w) weighted graph and wij isviewed a spring constant.

• M is a diagonal matrix of mass constants• Fact: Modes of vibration of Spring-Mass

system G, M are:Eigen-pairs of LGx = λMx.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Graph LaplaciansApplications

Spring Mass System

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Graph LaplaciansApplications

Graph Laplacian’s and MaximumFlow

Graph Maximum Flow: O((m + n)3/2) (Daitch Spielman)

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Graph LaplaciansApplications

Graph Laplacian’s and ConvexProgramming

Nonuniform TV Denoising: O((m + n)3/2) (Koutis M SinopTolliver)

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Spanning Tree Preconditioners• Vaidya ’93: Use Maximum Weight Spanning

Tree (MST) plus a few edges.

b

c

a

d

e f g

h

j

i

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Spanning Tree Preconditioners• Advantages: Easy to find and Easy to solve

their systems.

b

c

a

d

e f g

h

j

i

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Spanning Tree Preconditioners• Problem: Small edge weights differences

can make MST bad.

b

c

a

d

e f g

h

j

i

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Low Stretch Spanning Trees• EEST ’05: Use low stretch spanning trees

plus a few edges.

b

c

a

d

e f g

h

j

i

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Low Stretch Spanning Trees• Advantages: Better condition number.

b

c

a

d

e f g

h

j

i

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Low Stretch Spanning Trees• Problem: Super linear time to find.

b

c

a

d

e f g

h

j

i

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Low Stretch Spanning Trees• Richter-M ’04: There are no good spanning

trees even for square mesh.

b

c

a

d

e f g

h

j

i

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Steiner Tree Preconditioners

• Gremban-M ’94: Use Steiner trees.

b

c

a

d

e f g

h

j

i

g

a b

c d

h i je

f

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Steiner Tree Preconditioners

• Advantages: Better condition number forgraphs like square mesh.

b

c

a

d

e f g

h

j

i

g

a b

c d

h i je

f

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Steiner Tree Preconditioners

• Advantages: Good experimental results.

b

c

a

d

e f g

h

j

i

g

a b

c d

h i je

f

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Steiner Tree Preconditioners

• Problem: Hard to construct in general andanalyze.

b

c

a

d

e f g

h

j

i

g

a b

c d

h i je

f

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Steiner Forest Preconditioners

• Koutis-M ’07: Use Steiner Forest.

b

c

a

d

e f g

h

j

i

g

a b

c d

h i je

f

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Steiner Forest Preconditioners

• Advantages: Easy to find and works wellwith recursive solvers.

b

c

a

d

e f g

h

j

i

g

a b

c d

h i je

f

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Steiner Forest Preconditioners

• Advantages: Good experimental results.

b

c

a

d

e f g

h

j

i

g

a b

c d

h i je

f

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Steiner Forest Preconditioners

• Problem: Analysis only for planar systems.

b

c

a

d

e f g

h

j

i

g

a b

c d

h i je

f

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Recursive Methods

• Vaidya ’93: Idea: Tree have a lot of degree1-2 degree nodes. Pivot on these nodes andthen find a preconditioner for this graph.

d h i jf ga b ec

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Recursive Methods

• Vaidya ’93: Idea: Tree have a lot of degree1-2 degree nodes. Pivot on these nodes andthen find a preconditioner for this graph.

The reduced graph.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

The Error

• Recurrence: u(i+1) = Gu(i) + b, for usG = I − A.

• Error: e(i) = u(i) − u, where u = Gu + b

• Fact: e(i) = Gie(0)

• We need: limi→∞ Gi = 0

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

The Error

• Recurrence: u(i+1) = Gu(i) + b, for usG = I − A.

• Error: e(i) = u(i) − u, where u = Gu + b

• Fact: e(i) = Gie(0)

• We need: limi→∞ Gi = 0

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

The Error

• Recurrence: u(i+1) = Gu(i) + b, for usG = I − A.

• Error: e(i) = u(i) − u, where u = Gu + b

• Fact: e(i) = Gie(0)

• We need: limi→∞ Gi = 0

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

The Error

• Recurrence: u(i+1) = Gu(i) + b, for usG = I − A.

