Laplacian Paradigm 2 - sachdevasushant.github.io · Laplacian Paradigm 2.0 8:40-9:10: Merging...
Transcript of Laplacian Paradigm 2 - sachdevasushant.github.io · Laplacian Paradigm 2.0 8:40-9:10: Merging...
LaplacianParadigm2.08:40-9:10:MergingContinuousandDiscrete(RichardPeng)9:10-9:50:BeyondLaplacianSolvers(AaronSidford)9:50-10:30:ApproximateGaussianElimination(Sushant Sachdeva)10:30-11:00:coffeebreak11:00-12:00:AnalysisusingmatrixMartingales(RasmusKyng)12:00-14:00 lunch14:00-15:00GraphStructureviaEliminations(AaronSchild)
Website:bit.ly/laplacian2
LargeNetworks• Datamining:centrality,clustering…• Image/videoprocessing:segmentation,denoising …• Scientificcomputing:stress,fluids,waves…
• \ (linearsystemsolve)• CVX(convexoptimization)• Eigenvectorsolvers
GraphsandMatricesHighperformancecomputing:non-zerosó edges,design/analyzematrixalgorithmsusinggraphtheory
nrows/columnsO(m)non-zeros
nverticesmedges
2-1-1-110-101
GraphsandMatricesHighperformancecomputing:non-zerosó edges,design/analyzematrixalgorithmsusinggraphtheory
nrows/columnsO(m)non-zeros
1
1
nverticesmedges
2-1-1-110-101
graphLaplacianmatrixL• Diagonal:degrees• Off-diagonal:-edgeweights
SourceofLaplacians2-1-1-110-101
graphLaplacianmatrixL• Diagonal:degrees• Off-diagonal:-edgeweights
d-Regulargraphs:L =dI – A,A:adjacencymatrix
SourceofLaplacians
Graphcuts:xTLx =∑u~v wuv(xu - xv)2
2-1-1-110-101
graphLaplacianmatrixL• Diagonal:degrees• Off-diagonal:-edgeweights
d-Regulargraphs:L =dI – A,A:adjacencymatrix
xb=0
xa=1
xc=1
(1-0)2=1
(1-1)2=0
xindicatorvectorofcutè weightofcut
SourceofLaplacians
L =BTWBwhereB isedge-vertexincidencematrix
Graphcuts:xTLx =∑u~v wuv(xu - xv)2
2-1-1-110-101
graphLaplacianmatrixL• Diagonal:degrees• Off-diagonal:-edgeweights
d-Regulargraphs:L =dI – A,A:adjacencymatrix
xb=0
xa=1
xc=1
(1-0)2=1
(1-1)2=0
xindicatorvectorofcutè weightofcut
[Spielman Teng `04]Input:graphLaplacianL
vectorbOutput:vectorxs.t. Lx ≅ bRuntime:O(mlogO(1)nlog(1/ε))
OriginoftheLaplacianParadigm
[Spielman Teng `04]Input:graphLaplacianL
vectorbOutput:vectorxs.t. Lx ≅ bRuntime:O(mlogO(1)nlog(1/ε))
[Cohen-Kyng-Miller-Pachocki-P-Rao-Xu`14]:≤1/2
OriginoftheLaplacianParadigm
[Spielman Teng `04]Input:graphLaplacianL
vectorbOutput:vectorxs.t. Lx ≅ bRuntime:O(mlogO(1)nlog(1/ε))
[Cohen-Kyng-Miller-Pachocki-P-Rao-Xu`14]:≤1/2
OriginoftheLaplacianParadigm
Wallclock:m≤107 in≤20s
TheLaplacianParadigm
Directlyrelated:Elliptic systems
Fewiterations:Eigenvectors,Heatkernels
Manyiterations/modifyalgorithmGraphproblemsImageprocessing
Lx =b asagraphproblemx:voltagevectorsDual:electricalflowf
Unifiedformulation:minf withresdiual b ‖f‖p:• p=2:solvingLx =b• p=1:shortestpath/transshipment• p=∞:max-flow/min-cut
DirectMethods(combinatorial)
M(2) ß Eliminate(M(1),i1)M(3) ß Eliminate(M(2),i2)…
Repeatedlyremoveverticesbycreatingequivalentgraphsontheirneighborhoods
DirectMethods(combinatorial)
M(2) ß Eliminate(M(1),i1)M(3) ß Eliminate(M(2),i2)…
• Parallelgraphalgorithms• Matrixmultiplication/densesolves• Sparsified squaring
Repeatedlyremoveverticesbycreatingequivalentgraphsontheirneighborhoods
IterativeMethods(numerical)
SolveAx=b byxß x - (Ax – b)
• SimpleB:B =I,manyiterations• B =A:1iteration,butsameproblem
Fixedpoint:Ax – b =0
Preconditioning:SolveB-1Ax =B-1b by:xß x - B-1(Ax – b)
IterativeMethods(numerical)
SolveAx=b byxß x - (Ax – b)
• Krylov spacemethods/PCG• Convexoptimizationalgorithms
• SimpleB:B =I,manyiterations• B =A:1iteration,butsameproblem
Fixedpoint:Ax – b =0
Preconditioning:SolveB-1Ax =B-1b by:xß x - B-1(Ax – b)
HardinstancesDirectmethodscreatetoomuchfillonhighlyconnectedgraphs
Iterativemethodstaketoomanyiterationspaths
HardinstancesDirectmethodscreatetoomuchfillonhighlyconnectedgraphs
Iterativemethodstaketoomanyiterationspaths
Still`easy’bythemselves
Easyforiterativemethods Easyfordirectmethods
