Related Sensitivity Notions โ Diff ffer erent ential ial Priva vacy cy [Dwork McSherry Nissim Smith '06] โ Edge Differential Privacy [Nissim Raskhodnikova Smith '07] An algorithm ๐ต on a graph ๐ป is differentially private if for all ๐ โ ๐น โ the distributions ๐ต(๐ป) and ๐ต(๐ป โ ๐) are close to each other โ Much stricter notion than average sensitivity โ Some of our algorithms inspired by differentially private algorithms 33
Related Sensitivity Notions โ Diff ffer erent ential ial Priva vacy cy [Dwork McSherry Nissim Smith '06] โ Edge Differential Privacy [Nissim Raskhodnikova Smith '07] An algorithm ๐ต on a graph ๐ป is differentially private if for all ๐ โ ๐น โ the distributions ๐ต(๐ป) and ๐ต(๐ป โ ๐) are close to each other โ Much stricter notion than average sensitivity โ Some of our algorithms inspired by differentially private algorithms โ Stabi bilit lity y of f Learnin rning g Algo gorit rithms hms [Bousquet Elisseeff '02] 34
Related Sensitivity Notions โ Diff ffer erent ential ial Priva vacy cy [Dwork McSherry Nissim Smith '06] โ Edge Differential Privacy [Nissim Raskhodnikova Smith '07] An algorithm ๐ต on a graph ๐ป is differentially private if for all ๐ โ ๐น โ the distributions ๐ต(๐ป) and ๐ต(๐ป โ ๐) are close to each other โ Much stricter notion than average sensitivity โ Some of our algorithms inspired by differentially private algorithms โ Stabi bilit lity y of f Learnin rning g Algo gorit rithms hms [Bousquet Elisseeff '02] โ A learner is stable if empirical loss does not change much by replacing any sample in the training data 35
Related Sensitivity Notions โ Diff ffer erent ential ial Priva vacy cy [Dwork McSherry Nissim Smith '06] โ Edge Differential Privacy [Nissim Raskhodnikova Smith '07] An algorithm ๐ต on a graph ๐ป is differentially private if for all ๐ โ ๐น โ the distributions ๐ต(๐ป) and ๐ต(๐ป โ ๐) are close to each other โ Much stricter notion than average sensitivity โ Some of our algorithms inspired by differentially private algorithms โ Stabi bilit lity y of f Learnin rning g Algo gorit rithms hms [Bousquet Elisseeff '02] โ A learner is stable if empirical loss does not change much by replacing any sample in the training data โ Stable learners have low generalization error 36
T alk Outline โ Our definition of average sensitivity for graph algorithms โ Key properties of our definition โ Main results โ Algorithm with low sensitivity for the global minimum cut problem โ Conclusions and open directions 37
k-Average Sensitivity from Average Sensitivity Theorem: em: If ๐ต has average sensitivity ๐(๐, ๐) , it has ๐ -average sensitivity at most ฯ ๐โ[๐] ๐(๐, ๐ โ ๐ + 1) . 38
Average Sensitivity Composes Algorithms ๐ต, ๐ถ, ๐ท such that ๐ต(๐ป) = ๐ถ(๐ป, ๐ท ๐ป ) 39