• Error: e(i) = u(i) − u, where u = Gu + b

• Fact: e(i) = Gie(0)

• We need: limi→∞ Gi = 0

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Condition Number

• Ax = λx

• DEF: λ Eigenvalue and x Eigenvector.• Λ(A) = {0 ≤ λ1 ≤ · · ·λn}.• Condition Number: κ(A) = λn/λ1

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Condition Number

• Ax = λx

• DEF: λ Eigenvalue and x Eigenvector.

• Λ(A) = {0 ≤ λ1 ≤ · · ·λn}.• Condition Number: κ(A) = λn/λ1

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Condition Number

• Ax = λx

• DEF: λ Eigenvalue and x Eigenvector.• Λ(A) = {0 ≤ λ1 ≤ · · ·λn}.

• Condition Number: κ(A) = λn/λ1

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Condition Number

• Ax = λx

• DEF: λ Eigenvalue and x Eigenvector.• Λ(A) = {0 ≤ λ1 ≤ · · ·λn}.• Condition Number: κ(A) = λn/λ1

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Convergence Rates

• Basic Method: Convergence Rate= O(1/κ(A))

• Conjugate Gradient: O(1/√

κ(A))

• Conjugate Gradient: ≈ 1/diameter(A).

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Convergence Rates

• Basic Method: Convergence Rate= O(1/κ(A))

• Conjugate Gradient: O(1/√

κ(A))

• Conjugate Gradient: ≈ 1/diameter(A).

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Convergence Rates

• Basic Method: Convergence Rate= O(1/κ(A))

• Conjugate Gradient: O(1/√

κ(A))

• Conjugate Gradient: ≈ 1/diameter(A).

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Generalized Condition Number

• Goal: Bound condition number of B−1A

• Note: B−1Ax = λx iff Ax = λBx

• DEF: λ is a generalized eigenvalue.• Condition Number: κ(B−1A) = λn/λ1

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Generalized Condition Number

• Goal: Bound condition number of B−1A

• Note: B−1Ax = λx iff Ax = λBx

• DEF: λ is a generalized eigenvalue.• Condition Number: κ(B−1A) = λn/λ1

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Generalized Condition Number

• Goal: Bound condition number of B−1A

• Note: B−1Ax = λx iff Ax = λBx

• DEF: λ is a generalized eigenvalue.

• Condition Number: κ(B−1A) = λn/λ1

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Generalized Condition Number

• Goal: Bound condition number of B−1A

• Note: B−1Ax = λx iff Ax = λBx

• DEF: λ is a generalized eigenvalue.• Condition Number: κ(B−1A) = λn/λ1

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

The Support

• Positive Semi-Definite: ∀x xTAx ≥ 0

• Support of A by B:σ(A/B) = min{τ : τB − A is PSD}

• Fact: κ(A, B) = σ(A, B) · σ(B, A).

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

The Support

• Positive Semi-Definite: ∀x xTAx ≥ 0

• Support of A by B:σ(A/B) = min{τ : τB − A is PSD}

• Fact: κ(A, B) = σ(A, B) · σ(B, A).

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

The Support

• Positive Semi-Definite: ∀x xTAx ≥ 0

• Support of A by B:σ(A/B) = min{τ : τB − A is PSD}

• Fact: κ(A, B) = σ(A, B) · σ(B, A).

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Estimating Support

• G and H are graphs and V (G) = V (H)

• Path Embedding: φ : E(G) then paths(H)s.t. φ(eij) = Vi · · ·Vj.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

Estimating Support ExampleEG: G ≡ K4 and H ≡ 4-cycle

Congestion ≡ 3 and Dilation ≡ 2Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

History of Planar Solvers

• 1950’s O(n2) (Conjugate Gradient)

• 1970’s O(n1.5) (Nested Dissection) (LRT)• 1990’s O(n1.2) (Combinatorial

Preconditioners) (Vaidya)• 2000’s O(n log2 n) (Low stretch spanning

trees) (ST)• 2006’s O(n) (separator based

preconditioners) (KM)

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

History of Planar Solvers

• 1950’s O(n2) (Conjugate Gradient)• 1970’s O(n1.5) (Nested Dissection) (LRT)

