planted cliques iterative thresholding and message
play

Planted Cliques, Iterative Thresholding and Message Passing - PowerPoint PPT Presentation

Planted Cliques, Iterative Thresholding and Message Passing Algorithms Yash Deshpande and Andrea Montanari Stanford University November 5, 2013 Deshpande, Montanari Planted Cliques November 5, 2013 1 / 49 Problem Definition Given


  1. Seems much harder than it looks! ◮ “Statistical algorithms” fail if k = n 1 ❂ 2 � ✍ : [Feldman et al., 2012] ♣ ◮ r -Lovász-Schrijver fails for k ✔ n ❂ 2 r : [Feige, Krauthgamer, 2002] ♣ n ◮ r -Lasserre fails for k ✔ ( log n ) r 2 : [Widgerson and Meka, 2013] Deshpande, Montanari Planted Cliques November 5, 2013 21 / 49

  2. Our result Theorem (Deshpande, Montanari, 2013) ♣ n ❂ e , there exists an O ( n 2 log n ) time algorithm that If ❥ S ❥ = k ✕ ( 1 + ✧ ) identifies S with high probability. I will present: 1 A (wrong) heuristic analysis 2 How to fix the heuristic 3 Lower bounds Deshpande, Montanari Planted Cliques November 5, 2013 22 / 49

  3. Our result Theorem (Deshpande, Montanari, 2013) ♣ n ❂ e , there exists an O ( n 2 log n ) time algorithm that If ❥ S ❥ = k ✕ ( 1 + ✧ ) identifies S with high probability. I will present: 1 A (wrong) heuristic analysis 2 How to fix the heuristic 3 Lower bounds Deshpande, Montanari Planted Cliques November 5, 2013 22 / 49

  4. Our result Theorem (Deshpande, Montanari, 2013) ♣ n ❂ e , there exists an O ( n 2 log n ) time algorithm that If ❥ S ❥ = k ✕ ( 1 + ✧ ) identifies S with high probability. I will present: 1 A (wrong) heuristic analysis 2 How to fix the heuristic 3 Lower bounds Deshpande, Montanari Planted Cliques November 5, 2013 22 / 49

  5. Iterative Thresholding The power iteration: v t + 1 = A v t ✿ Improvement: v t + 1 = AF t ( v t ) ✿ where F t ( v ) = ( f t ( v 1 ) ❀ f t ( v 2 ) ❀ ✁ ✁ ✁ ❀ f t ( v n )) T Choose f t ( ✁ ) to exploit sparsity of e S Deshpande, Montanari Planted Cliques November 5, 2013 23 / 49

  6. Iterative Thresholding The power iteration: v t + 1 = A v t ✿ Improvement: v t + 1 = AF t ( v t ) ✿ where F t ( v ) = ( f t ( v 1 ) ❀ f t ( v 2 ) ❀ ✁ ✁ ✁ ❀ f t ( v n )) T Choose f t ( ✁ ) to exploit sparsity of e S Deshpande, Montanari Planted Cliques November 5, 2013 23 / 49

  7. Iterative Thresholding The power iteration: v t + 1 = A v t ✿ Improvement: v t + 1 = AF t ( v t ) ✿ where F t ( v ) = ( f t ( v 1 ) ❀ f t ( v 2 ) ❀ ✁ ✁ ✁ ❀ f t ( v n )) T Choose f t ( ✁ ) to exploit sparsity of e S Deshpande, Montanari Planted Cliques November 5, 2013 23 / 49

  8. Iterative Thresholding The power iteration: v t + 1 = A v t ✿ Improvement: v t + 1 = AF t ( v t ) ✿ where F t ( v ) = ( f t ( v 1 ) ❀ f t ( v 2 ) ❀ ✁ ✁ ✁ ❀ f t ( v n )) T Choose f t ( ✁ ) to exploit sparsity of e S Deshpande, Montanari Planted Cliques November 5, 2013 23 / 49

  9. Iterative Thresholding The power iteration: v t + 1 = A v t ✿ Improvement: v t + 1 = AF t ( v t ) ✿ where F t ( v ) = ( f t ( v 1 ) ❀ f t ( v 2 ) ❀ ✁ ✁ ✁ ❀ f t ( v n )) T Choose f t ( ✁ ) to exploit sparsity of e S Deshpande, Montanari Planted Cliques November 5, 2013 23 / 49

