matching size and matrix rank estimation in data streams
play

Matching Size and Matrix Rank Estimation in Data Streams Sepehr - PowerPoint PPT Presentation

Feb 10, 2024 •370 likes •1.74k views

Matching Size and Matrix Rank Estimation in Data Streams Sepehr Assadi University of Pennsylvania Joint work with Sanjeev Khanna (Penn), and Yang Li (Penn) Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar The Streaming Model

  1. Estimating Maximum Matching Size Question. Is matching size estimation strictly easier than finding an approximate matching? Yes! Theorem There is a randomized algorithm that outputs an – -approximate estimate of maximum matching size in: Â O ( n/ – 2 ) space in insertion-only streams. Â O ( n 2 / – 4 ) space in dynamic streams. In constrast, to find an – -approximate matching, the space necessary is: Ω ( n/ – ) in insertion-only streams. Ω ( n 2 / – 3 ) in dynamic streams. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  2. Algorithms for – -Estimation of Matching Size The main ingredient of our algorithms is the following sampling lemma: Lemma (Vertex Sampling Lemma) Let H be a subgraph of G obtained by sampling each vertex independently w.p. 1 / – . Define: µ G : the maximum matching size in G , µ H : the maximum matching size in H . Then, w.h.p., µ G – 2 Æ µ H Æ 2 µ G – . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  3. Algorithms for – -Estimation of Matching Size The main ingredient of our algorithms is the following sampling lemma: Lemma (Vertex Sampling Lemma) Let H be a subgraph of G obtained by sampling each vertex independently w.p. 1 / – . Define: µ G : the maximum matching size in G , µ H : the maximum matching size in H . Then, w.h.p., µ G – 2 Æ µ H Æ 2 µ G – . Therefore, maximum matching size in H is an – -estimation for the maximum matching size in G . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  4. Proof by Picture Any graph G with a maximum matching size of µ G looks as follows: A matching of size µ G G between the blue vertices. L R No edges between the green vertices. µ G Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  5. Proof by Picture The vertex sampled graph H then look as follows: A matching of size µ G / – 2 H between the blue vertices L R ∆ µ H Ø µ G / – 2 . = All edges are incident on µ G / – blue vertices µ G / – = ∆ µ H Æ 2 µ G / – . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  6. An – -Estimation Algorithm To distinguish between graphs with maximum matching of size Ø k and o ( k/ – ) : Sample each vertex in G w.p. 1 / – to obtain H . 1 Test whether H has a matching of size at least Ω ( k/ – 2 ) or not. 2 Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  7. An – -Estimation Algorithm To distinguish between graphs with maximum matching of size Ø k and o ( k/ – ) : Sample each vertex in G w.p. 1 / – to obtain H . 1 Test whether H has a matching of size at least Ω ( k/ – 2 ) or not. 2 Can be implemented in: O ( k/ – 2 ) = Â Â O ( n/ – 2 ) in insertion-only streams. O ( k 2 / – 4 ) = Â Â O ( n 2 / – 4 ) in dynamic streams. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  8. Estimating Maximum Matching Size Matching size estimation is indeed easier than finding an approximate matching! Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  9. Estimating Maximum Matching Size Matching size estimation is indeed easier than finding an approximate matching! Question. Is it possible to achieve an arbitrary good estimation of matching size in sub-quadratic space? Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  10. Estimating Maximum Matching Size Matching size estimation is indeed easier than finding an approximate matching! Question. Is it possible to achieve an arbitrary good estimation of matching size in sub-quadratic space? Question. In general, what is the space-approximation tradeo ff for matching size estimation? Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  11. Estimating Maximum Matching Size Matching size estimation is indeed easier than finding an approximate matching! Question. Is it possible to achieve an arbitrary good estimation of matching size in sub-quadratic space? Question. In general, what is the space-approximation tradeo ff for matching size estimation? We make progress on each of these questions. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  12. Our Main Results Near-optimal approximation of maximum matching size may require almost quadratic space even in insertion-only streams. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  13. Our Main Results Near-optimal approximation of maximum matching size may require almost quadratic space even in insertion-only streams. Theorem Any randomized (1 + Á ) -approximate estimation of maximum matching size requires: RS ( n ) · n 1 − O ( Á ) space in insertion-only streams. n 2 − O ( Á ) space in dynamic streams. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  14. Our Main Results Near-optimal approximation of maximum matching size may require almost quadratic space even in insertion-only streams. Theorem Any randomized (1 + Á ) -approximate estimation of maximum matching size requires: RS ( n ) · n 1 − O ( Á ) space in insertion-only streams. n 2 − O ( Á ) space in dynamic streams. RS ( n ) denotes the maximum number of edge-disjoint induced matchings of size Θ ( n ) in an n -vertex graph: [ Fischer et al., 2002 ] n Ω (1 / log log n ) Æ RS ( n ) Æ n/ log n [ Fox et al., 2015 ] Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  15. Our Main Results We further establish the first non-trivial lower bound for super-constant approximation of matching size. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  16. Our Main Results We further establish the first non-trivial lower bound for super-constant approximation of matching size. Theorem Any randomized – -approximate estimate of maximum matching size requires Ω ( n/ – 2 ) in dynamic streams. Furthermore, even if we restrict to sparse graphs with arboricity O ( – ) , Ω ( Ô n/ – 2 . 5 ) space is necessary. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  17. Our Main Results We further establish the first non-trivial lower bound for super-constant approximation of matching size. Theorem Any randomized – -approximate estimate of maximum matching size requires Ω ( n/ – 2 ) in dynamic streams. Furthermore, even if we restrict to sparse graphs with arboricity O ( – ) , Ω ( Ô n/ – 2 . 5 ) space is necessary. There is an active line of research on estimating matching size of bounded arboricity graphs in graph streams [Chitnis et al., 2016] [Bury and Schwiegelshohn, 2015] [Esfandiari et al., 2015] [McGregor and Vorotnikova, 2016b] [Cormode et al., 2016] [McGregor and Vorotnikova, 2016a] . . . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  18. Schatten p -Norms Given an n ◊ n matrix A , for any p œ [0 , Œ ) : Schatten p -norm of A is the p -th frequency moment of vector of singular values ( ‡ 1 , . . . , ‡ n ) of A . A n B 1 /p ÿ ‡ p Î A Î p := i i =1 Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  19. Schatten p -Norms Given an n ◊ n matrix A , for any p œ [0 , Œ ) : Schatten p -norm of A is the p -th frequency moment of vector of singular values ( ‡ 1 , . . . , ‡ n ) of A . A n B 1 /p ÿ ‡ p Î A Î p := i i =1 Î A Î 0 = Rank of A . Î A Î 1 = Trace norm of A . Î A Î 2 = Frobenius norm of A . Î A Î ∞ = Operator norm of A . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  20. Schatten p -Norms in Data Streams Question. What is the space complexity of (1 + Á ) -approximating the Schatten p -norm of a matrix which its entries are revealed in a data stream? Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  21. Schatten p -Norms in Data Streams Question. What is the space complexity of (1 + Á ) -approximating the Schatten p -norm of a matrix which its entries are revealed in a data stream? Previous work: For p = 0 , Ω ( n 1 − g ( Á ) ) space is necessary [Bury and Schwiegelshohn, 2015]. For p œ (0 , Œ ) \ 2 Z , Ω ( n 1 − g ( Á ) ) space is necessary [Li and Woodru ff , 2016]. For p œ 2 Z \ { 0 } , Ω ( n 1 − 2 /p ) space is necessary [Li and Woodru ff , 2016] (and is su ffi cient for sparse matrices). Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  22. Schatten p -Norms in Data Streams Question. What is the space complexity of (1 + Á ) -approximating the Schatten p -norm of a matrix which its entries are revealed in a data stream? Previous work: For p = 0 , Ω ( n 1 − g ( Á ) ) space is necessary [Bury and Schwiegelshohn, 2015]. For p œ (0 , Œ ) \ 2 Z , Ω ( n 1 − g ( Á ) ) space is necessary [Li and Woodru ff , 2016]. For p œ 2 Z \ { 0 } , Ω ( n 1 − 2 /p ) space is necessary [Li and Woodru ff , 2016] (and is su ffi cient for sparse matrices). We answer this question for the case of rank computation. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  23. Matrix Rank Computation in Data Streams It is well-known that computing maximum matching size of a graph is equivalent to computing the rank of the (symbolic) Tutte matrix. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  24. Matrix Rank Computation in Data Streams It is well-known that computing maximum matching size of a graph is equivalent to computing the rank of the (symbolic) Tutte matrix. As a corollary, all our lower bounds for matching size estimation also extend to the matrix rank computation problem. In particular, An Ω ( n 2 − O ( Á ) ) space lower bound for (1 + Á ) -estimation of rank in dense matrices. Ω ( Ô n ) space lower bound for any polylog( n ) -estimation of An  rank in sparse matrices. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  25. An Ω ( n 2 ≠ O ( Á ) ) Lower Bound for Dynamic Streams Theorem Any randomized (1 + Á ) -approximate estimation of maximum matching size requires Ω ( n 2 − O ( Á ) ) space in dynamic streams. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  26. Previous Approaches: An n 1 ≠ O ( Á ) Lower Bound Consider the following two-player one-way communication problem. MaxMatching : Alice is given a matching M on vertices V . 