1. Distance sensitivity oracles in weighted directed graphs Raphael Yuster University of Haifa Joint work with Oren Weimann Weizmann inst.

2. 2 Distance Sensitivity Oracles A data structure that reports shortest paths when edges of a (possibly directed, possibly weighted) graph G = (V, E) fail. A query (u,v  V, F  E) to the oracle returns the length of the shortest u-to-v path in the graph G’ = (V, E-F). When measuring distance oracles we are mostly concerned with the tradeoff between: Construction time / space vs.Query time Related variant: Replacement paths. Here we are given u,v and a shortest path p between them in advance, and need to compute all “replacement paths’’ resulting from single edge failures on p.

3. 3 Previous research NameVariant Construction Query timespace Demetrescu et al. [SICOMP – 2008] |F|=1O(n4)O(n4)n 2 log nO(1) Bernstein & Carger [STOC – 2009] |F|=1O(n3)O(n3)n 2 log nO(1) Duan & Pettie [SODA – 2009] |F|=2polynomialn 2 log 3 nlog n Y. & Weimann [FOCS – 2010] |F|=k|F|=ksub-cubicn2n2 sub- quadratic Roditty & Zwick [ICALP – 2005] replacement, unweighted n 2.5 Y. & Weimann [FOCS – 2010] replacement n 2.58 Vassilevska [SODA-2011] replacement n 2.38

4. 4 For any 0<α<1, w.h.p. a distance sensitivity oracle for k edge failures is constructed in O(n 2 ) space, O(n 1+ω-α ) < O(n 3.38-α ) time, and a query (u,v,F) is answered in O(n 2-(1-α)/k ) time. We assume that the edges have integer weights. Here ω < 2.38 is the “matrix multiplication exponent”. For simplicity, we will outline the proof of the case k=1: Main result For any 0 < α < 1, w.h.p. a distance sensitivity oracle for avoiding a single edge failure is constructed in O(n 2 ) space, O(n 3.38-α ) time, and a query (u,v,e) is answered in O(n 1+α ) time.

5. 5 For each possible query (u,v,e), let p(u,v,e) denote an (unknown) shortest path from u to v in G-e. We construct a set of random subgraphs R={G 1,…,G r } and prove that R has the following properties: For every edge e, the subset R e  R of subgraphs missing e is not too small and not too large. Every small-enough interval of any p(u,v,e) survives in at least one subgraph from R e. Choose at random a small-enough subset of vertices B. Every small-enough interval of any p(u,v,e) is hit by B. We compute the “all pairs” B*B distance matrix in every G i. Call it D i. Rough outline

6. 6 The preprocessing consists of R={G 1,…,G r } and of {D 1,…,D r } as well as a “reduced length function”. Given a query Q = (u,v,e): We identify the set of subgraphs R e. For every G i  R e we extend D i from an all-pairs distance matrix of B to an all pairs matrix of B  {u,v}. We construct a weighted complete graph G Q : Its vertex set is B  {u,v}. The weight of (x,y) is the length of the shortest x to y path in all the G i  R e. The answer to Q is the distance from u to v in G Q. It can be efficiently computed using reduced lengths. Rough outline – cont.

7. 7 The set of random subgraphs R={G 1,…,G r } is generated as follows: r = 42n 1-α log n. An edge is not taken to G i with probability n α-1 (independently for each edge and for each G i ). The random subgraphs Lemma 1 The probability that 21 log n < |R e | < 70 log n for all e  E is 1-o(1).

8. 8 Recall: p(u,v,e) denotes a shortest path from u to v in G-e. A short interval is a subpath with n 1-α vertices of p(u,v,e). Observe: there are only O(n 4 ) possible short intervals. We say that a short interval of p(u,v,e) survives if all its edges appear in some random subgraph of R e. The random subgraphs Lemma 2 The probability that all short intervals survive is 1-o(1). Proof: Consider short interval I of p(u,v,e) and some G i  R e. Survival probability: Non-survival in all G i  R e : By Lemma 1:

9. 9 Let B be a random subset of 5n α log n vertices. Hitting sets for shortest paths Lemma 3 With probability 1-o(1) every short interval is “hit” by B. We need to compute, for each G i  R, the distance between any pair of vertices of B (in fact, we care only about relevant pairs: distances realized by paths using at most n 1-α vertices). 1 2 34 5 6 0 9 8 7 B={1,3,5,8} n 1-α =3 02∞∞ ∞02∞ ∞∞02 21∞0 13581358 1 3 5 8

10. 10 The following is a corollary of a result of [Y. & Zwick, FOCS’05] : Lemma 4 Let G be a directed graph with integer edge weights, and let B be a subset of vertices of size O(n α ). A square matrix D indexed by B where D[x,y] is the distance from x to y whenever (x,y) is a relevant pair can be computed in matrix multiplication time O(n ω ) < O(n 2.38 ).

11. 11 Summary of preprocessing stage Corollary 5 Distance matrices D i for all G i  R can be computed in O(|R|n ω )=O(n ω+1-α ). O(n ω+1-α ) is also the time of the preprocessing stage. Space: The matrices D i together require |B| 2 |R|= O(n 1+α ) space, the random subgraphs require (whp) O(|R|n 1+α ) space (just store the edges that don’t appear in it). Total: O(n 2 ) space. “reduced length function”: a trick to handle negative weights reducing to the nonnegative edge-weight case. Can also be done in matrix multiplication time.

12. 12 Query stage Upon query Q=(u,v,e): Check each random subgraph whether belongs to R e. This takes only O(log n) time per subgraph. F or each G i  R e, (and recall that |R e | < 70 log n whp) extend D i from a {B}*{B} distance matrix to a {B [ u}* {B [ v} matrix (just add one row and one column recording the shortest path from u to vertices of B and from B to vertices of v). This can be done in O(n|B|)= O(n 1+α ) time. Construct a weighted complete graph G Q : Its vertex set is B  {u,v}. The weight of (x,y) is the smallest of the entries D i [x,y] ranging over all G i  R e. G Q is constructed in O(|B| 2 |R e |)= O(n 2α ) time.

13. 13 Finalizing the query Lemma 6 The distance from u to v in G-e is their distance in G Q. This distance is found in O(n 2α ) time by Dijkstra’s algorithm. Proof: Consider p(u,v,e) (a shortest u-v path in G-e). uv = member of B survives 1 11 1 1 11 1 1 1 1 11 1 (assume unit lengths) Between any two consecutive vertices of B on p(u,v,e) there is at most a short interval, so each interval survives in some G i  R e. Therefore, in G Q we have uv 2 32 43

14. 14 Thanks!