Average Sensitivity Composes Algorithms ๐ต, ๐ถ, ๐ท such that ๐ต(๐ป) = ๐ถ(๐ป, ๐ท ๐ป ) Theorem em (Info formal) mal): : Average sensitivity of ๐ต on ๐ป = (๐, ๐น) can be bounded by the sum of: a term for average sensitivity of ๐ถ , and โข a term for average sensitivity of ๐ท . โข 40
Average Sensitivity Composes Algorithms ๐ต, ๐ถ, ๐ท such that ๐ต(๐ป) = ๐ถ(๐ป, ๐ท ๐ป ) Theorem em (Info formal) mal): : Average sensitivity of ๐ต on ๐ป = (๐, ๐น) can be bounded by the sum of: a term for average sensitivity of ๐ถ , and โข a term for average sensitivity of ๐ท . โข Can be used to bound the average sensitivity of a distribution over multiple stable-on-average algorithms. 41
Connection to Deterministic ๐ป ๐ต(๐ป) Sublinear Algorithms Algorithm ๐ต 42
Connection to Deterministic ๐ป ๐ต(๐ป) Sublinear Algorithms Algorithm ๐ต Local simulator ๐ Graph ๐ป 43
Connection to Deterministic ๐ป ๐ต(๐ป) Sublinear Algorithms Algorithm ๐ต 1 if ๐ โ ๐ต(๐ป) Local ๐ โ ๐น simulator ๐ 0, otherwise Graph ๐ป 44
Connection to Deterministic ๐ป ๐ต(๐ป) Sublinear Algorithms Algorithm ๐ต ๐ ๐ป โ ๐ฝ ๐โ๐น [# queries by ๐] 1 if ๐ โ ๐ต(๐ป) Local ๐ โ ๐น simulator ๐ 0, otherwise Graph ๐ป 45
Connection to Deterministic ๐ป ๐ต(๐ป) Sublinear Algorithms Algorithm ๐ต ๐ ๐ป โ ๐ฝ ๐โ๐น [# queries by ๐] 1 if ๐ โ ๐ต(๐ป) Local ๐ โ ๐น simulator ๐ Avera erage ge se sensiti sitivit vity y of f ๐ต 0, otherwise on on ๐ป is s โค ๐(๐ป) Graph ๐ป 46
๐ is the random string Connection to ๐ป Algorithm ๐ต ๐ต ๐ (๐ป) Sublinear Algorithms ๐ ๐ ๐ป โ ๐ฝ ๐,๐โ๐น [# queries by ๐] ๐ โ ๐น 1 if ๐ โ ๐ต ๐ (๐ป) Local simulator ๐ Avera erage ge se sensiti sitivit vity y of f ๐ต 0, otherwise ๐ on on ๐ป is s โค ๐(๐ป) Graph ๐ป 47
Graph problem ๐ Connection to Local Computation 1 if ๐ค is part of LCA ๐ a solution to ๐ Algorithms (LCAs) ๐ค โ ๐ on ๐ป ๐ โ 0,1 ๐ 0, otherwise Graph ๐ป Answers of ๐ are consistent with a single feasible solution of ๐ on ๐ป 48
Graph problem ๐ Connection to Local Computation 1 if ๐ค is part of LCA ๐ a solution to ๐ Algorithms (LCAs) ๐ค โ ๐ on ๐ป ๐ โ 0,1 ๐ 0, otherwise If f a pr probl blem m ๐ has s an LCA of f qu query co comp mplexit xity y ๐(๐ป) , then it Graph ๐ป has s an algo gorit ithm hm with avera rage ge sensi se sitivi ivity y โค ๐(๐ป) Answers of ๐ are consistent with a single feasible solution of ๐ on ๐ป 49
Graph problem ๐ Connection to Local Computation 1 if ๐ค is part of LCA ๐ a solution to ๐ Algorithms (LCAs) ๐ค โ ๐ on ๐ป ๐ โ 0,1 ๐ 0, otherwise If f a pr probl blem m ๐ has s an LCA of f qu query co comp mplexit xity y ๐(๐ป) , then it Graph ๐ป has s an algo gorit ithm hm with avera rage ge se sensi sitivi ivity y โค ๐(๐ป) Answers of ๐ are consistent with a Lower r bo bound d on average ge single feasible solution of ๐ on ๐ป sensi se sitivi ivity y imp mplies es lower er bo bound d on LCA A qu query y co comp mplexit xity! y! 50
T alk Outline โ Our definition of average sensitivity for graph algorithms โ Key properties of our definition โ Main results โ Algorithm with low sensitivity for the global minimum cut problem โ Conclusions and open directions 51