  10. (Wrong) Analysis ❳ 1 v t + 1 A ij f t ( v t ♣ n = j ) ✿ i j A ij are random ✝ 1 r.v. Use Central Limit Theorem for v t + 1 i Deshpande, Montanari Planted Cliques November 5, 2013 24 / 49

  11. (Wrong) Analysis ❳ 1 v t + 1 A ij f t ( v t ♣ n = j ) ✿ i j A ij are random ✝ 1 r.v. Use Central Limit Theorem for v t + 1 i Deshpande, Montanari Planted Cliques November 5, 2013 24 / 49

  12. (Wrong) Analysis If i ❂ ✷ S : ❳ 1 v t + 1 A ij f t ( v t ♣ n = j ) i j ✒ ✓ ❳ 0 ❀ 1 f t ( v t j ) 2 ✙ N n j Letting v t i ✙ N ( 0 ❀ ✛ 2 t ) . . . ❳ t + 1 = 1 ✛ 2 f t ( v t j ) 2 n j = E ❢ f t ( ✛ t ✘ ) 2 ❣ ❀ ✘ ✘ N ( 0 ❀ 1 ) Deshpande, Montanari Planted Cliques November 5, 2013 25 / 49

  13. (Wrong) Analysis If i ❂ ✷ S : ❳ 1 v t + 1 A ij f t ( v t ♣ n = j ) i j ✒ ✓ ❳ 0 ❀ 1 f t ( v t j ) 2 ✙ N n j Letting v t i ✙ N ( 0 ❀ ✛ 2 t ) . . . ❳ t + 1 = 1 ✛ 2 f t ( v t j ) 2 n j = E ❢ f t ( ✛ t ✘ ) 2 ❣ ❀ ✘ ✘ N ( 0 ❀ 1 ) Deshpande, Montanari Planted Cliques November 5, 2013 25 / 49

  14. (Wrong) Analysis If i ❂ ✷ S : ❳ 1 v t + 1 A ij f t ( v t ♣ n = j ) i j ✒ ✓ ❳ 0 ❀ 1 f t ( v t j ) 2 ✙ N n j Letting v t i ✙ N ( 0 ❀ ✛ 2 t ) . . . ❳ t + 1 = 1 ✛ 2 f t ( v t j ) 2 n j = E ❢ f t ( ✛ t ✘ ) 2 ❣ ❀ ✘ ✘ N ( 0 ❀ 1 ) Deshpande, Montanari Planted Cliques November 5, 2013 25 / 49

  15. (Wrong) Analysis If i ❂ ✷ S : ❳ 1 v t + 1 A ij f t ( v t ♣ n = j ) i j ✒ ✓ ❳ 0 ❀ 1 f t ( v t j ) 2 ✙ N n j Letting v t i ✙ N ( 0 ❀ ✛ 2 t ) . . . ❳ t + 1 = 1 ✛ 2 f t ( v t j ) 2 n j = E ❢ f t ( ✛ t ✘ ) 2 ❣ ❀ ✘ ✘ N ( 0 ❀ 1 ) Deshpande, Montanari Planted Cliques November 5, 2013 25 / 49

  16. (Wrong) Analysis If i ✷ S : ❳ ❳ 1 + 1 v t + 1 f t ( v t A ij f t ( v t = ♣ n j ) ♣ n j ) i j ✷ S j ❂ ✷ S ⑤ ④③ ⑥ ⑤ ④③ ⑥ ✖ t + 1 ✙ N ( 0 ❀✛ 2 t + 1 ) where ❳ 1 f t ( v t ♣ n ✖ t + 1 = j ) j ✷ S ✒ k ✓ ✒ 1 ✓ ❳ f t ( v t ♣ n = j ) k j ✷ S = ✔ E ❢ f t ( ✖ t + ✛ t ✘ ) ❣ ❀ ✘ ✘ N ( 0 ❀ 1 ) Deshpande, Montanari Planted Cliques November 5, 2013 26 / 49

