Distributed Approximate Single-Source Shortest Paths

Transcription

1 Distributed Approximate Single-Source Shortest Paths Sebastian Krinninger University of Vienna joint works with Ruben Monika Andreas Christoph Danupon Becker Henzinger Karrenbauer Lenzen Nanongkai 1 / 26

2 One Problem Two Results Distributed (1 + ε)-approximate single-source shortest paths (SSSP) 2 / 26

3 One Problem Two Results Distributed (1 + ε)-approximate single-source shortest paths (SSSP) 1 Deterministically compute approximate shortest paths in ( n + Diam) n o(1) rounds for ε 1/polylog(n) [Henzinger/K/Nanongkai 16] 2 / 26

4 One Problem Two Results Distributed (1 + ε)-approximate single-source shortest paths (SSSP) 1 Deterministically compute approximate shortest paths in ( n + Diam) n o(1) rounds for ε 1/polylog(n) [Henzinger/K/Nanongkai 16] 2 Deterministically compute approximate shortest paths in ( n + Diam) poly(log n, ε) rounds [Becker/Lenzen/Karrenbauer/K 16] 2 / 26

5 One Problem Two Results Distributed (1 + ε)-approximate single-source shortest paths (SSSP) 1 Deterministically compute approximate shortest paths in ( n + Diam) n o(1) rounds for ε 1/polylog(n) [Henzinger/K/Nanongkai 16] 2 Deterministically compute approximate shortest paths in ( n + Diam) poly(log n, ε) rounds [Becker/Lenzen/Karrenbauer/K 16] Comparison: Lower bound: Ω( n + Diam) rounds [Das Sarma et al 11] Exact SSSP: O((n log n) 2/3 Diam 1/3 ) rounds (randomized) [Elkin 17] 1 + ε: O(n 1/2 Diam 1/4 + Diam) (randomized) [Nanongkai 14] 2 / 26

6 One Problem Two Results Distributed (1 + ε)-approximate single-source shortest paths (SSSP) 1 Deterministically compute approximate shortest paths in ( n + Diam) n o(1) rounds for ε 1/polylog(n) [Henzinger/K/Nanongkai 16] 2 Deterministically compute approximate shortest paths in ( n + Diam) poly(log n, ε) rounds [Becker/Lenzen/Karrenbauer/K 16] Comparison: Lower bound: Ω( n + Diam) rounds [Das Sarma et al 11] Exact SSSP: O((n log n) 2/3 Diam 1/3 ) rounds (randomized) [Elkin 17] 1 + ε: O(n 1/2 Diam 1/4 + Diam) (randomized) [Nanongkai 14] Today: Weighted undirected graphs 2 / 26

7 Tight and Tighter 3 / 26

8 Tight and Tighter Combinatorics & Optimization 3 / 26

9 Model and Problem Statement Many distributed models measure amount of communication Running time = number of rounds 4 / 26

10 Model and Problem Statement Many distributed models measure amount of communication Running time = number of rounds CONGEST model: Synchronous rounds (global clock) Message size O(log n) In each round, every node sends (at most) one message to each neighbor Local computation is free 4 / 26

11 Model and Problem Statement Many distributed models measure amount of communication Running time = number of rounds CONGEST model: Synchronous rounds (global clock) Message size O(log n) In each round, every node sends (at most) one message to each neighbor Local computation is free Problem statement: Initially, each node knows whether it is the source or not Finally: Every node knows its approximate distance to the source Often also: Implicit tree; every node knows next edge on approximate shortest path to source 4 / 26

12 Unweighted Graphs: BFS s 5 / 26

13 Unweighted Graphs: BFS 0 5 / 26

14 Unweighted Graphs: BFS 0 5 / 26

15 Unweighted Graphs: BFS / 26

16 Unweighted Graphs: BFS / 26

17 Unweighted Graphs: BFS / 26

18 Unweighted Graphs: BFS / 26

19 Unweighted Graphs: BFS BFS tree can be computed in O(Diam) rounds 5 / 26

20 Reduce to SSSP on Overlay Network [Nanongkai 14] 6 / 26

21 Reduce to SSSP on Overlay Network [Nanongkai 14] 1 Solve SSSP on overlay network and make global knowledge 2 Combine local knowledge of local neighborhoods with global knowledge 6 / 26