Minimum Spanning Forest For graphs on ๐ vertices and ๐ edges Algo gorit rithm hm Avera erage ge Sensit nsitivi ivity Krusk skal' al's s Algo gorit rithm hm Prim' m's s Algo gorit rithm hm 52
Minimum Spanning Forest For graphs on ๐ vertices and ๐ edges Algo gorit rithm hm Avera erage ge Sensit nsitivi ivity Krusk skal' al's s Algo gorit rithm hm ๐(๐/๐) Prim' m's s Algo gorit rithm hm 53
Minimum Spanning Forest For graphs on ๐ vertices and ๐ edges Algo gorit rithm hm Avera erage ge Sensit nsitivi ivity Krusk skal' al's s Algo gorit rithm hm ๐(๐/๐) Prim' m's s Algo gorit rithm hm ฮฉ(๐) For a specific tie- breaking rule 54
Other Problems We Study โ Maximu mum m Cardi dina nality lity Match ching ng โ Output an independent set of edges with maximum cardinality 55
Other Problems We Study โ Maximu mum m Cardi dina nality lity Match ching ng โ Output an independent set of edges with maximum cardinality โ Globa bal l Minim imum um Cut โ Output a subset ๐ of vertices with minimum number of edges between ๐ and ๐ โ ๐ 56
Other Problems We Study โ Maximu mum m Cardi dina nality lity Match ching ng โ Output an independent set of edges with maximum cardinality โ Globa bal l Minim imum um Cut โ Output a subset ๐ of vertices with minimum number of edges between ๐ and ๐ โ ๐ โ ๐ก - ๐ข Minimu mum m Cut โ 2-Colo oloring ring 57
Maximum Cardinality Matching For graphs on ๐ vertices with max. matching size OP OPT 58
Maximum Cardinality Matching For graphs on ๐ vertices with max. matching size OP OPT App pproxima ximatio tion n Ra Ratio Avera erage ge Sensit nsitivi ivity 1 ฮฉ(๐) 59
Maximum Cardinality Matching For graphs on ๐ vertices with max. matching size OP OPT App pproxima ximatio tion n Ra Ratio Avera erage ge Sensit nsitivi ivity 1 ฮฉ(๐) 1/2 1 60
Maximum Cardinality Matching For graphs on ๐ vertices with max. matching size OP OPT App pproxima ximatio tion n Ra Ratio Avera erage ge Sensit nsitivi ivity 1 ฮฉ(๐) 1/2 1 Corollar llary: 2-approximation algorithm for minimum vertex cover with average sensitivity 2 . 61
Maximum Cardinality Matching For graphs on ๐ vertices with max. matching size OP OPT App pproxima ximatio tion n Ra Ratio Avera erage ge Sensit nsitivi ivity 1 ฮฉ(๐) 1/2 1 1 1+๐ 2 ๐๐๐ 1 โ ๐ ๐ ๐ 3 Corollar llary: 2-approximation algorithm for minimum vertex cover with average sensitivity 2 . 62
Global Minimum Cut For graphs on ๐ vertices with global min. cut of size OP OPT 63
Global Minimum Cut For graphs on ๐ vertices with global min. cut of size OP OPT App pproxima ximatio tion n Ra Ratio Avera erage ge Sensiti nsitivit vity 1 ฮฉ(๐) 64
Global Minimum Cut For graphs on ๐ vertices with global min. cut of size OP OPT App pproxima ximatio tion n Ra Ratio Avera erage ge Sensiti nsitivit vity 1 ฮฉ(๐) 1 2 + ๐ ๐( ๐ OPT ) ๐ 65