  17. (Wrong) Analysis If i ✷ S : ❳ ❳ 1 + 1 v t + 1 f t ( v t A ij f t ( v t = ♣ n j ) ♣ n j ) i j ✷ S j ❂ ✷ S ⑤ ④③ ⑥ ⑤ ④③ ⑥ ✖ t + 1 ✙ N ( 0 ❀✛ 2 t + 1 ) where ❳ 1 f t ( v t ♣ n ✖ t + 1 = j ) j ✷ S ✒ k ✓ ✒ 1 ✓ ❳ f t ( v t ♣ n = j ) k j ✷ S = ✔ E ❢ f t ( ✖ t + ✛ t ✘ ) ❣ ❀ ✘ ✘ N ( 0 ❀ 1 ) Deshpande, Montanari Planted Cliques November 5, 2013 26 / 49

  18. (Wrong) Analysis If i ✷ S : ❳ ❳ 1 + 1 v t + 1 f t ( v t A ij f t ( v t = ♣ n j ) ♣ n j ) i j ✷ S j ❂ ✷ S ⑤ ④③ ⑥ ⑤ ④③ ⑥ ✖ t + 1 ✙ N ( 0 ❀✛ 2 t + 1 ) where ❳ 1 f t ( v t ♣ n ✖ t + 1 = j ) j ✷ S ✒ k ✓ ✒ 1 ✓ ❳ f t ( v t ♣ n = j ) k j ✷ S = ✔ E ❢ f t ( ✖ t + ✛ t ✘ ) ❣ ❀ ✘ ✘ N ( 0 ❀ 1 ) Deshpande, Montanari Planted Cliques November 5, 2013 26 / 49

  19. (Wrong) Analysis If i ✷ S : ❳ ❳ 1 + 1 v t + 1 f t ( v t A ij f t ( v t = ♣ n j ) ♣ n j ) i j ✷ S j ❂ ✷ S ⑤ ④③ ⑥ ⑤ ④③ ⑥ ✖ t + 1 ✙ N ( 0 ❀✛ 2 t + 1 ) where ❳ 1 f t ( v t ♣ n ✖ t + 1 = j ) j ✷ S ✒ k ✓ ✒ 1 ✓ ❳ f t ( v t ♣ n = j ) k j ✷ S = ✔ E ❢ f t ( ✖ t + ✛ t ✘ ) ❣ ❀ ✘ ✘ N ( 0 ❀ 1 ) Deshpande, Montanari Planted Cliques November 5, 2013 26 / 49

  20. Summarizing . . . v t i , i / ∈ S 0 . 3 0 . 2 0 . 1 v t i , i ∈ S 0 − 4 − 2 0 2 4 Deshpande, Montanari Planted Cliques November 5, 2013 27 / 49

  21. State Evolution ✖ t + 1 = ✔ E ❢ f t ( ✖ t + ✛ t ✘ ) ❣ ✛ 2 t + 1 = E ❢ f t ( ✛ t ✘ ) 2 ❣ ✿ Using the optimal function f t ( x ) = e ✖ t x � ✖ 2 t ✖ t + 1 = ✔ e ✖ 2 t ❂ 2 ✛ 2 t + 1 = 1 Deshpande, Montanari Planted Cliques November 5, 2013 28 / 49

  22. State Evolution ✖ t + 1 = ✔ E ❢ f t ( ✖ t + ✛ t ✘ ) ❣ ✛ 2 t + 1 = E ❢ f t ( ✛ t ✘ ) 2 ❣ ✿ Using the optimal function f t ( x ) = e ✖ t x � ✖ 2 t ✖ t + 1 = ✔ e ✖ 2 t ❂ 2 ✛ 2 t + 1 = 1 Deshpande, Montanari Planted Cliques November 5, 2013 28 / 49

  23. Fixed points develop below threshold! 1 If ✔ ❃ ♣ e µ t +1 µ t Deshpande, Montanari Planted Cliques November 5, 2013 29 / 49

  24. Fixed points develop below threshold! 1 If ✔ ❃ ♣ e µ t +1 µ t Deshpande, Montanari Planted Cliques November 5, 2013 29 / 49

  25. Fixed points develop below threshold! 1 If ✔ ❃ ♣ e µ t +1 µ t Deshpande, Montanari Planted Cliques November 5, 2013 29 / 49

  26. Fixed points develop below threshold! 1 If ✔ ❃ ♣ e µ t +1 µ t Deshpande, Montanari Planted Cliques November 5, 2013 29 / 49

