Improved bonferroni inequalities via abstract tubes

Lecture Notes in Mathematics Editors: J.--M. Morel, Cachan F. Takens, Groningen B. Teissier, Paris

1826

3 Berlin Heidelberg New York Hong Kong London Milan Paris Tokyo

Klaus Dohmen

Improved Bonferroni Inequalities via Abstract Tubes Inequalities and Identities of Inclusion-Exclusion Type

13

Author Klaus Dohmen Department of Mathematics Mittweida University of Applied Sciences Technikumplatz 17 09648 Mittweida Germany e-mail: [email protected] http://www.dohmen.htwm.de

Cataloging-in-Publication Data applied for Bibliographic information published by Die Deutsche Bibliothek Die Deutsche Bibliothek lists this publication in the Deutsche Nationalbibliografie; detailed bibliographic data is available in the Internet at http://dnb.ddb.de

Mathematics Subject Classification (2000): primary: 05A19, 05A20, 60C05, 60E15 secondary: 05A15, 05B35, 05C15, 06A15, 62N05, 68M15, 68R10, 90B15, 90B25 ISSN 0075-8434 ISBN 3-540-20025-8 Springer-Verlag Berlin Heidelberg New York This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specif ically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microf ilm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer-Verlag. Violations are liable for prosecution under the German Copyright Law. Springer-Verlag Berlin Heidelberg New York a member of BertelsmannSpringer Science + Business Media GmbH springer.de c Springer-Verlag Berlin Heidelberg 2003
Printed in Germany

The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specif ic statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Typesetting: Camera-ready TEX output by the authors SPIN: 10957806

41/3142/du - 543210 - Printed on acid-free paper

Preface This work is based on my habilitation thesis which I prepared at Berlin’s Humboldt-University while I was an assistant professor at the computer science department in the years 1994–2000. The material has been re-arranged, some parts have been significantly shortened, while other parts have been expanded. I am especially grateful to Professor Egmar R¨odel at Humboldt-University for giving me the opportunity to work on this subject. Special thanks go to the referees of my habilitation thesis as well as to the anonymous referees of Springer-Verlag for their worthy suggestions that resulted in an improvement of the manuscript. Particular thanks are owed to Professor Douglas Shier (Clemson University) for his valuable comments. Finally, I would like to thank my wife Imke and my children Sophie, Emily and Justus for their never-ending patience. Mittweida, Germany July, 2003

Klaus Dohmen

Contents

1 Introduction and Overview

1

2 Preliminaries 2.1 Graphs and posets . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5 5 7

3 Bonferroni Inequalities via Abstract Tubes 3.1 An outline of abstract tube theory . . . . . . . . . . . 3.1.1 The notion of an abstract tube . . . . . . . . . 3.1.2 Two fundamental theorems on abstract tubes . 3.1.3 An importance sampling scheme . . . . . . . . 3.2 Examples of abstract tubes . . . . . . . . . . . . . . . 3.2.1 Abstract tubes for convex polyhedra . . . . . . 3.2.2 Abstract tubes for unions of balls and spherical

. . . . . . . . . . . . . . . . . . caps

. . . . . . .

. . . . . . .

. . . . . . .

9 9 9 11 13 14 14 15

4 Abstract Tubes via Closure and Kernel Operators 4.1 Abstract tubes via closure operators . . . . . . . . . 4.1.1 Closure operators . . . . . . . . . . . . . . . . 4.1.2 How closure operators lead to abstract tubes 4.2 Abstract tubes via kernel operators . . . . . . . . . . 4.2.1 Kernel operators . . . . . . . . . . . . . . . . 4.2.2 How kernel operators lead to abstract tubes . 4.3 Alternative proofs . . . . . . . . . . . . . . . . . . . 4.3.1 Improved identities via closure operators . . . 4.3.2 Improved inequalities via kernel operators . . 4.3.3 Additional proofs of identities . . . . . . . . . 4.4 The chordal graph sieve . . . . . . . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

19 19 19 21 25 25 26 28 28 33 35 37

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

5 Recursive Schemes 44 5.1 Recursive schemes for semilattices . . . . . . . . . . . . . . . . . 44 5.2 Dynamic programming approach . . . . . . . . . . . . . . . . . . 45

VIII

Contents

6 Reliability Applications 6.1 System reliability . . . . . . . . . . . . . . . . . . . 6.1.1 Terminology . . . . . . . . . . . . . . . . . 6.1.2 Abstract tubes for system reliability . . . . 6.1.3 Shier’s pseudopolynomial algorithm . . . . 6.1.4 Inclusion-exclusion and domination theory . 6.2 Special reliability systems . . . . . . . . . . . . . . 6.2.1 Communications networks . . . . . . . . . . 6.2.2 k-out-of-n systems . . . . . . . . . . . . . . 6.2.3 Consecutive k-out-of-n systems . . . . . . . 6.3 Reliability covering problems . . . . . . . . . . . . 6.3.1 The hypergraph model . . . . . . . . . . . . 6.3.2 Abstract tubes and polynomial algorithms . 6.4 Chapter notes . . . . . . . . . . . . . . . . . . . . . 7 Combinatorial Applications and Related Topics 7.1 Inclusion-exclusion on partition lattices . . . . . 7.2 Chromatic polynomials and broken circuits . . . 7.2.1 The usual chromatic polynomial . . . . . 7.2.2 The generalized chromatic polynomial . . 7.3 Sums over partially ordered sets . . . . . . . . . . 7.3.1 A general theorem on sums . . . . . . . . 7.3.2 Application to inclusion-exclusion . . . . 7.4 Matroid polynomials and the β invariant . . . . . 7.4.1 The Tutte polynomial . . . . . . . . . . . 7.4.2 The characteristic polynomial . . . . . . . 7.4.3 The β invariant . . . . . . . . . . . . . . . 7.5 Euler characteristics and M¨obius functions . . . . 7.5.1 Euler characteristics . . . . . . . . . . . . 7.5.2 M¨ obius functions . . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . .

47 47 47 48 49 52 53 53 66 70 75 75 75 81

. . . . . . . . . . . . . .

82 82 83 84 86 89 89 90 91 91 93 94 95 95 96

Bibliography

100

Author Index

111

Subject Index

113

1

Introduction and Overview

Many problems in combinatorics, number theory, probability theory, reliability theory and statistics can be solved by applying a unifying method, which is known as the principle of inclusion-exclusion. The principle of inclusionexclusion expresses the indicator function of a union of finitely many sets as an alternating sum of indicator functions of their intersections. More precisely, for any finite family of sets {Av }v∈V the principle of inclusion-exclusion states that
(1.1)

χ

 v∈V

 Av

=

 I⊆V I=∅


|I|−1

(−1)

χ



 Ai

,

i∈I

where χ(A) denotes the indicator function of A with respect to some universal set Ω, that is, χ(A)(ω) = 1 if ω ∈ A, and = 0 if ω ∈Ω\A. Equivalently,  χ(A)(ω)  (1.1) can be expressed as χ v∈V Av =  I⊆V (−1)|I| χ i∈I Ai , where Av denotes the complement of Av in Ω and i∈∅ Ai = Ω. A proof by induction on the number of sets is a common task in undergraduate courses. Usually, the Av ’s are measurable with respect to some finite measure µ on a σ-field of subsets of Ω. Integration of the indicator function identity (1.1) with respect to µ then gives
 
   |I|−1 µ (1.2) Av = (−1) µ Ai , v∈V

I⊆V I=∅

i∈I

which now expresses the measure of a union of finitely many sets as an alternating sum of measures of their intersections. The step leading from (1.1) to (1.2) is referred to as the method of indicators [GS96b]. Naturally, two special cases are of particular interest, namely the case where µ is the counting measure on the power set of Ω, and the case where µ is a probability measure on a σ-field of subsets of Ω. These special cases are sometimes attributed to Sylvester (1883) and Poincar´e (1896), although the second edition of Montmort’s book “Essai d’Analyse sur les Jeux de Hazard”, which appeared in 1714, already contains an implicit use of the method, based on a letter by N. Bernoulli in 1710. A first K. Dohmen: LNM 1826, pp. 1–4, 2003. c Springer-Verlag Berlin Heidelberg 2003


1 Introduction and Overview

2

explicit description of the method was given by Da Silva (1854). For references to these sources and additional historical notes, we refer to Tak´ acs [Tak67]. Since the identities (1.1) and (1.2) contain 2|V | − 1 terms and intersections of up to |V | sets, one often resorts to inequalities like that of Boole [Boo54]:
   χ (1.3) Av ≤ χ(Ai ) . v∈V

i∈V

A more general result, first discovered by Ch. Jordan [Jor27] and later by Bonferroni [Bon36], states that for any finite family of sets {Av }v∈V and any r ∈ ,  

   (1.4) (r odd), Av ≤ (−1)|I|−1 χ Ai χ


v∈V



χ

(1.5)

I⊆V 00 fr >0 fr which completes the proof. 2

3.2

Examples of abstract tubes

The original motivation of Naiman and Wynn [NW92, NW97] that led to the theory of abstract tubes was the problem of computing or bounding the volume or probability content of certain geometric objects such as convex polyhedra and unions of finitely many Euclidean balls or spherical caps, which has applications to the statistical theory of multiple comparisons [NW92] and to computational biology where protein molecules are modeled as unions of balls in three-dimensional Euclidean space [Ede95]. Related problems have also been considered by Edelsbrunner [Ede95] as well as Edelsbrunner and Ramos [ER97].

3.2.1

Abstract tubes for convex polyhedra

In this section we deal with d-dimensional polyhedra in


d .

 Definition 3.2.1 A polyhedron in
d is a set P = v∈V Hv where {Hv }v∈V is a finite family of open half-spaces in
d and Hv denotes the complement of Hv in
d . P is d-dimensional if it contains d + 1 affinely independent points. 

 Note that due to a law of De Morgan, χ(P ) = 1 − χ v∈V Hv . Thus, by Theorem 3.1.9, the following result of Naiman and Wynn [NW97] gives rise to improved inclusion-exclusion identities and Bonferroni inequalities for the indicator function (and hence for the volume or probability content) of a ddimensional polyhedron in
d . For any H ⊆
d we use ∂H to denote the topological boundary of H, that is, the difference between its topological closure and interior.

3.2 Examples of abstract tubes

15

 Theorem 3.2.2 [NW97] Let P = v∈V Hv be a d-dimensional polyhedron in  
d . Then, {Hv }v∈V , {I ∈
∗ (V )| i∈I (P ∩ ∂Hi ) = ∅} is an abstract tube.   

Proof. (Sketch) Let := I ∈
∗ (V )  i∈I (P ∩ ∂Hi ) = ∅ and ω ∈ v∈V Hv . By Definition 3.1.1 we have to prove that (ω), with Hi instead of Ai , is contractible. Observe that (ω) is the nerve of C(ω) := {P ∩ ∂Hi | ω ∈ Hi }, whence by a classical theorem of Borsuk [Bor48] the contractibility of (ω) follows from

that of C(ω). To

prove contractibility of C(ω), we show that some homeomorphic image of C(ω) is contractible: Let π be the homeomorphism mapping

each ν ∈ C(ω) to the intersection of the line segment νω with some hyperplane K separating ω from P . Then, the image of C(ω) under π is convex and hence contractible. 2 Remark. Note that if P is in general position, that is, if no d + 1 facets P ∩ ∂Hi share a common point, then the abstract simplicial complex in the abstract tube of Theorem 3.2.2 is at most (d − 1)-dimensional. In this case, the improved inclusion-exclusion identity associated with the abstract tube of Theorem 3.2.2 involves intersections of up to d half-spaces only and thus contains at most         |V | |V | |V | + + ··· + = O |V |d 1 2 d terms, whereas the classical inclusion-exclusion identity still contains 2|V | − 1 terms and involves intersections of up to |V | half-spaces. If P is not in general position, then due to Naiman  and ˜Wynn [NW97] it can be perturbed slightly to give a polyhedron P˜ = v∈V H , which is in general position and which  v   ˜ i ) = ∅} gives rise to a weak abstract tube {Hv }v∈V , {I ∈
∗ (V )| i∈I (P˜ ∩ ∂ H (involving the original half-spaces Hv ) with respect to Lebesgue measure. Example 3.2.3 Consider the two-dimensional polyhedron P = H1 ∩ H2 ∩ H3 ∩ H4 ∩ H5 which is displayed in Figure 3.1. By combining Theorem 3.2.2 with Theorem 3.1.9 we obtain the improved inclusion-exclusion identity χ(P ) = 1 − χ(H1 ) − χ(H2 ) − χ(H3 ) − χ(H4 ) − χ(H5 ) + χ(H1 ∩ H2 ) + χ(H2 ∩ H3 ) + χ(H3 ∩ H4 ) + χ(H4 ∩ H5 ) , which contains ten terms and intersection of up to two sets only. In contrast, the traditional inclusion-exclusion formula for the indicator function of the same polyhedron contains 25 = 32 terms and intersection of up to five sets.

3.2.2

Abstract tubes for unions of balls and spherical caps

The abstract tube of Theorem 3.2.5 below derives from a Voronoi subdivision of the underlying space and its associated Delauney complex. We briefly review these notions in the following definition.

16

3

Bonferroni Inequalities via Abstract Tubes

H5

H4

P

H3 H2 H1

Figure 3.1: A two-dimensional polyhedron.

Definition 3.2.4 Let d ∈ . For any x ∈
d and r > 0 let Bd (x, r) denote the open ball in
d with center x and radius r, that is, Bd (x, r) := {y ∈
d | δ(x, y) < r} where δ(·, ·) denotes Euclidean distance. With any finite set V ⊆
d we associate the Voronoi subdivision of
d into non-empty closed convex polyhedra (3.6)

Dv := {x ∈
d | δ(x, v) = min δ(x, u)} u∈V

(v ∈ V )

consisting of points closest in Euclidean distance to v, and the Delauney complex     ∗  (V ) := I ∈
(V )  Di = ∅ . (3.7) i∈I

Theorem 3.2.5 [NW92, NW97] Let V be a finite set of points in
d and r > 0. Then, ({Bd (v, r)}v∈V , (V )) is an abstract tube.