22 Reduce to SSSP on Overlay Network [Nanongkai 14] 1 Solve SSSP on overlay network and make global knowledge 2 Combine local knowledge of local neighborhoods with global knowledge Sample N = Õ(n 1/2 ) centers (+ source s) Every shortest path with n 1/2 edges contains center whp 6 / 26

23 Derandomization of Overlay Network [HKN 16] Randomization: Hit every shortest path with n edges u v 7 / 26

24 Derandomization of Overlay Network [HKN 16] Randomization: Hit every shortest path with n edges u v Deterministic relaxation: Almost hit every path n edges εd(u, v) u v 7 / 26

25 Congested Clique Special model: Communication not restricted to neighbors In each round, each node can send one message to each other node Heavily studied in recent years! 8 / 26

26 Congested Clique Special model: Communication not restricted to neighbors In each round, each node can send one message to each other node Heavily studied in recent years! Simulation: Overlay network as congested clique t rounds in Congested Clique Õ(t ( n + Diam)) rounds in CONGEST 8 / 26

27 Hop Reduction 9 / 26

28 Well Known: Spanners Definition A k-spanner is a subgraph H of G such that, for all pairs of nodes u and v, dist H (u, v) k dist G (u, v). 10 / 26

29 Well Known: Spanners Definition A k-spanner is a subgraph H of G such that, for all pairs of nodes u and v, dist H (u, v) k dist G (u, v). 10 / 26

30 Well Known: Spanners Definition A k-spanner is a subgraph H of G such that, for all pairs of nodes u and v, dist H (u, v) k dist G (u, v). 10 / 26

31 Well Known: Spanners Definition A k-spanner is a subgraph H of G such that, for all pairs of nodes u and v, dist H (u, v) k dist G (u, v). Fact: Every graph has a k-spanner of size n 1+1/k [Folklore] Application: Running time T(m, n) T(n 1+1/k, n) 10 / 26

32 Less Known: Hop Sets Definition An (h, ε)-hop set is a set of weighted edges F such that, for all pairs of nodes u and v, in the shortcut graph G F there is a path from u to v with at most h edges of weight at most (1 + ε)dist(u, v). 11 / 26

33 Less Known: Hop Sets Definition An (h, ε)-hop set is a set of weighted edges F such that, for all pairs of nodes u and v, in the shortcut graph G F there is a path from u to v with at most h edges of weight at most (1 + ε)dist(u, v). 11 / 26

34 Less Known: Hop Sets Definition An (h, ε)-hop set is a set of weighted edges F such that, for all pairs of nodes u and v, in the shortcut graph G F there is a path from u to v with at most h edges of weight at most (1 + ε)dist(u, v). 11 / 26

35 Less Known: Hop Sets Definition An (h, ε)-hop set is a set of weighted edges F such that, for all pairs of nodes u and v, in the shortcut graph G F there is a path from u to v with at most h edges of weight at most (1 + ε)dist(u, v). Fact: Every graph has a (n o(1), ε)-hop set of size n 1+o(1) [Cohen 94] (for ε 1/polylogn) 11 / 26

36 Less Known: Hop Sets Definition An (h, ε)-hop set is a set of weighted edges F such that, for all pairs of nodes u and v, in the shortcut graph G F there is a path from u to v with at most h edges of weight at most (1 + ε)dist(u, v). Application? SSSP up to h hops (Bellman-Ford) RAM: O(mh) time 11 / 26

37 Less Known: Hop Sets Definition An (h, ε)-hop set is a set of weighted edges F such that, for all pairs of nodes u and v, in the shortcut graph G F there is a path from u to v with at most h edges of weight at most (1 + ε)dist(u, v). Application? SSSP up to h hops (Bellman-Ford) RAM: O(mh) time PRAM: O(mh) with O(h) depth 11 / 26

38 Less Known: Hop Sets Definition An (h, ε)-hop set is a set of weighted edges F such that, for all pairs of nodes u and v, in the shortcut graph G F there is a path from u to v with at most h edges of weight at most (1 + ε)dist(u, v). Application? SSSP up to h hops (Bellman-Ford) RAM: O(mh) time PRAM: O(mh) with O(h) depth Congested Clique: O(h) rounds 11 / 26

