avera rage ge se sensitivit sitivity of of gra graph ph
play

AVERA RAGE GE SE SENSITIVIT SITIVITY OF OF GRA GRAPH PH ALGO - PowerPoint PPT Presentation

AVERA RAGE GE SE SENSITIVIT SITIVITY OF OF GRA GRAPH PH ALGO GORITHM RITHMS Nithin in Varma rma Joint nt work k with Yuich ichi Yos oshid ida 1 Sensitivity of an Algorithm Measure of change in output as a function of change


  1. 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 ๐ต on a graph ๐ป is differentially private if for all ๐‘“ โˆˆ ๐น โ–  the distributions ๐ต(๐ป) and ๐ต(๐ป โˆ’ ๐‘“) are close to each other โ€“ Much stricter notion than average sensitivity โ€“ Some of our algorithms inspired by differentially private algorithms 33

  2. 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 ๐ต on a graph ๐ป is differentially private if for all ๐‘“ โˆˆ ๐น โ–  the distributions ๐ต(๐ป) and ๐ต(๐ป โˆ’ 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

  3. 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 ๐ต on a graph ๐ป is differentially private if for all ๐‘“ โˆˆ ๐น โ–  the distributions ๐ต(๐ป) and ๐ต(๐ป โˆ’ 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

  4. 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 ๐ต on a graph ๐ป is differentially private if for all ๐‘“ โˆˆ ๐น โ–  the distributions ๐ต(๐ป) and ๐ต(๐ป โˆ’ 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

  5. 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

  6. k-Average Sensitivity from Average Sensitivity Theorem: em: If ๐ต has average sensitivity ๐‘”(๐‘œ, ๐‘›) , it has ๐‘™ -average sensitivity at most ฯƒ ๐‘—โˆˆ[๐‘™] ๐‘”(๐‘œ, ๐‘› โˆ’ ๐‘— + 1) . 38

  7. Average Sensitivity Composes Algorithms ๐ต, ๐ถ, ๐ท such that ๐ต(๐ป) = ๐ถ(๐ป, ๐ท ๐ป ) 39

  8. Average Sensitivity Composes Algorithms ๐ต, ๐ถ, ๐ท such that ๐ต(๐ป) = ๐ถ(๐ป, ๐ท ๐ป ) Theorem em (Info formal) mal): : Average sensitivity of ๐ต on ๐ป = (๐‘Š, ๐น) can be bounded by the sum of: a term for average sensitivity of ๐ถ , and โ€ข a term for average sensitivity of ๐ท . โ€ข 40

  9. Average Sensitivity Composes Algorithms ๐ต, ๐ถ, ๐ท such that ๐ต(๐ป) = ๐ถ(๐ป, ๐ท ๐ป ) Theorem em (Info formal) mal): : Average sensitivity of ๐ต on ๐ป = (๐‘Š, ๐น) can be bounded by the sum of: a term for average sensitivity of ๐ถ , and โ€ข a term for average sensitivity of ๐ท . โ€ข Can be used to bound the average sensitivity of a distribution over multiple stable-on-average algorithms. 41

  10. Connection to Deterministic ๐ป ๐ต(๐ป) Sublinear Algorithms Algorithm ๐ต 42

  11. Connection to Deterministic ๐ป ๐ต(๐ป) Sublinear Algorithms Algorithm ๐ต Local simulator ๐‘€ Graph ๐ป 43

  12. Connection to Deterministic ๐ป ๐ต(๐ป) Sublinear Algorithms Algorithm ๐ต 1 if ๐‘“ โˆˆ ๐ต(๐ป) Local ๐‘“ โˆˆ ๐น simulator ๐‘€ 0, otherwise Graph ๐ป 44

  13. Connection to Deterministic ๐ป ๐ต(๐ป) Sublinear Algorithms Algorithm ๐ต ๐‘Ÿ ๐ป โ‰œ ๐”ฝ ๐‘“โˆˆ๐น [# queries by ๐‘€] 1 if ๐‘“ โˆˆ ๐ต(๐ป) Local ๐‘“ โˆˆ ๐น simulator ๐‘€ 0, otherwise Graph ๐ป 45