Proof. (Sketch) Let := (V ) and ω ∈ v∈V Bd (v, r). By Definition 3.1.1 we have to show that (ω), with Bd (i, r) in place of Ai , is contractible. We first observe that (ω) is the nerve of C(ω) := {Di | ω ∈ Bd (i, r)}, whence similar to the proof of

Theorem 3.2.2 the contractibility of (ω) follows from that of C(ω). Indeed, C(ω) is contractible since it is star-shaped with respect to ω. 2 Remark. If the centers of the balls are in general position, that is, if no d + 1 centers of the balls lie in a (d − 1)-dimensional affine subspace of
d and no

3.2 Examples of abstract tubes

17

x ∈
d is equidistant from more than d + 1 of the centers of the balls, then the intersection of more than d + 1 of the sets Dv is empty and hence the dimension of (V ) is at most d, regardless of the radius r. In this case, the improved inclusion-exclusion identity associated with the abstract tube of Theorem 3.2.5 involves intersections of up to d + 1 balls and therefore contains at most         |V | |V | |V | + + ···+ = O |V |d+1 1 2 d+1 terms, whereas the classical inclusion-exclusion identity still contains 2|V | − 1 terms and involves intersections of up to |V | balls. As noted in [NW92, NW97], the result can be generalized so that the balls may have different radii. The following example explains the word tube in Definition 3.1.1. Example 3.2.6 Figure 3.2 shows ten disks in
2 of equal radius with equidistant centers on a straight line. By Theorem 3.2.5, these ten disks together with the path comprising the centers of the disks constitute an abstract tube.

Figure 3.2: Ten disks of equal radius with centers on a straight line.

Example 3.2.7 Consider the disks in Figure 3.3(a). Clearly, the classical inclusion-exclusion identity for the indicator function of their union contains 220 − 1 = 1048575 terms and intersections of up to 20 disks. In order to apply Theorem 3.2.5 we first form the Voronoi subdivision according to (3.6) and then the Delauney complex according to (3.7), see Figure 3.3(b). The Delauney complex contains 20 vertices, 47 edges and 28 triangles. Hence, the improved inclusion-exclusion identity associated with the abstract tube of Theorem 3.2.5 contains only 20 + 47 + 28 = 95 terms and intersections of up to three disks. The preceding theorem has an analogue for the spherical case as described below. This spherical analogue is used in [NW92] to give a simulation method for finding critical probabilities for multiple-comparisons procedures. Definition 3.2.8 For any x ∈ d, where d denotes the unit d-sphere in
d+1 , and any r > 0 we use Bd∗ (x, r) to denote the spherical cap in d with center x and radius r ∈ (0, π/2), that is, Bd∗ (x, r) := {y ∈ d | δ ∗ (x, y) ≤ r}

18

3

Bonferroni Inequalities via Abstract Tubes

(a) 20 disks of equal radius.

(b) Voronoi subdivision and Delauney complex.

Figure 3.3: 20 disks of equal radius and their Delauney complex. where δ ∗ (x, y) := cos−1 x, y is the angular distance between x and y. With any finite V ⊆ d we associate the spherical Voronoi subdivision of d into regions Dv∗ := {x ∈ d | δ ∗ (x, v) = min δ ∗ (x, u)} u∈V

(v ∈ V )

and the spherical Delauney complex      ∗ ∗  (V ) := I ∈
(V )  Di∗ = ∅ . i∈I

Theorem 3.2.9 [NW92, NW97] Let V be a finite set of points in d and r ∈  (0, π/2). If v∈V Bd∗ (v, r) = ∅, then ({Bd∗ (v, r)}v∈V , ∗ (V )) is an abstract tube.

Proof. (Sketch) Let := ∗ (V ) and ω ∈ v∈V Bd∗ (v, r). We have to show that (ω), with Bd∗ (i, r) in place of Ai , is contractible. Similar to the preceding proof, (ω) is the nerve of C ∗ (ω)

:= {Di∗ | ω ∈ Bd∗ (i, r)}, whence the contractibil

ity of (ω) follows from that of C ∗ (ω). In fact, it can be shown that C ∗ (ω) contains every geodesic arc between any of its points and ω, whence it is contractible or equal to the sphere. The second alternative does not apply, since by the requirements the intersection of the spherical caps is empty. 2 Remark. Similar remarks as for the Euclidean case apply to the spherical case. In particular, if the centers of the spherical caps are in general position, that is, if no d + 1 centers lie in a d-dimensional subspace and no x ∈ d is equidistant from more than d + 1 of the centers, then the dimension of ∗ (V ) is at most d.

4 Abstract Tubes via Closure and Kernel Operators In this chapter, the fundamental theorems of abstract tube theory are applied in establishing improved inclusion-exclusion identities and Bonferroni inequalities. We do not state these improved identities and inequalities explicitly, since they can easily be read from Theorem 3.1.9, once an abstract tube is specified. The key concepts in this chapter are the concept of a closure operator as well as the concept of a kernel operator. In the first two sections we review the basic facts on closure operators and kernel operators, and show how these concepts give rise to new abstract tubes and thus to new improvements of the classical inclusion-exclusion principle and its associated truncation inequalities. In the third section we present some elementary alternative proofs (not using abstract tubes) and a generalization of one of the identities. One of the alternative proofs is based on Zeilberger’s abstract lace expansion [Zei97], which turns out to be a valuable tool in deriving new inclusion-exclusion identities. The last section of this chapter is devoted to a Bonferroni-type inequality inspired by one of our results on closure operators and abstract tubes. The main result of this section—the chordal graph sieve—associates a Bonferronitype inequality with any chordal graph G. By varying G, several well-known and new Bonferroni-type inequalities are obtained in a concise and unified way.

4.1 4.1.1

Abstract tubes via closure operators Closure operators

In what follows it is intuitive to imagine that c is the convex hull operator in
d . Here and elsewhere in this text,
(V ) is used to denote the power set of V . Definition 4.1.1 Let V be a set. A mapping c :
(V ) →
(V ) is called a K. Dohmen: LNM 1826, pp. 19–43, 2003. c Springer-Verlag Berlin Heidelberg 2003


20

4

Abstract Tubes via Closure and Kernel Operators

closure operator on V if for all subsets X and Y of V , (i) (ii) (iii)

X ⊆ c(X) (extensionality), X ⊆ Y ⇒ c(X) ⊆ c(Y ) (monotonicity), c(c(X)) = c(X) (idempotence).

A subset X of V is called c-closed if c(X) = X, and c-free if all subsets of X are c-closed. A minimal subset B of X such that c(B) = X is referred to as a c-basis of X. If there are no ambiguities, we supress c in these notions. The following proposition characterizes the free sets by means of their bases. Proposition 4.1.2 [Doh00b] Let c be a closure operator on some finite set V . Then, any subset J of V is free if and only if it is a basis of itself. Proof. We show that any set which is not free is not a basis of itself (the opposite direction is trivial). So assume that J is not free, that is, K ⊂ J for some non-closed set K. If J is not closed, then obviously it can’t be a basis of itself. Thus, we may assume that J is closed. For each k ∈ c(K) \ K we find that k ∈ c(K) = c(K \ {k}) ⊆ c(J) = J and hence, J ⊆ c(J \ {k} ∪ {k}) ⊆ c(c(J \ {k}) ∪ {k}) = c(c(J \ {k})) = c(J \ {k}) ⊆ c(J) = J. Therefore, k ∈ J and c(J \ {k}) = J, showing that J is not a basis of itself. 2 The following definition is due to Edelman and Jamison [EJ85]. Definition 4.1.3 A convex geometry is a pair (V, c) consisting of a finite set V and a closure operator c on V such that c(∅) = ∅ and such that any c-closed subset of V has a unique c-basis. Convex geometries are sometimes referred to as dual antimatroids (see e.g., [BZ92]). The most prominent example of a convex geometry is the following: Example 4.1.4 [EJ85] Let V be a finite subset of
d , and for any subset X = {x1 , . . . , xn } of V define c(X) := conv(X) ∩ V where  

n n     λi xi  λ1 , . . . , λn ≥ 0 and λi = 1 . (4.1) conv(X) :=  i=1

i=1

is the convex hull of X. By the Minkowski-Krein-Milman theorem the extremal points of conv(X) form a unique c-basis for any c-closed subset X of V . Thus, (V, c) is a convex geometry. Some further examples associated with graphs and semilattices follow. Example 4.1.5 [EJ85] For any subset X of some finite upper semilattice V let c(X) be the smallest upper subsemilattice of V containing X. Then, X is c-closed if and only if X is an upper subsemilattice, and X is c-free if and only

4.1 Abstract tubes via closure operators

21

if X is a chain. Since the set of join-irreducibles of X forms a unique c-basis of any c-closed subset X of V , we conclude that (V, c) is a convex geometry. In a similar way, any lower semilattice gives rise to a convex geometry. Example 4.1.6 [EJ85] For any tree G = (V, E) and any X ⊆ V let c(X) be the set of all z ∈ V lying on the path between x and y for some x, y ∈ X. Then, X is c-closed if and only if G[X] is a tree, and c-free if and only if X = {v, w} for some {v, w} ∈ E or X = {v} for some v ∈ V . Since the leaves of G[X] form a unique c-basis of any c-closed subset X of V , (V, c) is a convex geometry. Example 4.1.7 [EJ85] For any subset X of a connected block graph G = (V, E) let c(X) be the minimal superset Y ⊇ X such that G[Y ] is connected. Then, X is c-closed if and only if G[X] is connected, and c-free if and only if X is a clique. Since the simplicial vertices of G[X] constitute a unique c-basis of any c-closed subset X of V , we conclude that (V, c) is a convex geometry.

4.1.2

How closure operators lead to abstract tubes

The main result of this section is based on the following statement, which can be found as Exercise 8.23c in the textbook of Bj¨ orner and Ziegler [BZ92]. Proposition 4.1.8 [BZ92] Let (V, c) be a convex geometry. Then the abstract simplicial complex consisting of all non-empty c-free subsets of V is contractible. Proof. For a rigorous proof we refer to Edelman and Reiner [ER00]. 2 We are now ready to state the main result of this section. Theorem 4.1.9 [Doh03] Let (V, c) be a convex geometry, and let {Av }v∈V be a finite family of sets such that for any non-empty and non-closed subset X of V ,   (4.2) Ax ⊆ Av . x∈X

v ∈X /

  Then, {Av }v∈V , {I ∈
∗ (V ) | I c-free} is an abstract tube.

Proof. For ω ∈ v∈V Av let Vω := {v ∈ V | ω ∈ Av } , and for any I ⊆ V define cω (I) := c(I). Then, by (4.2), Vω is c-closed, whence (Vω , cω ) is a convex geometry. The contractibility of {I ∈
∗ (V ) | I c-free}(ω) follows from Proposition 4.1.8 and {I ∈
∗ (V ) | I c-free}(ω) = {I ∈
∗ (Vω ) | I cω -free}. 2 Remarks.  1. The requirements of Theorem 4.1.9 are already satisfied if x∈X Ax ⊆ Av for any non-empty subset X of V and any v ∈ c(X). 2. For c : X → X the abstract tube of Theorem 4.1.9 equals ({Av }v∈V ,
∗ (V )). In this case, the associated inequalities are the classical Bonferroni inequalities.

22

4

Abstract Tubes via Closure and Kernel Operators

3. The improved inclusion-exclusion identity associated with the abstract tube of Theorem 4.1.9 involves intersections of at most h(c) := max{|J| : J c-free} sets. A general result of Jamison [JW81] states that h(c) equals the Helly number of the family of c-closed subsets of V . Thus, h(c) is the smallest integer h such that any family of c-closed subsets of V whose intersection is empty has a subfamily of h or fewer sets whose intersection is also empty. 4. Let P be a probability measure on any σ-field containing the events Av , v ∈ V , such that for any non-empty and non-closed subset X of V ,  

     (4.3) P Ax > 0 and P Av  Ax = 1  x∈X

v ∈X /

x∈X

  Then, {Av }v∈V , {I ∈
∗ (V ) | I c-free} is a weak abstract tube in the sense of Definition 3.1.6. The proof is quite similar and left as an option to the reader. 5. For closure operators c and c on V define c ≤ c :⇔ c(I) ⊆ c (I) for any subset I of V , or equivalently, c ≤ c :⇔ all c -closed subsets of V are c-closed. Thus, a partial order on the set of closure operators on V is defined. The  significance of this partial order is as follows. c and c satisfy  If c ≤ c and both  ∗ the requirements then {Av }v∈V , {I ∈
(V ) | I c -free} is  of Theorem 4.1.9, a subtube of {Av }v∈V , {I ∈
∗ (V ) | I c-free} and hence, by Theorem 3.1.10, the improved Bonferroni inequalities associated with c are at least as sharp as those associated with c. In particular, since the closure operator I → I on V is largest with respect to ≤, the inequalities associated with the abstract tube of Theorem 4.1.9 are at least as sharp as the classical Bonferroni inequalities. We now deduce some results from Theorem 4.1.9. The first one is an abstract tube generalization of Narushima’s semilattice sieve [Nar74, Nar77]. Before stating this result, we review the following definition due to Alexandroff [Ale37]. Definition 4.1.10 The order complex of a finite partially ordered set V , (V ) for short, is the abstract simplicial complex of all non-empty chains of V . Now the abstract tube generalization of Narushima’s semilattice sieve [Nar74, Nar77] reads as follows. Corollary 4.1.11 [Doh99d] Let {Av }v∈V be a finite family of sets, where V is an upper semilattice such that Ax ∩ Ay ⊆ Ax∨y for any x, y ∈ V . Then, {Av }v∈V and the order complex of V constitute an abstract tube. Proof. Apply Theorem 4.1.9 in connection with Example 4.1.5. 2

4.1 Abstract tubes via closure operators

23

Remark. Corollary 4.1.11 gives the trivial abstract tube if V is a chain. The particular case where Ax ∩ Ay = Ax∨y for any x, y ∈ V is treated in [Doh01a]. The following result is due to Naiman and Wynn [NW92]: Corollary 4.1.12 [NW92] Let {Av }v∈V be a finite family of sets, where the indices form the vertices of a tree G = (V, E) such that Ax ∩ Ay ⊆ Az for any x, y ∈ V and any z on the path between x and y in G. Then, {Av }v∈V and the tree (considered as an abstract simplicial complex) constitute an abstract tube. Proof. Apply Theorem 4.1.9 in connection with Example 4.1.6. 2 Remark. The case where G is a path on n vertices gives rise to the identity
n  n n    χ Ai = χ(Ai ) − χ(Ai−1 ∩ Ai ) , i=1

i=1

i=2

which is valid for all finite sequences of sets A1 , . . . , An that satisfy Ai ∩Aj ⊆ Ak for i, j = 1, . . . , n and k = i, . . . , j. This latter consequence is again due to Naiman and Wynn [NW92] and explains the cancellations in Figure 1.1. Definition 4.1.13 The clique complex of a graph G is the abstract simplicial complex of all non-empty cliques of G. Since any tree is a connected chordal graph, the following corollary generalizes the preceding one. Corollary 4.1.14 [Doh03] Let {Av }v∈V be a finite family of sets, where the indices form the vertices of a connected chordal graph G = (V, E) such that Ax ∩ Ay ⊆ Az for any x, y ∈ V and any z on any chordless path between x and y. Then, {Av }v∈V and the clique complex of G form an abstract tube. Proof. For any X ⊆ V let c(X) be the set of all z ∈ V lying on a chordless path between x and y for some x, y ∈ X. Then, (V, c) is a convex geometry, where X is free if and only if X is a clique [EJ85, FJ86]. Now apply Theorem 4.1.9. 2 Remarks. 1. Corollary 4.1.14 gives the trivial abstract tube ({Av }v∈V ,
∗ (V )) if G is complete, since in this case all subsets of the vertex-set are cliques. 2. Howorka [How81] showed that in chordal graphs where all cycles of length five have at least three chords the chordless paths are precisely the shortest paths, and that chordal graphs with this property are unions of Ptolemaic graphs and vice versa. Ptolemaic graphs are defined below. For a discussion of this class of graphs and other restricted classes of chordal graphs in connection with convex geometries, the reader is referred to Farber and Jamison [FJ86].

