1 / 43

Degree bounded Network Design

Degree bounded Network Design. Nikhil Bansal (IBM) Rohit Khandekar (IBM) Viswanath Nagarajan (CMU). Degree Bounded Network Design. Network design: Given a graph G, find min cost spanning tree 2-edge connected subgraph Steiner Tree on a subset of terminals T …

kevlyn
Download Presentation

Degree bounded Network Design

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Degree bounded Network Design Nikhil Bansal (IBM) Rohit Khandekar (IBM) Viswanath Nagarajan (CMU)

  2. Degree Bounded Network Design Network design: Given a graph G, find min cost spanning tree 2-edge connected subgraph Steiner Tree on a subset of terminals T … Degree Bounds: No degree more than 3 Arbitrary degree bounds bv Applications: Models workload on nodes Networking, VLSI, vehicle routing, …

  3. 10 10 20 10 10 10 10 trees Additional 10 trees Motivating Problem Input: Rooted directed graph with out-bandwidth bv on vertices Problem: At what max. rate can root broadcast? Problem: Pack max. number of arborescences T1,…,Tk rooted at r, s.t. for each v, i outdeg (v, Ti) · bv 30 r 20 10 10 10 (numberof arborescences = rate of traffic broadcast from r)

  4. Equivalent Formulation More generally:Edge Disjoint Version K-arborescence problem: Given k, can we pack k edge-disjoint arborescences subject to degree bounds? K=1: Can we find one arborescence or not? NP-Hard: if all bv = 1, yes iff a Hamiltonian Path.

  5. Previous Work Undirected Case (k=1): Spanning Tree w/ degree bounds bv Just minimize the maximum degree B* NP-Hard: b* =2 iff Hamiltonian Path +1 (local search) [Furer Raghavachari 94] What if edge costs? O(b* + log n) degree, O(1) cost [ Konemann Ravi 00] O(bv + log n) , O(1) cost [Konemann Ravi 03] (1+) b* + 2 log n/, 1 cost [Chaudhuri Rao Riesenfeld Talwar 05]

  6. Undirected Case (k=1) bv+2, Opt cost [Goemans 06] bv+1, Opt cost [Singh, Lau 07] Very simple algorithm (iterative relaxation) Singh & Lau give very easy proof of +2, but +1 proof not so simple. We give a very simple analysis of +1, Easily generalizes to crossing spanning tree, matroids. Implies (bv+1, Opt) even when packing k MST’s.

  7. General Connectivity Steiner Network:rijedge disjoint paths between i & j (Spanning tree: rij=1 for every vertex pair i and j.) Design min cost Steiner network: 2 approximation [Jain 98] New, very easy proof [Nagarajan, Ravi, Singh, manuscript] Degree bounded Steiner Network: (2 bv + 3, 2) [Lau Naor Salavatipour Singh 07] Recently, ( bv + O(rmax) , 2 ) [Lau Singh 08]

  8. Previous Work: Directed Case 1-arborescence (no cost): (1+) bv + O(log n)/ [Klein Krishnan Raghavachari Ravi 04] 1-arbor. with cost: (2bv+3, 2)[Lau Naor Salavatipour Singh 07] k-arborescences with cost: (3bv+ 4, 3) Intersecting and crossing supermodular requirements Remark: General connectivity problems for directed graphs are much more difficult even without degree bounds.

  9. Our Results: Directed case 1-arborescence (no cost): b(v)+2 K-arborescences (no cost): b(v) + 4 With cost: For k arborescence problem (b(v)/(1-) + 4, 1/Opt) Note: For  = ½, (2b(v)+4, 2) improves (3 b(v) + 4, 3) of [Lau et al. 07] Best possible: Integrality gap instance, if out-degree violation b(v)/(1-) + O(1), cost violation is (1-o(1))/ Intersecting (crossing) supermodular, both indegree & outdegree bounds

  10. Various other extensions Other network design problems. Degree bounds ! bounds on arbitrary subsets of edges structured families of edge subsets More general settings. Matroids : Generalize Spanning Trees Matroid intersection : Generalize Arborescences Lattice Polyhedra : Generalize above [B, Khandekar, Konemann, Nagarajan, Peis]

  11. Outline of the talk Undirected Case 1) +1 Algorithm of Singh and Lau 2) Basic Solution, Submodularity, Uncrossing 3) Our simple proof 4) Extension to Matroids General Connectivity requirements Directed Case 1) +2 guarantee for arborescence without costs 2) Integrality Gap

  12. LP formulation Variables xe for each edge For E’ ½ E, x(E’) = e 2 E’ xe For S½V, let E(S) be edges completely in S min e ce xe x(E(S)) · |S|-1 8 S ½ V x(E(V)) = |V|-1 e 2(v) xe· bv8 v (degree constraints) 0 · xe· 1 S Spanning Tree Polytope

  13. Singh and Lau Algorithm Iterative relaxation: 0) Obtain a basic solution to the current LP 1) Apply one of these steps a) If xe = 0 for some e, remove e from graph, E = E\{e} b) If xe = 1 for some e, choose it (and update LP) c) If 9 vertex v, s.t. at most bv+1 adjacent edges with xe > 0, then drop degree constraint. 2) Go to step 0. 0.2 0.5 bv=2 0.6