  14. Connection to Deterministic ๐ป ๐ต(๐ป) Sublinear Algorithms Algorithm ๐ต ๐‘Ÿ ๐ป โ‰œ ๐”ฝ ๐‘“โˆˆ๐น [# queries by ๐‘€] 1 if ๐‘“ โˆˆ ๐ต(๐ป) Local ๐‘“ โˆˆ ๐น simulator ๐‘€ Avera erage ge se sensiti sitivit vity y of f ๐ต 0, otherwise on on ๐ป is s โ‰ค ๐‘Ÿ(๐ป) Graph ๐ป 46

  15. ๐œŒ is the random string Connection to ๐ป Algorithm ๐ต ๐ต ๐œŒ (๐ป) Sublinear Algorithms ๐œŒ ๐‘Ÿ ๐ป โ‰œ ๐”ฝ ๐œŒ,๐‘“โˆˆ๐น [# queries by ๐‘€] ๐‘“ โˆˆ ๐น 1 if ๐‘“ โˆˆ ๐ต ๐œŒ (๐ป) Local simulator ๐‘€ Avera erage ge se sensiti sitivit vity y of f ๐ต 0, otherwise ๐œŒ on on ๐ป is s โ‰ค ๐‘Ÿ(๐ป) Graph ๐ป 47

  16. Graph problem ๐‘„ Connection to Local Computation 1 if ๐‘ค is part of LCA ๐‘€ a solution to ๐‘„ Algorithms (LCAs) ๐‘ค โˆˆ ๐‘Š on ๐ป ๐œŒ โˆˆ 0,1 ๐‘  0, otherwise Graph ๐ป Answers of ๐‘€ are consistent with a single feasible solution of ๐‘„ on ๐ป 48

  17. Graph problem ๐‘„ Connection to Local Computation 1 if ๐‘ค is part of LCA ๐‘€ a solution to ๐‘„ Algorithms (LCAs) ๐‘ค โˆˆ ๐‘Š on ๐ป ๐œŒ โˆˆ 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‘Ÿ(๐ป) , then it Graph ๐ป has s an algo gorit ithm hm with avera rage ge sensi se sitivi ivity y โ‰ค ๐‘Ÿ(๐ป) Answers of ๐‘€ are consistent with a single feasible solution of ๐‘„ on ๐ป 49

  18. Graph problem ๐‘„ Connection to Local Computation 1 if ๐‘ค is part of LCA ๐‘€ a solution to ๐‘„ Algorithms (LCAs) ๐‘ค โˆˆ ๐‘Š on ๐ป ๐œŒ โˆˆ 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‘Ÿ(๐ป) , then it Graph ๐ป has s an algo gorit ithm hm with avera rage ge se sensi sitivi ivity y โ‰ค ๐‘Ÿ(๐ป) Answers of ๐‘€ are consistent with a Lower r bo bound d on average ge single feasible solution of ๐‘„ on ๐ป 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

  19. 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

  20. Minimum Spanning Forest For graphs on ๐‘œ 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

  21. Minimum Spanning Forest For graphs on ๐‘œ vertices and ๐‘› edges Algo gorit rithm hm Avera erage ge Sensit nsitivi ivity Krusk skal' al's s Algo gorit rithm hm ๐‘ƒ(๐‘œ/๐‘›) Prim' m's s Algo gorit rithm hm 53

  22. Minimum Spanning Forest For graphs on ๐‘œ vertices and ๐‘› edges Algo gorit rithm hm Avera erage ge Sensit nsitivi ivity Krusk skal' al's s Algo gorit rithm hm ๐‘ƒ(๐‘œ/๐‘›) Prim' m's s Algo gorit rithm hm ฮฉ(๐‘œ) For a specific tie- breaking rule 54

  23. 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

  24. 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‘Š โˆ– ๐‘‡ 56

  25. 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‘Š โˆ– ๐‘‡ โ–  ๐‘ก - ๐‘ข Minimu mum m Cut โ–  2-Colo oloring ring 57

  26. Maximum Cardinality Matching For graphs on ๐‘œ vertices with max. matching size OP OPT 58

  27. Maximum Cardinality Matching For graphs on ๐‘œ vertices with max. matching size OP OPT App pproxima ximatio tion n Ra Ratio Avera erage ge Sensit nsitivi ivity 1 ฮฉ(๐‘œ) 59