39 Less Known: Hop Sets Definition An (h, ε)-hop set is a set of weighted edges F such that, for all pairs of nodes u and v, in the shortcut graph G F there is a path from u to v with at most h edges of weight at most (1 + ε)dist(u, v). Application? SSSP up to h hops (Bellman-Ford) RAM: O(mh) time PRAM: O(mh) with O(h) depth Congested Clique: O(h) rounds Streaming: h passes with O(n) space 11 / 26

40 Less Known: Hop Sets Definition An (h, ε)-hop set is a set of weighted edges F such that, for all pairs of nodes u and v, in the shortcut graph G F there is a path from u to v with at most h edges of weight at most (1 + ε)dist(u, v). Application? SSSP up to h hops (Bellman-Ford) RAM: O(mh) time PRAM: O(mh) with O(h) depth Congested Clique: O(h) rounds Streaming: h passes with O(n) space Incremental/Decremental O(mh) [Even/Shiloach 81, HKN 14] 11 / 26

41 Less Known: Hop Sets Definition An (h, ε)-hop set is a set of weighted edges F such that, for all pairs of nodes u and v, in the shortcut graph G F there is a path from u to v with at most h edges of weight at most (1 + ε)dist(u, v). Application? SSSP up to h hops (Bellman-Ford) RAM: O(mh) time PRAM: O(mh) with O(h) depth Congested Clique: O(h) rounds Streaming: h passes with O(n) space Incremental/Decremental O(mh) [Even/Shiloach 81, HKN 14] Hopset with h = n o(1) and size n 1+o(1) gives almost tight algorithms 11 / 26

42 Less Known: Hop Sets Definition An (h, ε)-hop set is a set of weighted edges F such that, for all pairs of nodes u and v, in the shortcut graph G F there is a path from u to v with at most h edges of weight at most (1 + ε)dist(u, v). Application? SSSP up to h hops (Bellman-Ford) RAM: O(mh) time PRAM: O(mh) with O(h) depth Congested Clique: O(h) rounds Streaming: h passes with O(n) space Incremental/Decremental O(mh) [Even/Shiloach 81, HKN 14] Hopset with h = n o(1) and size n 1+o(1) gives almost tight algorithms Remaining challenge: Compute hop set efficiently 11 / 26

43 Hop Sets: Approaching Optimality Authors Stretch α Hopbound h Size [Baseline] 1 1 O(n 2 ) 12 / 26

44 Hop Sets: Approaching Optimality Authors Stretch α Hopbound h Size [Baseline] 1 1 O(n 2 ) [Klein/Subramanian 97] 1 O( n log n t ) O(t 2 ) [Shi/Spencer 99] 1 O( n t ) O(nt) 12 / 26

45 Hop Sets: Approaching Optimality Authors Stretch α Hopbound h Size [Baseline] 1 1 O(n 2 ) [Klein/Subramanian 97] 1 O( n log n t ) O(t 2 ) [Shi/Spencer 99] 1 O( n t ) O(nt) [Thorup/Zwick 01] 2k 1 2 O(kn 1+ 1 k ) 12 / 26

46 Hop Sets: Approaching Optimality Authors Stretch α Hopbound h Size [Baseline] 1 1 O(n 2 ) [Klein/Subramanian 97] 1 O( n log n t ) O(t 2 ) [Shi/Spencer 99] 1 O( n t ) O(nt) [Thorup/Zwick 01] 2k 1 2 O(kn 1+ 1 k ) [Cohen 94] 1 + ε ( log n ε ) O(log k) O(n 1+ 1 k log n) [Bernstein 09] 1 + ε O( 3 ε )k log n O(kn 1+ 1 k ) [Elkin/Neiman 16] 1 + ε ( log k ε ) O(log k) O(n 1+ 1 k log n log k) [Elkin/Neiman 17] 1 + ε O( k+1 ε )k+1 O(n k+1 1 ) [Huang/Pettie 17] 1 + ε O( k ε )k O(n k+1 1 ) 12 / 26