  14. Intermediate LP F(S) denote edges fixed to 1 in set S. W½ V, vertices with active degree constraint min e ce xe x(E(S)) · |S|-1 - |F(S)| 8 S x(E(V)) = |V|-1- |F(V)| e 2(v) xe· bv – |F(v)| 8v 2 W 0 · xe· 1 Remark: Previous solution is still feasible for new LP. No need to resolve, just convert to a basic feasible solution

  15. Correctness If it does not get stuck, easy to see that solution 1) Satisfies degree bound + 1 2) Cost · original LP cost (at most OPT) Main question: Why can we apply one of the steps above? If 0 < xe < 1 for all edges, can drop some degree constraint. Key Idea: LP solution is sparse (few fractional variables) Essentially only technique (?): LP has few constraints But we have exponentially many ?!

  16. Outline of the talk Undirected Case 1) +1 Algorithm of Singh and Lau 2) Basic Solution, Submodularity, Uncrossing 3) Our simple proof 4) Extension to Matroids General Connectivity requirements Directed Case 1) +2 guarantee for arborescence without costs 2) Integrality Gap

  17. Basic Feasible Solution Vertex solution of polytope System: n variables, m ¸ n constraints n linearly independent constraints ! vertex. polytope

  18. Basic Feasible Solution Our Goal: If no xe = 0 or 1, can drop some degree constraint Basic LP solution, some constraints hold with equality. e 2 E(S) xe = |S|-1 - F(S) S (1) e 2(v) xe = bv – F(v) v 2 W (2) xe = 0,1 Determined by some linearly independent constraints of type (1) and (2) Plan: Cannot have many linearly independent tight constraints. Any linearly independent subset suffices

  19. Submodularity F: 2S! R Submodular if for all S,T s.t. S Å T ; F(S [ T) + F(S Å T) · F(S) + F(T) Supermodular if: F(S [ T) + F(S Å T) ¸ F(S) + F(T) If f is supermodular, then -f is submodular Example: x: E ! R+f(S) =x(E(S)) is supermodular x(E(S[T)) + x(E(SÅT)) ¸ x(E(S)) + x(E(T)) S T

  20. Uncrossing Our constraints: x(E(S)) · |S|-1-F(S) Claim: If S, T are tight, then S[T and SÅT also tight Pf:x(E(S)) + x(E(T)) = |S|-1-F(S) + |T|-1- F(T) x(E(S[T)) + x(E(SÅT)) |S[T|-1-F(S[T) + |SÅT|-1-F(SÅT) Any collection of tight sets can be replaced by laminar family of sets submodular supermodular (submodularity) ¸ ¸ (supermodularity) · (feasibility)

  21. Basic Feasible Solution Our Goal: If no xe = 0 or 1, can drop some degree constraint Basic LP solution, some constraints hold with equality. e 2 E(S) xe = |S|-1 - F(S) S (1) e 2(v) xe = bv – F(v) v 2 W (2) xe = 0,1 E: edges in support Determined by some |E| lin. indep. constraints of type (1) and (2) By uncrossing, (1) form a laminar family L Let T be the set of tight constraints of type (2). Any linearly independent subset suffices |E| = |L| + |T|

  22. Intuition Number of edges |E| = |L| + |T| |T| is at most n. Turns out |L| can be at most n-1. Laminar family is a tree with at most n/2 leaves. Each leaf in L corresponds to a set S, with |S| ¸ 2. (Note: constraint x(E(S)) · |S|-1-|F(S)| is void if |S|=1) In general: Key issue is always how to bound |L|.

  23. Outline of the talk Undirected Case 1) +1 Algorithm of Singh and Lau 2) Basic Solution, Submodularity, Uncrossing 3) Our simple proof 4) Extension to Matroids General Connectivity requirements Directed Case 1) +2 guarantee for arborescence without costs 2) Integrality Gap

  24. Proof: 0.2 0.5 bv=2 Define Spare, sp(v) = e 2 N(v) (1-xe ) Claim: If drop constraint for v, degree violation·sp(v) Pf: Violation = (e 2 N(v)1) - bv • ·e 2 N(v)1 - e 2 N(v) xe = sp(v) Need to show: 9 v with degree violation (v) · 1 (suffices to show that some v with sp(v) < 2) 0.6

  25. S T1 T2 Claim:|L| · x(E(V)) Pf: Assign x(e) tokens to edge e, and give to smallest set containing e. Show, each laminar set gets at least 1 So, E = |L| + |T| implies E · x(E(V)) + |T| v 2 W sp(v) = v 2 We 2 N(v) (1-x(e)) ·E 2(1-x(e)) = 2|E| - 2x(E(V)) · 2|T| ·2 |W| Strictness: For equality, each edge has both end pts in W, T=W and |L| = x(E(V)) (so, each edge contained in L) Consider maximal laminar sets M  v 2 T(v) = 2 ME(M)(linear dependence) x(E(S)) – x(E(T1)) – x(E(T2)) = integer > 0