24

4

Abstract Tubes via Closure and Kernel Operators

Definition 4.1.15 A graph G is called Ptolemaic if it is connected and if for any four vertices v, w, x, y of G, d(v, w)d(x, y) ≤ d(v, x)d(w, y) + d(w, x)d(v, y) where d(a, b) denotes the length of a shortest path between a and b in G. Corollary 4.1.16 Let {Av }v∈V be a finite family of sets, where the indices form the vertices of a Ptolemaic graph G = (V, E) such that Ax ∩ Ay ⊆ Az for any x, y ∈ V and any z ∈ V on any shortest path between x and y in G. Then, {Av }v∈V and the clique complex of G constitute an abstract tube. Proof. Corollary 4.1.16 follows from Corollary 4.1.14 and the preceding discussion on Ptolemaic graphs. 2 We proceed with deducing some further consequences of Theorem 4.1.9. Corollary 4.1.17 [Doh03] Let {Av }v∈V be a finite family of sets such that for any non-empty X ⊆ V there is a unique minimal non-empty Y ⊆ X such that   (4.4) Ax = Ay . x∈X

y∈Y

  Then, {Av }v∈V and the set of all I ∈
∗ (V ) satisfying i∈I Ai = j∈J Aj for all non-empty proper subsets and supersets J of I constitute an abstract tube.    Proof. For any subset X of V define c(X) = v ∈ V  x∈X Ax ⊆ Av if X = ∅, and c(∅) := ∅. Then, c is a closure operator on V . In order to verify that each c-closed subset of V has a unique c-basis, assume that X is non-empty and cclosed. By the above requirement there is a unique minimal non-empty Y ⊆ X satisfying (4.4). By definition of c, c(X) = c(Y ) ⊆ X and hence, c(Y ) = X. In  order to show that Y is smallest with c(Y ) = X, suppose ) = X for  that c(Y   some other non-empty subset Y of X. By definition of c, x∈X Ax = y∈Y  Ay , whence by the choice of Y , Y  ⊇ Y . Consequently, Y is the unique c-basis of X. The corollary now follows from Theorem 4.1.9 and Proposition 4.1.2. 2 Definition 4.1.18 With any finite set V and any system  of non-empty subsets of V we associate the abstract simplicial complex

(V, ) := {I ∈
∗ (V ) | I ⊇ X for any X ∈ } , consisting of all non-empty subsets of V not including any X ∈  as a subset. Corollary 4.1.19 [Doh99d] Let {Av }v∈V be a finite family of sets where V is linearly ordered. Furthermore, let  ⊆
∗ (V ) such that for any X ∈ ,  Ax ⊆ Av for some v > max X. (4.5) x∈X

Then, ({Av }v∈V , (V, )) is an abstract tube.

4.2 Abstract tubes via kernel operators

25

Proof. For any X ∈  choose vX such that (4.5) is satisfied for v = vX . For any subset I of V define c1 (I) := c(I) := I ∪ {vX | X ∈ , X ⊆ I}, cn (I) := ∞ cn−1 (c(I)) (n ≥ 2), and c∗ (I) := n=1 cn (I). Then, c∗ is a closure operator on V , and it is easily seen that I \ {vX | X ∈ , X ⊆ I} is the unique c∗ -basis of any c∗ -closed subset I of V . Hence, (V, c∗ ) is a convex geometry. In this convex geometry, any subset I of V is c∗ -free if and only if I ⊇ X for any X ∈ . The result now follows by applying Theorem 4.1.9 with c∗ instead of c. 2 Remark. Note that the corollary gives the trivial abstract tube if  is empty, and that the corollary can be dualized by replacing v > max X with v < min X. As a consequence of Corollary 4.1.19 we deduce the following generalization of Corollary 4.1.11. Corollary 4.1.20 [Doh99d] Let {Av }v∈V be a finite family of sets, where V is endowed with a partial ordering relation such that for any x, y ∈ V , Ax ∩Ay ⊆ Az for some upper bound z of x and y. Then, {Av }v∈V and the order complex of V constitute an abstract tube. Proof. The result follows from Corollary 4.1.19 by setting  := {{x, y} ⊆ V | x, y incomparable} and considering some arbitrary linear extension. 2 Remark. The inclusion-exclusion identity associated with the abstract tube of Corollary 4.1.20 is due to Narushima [Nar82]. Note, however, that the original requirements of Narushima [Nar82] are slightly stronger in that he requires that for any x, y ∈ V , Ax ∩ Ay ⊆ Az for some minimal upper bound z of x and y.

4.2

Abstract tubes via kernel operators

Similar results as for closure operators are now established for kernel operators. There is, however, no duality between the results of the preceding section and the results of the present section.

4.2.1

Kernel operators

Definition 4.2.1 Let V be a set. A mapping k :
(V ) →
(V ) is called a kernel operator on V if for all subsets X and Y of V , (i) k(X) ⊆ X

(intensionality),

(ii) X ⊆ Y ⇒ k(X) ⊆ k(Y ) (monotonicity), (iii) k(k(X)) = k(X) (idempotence). A subset X of V is called k-open if k(X) = X.

4

26

Abstract Tubes via Closure and Kernel Operators

There is a well-known correspondence between kernel operators on V and union-closed subsets of the power set of V . Recall that a subset  of the power set of V is called union-closed if X ∪ Y ∈  for any X, Y ∈ . For a proof of the following proposition, the reader is referred to the textbook of Ern´e [Ern82]. Proposition 4.2.2 [Ern82] Let V be a finite set. If k is a kernel operator on V , then {X ⊆ V | X k-open} is union-closed. If  ⊆
(V ) is union-closed, then  I → {X ∈  | X ⊆ I} (I ⊆ V ) defines a kernel operator k on V such that X is k-open if and only if X ∈ .

4.2.2

How kernel operators lead to abstract tubes

Our main result on kernel operators is as follows. Theorem 4.2.3 [Doh00a] Let {Av }v∈V be a finite family of sets, and let k be a kernel operator on V such that for any non-empty and k-open subset X of V ,   Ax ⊆ Av . x∈X

v ∈X /

  Then, {Av }v∈V , {I ∈
∗ (V ) | k(I) = ∅} is an abstract tube. The proof of Theorem 4.2.3 is based on the following proposition: Proposition 4.2.4 [Doh00a] Let V be a finite set, and let k be a kernel operator on V . Then, the complex {I ∈
∗ (V ) | k(I) = ∅} is contractible or V is k-open. Proof. Assume that V is not k-open, and let v ∈ V \ k(V ). Then, for any I ⊆ V , v∈ / k (I ∪ {v}), and hence the implication k(I) = ∅ ⇒ k (I ∪ {v}) = ∅ holds for any I ⊆ V . From this we conclude that v is contained in every maximal face of := {I ∈
∗ (V ) | k(I) = ∅}, whence is a cone and hence contractible. (Recall from topology that a cone is an abstract simplicial complex having a vertex v which is contained in every maximal face of . The geometric realization (2.2) of each such complex is contractible by means of (x, t) → (1 − t)x + teπv .) 2 Proof of Theorem 4.2.3. To obtain a contradiction, assume there is some ω ∈ v∈V Av such that {I ∈
∗ (V ) | k(I) = ∅}(ω) is not contractible. From {I ∈
∗ (V ) | k(I) = ∅}(ω) = {I ∈
∗ (Vω ) | k(I) = ∅} , where Vω := {v ∈ V | ω ∈ Av }, the assumption and Proposition 4.2.4 we conclude that Vω is k-open. On the contrary, the definition of Vω and the requirements imply that Vω is not k-open. Thus, Theorem 4.2.3 is proved. 2 Remarks. The following remarks relate to Theorem 4.2.3 above.

4.2 Abstract tubes via kernel operators

27

1. By setting k(X) := ∅ for any subset X of V , the abstract tube of Theorem 4.2.3 specializes to the trivial abstract tube. The associated Bonferroni inequalities are the classical  Bonferroni inequalities.  2. One can prove that {Av }v∈V , {I ∈
∗ (V ) | k(I) = ∅} is a weak abstract tube with respect to any probability measure P on the σ-field generated by {Av }v∈V such that  (4.3) holds for any non-empty and  k-open subset X of V . 3. Note that {Av }v∈V , {I ∈
∗ (V ) | k  (I) = ∅} is a subtube of {Av }v∈V , {I ∈
∗ (V ) | k(I) = ∅} if k and k  are as in Theorem 4.2.3 and k  ≤ k, where (4.6)

k  ≤ k :⇔ k(I) ⊆ k  (I) for any subset I of V

or equivalently, (4.7)

k  ≤ k :⇔ all k-open subsets of V are k  -open.

By this and Theorem 3.1.10, it follows that the improved Bonferroni inequalities associated with k  are at least as sharp as those associated with k if k  ≤ k. In particular, since the kernel operator I → ∅ on V is largest with respect to ≤, the improved inequalities are at least as sharp as their classical counterparts. The correspondence between kernel operators and union-closed sets leads to the following equivalent formulation of Theorem 4.2.3. Recall from Definition 4.1.18 that (V, ) := {I ∈
∗ (V ) | I ⊇ X for any X ∈ }. Theorem 4.2.5 [Doh00a] Let {Av }v∈V a finite family of sets, and let  be a union-closed set of non-empty subsets of V such that for any X ∈ ,   Ax ⊆ Av . x∈X

v ∈X /

Then, ({Av }v∈V , (V, )) is an abstract tube. Proof. The result follows from Proposition 4.2.2 and Theorem 4.2.3. 2 Corollary 4.2.6 [Doh00a] Let {Av }v∈V be a finite family of sets, c a closure operator on V and  a set of non-empty subsets of V such that {c(X) | X ∈ } is a chain and such that for any X ∈ ,   Ax ⊆ Av . (4.8) x∈X

v ∈c(X) /

Then, ({Av }v∈V , (V, )) is an abstract tube.

Proof. Let  := {  | ∅ =  ⊆ }. Then,  is union-closed, and for any Y ∈ ,   Y = {X | X ∈ , X ⊆ Y } ⊆ {c(X) | X ∈ , X ⊆ Y } = c(Z)

28

4

Abstract Tubes via Closure and Kernel Operators

for some Z ∈ , Z ⊆ Y . Hence, Z ⊆ Y ⊆ c(Z) and therefore,     Ax ⊆ Ax ⊆ Av ⊆ Av . x∈Y

x∈Z

v ∈c(Z) /

v ∈Y /

The result now follows by applying Theorem 4.2.5 with  instead of . 2 The following result generalizes Corollary 4.1.19. We refer to [Doh99a] for an exemplary application of the associated inclusion-exclusion identity to counting arrangements with forbidden positions on a chess-like board. Corollary 4.2.7 [Doh99d] Let {Av }v∈V be a finite family of sets, where V is a linearly ordered set, and let  ⊆
∗ (V ) such that for any X ∈ ,   (4.9) Ax ⊆ Av . x∈X

v>max X

Then, ({Av }v∈V , (V, )) is an abstract tube. Proof. Corollary 4.2.7 follows from Corollary 4.2.6 by using the closure operator X → c(X) where c(X) := {v ∈ V | v ≤ max X} if X = ∅, and c(∅) := ∅. 2

4.3

Alternative proofs

By virtue of Theorem 3.1.9 each abstract tube of the preceding section gives rise to an improved inclusion-exclusion identity and a series of improved Bonferroni inequalities. In this section, we present some elementary alternative proofs for some of these identities and inequalities without making use of abstract tubes.

4.3.1

Improved identities via closure operators

We start with an alternative proof of the inclusion-exclusion identity associated with the abstract tube of Theorem 4.1.9 by using Zeilberger’s abstract lace expansion [Zei97]. We then proceed by generalizing this inclusion-exclusion identity to arbitrary closure operators. Alternative proof by Zeilberger’s abstract lace expansion The following definition is due to Zeilberger [Zei97]. Definition 4.3.1 [Zei97] Let V be a set. A lace map on V is a mapping l from the power set of V into itself such that for any subsets X and Y of V , (i) l(X) ⊆ X, (ii) l(X) ⊆ Y ⊆ X ⇒ l(X) = l(Y ), (iii)

l(X) = l(Y ) ⇒ l(X ∪ Y ) = l(X).

If l is a lace map on V , then a subset X of V is called an l-lace if l(X) = X.