47 Hop Sets: Approaching Optimality Authors Stretch α Hopbound h Size [Baseline] 1 1 O(n 2 ) [Klein/Subramanian 97] 1 O( n log n t ) O(t 2 ) [Shi/Spencer 99] 1 O( n t ) O(nt) [Thorup/Zwick 01] 2k 1 2 O(kn 1+ 1 k ) [Cohen 94] 1 + ε ( log n ε ) O(log k) O(n 1+ 1 k log n) [Bernstein 09] 1 + ε O( 3 ε )k log n O(kn 1+ 1 k ) [Elkin/Neiman 16] 1 + ε ( log k ε ) O(log k) O(n 1+ 1 k log n log k) [Elkin/Neiman 17] 1 + ε O( k+1 ε )k+1 O(n k+1 1 ) [Huang/Pettie 17] 1 + ε O( k ε )k O(n k+1 1 ) [Abboud/Bodwin/Pettie 16] 1 + ε Ω k ( 1 ε )k n k 1 δ 12 / 26

48 Hop Sets: Approaching Optimality Authors Stretch α Hopbound h Size [Baseline] 1 1 O(n 2 ) [Klein/Subramanian 97] 1 O( n log n t ) O(t 2 ) [Shi/Spencer 99] 1 O( n t ) O(nt) [Thorup/Zwick 01] 2k 1 2 O(kn 1+ 1 k ) [Cohen 94] 1 + ε ( log n ε ) O(log k) O(n 1+ 1 k log n) [Bernstein 09] 1 + ε O( 3 ε )k log n O(kn 1+ 1 k ) [Elkin/Neiman 16] 1 + ε ( log k ε ) O(log k) O(n 1+ 1 k log n log k) [Elkin/Neiman 17] 1 + ε O( k+1 ε )k+1 O(n k+1 1 ) [Huang/Pettie 17] 1 + ε O( k ε )k O(n k+1 1 ) [Abboud/Bodwin/Pettie 16] 1 + ε Ω k ( 1 ε )k n k 1 δ Cannot have α = 1 + ε, h = poly(1/ε) and size n polylog(n). 12 / 26

49 It was too good to be true / 26

50 Hop Set Example 14 / 26

51 Simple Hop Set Based on Balls (following [Thorup/Zwick 06]) V = A 0 A 1 A k = where node of A i goes to A i+1 with probability 1/n 1/k v has priority i if v A i \ A i+1 15 / 26

52 Simple Hop Set Based on Balls (following [Thorup/Zwick 06]) V = A 0 A 1 A k = where node of A i goes to A i+1 with probability 1/n 1/k v has priority i if v A i \ A i+1 For every node u of priority i: Ball(u) = {v V dist(u, v) < dist(u, A i+1 )} 15 / 26

53 Simple Hop Set Based on Balls (following [Thorup/Zwick 06]) V = A 0 A 1 A k = where node of A i goes to A i+1 with probability 1/n 1/k v has priority i if v A i \ A i+1 pr. i + 1 For every node u of priority i: Ball(u) = {v V dist(u, v) < dist(u, A i+1 )} pr. i 15 / 26

54 Simple Hop Set Based on Balls (following [Thorup/Zwick 06]) V = A 0 A 1 A k = where node of A i goes to A i+1 with probability 1/n 1/k v has priority i if v A i \ A i+1 pr. i + 1 For every node u of priority i: Ball(u) = {v V dist(u, v) < dist(u, A i+1 )} pr. i 15 / 26

55 Simple Hop Set Based on Balls (following [Thorup/Zwick 06]) V = A 0 A 1 A k = where node of A i goes to A i+1 with probability 1/n 1/k v has priority i if v A i \ A i+1 pr. i + 1 For every node u of priority i: Ball(u) = {v V dist(u, v) < dist(u, A i+1 )} pr. i Hop set: (u, v) F iff v Ball(u) w(u, v) = dist G (u, v) 15 / 26