  28. Maximum Cardinality Matching For graphs on ๐‘œ vertices with max. matching size OP OPT App pproxima ximatio tion n Ra Ratio Avera erage ge Sensit nsitivi ivity 1 ฮฉ(๐‘œ) 1/2 1 60

  29. Maximum Cardinality Matching For graphs on ๐‘œ vertices with max. matching size OP OPT App pproxima ximatio tion n Ra Ratio Avera erage ge Sensit nsitivi ivity 1 ฮฉ(๐‘œ) 1/2 1 Corollar llary: 2-approximation algorithm for minimum vertex cover with average sensitivity 2 . 61

  30. Maximum Cardinality Matching For graphs on ๐‘œ vertices with max. matching size OP OPT App pproxima ximatio tion n Ra Ratio Avera erage ge Sensit nsitivi ivity 1 ฮฉ(๐‘œ) 1/2 1 1 1+๐œ— 2 ๐‘ƒ๐‘„๐‘ˆ 1 โˆ’ ๐œ— ๐‘ƒ ๐œ— 3 Corollar llary: 2-approximation algorithm for minimum vertex cover with average sensitivity 2 . 62

  31. Global Minimum Cut For graphs on ๐‘œ vertices with global min. cut of size OP OPT 63

  32. Global Minimum Cut For graphs on ๐‘œ 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‘œ) 64

  33. Global Minimum Cut For graphs on ๐‘œ 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‘œ) 1 2 + ๐œ— ๐‘ƒ( ๐œ— OPT ) ๐‘œ 65

  34. Global Minimum Cut For graphs on ๐‘œ 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‘œ) 1 2 + ๐œ— ๐‘ƒ( ๐œ— OPT ) ๐‘œ If OP OPT = ๐œ•(log ๐‘œ) , average sensitivity is ๐‘ƒ(1) 66

  35. Global Minimum Cut For graphs on ๐‘œ 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‘œ) 1 2 + ๐œ— ๐‘ƒ( ๐œ— OPT ) ๐‘œ ฮฉ(๐‘œ 1/ OPT / OPT 2 ) < โˆž If OP OPT = ๐œ•(log ๐‘œ) , average sensitivity is ๐‘ƒ(1) 67

  36. Global Minimum Cut For graphs on ๐‘œ 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‘œ) 1 2 + ๐œ— ๐‘ƒ( ๐œ— OPT ) ๐‘œ ฮฉ(๐‘œ 1/ OPT / OPT 2 ) < โˆž If OP OPT = ๐œ•(log ๐‘œ) , average sensitivity is ๐‘ƒ(1) If OP OPT = = O log ๐‘œ , average sensitivity is (nearly) optimal 68

  37. s-t Minimum Cut Probl blem em: Given graph ๐ป and vertices ๐‘ก, ๐‘ข , find output a subset ๐‘‡ of vertices with minimum number of edges between ๐‘‡ and ๐‘Š โˆ– ๐‘‡ such For graphs on ๐‘œ vertices with s-t min. cut of size OP OPT that ๐‘ก โˆˆ ๐‘‡ and ๐‘ข โˆˆ ๐‘Š โˆ– ๐‘‡ App pproxima ximatio tion Avera erage ge Sensiti nsitivit vity (mu mult ltip iplic licative, ive, add dditive) e) (1, ๐‘ƒ(๐‘œ 2/3 )) ๐‘ƒ(๐‘œ 2/3 ) 69

  38. 2-Coloring Probl blem em: Given a bipartite graph ๐ป , , 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) โˆ’ ฮฉ(๐‘œ) Every LCA for 2 -coloring has query complexity ฮฉ(๐‘œ) Answers an open question raised by [Czumaj, Mansour, Vardi 18] on existence of sublinear-query LCAs for the problem of 2-coloring. 70

  39. 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

  40. Global Minimum Cut Problem Given ๐ป = (๐‘Š, ๐น) and ๐‘‡ โŠ† ๐‘Š , size( ๐‘‡, ๐ป ): number of edges crossing (๐‘‡, ๐‘Š โˆ– ๐‘‡) 72