• 1990’s O(n1.2) (CombinatorialPreconditioners) (Vaidya)

• 2000’s O(n log2 n) (Low stretch spanningtrees) (ST)

• 2006’s O(n) (separator basedpreconditioners) (KM)

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

History of Planar Solvers

• 1950’s O(n2) (Conjugate Gradient)• 1970’s O(n1.5) (Nested Dissection) (LRT)• 1990’s O(n1.2) (Combinatorial

Preconditioners) (Vaidya)

• 2000’s O(n log2 n) (Low stretch spanningtrees) (ST)

• 2006’s O(n) (separator basedpreconditioners) (KM)

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

History of Planar Solvers

• 1950’s O(n2) (Conjugate Gradient)• 1970’s O(n1.5) (Nested Dissection) (LRT)• 1990’s O(n1.2) (Combinatorial

Preconditioners) (Vaidya)• 2000’s O(n log2 n) (Low stretch spanning

trees) (ST)

• 2006’s O(n) (separator basedpreconditioners) (KM)

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

History of Planar Solvers

• 1950’s O(n2) (Conjugate Gradient)• 1970’s O(n1.5) (Nested Dissection) (LRT)• 1990’s O(n1.2) (Combinatorial

Preconditioners) (Vaidya)• 2000’s O(n log2 n) (Low stretch spanning

trees) (ST)• 2006’s O(n) (separator based

preconditioners) (KM)

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

General Laplacian Solver

O(n + m) (Spielman Teng)

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.80

2

4

6

8

10

12

14

16

18

20

Number of Pixels (in millions)

Ru

nn

ing

Tim

e (s

ecs)

Two dimensional images

Our solverMATLAB’s direct solver

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Combinatorial PreconditionersRecursive Preconditioned MethodsAnalysisRun Times

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.90

2

4

6

8

10

12

Number of Pixels (in millions)

Ru

nn

ing

Tim

e (s

eco

nd

s)Three dimensional images

Our solverdirect solver

Out of memory

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Symmetric Diagonally DominateSystems

• Goal: Show how to solve SDD using regularLaplacians.

• Let G = (V, E, w) such that ij ∈ E:wij 6= 0• Weighted incidence matrix: A.• Degree of vi: di =

∑j |wij|

• Generalized Laplacian: L = D − A• Note: Every SDD is a Generalized

Laplacian.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Symmetric Diagonally DominateSystems

• Goal: Show how to solve SDD using regularLaplacians.

• Let G = (V, E, w) such that ij ∈ E:wij 6= 0

• Weighted incidence matrix: A.• Degree of vi: di =

∑j |wij|

• Generalized Laplacian: L = D − A• Note: Every SDD is a Generalized

Laplacian.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Symmetric Diagonally DominateSystems

• Goal: Show how to solve SDD using regularLaplacians.

• Let G = (V, E, w) such that ij ∈ E:wij 6= 0• Weighted incidence matrix: A.

• Degree of vi: di =∑

j |wij|• Generalized Laplacian: L = D − A• Note: Every SDD is a Generalized

Laplacian.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Symmetric Diagonally DominateSystems

• Goal: Show how to solve SDD using regularLaplacians.

• Let G = (V, E, w) such that ij ∈ E:wij 6= 0• Weighted incidence matrix: A.• Degree of vi: di =

∑j |wij|

• Generalized Laplacian: L = D − A• Note: Every SDD is a Generalized

Laplacian.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Symmetric Diagonally DominateSystems

• Goal: Show how to solve SDD using regularLaplacians.

• Let G = (V, E, w) such that ij ∈ E:wij 6= 0• Weighted incidence matrix: A.• Degree of vi: di =

∑j |wij|

• Generalized Laplacian: L = D − A

• Note: Every SDD is a GeneralizedLaplacian.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Symmetric Diagonally DominateSystems

• Goal: Show how to solve SDD using regularLaplacians.

