Kruskal for maximum spanning forest: Difference between revisions

From Algowiki
Jump to navigation Jump to search
No edit summary
 
(6 intermediate revisions by 2 users not shown)
Line 1: Line 1:
[[Category:Videos]]
[[Category:Videos]]
{{#ev:youtube|https://www.youtube.com/watch?v=Pz6x3BB86YA|500|right|Algorithmus von Kruskal|frame}}
{{#ev:youtube|https://www.youtube.com/watch?v=Pz6x3BB86YA|500|right|Chapters
#[00:00] Einführung
#[01:59] Der Algorithmus von Kruskal anhand eines Beispiels
#[06:56] Und wie ist das bei mehreren Zusammenhangskomponenten?
#[07:29] Welches Problem löst der Algorithmus von Kruskal genau?
#[08:05] Wie lautet die Invariante?
#[08:33] Warum ist der Algorithmus korrekt?
#[08:59] Wie wird die Invariante sichergestellt?
#[10:25] Was ist die asymptotische Komplexität des Algorithmus?
|frame}}
 
== General information ==
 
'''Algorithmic problem:''' [[Maximum spanning forest]]
'''Algorithmic problem:''' [[Maximum spanning forest]]


Line 47: Line 59:


==Complexity==
==Complexity==
'''Statement:''' The asymptotic complexity is in <math>\mathcal{O}(m \cdot T(n) + m\cdot log\cdot m)</math> in the worst case, where <math>n = \vert V \vert</math> and <math>m = \vert E \vert</math>, and <math>T(n)</math> is the complexity of a single [[union-find]] operation.
'''Statement:''' The asymptotic complexity is in <math>\mathcal{O}(n \cdot T(n) + m\cdot\log m )</math> in the worst case, where <math>n = \vert V \vert</math> and <math>m = \vert E \vert</math>, and <math>T(n)</math> is the complexity of a single unite operation.


'''Proof:''' The summand <math>\mathcal{O}(m\cdot log\cdot m)</math> results from sorting all edges in the preprocessing. The main loop has <math>m</math> iterations, and each iteration takes <math>\mathcal{O}(T(n))</math> time.
'''Proof:''' The summand <math>\mathcal{O}(m\cdot \log m)</math> results from sorting all edges in the preprocessing. The main loop has <math>m</math> iteration. There are <math>n-1</math> unite operations.

Latest revision as of 18:31, 20 September 2015

Chapters
  1. [00:00] Einführung
  2. [01:59] Der Algorithmus von Kruskal anhand eines Beispiels
  3. [06:56] Und wie ist das bei mehreren Zusammenhangskomponenten?
  4. [07:29] Welches Problem löst der Algorithmus von Kruskal genau?
  5. [08:05] Wie lautet die Invariante?
  6. [08:33] Warum ist der Algorithmus korrekt?
  7. [08:59] Wie wird die Invariante sichergestellt?
  8. [10:25] Was ist die asymptotische Komplexität des Algorithmus?

General information

Algorithmic problem: Maximum spanning forest

Prerequisites:

Type of algorithm: loop

Auxiliary data:

  1. A sorted sequence [math]\displaystyle{ S }[/math] whose items are the edges of [math]\displaystyle{ E }[/math]. The edges are sorted in descending order, that is, the edge [math]\displaystyle{ e \in E }[/math] with the largest value of [math]\displaystyle{ \ell(e) }[/math] is the first element of the sequence (ties arbitrarily broken). Below, [math]\displaystyle{ e_i }[/math] denotes the [math]\displaystyle{ i }[/math]-th edge in [math]\displaystyle{ S }[/math].
  2. A union-find data structure [math]\displaystyle{ U F }[/math] on [math]\displaystyle{ V }[/math].

Abstract view

Invariant: After [math]\displaystyle{ i \ge 0 }[/math] iterations:

  1. The undirected graph [math]\displaystyle{ F_i = (V,E_i) }[/math], where [math]\displaystyle{ E_i }[/math] is the set of all selected edges immediately after the [math]\displaystyle{ i }[/math]-th iteration, is cycle-free; in particular, it is maximally cycle-free in the sense that, for [math]\displaystyle{ j \in \{1,...,i\} }[/math], adding an edge [math]\displaystyle{ e_J \notin F_i }[/math] would close a cycle in [math]\displaystyle{ F_i }[/math].
  2. For the undirected graph [math]\displaystyle{ F_i = (V,E_i) }[/math], there is a maximum spanning forest [math]\displaystyle{ F_i' = (V,E'_i) }[/math] of [math]\displaystyle{ G }[/math] such that [math]\displaystyle{ E_i \subseteq E'_i }[/math].
  3. The sets of [math]\displaystyle{ U F }[/math] are the connected components of [math]\displaystyle{ F_i }[/math].