  41. Global Minimum Cut Problem Given ๐ป = (๐‘Š, ๐น) and ๐‘‡ โŠ† ๐‘Š , size( ๐‘‡, ๐ป ): number of edges crossing (๐‘‡, ๐‘Š โˆ– ๐‘‡) Probl blem: em: Output set ๐‘‡ โŠ† ๐‘Š with the minimum size. 73

  42. Global Minimum Cut Problem Given ๐ป = (๐‘Š, ๐น) and ๐‘‡ โŠ† ๐‘Š , size( ๐‘‡, ๐ป ): number of edges crossing (๐‘‡, ๐‘Š โˆ– ๐‘‡) Probl blem: em: Output set ๐‘‡ โŠ† ๐‘Š with the minimum size. Polynomial time exact algorithms exist. 74

  43. Global Minimum Cut Problem Given ๐ป = (๐‘Š, ๐น) and ๐‘‡ โŠ† ๐‘Š , size( ๐‘‡, ๐ป ): number of edges crossing (๐‘‡, ๐‘Š โˆ– ๐‘‡) Probl blem: em: Output set ๐‘‡ โŠ† ๐‘Š with the minimum size. Polynomial time exact algorithms exist. Theorem em [Karge ger r 93]: : For ๐›ฝ โ‰ฅ 1 , the number of cuts of size at most ๐›ฝ โ‹… OPT is at most ๐‘œ 2๐›ฝ and they can be enumerated in polynomial time (per cut). 75

  44. Global Minimum Cut Theorem em: : There exists a polynomial time (2 + ๐œ—) -approximation algorithm with average sensitivity 1 ๐‘ƒ ๐œ— OPT for the global minimum cut problem for all ๐œ— > 0 . ๐‘œ 76

  45. Stable Algorithm for Global Minimum Cut On input ๐ป = (๐‘Š, ๐น) and parameter ๐œ— > 0 : โ€ข Compute the value OPT; log ๐‘œ โ€ข Let ๐›ฝ โ† ๐œ„( ๐œ— OPT ) ; โ€ข Enumerate all cuts of size at most 2 + ๐œ— โ‹… OPT ; โ€ข Output a cut ๐‘‡ โŠ† ๐‘Š with probability proportional to exp(โˆ’๐›ฝ โ‹… size ๐‘‡, ๐ป ) 77

  46. Stable Algorithm for Global Minimum Cut On input ๐ป = (๐‘Š, ๐น) and parameter ๐œ— > 0 : โ€ข Compute the value OPT; log ๐‘œ โ€ข Let ๐›ฝ โ† ๐œ„( ๐œ— OPT ) ; โ€ข Enumerate all cuts of size at most 2 + ๐œ— โ‹… OPT ; โ€ข Output a cut ๐‘‡ โŠ† ๐‘Š with probability proportional to exp(โˆ’๐›ฝ โ‹… size ๐‘‡, ๐ป ) 78

  47. Stable Algorithm for Global Minimum Cut On input ๐ป = (๐‘Š, ๐น) and parameter ๐œ— > 0 : โ€ข Compute the value OPT; log ๐‘œ โ€ข Let ๐›ฝ โ† ๐œ„( ๐œ— OPT ) ; โ€ข Enumerate all cuts of size at most 2 + ๐œ— โ‹… OPT ; โ€ข Output a cut ๐‘‡ โŠ† ๐‘Š with probability proportional to exp(โˆ’๐›ฝ โ‹… size ๐‘‡, ๐ป ) 79

  48. Stable Algorithm for Global Minimum Cut On input ๐ป = (๐‘Š, ๐น) and parameter ๐œ— > 0 : โ€ข Compute the value OPT; log ๐‘œ โ€ข Let ๐›ฝ โ† ๐œ„( ๐œ— OPT ) ; โ€ข Enumerate all cuts of size at most 2 + ๐œ— โ‹… OPT ; โ€ข Output a cut ๐‘‡ โŠ† ๐‘Š with probability proportional to exp(โˆ’๐›ฝ โ‹… size ๐‘‡, ๐ป ) 80