4.3 Alternative proofs

29

We now state Zeilberger’s abstract lace expansion [Zei97]. Theorem 4.3.2 [Zei97] Let {Av }v∈V be a finite family of sets, and let l be a lace map on V . Then,  
     χ Av = (−1)|I|−1 χ  Ai ∩ Ak  . (4.10)


I∈ ∗ (V ) I is an l-lace

v∈V

k∈Cl (I)

i∈I

where Cl (I) := {v ∈ V \ I | l(I ∪ {v}) = I} Proof. Applying the inclusion-exclusion principle to the left-hand side we obtain  

   |I|−1 Av = (−1) χ Ai . χ I∈
∗ (V )

v∈V

i∈I

From this we conclude that   
    χ Av = (−1)|J|−1 χ  Aj 















I∈ ∗ (V ) J∈ ∗ (V ) I is an l-lace l(J)=I

v∈V

=

j∈J



(−1)|I|+|J\I|−1 χ 

I∈ ∗ (V ) J∈ ∗ (V ) I is an l-lace l(J)=I

=






I∈ ∗ (V ) I is an l-lace

=




|I|−1

(−1)




χ

I∈ ∗ (V ) I is an l-lace

=




|I|−1

(−1)




I∈ ∗ (V ) I is an l-lace

Ai ∩

i∈I

(−1)|I|−1 χ






χ



 Ai

i∈I



 i∈I




 Ai

i∈I

 Ai



 j∈J\I

 Aj  



(−1)|J\I|−1 χ 

J∈ ∗ (V ) l(J)=I




|K|−1

χ

(−1)

K⊆V \I l(I∪K)=I




|K|−1

(−1)

χ

K⊆Cl (I)

 Aj 

j∈J\I



Ak

k∈K





 Ak

.

k∈K

Now, a second application of the inclusion-exclusion principle gives 
  
    χ Av = (−1)|I|−1 χ Ai χ  Ak  , v∈V




I∈ ∗ (V ) I is an l-lace

i∈I

k∈Cl (I)

from which (4.10) immediately follows. 2 In the following, we show that any convex geometry gives rise to a lace map and thus deduce the inclusion-exclusion identity associated with the abstract tube of Theorem 4.1.9 from Zeilberger’s abstract lace expansion.

30

4

Abstract Tubes via Closure and Kernel Operators

Theorem 4.3.3 [Doh02a] Let (V, c) be a convex geometry. For any subset X of V define l(X) := unique c-basis of c(X).

(4.11) Then, l is a lace map.

Proof. We show that l satisfies (i)–(iii) in the definition of a lace map. (i) Clearly, l(X) ⊆ X. (ii) If l(X) ⊆ Y ⊆ X, then c(l(X)) ⊆ c(Y ) ⊆ c(X) by the monotonicity of c. From this and c(l(X)) = c(X) we conclude that c(X) = c(Y ) and hence, l(X) = l(Y ). (iii) From l(X) = l(Y ) we get c(l(X)) = c(l(Y )) and hence, c(X) = c(Y ). Since c is an extensional mapping, X ⊆ c(X) and Y ⊆ c(Y ). Hence, X ∪ Y ⊆ c(X) ∪ c(Y ) = c(X). By this, the monotonicity and idempotence of c we find c(X ∪ Y ) = c(X) and hence, l(X ∪ Y ) = l(X). 2 We are now ready to prove the inclusion-exclusion identity
 
   |I|−1 (4.12) χ Av = (−1) χ Ai


I∈ ∗ (V ) I c-free

v∈V

i∈I

associated with the abstract tube of Theorem 4.1.9 under the requirements that {Av }v∈V is a finite family of sets, (V, c) is a convex geometry, and   (4.13) Ax ⊆ Av x∈X

v∈c(X)\X

for any non-empty and non-closed subset X of V . Note that (4.13) is more restrictive than the corresponding condition (4.2) of Theorem 4.1.9. Proof of (4.12). For any subset X of V let l(X) be the unique c-basis of c(X) as in (4.11). Then, by Theorem 4.3.3 l is a lace map. We observe that for any subset I ⊆ V , I is an l-lace if and only if I is a c-basis, and that for any l-lace I ⊆ V , Cl (I) = c(I) \ I. Therefore, by applying Theorem 4.3.2 we obtain   
    χ Av = (−1)|I|−1 χ  Ai ∩ Ak  .


I∈ ∗ (V ) I is a c-basis

v∈V

i∈I

k∈c(I)\I

Now, if I  is not a c-basis of itself, then it is not c-closed and hence by (4.13),  A ∩ i i∈I k∈c(I)\I Ak = ∅. On the other hand, if I is a c-basis of itself, then it    is c-closed and therefore, i∈I Ai ∩ k∈c(I)\I Ak = i∈I Ai . Hence,  

   |I|−1 Av = (−1) χ Ai . χ v∈V




I∈ ∗ (V ) I is a c-basis of itself

i∈I

From this, our claim (4.12) immediately follows since by virtue of Proposition 4.1.2, I is a c-basis of itself if and only if I is c-free. 2

4.3 Alternative proofs

31

Alternative proof for arbitrary closure operators We now establish a generalization of the identity associated with the abstract tube of Theorem 4.1.9 (see (4.12)) for closure operators which not necessarily have the unique basis property. To this end, we need the following definition. Definition 4.3.4 Let V be a finite set and c a closure operator on V . A cformation of a subset X of V is any non-empty set of c-bases of X such that

= X. A c-formation of X is odd resp. even if | | is odd resp. even. The c-domination of X, domc (X), is the number of odd c-formations of X minus the number of even c-formations of X. Evidently, if X is not c-closed, then domc (X) = 0, and if X is c-free, then domc (X) = 1. If (V, c) is a convex geometry, then domc (X) = 1 resp. 0 depending on whether X is c-free or not. The following proposition generalizes a result of Edelman and Jamison [EJ85] on convex geometries and a result of Narushima [Nar74, Nar77] on semilattices. Proposition 4.3.5 [Doh99b] Let V be a finite set and c a closure operator on V . Then, for any c-closed subset J of V ,  (−1)|I| = (−1)|J| domc (J) . I⊆J c(I)=J

Proof. Let J0 , . . . , Jn be the distinct bases of J. Evidently, c(I) = J if and only if Jk ⊆ I ⊆ J for some k ∈ {0, . . . , n}. Thus, by including and excluding terms,     (−1)|I| = (−1)||−1 (−1)|I| = (−1)||−1 (−1)|J| . Ë I⊆J ⊆{J0 ,...,Jn }  ⊆{J ,...,J n} 0 I : ⊆I⊆J c(I)=J =∅ =∅, Ë=J The result now follows from the definition of the c-domination domc (J). 2 The following proposition will be used later to derive the main result of this subsection. It may have applications not only to inclusion-exclusion. Proposition 4.3.6 [Doh99b] Let V be a finite set, c a closure operator on V and g a mapping from
(V ) into an abelian group such that g = g ◦ c. Then,   (−1)|I| g(I) = (−1)|J| domc (J) g(J) . J⊆V J closed

I⊆V

Proof. Since g(I) = g(c(I)) for any subset I of V we find that     (−1)|I| g(I) = (−1)|I| g(c(I)) = (−1)|I| g(J) . I⊆V

I⊆V

J⊆V I⊆J c(J)=J c(I)=J

By Proposition 4.3.5, the inner sum on the right coincides with (−1)|J| domc (J), which gives the result. 2

4

32

Abstract Tubes via Closure and Kernel Operators

Remarks. In the following remarks, it is assumed that (V, c) is a convex geometry. Thus, domc (X) = 1 if X is c-free, and domc (X) = 0 otherwise. 1. By defining g(I) := 1 for any I ⊆ V where V = ∅, Proposition 4.3.6 gives 

(4.14)

(−1)|J| = 0 ,

J⊆V J free

which is an unpublished result due to Lawrence (cf. [EJ85]). , Proposition 4.3.6 shows that 2. By defining g(I) := (−1)|c(I)| for any I ⊆ V in each convex geometry (V, c) there are exactly I⊆V (−1)|c(I)\I| free sets. 3. By defining g(I) := |c(I)| for any I ⊆ V , Proposition 4.3.6 specializes to (4.15)



(−1)|I| |c(I)| =



(−1)|J| |J| ,

J⊆V J free

I⊆V

which in turn specializes to a result of Gordon [Gor97] if c is derived from the convex hull operator in
d as in Example 4.1.4. Edelman and Reiner [ER00] showed that in this case either side of (4.15) agrees in absolute value with the number of points in V which are in the interior of conv(V ) if |V | > 1. We continue with a further preliminary proposition. Proposition 4.3.7 [Doh99b] Let {Av }v∈V be a finite family of sets, and let c be a closure operator on V such that for any non-empty and non-closed subset X of V ,   (4.16) Ax ⊆ Av . x∈X

v ∈X /

Then, for any non-empty subset I of V ,   Ai = Ai . i∈I

i∈c(I)

 show that Proof.  Let I ⊆ V , I = ∅. If i∈I Ai = ∅, we are done. Otherwise  ω ∈ i∈c(I) Ai for each ω ∈ i∈I Ai . Evidently, for any ω ∈ i∈I Ai , I ⊆ Vω where again Vω is defined as {v ∈ V | ω ∈ Av }. From  (4.16) we conclude that Vω is closed and hence, c(I) ⊆ Vω or equivalently, ω ∈ i∈c(I) Ai . 2 We now state and prove the announced generalization of (4.12). Theorem 4.3.8 [Doh99b] Let {Av }v∈V be a finite family of sets and c a closure operator on V such that for any non-empty and non-closed subset X of V ,   Ax ⊆ Av x∈X

v ∈X /

4.3 Alternative proofs

33

and such that c(∅) = ∅. Then,   
   Av = (−1)|J|−1 domc (J) χ  Aj  . χ


J∈ ∗ (V ) J closed

v∈V

j∈J

 

Proof. By the classical inclusion-exclusion identity (1.1) we have χ v∈V Av =   |I|−1 g(I) where g(I) := χ i∈I Ai if I = ∅ and g(∅) := 0. By I⊆V (−1) Proposition 4.3.7 and since c(∅) = ∅ we find that g = g ◦ c. Now, by applying Proposition 4.3.6, the statement of the theorem immediately follows. 2

4.3.2

Improved inequalities via kernel operators

In the following, we present an elementary proof of the improved Bonferroni inequalities associated with the abstract tube of Theorem 4.2.3 as well as an elementary proof of the fact that these inequalities are at least as sharp as their classical counterparts. The proof given below is new even in the traditional case where k(I) = ∅ for any subset I of V . In this case, it generalizes Garsia and Milne’s bijective proof of the classical inclusion-exclusion principle [GM81, Zei84, Pau86] to an “injective proof” of the classical Bonferroni inequalities. Theorem 4.3.9 [Doh02c] Let {Av }v∈V be a finite family of sets and k a kernel operator on V such that for any non-empty and k-open subset X of V ,   Ax ⊆ Av . v ∈X /

x∈X

Then, for any r ∈ , 
 χ Av

≥

χ







≤

Av

(−1)|I|−1 χ




I∈ ∗ (V ) k(I)=∅ |I|≤r

v∈V









 Ai

(r even),

i∈I


|I|−1

(−1)




χ

I∈ ∗ (V ) k(I)=∅ |I|≤r

v∈V





 Ai

(r odd).

i∈I

Proof. Obviously, the preceding two inequalities are equivalent to   


     (4.17) χ Av + χ Ai ≥ χ Ai


I∈ ∗ (V ) k(I)=∅ |I|≤r |I| even

v∈V


(4.18) χ

 v∈V

 Av

+




I∈ ∗ (V ) k(I)=∅ |I|≤r |I| even




I∈ ∗ (V ) k(I)=∅ |I|≤r |I| odd

i∈I


χ

 i∈I

 Ai

≤




I∈ ∗ (V ) k(I)=∅ |I|≤r |I| odd

(r even),

i∈I


χ

 i∈I

 Ai

(r odd).

4

34

Abstract Tubes via Closure and Kernel Operators

For any ω ∈



Av define Vω := {v ∈ V | ω ∈ Av }, and for any r ∈ ,    r (ω) := I ∈
(Vω )  k(I) = ∅, |I| ≤ r, |I| even ,   

r (ω) := I ∈
(Vω )  k(I) = ∅, |I| ≤ r, |I| odd . v∈V

To prove (4.17) and (4.18), it suffices to show that |r (ω)| ≥ | r (ω)| if r is even, and |r (ω)| ≤ | r (ω)| if r is odd. Evidently, Vω is not k-open, so some v ∈ Vω \ k(Vω ) exists. For any I ⊆ Vω it follows that v ∈ / k(I ∪ {v}) since otherwise v ∈ k(I ∪ {v}) ⊆ k(Vω ∪ {v}) = k(Vω ), contradicting v ∈ / k(Vω ). From v∈ / k(I ∪ {v}) and k(I ∪ {v}) ⊆ I ∪ {v} we conclude that k(I ∪ {v}) ⊆ I and hence, k(I ∪ {v}) ⊆ k(I). Therefore, k(I ∪ {v}) = ∅ ⇔ k(I) = ∅. By this, I → I
{v}, where
denotes symmetric difference, is an injective mapping from r (ω) into r (ω) if r is even and from r (ω) into r (ω) if r is odd. 2 Theorem 4.3.10 [Doh02c] Let {Av }v∈V be a finite family of sets, and let k and k  be kernel operators on V such that k  ≤ k with respect to (4.6) or (4.7) and such that for any non-empty and k  -open subset X of V ,   Ax ⊆ Av . v ∈X /

x∈X

Then, for any r ∈ , 
  |I|−1 (−1) χ Ai


I∈ ∗ (V ) k (I)=∅ |I|≤r




≥

(−1)|I|−1 χ

I∈ ∗ (V ) k (I)=∅ |I|≤r





(−1)






≤

Ai


|I|−1

χ

I∈ ∗ (V ) k(I)=∅ |I|≤r

i∈I









 Ai

(r even),

i∈I


(−1)|I|−1 χ

I∈ ∗ (V ) k(I)=∅ |I|≤r

i∈I





 Ai

(r odd).

