Minimum cost spanning tree

Transcription

1 Minimum cost spanning tree Doctoral course Optimization on graphs - Lecture 2.2 Giovanni Righini January 15 th, 2013

2 Definitions - 1 A graph G = (V,E) is a tree if and only if it is connected and acyclic. Connectivity: for each cut, at least one edge belongs to the tree. Acyclicity: for each cycle, at least one edge does not belong to the tree. Given a G = (V,E) a subset F E is: a forest if it does not contain cycles; a connector if (V,F) is connected; a spanning tree if (V,F) is a tree; a maximal forest if there is no forest containing it.

3 Definitions - 2 A graph G = (V,E) has a spanning tree if and only if it is connected. Given a connected graph G = (V,E), F is a spanning tree if and only if: F is a maximal forest; F is a minimal connector; F is a forest with F = V 1 edges; F is a connector with F = V 1 edges.

4 Definitions - 3 Given a graph G = (V,E) with k connected components, every maximal forest in it has V k edges. It forms a spanning tree in each connected component of G. Every maximal forest is also a maximum cardinality forest. Analogously every connector contains a connector of minimum cardinality.

5 The Minimum Spanning Tree Problem Let G = (V,E) be a connected graph. Let c : E R a cost function. We define F E: c(f) := e F c e Problem (Minimum Spanning Tree Problem). Find a spanning tree of minimum cost in G.

6 Properties Definition. F is a good forest if it belongs to a minimum cost spanning tree. Theorem. Given a good forest F and given an edge e F, F {e} is a good forest if and only if there is a cut C disjoint from F such that e is the edge with minimum cost in C.

7 Proof - 1 Necessity. Let T be a minimum cost spanning tree containing F {e}. Let C be the (unique) cut disjoint from T \{e}. Consider any edge f C. The edge set T = T \{e} {f} is a spanning tree. Since T is optimal, c(t ) c(t ) and then c e c f. Therefore e is an edge of minimum cost in C. C e f F e T*

8 Proof - 2 Sufficiency. Let T be a minimum cost spanning tree containing F. Let P the (unique) path in T between the two endpoints of edge e. P must contain at least one edge f C, where C is a cut disjoint from F. Hence T = T \{f} {e} is also a spanning tree. Since c e c f we have c(t ) c(t ). Then T is also a minimum cost spanning tree. Since F {e} is contained in T, it is a good forest. P e C f F T*

9 A mathematical programming model (A) min z = e E c e x e s.t. e δ(s) x e {0, 1} x e 1 S V (1) e E A cut δ(s) is the subset of edges with one endpoint in S. Constraints 1 impose connectivity. Acyclicity comes for free from cost minimization. Integrality conditions are redundant.

10 The dual model (A) min z = e E c e x e s.t. x e 1 e δ(s) x e 0 S V e E This linear program has a dual. max w = S V y S s.t. y S c e S V:e δ(s) y S 0 e E S V

11 Complementary slackness conditions (A) Primal C.S.C.: x e (c e S V:e δ(s) y S) = 0 e E Dual C.S.C.: y S ( e δ(s) x e 1) = 0 S V The initial primal solution x e = 0 e E is primal infeasible (and super-optimal): all connectivity constraints are violated. The corresponding dual solution y S = 0 S V is dual feasible (and sub-optimal).

12 Another mathematical programming model (B) min z = e E c e x e s.t. e E x e = n 1 (2) e E(S) x e {0, 1} x e S 1 S V (3) e E E(S) is the edge subset of the subgraph induced by S. Constraints 3 impose acyclicity instead of connectivity. Hence we also need constraint 2 to impose connectivity. Integrality conditions are redundant.

13 The dual model (B) min z = e E c e x e s.t. e E x e = n 1 x e ( S 1) e E(S) x e 0 This linear program has a dual: S V e E max w = S V( S 1)y S +(n 1)y V s.t. y S + y V c e S V:e E(S) y S 0 y V free e E S V

14 Complementary slackness conditions (B) Primal C.S.C.: x e (c e + y S y V ) = 0 e E. S V:e E(S) Dual C.S.C.: y V ( e E x e (n 1)) = 0. y S ( S 1 e E(S) x e ) = 0 S V. The initial primal solution x e = 0 e E is primal infeasible (and super-optimal): the cardinality constraint is violated. The initial dual solution y S = y V = 0 S V is dual feasible (and sub-optimal).