HardinstancesDirectmethodscreatetoomuchfillonhighlyconnectedgraphs
Iterativemethodstaketoomanyiterationspaths
Musthandlebothsimultaneously,butavoidpayingniterations⨉mperiteration
Still`easy’bythemselves
Easyforiterativemethods Easyfordirectmethods
Hybridalgorithms(aka.v1.0)• Scientificcomputing:iChol,multigrid• [Vaidya`89]preconditionwithgraphs
Focus:howtocombine• [Gemban-Miller`96]:spectralgraphtheory• [Spielman-Teng `04]:spectral(ultra-)sparsify
Key“glue”:sparsification[Spielman-Teng `04]: foranyG,canfindHwithO(nlogO(1)n)edgess.t. xTLGx ≈ xTLHx∀x
Key“glue”:sparsification[Spielman-Teng `04]: foranyG,canfindHwithO(nlogO(1)n)edgess.t. xTLGx ≈ xTLHx∀x
• Combinatorialparameter:#edges• Numericalparameter:approximations
Key“glue”:sparsification[Spielman-Teng `04]: foranyG,canfindHwithO(nlogO(1)n)edgess.t. xTLGx ≈ xTLHx∀x
• Combinatorialparameter:#edges• Numericalparameter:approximations
[Spielman-Srivatava`08]:samplebyeffectiveresistancesgivesHwithO(nlogn)edges
s
Max-Flowproblem
t
Maximumnumberofdisjoints-tpaths
Applications:• Routing• Scheduling
Recall:minf withresdiual b ‖f‖p:• p=2:solvingLx =b• p=∞:max-flow/min-cut
s
Max-Flowproblem
t
Maximumnumberofdisjoints-tpaths
Applications:• Routing• Scheduling
Dual: separatesandtbyremovingfewestedges
Applications:• Partitioning• Clustering
Recall:minf withresdiual b ‖f‖p:• p=2:solvingLx =b• p=∞:max-flow/min-cut
HybridAlgorithmsforMax-Flow[Daitch-Spielman `08][Christiano-Kelner-Madry-Spielman-Teng `10]:[Lee-Sidford `14]Max-flow/Min-cutvia(several)electricalflows
s t
Repeataboutm1/3 iters• Solvelinearsystems• Re-adjustedgeweights
HybridAlgorithmsforMax-Flow[Daitch-Spielman `08][Christiano-Kelner-Madry-Spielman-Teng `10]:[Lee-Sidford `14]Max-flow/Min-cutvia(several)electricalflows
s t
Repeataboutm1/3 iters• Solvelinearsystems• Re-adjustedgeweights
[Madry `10][Racke-Shah-Taubig `14]:cutapproximator /obliviousroutingO(no(1))-approx.inO(m1+o(1))
HybridAlgorithmsforMax-Flow[Daitch-Spielman `08][Christiano-Kelner-Madry-Spielman-Teng `10]:[Lee-Sidford `14]Max-flow/Min-cutvia(several)electricalflows
s t
Repeataboutm1/3 iters• Solvelinearsystems• Re-adjustedgeweights
[Madry `10][Racke-Shah-Taubig `14]:cutapproximator /obliviousroutingO(no(1))-approx.inO(m1+o(1))
[Lee-Rao-Srivastava`13][Sherman`13,`17][Kelner-Lee-Orecchia-Sidford `14]:Preconditioning,(1+ε)-approx
HybridAlgorithmsforMax-Flow[Daitch-Spielman `08][Christiano-Kelner-Madry-Spielman-Teng `10]:[Lee-Sidford `14]Max-flow/Min-cutvia(several)electricalflows
s t
Repeataboutm1/3 iters• Solvelinearsystems• Re-adjustedgeweights
[Madry `10][Racke-Shah-Taubig `14]:cutapproximator /obliviousroutingO(no(1))-approx.inO(m1+o(1))
[Lee-Rao-Srivastava`13][Sherman`13,`17][Kelner-Lee-Orecchia-Sidford `14]:Preconditioning,(1+ε)-approx
[P`16]:recurse themintoeachother:O(mlog41n),optimisticallymlog6n
LaplacianParadigm2.0
NewIntermediatestructures/theoremsmotivatedbytheoverallalgorithms
Motivatedbythegoalofhybridalgorithms,modifydirectanditerativemethods
Sparsified/ApproximateGaussianElimination
ExamplesDirectedgraphs/asymmetricmatrices
F
C A[CF]
SC(A,C)
A[CC]
A[FF] A[FC]
c
b
a
1
1
21
Partitioning/LocalizationsofRandomWalks
UnderthehoodMatrix(martingale) concentration
V1 V2
G[V1] Sc(G,V2)
Interactionswithdatastructures
NotcoveredL
[Kelner-Orecchia-Sidford-Zhu `13][Nanongkai-Saranuk `17][Wulff-Nilsen`17][Durfee-Kyng-Peebles-Rao-Sachdeva `17]
G
H
events
MatrixZoofromScientificComputing
[Boman-Hendrickson-Vavasis `04][Kyng-Lee-P-Sachdeva-Spielman `16][Kyng-Zhang `17][Kyng-P-Schweiterman-Zhang `18]
Questions
Approximateeliminationsbeyondspectralcondition#
Non-linear(preconditioned)iterativemethods
[Adil-Kyng-P-Sachdeva `19]:p-normiterativerefinment
Unreasonableeffectivenessofpcg(ichol(A),b),multigrid
Wist list Direct Iterative HybridConvex functions ? J JJ??Arbitrary values J ? ?LJ!Dynamic/streaming J L J???