i∈I

Proof. Since k  ≤ k the preceding two inequalities are equivalent to  

    (4.19) ≥ (r even), χ Ai χ Ai


I∈ ∗ (V ) k(I)=∅ k (I)=∅ |I|≤r |I| even



(4.20)




For any ω ∈



I∈ ∗ (V ) k(I)=∅ k (I)=∅ |I|≤r |I| even




I∈ ∗ (V ) k(I)=∅ k (I)=∅ |I|≤r |I| odd

i∈I


χ

 i∈I

 Ai

≤




I∈ ∗ (V ) k(I)=∅ k (I)=∅ |I|≤r |I| odd

i∈I


χ



 Ai

(r odd).

i∈I

Av define Vω as above, and for any r ∈ define    ∗r (ω) := I ∈
∗ (Vω )  k(I) = ∅, k  (I) = ∅, |I| ≤ r, |I| even ,   

∗r (ω) := I ∈
∗ (Vω )  k(I) = ∅, k  (I) = ∅, |I| ≤ r, |I| odd . v∈V

4.3 Alternative proofs

35

To prove (4.19) and (4.20), we show that |∗r (ω)| ≥ | ∗r (ω)| if r is even, and |∗r (ω)| ≤ | ∗r (ω)| if r is odd. Since Vω is not k  -open, some v ∈ Vω \ k  (Vω ) exists, and since k  ≤ k, v ∈ Vω \ k(Vω ). Similar to the preceding proof, we conclude that k(I
{v}) = ∅ ⇔ k(I) = ∅ and k  (I
{v}) = ∅ ⇔ k  (I) = ∅. Hence, I → I
{v} is an injective mapping from ∗r (ω) into ∗r (ω) if r is even and from ∗r (ω) into ∗r (ω) if r is odd, thus establishing our claim. 2

4.3.3

Additional proofs of identities

We finally give two self-contained proofs of the inclusion-exclusion identities associated with the abstract tubes of Corollary 4.2.6 and Corollary 4.2.7. These self-contained proofs only require the traditional form of the inclusion-exclusion principle. The presentation here closely follows [Doh00c] and [Doh99a]. Our first proof makes use of a partial ordering relation, which we introduce in the following lemma. Lemma 4.3.11 Let c be a closure operator on some set V . Then, by A ≤c B :⇔ c(A) ∩ B = A

(A, B ⊆ V )

a partial ordering relation on V is defined. Proof. i) ≤c is reflexive: A ⊆ c(A) ⇒ c(A) ∩ A = A ⇒ A ≤c A. ii) ≤c is antisymmetric: (A ≤c B) ∧ (B ≤c A) ⇒ (c(A) ∩ B = A) ∧ (c(B) ∩ A = B) ⇒ A = c(A) ∩ B = c(A) ∩ c(B) ∩ A = c(B) ∩ A = B. iii) ≤c is transitive: (A ≤c B) ∧ (B ≤c C) ⇒ (c(A) ∩ B = A) ∧ (c(B) ∩ C = B) ⇒ c(A) ∩ C = c(c(A) ∩ B) ∩ C ⊆ c(A) ∩ c(B) ∩ C = c(A) ∩ B = A ⇒ A ≤c C. 2 Alternative proof of the inclusion-exclusion identity associated with the abstract tube of Corollary 4.2.6. Define  := {I ∈
∗ (V ) | I ⊇ X for some X ∈ }. Then, by the classical inclusion-exclusion principle, it suffices to prove that 
  |I|−1 (4.21) (−1) χ Ai = 0 . I∈

i∈I

We first establish that for any Y1 , Y2 ∈  which are ≤c -minimal in , (4.22)

{I ⊆ V | I ≥c Y1 } ∩ {I ⊆ V | I ≥c Y2 } = ∅ ⇒ Y1 = Y2 .

Let I ≥c Y1 and I ≥c Y2 . Then, Y1 = I ∩ c(Y1 ), Y2 = I ∩ c(Y2 ), and hence, c(c(Y1 ) ∩ Y2 ) ∩ Y1 ⊆ Y1 ∩ c(Y2 ) = I ∩ c(Y1 ) ∩ c(Y2 ) = c(Y1 ) ∩ Y2 = c(Y1 ) ∩ Y2 ∩ I ∩ c(Y1 ) ⊆ c(c(Y1 ) ∩ Y2 ) ∩ Y1 . So c(c(Y1 ) ∩ Y2 ) ∩ Y1 = c(Y1 ) ∩ Y2 , which is equivalent to c(Y1 ) ∩ Y2 ≤c Y1 by definition of ≤c . Since Y1 , Y2 ∈ , there are X1 , X2 ∈  satisfying Y1 ⊇ X1 and Y2 ⊇ X2 . By the requirements of Corollary 4.2.6  is a chain, whence we may assume that c(X1 ) ⊇ c(X2 ). Thus,

36

4

Abstract Tubes via Closure and Kernel Operators

c(Y1 ) ∩ Y2 ⊇ c(X1 ) ∩ X2 ⊇ c(X2 ) ∩ X2 = X2 ∈  and hence, c(Y1 ) ∩ Y2 ∈ . By this and the ≤c -minimality of Y1 in , c(Y1 )∩Y2 = Y1 and hence, Y1 ≤c Y2 . Now, by the ≤c -minimality of Y2 in  we conclude that Y1 = Y2 , thus establishing the implication in (4.22). By virtue of (4.22), our claim (4.21) is proved if 
  |I|−1 (4.23) (−1) χ Ai = 0 I≥c Y

i∈I

for each ≤c -minimal Y ∈ . By definition of ≤c we have   
     . (−1)|I|−1 χ Ai = (−1)|I|−1 χ  Ay ∩ Ai  I≥c Y

I⊆V \c(Y )

i∈I

y∈Y

i∈I

. where a = b means that a = b or a = −b. By the classical principle of inclusionexclusion we may rewrite the preceding equation as     
     . (−1)|I|−1 χ Ai = χ  Ay ∩ Ai  − χ  Ay  . I≥c Y

i∈I

i∈c(Y / )

y∈Y

y∈Y

Now, (4.23) (and hence our claim (4.21)) is proved if   Ay ⊆ Ai . (4.24) y∈Y

i∈c(Y / )

Since Y ∈ , there is some X ∈  such that Y ⊇ X and hence, c(X) ∩ Y ∈ . It is easy to see that c(c(X) ∩ Y ) ∩ Y = c(X) ∩ Y , whence c(X) ∩ Y ≤c Y and thus c(X) ∩ Y = Y , which is implied by the ≤c -minimality of Y in . Hence, c(Y ) ⊆ c(X) (in fact: c(Y ) = c(X)). This in combination with Y ⊇ X gives     Ay ⊆ Ax ⊆ Ai ⊆ Ai , y∈Y

x∈X

i∈c(X) /

i∈c(Y / )

where the second inclusion is justified by requirement (4.8) of Corollary 4.2.6. Thus, (4.24) holds, and the proof is complete. 2 Alternative proof of the inclusion-exclusion identity associated with the abstract tube of Corollary 4.2.7. Again, it suffices to prove (4.21). The idea is to define a partition of  such that for any  ∈ , 
  (4.25) (−1)|I|−1 χ Ai = 0 . I∈

i∈I

For any X ∈  define X + := {v ∈ V | v > max X}. Without loss of generality we may assume that X + = ∅ for any X ∈ . For any M, N ∈
∗ (V ) define

4.4 The chordal graph sieve

37

M  N if max M < max N or M = N , and let ∗ be a linear # extension of .$ For any I ∈  define XI := min ∗ {X ∈  | X ⊆ I} and [I] := I \ XI+ , I ∪ XI+ . We now show that := {[I] | I ∈ } is a partition of . Obviously, I ∈ [I] ⊆  for any I ∈  and hence,  = I∈ [I]. It remains to show that I ∈ [J] if J ∈ [I]. For J ∈ [I], XJ ∗ XI since J ⊇ XI . By this, max XJ ≤ max XI < min XI+ and therefore, XJ ∩ XI+ = ∅. We conclude that XJ = XJ \ XI+ ⊆ J \ XI+ ⊆ (I ∪ XI+ ) \ XI+ ⊆ I and hence, XI ∗ XJ . From this and XJ ∗ XI it follows that XJ = XI and hence, XJ+ = XI+ . Therefore, J \ XJ+ ⊆ (I ∪ XI+ ) \ XI+ ⊆ I ⊆ (I \ XI+ ) ∪ XI+ ⊆ J ∪ XJ+ , whence I ∈ [J]. We now prove that (4.25) holds for any  ∈ . For  = [J] we have   
     . (−1)|I|−1 χ Ai = (−1)|I| χ  Aj ∩ Ai  I∈

I⊆XJ+

i∈I

j∈J\XJ+

i∈I

. where, as in the preceding proof, = means equality up to sign. By applying the classical inclusion-exclusion principle to the right-hand side we obtain   
    . (−1)|I|−1 χ Ai = χ  Aj ∩ Ax  . I∈

i∈I

j∈J\XJ+

x∈XJ+

From J \ XJ+ ⊇ XJ and precondition (4.9) we conclude that         χ Aj ∩ Ax  ≤ χ  Ax ∩ Ax  = 0 . j∈J\XJ+

x∈XJ+

x∈XJ

x∈XJ+

Now, (4.25) immediately follows from the preceding two equations. 2

4.4

The chordal graph sieve

In Corollary 4.1.14 we saw how chordal graphs give rise to abstract tubes and thus to improved Bonferroni inequalities if some structural requirement on the collection of events applies. In the following theorem this requirement is dropped. The result is a generally-valid inequality where the selection of the intersections in the estimates is determined by a graph. In the literature, such inequalities are referred to as graph sieves [GS96a, GS96b]. The first graph sieves were obtained by R´enyi [R´en61] and Galambos [Gal66, Gal72], followed by Hunter [Hun76], Worsley [Wor82], Galambos and Simonelli [GS96a, GS96b], McKee [McK97, McK98] and Buksz´ar and Pr´ekopa [BP01]. The hypertree sieve of Tomescu [Tom86] and related results by Grable [Gra93, Gra94] and Buksz´ar [Buk01] fall into the category of hypergraph sieves, which we do not consider here. We start with the main result of this section—the chordal graph sieve— which associates a Bonferroni-type inequality with any chordal graph G.

38

4

Abstract Tubes via Closure and Kernel Operators

Theorem 4.4.1 [Doh02b] Let {Av }v∈V be a finite family of sets, where V is the vertex-set of a chordal graph G. Then, for any odd r ∈ ,  

   (4.26) χ Av ≤ (−1)|I|−1 χ Ai .


I∈ ∗ (V ) I clique of G |I|≤r

v∈V

i∈I

In particular,
χ

(4.27)

 v∈V

Proof. Let ω ∈

v∈V

(4.28)

 Av



≤





(−1)|I|−1 χ

I∈ ∗ (V ) I clique of G



 Ai

.

i∈I

Av and Vω := {v ∈ V | ω ∈ Av }. We have to show that   1 ≤ (−1)|I|−1 H

I clique of H 0 1. Then (5.1) holds where Λ is defined by  (5.4) Λ(v) := P (Av ) − Λ(w) P (Av |Aw ) . w” with “ j (i, j = 1, . . . , n) 2: prob ← 0 3: for i = 1 to n do 4: acc ← 0 5: for j = 1 to i − 1 do 6: if vj > vi then 7: acc ← acc + a[j] P (Avi |Avj ) 8: end if 9: end for 10: a[i] ← P (Avi ) − acc 11: prob ← prob + a[i] 12: end for Under the requirements of Theorem 5.1.1, the next theorem shows that the 

partial sums of v∈V Λ(v) provide lower bounds on P . Thus, if AlA v v∈V gorithm I is stopped at an arbitrary instant of time, prob provides a lower bound 

 to P v∈V Av . This is not the case under the requirements of Theorem 5.1.2. Theorem 5.2.1 Under the requirements of Theorem 5.1.1, 
  P Av ≥ Λ(v) for any V  ⊆ V . v∈V 

v∈V

Proof. For any v ∈ V define V >v := {i ∈ V | i > v}. Then, by (5.3), we see that 
  Λ(v) = P (Av ) − (−1)|I|−1 P Av ∩ Ai .


I∈ ∗ (V >v ) I is a chain

i∈I

By applying the identity associated with Corollary 4.1.11 with V >v instead of V and integrating the result with respect to the measure P (Av ∩ · ), we obtain 

%   Λ(v) = P (Av ) − P Av ∩ Ai = P Av Ai . i>v

i>v

So Λ(v) is non-negative, whence the result follows from Theorem 5.1.1. 2

6

Reliability Applications

In many practical situations one is interested in the probability that a technical system with randomly failing components is operating. Examples include transportation networks, electrical power systems, pipeline networks and nuclear power plants. Recently, the study of system reliability has received considerable attention from its applicability to computer and telecommunications networks. One of the standard methods in reliability theory is the principle of inclusionexclusion and the associated Bonferroni inequalities. This chapter deals with improvements of this method derived from the results of the preceding chapters. In this way, we rediscover Shier’s semilattice expression and recursive scheme for system reliability [Shi88, Shi91] and establish related inequalities based on abstract tubes. The results are then applied in the more specific context of network reliability, k-out-of-n systems, consecutive k-out-of-n systems and covering problems, where several results from the literature are rediscovered in a concise way. Examples demonstrate that the new reliability bounds are much sharper than the usual Bonferroni bounds, although less computational effort is needed to compute them. We finally identify a new class of hypergraphs for which the reliability covering problem can be solved in polynomial time. For an introduction to reliability theory with emphasis on networks we refer to the monographs of Colbourn [Col87] and Shier [Shi91]. For a more general introduction we recommend the monograph of Aven and Jensen [AJ99].