1 Bob is given a collection of edges E B on vertices V . 2 Alice sends a single message to Bob and Bob outputs an 3 estimation of maximum matching size in G ( V, M fi E B ) . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  27. Previous Approaches: An n 1 ≠ O ( Á ) Lower Bound Consider the following two-player one-way communication problem. MaxMatching : Alice is given a matching M on vertices V . 1 Bob is given a collection of edges E B on vertices V . 2 Alice sends a single message to Bob and Bob outputs an 3 estimation of maximum matching size in G ( V, M fi E B ) . CC( MaxMatching ): minimum length message to solve this problem with probability, say, 2 / 3 . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  28. Previous Approaches: An n 1 ≠ O ( Á ) Lower Bound Consider the following two-player one-way communication problem. MaxMatching : Alice is given a matching M on vertices V . 1 Bob is given a collection of edges E B on vertices V . 2 Alice sends a single message to Bob and Bob outputs an 3 estimation of maximum matching size in G ( V, M fi E B ) . CC( MaxMatching ): minimum length message to solve this problem with probability, say, 2 / 3 . Fact. CC( MaxMatching ) Æ space complexity of any streaming algorithm for estimating maximum matching size. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  29. Previous Approaches: An n 1 ≠ O ( Á ) Lower Bound [Bury and Schwiegelshohn, 2015]: CC( MaxMatching ) = Ω ( n 1 − O ( Á ) ) . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  30. Previous Approaches: An n 1 ≠ O ( Á ) Lower Bound [Bury and Schwiegelshohn, 2015]: CC( MaxMatching ) = Ω ( n 1 − O ( Á ) ) . Proof Sketch. (for Á = 1 / 2 ) Alice is given a random subset of size n/ 2 from a fixed perfect matching between L and R . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  31. Previous Approaches: An n 1 ≠ O ( Á ) Lower Bound [Bury and Schwiegelshohn, 2015]: CC( MaxMatching ) = Ω ( n 1 − O ( Á ) ) . Proof Sketch. (for Á = 1 / 2 ) Alice is given a random subset of size n/ 2 from a fixed perfect matching between L and R . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  32. Previous Approaches: An n 1 ≠ O ( Á ) Lower Bound [Bury and Schwiegelshohn, 2015]: CC( MaxMatching ) = Ω ( n 1 − O ( Á ) ) . Proof Sketch. (for Á = 1 / 2 ) Alice is given a random subset of size n/ 2 from a fixed perfect matching between L and R . Bob is given a matching of size n/ 2 incident on R . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  33. Previous Approaches: An n 1 ≠ O ( Á ) Lower Bound [Bury and Schwiegelshohn, 2015]: CC( MaxMatching ) = Ω ( n 1 − O ( Á ) ) . Proof Sketch. (for Á = 1 / 2 ) Alice is given a random subset of size n/ 2 from a fixed perfect matching between L and R . Bob is given a matching of size n/ 2 incident on R . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  34. Previous Approaches: An n 1 ≠ O ( Á ) Lower Bound [Bury and Schwiegelshohn, 2015]: CC( MaxMatching ) = Ω ( n 1 − O ( Á ) ) . Proof Sketch. (for Á = 1 / 2 ) Alice is given a random subset of size n/ 2 from a fixed perfect matching between L and R . Bob is given a matching of size n/ 2 incident on R . Yes case: Each edge of Bob’s matching is incident on even number of Alice’s matching = ∆ MaxMatching = 3 n/ 4 . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  35. Previous Approaches: An n 1 ≠ O ( Á ) Lower Bound [Bury and Schwiegelshohn, 2015]: CC( MaxMatching ) = Ω ( n 1 − O ( Á ) ) . Proof Sketch. (for Á = 1 / 2 ) Alice is given a random subset of size n/ 2 from a fixed perfect matching between L and R . Bob is given a matching of size n/ 2 incident on R . No case: Each edge of Bob’s matching is incident on odd number of Alice’s matching = ∆ MaxMatching = n/ 2 . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  36. Previous Approaches: An n 1 ≠ O ( Á ) Lower Bound [Bury and Schwiegelshohn, 2015]: CC( MaxMatching ) = Ω ( n 1 − O ( Á ) ) . Proof Sketch. (for Á = 1 / 2 ) A better than 3 / 2 -approximation distinguishes between the two cases. Distinguishing between the two cases requires Ω ( Ô n ) communication by a reduction from the boolean hidden matching problem of [Gavinsky et al., 2007]. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  37. Our Approach A natural idea to boost the previous lower bound: Instead of one matching M , provide Alice with t independently 1 chosen matchings M 1 , . . . , M t . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  38. Our Approach A natural idea to boost the previous lower bound: Instead of one matching M , provide Alice with t independently 1 chosen matchings M 1 , . . . , M t . Provide Bob with a single set E B of edges as before. 2 Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  39. Our Approach A natural idea to boost the previous lower bound: Instead of one matching M , provide Alice with t independently 1 chosen matchings M 1 , . . . , M t . Provide Bob with a single set E B of edges as before. 