• Let G = (V, E, w) such that ij ∈ E:wij 6= 0• Weighted incidence matrix: A.• Degree of vi: di =

∑j |wij|

• Generalized Laplacian: L = D − A• Note: Every SDD is a Generalized

Laplacian.Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Basic Properties: GeneralizedLaplacian

• xTLx =∑wij>0 wij(xi − xj)

2 −∑

wij<0 wij(xi + xj)2

• Thus L is positive semidefinite.• Claim: Rank = n− 1 if G is connected.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Basic Properties: GeneralizedLaplacian

• xTLx =∑wij>0 wij(xi − xj)

2 −∑

wij<0 wij(xi + xj)2

• Thus L is positive semidefinite.

• Claim: Rank = n− 1 if G is connected.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Basic Properties: GeneralizedLaplacian

• xTLx =∑wij>0 wij(xi − xj)

2 −∑

wij<0 wij(xi + xj)2

• Thus L is positive semidefinite.• Claim: Rank = n− 1 if G is connected.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Solving Generalized Laplacian byChange of Variables

• First Idea: Find a change of variables it get aregular Laplacian.

• Note: multiplying the ith column and row by−1 preserves Laplacian.

• This is just Flipping xi and bi to −xi and −bi.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Solving Generalized Laplacian byChange of Variables

• First Idea: Find a change of variables it get aregular Laplacian.

• Note: multiplying the ith column and row by−1 preserves Laplacian.

• This is just Flipping xi and bi to −xi and −bi.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Solving Generalized Laplacian byChange of Variables

• First Idea: Find a change of variables it get aregular Laplacian.

• Note: multiplying the ith column and row by−1 preserves Laplacian.

• This is just Flipping xi and bi to −xi and −bi.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Generalized Laplacians Example

+1

V1

−1

V2−1V3 2 +1 −1

+1 2 +1−1 +1 2

x1

x2

x3

=

b1

b2

b3

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Change of Variables for Laplacians

+1

V1

−1

V2−1V3

+1−1

2 −1 +1−1 2 +1+1 +1 2

−x1

x2

x3

=

−b1

b2

b3

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Orientable Generalized Laplacians

• DEF: G = (V, E, w) is orientable is ∃sequence of flips s.t. w > 0.

• DEF: LG is orientable if G is.• Note: Orientability is linear testable, greedy.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Orientable Generalized Laplacians

• DEF: G = (V, E, w) is orientable is ∃sequence of flips s.t. w > 0.

• DEF: LG is orientable if G is.

• Note: Orientability is linear testable, greedy.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Orientable Generalized Laplacians

• DEF: G = (V, E, w) is orientable is ∃sequence of flips s.t. w > 0.

• DEF: LG is orientable if G is.• Note: Orientability is linear testable, greedy.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Orientable Generalized Laplacians

• Claim: If G is connected and not orientablethen L is SPD.

• Proof: Suppose xTLx = 0 and x 6= 0

• Pick a spanning tree T of G and orient it andflipping x.

• WLOG: x is the all-ones vector, a contra!,since G still has negative edges.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Orientable Generalized Laplacians

• Claim: If G is connected and not orientablethen L is SPD.

• Proof: Suppose xTLx = 0 and x 6= 0

• Pick a spanning tree T of G and orient it andflipping x.

• WLOG: x is the all-ones vector, a contra!,since G still has negative edges.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Orientable Generalized Laplacians

• Claim: If G is connected and not orientablethen L is SPD.

• Proof: Suppose xTLx = 0 and x 6= 0

• Pick a spanning tree T of G and orient it andflipping x.

• WLOG: x is the all-ones vector, a contra!,since G still has negative edges.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Orientable Generalized Laplacians

• Claim: If G is connected and not orientablethen L is SPD.

• Proof: Suppose xTLx = 0 and x 6= 0

• Pick a spanning tree T of G and orient it andflipping x.