Global Minimum Cut For graphs on ๐ vertices with global min. cut of size OP OPT App pproxima ximatio tion n Ra Ratio Avera erage ge Sensiti nsitivit vity 1 ฮฉ(๐) 1 2 + ๐ ๐( ๐ OPT ) ๐ If OP OPT = ๐(log ๐) , average sensitivity is ๐(1) 66
Global Minimum Cut For graphs on ๐ vertices with global min. cut of size OP OPT App pproxima ximatio tion n Ra Ratio Avera erage ge Sensiti nsitivit vity 1 ฮฉ(๐) 1 2 + ๐ ๐( ๐ OPT ) ๐ ฮฉ(๐ 1/ OPT / OPT 2 ) < โ If OP OPT = ๐(log ๐) , average sensitivity is ๐(1) 67
Global Minimum Cut For graphs on ๐ vertices with global min. cut of size OP OPT App pproxima ximatio tion n Ra Ratio Avera erage ge Sensiti nsitivit vity 1 ฮฉ(๐) 1 2 + ๐ ๐( ๐ OPT ) ๐ ฮฉ(๐ 1/ OPT / OPT 2 ) < โ If OP OPT = ๐(log ๐) , average sensitivity is ๐(1) If OP OPT = = O log ๐ , average sensitivity is (nearly) optimal 68
s-t Minimum Cut Probl blem em: Given graph ๐ป and vertices ๐ก, ๐ข , find output a subset ๐ of vertices with minimum number of edges between ๐ and ๐ โ ๐ such For graphs on ๐ vertices with s-t min. cut of size OP OPT that ๐ก โ ๐ and ๐ข โ ๐ โ ๐ App pproxima ximatio tion Avera erage ge Sensiti nsitivit vity (mu mult ltip iplic licative, ive, add dditive) e) (1, ๐(๐ 2/3 )) ๐(๐ 2/3 ) 69
2-Coloring Probl blem em: Given a bipartite graph ๐ป , , output the set of vertices in one of the bipartitions. App pproxima ximatio tion Avera erage ge Sensiti nsitivit vity (mu mult ltip iplic licative, ive, add dditive) e) โ ฮฉ(๐) Every LCA for 2 -coloring has query complexity ฮฉ(๐) Answers an open question raised by [Czumaj, Mansour, Vardi 18] on existence of sublinear-query LCAs for the problem of 2-coloring. 70
T alk Outline โ Our definition of average sensitivity for graph algorithms โ Key properties of our definition โ Main results โ Algorithm with low sensitivity for the global minimum cut problem โ Conclusions and Open directions 71
Global Minimum Cut Problem Given ๐ป = (๐, ๐น) and ๐ โ ๐ , size( ๐, ๐ป ): number of edges crossing (๐, ๐ โ ๐) 72
Global Minimum Cut Problem Given ๐ป = (๐, ๐น) and ๐ โ ๐ , size( ๐, ๐ป ): number of edges crossing (๐, ๐ โ ๐) Probl blem: em: Output set ๐ โ ๐ with the minimum size. 73
Global Minimum Cut Problem Given ๐ป = (๐, ๐น) and ๐ โ ๐ , size( ๐, ๐ป ): number of edges crossing (๐, ๐ โ ๐) Probl blem: em: Output set ๐ โ ๐ with the minimum size. Polynomial time exact algorithms exist. 74
Global Minimum Cut Problem Given ๐ป = (๐, ๐น) and ๐ โ ๐ , size( ๐, ๐ป ): number of edges crossing (๐, ๐ โ ๐) Probl blem: em: Output set ๐ โ ๐ with the minimum size. Polynomial time exact algorithms exist. Theorem em [Karge ger r 93]: : For ๐ฝ โฅ 1 , the number of cuts of size at most ๐ฝ โ OPT is at most ๐ 2๐ฝ and they can be enumerated in polynomial time (per cut). 75
Global Minimum Cut Theorem em: : There exists a polynomial time (2 + ๐) -approximation algorithm with average sensitivity 1 ๐ ๐ OPT for the global minimum cut problem for all ๐ > 0 . ๐ 76