6.1 6.1.1

System reliability Terminology

The following definition essentially goes back to Esary and Proschan [EP63]. Definition 6.1.1 A coherent system or—to be more precise—a coherent binary system is a pair Σ = (E, φ) where E is a finite set and φ :
(E) → {0, 1} is a function satisfying φ(∅) = 0, φ(E) = 1 and X ⊆ Y ⇒ φ(X) ≤ φ(Y ) for any X, Y ⊆ E. We refer to E and φ as the component set resp. the structure function of the system. Each component e ∈ E assumes randomly and independently K. Dohmen: LNM 1826, pp. 47–81, 2003. c Springer-Verlag Berlin Heidelberg 2003


6

48

Reliability Applications

one of two states, operating or failing, with probabilities pe and qe = 1 − pe , respectively. The set of operating components of Σ is referred to as the state of Σ. Σ is said to be operating if φ applied to its state gives 1; otherwise, Σ is said to be failing. The probability for the event that Σ is operating is referred to as the reliability of Σ. We use the abbreviation RelΣ (p), where p = (pe )e∈E , to denote the reliability of Σ for given component operation probabilities (pe )e∈E . Remark. Note that by the preceding definition the components of a coherent system fail in a statistically independent fashion. Of course, this definition could be relaxed by allowing dependent component failures, and indeed most of our inclusion-exclusion expressions for system and network reliability can easily be adapted to this more general setting. For ease of presentation, however, we prefer to use the more restrictive definition. A generalization to dependent component failures is (whenever possible) left as an option to the reader. The following notions are of vital importance in reliability analysis. Definition 6.1.2 Let Σ = (E, φ) be a coherent binary system. A minpath of Σ = (E, φ) is a minimal set P ⊆ E such that φ(P ) = 1 (that is, φ(P ) = 1 and φ(Q) = 0 for any Q ⊂ P ). A mincut of Σ is a minimal set C ⊆ E such that φ(E \ C) = 0 (that is, φ(E \ C) = 0 and φ(E \ D) = 1 for any D ⊂ C). Thus, with denoting the set of minpaths resp. mincuts,

   {F operates} resp. 1 − RelΣ (p) = P {F fails} , RelΣ (p) = P F ∈

F ∈

where P denotes the probability measure on the set of system states, and where   {e operates}; {F fails} := {e fails} {F operates} := e∈F

e∈F

for any F ∈ . In the following, ( ) again denotes the order complex of .

6.1.2

Abstract tubes for system reliability

The following theorem is of vital importance in this chapter. For a generalization of this theorem from semilattices to convex geometries, see [Doh03]. Theorem 6.1.3 [Doh99d] Let Σ = (E, φ) be a coherent binary system, whose set of minpaths resp. mincuts is given the structure of a lower semilattice such that X ∧ Y ⊆ X ∪ Y for any X, Y ∈ . Then, & ' & '   {F operates} F ∈ , ( ) resp. {F fails} F ∈ , ( ) is an abstract tube.

6.1 System reliability

49

Proof. Apply Corollary 4.1.11 (dualized) with V := and AF := {F operates} resp. AF := {F fails} for any F ∈ . 2 Remark. The improved inclusion-exclusion identity associated with the abstract tube of the preceding theorem is due to Shier [Shi88, Shi91]. Subsequently, the improved inequalities associated with this abstract tube are explicitly stated. Corollary 6.1.4 [Doh99c] Let Σ = (E, φ) be a coherent system, whose set of minpaths resp. mincuts is a lower semilattice such that X ∧ Y ⊆ X ∪ Y for any X, Y ∈ , and let r ∈ . Then, in case that denotes the set of minpaths,  ( (−1)||−1 pe (r even), RelΣ (p) ≥ ∈() ||≤r

RelΣ (p) ≤



e∈

Ë

e∈

Ë

(

(−1)||−1

∈() ||≤r

pe

(r odd),

and in case that denotes the set of mincuts, 1 − RelΣ (p) ≥ 1 − RelΣ (p) ≤



(−1)||−1

∈() ||≤r



∈() ||≤r

(−1)||−1

(

Ë

qe

(r even),

e∈

Ë

qe

(r odd),

e∈

(

where in both cases p = (pe )e∈E ∈ [0, 1]E and qe = 1 − pe for any e ∈ E. Proof. The result directly follows from Theorem 6.1.3 and Theorem 3.1.9. 2 Remarks. In view of Theorem 3.1.10 the inequalities of Corollary 6.1.4 are at least as sharp as their classical counterparts, which are obtained if is a chain, that is, if X ∧ Y = X or X ∧ Y = Y for any X, Y ∈ . Note that Corollary)6.1.4 can easily) be generalized to dependent component failures by replacing e∈Ë  pe resp. e∈Ë  qe with the probability that all

components in  work resp. fail.

6.1.3

Shier’s pseudopolynomial algorithm

The following theorem, due to Shier [Shi88, Shi91], generalizes an important algorithm of Provan and Ball [PB84, BP87] for computing network reliability in pseudopolynomial time. The theorem requires the following notion of convexity: Definition 6.1.5 A subset X of a partially ordered set V is convex if [x, y] ⊆ X for any x, y ∈ X, where [x, y] denotes the interval {z ∈ V | x ≤ z ≤ y}.

50

6

Reliability Applications

Theorem 6.1.6 [Shi88, Shi91] Let Σ = (E, φ) be a coherent binary system, whose set of minpaths resp. mincuts is a lower semilattice such that X ∧ Y ⊆ X ∪ Y for any X, Y ∈ and {F ∈ | e ∈ F } is convex for any e ∈ E. Then,   (6.1) Λ(F, p) resp. 1 − RelΣ (p) = Λ(F, q) , RelΣ (p) = F ∈

F ∈

where q = 1 − p ∈ [0, 1]E , and where in both cases Λ is defined recursively by  ( ( xe − Λ(G, x) xe ; x = (xe )e∈E ∈ [0, 1]E . (6.2) Λ(F, x) := e∈F

e∈F \G

G · · · > Ik in where k > 1 and where P again denotes the induced probability measure on the set of system states. Since the components of the system are assumed to operate and fail independently, we find that   
       P = P Bi  Bi ∩ · · · ∩ Bi Bi  .  i∈I1

i∈I2

i∈Ik

i∈I1 \(I2 ∪···∪Ik )

By the convexity assumption, I1 \ (I2 ∪ · · · ∪ Ik ) = I1 \ I2 and therefore,      
         P Bi Bi Bi  Bi , = P = P  i∈I1 \(I2 ∪···∪Ik )

i∈I1 \I2

i∈I1

i∈I2

where the last equals sign again follows from the independence assumption. 2 Remarks. In view of Theorem 5.1.2, replacing “” in (6.2) would result in a different recursive scheme for RelΣ (p). By suitably adapting Algorithm I, we obtain Shier’s dynamic programming solution to (6.1) and (6.2), which is restated as Algorithm II. As pointed out by Shier [Shi88, Shi91], this algorithm has a space complexity of O(| |) and a time complexity of O(|E| × | |2), since there are at most O(| |2 ) products to be calculated in line 7 and the calculation of each requires at most O(|E|) work. Thus, the algorithm is pseudopolynomial, that is, its running time is bounded by a polynomial in the number of

6.1 System reliability

51

components and the number of minpaths resp. mincuts. The classical inclusionexclusion method for the same problem has a time complexity of O(|E| × 2||). While the classical inclusion-exclusion method and the improved Bonferroni inequalities provided by Corollary 6.1.4 can easily be adapted to work for statistically dependent component failures, Theorem 6.1.6 strongly relies on the independence assumption. Algorithm II Pseudopolynomial algorithm for computing system reliability Require: Same requirements as in Theorem 6.1.6; x = (xe )e∈E ∈ [0, 1]E Ensure: prob = F ∈ Λ(F, x) 1: Find an ordering F1 , . . . , Fn of such that Fi < Fj ⇒ i < j (i, j = 1, . . . , n) 2: prob ← 0 3: for i = 1 to n do 4: acc ← 0 5: for j = 1 to i − 1 do 6: if Fj < Fi then ) 7: acc ← acc + a[j] e∈Fi \Fj xe 8: end if 9: end for ) 10: a[i] ← e∈Fi xe − acc 11: prob ← prob + a[i] 12: end for In the same way as Theorem 6.1.6 follows from Theorem 5.1.1, the following inequalities follow from Theorem 5.2.1. These inequalities ensure that during execution of Algorithm II, prob provides a lower bound to the desired value. Theorem 6.1.7 Under the requirements of Theorem 6.1.6,   RelΣ (p) ≥ Λ(F, p) resp. 1 − RelΣ (p) ≥ Λ(F, q) F ∈ 

F ∈ 

for any subset  of , where q = 1 − p and Λ is defined recursively as in (6.2). Remark. Without modifying the proofs, the preceding theorems can be generalized by imposing the somewhat weaker requirement that is an extended set of minpaths resp. mincuts of Σ. These concepts are introduced subsequently. Definition 6.1.8 An extended set of minpaths of a coherent binary system Σ is an upper set of the set of minpaths of Σ such that any F ∈ is an upper set of some minpath of Σ. An extended set of mincuts of Σ is an upper set of the set of mincuts of Σ such that any F ∈ is an upper set of some mincut of Σ. We will make use of these notions when considering particular reliability systems in Section 6.2.

52

6.1.4

6

Reliability Applications

Inclusion-exclusion and domination theory

The following concept of domination was introduced by Satyanarayana and Prabhakar [SP78] in the specific context of source-to-terminal reliability. Definition 6.1.9 Let Σ = (E, φ) be a coherent binary system, whose set of minpaths resp. mincuts is denoted by . An -formation of a subset X of E is

any subset  of such that  = X. An -formation  of X is odd resp. even if || is odd resp. even. The -domination of X, dom (X), is the number of odd -formations of X minus the number of even -formations of X. The subsequent theorem provides an improved inclusion-exclusion formula for the reliability of a coherent system based on the concept of domination. Note that this formula contains no cancelling terms, and in this sense it is best possible among all inclusion-exclusion expansions. However, apart from some particular measures associated with directed networks [SP78, SH81, Sat82] and consecutively connected systems [KP89, SM91] the determination of dom (X) is a tiresome task and in fact amounts to computing the M¨ obius function of a lattice. Theorem 6.1.10 [SP78] Let Σ = (E, φ) be a coherent binary system, whose set of minpaths resp. mincuts is denoted by . Then,  ( RelΣ (p) = dom (X) pe , X∈
∗ (E)

e∈X



(

respectively 1 − RelΣ (p) =

dom (X)

X∈
∗ (E)

qe ,

e∈X

where p = (pe )e∈E ∈ [0, 1]E and qe = 1 − pe for any e ∈ E. Proof. Theorem 6.1.10 is proved by applying the classical inclusion-exclusion principle. 2 ∗ Definition 6.1.11 For any set of minpaths or mincuts we  use todenote ∗  the set of all unions of sets in including ∅, that is, :=  ⊆ .

Clearly, ∗ is a lattice with respect to the usual inclusion order. The following interpretation of dom (X) in terms of the M¨ obius function of ∗ is due to Manthei [Man90, Man91] and proved here in a new and simplified way without using Rota’s crosscut theorem (see p. 96). We first define the M¨obius function. Definition 6.1.12 The M¨ obius function of a finite partially ordered set P with least element ˆ 0 is the unique -valued function µP on P such that for any x ∈ P ,  (6.3) µP (y) = δˆ0x , y≤x

where δ is the usual Kronecker delta.

6.2 Special reliability systems

53

Theorem 6.1.13 [Man90, Man91] Let Σ = (E, φ) be a coherent binary system, whose set of minpaths resp. mincuts is denoted by . Then, for any X ∈ ∗, dom (X) = −µ∗ (X) . Proof. By the definition of the M¨ obius function it suffices to prove that

 −1 if X = ∅, dom (Y ) = 0 otherwise, Y ∈∗ Y ⊆X

which is clear if X = ∅. Otherwise, |X := {F ∈ | F ⊆ X} = ∅ and hence,      dom (Y ) = (−1)||−1 = (−1)||−1 = (−1)||−1 = 0 ,  Y ∈∗ Y ∈∗ Ë⊆ ⊆|X Ë⊆⊆X Y ⊆X Y ⊆X =Y which gives the result. 2 The domination concepts of Definition 4.3.4 and 6.1.9 are related as follows. Theorem 6.1.14 [Doh99b] Let Σ = (E, φ) be a coherent binary system, whose set of minpaths  resp. mincuts

is denoted by , and let c denote the closure operator  → F ∈  F ⊆  on . Then, for any X ∈ ∗ , dom (X) = (−1)||X|−1domc ( |X) , where |X := {F ∈ | F ⊆ X}. Proof. Evidently, |X is c -closed. Hence, by Proposition 4.3.5, we obtain  (−1)||−1 = (−1)||X|−1domc ( |X) . c

⊆|X  ()=|X

Now, c () = |X if and only if  is an -formation of X. Thus, the left-hand side of the preceding equation coincides with the -domination of X. 2

6.2 6.2.1

Special reliability systems Communications networks

Throughout this section, a network is viewed as a finite graph or digraph, whose nodes or vertices are perfectly reliable and whose edges, which are either undirected or directed, are subject to random and independent failure, where all failure probabilities are assumed to be known in advance. The reader should keep in mind that the improved inclusion-exclusion identities and Bonferroni inequalities in this section can easily be generalized to dependent edge failures