  26. Extensions • Crossing MST: E1,…,Es edge subsets, bounds b1,…,bs. Find min-cost MST with at most bi edges from Ei. Additive r-1 approximation, if an edge lies in ·r sets Ei (Deg. bounded Spanning tree: Ei’s = Nv’s, then r = 2 ) Previously: O(r log n) mult. approx [Bilo, Goyal, Ravi Singh 04] • Min-cost basis in a matroid: E1,…,Es sets of elements with bounds b1,…,bs. Each element in at most r sets Ei. Additive r-1 approx. (Similar results independently obtained by [Kiraly, Lau, Singh 08])

  27. In hindsight |E| = |L| + |T| (basic solution) Bound |L| by some function of xe (often a natural way) |E| - E f(xe) ·|T| · |W| (view as some function of overall spare)

  28. Outline of the talk Undirected Case 1) +1 Algorithm of Singh and Lau 2) Basic Solution, Submodularity, Uncrossing 3) Our simple proof 4) Extension to Matroids General Connectivity requirements Directed Case 1) +2 guarantee for arborescence without costs 2) Integrality Gap

  29. General Connectivity Requirements Steiner Network: rij edge-disjoint path between i,j Min e ce xe (cut formulation) x((S)) ¸ r(S) for all S r(S) = max(i,j) 2(S) rij 0 · xe· 1 Cut formulations problematic. Can only guarantee that some variable ¸ ½ Eg: Min. Spanning Tree Cut LP: x((S)) ¸ 1 for all S G = Cycle, x(e) =1/2 is feasible.

  30. Steiner Network Jain [98]: Can do uncrossing. x((S)) submodular r(S) Weakly Supermodular a) A[B & AÅB tight Or, b) A\B & B\A tight. Jain (iterated rounding): Round xe¸ ½ to 1 Degree bounds: (2bv +3, 2)[ Lau Naor Salavatipour Singh 07] Algorithm: Either some xe¸ ½ or, can drop some degree constraint.

  31. General Connectivity + deg bounds Surprising recent result (bv+ O(rmax),2) [Lau Singh 08] • Some xe¸ ½ s.t. |support| on endpoints of e = O(rmax) 2) Or, can drop some degree constraint “Idea”: Many edges xe¸ ½, few vertices with high support. Tight: Integrality gap of (bv + (rmax), 2)

  32. Outline of the talk Undirected Case 1) +1 Algorithm of Singh and Lau 2) Submodularity, Uncrossing 3) Our simple proof 4) Extension to Matroids General Connectivity requirements Directed Case 1) +2 guarantee for arborescence without costs 2) Integrality Gap

  33. Directed graphs r Cut Formulation for Arborescence x(in(S)) ¸ 1 S µ V \ {r} x(out(v)) · bv Thm: No degree constraints, Polytope integral[Edmonds 70] (just 1 degree constraint ! |E| = 2n-2, possibly ½ integral) Guarantee: (2bv+ 3, 2 cost) [Lau et al 07] (a) Either some xi¸ 1/2 (b) can drop degree constraint

  34. Our Algorithms No cost, guarantee bv + 2 Algorithm (iterated rounding): 1) If tail of e has no degree constraint, set x(e)=1 2) If tail of edge e has a degree constraint a) Either x(e) =1 b) Or, some degree constraint can be dropped

  35. k-Arborescences with costs In the cost version, guarantee (bv/1- + 4 , 1/) Algorithm: If tail of e has no deg. const. and x(e) ¸, set x(e) =1 If tail has degree constraint • Either some x(e) ¸ 1- • Some degree constraint can be dropped

  36. Guarantee, best possible k-ary tree rooted at r bv = k(1-) at internal vertices Solid edge: cost 0, x(e) = 1- Dotted edges: cost 1, x(e) =  LP cost =  (# of leaves) 1- 1- 1- 1- 1- 1- Integral Cost =  (# leaves) +  ( 1- ) (# leaves) +  (1- )2 (# leaves) + … ¼ LP cost / 

  37. Conclusions Iterative relaxation/rounding seems powerful technique Simpler proofs for several traditional results. 1) Can often get around limitations of cut formulations. 2) Is it NP-Hard to improve (bv/(1-), 1/) tradeoff for arborescence with costs? 3) Can we get +1 for arborescence without costs.

  38. Thank You!

  39. Result for structured laminar families Result for matroid intersection Crossing Spanning Tree problem, open problems.

  40. Proof for no cost version E = L + T (uncrossing) Will show: L < x+2W So, E < x+2W+T · x+3W implies 9 v with sp(v) < 3 Give each edge e, xe tokens which lie on head And each W vertex 2 tokens. Show each set in L can get 1 token. Key point: The tail of any edge lies on a W-vertex xe

  41. Make picture of integrality gap instance. to explain …

  42. Node is marked if contains W vertex. : subtree of marked nodes Unmarked leaf: 1 token Unmarked non-leaf does not exist Marked leaf in  has private W-vertex E = L + T L < x + 2 W Marked single child node, has a private W-vertex

  43. Gadget: 1-  1- 1- 1-    

More Related