Stable Algorithm for Global Minimum Cut On input ๐ป = (๐, ๐น) and parameter ๐ > 0 : โข Compute the value OPT; log ๐ โข Let ๐ฝ โ ๐( ๐ OPT ) ; โข Enumerate all cuts of size at most 2 + ๐ โ OPT ; โข Output a cut ๐ โ ๐ with probability proportional to exp(โ๐ฝ โ size ๐, ๐ป ) 77
Stable Algorithm for Global Minimum Cut On input ๐ป = (๐, ๐น) and parameter ๐ > 0 : โข Compute the value OPT; log ๐ โข Let ๐ฝ โ ๐( ๐ OPT ) ; โข Enumerate all cuts of size at most 2 + ๐ โ OPT ; โข Output a cut ๐ โ ๐ with probability proportional to exp(โ๐ฝ โ size ๐, ๐ป ) 78
Stable Algorithm for Global Minimum Cut On input ๐ป = (๐, ๐น) and parameter ๐ > 0 : โข Compute the value OPT; log ๐ โข Let ๐ฝ โ ๐( ๐ OPT ) ; โข Enumerate all cuts of size at most 2 + ๐ โ OPT ; โข Output a cut ๐ โ ๐ with probability proportional to exp(โ๐ฝ โ size ๐, ๐ป ) 79
Stable Algorithm for Global Minimum Cut On input ๐ป = (๐, ๐น) and parameter ๐ > 0 : โข Compute the value OPT; log ๐ โข Let ๐ฝ โ ๐( ๐ OPT ) ; โข Enumerate all cuts of size at most 2 + ๐ โ OPT ; โข Output a cut ๐ โ ๐ with probability proportional to exp(โ๐ฝ โ size ๐, ๐ป ) 80
Stable Algorithm for Global Minimum Cut Sa Samplin ling g from On input ๐ป = (๐, ๐น) and parameter ๐ > 0 : an approxim ximat ate e Gibb bbs s โข Compute the value OPT; distribu tributio tion log ๐ โข Let ๐ฝ โ ๐( ๐ OPT ) ; โข Enumerate all cuts of size at most 2 + ๐ โ OPT ; โข Output a cut ๐ โ ๐ with probability proportional to exp(โ๐ฝ โ size ๐, ๐ป ) 81
Stable Algorithm for Global Minimum Cut Samplin Sa ling g from On input ๐ป = (๐, ๐น) and parameter ๐ > 0 : an approxim ximat ate e Gibb bbs s โข Compute the value OPT; distribu tributio tion log ๐ โข Let ๐ฝ โ ๐( ๐ OPT ) ; โข Enumerate all cuts of size at most 2 + ๐ โ OPT ; โข Output a cut ๐ โ ๐ with probability proportional to exp(โ๐ฝ โ size ๐, ๐ป ) Inspired from a differentially private algorithm for global minimum cut [Gupta Ligett McSherry Roth T alwar '10] 82
Analysis App pproxima ximatio tion n Ra Ratio Clear from algorithm description Ru Runnin ning g time me Follows from Karger's theorem Avera erage ge Sensiti nsitivit vity Will analyze now 83
Analysis: A (Slightly) Different Algorithm On input ๐ป = (๐, ๐น) and parameter ๐ > 0 : Sa Samplin ling g from Gibb bbs s โข Compute the value OPT; distrib tributio tion log ๐ โข Let ๐ฝ โ ๐( ๐ OPT ) ; โข Output cut ๐ โ ๐ with prob. proportional to exp(โ๐ฝ โ size ๐, ๐ป ) 84
Analysis: A (Slightly) Different Algorithm On input ๐ป = (๐, ๐น) and parameter ๐ > 0 : Sa Samplin ling g from Gibb bbs s โข Compute the value OPT; distrib tributio tion log ๐ โข Let ๐ฝ โ ๐( ๐ OPT ) ; โข Output cut ๐ โ ๐ with prob. proportional to exp(โ๐ฝ โ size ๐, ๐ป ) Obse Ob servation ion: Enough to bound average sensitivity of above inefficient algorithm, since its output distribution is close to original algorithm 85