  49. Stable Algorithm for Global Minimum Cut Sa Samplin ling g from On input ๐ป = (๐‘Š, ๐น) and parameter ๐œ— > 0 : an approxim ximat ate e Gibb bbs s โ€ข Compute the value OPT; distribu tributio tion log ๐‘œ โ€ข Let ๐›ฝ โ† ๐œ„( ๐œ— OPT ) ; โ€ข Enumerate all cuts of size at most 2 + ๐œ— โ‹… OPT ; โ€ข Output a cut ๐‘‡ โŠ† ๐‘Š with probability proportional to exp(โˆ’๐›ฝ โ‹… size ๐‘‡, ๐ป ) 81

  50. Stable Algorithm for Global Minimum Cut Samplin Sa ling g from On input ๐ป = (๐‘Š, ๐น) and parameter ๐œ— > 0 : an approxim ximat ate e Gibb bbs s โ€ข Compute the value OPT; distribu tributio tion log ๐‘œ โ€ข Let ๐›ฝ โ† ๐œ„( ๐œ— OPT ) ; โ€ข Enumerate all cuts of size at most 2 + ๐œ— โ‹… OPT ; โ€ข Output a cut ๐‘‡ โŠ† ๐‘Š with probability proportional to exp(โˆ’๐›ฝ โ‹… size ๐‘‡, ๐ป ) Inspired from a differentially private algorithm for global minimum cut [Gupta Ligett McSherry Roth T alwar '10] 82

  51. 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

  52. Analysis: A (Slightly) Different Algorithm On input ๐ป = (๐‘Š, ๐น) and parameter ๐œ— > 0 : Sa Samplin ling g from Gibb bbs s โ€ข Compute the value OPT; distrib tributio tion log ๐‘œ โ€ข Let ๐›ฝ โ† ๐œ„( ๐œ— OPT ) ; โ€ข Output cut ๐‘‡ โŠ† ๐‘Š with prob. proportional to exp(โˆ’๐›ฝ โ‹… size ๐‘‡, ๐ป ) 84

  53. Analysis: A (Slightly) Different Algorithm On input ๐ป = (๐‘Š, ๐น) and parameter ๐œ— > 0 : Sa Samplin ling g from Gibb bbs s โ€ข Compute the value OPT; distrib tributio tion log ๐‘œ โ€ข Let ๐›ฝ โ† ๐œ„( ๐œ— OPT ) ; โ€ข Output cut ๐‘‡ โŠ† ๐‘Š with prob. proportional to exp(โˆ’๐›ฝ โ‹… size ๐‘‡, ๐ป ) Obse Ob servation ion: Enough to bound average sensitivity of above inefficient algorithm, since its output distribution is close to original algorithm 85

  54. Analysis Denote the inefficient algorithm Overview using ๐ต โ–  Average sensitivity = Average On input ๐ป = (๐‘Š, ๐น) and (over ๐‘“ โˆˆ ๐น) earth mover's parameter ๐œ— > 0 : distance between ๐ต(๐ป) and ๐ต(๐ป โˆ’ ๐‘“) โ€ข Compute the value OPT; log ๐‘œ โ€ข Let ๐›ฝ โ† ๐œ„( ๐œ— OPT ) ; โ€ข Output cut ๐‘‡ โŠ† ๐‘Š with prob. proportional to exp(โˆ’๐›ฝ โ‹… size ๐‘‡, ๐ป ) 86

  55. Analysis ๐‘ž ๐‘‡, ๐ป :Probability that ๐ต outputs cut ๐‘‡ Overview on input ๐ป Fix ๐‘“ โˆˆ ๐น . โ–  For cuts ๐‘‡ such that ๐‘“ crosses ๐‘‡ , ๐‘ž ๐‘‡, ๐ป โˆ’ ๐‘“ โ‰ˆ ๐‘ž ๐‘‡, ๐ป โ‹… exp ๐›ฝ On input ๐ป = (๐‘Š, ๐น) and parameter ๐œ— > 0 : โ–  Earth mover's distance between ๐ต(๐ป) and ๐ต ๐ป โˆ’ ๐‘“ โ€ข Compute the value OPT; log ๐‘œ โ€ข Let ๐›ฝ โ† ๐œ„( ๐œ— OPT ) ; โ‰ˆ ๐‘œ โ‹… เท ๐‘ž ๐‘‡, ๐ป โˆ’ ๐‘“ โˆ’ ๐‘ž ๐‘‡, ๐ป โ€ข Output cut ๐‘‡ โŠ† ๐‘Š with ๐‘‡:๐‘“ crosses ๐‘‡ prob. proportional to = ๐‘œ โ‹… exp ๐›ฝ โˆ’ 1 โ‹… ๐‘ž ๐‘‡, ๐ป เท exp(โˆ’๐›ฝ โ‹… size ๐‘‡, ๐ป ) ๐‘‡:๐‘“ crosses ๐‘‡ 87