56 Simple Hop Set Based on Balls (following [Thorup/Zwick 06]) V = A 0 A 1 A k = where node of A i goes to A i+1 with probability 1/n 1/k v has priority i if v A i \ A i+1 pr. i + 1 For every node u of priority i: Ball(u) = {v V dist(u, v) < dist(u, A i+1 )} pr. i priority # nodes 0 n 1 n 1 1/k.. k 1 n 1/k Hop set: (u, v) F iff v Ball(u) w(u, v) = dist G (u, v) 15 / 26

57 Simple Hop Set Based on Balls (following [Thorup/Zwick 06]) V = A 0 A 1 A k = where node of A i goes to A i+1 with probability 1/n 1/k v has priority i if v A i \ A i+1 For every node u of priority i: Ball(u) = {v V dist(u, v) < dist(u, A i+1 )} Expected size: n (i+1)/k priority # nodes Ball(u) 0 n n 1/k 1 n 1 1/k n 2/k... k 1 n 1/k n pr. i + 1 pr. i Hop set: (u, v) F iff v Ball(u) w(u, v) = dist G (u, v) 15 / 26

58 Simple Hop Set Based on Balls (following [Thorup/Zwick 06]) V = A 0 A 1 A k = where node of A i goes to A i+1 with probability 1/n 1/k v has priority i if v A i \ A i+1 For every node u of priority i: Ball(u) = {v V dist(u, v) < dist(u, A i+1 )} Expected size: n (i+1)/k priority # nodes Ball(u) # edges 0 n n 1/k n 1+1/k 1 n 1 1/k n 2/k n 1+1/k.... k 1 n 1/k n n 1+1/k pr. i + 1 pr. i Hop set: (u, v) F iff v Ball(u) w(u, v) = dist G (u, v) 15 / 26

59 Simple Hop Set Based on Balls (following [Thorup/Zwick 06]) V = A 0 A 1 A k = where node of A i goes to A i+1 with probability 1/n 1/k v has priority i if v A i \ A i+1 For every node u of priority i: Ball(u) = {v V dist(u, v) < dist(u, A i+1 )} Expected size: n (i+1)/k priority # nodes Ball(u) # edges 0 n n 1/k n 1+1/k 1 n 1 1/k n 2/k n 1+1/k.... k 1 n 1/k n n 1+1/k kn 1+1/k pr. i + 1 pr. i Hop set: (u, v) F iff v Ball(u) w(u, v) = dist G (u, v) 15 / 26

60 Parameter Choice k = log n log 4/ε ( ) k 4 = n 1/k ε 16 / 26

61 Parameter Choice k = log n log 4/ε ( ) k 4 = n 1/k = n o(1) ε 16 / 26

62 (n 1/2+o(1), ε)-hop set Case 1: dist(u 0, v) n 1/2+1/k /ε u 0 v 17 / 26

63 (n 1/2+o(1), ε)-hop set Case 2: dist(u 0, v) > n 1/2+1/k /ε u 0 v 17 / 26

64 (n 1/2+o(1), ε)-hop set Case 2: dist(u 0, v) > n 1/2+1/k /ε r 0 = n 1/2 u 0 v 0 v r 0 17 / 26

65 (n 1/2+o(1), ε)-hop set Case 2: dist(u 0, v) > n 1/2+1/k /ε r 0 = n 1/2 u 1 r 0 u 0 r 0 v 0 v For every node u of priority i and every node v, either (u, v) H, or u of priority i + 1 s. t. dist(u, u ) dist(u, v). 17 / 26

66 (n 1/2+o(1), ε)-hop set Case 2: dist(u 0, v) > n 1/2+1/k /ε u 2 r 1 r 0 = n 1/2 ( r i+1 = ) ε 0 j i r j r u 1 1 r 0 u 0 r 0 v 0 v 1 v For every node u of priority i and every node v, either (u, v) H, or u of priority i + 1 s. t. dist(u, u ) dist(u, v). 17 / 26

67 (n 1/2+o(1), ε)-hop set Case 2: dist(u 0, v) > n 1/2+1/k /ε u 2 r 1 r 2 r 0 = n 1/2 ( r i+1 = ) ε 0 j i r j r u 1 1 r 0 u 0 r 0 v 0 v 1 v 2 v For every node u of priority i and every node v, either (u, v) H, or u of priority i + 1 s. t. dist(u, u ) dist(u, v). 17 / 26

