Algorithm Design Using Spectral Graph Theory

download Algorithm  Design  Using  Spectral Graph Theory

If you can't read please download the document

description

Algorithm Design Using Spectral Graph Theory. Richard Peng. Joint Work with Guy Blelloch, HuiHan Chin, Anupam Gupta, Jon Kelner, Yiannis Koutis, Aleksander M ą dry, Gary Miller and Kanat Tangwongsan. Outline. Motivating problem: image denoising Fast solvers for SDD linear systems - PowerPoint PPT Presentation

Transcript of Algorithm Design Using Spectral Graph Theory

Slide 1

Algorithm Design Using Spectral Graph Theory

Richard PengJoint Work with Guy Blelloch, HuiHan Chin, Anupam Gupta, Jon Kelner, Yiannis Koutis, Aleksander Mdry, Gary Miller and Kanat Tangwongsan11OutlineMotivating problem: image denoisingFast solvers for SDD linear systemsUsing solver for L1 minimization and graph problems.2

Image DenoisingGiven image + noise, recover image.

CMUs space programRocket takes picture of moon surfaceNoisy picture, want to finding landing site from it.3Image Denoising: the Modeloriginal noiseless image.noise from some distribution added.input: original + noise, s.goal: recover original, x.Denoised Image:Noise:Input:s-xsxNoise is s-x in this setting4Explicit vs. Implicit ApproachesExplicitImplicitGoalRecover x directlyDefine conditions on x and s, solve for xBasic OperationAveraging a set of pixelsFilteringMinimize objective functionRuntimeO(n)O(n2) or higherQualityReasonableHighn > 106 for most imagesFirst give a simplified objective that can be optimized fast5Solution recovered has quality issues, will come back to this later.Simple Objective FunctionGradient: 2Ax 2sOptimal: 0 = 2Ax 2sAx = sEqual to xTAx-2sTx where x, s are length n vectors, A is n-by-n matrix

x = A-1sminimizei(xi-si)2 + i~j(xi-xj)2Same as minimizing quadratics in a single variable6Special Structure of AA is Symmetric Diagonally Dominant (SDD) if: Its symmetric In each row, diagonal entry at least sum of absolute values of all off diagonal entries

7OutlineMotivating problem: image denoisingFast solvers for SDD linear systemsUsing solver for L1 minimization and graph problems.8Fundamental Problem:Solving Linear SystemsGiven matrix A, vector bFind vector x such that Ax=b

Size of A:n-by-nm non-zero entriesFormally the running time can be characterized using n and m.9Solving Linear Systems:Explicit and ImplicitDirect (explicit)Iterative (implicit)Unit OperationModifying entryMatrix-vector multiplyMain goalOperations applied on matrix are reversibleExplored large portion of rank spaceCost per stepO(1)O(m)Numer of StepsO(n)O(n)Total RuntimeO(n)O(nm)10Explicit Algorithms[1st century CE] Gaussian Elimination: O(n3)[Strassen `69] O(n2.8)[Coppersmith-Winograd `90] O(n2.3755)[Stothers `10] O(n2.3737)[Vassilevska Williams`11] O(n2.3727)

Chinese math text, also studied by Newton11SDD Linear SystemsDirect (explicit)Iterative (implicit)Unit OperationModifying entryMatrix-vector multiplyMain ideaOperations applied on matrix are reversibleExplored large portion of rank spaceCost per stepO(1)O(m)Numer of StepsO(n)O(n)Total RuntimeO(n)O(nm)[Vaidya `91]: Hybrid methods12Nearly Linear Time Solvers[Spielman-Teng 04]

Input: n by n SDD matrix A with m non-zerosvector bWhere: b = Ax for some xOutput: Approximate solution x s.t.|x-x|A 106 Key issue: exponent on logAlgorithm isnt much faster than a O(n2) algorithm on reasonably sized problems16Practical Nearly Linear Time Solvers[Koutis-Miller-P `10, `11]

Input: n by n SDD matrix A with m non-zerosvector bWhere: b = Ax for some xOutput: Approximate solution x s.t.|x-x|A