  56. Analysis ๐‘ž ๐‘‡, ๐ป :Probability that ๐ต outputs cut ๐‘‡ Overview on input ๐ป Fix ๐‘“ โˆˆ ๐น . โ–  For cuts ๐‘‡ such that ๐‘“ crosses ๐‘‡ , ๐‘ž ๐‘‡, ๐ป โˆ’ ๐‘“ โ‰ˆ ๐‘ž ๐‘‡, ๐ป โ‹… exp ๐›ฝ On input ๐ป = (๐‘Š, ๐น) and parameter ๐œ— > 0 : โ–  Earth mover's distance between ๐ต(๐ป) and ๐ต ๐ป โˆ’ ๐‘“ โ€ข Compute the value OPT; log ๐‘œ โ€ข Let ๐›ฝ โ† ๐œ„( ๐œ— OPT ) ; โ‰ˆ ๐‘œ โ‹… เท ๐‘ž ๐‘‡, ๐ป โˆ’ ๐‘“ โˆ’ ๐‘ž ๐‘‡, ๐ป โ€ข Output cut ๐‘‡ โŠ† ๐‘Š with ๐‘‡:๐‘“ crosses ๐‘‡ prob. proportional to = ๐‘œ โ‹… exp ๐›ฝ โˆ’ 1 โ‹… ๐‘ž ๐‘‡, ๐ป เท exp(โˆ’๐›ฝ โ‹… size ๐‘‡, ๐ป ) ๐‘‡:๐‘“ crosses ๐‘‡ 88

  57. Analysis ๐‘ž ๐‘‡, ๐ป :Probability that ๐ต outputs cut ๐‘‡ Overview on input ๐ป Fix ๐‘“ โˆˆ ๐น . โ–  For cuts ๐‘‡ such that ๐‘“ crosses ๐‘‡ , ๐‘ž ๐‘‡, ๐ป โˆ’ ๐‘“ โ‰ˆ ๐‘ž ๐‘‡, ๐ป โ‹… exp ๐›ฝ On input ๐ป = (๐‘Š, ๐น) and parameter ๐œ— > 0 : โ–  Earth mover's distance between ๐ต(๐ป) and ๐ต ๐ป โˆ’ ๐‘“ โ€ข Compute the value OPT; log ๐‘œ โ€ข Let ๐›ฝ โ† ๐œ„( ๐œ— OPT ) ; โ‰ˆ ๐‘œ โ‹… เท ๐‘ž ๐‘‡, ๐ป โˆ’ ๐‘“ โˆ’ ๐‘ž ๐‘‡, ๐ป โ€ข Output cut ๐‘‡ โŠ† ๐‘Š with ๐‘‡:๐‘“ crosses ๐‘‡ prob. proportional to = ๐‘œ โ‹… exp ๐›ฝ โˆ’ 1 โ‹… ๐‘ž ๐‘‡, ๐ป เท exp(โˆ’๐›ฝ โ‹… size ๐‘‡, ๐ป ) ๐‘‡:๐‘“ crosses ๐‘‡ 89