68 (n 1/2+o(1), ε)-hop set Case 2: dist(u 0, v) > n 1/2+1/k /ε increasing priority u 2 r 1 r 2 r u 1 1 r 0 u 0 r 0 v 0 v 1 v 2 v decreasing distance to v r 0 = n 1/2 ( r i+1 = ) ε 0 j i For every node u of priority i and every node v, either (u, v) H, or u of priority i + 1 s. t. dist(u, u ) dist(u, v). r j 17 / 26

69 (n 1/2+o(1), ε)-hop set Case 2: dist(u 0, v) > n 1/2+1/k /ε increasing priority u 2 r 1 r 2 r u 1 1 r 0 u 0 r 0 v 0 v 1 v 2 v decreasing distance to v r 0 = n 1/2 ( r i+1 = ) ε n 1/2 n 1/k 0 j i r j k = log n/ log 4/ε For every node u of priority i and every node v, either (u, v) H, or u of priority i + 1 s. t. dist(u, u ) dist(u, v). Weight (1 + ε)dist(u 0, v) 17 / 26

70 (n 1/2+o(1), ε)-hop set Case 2: dist(u 0, v) > n 1/2+1/k /ε increasing priority u 2 r 1 r 2 r u 1 1 r 0 u 0 r 0 v 0 v 1 v 2 v decreasing distance to v r 0 = n 1/2 ( r i+1 = ) ε n 1/2 n 1/k 0 j i r j k = log n/ log 4/ε For every node u of priority i and every node v, either (u, v) H, or u of priority i + 1 s. t. dist(u, u ) dist(u, v). Weight (1 + ε)dist(u 0, v) #Edges k dist(u, v) n 1/2 k n = kn1/2 n1/2 17 / 26

71 Chicken-Egg Problem? 1 Goal: Faster SSSP via hop set 2 Compute hop set by computing balls 3 Computing balls at least as hard as SSSP Back at problem we wanted to solve initially? 18 / 26

72 Chicken-Egg Problem? 1 Goal: Faster SSSP via hop set 2 Compute hop set by computing balls 3 Computing balls at least as hard as SSSP Back at problem we wanted to solve initially? No! (n 1/2+o(1), ε)-hop set only requires balls up to n 1/2+o(1) hops 18 / 26

73 (n 1/2+o(1), ε)-hop set Iterative computation In each iteration number of hops is reduced by a factor of n 1/k 19 / 26

74 (n 1/2+o(1), ε)-hop set Iterative computation In each iteration number of hops is reduced by a factor of n 1/k Algorithm: for i = 1 to k do H i = G 1 j i 1 F j Compute balls with k priorities in H i up to n 2/k hops F i = {(u, v) v Ball(u)} end return F = 1 i k F i 19 / 26

75 (n 1/2+o(1), ε)-hop set Iterative computation In each iteration number of hops is reduced by a factor of n 1/k Algorithm: for i = 1 to k do H i = G 1 j i 1 F j Compute balls with k priorities in H i up to n 2/k hops F i = {(u, v) v Ball(u)} end return F = 1 i k F i Error amplification: (1 + ε ) k (1 + ε) for ε = 1/(2ε log n) 19 / 26

76 (n 1/2+o(1), ε)-hop set Iterative computation In each iteration number of hops is reduced by a factor of n 1/k Algorithm: for i = 1 to k do H i = G 1 j i 1 F j Compute balls with k priorities in H i up to n 2/k hops F i = {(u, v) v Ball(u)} end return F = 1 i k F i Error amplification: (1 + ε ) k (1 + ε) for ε = 1/(2ε log n) Omitted detail: weighted graphs, use rounding technique 19 / 26

77 Beyond Hop Sets 20 / 26

78 New Distributed Algorithm Theorem ([Becker/Karrenbauer/K/Lenzen arxiv 16]) There is a deterministic algorithm for computing (1 + ε) approximate SSSP in ( n + Diam)poly(log n, ε) rounds. 21 / 26

