41
Improved Decremental Algorithms for Transitive Closure All-Pairs Shortest Paths in Directed Graphs

Improved Decremental Algorithms for

  • Upload
    zeke

  • View
    33

  • Download
    0

Embed Size (px)

DESCRIPTION

Improved Decremental Algorithms for. Transitive Closure All-Pairs Shortest Paths in Directed Graphs. Surender Baswana work supported by a fellowship from Infosys Tech. Bangalore Sandeep Sen work supported in part by an IBM UPP award I.I.T. Delhi India. - PowerPoint PPT Presentation

Citation preview

Page 1: Improved Decremental Algorithms for

Improved Decremental Algorithms for

• Transitive Closure• All-Pairs Shortest Paths in Directed Graphs

Page 2: Improved Decremental Algorithms for

34th ACM STOC 2002

• Surender Baswana work supported by a fellowship from Infosys Tech. Bangalore

• Sandeep Senwork supported in part by

an IBM UPP award

I.I.T. Delhi

India

• Ramesh Hariharan

I.I.Sc. Bangalore

India

Page 3: Improved Decremental Algorithms for

Two Fundamental Algorithmic Graph Problems

• Transitive Closure

• All-Pairs Shortest Paths

Page 4: Improved Decremental Algorithms for

Transitive Closure

• Process a given directed graph G(V,E) so that any query

of the form,

“ Is there a path from u to v in the graph?”

can be answered efficiently

Page 5: Improved Decremental Algorithms for

Existing Static Algorithms

• Classical Algorithms : O(mn)

• Algorithms based on fast matrix multiplications : O(n2.575)

(Zwick, FOCS 1998)

Page 6: Improved Decremental Algorithms for

Dynamic Maintenance of APSP (Transitive Closure)

Given a directed graph G(V,E) followed by an online sequence of updates interspersed with queries

Process each query and update efficiently

Page 7: Improved Decremental Algorithms for

Aim of Dynamic Algorithms for APSP (Transitive Closure)

To Maintain a data-structure that

• can answer each query efficiently

• can be updated efficiently after each update (in time less than the best static

algorithm)

Page 8: Improved Decremental Algorithms for

Two Types of Dynamic Algorithms

• Fully Dynamic ( both insertion and deletion of edges allowed)

• Partially Dynamic Incremental (only insertion) Deletion (only deletion)

Page 9: Improved Decremental Algorithms for

Decremental Algorithms for Transitive Closure

Query Time Update time(per edge deletion)

Leeuwen et al.

1998O(1) O(m)

Henzinger & King, STOC 1995

O(n/log n) O(n log n)

Demetrescu et al.

FOCS 2000O(1) O(n/m)

2

3

Page 10: Improved Decremental Algorithms for

Contribution of our Paper

Decremental Algorithms for

• Transitive Closure

• APSP

that achieve improved update time while keeping optimal query time

Page 11: Improved Decremental Algorithms for

Decremental Algorithm for transitive closure

Page 12: Improved Decremental Algorithms for

Main Idea

A new data-structure that

• stores transitive closure imformation implicitly

• is more suitable for maintaining reachability for long paths

Page 13: Improved Decremental Algorithms for

Main Idea Contd.

Suitably combine two data-structures

• S:

suitable for reachability corresponding to short paths.

• L :

suitable for reachability corresponding to long paths.

Page 14: Improved Decremental Algorithms for

Basic Data-structure Used

Theorem : A BFS tree of depth d, rooted at a vertex can be maintained under deletion of edges that

• takes O(1) query time

• requires O(d) amortized update time per edge deletion.

Page 15: Improved Decremental Algorithms for

Explicit Maintenance of Transitive Closure

uv

Page 16: Improved Decremental Algorithms for

Explicit Maintenance of Transitive Closure contd.

uv

Page 17: Improved Decremental Algorithms for

Implicit Maintenance of Transitive Closure

uv

Page 18: Improved Decremental Algorithms for

Implicit Maintenance of Transitive Closure

uv

Page 19: Improved Decremental Algorithms for

Implicit Maintenance of Transitive Closure

uv

In_treeOut_tree

Page 20: Improved Decremental Algorithms for

Transitive closure corresponding to paths of length d passing through

set of witnesses W

Page 21: Improved Decremental Algorithms for

Transitive closure corresponding to paths of length d passing through

set of witnesses W

Page 22: Improved Decremental Algorithms for

Transitive closure corresponding to paths of length d passing through

set of witnesses W

u

vM

Page 23: Improved Decremental Algorithms for

Transitive closure corresponding to paths of length d passing through

set of witnesses W

u

v

w

M

Page 24: Improved Decremental Algorithms for

Handling Edge Deletion

u

v

w

u vw

W’

