1. The Shortest-Path problem in graphs with selfish-edges

2. The selfish-edges SP problem Given: an undirected graph G=(V,E) such that each edge is owned by a distinct selfish agent, a source node s and a destination node z; we assume that agent’s private type is the positive cost (length) of the edge, and his valuation function is equal to his negated type if edge is selected in the solution, and 0 otherwise. Question: design a truthful mechanism in order to find a shortest path in G t =(V,E,t) between s and z.

3. Notation and assumptions n=|V|, m=|E| d G (s,z): distance in G=(V,E,r) between s ans z (sum of reported costs of edges on a shortest path P G (s,z)) Nodes s and z are 2-edge-connected in G, i.e., there exists in G at least 2 edge disjoint paths between s and z  for any edge of P G (s,z) removed from the graph there exists at least one replacement path in G-e between s and z (this will bound the problem, since otherwise a bridge-edge might have an unbounded marginal utility)

4. VCG mechanism The problem is utilitarian (indeed, the (negated) cost of a solution is given by the sum of valuations)  VCG-mechanism M= : g(r): computes arg max y  X  i v i (r i,y), i.e., P G (s,z) in G=(V,E,r); indeed, valuation functions are negative, and so maximizing their sum means to compute a shortest path; p e : (Clarke payments) for each e  E: p e =-  j≠e v j (r j,x(r - e )) +  j≠e v j (r j,x), namely d G-e (s,z)-[d G (s,z)-r e ] = d G-e (s,z)-d G (s,z) + r e if e  P G (s,z) d G (s,z)-d G (s,z) = 0 otherwise  For each e  P G (s,z), we have to compute d G-e (s,z), namely the length of a shortest path in G-e =(V,E\{e},r -e ) between s and z. pe=pe= {

5. The best replacement path s z e 2 2 3 4 5 6 5 10 5 12 P G-e (s,z)  d G-e (s,z)=12 P G (s,z)  d G (s,z)=11 Remark: u e =p e +v e =p e -r e ≥0 since d G-e (s,z)  d G (s,z) and so p e  r e p e =d G-e (s,z)-d G (s,z) + r e = 12-11+2=3

6. A trivial but costly implementation Step 1: First of all, apply Dijkstra to compute P G (s,z)  this costs O(m + n log n) time by using Fibonacci heaps. Step 2: Then,  e  P G (s,z) apply Dijkstra in G-e to compute P G-e (s,z)  we spend O(m + n log n) time for each of the O(n) edges in P G (s,z), i.e., O(mn + n 2 logn) time  Overall complexity: O(mn + n 2 log n) time We will see an efficient solution costing O(m + n log n) time

7. Notation S G (s), S G (z): single-source shortest paths trees rooted at s and z M s (e): set of nodes in S G (s) not descending from edge e (i.e., the set of nodes whose shortest path from s does not use e) N s (e)=V/M s (e) M z (e), N z (e) defined analogously

8. A picture s u v z e Ms(e)Ms(e) Ns(e)Ns(e) S G (s)

9. Crossing edges (M s (e),N s (e)) is a cut in G C s (e)={(x,y)  E\{e}: x  M s (e), y  N s (e)} edges “belonging” to the cut: crossing edges

10. Crossing edges s u v z e Ms(e)Ms(e) Ns(e)Ns(e) S G (s) Cs(e)Cs(e)

11. What about P G-e (s,z)? Trivial: it does not use e It must cross the cut C s (e) It is shortest among all the paths between s and z not using e Its length is: where w(f) denotes the cost returned for f. d G-e (s,z)= min {d G-e (s,x)+w(f)+d G-e (y,z)} f=(x,y)  C s (e)

12. The path P G-e (s,z) s u v z e x y d G-e (s,z)= min {d G-e (s,x)+w(f)+d G-e (y,z)} f=(x,y)  C s (e)

13. How to compute d G-e (s,z) Let f=(x,y)  C s (e); we will show that d G-e (s,x)+w(f)+d G-e (y,z)=d G (s,x)+w(f)+d G (y,z) Remark: d G-e (s,x)=d G (s,x), since x  M s (e) Lemma: Let f=(x,y)  C s (e) be a crossing edge (x  M s (e)). Then y  M z (e) (from which it follows that d G-e (y,z)=d G (y,z)).