• WLOG: x is the all-ones vector, a contra!,since G still has negative edges.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Two-Fold Covers• DEF: The 2-fold cover G = (V , W , w) ofG = (V, W, w) is:

• V = {V1, V1, . . . , Vn, Vn}• E: If wij > 0

add edges < Vi, Vj > and < Vi, Vj > withweight wij

• E: If wij < 0add edges < Vi, Vj > and < Vi, Vj > withweight −wij

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Two-Fold Covers• DEF: The 2-fold cover G = (V , W , w) ofG = (V, W, w) is:

• V = {V1, V1, . . . , Vn, Vn}

• E: If wij > 0add edges < Vi, Vj > and < Vi, Vj > withweight wij

• E: If wij < 0add edges < Vi, Vj > and < Vi, Vj > withweight −wij

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Two-Fold Covers• DEF: The 2-fold cover G = (V , W , w) ofG = (V, W, w) is:

• V = {V1, V1, . . . , Vn, Vn}• E: If wij > 0

add edges < Vi, Vj > and < Vi, Vj > withweight wij

• E: If wij < 0add edges < Vi, Vj > and < Vi, Vj > withweight −wij

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Two-Fold Covers• DEF: The 2-fold cover G = (V , W , w) ofG = (V, W, w) is:

• V = {V1, V1, . . . , Vn, Vn}• E: If wij > 0

add edges < Vi, Vj > and < Vi, Vj > withweight wij

• E: If wij < 0add edges < Vi, Vj > and < Vi, Vj > withweight −wij

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Example: Two-Fold Cover

+1

V1

−1

V2−1V3

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Two-Fold Covers and SolvingGeneralized Laplacians

• Let L = L(G) and L = G

• Note:

Lx = b then L

(x

−x

)=

(b

−b

)•

L

(xy

)=

(b

−b

)then L(x/2− y/2) = b

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Two-Fold Covers and SolvingGeneralized Laplacians

• Let L = L(G) and L = G• L is a regular Laplacian which we can solve

quickly.

• Note:

Lx = b then L

(x

−x

)=

(b

−b

)•

L

(xy

)=

(b

−b

)then L(x/2− y/2) = b

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Two-Fold Covers and SolvingGeneralized Laplacians

• Let L = L(G) and L = G

• Note:

Lx = b then L

(x

−x

)=

(b

−b

)

L

(xy

)=

(b

−b

)then L(x/2− y/2) = b

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Two-Fold Covers and SolvingGeneralized Laplacians

• Let L = L(G) and L = G

• Note:

Lx = b then L

(x

−x

)=

(b

−b

)•

L

(xy

)=

(b

−b

)then L(x/2− y/2) = b

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

SDD systems

• The 2-fold trick can be run without the factorof two in space and time.

• There should be uses of negative weights inrecommendation problems.

• Naive approach does not seem to work right.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

SDD systems

• The 2-fold trick can be run without the factorof two in space and time.

• There should be uses of negative weights inrecommendation problems.

• Naive approach does not seem to work right.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

SDD systems

• The 2-fold trick can be run without the factorof two in space and time.

• There should be uses of negative weights inrecommendation problems.

• Naive approach does not seem to work right.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Spectral Graph Partitioning• Idea: Pick a few low frequency eigenvectors.

• Use these vectors to embed the graph in Rd

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Spectral Graph Partitioning• Idea: Pick a few low frequency eigenvectors.• Use these vectors to embed the graph in Rd

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

The Blue Sky ProblemShi Malik applied to an image:

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

The Blue Sky ProblemShi Malik solution:

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

The Blue Sky ProblemSpectral Rounding applied to Image:

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Spectral RoundingEdge Reweighting

Algorithm:• Solve Lf = λ2Df .• Reweight graph edges getting L′ and D′.• Solve L′f = λ2D

′f

• Repeat while λ2 6= 0.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

SR: The reweighting scheme

• View D, f , and λ2 as a function of L

• Subject to Lf = λ2Df and fTDf = 1.• We get: ∂λ2

∂eij= (fi − fj)

2 − λ2(f2i + f 2

j )

• Take a “small” step in the direction of thegradient.

• If an edge goes negative set it to zero.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Medical Examples of SR

Breast Tumors

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

Medical Examples of SR

Breast Tumors

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications

IntroductionGraph Based Methods

Iterative Methods for LaplaciansSymmetric Diagonally Dominate Systems

Spectral RoundingOpen Questions

• Find fast methods for any SPD system.• Find spectral methods that find better cut by

using more than one eigenvector.• Find solvers that work in the L2 norm.• A implementable solver with near linear time

guarantees.

Gary L. Miller Spectral Graph Theory, Linear Solvers, and Applications