2 “Ask” Alice and Bob to solve the MaxMatching problem for 3 a uniformly at random chosen matching M j ı and E B (the index j ı is unknown to Alice). Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  40. Our Approach A natural idea to boost the previous lower bound: Instead of one matching M , provide Alice with t independently 1 chosen matchings M 1 , . . . , M t . Provide Bob with a single set E B of edges as before. 2 “Ask” Alice and Bob to solve the MaxMatching problem for 3 a uniformly at random chosen matching M j ı and E B (the index j ı is unknown to Alice). The hope is that communication complexity of this problem is now Ø t · CC( MaxMatching ). Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  41. Our Approach There are three main obstacles in implementing this idea: The matchings M 1 , . . . , M t should be supported on Θ ( n ) 1 vertices as opposed to trivial Θ ( t · n ) vertices. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  42. Our Approach There are three main obstacles in implementing this idea: The matchings M 1 , . . . , M t should be supported on Θ ( n ) 1 vertices as opposed to trivial Θ ( t · n ) vertices. The matchings should be chosen independently even though 2 they are supported on the same set of Θ ( n ) vertices. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  43. Our Approach There are three main obstacles in implementing this idea: The matchings M 1 , . . . , M t should be supported on Θ ( n ) 1 vertices as opposed to trivial Θ ( t · n ) vertices. The matchings should be chosen independently even though 2 they are supported on the same set of Θ ( n ) vertices. The reduction should ensure that Alice and Bob indeed need to 3 solve the j ı -th embedded instance. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  44. Our Approach There are three main obstacles in implementing this idea: The matchings M 1 , . . . , M t should be supported on Θ ( n ) 1 vertices as opposed to trivial Θ ( t · n ) vertices. The matchings should be chosen independently even though 2 they are supported on the same set of Θ ( n ) vertices. The reduction should ensure that Alice and Bob indeed need to 3 solve the j ı -th embedded instance. (1) + (2) = ∆ Ruzsa-Szemer´ edi graphs (RS graphs). Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  45. Our Approach There are three main obstacles in implementing this idea: The matchings M 1 , . . . , M t should be supported on Θ ( n ) 1 vertices as opposed to trivial Θ ( t · n ) vertices. The matchings should be chosen independently even though 2 they are supported on the same set of Θ ( n ) vertices. The reduction should ensure that Alice and Bob indeed need to 3 solve the j ı -th embedded instance. (1) + (2) = ∆ Ruzsa-Szemer´ edi graphs (RS graphs). RS graphs + (3) = ∆ characterization of dynamic streaming algorithms via simultaneous communication complexity. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  46. Our Approach There are three main obstacles in implementing this idea: The matchings M 1 , . . . , M t should be supported on Θ ( n ) 1 vertices as opposed to trivial Θ ( t · n ) vertices. The matchings should be chosen independently even though 2 they are supported on the same set of Θ ( n ) vertices. The reduction should ensure that Alice and Bob indeed need to 3 solve the j ı -th embedded instance. (1) + (2) = ∆ Ruzsa-Szemer´ edi graphs (RS graphs). RS graphs + (3) = ∆ characterization of dynamic streaming algorithms via simultaneous communication complexity. Formalizing the lower bound = ∆ a direct-sum style argument using information complexity. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  47. Ruzsa-Szemer´ edi Graphs Definition ( ( r, t ) -RS graphs) A graph G ( V, E ) whose edges can be partitioned into t induced matchings of size r each. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  48. Ruzsa-Szemer´ edi Graphs Definition ( ( r, t ) -RS graphs) A graph G ( V, E ) whose edges can be partitioned into t induced matchings of size r each. Example. A (2 , 4) -RS graph 1 on 8 vertices: Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  49. Ruzsa-Szemer´ edi Graphs Definition ( ( r, t ) -RS graphs) A graph G ( V, E ) whose edges can be partitioned into t induced matchings of size r each. Example. A (2 , 4) -RS graph 1 on 8 vertices: Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  50. Ruzsa-Szemer´ edi Graphs Definition ( ( r, t ) -RS graphs) A graph G ( V, E ) whose edges can be partitioned into t induced matchings of size r each. Example. A (2 , 4) -RS graph 1 on 8 vertices: Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  51. Ruzsa-Szemer´ edi Graphs Definition ( ( r, t ) -RS graphs) A graph G ( V, E ) whose edges can be partitioned into t induced matchings of size r each. Example. A (2 , 4) -RS graph 1 on 8 vertices: Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  52. Ruzsa-Szemer´ edi Graphs Definition ( ( r, t ) -RS graphs) A graph G ( V, E ) whose edges can be partitioned into t induced matchings of size r each. Example. A (2 , 4) -RS graph 1 on 8 vertices: Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  53. Ruzsa-Szemer´ edi Graphs We are typically interested in RS graphs with large values of r and t as a function of n . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  54. Ruzsa-Szemer´ edi Graphs We are typically interested in RS graphs with large values of r and t as a function of n . How dense a graph with many large induced matching can be? Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  55. Ruzsa-Szemer´ edi Graphs We are typically interested in RS graphs with large values of r and t as a function of n . How dense a graph with many large induced matching can be? Theorem ([Fischer et al., 2002]) There exists an ( r, t ) -RS graph on n vertices with t = n Ω (1 / log log n ) induced matchings of size r = (1 ≠ Á ) · n/ 4 . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  56. Ruzsa-Szemer´ edi Graphs We are typically interested in RS graphs with large values of r and t as a function of n . How dense a graph with many large induced matching can be? Theorem ([Fischer et al., 2002]) There exists an ( r, t ) -RS graph on n vertices with t = n Ω (1 / log log n ) induced matchings of size r = (1 ≠ Á ) · n/ 4 . Theorem ([Alon et al., 2012]) There exists an ( r, t ) -RS graph on n vertices with t = n 1+ o (1) induced matchings of size r = n 1 − o (1) . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  57. A Simple Lower Bound for Insertion-Only Streams Let G 1 be an ( r, t ) -RS bipartite graph on n vertices on each side. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  58. A Simple Lower Bound for Insertion-Only Streams Let G 1 be an ( r, t ) -RS bipartite graph on n vertices on each side. To Alice, we give random subset of size r/ 2 from each induced matchings M 1 , . . . , M t of G 1 . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  59. A Simple Lower Bound for Insertion-Only Streams Let G 1 be an ( r, t ) -RS bipartite graph on n vertices on each side. To Alice, we give random subset of size r/ 2 from each induced matchings M 1 , . . . , M t of G 1 . Choose M j ı uniformly at random. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  60. A Simple Lower Bound for Insertion-Only Streams Let G 1 be an ( r, t ) -RS bipartite graph on n vertices on each side. To Alice, we give random subset of size r/ 2 from each induced matchings M 1 , . . . , M t of G 1 . Choose M j ı uniformly at random. To Bob we give the following input: I A matching between vertices not in M j ı and a new set of vertices. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  61. A Simple Lower Bound for Insertion-Only Streams Let G 1 be an ( r, t ) -RS bipartite graph on n vertices on each side. To Alice, we give random subset of size r/ 2 from each induced matchings M 1 , . . . , M t of G 1 . Choose M j ı uniformly at random. To Bob we give the following input: I A matching between vertices not in M j ı and a new set of vertices. I A graph E B over the set of vertices in M j ı . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  62. A Simple Lower Bound for Insertion-Only Streams Size of the maximum matching in this graph: 2( n ≠ r ) + MaxMatching ( M j ı , E B ) Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  63. A Simple Lower Bound for Insertion-Only Streams Size of the maximum matching in this graph: 2( n ≠ r ) + MaxMatching ( M j ı , E B ) For r = Θ ( n ) , Alice and Bob need to solve MaxMatching ( M j ı , E B ) for (1 + Á ) -approximation. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  64. A Simple Lower Bound for Insertion-Only Streams Size of the maximum matching in this graph: 2( n ≠ r ) + MaxMatching ( M j ı , E B ) For r = Θ ( n ) , Alice and Bob need to solve MaxMatching ( M j ı , E B ) for (1 + Á ) -approximation. To solve this for an unknown matching M j ı , the message length must be Ø t · CC( MaxMatching ). Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  65. A Simple Lower Bound for Insertion-Only Streams Main limitation of this approach: Requires r = Θ ( n ) = ∆ t = RS ( n ) . RS ( n ) maybe as large as n/ log n . However, best known bound for RS ( n ) is only n Ω (1 / log log n ) . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  66. A Simple Lower Bound for Insertion-Only Streams Main limitation of this approach: Requires r = Θ ( n ) = ∆ t = RS ( n ) . RS ( n ) maybe as large as n/ log n . However, best known bound for RS ( n ) is only n Ω (1 / log log n ) . We bypass the r = Θ ( n ) limitation in dynamic streams using the characterization result of [Li et al., 2014, Ai et al., 2016] in terms of the simultaneous communication complexity. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  67. The Simultaneous Communication Model The input is partitioned between k players P (1) , . . . , P ( k ) . Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  68. The Simultaneous Communication Model The input is partitioned between k players P (1) , . . . , P ( k ) . There exists an additional party called the referee. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