  27. Fixed points develop below threshold! 1 If ✔ ❃ ♣ e µ t +1 µ t Deshpande, Montanari Planted Cliques November 5, 2013 29 / 49

  28. Fixed points develop below threshold! 1 If ✔ ❃ ♣ e µ t +1 µ t Deshpande, Montanari Planted Cliques November 5, 2013 29 / 49

  29. Fixed points develop below threshold! 1 If ✔ ❃ ♣ e µ t +1 µ t Deshpande, Montanari Planted Cliques November 5, 2013 29 / 49

  30. Fixed points develop below threshold! 1 If ✔ ❃ ♣ e µ t +1 µ t Deshpande, Montanari Planted Cliques November 5, 2013 29 / 49

  31. Fixed points develop below threshold! 1 If ✔ ❃ ♣ e µ t +1 µ t Deshpande, Montanari Planted Cliques November 5, 2013 29 / 49

  32. Fixed points develop below threshold! 1 If ✔ ❃ ♣ e µ t +1 µ t Deshpande, Montanari Planted Cliques November 5, 2013 29 / 49

  33. Fixed points develop below threshold! 1 1 If ✔ ❃ If ✔ ❁ ♣ e ♣ e µ t +1 µ t +1 µ t µ t Deshpande, Montanari Planted Cliques November 5, 2013 29 / 49

  34. Fixed points develop below threshold! 1 1 If ✔ ❃ If ✔ ❁ ♣ e ♣ e µ t +1 µ t +1 µ t µ t Deshpande, Montanari Planted Cliques November 5, 2013 29 / 49

  35. Fixed points develop below threshold! 1 1 If ✔ ❃ If ✔ ❁ ♣ e ♣ e µ t +1 µ t +1 µ t µ t Deshpande, Montanari Planted Cliques November 5, 2013 29 / 49

  36. Fixed points develop below threshold! 1 1 If ✔ ❃ If ✔ ❁ ♣ e ♣ e µ t +1 µ t +1 µ t µ t Deshpande, Montanari Planted Cliques November 5, 2013 29 / 49

  37. Fixed points develop below threshold! 1 1 If ✔ ❃ If ✔ ❁ ♣ e ♣ e µ t +1 µ t +1 µ t µ t Deshpande, Montanari Planted Cliques November 5, 2013 29 / 49

  38. Analysis is wrong but. . . Theorem (Deshpande, Montanari, 2013) ♣ n ❂ e , there exists an O ( n 2 log n ) time algorithm that If ❥ S ❥ = k ✕ ( 1 + ✧ ) identifies S with high probability. . . . so we modify the algorithm. Deshpande, Montanari Planted Cliques November 5, 2013 30 / 49

  39. What algorithm? Slight modification to iterative scheme: ( v t i ) i ✷ [ n ] ✦ ( v t i ✦ j ) i ❀ j ✷ [ n ] ❳ 1 v t + 1 A i ❵ f t ( v t i ✦ j = ♣ n ❵ ✦ i ) ✿ ❵ ✻ = i ❀ j Analysis is exact as n ✦ ✶ . Deshpande, Montanari Planted Cliques November 5, 2013 31 / 49

  40. What algorithm? Slight modification to iterative scheme: ( v t i ) i ✷ [ n ] ✦ ( v t i ✦ j ) i ❀ j ✷ [ n ] ❳ 1 v t + 1 A i ❵ f t ( v t i ✦ j = ♣ n ❵ ✦ i ) ✿ ❵ ✻ = i ❀ j Analysis is exact as n ✦ ✶ . Deshpande, Montanari Planted Cliques November 5, 2013 31 / 49

  41. What algorithm? Slight modification to iterative scheme: ( v t i ) i ✷ [ n ] ✦ ( v t i ✦ j ) i ❀ j ✷ [ n ] ❳ 1 v t + 1 A i ❵ f t ( v t i ✦ j = ♣ n ❵ ✦ i ) ✿ ❵ ✻ = i ❀ j Analysis is exact as n ✦ ✶ . Deshpande, Montanari Planted Cliques November 5, 2013 31 / 49