14. A simple lemma Proof (by contr.) Let y  M z (e), then y  N z (e). Hence, y is a descendant of u in S G (z), and P G (z,y) uses e. But P G (v,y) is a subpath of P G (z,y), and then: d G (v,y)=w(e) + d G (u,y) > d G (u,y). But y  N s (e), then P G (s,y) uses e. But P G (u,y) is a subpath of P G (s,y), and then: d G (u,y)=w(e) + d G (v,y) > d G (v,y).

15. A picture s z N s (e)  M z (e) Ms(e)Ms(e)

16. Computing the length of replacement paths Given S G (s) and S G (z), in O(1) time we can compute the length of a shortest path between s and z passing through f and avoiding e as follows: k(f):= d G-e (s,x) + w(f) + d G-e (y,z) d G (s,x) given by S G (s) d G (y,z) given by S G (z)

17. A corresponding algorithm Step 1: Compute S G (s) and S G (z) Step 2:  e  P G (s,z) check all the crossing edges in C s (e), and take the minimum w.r.t. the key k. Time complexity Step 1: O(m + n log n) time Step 2: O(m) crossing edges for each of the O(n) edges on P G (s,z): since in O(1) we can establish whether an edge of G is currently a crossing edge (can you guess how?), Step 2 costs O(mn) time  Overall complexity: O(mn) time Improves on O(mn + n 2 log n) if m=o(n log n)

18. A more efficient solution: the Malik, Mittal and Gupta algorithm (1989) MMG have solved in O(m + n log n) time the following related problem: given a SP P G (s,z), what is its most vital edge, namely an edge whose removal induces the worst (i.e., longest) best replacement path between s and z? Their approach computes efficiently all the best replacement paths between s and z… …but this is exactly what we are looking for in our VCG-mechanism!

19. The MMG algorithm at work The basic idea of the algorithm is that when an edge e on P G (s,z) is considered, then we have a priority queue H containing the set of nodes in N s (e); with each node y  H, it remains associated a key k(y) and a corresponding crossing edge, defined as follows: k(y)= min {d G (s,x)+w(x,y)+d G (y,z)}  k(y) is the length of a SP in G-e from s to z passing through the node y, and so the minimum key is associated with a best replacement shortest-path for e (x,y)  E, x  M s (e i )

20. The MMG algorithm at work (2) Initially, H =V, and k(y)=+  for each y  V Consider e 1, e 2,…, e q one after the other. When edge e i is considered, modify H as follows: Remove from H all the nodes in W s (e i )=N s (e i-1 )\N s (e i ) (for i=1, set N s (e i-1 )=V) Consider all the edges (x,y) s.t. x  W s (e i ) and y  H, and compute k’(y)=d G (s,x)+w(x,y)+d G (y,z). If k’(y)<k(y), decrease k(y) to k’(y), and update the corresponding crossing edge to (x,y) Then, find the minimum in H w.r.t. k, which returns the length of a best replacement path for e i (i.e., d G-e i (s,z)), along with the selected crossing edge

21. An example N s (e 1 ) e1e1 e2e2 e3e3 e5e5 e4e4 s z W s (e 1 )

22. An example (2) N s (e 2 ) e1e1 e2e2 e3e3 e5e5 e4e4 s z W s (e 2 )

23. Time complexity of MMG Theorem: Given a shortest path between two nodes s and z in a graph G with n vertices and m edges, all the best replacement paths between s and z can be computed in O(m + n log n) time.

24. Time complexity of MMG Proof: Compute S G (s) and S G (z) in O(m + n log n) time. Then, use a Fibonacci heap to maintain H (observe that W s (e i ) can be computed in O(|W s (e i )|) time), on which the following operations are executed: A single make_heap n insert q=O(n) find_min O(n) delete O(m) decrease_key In a Fibonacci heap, the amortized cost of a delete is O(log n), the amortized cost of a decrease_key is O(1), while insert, find_min, and make_heap cost O(1), so O(m + n log n) total time

25. Plugging-in the MMG algorithm into the VCG-mechanism Corollary There exists a VCG-mechanism for the selfish- edges SP problem running in O(m + n log n) time. Proof. Running time for g: O(m + n log n) (Dijkstra). Running time of p: O(m + n log n), by applying MMG to compute all the distances d G-e (s,z).

26. This is the end My only friend, the end… (The Doors, 1967) Thanks for your attention!