79 New Distributed Algorithm Theorem ([Becker/Karrenbauer/K/Lenzen arxiv 16]) There is a deterministic algorithm for computing (1 + ε) approximate SSSP in ( n + Diam)poly(log n, ε) rounds. Key insight: Solve more general problem Shortest Transshipment Problem Find the cheapest route for sending units of a single good from sources to sinks along the edges of a graph as specified by demands on nodes. 21 / 26

80 New Distributed Algorithm Theorem ([Becker/Karrenbauer/K/Lenzen arxiv 16]) There is a deterministic algorithm for computing (1 + ε) approximate SSSP in ( n + Diam)poly(log n, ε) rounds. Key insight: Solve more general problem Shortest Transshipment Problem Find the cheapest route for sending units of a single good from sources to sinks along the edges of a graph as specified by demands on nodes. Uncapacitated minimum-cost flow 21 / 26

81 New Distributed Algorithm Theorem ([Becker/Karrenbauer/K/Lenzen arxiv 16]) There is a deterministic algorithm for computing (1 + ε) approximate SSSP in ( n + Diam)poly(log n, ε) rounds. Key insight: Solve more general problem Shortest Transshipment Problem Find the cheapest route for sending units of a single good from sources to sinks along the edges of a graph as specified by demands on nodes. Uncapacitated minimum-cost flow SSSP: source has demand (n 1), other nodes have demand 1 21 / 26

82 Shortest Transshipment Problem Shortest transshipment as linear program: minimize Wx 1 s.t. Ax = b 22 / 26

83 Shortest Transshipment Problem Shortest transshipment as linear program: minimize Wx 1 s.t. Ax = b Dual program: maximize b T y s.t. W 1 A T y 1 22 / 26

84 Shortest Transshipment Problem Shortest transshipment as linear program: minimize Wx 1 s.t. Ax = b Dual program: maximize b T y s.t. W 1 A T y 1 Equivalent: minimize W 1 A T y s.t. b T π = 1 22 / 26

85 Shortest Transshipment Problem Shortest transshipment as linear program: minimize Wx 1 s.t. Ax = b Dual program: maximize b T y s.t. W 1 A T y 1 Equivalent: minimize W 1 A T y s.t. b T π = 1 We approximate by soft-max: lse β (x) := 1 β ln i [d] ( e βx i + e βx i ) 22 / 26

86 Gradient Descent Algorithm at a glance: 1 Soft-max is differentiable apply gradient descent 23 / 26

87 Gradient Descent Algorithm at a glance: 1 Soft-max is differentiable apply gradient descent 2 Each iteration: solve transshipment problem with different demand vector b depending on current gradient 23 / 26

88 Gradient Descent Algorithm at a glance: 1 Soft-max is differentiable apply gradient descent 2 Each iteration: solve transshipment problem with different demand vector b depending on current gradient 3 Key observation: For b, bad approximation is sufficient 23 / 26

89 Gradient Descent Algorithm at a glance: 1 Soft-max is differentiable apply gradient descent 2 Each iteration: solve transshipment problem with different demand vector b depending on current gradient 3 Key observation: For b, bad approximation is sufficient 4 Compute spanner on overlay network and solving transshipment on overlay spanner Spanner has stretch O(log n) and size Õ(n) 23 / 26

90 Gradient Descent Algorithm at a glance: 1 Soft-max is differentiable apply gradient descent 2 Each iteration: solve transshipment problem with different demand vector b depending on current gradient 3 Key observation: For b, bad approximation is sufficient 4 Compute spanner on overlay network and solving transshipment on overlay spanner Spanner has stretch O(log n) and size Õ(n) 5 Overall: Polylog iterations, each solving O(log n)-approximate transshipment on graph of Õ(n) edges 23 / 26

91 Gradient Descent Algorithm at a glance: 1 Soft-max is differentiable apply gradient descent 2 Each iteration: solve transshipment problem with different demand vector b depending on current gradient Congested Clique: Compute gradient in O(1) rounds 3 Key observation: For b, bad approximation is sufficient 4 Compute spanner on overlay network and solving transshipment on overlay spanner Spanner has stretch O(log n) and size Õ(n) 5 Overall: Polylog iterations, each solving O(log n)-approximate transshipment on graph of Õ(n) edges 23 / 26