  42. What algorithm? Slight modification to iterative scheme: ( v t i ) i ✷ [ n ] ✦ ( v t i ✦ j ) i ❀ j ✷ [ n ] ❳ 1 v t + 1 A i ❵ f t ( v t i ✦ j = ♣ n ❵ ✦ i ) ✿ ❵ ✻ = i ❀ j Analysis is exact as n ✦ ✶ . Deshpande, Montanari Planted Cliques November 5, 2013 31 / 49

  43. Fixing the heuristic Lemma Let ( f t ( z )) t ✕ 0 be a sequence of polynomials. Then, for every fixed t , and bounded, continuous function ✥ : R ✦ R the following limit holds in probability: ❳ 1 ✥ ( v t ♣ n lim i ✦ j ) = ✔ E ❢ ✥ ( ✖ t + ✛ t ✘ ) ❣ ❀ n ✦✶ i ✷ S ❳ 1 ✥ ( v t lim i ✦ j ) = E ❢ ✥ ( ✛ t ✘ ) ❣ ❀ n ✦✶ n i ✷ [ n ] ♥ S where ✘ ✘ N ( 0 ❀ 1 ) . Deshpande, Montanari Planted Cliques November 5, 2013 32 / 49

  44. Proof Technique Key ideas: Expand v t i ✦ j for polynomial f t ( ✁ ) Wrong analysis works if A ✦ A t Deshpande, Montanari Planted Cliques November 5, 2013 33 / 49

  45. Proof Technique Key ideas: Expand v t i ✦ j for polynomial f t ( ✁ ) Wrong analysis works if A ✦ A t Deshpande, Montanari Planted Cliques November 5, 2013 33 / 49

  46. Proof Technique Key ideas: Expand v t i ✦ j for polynomial f t ( ✁ ) Wrong analysis works if A ✦ A t Deshpande, Montanari Planted Cliques November 5, 2013 33 / 49

  47. Proof Technique - Expanding v t Let f t ( x ) = x 2 ❀ v 0 i ✦ j = 1 j ❳ i v 1 i ✦ j = A ik ✿ k ✻ = j k Deshpande, Montanari Planted Cliques November 5, 2013 34 / 49

  48. Proof Technique - Expanding v t j ❳ v 2 A ik ( v 1 k ✦ i ) 2 i ✦ j = i k ✻ = j ✒ ❳ ✓✒ ❳ ✓ ❳ = A ik A k ❵ A km k ✻ = j ❵ ✻ = i m ✻ = i k ❳ ❳ ❳ = A ik A k ❵ A km ✿ k ✻ = j ❵ ✻ = i m ✻ = i m ℓ Deshpande, Montanari Planted Cliques November 5, 2013 35 / 49

  49. Proof Technique - Expanding v t j ❳ v 2 A ik ( v 1 k ✦ i ) 2 i ✦ j = i k ✻ = j ✒ ❳ ✓✒ ❳ ✓ ❳ = A ik A k ❵ A km k ✻ = j ❵ ✻ = i m ✻ = i k ❳ ❳ ❳ = A ik A k ❵ A km ✿ k ✻ = j ❵ ✻ = i m ✻ = i m ℓ Deshpande, Montanari Planted Cliques November 5, 2013 35 / 49

  50. Proof Technique - Expanding v t j ❳ v 2 A ik ( v 1 k ✦ i ) 2 i ✦ j = i k ✻ = j ✒ ❳ ✓✒ ❳ ✓ ❳ = A ik A k ❵ A km k ✻ = j ❵ ✻ = i m ✻ = i k ❳ ❳ ❳ = A ik A k ❵ A km ✿ k ✻ = j ❵ ✻ = i m ✻ = i m ℓ Deshpande, Montanari Planted Cliques November 5, 2013 35 / 49

  51. Proof Technique ❳ ❳ v t + 1 A ik f t ( v t ✘ t + 1 A t ik f t ( ✘ t i ✦ j = k ✦ i ) ✿ i ✦ j = k ✦ i ) ✿ k ✻ = i k ✻ = i i i k k o o ℓ ℓ m m p p r r q q s w s w Deshpande, Montanari Planted Cliques November 5, 2013 36 / 49