Analysis Denote the inefficient algorithm Overview using ๐ต โ Average sensitivity = Average On input ๐ป = (๐, ๐น) and (over ๐ โ ๐น) earth mover's parameter ๐ > 0 : distance between ๐ต(๐ป) and ๐ต(๐ป โ ๐) โข Compute the value OPT; log ๐ โข Let ๐ฝ โ ๐( ๐ OPT ) ; โข Output cut ๐ โ ๐ with prob. proportional to exp(โ๐ฝ โ size ๐, ๐ป ) 86
Analysis ๐ ๐, ๐ป :Probability that ๐ต outputs cut ๐ Overview on input ๐ป Fix ๐ โ ๐น . โ For cuts ๐ such that ๐ crosses ๐ , ๐ ๐, ๐ป โ ๐ โ ๐ ๐, ๐ป โ exp ๐ฝ On input ๐ป = (๐, ๐น) and parameter ๐ > 0 : โ Earth mover's distance between ๐ต(๐ป) and ๐ต ๐ป โ ๐ โข Compute the value OPT; log ๐ โข Let ๐ฝ โ ๐( ๐ OPT ) ; โ ๐ โ เท ๐ ๐, ๐ป โ ๐ โ ๐ ๐, ๐ป โข Output cut ๐ โ ๐ with ๐:๐ crosses ๐ prob. proportional to = ๐ โ exp ๐ฝ โ 1 โ ๐ ๐, ๐ป เท exp(โ๐ฝ โ size ๐, ๐ป ) ๐:๐ crosses ๐ 87
Analysis ๐ ๐, ๐ป :Probability that ๐ต outputs cut ๐ Overview on input ๐ป Fix ๐ โ ๐น . โ For cuts ๐ such that ๐ crosses ๐ , ๐ ๐, ๐ป โ ๐ โ ๐ ๐, ๐ป โ exp ๐ฝ On input ๐ป = (๐, ๐น) and parameter ๐ > 0 : โ Earth mover's distance between ๐ต(๐ป) and ๐ต ๐ป โ ๐ โข Compute the value OPT; log ๐ โข Let ๐ฝ โ ๐( ๐ OPT ) ; โ ๐ โ เท ๐ ๐, ๐ป โ ๐ โ ๐ ๐, ๐ป โข Output cut ๐ โ ๐ with ๐:๐ crosses ๐ prob. proportional to = ๐ โ exp ๐ฝ โ 1 โ ๐ ๐, ๐ป เท exp(โ๐ฝ โ size ๐, ๐ป ) ๐:๐ crosses ๐ 88
Analysis ๐ ๐, ๐ป :Probability that ๐ต outputs cut ๐ Overview on input ๐ป Fix ๐ โ ๐น . โ For cuts ๐ such that ๐ crosses ๐ , ๐ ๐, ๐ป โ ๐ โ ๐ ๐, ๐ป โ exp ๐ฝ On input ๐ป = (๐, ๐น) and parameter ๐ > 0 : โ Earth mover's distance between ๐ต(๐ป) and ๐ต ๐ป โ ๐ โข Compute the value OPT; log ๐ โข Let ๐ฝ โ ๐( ๐ OPT ) ; โ ๐ โ เท ๐ ๐, ๐ป โ ๐ โ ๐ ๐, ๐ป โข Output cut ๐ โ ๐ with ๐:๐ crosses ๐ prob. proportional to = ๐ โ exp ๐ฝ โ 1 โ ๐ ๐, ๐ป เท exp(โ๐ฝ โ size ๐, ๐ป ) ๐:๐ crosses ๐ 89
Analysis ๐ ๐, ๐ป :Probability that ๐ต outputs cut ๐ Overview on input ๐ป Fix ๐ โ ๐น . โ For cuts ๐ such that ๐ crosses ๐ , ๐ ๐, ๐ป โ ๐ โ ๐ ๐, ๐ป โ exp ๐ฝ On input ๐ป = (๐, ๐น) and parameter ๐ > 0 : โ Earth mover's distance between ๐ต(๐ป) and ๐ต ๐ป โ ๐ โข Compute the value OPT; log ๐ โข Let ๐ฝ โ ๐( ๐ OPT ) ; โ ๐ โ เท ๐ ๐, ๐ป โ ๐ โ ๐ ๐, ๐ป โข Output cut ๐ โ ๐ with ๐:๐ crosses ๐ prob. proportional to = ๐ โ exp ๐ฝ โ 1 โ ๐ ๐, ๐ป เท exp(โ๐ฝ โ size ๐, ๐ป ) ๐:๐ crosses ๐ 90
Analysis ๐ ๐, ๐ป :Probability that ๐ต outputs cut ๐ Overview on input ๐ป Fix ๐ โ ๐น . โ For cuts ๐ such that ๐ crosses ๐ , ๐ ๐, ๐ป โ ๐ โ ๐ ๐, ๐ป โ exp ๐ฝ On input ๐ป = (๐, ๐น) and parameter ๐ > 0 : โ Earth mover's distance between ๐ต(๐ป) and ๐ต ๐ป โ ๐ โข Compute the value OPT; log ๐ โข Let ๐ฝ โ ๐( ๐ OPT ) ; โ ๐ โ เท ๐ ๐, ๐ป โ ๐ โ ๐ ๐, ๐ป โข Output cut ๐ โ ๐ with ๐:๐ crosses ๐ prob. proportional to = ๐ โ exp ๐ฝ โ 1 โ ๐ ๐, ๐ป เท exp(โ๐ฝ โ size ๐, ๐ป ) ๐:๐ crosses ๐ 91