15 Algorithms All MSTP algorithms exploit the previous theorem. The idea is to start with an empty (good) forest F and to extend it iteratively with an edge satisfying the theorem requirement, i.e. an edge of minimum cost in a cut C disjoint from F. Different algorithms are obtained by different choices of C. The two most common algorithms are Jarnik (1930), Kruskal (1956), Prim (1957), Dijkstra (1959): C is the cut that separates the connected component including a predefined vertex; Kruskal (1956), Loberman e Weinberger (1957), Prim (1957): C is the cut that separates the two connected components including the endpoints of the minimum cost edge e in a sorted list. Both algorithms can be seen as dual ascent algorithms.

16 Prim algorithm (1957) begin T := ; z:=0; for v:=1 to n do flag[v]:=0; flag[r]:=1; for v:=1 to n do cost[v]:=c[r, v]; pred[v]:=r; for k:=1 to n 1 do mincost:= ; for v:=1 to n do if (flag[v] = 0) and (cost[v] < mincost) then v:=v; mincost:=cost[v]; T :=T {[pred[v], v]}; z:=z + mincost; flag[v] := 1; for v:=1 to n do if (flag[v] = 0) and (c[v, v] < cost[v]) then pred[v] := v; cost[v]:=c[v, v]; end The complexity is O(n 2 ). With 2-heaps it is possible to obtain O(m + log n). With Fibonacci heaps it is possible to obtain O(m+n log n).

17 Kruskal algorithm (1956) begin T := ; z:=0; E:=E; Sort E by non-decreasing edge costs; for v:=1 to n do List[v]:={v}; head[v]:=v; card[v]:=1; while T < n 1 do [u, v]:=argmin e E {c e }; E:=E\{[u, v]}; if (head[u] head[v]) then T :=T {[u, v]}; z:=z + c[u, v]; if (card[v] > card[u]) then Swap (u, v); L[u] := L[u] L[v]; card[u] := card[u]+card[v]; for i L[v] do head[i] := head[u]; end

18 Kruskal algorithm (1956) Sorting the edges requires O(m log n). After sorting the complexity is O(m + n log n) and it can be obtained with linked lists. A list L is associated with each component of the current forest. For each vertex v V, let head(v) be the head of the list L v which v belongs to. Initially head(v):=v and L v :={v} for all vertices. At each iteration: it is necessary to test whether the next edge e = [u, v] in the list would close a cycle or not; if not, the data-structure must be updated.

19 Kruskal algorithm complexity The test is: head(u) = head(v)? It is executed in constant time at most m times. Hence it requires O(m) overall. When extending the current forest with e = [u, v], the shortest list among L u and L v (detectable in O(1) with an items counter associated with each list) is appended to the longest one (in O(1)). The head of the list is updated: head(u ) := head(v) for every u L u in O(n). No vertex can belong to the shortest list more than log n times, because the size of the shortest list at least doubles every time. Therefore the update operation requires at most O(log n) for each node, i.e. O(n log n) overall.

20 Boruvka algorithm (1926) It requires that all edge costs are different from each other and it allows for a parallel implementation. F:= ; while ( F < n 1) do For each component K of F Select e := argmin f δ(k) {c f }; F := F {e}; F remains a good forest after each iteration. Let e 1, e 2,...,e k the edges inserted into F with c e1 < c e2 <... < c ek. For each i = 1,...,k, e i is the minimum cost edge leaving F {e 1, e 2,...,e i 1 }, because none of {e 1, e 2,...,e i 1 } leaves component K i. Then F remains a good forest as if the edges would have been inserted sequentially.

21 Dual greedy algorithm It deletes edges from a connector instead of inserting them into a forest. Definition. A connector is good if it contains at least one minimum cost spanning tree. Theorem. Given a good connector K and an edge e K, K\{e} is a good connector if and only if K contains a cycle C such that e is the maximum cost edge in C. Kruskal (1956): sort the edges and starting from the connector K = E, iteratively delete the maximum cost edge which is not a bridge (i.e. without disconnecting the graph).

22 More algorithms Dijkstra algorithm (1960): Arbitrarily sort the edges and iteratively insert them into a forest (initially empty). When an edge which forms a cycle C is inserted, delete the maximum cost edge in C. Kalaba algorithm (1960): The same, but starting from an arbitrary spanning tree.