54

6

Reliability Applications

provided the joint probability distribution of the edge failures is known. The general objective is to assess the overall reliability of the network relative to a given reliability measure. Some reliability measures are reviewed subsequently. The source-to-terminal reliability or two-terminal reliability of a network is the probability that a message can pass from a distinguished source node s to a distinguished terminal node t along a path of operating edges. Here and subsequently, paths are considered as directed if the network is directed and as undirected if the network is undirected, and it is assumed that there is always an operating path from each node to itself. The problem of computing the sourceto-terminal reliability for a distinguished source node s and a distinguished terminal node t is usually referred to as the s, t-connectedness problem. It has been studied extensively in the literature (see e.g., [Shi91] for a comprehensive account). An appropriate model for the source-to-terminal measure is a coherent binary system Σ = (E, φ), where E is the edge-set of the network and φ(X) = 1 if and only if X contains the edges of an s, t-path. Evidently, the minpaths and mincuts of the system correspond to the s, t-paths and s, t-cutsets (= minimal sets of edges whose removal disconnects s from t) of the network, respectively. As a generalization of the s, t-connectedness problem, the s, T -connectedness problem asks for the probability that a message can be sent from a distinguished source node s to each node of some specified set T along a path of operating edges. This probability is usually referred to as the source-to-T -terminal reliability. An appropriate model for dealing with this network reliability measure is a coherent binary system Σ = (E, φ), where E is the edge-set of the network and φ(X) = 1 if and only if X contains the edges of an s, T -tree (= minimal subnetwork containing an s, t-path for all t ∈ T ). In this case, the minpaths of the system correspond to the s, T -trees of the network and the mincuts to its s, T -cutsets (= minimal sets of edges whose removal disconnects s from at least one node in T ). A huge literature exists on the all-terminal reliability (see e.g., [Col87] for an extensive account), which expresses the probability that a message can be sent between any two nodes of the network along a path of operating edges. In the case of an undirected network, for instance, the appropriate model is a coherent binary system Σ = (E, φ), where E is the edge-set of the network and φ(X) = 1 if and only if the subnetwork induced by X is connected. Thus, in the undirected case the minpaths correspond to the spanning trees of the network and the mincuts to its cutsets (= minimal sets of edges whose removal disconnects the network). The methods of the preceding section require a generation of the minpaths resp. mincuts as well as an appropriate semilattice structure on the set of these objects. For most relevant network reliability measures these key objects can be generated quite efficiently, that is, their generation time grows only polynomially (or even linearly) with their number. For instance, Tsukiyama et al. [TSOA80] devised an efficient algorithm for generating all s, t-cutsets in an undirected network that has a time complexity of only O((n + m)cst ) where n, m and cst denote the number of nodes, edges and s, t-cutsets of the network, respectively. More recently, Provan and Shier [PS96] established a unifying paradigm for

6.2 Special reliability systems

55

generating all s, t-cutsets and several other classes of cuts in directed and undirected networks that exhibits a worst-case time complexity growing only linearly with the number of objects generated. Efficient algorithms for generating s, tpaths are provided by Read and Tarjan [RT75] as well as Colbourn [Col87]. Shier [Shi91] describes several algebraic enumeration techniques for generating these key objects based on a symbolic version of the Gauss-Jordan algorithm. Unfortunately, the number of key objects can grow exponentially with the size (= number of nodes and edges) of the network, whence the methods of the preceding section exhibit an exponential time behaviour in the worst case. On the other hand, the computation of nearly all relevant network reliability measures is known to be #P -hard [Val79, Bal86] (see also Garey and Johnson [GJ79] for notions of computational complexity), whence a polynomial time algorithm (that is, an algorithm whose time complexity is bounded by a polynomial in the size of the network) for exactly computing any of these measures is unlikely to exist. Fortunately, though, in many practical situations the networks are sparse and thus do not have too many minpaths or mincuts. In view of this and the fact that the method of inclusion-exclusion is a standard method in system and network reliability analysis, it is reasonable to investigate improvements of this method. Now, in order to apply the results of the preceding section to a network, an appropriate partial ordering relation on the set of minpaths resp. mincuts (or, more generally, on an extended set of minpaths resp. mincuts) must be imposed. The following partial ordering relations, which are adopted from Shier [Shi88, Shi91], are appropriate for dealing with the source-to-terminal reliability measure. I. For edge-sets X and Y of s, t-paths in a planar network define (6.4)

X≤Y

:⇔

X lies below Y .

Of course, this partial ordering relation depends on a specific drawing of the network in the plane with no edges crossing and with s and t lying on the boundary of the exterior region. Figure 6.1 clarifies this concept. II. For s, t-cutsets X and Y of an arbitrary network define (6.5)

X ≤Y

:⇔

Ns (X) ⊆ Ns (Y ) ,

where Ns (X) is the set of nodes reachable from s after removing X. It is easy to see that these partial ordering relations induce a lattice structure, where X ∧ Y and X ∨ Y are included in X ∪ Y . Moreover, both partial ordering relations satisfy the convexity requirement of Theorem 6.1.6. We thus conclude that Theorems 6.1.3–6.1.7 can be applied to networks whose s, t-paths resp. s, t-cutsets are ordered as in (6.4) resp. (6.5). For s, t-paths in planar networks, ordered as in (6.4), the improved inclusion-exclusion identity associated with the abstract tube of Theorem 6.1.3 is due to Shier [Shi88, Shi91], while the associated improved Bonferroni inequalities are due to the author [Doh99d]. For s, t-cutsets of arbitrary networks, ordered as in (6.5), the improved inclusionexclusion identity associated with the abstract tube of Theorem 6.1.3 coincides

56

6

Reliability Applications

with Buzacott’s node partition formula [BC84, Buz87], whereas the improved inequalities were first established in [Doh99d]. For complete undirected networks on n nodes the classical and improved inclusion-exclusion identity for two-terminal reliability based on the s, t-cutsets of the network are compared in [Vog99] for some small values of n, see Table 6.2 for details. In [Vog99] similar results were reported for random undirected networks. We further remark that Theorem 6.1.6, when applied to s, t-cutsets, specializes to a wellknown result of Provan and Ball [PB84], whereas for s, t-paths it is again due to Shier [Shi88, Shi91]. For further informations on these partial ordering relations, the reader is referred to Shier [Shi88, Shi91]. X

Y

s

t inf(X,Y)

Figure 6.1: Two s, t-paths and their infimum [Shi91].

n

number of terms n−2
classical 22 improved

3 4 5 6 7 8

4 16 256 65535 4294967296 18446744073709551616

4 12 52 300 2164 18732

Figure 6.2: Classical and improved inclusion-exclusion [Vog99]. A partial ordering relation on the set of s, t-paths or s, T -cutsets of a directed network satisfying the requirements of Theorem 6.1.6 is difficult to find, since such a partial ordering relation gives rise to a pseudopolynomial algorithm for computing two-terminal resp. source-to-T -terminal reliability. Here, pseudopolynomial means that the worst-case time complexity of the algorithm is bounded by a polynomial in the number of s, t-paths resp. s, T -cutsets of the network. However, by a well-known result due to Provan and Ball [PB84], such an algorithm cannot exist unless the complexity classes P and N P coincide. For complete networks, however, we can establish a partial ordering relation on the set of s, T -cutsets that satisfies the requirements of Theorem 6.1.6 by defining X ≤ Y as in (6.5) for any s, T -cutsets X and Y of the network. Indeed, this partial ordering relation induces a lower semilattice such that X∧Y ⊆ X∪Y ; see [Doh98] for details. The convexity requirement is easily verified. We remark that for s, t-cutsets of arbitrary networks and s, T -cutsets of

6.2 Special reliability systems

57

complete networks, the recursive scheme of Theorem 6.1.6 goes back to Ball and Provan [PB84, BP87]. As in [Doh99c], we now establish a partial ordering relation, which is appropriate for dealing with the all-terminal reliability of an undirected network and which leads to an improvement upon the classical inclusion-exclusion identities and inequalities if the network is sufficiently dense. To this end, let G be an undirected network with vertex-set V , and for any non-empty proper subset W of V let W  be the quasi-cut associated with W , that is, the set of edges linking some node in W to some node in V \ W . Then, := {W  | ∅ = W ⊂ V } is an extended set of mincuts of the coherent system associated with G and the all-terminal reliability measure. Note that coincides with the set of cutsets of the network if and only if the network is complete. Now, fix some v ∈ V , and for any non-empty proper subsets W1 and W2 of V containing v define W1  ≤ W2  :⇔ W1 ⊆ W2 .

(6.6)

In this way, becomes a lower semilattice where W1  ∧ W2  = W1 ∩ W2  ⊆ W1  ∪ W2  for any non-empty proper subsets W1 and W2 of V containing v. For complete undirected networks, whose cutsets are ordered as in (6.6), the improved identity associated with the abstract tube of Theorem 6.1.3 coincides with a particular case of Buzacott’s node partition formula [BC84]. As noted by Buzacott [BC84], the formula contains no cancelling terms in this case and thus is best possible among all cutset-based inclusion-exclusion expansions. Example 6.2.1 [Doh99c, Doh99d] Consider the network in Figure 6.3. We are interested in bounds for its two-terminal reliability with respect to s and t. For simplicity, assume that all edges fail independently with probability q = 1 − p. Let’s first consider the classical and improved approach based on

6

1

4 2

s

7

t

5 3

8

Figure 6.3: A sample network with terminal nodes s and t. the s, t-paths of the network, which are assumed to be partially ordered as proposed in (6.4). The Hasse diagram corresponding to this partial ordering relation is shown in Figure 6.4(a), and the corresponding bounds (both classical and improved ones) are listed in Table 6.1 together with the number of sets

58

6

Reliability Applications

16

147

1458

678

246

27

258

3567

24568

1478

2346

13457

1258

3546

357

38

123

(a) Hasse diagram of s, t-paths

(b) Hasse diagram of s, t-cutsets

Figure 6.4: Hasse diagrams for the network in Figure 6.3.

inspected during the computation of each bound. Note that the classical bounds come from the classical Bonferroni inequalities, whereas the improved bounds are those of Corollary 6.1.4. Similarly, the s, t-cutsets of the network, which are assumed to be partially ordered as in (6.5), give rise to the Hasse diagram in Figure 6.4(b) and the bounds in Table 6.2. Note that in Table 6.1 even and odd values of r correspond to lower and upper bounds on the reliability of the network, respectively, whereas in Table 6.2 the correspondence is vice versa. In each case, the last bound represents the exact reliability of the network. As expected, the improved bounds involve much fewer terms (sets) than their classical counterparts, although they are much closer to the exact reliability. A numerical comparison of classical and improved bounds is shown in Tables 6.3 and 6.4, and in Figures 6.6 and 6.7 some of these bounds are plotted. To illustrate the bounds provided by Theorem 6.1.7, we consider the Hasse diagram of s, t-paths in Figure 6.4(a). (Of course, we could also employ the s, t-cutsets.) Straightforward application of the recursive scheme (6.2) gives Λ(38) = p2 , Λ(258) = p3 − p4 , Λ(357) = p3 − p4 , Λ(1458) = p4 − 2p5 + p6 ,

Λ(27) = p2 − 3p4 + 2p5 , Λ(3546) = p4 − 2p5 + p6 , Λ(147) = p3 − p4 − 3p5 + 5p6 − 2p7 , Λ(246) = p3 − p4 − 3p5 + 5p6 − 2p7 ,

where the expression Λ(e1 . . . en ) is used as an abbreviation for Λ({e1 , . . . , en }, p). Now, by virtue of Theorem 6.1.7, these Λ-values give rise to lower bounds for the two-terminal reliability of our sample network in Figure 6.3. We thus obtain

6.2 Special reliability systems

59

e.g. the following lower bounds, which are plotted in Figure 6.8: c0 := Λ(38) = p2 , c1 := c0 + Λ(258) + Λ(357) = p2 + 2p3 − 2p4 , c2 := c1 + Λ(1458) + Λ(27) + Λ(3546) = 2p2 + 2p3 − 3p4 − 2p5 + 2p6 , c3 := c2 + Λ(147) + Λ(246) = 2p2 + 4p3 − 5p4 − 8p5 + 12p6 − 4p7 . In a similar way, bounds for the all-terminal reliability of the network in Figure 6.3 are obtained by employing the cutsets and quasi-cuts of the network. The Hasse diagram of the quasi-cuts, which are assumed to be partially ordered as in (6.6), is shown in Figure 6.5, and the corresponding bounds can be read from Table 6.5 together with the number of sets inspected during the computation of each bound. Here, the classical bounds are the classical Bonferroni bounds based on the cutsets of the network, whereas the improved bounds are obtained by applying Corollary 6.1.4 to the quasi-cuts of the network. Note that in Table 6.5 even and odd values of r correspond to upper and lower bounds, respectively. Again, we observe that the improved bounds are much sharper than the classical ones. Some numerical values of the bounds in Table 6.5 are shown in Table 6.6, and in Figure 6.9 some of these bounds are plotted.

23478

358

678

2457

146

3567

24568

1478

134568

2346

13457

1258

123678

12567

123

Figure 6.5: Hasse diagram of quasi-cuts of the network in Figure 6.3.

4p3

2p4 9 45 129 255 381 465 501 510 511

sets 4p3

2p4

+ + 3p2 + 4p3 − 9p4 − 14p5 3p2 + 4p3 − 9p4 − 10p5 3p2 + 4p3 − 9p4 − 10p5 3p2 + 4p3 − 9p4 − 10p5

3p2 − 2p6 + 27p6 + 4p7 + 27p6 − 18p7 − 2p8 + 27p6 − 18p7 + 4p8

improved bounds a∗r 9 36 73 97 103

sets

4q 4

2q 5

1− − − 1 − 2q 3 − 4q 4 + 2q 5 + 13q 6 1 − 2q 3 − 4q 4 + 2q 5 + 13q 6 1 − 2q 3 − 4q 4 + 2q 5 + 13q 6 1 − 2q 3 − 4q 4 + 2q 5 + 13q 6 1 − 2q 3 − 4q 4 + 2q 5 + 13q 6 1 − 2q 3 − 4q 4 + 2q 5 + 13q 6 1 − 2q 3 − 4q 4 + 2q 5 + 13q 6

2q 3 + 10q 7 − 22q 7 − 14q 7 − 14q 7 − 14q 7 − 14q 7 − 14q 7

classical bounds br + q8 − 23q 8 + 39q 8 − 17q 8 + 11q 8 + 3q 8 + 4q 8

9 37 93 163 219 247 255 256

sets 4q 4

2q 5

1− − − 1 − 2q 3 − 4q 4 + 2q 5 + 13q 6 + 2q 7 1 − 2q 3 − 4q 4 + 2q 5 + 13q 6 − 14q 7 − 2q 8 1 − 2q 3 − 4q 4 + 2q 5 + 13q 6 − 14q 7 + 4q 8

2q 3

improved bounds b∗r

9 28 46 52

sets

Table 6.2: Cutset-based bounds for the two-terminal reliability of the network in Figure 6.3.

1 2 3 4 5 6 7 8

r

Table 6.1: Path-based bounds for the two-terminal reliability of the network in Figure 6.3.