Variant: [math]\displaystyle{ i }[/math] increases by [math]\displaystyle{ 1 }[/math].

Break condition: [math]\displaystyle{ i = \vert E \vert }[/math].

Induction basis

Abstract view:

  1. [math]\displaystyle{ E_0 := \emptyset }[/math]
  2. [math]\displaystyle{ U F }[/math] is initialized by [math]\displaystyle{ V }[/math], so each node [math]\displaystyle{ v \in V }[/math] is a singleton in [math]\displaystyle{ U F }[/math].

Implementation: Obvious.

Proof: Nothing to show.

Induction step

Abstract view: If inserting [math]\displaystyle{ e_i }[/math] would close a cycle, reject [math]\displaystyle{ e_i }[/math]; otherwise, insert [math]\displaystyle{ e_i }[/math].

Implementation:

  1. Let [math]\displaystyle{ v,w \in V }[/math] denote the endnodes of [math]\displaystyle{ e_i }[/math], that is, [math]\displaystyle{ e_i = \{v,w\} }[/math].
  2. If [math]\displaystyle{ v }[/math] and [math]\displaystyle{ w }[/math] belong to different sets in [math]\displaystyle{ U F }[/math]:
    1. Insert [math]\displaystyle{ e_i }[/math] in the result.
    2. Unite the sets of [math]\displaystyle{ v }[/math] and [math]\displaystyle{ w }[/math] in [math]\displaystyle{ U F }[/math].

Correctness: The first and third invariants are obvious. So we focus on the second invariant. We may safely set [math]\displaystyle{ F'_i := F'_{i-1} }[/math] if

  1. either [math]\displaystyle{ e_i \in E'_{i-1} }[/math] and [math]\displaystyle{ e_i }[/math] is inserted in the [math]\displaystyle{ i }[/math]-th iteration
  2. or [math]\displaystyle{ e_i \notin E'_{i-1} }[/math] and [math]\displaystyle{ e_i }[/math] is rejected in the -th iteration.

Note that the case that [math]\displaystyle{ e_i \in E'_{i-1} }[/math] and [math]\displaystyle{ e_i }[/math] is rejected cannot occur. In fact, [math]\displaystyle{ e_i \in E'_{i-1} }[/math] implies that [math]\displaystyle{ (V,E_{i-1} \cup{\{e_i\}}) }[/math] is cycle-free, but rejecting [math]\displaystyle{ e_i }[/math] implies that [math]\displaystyle{ (V,E_{i-1} \cup{\{e_i\}}) }[/math] contains a cycle. So it remains to consider the case that [math]\displaystyle{ e_i }[/math] is accepted but [math]\displaystyle{ e_i \notin E'_{i-1} }[/math].

As [math]\displaystyle{ \ell(e_i) \gt 0 }[/math] and [math]\displaystyle{ F'_{i-1} }[/math] is maximal, [math]\displaystyle{ (V,E'_{i-1} \cup{\{e_i\}}) }[/math] contains a cycle. Since [math]\displaystyle{ F_{i-1} }[/math] is cycle-free, there must be an edge [math]\displaystyle{ e \in E'_{i-1} \setminus E_{i-1} }[/math] on that cycle. By assumption, [math]\displaystyle{ (V,E_{i-1} \cup{\{e_i\}}) }[/math] is cycle-free as well. Therefore, [math]\displaystyle{ e }[/math] is none of the edges considered so far but one of [math]\displaystyle{ e_{i+1},...,e_{\vert E \vert } }[/math], which implies [math]\displaystyle{ \ell(e) \leq \ell(e_i) }[/math] and, hence, maximality of the spanning forest [math]\displaystyle{ F'_i := (V,E'_{i-1} \setminus \{e\} \cup{\{ e_i \}}) }[/math]. Clearly, this [math]\displaystyle{ F'_i }[/math] includes [math]\displaystyle{ F_i }[/math], so the second invariant is maintained again.

Complexity

Statement: The asymptotic complexity is in [math]\displaystyle{ \mathcal{O}(n \cdot T(n) + m\cdot\log m ) }[/math] in the worst case, where [math]\displaystyle{ n = \vert V \vert }[/math] and [math]\displaystyle{ m = \vert E \vert }[/math], and [math]\displaystyle{ T(n) }[/math] is the complexity of a single unite operation.

Proof: The summand [math]\displaystyle{ \mathcal{O}(m\cdot \log m) }[/math] results from sorting all edges in the preprocessing. The main loop has [math]\displaystyle{ m }[/math] iteration. There are [math]\displaystyle{ n-1 }[/math] unite operations.