  69. The Simultaneous Communication Model The input is partitioned between k players P (1) , . . . , P ( k ) . There exists an additional party called the referee. Players P (1) , . . . , P ( k ) simultaneously send a message to the referee who outputs the answer. Sepehr Assadi (Penn) JHU Algorithms and Complexity Seminar

Recommend

1 SVD applications: rank, column, row, and null spaces Rank : the rank of a matrix is equal to:

1 SVD applications: rank, column, row, and null spaces Rank : the rank of a matrix is equal to:

Statistical Modeling and Analysis of Neural Data (NEU 560) Princeton University, Spring 2018 Jonathan Pillow Lecture 3A notes: SVD and Linear Systems 1 SVD applications: rank, column, row, and null spaces Rank : the rank of a matrix is equal to:

462 views • 3 slides
Bayesian Estimation of Low-rank Matrices Pierre Alquier Journes de Statistique du Sud,

Bayesian Estimation of Low-rank Matrices Pierre Alquier Journes de Statistique du Sud,

Bayesian Estimation of Low-rank Matrices : Introduction Bayesian Reduced Rank Regression Bayesian Matrix Completion Bayesian Estimation of Low-rank Matrices Pierre Alquier Journes de Statistique du Sud, Barcelona 09/06/2014 Pierre Alquier

596 views • 44 slides
1 Low-rank approximations to a matrix using SVD First point: we can write the SVD as a sum of

1 Low-rank approximations to a matrix using SVD First point: we can write the SVD as a sum of

Statistical Modeling and Analysis of Neural Data (NEU 560) Princeton University, Spring 2018 Jonathan Pillow Lecture 4 notes: PCA Thurs, 2.15 1 Low-rank approximations to a matrix using SVD First point: we can write the SVD as a sum of rank-1

363 views • 3 slides
New matrix norms for structured matrix estimation Jean-Philippe Vert Optimization and Statistical

New matrix norms for structured matrix estimation Jean-Philippe Vert Optimization and Statistical

New matrix norms for structured matrix estimation Jean-Philippe Vert Optimization and Statistical Learning workshop Les Houches, France, Jan 11-16, 2015 Outline Atomic norms 1 Sparse matrices with disjoint column supports 2 Low-rank

638 views • 46 slides
Phase transitions in low-rank matrix estimation May 11, 2017 Marc Lelarge & L eo Miolane

Phase transitions in low-rank matrix estimation May 11, 2017 Marc Lelarge & L eo Miolane

Phase transitions in low-rank matrix estimation May 11, 2017 Marc Lelarge & L eo Miolane INRIA, ENS 1 / 14 Introduction The statistical model Spiked Wigner model n XX Y + Z =

516 views • 22 slides
Low-rank Matrix Estimation via Approximate Message Passing Andrea Montanari Ramji Venkataramanan

Low-rank Matrix Estimation via Approximate Message Passing Andrea Montanari Ramji Venkataramanan