Analysis โ Average sensitivity of ๐ต is Overview โ ๐ ๐ โ exp ๐ฝ โ 1 โ เท เท ๐ ๐, ๐ป ๐ ๐:๐ crosses ๐ โ Average sensitivity of ๐ต is On input ๐ป = (๐, ๐น) and ๐ parameter ๐ > 0 : ๐ โ exp ๐ฝ โ 1 โ (Expected โค size of cut output by ๐ต) โข Compute the value OPT; โ Expected size of cut log ๐ โข Let ๐ฝ โ ๐( ๐ OPT ) ; โค 2 + ๐ โ OPT + ๐(1) โข Output cut ๐ โ ๐ with 2๐ โ OPT โค ๐ , as min. cut size at most prob. proportional to average degree exp(โ๐ฝ โ size ๐, ๐ป ) (More Detailed Analysis Overview) 92
Analysis โ Average sensitivity of ๐ต is Overview โ ๐ ๐ โ exp ๐ฝ โ 1 โ เท เท ๐ ๐, ๐ป ๐ ๐:๐ crosses ๐ โ Average sensitivity of ๐ต is On input ๐ป = (๐, ๐น) and ๐ parameter ๐ > 0 : ๐ โ exp ๐ฝ โ 1 โ (Expected โค size of cut output by ๐ต) โข Compute the value OPT; โ Expected size of cut log ๐ โข Let ๐ฝ โ ๐( ๐ OPT ) ; โค 2 + ๐ โ OPT + ๐(1) โข Output cut ๐ โ ๐ with 2๐ โ OPT โค ๐ , as min. cut size at most prob. proportional to average degree exp(โ๐ฝ โ size ๐, ๐ป ) (More Detailed Analysis Overview) 93
Analysis โ Average sensitivity of ๐ต is Overview โ ๐ ๐ โ exp ๐ฝ โ 1 โ เท เท ๐ ๐, ๐ป ๐ ๐:๐ crosses ๐ โ Average sensitivity of ๐ต is On input ๐ป = (๐, ๐น) and ๐ parameter ๐ > 0 : ๐ โ exp ๐ฝ โ 1 โ (Expected โค size of cut output by ๐ต) โข Compute the value OPT; โ Expected size of cut log ๐ โข Let ๐ฝ โ ๐( ๐ OPT ) ; โค 2 + ๐ โ OPT + ๐(1) โข Output cut ๐ โ ๐ with 2๐ โ OPT โค ๐ , as min. cut size at most prob. proportional to average degree exp(โ๐ฝ โ size ๐, ๐ป ) (More Detailed Analysis Overview) 94
Analysis โ Average sensitivity of ๐ต is Overview โ ๐ ๐ โ exp ๐ฝ โ 1 โ เท เท ๐ ๐, ๐ป ๐ ๐:๐ crosses ๐ โ Average sensitivity of ๐ต is On input ๐ป = (๐, ๐น) and ๐ parameter ๐ > 0 : ๐ โ exp ๐ฝ โ 1 โ (Expected โค size of cut output by ๐ต) โข Compute the value OPT; โ Expected size of cut log ๐ โข Let ๐ฝ โ ๐( ๐ OPT ) ; โค 2 + ๐ โ OPT + ๐(1) โข Output cut ๐ โ ๐ with 2๐ โ OPT โค ๐ , as min. cut size at most prob. proportional to average degree exp(โ๐ฝ โ size ๐, ๐ป ) (More Detailed Analysis Overview) 95
Global Minimum Cut Theorem em: : There exists a polynomial time (2 + ๐) -approximation algorithm with average sensitivity 1 ๐ ๐ OPT for the global minimum cut problem for all ๐ > 0 . ๐ 96
Global Minimum Cut Theorem em: : There exists a polynomial time (2 + ๐) -approximation algorithm with average sensitivity 1 ๐ ๐ OPT for the global minimum cut problem for all ๐ > 0 . ๐ Samp Sa mplin ling g from m Gibb bbs s distri ributio ution gives es stabil ility ity 97
T alk Outline โ Our definition of average sensitivity for graph algorithms โ Key properties of our definition โ Main results โ Algorithm with low sensitivity for the global minimum cut problem โ Conclusions and open directions 98
Summary of our contributions โ Introduced a definition of sensitivity of graph algorithms with several useful properties 99
Summary of our contributions โ Introduced a definition of sensitivity of graph algorithms with several useful properties โ Design of stable algorithms for various combinatorial problems 100
Recommend
More recommend