23 Dual ascent A dual ascent algorithm starts from a pair (x, y) of primal infeasible and dual feasible solutions and it iteratively does the following: Dual iteration: a violated primal constraint is selected; the corresponding dual variable y S is increased as much as possible, until a dual constraint becomes active; it corresponds to a primal variable x e ; Primal iteration: the corresponding primal variable x e enters the basis (is set to 1), repairing the infeasibility of the selected primal constraint. These two steps are repeated until primal feasibility is achieved. A primal solution and a dual solution are kept during the algorithm and they correspond to one another through complementary slackness conditions.

24 Model A: dual ascent A dual ascent algorithm based on model A iteratively does the following: Dual iteration: a cut (S, S) for which e δ(s) x e = 0, is selected; the corresponding dual variable y S is increased as much as possible, to activate a dual constraint corresponding to a primal variable x e, i.e. to an edge; Primal iteration: the corresponding primal variable x e is set to 1, repairing the infeasibility of the selected connectivity constraint. Dual feasibility requires S V:e δ(s) y S c e e E. Therefore for each node subset S, the maximum value that the corresponding dual variable y S can take is equal to the minimum cost c e among all the edges in the cut δ(s). If the dual iteration selects y S where S has maximum cardinality, Prim algorithm is obtained.

25 Model B: dual ascent A dual ascent algorithm based on model B iteratively does the following: Dual iteration: the only violated primal constraint is e E x e = n 1; acyclicity constraints are already satisfied, when we start from x = 0, and they always remain satisfied; the corresponding dual variable y V is increased, until a dual constraint becomes active; it corresponds to a primal variable x e ; Primal iteration: the primal variable x e enters the basis in order to decrease the infeasibility of the cardinality constraint; the maximum value x e can take is 1, because of the acyclicity constraint with S = 2; this makes some acyclicity constraints active, hence allowing for some dual variables to enter the basis; this in turn allows to increase y V further on. The primal and the dual solutions always correspond to each other through C.S.C., with the exception of y V ( e E x e (n 1)) = 0, which corresponds to a violated equality constraint in the primal problem. This is equivalent to Kruskal s algorithm.

26 Model B (Kruskal algorithm): an example k = 0. x = 0. z = y V = 0. w = 0. Dual C.S.C.: y V (k (n 1)) = 0(0 5) = 0.

27 An example: dual iteration 1 29(19) 13(3) (21) 25(15) 17(7) 22(12) 20(10) 35(25) 10(0) k = 0. x = 0. z = y V = 10. w = 10 5 = 50. Dual C.S.C.: y V (k (n 1)) = 10(0 5) = 50.

28 An example: primal iteration 1 29(19) 13(3) (21) 25(15) 17(7) 22(12) 20(10) 35(25) 10(0) y 56 = 0 5, k = 1. x 56 = 1. z = y V = 10. w = = 50. Dual C.S.C.: y V (k (n 1)) = 10(1 5) = 40.

29 An example: dual iteration 2 29(16) 13(0) (18) 25(12) 17(4) 22(9) 20(7) 35(22) 10(0) y 56 = 3 5, k = 1. x 56 = 1. z = y V = 13. w = = 62. Dual C.S.C.: y V (k (n 1)) = 13(1 5) = 52.

30 An example: primal iteration 2 29(16) 13(0) (18) 25(12) 17(4) 22(9) 20(7) 35(22) 10(0) y 23 = 0 2,3 y 56 = 3 5, k = 2. x 56 = x 23 = 1. z = y V = 13. w = = 62. Dual C.S.C.: y V (k (n 1)) = 13(2 5) = 39.

31 An example: dual iteration 3 29(12) 13(0) (8) 31(14) 17(0) 22(5) 20(3) 35(18) 10(0) y 23 = 4 2,3 y 56 = 7 5, k = 2. x 56 = x 23 = 1. z = y V = 17. w = = 74. Dual C.S.C.: y V (k (n 1)) = 17(2 5) = 51.

32 An observation 29(12) 13(0) (8) 31(14) 17(0) 22(5) 35(18) 10(0) k = 3. x 56 = x 23 = x 25 = 1. z = (3) Several primal constraints are now active: several dual variables can enter the basis: {2, 5} : x 25 = 1 {2, 3, 5} : x 23 + x 25 = 2 {2, 5, 6} : x 25 + x 56 = 2 {2, 3, 5, 6} : x 23 + x 25 + x 56 = 3 Which dual variable should we choose to enter the basis?