Low-rank Matrix Estimation via Approximate Message Passing Andrea Montanari Ramji Venkataramanan Stanford University University of Cambridge WoLA 2018 1 / 25 The Spiked Model k i v i v T R n n A = i + W i =1 1 2

783 views • 36 slides
min such that y 1 2 Then we have: C q S is created by retaining the S

min such that y 1 2 Then we have: C q S is created by retaining the S

Suppose the matrix A= FY of size m by n (where sensing matrix F has size m by n , and basis matrix Y has size n by n ) has RIP property of order 2 S where d 2S < 0.41. Let the solution of the following be denoted as q * , (for signal f = Yq ,

722 views • 36 slides
Learning-Based* Frequency Estimation in Data Streams Chen-Yu Hsu Piotr Indyk Dina Katabi Ali

Learning-Based* Frequency Estimation in Data Streams Chen-Yu Hsu Piotr Indyk Dina Katabi Ali

1 2 3 4 5 6 7 8 9 10 Learning-Based* Frequency Estimation in Data Streams Chen-Yu Hsu Piotr Indyk Dina Katabi Ali Vakilian (+Anders Aamand) MIT *A.k.a. Automated / Data-Driven Data Streams A data stream is a (massive)

265 views • 14 slides
Matrix invertibility Rank-Nullity Theorem: For any n -column matrix A , nullity A + rank A = n

Matrix invertibility Rank-Nullity Theorem: For any n -column matrix A , nullity A + rank A = n

Matrix invertibility Rank-Nullity Theorem: For any n -column matrix A , nullity A + rank A = n Corollary: Let A be an R C matrix. Then A is invertible if and only if | R | = | C | and the columns of A are linearly independent. F R by f ( x )

409 views • 21 slides
Lower Bounds for Quantile Estimation in Random-Order and Multi-Pass Streams Sudipto Guha (UPenn)

Lower Bounds for Quantile Estimation in Random-Order and Multi-Pass Streams Sudipto Guha (UPenn)

Lower Bounds for Quantile Estimation in Random-Order and Multi-Pass Streams Sudipto Guha (UPenn) Andrew McGregor (UCSD) Data Stream Model Data Stream Model Stream: m elements from a universe of size n :

991 views • 82 slides
Online robust matrix factorization for dependent data streams Hanbaek Lyu Department of

Online robust matrix factorization for dependent data streams Hanbaek Lyu Department of

Online robust matrix factorization for dependent data streams Hanbaek Lyu Department of Mathematics, University of California, Los Angeles Seminar on applied math and data Science, HKUST Joint work with HanQin Cai and Deanna Needell Mar. 24,

963 views • 71 slides
Matrix Factorization DS-GA 1013 / MATH-GA 2824 Optimization-based Data Analysis

Matrix Factorization DS-GA 1013 / MATH-GA 2824 Optimization-based Data Analysis

Matrix Factorization DS-GA 1013 / MATH-GA 2824 Optimization-based Data Analysis http://www.cims.nyu.edu/~cfgranda/pages/OBDA_fall17/index.html Carlos Fernandez-Granda Low-rank models Matrix completion Structured low-rank models Motivation

586 views • 56 slides
Optimal Estimation of Matching Constraints 1 Motivation & general approach 2 Parametrization

Optimal Estimation of Matching Constraints 1 Motivation & general approach 2 Parametrization

Optimal Estimation of Matching Constraints 1 Motivation & general approach 2 Parametrization of matching constraints 3 Direct vs. reduced fitting 4 Numerical methods 5 Robustification 6 Summary Why Study Matching Constraint Estimation? :

392 views • 35 slides
Outline Morning program Preliminaries Text matching I Text matching II Afternoon program

Outline Morning program Preliminaries Text matching I Text matching II Afternoon program

Outline Morning program Preliminaries Text matching I Text matching II Afternoon program Learning to rank Modeling user behavior Generating responses Wrap up 49 Text matching I Supervised text matching Traditional IR data consists of

370 views • 26 slides
Te Testing sting th the e Bo Boolean lean Ran Rank Michal Parnas Joint work with: Dana

Te Testing sting th the e Bo Boolean lean Ran Rank Michal Parnas Joint work with: Dana

Te Testing sting th the e Bo Boolean lean Ran Rank Michal Parnas Joint work with: Dana Ron, Adi Shraibman The Real rank The real rank of a matrix M n n of size n n: Maximal # independent rows/columns of M. Minimal r such

581 views • 17 slides
Parallel Numerical Algorithms Chapter 6 Matrix Models Section 6.2 Low Rank Approximation