  52. Proof Technique - a Combinatorial Lemma Lemma ❳ v t A ( T )Γ( T ) v 0 ( T ) i ✦ j = T ✷❚ t i ✦ j where ❚ t i ✦ j consists rooted, labeled trees that: 1 have maximum depth t . 2 do not backtrack. (Similarly for the ✘ t i ✦ j ) Deshpande, Montanari Planted Cliques November 5, 2013 37 / 49

  53. Proof Technique - Moment Method ❳ ❳ v t + 1 A ik f t ( v t ✘ t + 1 A t ik f t ( ✘ t i ✦ j = k ✦ i ) ✿ i ✦ j = k ✦ i ) ✿ k ✻ = i k ✻ = i i i k k o o ℓ ℓ m m p p r r q q s w s w lim n ✦✶ Moments of v t + 1 = Moments of ✘ t + 1 . Deshpande, Montanari Planted Cliques November 5, 2013 38 / 49

  54. Proof Technique - Moment Method ❳ ❳ v t + 1 A ik f t ( v t ✘ t + 1 A t ik f t ( ✘ t i ✦ j = k ✦ i ) ✿ i ✦ j = k ✦ i ) ✿ k ✻ = i k ✻ = i i i k k o o ℓ ℓ m m p p r r q q s w s w lim n ✦✶ Moments of v t + 1 = Moments of ✘ t + 1 . Deshpande, Montanari Planted Cliques November 5, 2013 38 / 49

  55. Progress(4) Complexity Spectral threshold = √ n n 2 log n n 2 log n n 2 k 1 . 261 √ n C √ n 2 √ n log n � n 2 log 2 n e Deshpande, Montanari Planted Cliques November 5, 2013 39 / 49

  56. Is this threshold fundamental? Rest of the talk: perhaps Deshpande, Montanari Planted Cliques November 5, 2013 40 / 49

  57. The “Hidden Set” Problem A ij ∼ Q 1 Given G n = ([ n ] ❀ E n ) A Set ✦ S ✚ [ n ] ✭ Q 1 if i ❀ j ✷ S ❀ Data ✦ A ij ✘ otherwise. Q 0 A ij ∼ Q 0 Problem: Given edge labels ( A ij ) ( i ❀ j ) ✷ E n , identify S Deshpande, Montanari Planted Cliques November 5, 2013 41 / 49

  58. The “Hidden Set” Problem A ij ∼ Q 1 Given G n = ([ n ] ❀ E n ) A Set ✦ S ✚ [ n ] ✭ Q 1 if i ❀ j ✷ S ❀ Data ✦ A ij ✘ otherwise. Q 0 A ij ∼ Q 0 Problem: Given edge labels ( A ij ) ( i ❀ j ) ✷ E n , identify S Deshpande, Montanari Planted Cliques November 5, 2013 41 / 49

  59. The “Hidden Set” Problem S Given G n = ([ n ] ❀ E n ) S A Set ✦ S ✚ [ n ] ✭ A = if i ❀ j ✷ S ❀ Q 1 Data ✦ A ij ✘ Q 0 otherwise. Problem: Given edge labels ( A ij ) ( i ❀ j ) ✷ E n , identify S Deshpande, Montanari Planted Cliques November 5, 2013 42 / 49

  60. “Local” Algorithms A t -local algorithm computes: Estimate at i : u ( i ) = F ( A Ball ( i ❀ t ) ) ❜ Deshpande, Montanari Planted Cliques November 5, 2013 43 / 49

  61. The Sparse Graph Analogue A ij ∼ Q 1 G n = ([ n ] ❀ E n ) ❀ n ✕ 1 satisfies: ◮ locally tree-like ◮ regular degree ∆ Further ◮ Q 1 = ✍ + 1 , Q 0 = 1 2 ✍ + 1 + 1 2 ✍ � 1 A ij ∼ Q 0 What can local algorithms achieve? Deshpande, Montanari Planted Cliques November 5, 2013 44 / 49

  62. The Sparse Graph Analogue A ij ∼ Q 1 G n = ([ n ] ❀ E n ) ❀ n ✕ 1 satisfies: ◮ locally tree-like ◮ regular degree ∆ Further ◮ Q 1 = ✍ + 1 , Q 0 = 1 2 ✍ + 1 + 1 2 ✍ � 1 A ij ∼ Q 0 What can local algorithms achieve? Deshpande, Montanari Planted Cliques November 5, 2013 44 / 49

Recommend


More recommend