  58. Analysis ๐‘ž ๐‘‡, ๐ป :Probability that ๐ต outputs cut ๐‘‡ Overview on input ๐ป Fix ๐‘“ โˆˆ ๐น . โ–  For cuts ๐‘‡ such that ๐‘“ crosses ๐‘‡ , ๐‘ž ๐‘‡, ๐ป โˆ’ ๐‘“ โ‰ˆ ๐‘ž ๐‘‡, ๐ป โ‹… exp ๐›ฝ On input ๐ป = (๐‘Š, ๐น) and parameter ๐œ— > 0 : โ–  Earth mover's distance between ๐ต(๐ป) and ๐ต ๐ป โˆ’ ๐‘“ โ€ข Compute the value OPT; log ๐‘œ โ€ข Let ๐›ฝ โ† ๐œ„( ๐œ— OPT ) ; โ‰ˆ ๐‘œ โ‹… เท ๐‘ž ๐‘‡, ๐ป โˆ’ ๐‘“ โˆ’ ๐‘ž ๐‘‡, ๐ป โ€ข Output cut ๐‘‡ โŠ† ๐‘Š with ๐‘‡:๐‘“ crosses ๐‘‡ prob. proportional to = ๐‘œ โ‹… exp ๐›ฝ โˆ’ 1 โ‹… ๐‘ž ๐‘‡, ๐ป เท exp(โˆ’๐›ฝ โ‹… size ๐‘‡, ๐ป ) ๐‘‡:๐‘“ crosses ๐‘‡ 90

  59. Analysis ๐‘ž ๐‘‡, ๐ป :Probability that ๐ต outputs cut ๐‘‡ Overview on input ๐ป Fix ๐‘“ โˆˆ ๐น . โ–  For cuts ๐‘‡ such that ๐‘“ crosses ๐‘‡ , ๐‘ž ๐‘‡, ๐ป โˆ’ ๐‘“ โ‰ˆ ๐‘ž ๐‘‡, ๐ป โ‹… exp ๐›ฝ On input ๐ป = (๐‘Š, ๐น) and parameter ๐œ— > 0 : โ–  Earth mover's distance between ๐ต(๐ป) and ๐ต ๐ป โˆ’ ๐‘“ โ€ข Compute the value OPT; log ๐‘œ โ€ข Let ๐›ฝ โ† ๐œ„( ๐œ— OPT ) ; โ‰ˆ ๐‘œ โ‹… เท ๐‘ž ๐‘‡, ๐ป โˆ’ ๐‘“ โˆ’ ๐‘ž ๐‘‡, ๐ป โ€ข Output cut ๐‘‡ โŠ† ๐‘Š with ๐‘‡:๐‘“ crosses ๐‘‡ prob. proportional to = ๐‘œ โ‹… exp ๐›ฝ โˆ’ 1 โ‹… ๐‘ž ๐‘‡, ๐ป เท exp(โˆ’๐›ฝ โ‹… size ๐‘‡, ๐ป ) ๐‘‡:๐‘“ crosses ๐‘‡ 91

  60. Analysis โ–  Average sensitivity of ๐ต is Overview โ‰ˆ ๐‘œ ๐‘› โ‹… exp ๐›ฝ โˆ’ 1 โ‹… เท เท ๐‘ž ๐‘‡, ๐ป ๐‘“ ๐‘‡:๐‘“ crosses ๐‘‡ โ–  Average sensitivity of ๐ต is On input ๐ป = (๐‘Š, ๐น) and ๐‘œ parameter ๐œ— > 0 : ๐‘› โ‹… exp ๐›ฝ โˆ’ 1 โ‹… (Expected โ‰ค size of cut output by ๐ต) โ€ข Compute the value OPT; โ–  Expected size of cut log ๐‘œ โ€ข Let ๐›ฝ โ† ๐œ„( ๐œ— OPT ) ; โ‰ค 2 + ๐œ— โ‹… OPT + ๐‘(1) โ€ข Output cut ๐‘‡ โŠ† ๐‘Š with 2๐‘› โ–  OPT โ‰ค ๐‘œ , as min. cut size at most prob. proportional to average degree exp(โˆ’๐›ฝ โ‹… size ๐‘‡, ๐ป ) (More Detailed Analysis Overview) 92