M

Page 25: Improved Decremental Algorithms for

Handling Edge Deletion contd.

u

v

w

u vw

u vW’

W’

M

Page 26: Improved Decremental Algorithms for

Handling Edge Deletion contd.

u

v

w

u vw

u vW’

W’

M

Page 27: Improved Decremental Algorithms for

Handling Edge Deletion contd.

• For each w, update in_tree(w) and out_tree(w)

• For each pair (u,v) for which w has ceased to remain witness, search for new witness.

• Update M[u,v] appropriately

Page 28: Improved Decremental Algorithms for

Handling Edge Deletion contd.

• Invariant maintained :

M[u,v] points to the first vertex in the witness list W that is a witness of

reachability from u to v

Page 29: Improved Decremental Algorithms for

Total update time

• For each pair (u,v), the witness list is scanned exactly once.

O(n |W|)

• Total update time for maintaining in_trees

and out_trees

O(md|W|)

2

Page 30: Improved Decremental Algorithms for

Maintaining all-pairs reachability corresponding to paths of length <d

passing through W

Theorem :

All-pairs reachability corresponding to paths of length <d passing through vertices of set W can be maintained under deletion of edges in amortized time :

O((n2/m + d)W) per edge deletion

Page 31: Improved Decremental Algorithms for

Random sampling to pick the witness vertices

Theorem : Given a path puv of length l from u to v, if we sample cn/l log2 n vertices (for any c>0), then with probability 1-1/nc , at least one of the vertices will be picked from the path puv

Page 32: Improved Decremental Algorithms for

Maintaining all-pairs reachability corresponding to paths of length

[d/2,d]

• Form a set Wd of n/d log n vertices uniformly randomly.

• Build and maintain in_trees and out_trees on vertices of set Wd.

Update time : O((n3/(dm) + n)log n)

Page 33: Improved Decremental Algorithms for

Maintaining all-pairs reachability corresponding to paths of length

[d,n]

• For each interval [2id,2i+1d], build and maintain in_trees and out_trees on a collection of n/(2id) log n vertices. • Maintain the reachability matrix M Total Update time : O((n3/(dm) + n)log2 n)

Page 34: Improved Decremental Algorithms for

Comparing the two data-structures

• Explicit maintenance of transitive closure

for path of length

[1,d]

Update time :

O(nd)

• Implicit maintenance of transitive closure for paths of length

[d,n]

Update time :

O((n3/(dm)+n)log2n)

Page 35: Improved Decremental Algorithms for

Comparing the two data-structures

• Explicit maintenance of transitive closure

for path of length [1,d] Update time : O(nd)

Suitable for short paths

• Implicit maintenance of transitive closure for paths of length

[d,n] Update time : O((n3/(dm)

+n)log2n)

Suitable for long paths

Page 36: Improved Decremental Algorithms for

Improved algorithm for transitive closure

Previous New

Update time O(n3/2) O(n4/3)

Source Leeuwen et al. (1988) and Demetrescu (FOCS 2000)

Page 37: Improved Decremental Algorithms for

Improved Decremental Algorithm for maintaining all-

pairs 2-approximate paths

Page 38: Improved Decremental Algorithms for

u

v

w

Page 39: Improved Decremental Algorithms for

• Invariant maintained :

M[u,v] points to the first vertex in the witness list W that is a witness of

reachability from u to v

Page 40: Improved Decremental Algorithms for

Improved update bounds for maintaining 2-approximate

APSPPrevious New

Update time per edge deletion none O(n2/m)

Page 41: Improved Decremental Algorithms for

Improved Update bounds for maintaining exact APSP

Type of query Previous New

Report distance from u to v

O(n3m log3 n)

FOCS 2000

O(n3/m log2 n)

Report shortest path from u to v

O(n3/m log3 n)

FOCS 2000

Min(n3/m log2n,

n1.5 log n)