Parallel Numerical Algorithms Chapter 6 Matrix Models Section 6.2 Low Rank Approximation

Low Rank Approximation by SVD Computing Low Rank Approximations Randomness and Approximation Hierarchical Low-Rank Structure Parallel Numerical Algorithms Chapter 6 Matrix Models Section 6.2 Low Rank Approximation Edgar Solomonik

584 views • 37 slides
Matching of Matrix Elements and Parton Showers CKKW matching in e + e collisions Lecture 2:

Matching of Matrix Elements and Parton Showers CKKW matching in e + e collisions Lecture 2:

Lecture 2: Matching in e + e collisions Why Matching? Present matching approaches Matching of Matrix Elements and Parton Showers CKKW matching in e + e collisions Lecture 2: Matching in e + e collisions The MLM procedure Johan

610 views • 29 slides
Data Streams Many large sources of data are generated as streams of updates: IP Network

Data Streams Many large sources of data are generated as streams of updates: IP Network

Summarizing and Mining Skewed Data Streams Graham Cormode cormode@bell-labs.com Flip Korn, S. Muthukrishnan, Divesh Srivastava Data Streams Many large sources of data are generated as streams of updates: IP Network traffic data Text:

485 views • 44 slides
Data Streams Many large sources of data are generated as streams of updates: IP Network

Data Streams Many large sources of data are generated as streams of updates: IP Network

Summarizing and Mining Skewed Data Streams Graham Cormode cormode@bell-labs.com S. Muthukrishnan muthu@cs.rutgers.edu Data Streams Many large sources of data are generated as streams of updates: IP Network traffic data Text: email/

535 views • 20 slides
Dr. OLEKSANDR SAKHARENKO COUNSEL KESAREV CONSULTING ESTIMATION OF THE SIZE OF SHADOW

Dr. OLEKSANDR SAKHARENKO COUNSEL KESAREV CONSULTING ESTIMATION OF THE SIZE OF SHADOW

ESTIMATION OF THE SCALE OF THE SHADOW MARKET IN UKRAINES AGRICULTURE Dr. OLEKSANDR SAKHARENKO COUNSEL KESAREV CONSULTING ESTIMATION OF THE SIZE OF SHADOW LEASE OF AGRICULTURAL LAND BY COMPARING DATA OF STATEGEOCADASTRE AND

896 views • 25 slides
2 3 4 5 8 9 MINNEAPOLIS MILWAUKEE MSA RANK #16 MSA RANK #39 CHICAGO MSA RANK #3

2 3 4 5 8 9 MINNEAPOLIS MILWAUKEE MSA RANK #16 MSA RANK #39 CHICAGO MSA RANK #3

2 3 4 5 8 9 MINNEAPOLIS MILWAUKEE MSA RANK #16 MSA RANK #39 CHICAGO MSA RANK #3 INDIANAPOLIS MSA RANK #34 DETROIT MSA RANK #14 COLUMBUS MSA RANK #33 BALTIMORE/ DC CORRIDOR MSA RANK #21 CINCINNATI MSA RANK #28 ATLANTA MSA RANK

726 views • 44 slides
Multiple-Rank Updates to Matrix Factorizations Zack 8/30/2013 Outline u Introduction u

Multiple-Rank Updates to Matrix Factorizations Zack 8/30/2013 Outline u Introduction u

Multiple-Rank Updates to Matrix Factorizations Zack 8/30/2013 Outline u Introduction u Multiple-rank Update Theory u Sparse Matrix Factorization by Multiple-rank Update u Symmetric Factorization by Multiple-rank Update u Conclusion Introduction

894 views • 42 slides
PWA with full rank density matrix of the + and 0 0 systems at VES

PWA with full rank density matrix of the + and 0 0 systems at VES

Preface PWA method Fit results Conclusions Backup slides PWA with full rank density matrix of the + and 0 0 systems at VES setup Igor Kachaev, Dmitry Ryabchikov for VES group, Protvino, Russia Institute for High

196 views • 16 slides
Contents Introduction Approaches and Methods for Software Size Estimation FPA Method

Contents Introduction Approaches and Methods for Software Size Estimation FPA Method

Estimating Software Size and Effort Software Size and Effort Estimation Dr. Ale ivkovi , CISA, PRINCE 2 University of Maribor, Slovenia Faculty of Electrical Engineering and Computer Science e-mail: ales.zivkovic@uni-mb.si

800 views • 38 slides

More recommend