  61. Analysis โ–  Average sensitivity of ๐ต is Overview โ‰ˆ ๐‘œ ๐‘› โ‹… exp ๐›ฝ โˆ’ 1 โ‹… เท เท ๐‘ž ๐‘‡, ๐ป ๐‘“ ๐‘‡:๐‘“ crosses ๐‘‡ โ–  Average sensitivity of ๐ต is On input ๐ป = (๐‘Š, ๐น) and ๐‘œ parameter ๐œ— > 0 : ๐‘› โ‹… exp ๐›ฝ โˆ’ 1 โ‹… (Expected โ‰ค size of cut output by ๐ต) โ€ข Compute the value OPT; โ–  Expected size of cut log ๐‘œ โ€ข Let ๐›ฝ โ† ๐œ„( ๐œ— OPT ) ; โ‰ค 2 + ๐œ— โ‹… OPT + ๐‘(1) โ€ข Output cut ๐‘‡ โŠ† ๐‘Š with 2๐‘› โ–  OPT โ‰ค ๐‘œ , as min. cut size at most prob. proportional to average degree exp(โˆ’๐›ฝ โ‹… size ๐‘‡, ๐ป ) (More Detailed Analysis Overview) 93

  62. Analysis โ–  Average sensitivity of ๐ต is Overview โ‰ˆ ๐‘œ ๐‘› โ‹… exp ๐›ฝ โˆ’ 1 โ‹… เท เท ๐‘ž ๐‘‡, ๐ป ๐‘“ ๐‘‡:๐‘“ crosses ๐‘‡ โ–  Average sensitivity of ๐ต is On input ๐ป = (๐‘Š, ๐น) and ๐‘œ parameter ๐œ— > 0 : ๐‘› โ‹… exp ๐›ฝ โˆ’ 1 โ‹… (Expected โ‰ค size of cut output by ๐ต) โ€ข Compute the value OPT; โ–  Expected size of cut log ๐‘œ โ€ข Let ๐›ฝ โ† ๐œ„( ๐œ— OPT ) ; โ‰ค 2 + ๐œ— โ‹… OPT + ๐‘(1) โ€ข Output cut ๐‘‡ โŠ† ๐‘Š with 2๐‘› โ–  OPT โ‰ค ๐‘œ , as min. cut size at most prob. proportional to average degree exp(โˆ’๐›ฝ โ‹… size ๐‘‡, ๐ป ) (More Detailed Analysis Overview) 94

  63. Analysis โ–  Average sensitivity of ๐ต is Overview โ‰ˆ ๐‘œ ๐‘› โ‹… exp ๐›ฝ โˆ’ 1 โ‹… เท เท ๐‘ž ๐‘‡, ๐ป ๐‘“ ๐‘‡:๐‘“ crosses ๐‘‡ โ–  Average sensitivity of ๐ต is On input ๐ป = (๐‘Š, ๐น) and ๐‘œ parameter ๐œ— > 0 : ๐‘› โ‹… exp ๐›ฝ โˆ’ 1 โ‹… (Expected โ‰ค size of cut output by ๐ต) โ€ข Compute the value OPT; โ–  Expected size of cut log ๐‘œ โ€ข Let ๐›ฝ โ† ๐œ„( ๐œ— OPT ) ; โ‰ค 2 + ๐œ— โ‹… OPT + ๐‘(1) โ€ข Output cut ๐‘‡ โŠ† ๐‘Š with 2๐‘› โ–  OPT โ‰ค ๐‘œ , as min. cut size at most prob. proportional to average degree exp(โˆ’๐›ฝ โ‹… size ๐‘‡, ๐ป ) (More Detailed Analysis Overview) 95

  64. Global Minimum Cut Theorem em: : There exists a polynomial time (2 + ๐œ—) -approximation algorithm with average sensitivity 1 ๐‘ƒ ๐œ— OPT for the global minimum cut problem for all ๐œ— > 0 . ๐‘œ 96

  65. Global Minimum Cut Theorem em: : There exists a polynomial time (2 + ๐œ—) -approximation algorithm with average sensitivity 1 ๐‘ƒ ๐œ— OPT for the global minimum cut problem for all ๐œ— > 0 . ๐‘œ Samp Sa mplin ling g from m Gibb bbs s distri ributio ution gives es stabil ility ity 97

  66. 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

  67. Summary of our contributions โ–  Introduced a definition of sensitivity of graph algorithms with several useful properties 99

  68. 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