33 An observation Active dual constraints: [2, 5] : y 25 + y y y c 25 = y V [2, 3] : y 23 + y y c 23 = y V [5, 6] : y 56 + y y c 56 = y V For each unit increase of y V (providing value 5 in the dual objective), we can keep the dual constraints active with a unit increase of: y 56 and y 23 and y 25 : the cost is 1+1+1=3; y 56 and y 235 : the cost is 1+2=3; y 23 and y 256 : the cost is 1+2=3; y 2356 : the cost is 3. From the viewpoint of the dual active constraints and the dual objective function, all these possibilities are equivalent.

34 An observation But from the viewpoint of non-active dual constraints, they are not: if we choose y 2356 to enter the basis, no dual constraint corresponding to any other edge in {2, 3, 5, 6} will become active. Edges reduced costs are: c e = c e + S V:e E(S) y S y V. Reduced costs of all edges covered by basic dual variables do not decrease anymore. The dual variable y 2356 dominates the others; its corresponding primal constraint e E({2,3,5,6}) x e 3 dominates those corresponding to the others: i.e., if we do not select more edges in S = {2, 3, 5, 6}, we cannot select more edges in any subset of S.

35 An example: primal iteration 3 29(12) 13(0) (8) 31(14) 17(0) 22(5) 35(18) 10(0) (3) y 2356 = ,3,5,6 17 y 23 = 4 2,3 y 56 = 7 5, k = 3. x 56 = x 23 = x 25 = 1. z = y V = 17. w = = 74. Dual C.S.C.: y V (k (n 1)) = 17(3 5) = 34.

36 An example: dual iteration 4 29(4) 13(0) (0) 31(6) 17(0) 22(5) 35(10) 10(0) (3) y 2356 = ,3,5,6 17 y 23 = 4 2,3 y 56 = 7 5, k = 3. x 56 = x 23 = x 25 = 1. z = y V = 25. w = = 90. Dual C.S.C.: y V (k (n 1)) = 25(3 5) = 50.

37 An example: primal iteration 4 29(4) 13(0) (0) 31(6) 17(0) 22(5) 35(10) 10(0) (3) y 2356 = ,3,5,6 17 y 23 = 4 2,3 y 56 = 7 5,6 y 14 = 0 1, k = 4. x 56 = x 23 = x 25 = x 14 = 1. z = y V = 25. w = = 90. Dual C.S.C.: y V (k (n 1)) = 25(4 5) = 25.

38 An example: dual iteration 5 29(0) 13(0) (0) 31(2) 17(0) 22(5) 35(6) 10(0) (3) y 2356 = ,3,5,6 17 y 23 = 4 2,3 y 56 = 7 5,6 y 14 = 4 1, k = 4. x 56 = x 23 = x 25 = x 14 = 1. z = y V = 29. w = = 94. Dual C.S.C.: y V (k (n 1)) = 29(4 5) = 29.

39 An example: primal iteration 5 y V = 29 1,2,3,4,5,6 29(0) 13(0) (0) 31(2) 17(0) 22(5) 35(6) 10(0) (3) y 2356 = ,3,5,6 17 y 23 = 4 2,3 y 56 = 7 5,6 29 y 14 = 4 1, k = 5(feasible!). x 23 = x 56 = x 25 = x 14 = x 12 = 1. z = y V = 29. w = = 94. Dual C.S.C.: y V (k (n 1)) = 29(5 5) = 0.

40 Dual feasibility requires S V:e E(S) y S + c e y V e E. For each node subset S, the value that the corresponding dual variable y S takes is equal to the difference between the minimum edge cost in the cut δ(s) and the maximum edge cost in the minimum spanning tree of the induced subgraph E(S). Dual optimal solution y 2356 = 12 y V = ,3,5, y 23 = 4 2,3 y 56 = 7 5, ,2,3,4,5,6 29 y 14 = 4 1, y V = 29. w = = 94. Dual C.S.C.: y V (k (n 1)) = 29(5 5) = 0.

41 29(0) 13(0) (0) 31(2) 17(0) 35(6) 10(0) Primal optimal solution 22(5) 20(3) The reduced cost of each edge is the difference between its cost and the largest cost along the (unique) path between its endpoints in the spanning tree. Edges with positive reduced cost do not belong to the minimum cost spanning tree, because there is a cycle in which they have the largest cost. k = 5. x 23 = x 56 = x 25 = x 14 = x 12 = 1. z = 94. Owing to the dual ascent procedure (monotonic increase of y V ), edges are chosen in non-decreasing order according to their costs.