92 Gradient Descent Algorithm at a glance: 1 Soft-max is differentiable apply gradient descent 2 Each iteration: solve transshipment problem with different demand vector b depending on current gradient Congested Clique: Compute gradient in O(1) rounds 3 Key observation: For b, bad approximation is sufficient 4 Compute spanner on overlay network and solving transshipment on overlay spanner Spanner has stretch O(log n) and size Õ(n) Congested Clique: Spanner can be computed in O(log n) rounds [Baswana/Sen 03] 5 Overall: Polylog iterations, each solving O(log n)-approximate transshipment on graph of Õ(n) edges 23 / 26

93 Technical Details 1 Black-box reduction from SSSP to shortest transshipment only for exact solutions 24 / 26

94 Technical Details 1 Black-box reduction from SSSP to shortest transshipment only for exact solutions 2 Transshipment will only guarantee (1 + ε)-approximation on average 24 / 26

95 Technical Details 1 Black-box reduction from SSSP to shortest transshipment only for exact solutions 2 Transshipment will only guarantee (1 + ε)-approximation on average 3 How to obtain per-node guarantee: Solve with increased precision Inspect gradient to identify good nodes Repeat transshipment for bad nodes only Analysis: Total mass reduced by constant fraction in each run 24 / 26

96 Technical Details 1 Black-box reduction from SSSP to shortest transshipment only for exact solutions 2 Transshipment will only guarantee (1 + ε)-approximation on average 3 How to obtain per-node guarantee: Solve with increased precision Inspect gradient to identify good nodes Repeat transshipment for bad nodes only Analysis: Total mass reduced by constant fraction in each run Independent work: Approximate transshipment [Sherman 16] 24 / 26

97 Technical Details 1 Black-box reduction from SSSP to shortest transshipment only for exact solutions 2 Transshipment will only guarantee (1 + ε)-approximation on average 3 How to obtain per-node guarantee: Solve with increased precision Inspect gradient to identify good nodes Repeat transshipment for bad nodes only Analysis: Total mass reduced by constant fraction in each run Independent work: Approximate transshipment [Sherman 16] More general solvers based on generalized preconditioning 24 / 26

98 Technical Details 1 Black-box reduction from SSSP to shortest transshipment only for exact solutions 2 Transshipment will only guarantee (1 + ε)-approximation on average 3 How to obtain per-node guarantee: Solve with increased precision Inspect gradient to identify good nodes Repeat transshipment for bad nodes only Analysis: Total mass reduced by constant fraction in each run Independent work: Approximate transshipment [Sherman 16] More general solvers based on generalized preconditioning Linear preconditioner based on metric embeddings 24 / 26

99 Technical Details 1 Black-box reduction from SSSP to shortest transshipment only for exact solutions 2 Transshipment will only guarantee (1 + ε)-approximation on average 3 How to obtain per-node guarantee: Solve with increased precision Inspect gradient to identify good nodes Repeat transshipment for bad nodes only Analysis: Total mass reduced by constant fraction in each run Independent work: Approximate transshipment [Sherman 16] More general solvers based on generalized preconditioning Linear preconditioner based on metric embeddings With additional analysis: spanner-based oracle as non-linear preconditioner 24 / 26

100 Technical Details 1 Black-box reduction from SSSP to shortest transshipment only for exact solutions 2 Transshipment will only guarantee (1 + ε)-approximation on average 3 How to obtain per-node guarantee: Solve with increased precision Inspect gradient to identify good nodes Repeat transshipment for bad nodes only Analysis: Total mass reduced by constant fraction in each run Independent work: Approximate transshipment [Sherman 16] More general solvers based on generalized preconditioning Linear preconditioner based on metric embeddings With additional analysis: spanner-based oracle as non-linear preconditioner No straightforward way of obtaining per-node guarantee 24 / 26

101 Conclusion Main contributions: Two almost tight algorithms in distributed and streaming models Combinatorial and continuous tools 25 / 26

102 Conclusion Main contributions: Two almost tight algorithms in distributed and streaming models Combinatorial and continuous tools Open problems: PRAM: improve Cohen s m 1+o(1) work with polylog depth? Deterministic decremental SSSP algorithm 25 / 26

103 Tight and Tighter 26 / 26