+ + 3p2 + 4p3 − 9p4 − 16p5 − 9p6 3p2 + 4p3 − 9p4 − 8p5 + 34p6 + 30p7 + 3p8 3p2 + 4p3 − 9p4 − 10p5 + 27p6 − 50p7 − 34p8 3p2 + 4p3 − 9p4 − 10p5 + 27p6 − 12p7 + 54p8 3p2 + 4p3 − 9p4 − 10p5 + 27p6 − 18p7 − 24p8 3p2 + 4p3 − 9p4 − 10p5 + 27p6 − 18p7 + 12p8 3p2 + 4p3 − 9p4 − 10p5 + 27p6 − 18p7 + 3p8 3p2 + 4p3 − 9p4 − 10p5 + 27p6 − 18p7 + 4p8

3p2

classical bounds ar

6

1 2 3 4 5 6 7 8 9

r

60 Reliability Applications

6.2 Special reliability systems .

1 0.8 0.6

Rel a*4 a4 a*3 a3

0.4 0.2 p 0.2

0.4

0.6

0.8

1

Figure 6.6: Some path-based bounds of Table 6.1.

1 0.8 0.6 0.4

Rel b*3 b3 b*2 b2

0.2 q 0.2

0.4

0.6

0.8

1

Figure 6.7: Some cutset-based bounds of Table 6.2.

61

62

6

Reliability Applications

p

a2

a∗2

a4

a∗4

a∗5 †

a5

a∗3

a3

0.1 0.2 0.3 0.4 0.5

0.033 0.132 0.260 0.305 0.047

0.033 0.133 0.270 0.354 0.219

0.033 0.135 0.287 0.410 0.273

0.033 0.136 0.296 0.483 0.648

0.033 0.136 0.297 0.487 0.672

0.033 0.136 0.301 0.530 0.914

0.033 0.136 0.301 0.520 0.828

0.033 0.138 0.317 0.614 1.215

† two-terminal

reliability

Table 6.3: Numerical values of the bounds in Table 6.1.

q

b3

b∗3

b5

b∗4 †

b4

b∗2

b2

0.1 0.2 0.3 0.4 0.5

0.998 0.979 0.922 0.792 0.504

0.998 0.979 0.925 0.819 0.648

0.998 0.979 0.924 0.809 0.590

0.998 0.979 0.925 0.823 0.672

0.998 0.979 0.927 0.846 0.809

0.998 0.979 0.928 0.847 0.781

0.998 0.979 0.930 0.860 0.848

† two-terminal

reliability

Table 6.4: Numerical values of the bounds in Table 6.2.

1 0.8 0.6

Rel c3 c2 c1 c0

0.4 0.2 p 0.2

0.4

0.6

0.8

1

Figure 6.8: Path-based bounds c0 , c1 , c2 , c3 .

1− 1 − 4q 3 1 − 4q 3 1 − 4q 3 1 − 4q 3 1 − 4q 3 1 − 4q 3 1 − 4q 3 1 − 4q 3 1 − 4q 3 1 − 4q 3 1 − 4q 3 1 − 4q 3

4q 3

− − 5q 4 − 5q 4 − 5q 4 − 5q 4 − 5q 4 − 5q 4 − 5q 4 − 5q 4 − 5q 4 − 5q 4 − 5q 4 − 5q 4

5q 4

− + 8q 5 + 4q 5 + 4q 5 + 4q 5 + 4q 5 + 4q 5 + 4q 5 + 4q 5 + 4q 5 + 4q 5 + 4q 5 + 4q 5

4q 5

+ 44q 6 + 30q 6 + 30q 6 + 30q 6 + 30q 6 + 30q 6 + 30q 6 + 30q 6 + 30q 6 + 30q 6 + 30q 6 + 30q 6

+ 20q 7 − 160q 7 + 40q 7 − 68q 7 − 36q 7 − 40q 7 − 40q 7 − 40q 7 − 40q 7 − 40q 7 − 40q 7 − 40q 7

classical bounds dr + 2q 8 − 86q 8 + 429q 8 − 750q 8 + 934q 8 − 778q 8 + 509q 8 − 206q 8 + 80q 8 + 2q 8 + 15q 8 + 14q 8

13 91 377 1092 2379 4095 5811 7098 7813 8099 8177 8190 8191

sets 1− 1 − 4q 3 1 − 4q 3 1 − 4q 3

4q 3 − − 5q 4 − 5q 4 − 5q 4

5q 4 − + 4q 5 + 4q 5 + 4q 5

4q 5 − + 30q 6 + 8q 7 + 2q 8 + 30q 6 − 40q 7 − 10q 8 + 30q 6 − 40q 7 + 14q 8

2q 6

improved bounds d∗r 15 65 125 149

sets

0.995 0.959 0.842 0.575 0.063

0.1 0.2 0.3 0.4 0.5

d3 0.996 0.961 0.842 0.461 -0.805

reliability

0.995 0.959 0.840 0.567 0.031

d∗1 0.996 0.963 0.874 0.708 0.430

d∗3 0.996 0.963 0.875 0.723 0.523

d∗4 † 0.996 0.965 0.920 1.127 2.770

d4

0.996 0.963 0.885 0.794 0.852

d∗2

0.996 0.966 0.908 0.912 1.289

d2

Table 6.6: Numerical values of the bounds in Table 6.5.

† all-terminal

d1

q

Table 6.5: Cutset-based bounds for the all-terminal reliability of the network in Figure 6.3.

1 2 3 4 5 6 7 8 9 10 11 12 13

r

6.2 Special reliability systems 63

64

6

Reliability Applications

1 Rel

0.8

d2 0.6

d*2

0.4

d3 d*3

0.2

0.2

0.4

0.6

0.8

1

q

Figure 6.9: A plot of some of the bounds in Table 6.5.

Comparison with best linear Bonferroni bounds So far, we saw that our abstract tube bounds for two-terminal resp. all-terminal reliability of the network in Figure 6.3 are better than the classical Bonferroni bounds for network reliability, although fewer terms are taken into account. We now compare our abstract tube bounds with some well-known optimal Bonferroni bounds. A good source of information on optimal Bonferroni bounds is Pr´ekopa [Pr´e88, Pr´e90, Pr´e95] where such bounds are obtained by linear programming techniques. See also Galambos and Simonelli [GS96a]. Let (Ω, , P ) be a probability space, and let {Av }v∈V be a finite family of events with Av ∈  for any v ∈ V . Define the binomial moments S1 , S2 , . . . by 
  (k = 1, 2, . . . ). Sk := P Ai I⊆V |I|=k

i∈I

The first inequality, that we consider, is due to Dawson and Sankoff [DS67]:
 + *  2 2 2S2 P S1 − S2 where θ := 1 + (6.7) . Av ≥ θ+1 θ(θ + 1) S1 v∈V

The second inequality, which gives an upper bound, is due to Kwerel [Kwe75a]:
  2t − 1 t−1 P (6.8) Av ≤ S1 − t+1 S2 + t+1 S3 (t = 2, . . . , |V | − 1). v∈V

2

3

Here, the optimal choice for t is topt := min{2 + 3S3 /S2 , |V | − 1}. Note that the particular case where t = |V | − 1 can also be deduced from our chordal graph sieve (for details, see Section 4.4, p. 42).

6.2 Special reliability systems

65

Let Σ = (E, φ) be a general coherent binary system where each component e ∈ E fails randomly and independently with probability qe = 1 − pe , and let p = (pe )e∈E . The Dawson-Sankoff bound (6.7) gives RelΣ (p) ≥

(6.9)

2 2 S1 − S2 θ+1 θ(θ + 1)

where (6.10)

Sk =

 (   e∈Ë 

* pe

(k = 1, 2, . . . ) ,

θ := 1 +

⊆ | |=k

+ 2S2 , S1

and where denotes the set of minpaths of Σ. In a similar way, by considering mincuts instead of minpaths we obtain from (6.7) RelΣ (p) ≤ 1 −

(6.11)

2 2 S + S θ + 1 1 θ (θ + 1) 2

where (6.12)

Sk

=

 ( ⊆ e ∈ Ë  ||=k

qe

(k = 1, 2, . . . ) ,

+ 2S2 , θ := 1 + S1 *



and  denotes the set of mincuts of Σ. Likewise, by Kwerel’s inequality (6.8), (6.13)

2topt − 1 topt − 1 RelΣ (p) ≤ S1 − topt +1 S2 + topt +1 S3 2

3

where S1 , S2 , S3 are given by (6.10), topt = min{2 + 3S3 /S2 , | | − 1} and denotes the set of minpaths of Σ. Similar to the above, (6.8) gives (6.14)

2topt − 1 topt − 1 RelΣ (p) ≥ 1 − S1 + t +1 S2 − t +1 S3 opt

2

opt

3

where S1 , S2 , S3 are given by (6.12), topt = min{2 + 3S3 /S2 , | | − 1} and  denotes the set of mincuts of Σ. Now, for the two-terminal reliability of the network in Figure 6.3 we have (6.15)

S1 = 3p2 + 4p3 + 2p4 ,

S1 = 2q 3 + 4q 4 + 2q 5 ,

(6.16)

S2 = 11p4 + 16p5 + 9p6 ,

S2 = 4q 5 + 13q 6 + 10q 7 + q 8 ,

(6.17)

S3 = 8p5 + 43p6 + 30p7 + 3p8 , S3 = 32q 7 + 24q 8 ,

provided that all edges fail with equal probability q = 1 − p. Let aDS and bDS 2 2 be the Dawson-Sankoff bounds given by (6.9) and (6.11), respectively, where Si and Si are taken from (6.15), (6.16) and (6.17), and where we use the subscript 2 in aDS and bDS in order to signify that these bounds are of order 2. Likewise, 2 2

66

6

Reliability Applications

let aKw and bKw be the Kwerel bounds given by (6.13) and (6.14), respectively, 3 3 where use the subscript 3 in order to signify that these bounds are of order 3. For the all-terminal reliability of the network in Figure 6.3 we find that (6.18)

S1 = 4q 3 + 5q 4 + 4q 5 ,

(6.19)

S2 = 12q 5 + 44q 6 + 20q 7 + 2q 8 ,

(6.20)

S3 = 4q 5 + 14q 6 + 180q 7 + 88q 8 ,

provided that all edges fail with equal probability q = 1 − p. Let dDS be the 2 Dawson-Sankoff bound given by (6.11) and dKw be the Kwerel bound given 3 by (6.14), where this time S1 , S2 , S3 are taken from (6.18), (6.19) and (6.20), respectively. Again, subscripts are used to indicate the order of the bounds. In Table 6.7 some numerical values are given that allow us to compare our abstract tube bounds with Dawson-Sankoff and Kwerel bounds for the specific network in Figure 6.3. Note that the second and the third column in Table 6.7 provide lower bounds to the reliability of the network whereas the fifth and the sixth column provide upper bounds. It turns out that for small values of p resp. q the abstract tube bounds are at least as sharp as the DawsonSankoff and Kwerel bounds, whereas for large values of p resp. q the DawsonSankoff and Kwerel bounds are sharper than our abstract tube bounds. In most practical applications, p is large and q is small. Thus, it seems that the abstract tube bounds are advantageous if cutsets are used to estimate the reliability of the network, while the Dawson-Sankoff and Kwerel bounds are advantageous if pathsets are used instead. Note, however, that in both cases the abstract tube bounds require much fewer sets to be inspected during their computation.

6.2.2

k-out-of-n systems

The class of k-out-of-n systems was introduced by Birnbaum, Esary and Saunders [BES61]; see also Rushdi [Rus93] for an extensive survey. A k-out-of-n success (resp. failure) system operates (resp. fails) whenever k or more components operate (resp. fail). As for coherent systems, it is again assumed that the components fail randomly and independently (this assumption might be relaxed). A formal definition in terms of coherent binary systems follows. Definition 6.2.2 Let k, n ∈ and 1 ≤ k ≤ n. A k-out-of-n success (resp. failure) system is a coherent binary system Σ = (E, φ) where |E| = n and where for any subset X of E, φ(X) = 1 (resp. φ(E \ X) = 0) if and only if |X| ≥ k. Remark. The minpaths (resp. mincuts) of a k-out-of-n success (resp. failure) system Σ = (E, φ) are the k-subsets of E. Likewise, the mincuts (resp. minpaths) of a k-out-of-n success (resp. failure) system Σ = (E, φ) are the (n − k + 1)-subsets of E. Any k-out-of-n success (resp. failure) system is an (n − k + 1)-out-of-n failure (resp. success) system. Thus, we may restrict ourselves to either type of system.

6.2 Special reliability systems

p

a∗2

aDS 2

Rel†

aKw 3

a∗3

0.1 0.2 0.3 0.4 0.5 0.6 0.7

0.033 0.133 0.270 0.354 0.219 -0.404 -1.907

0.033 0.132 0.260 0.364 0.474 0.587 0.690

0.033 0.136 0.297 0.487 0.672 0.823 0.925

0.033 0.138 0.311 0.540 0.796 1.053 1.268

0.033 0.136 0.301 0.520 0.829 1.372 2.506

† two-terminal

reliability

q

b∗3

bKw 3

Rel†

bDS 2

b∗2

0.1 0.2 0.3 0.4 0.5 0.6 0.7

0.998 0.979 0.925 0.819 0.648 0.386 -0.049

0.998 0.979 0.922 0.808 0.621 0.363 0.063

0.998 0.979 0.925 0.823 0.672 0.487 0.297

0.998 0.979 0.930 0.860 0.762 0.649 0.515

0.998 0.979 0.928 0.847 0.781 0.868 1.384

† two-terminal

reliability

q

d∗3

dKw 3

Rel†

dDS 2

d∗2

0.1 0.2 0.3 0.4 0.5 0.6 0.7

0.996 0.963 0.874 0.708 0.430 -0.089 -1.241

0.996 0.962 0.868 0.687 0.408 0.045 -0.343

0.996 0.963 0.875 0.723 0.523 0.314 0.142

0.996 0.966 0.908 0.829 0.736 0.629 0.499

0.996 0.963 0.885 0.794 0.852 1.456 3.403

† all-terminal

67

reliability

Table 6.7: Abstract tube bounds vs. Dawson-Sankoff-Kwerel bounds.

As in [Doh98] we establish a partial ordering relation on the set of k-subsets of E, thereby assuming that E is endowed with a linear ordering relation ≤ . For k-subsets X and Y of a linearly ordered set E define (6.21)

X≤Y

:⇔

x ≤ y for all x ∈ X, y ∈ Y \ X .

Thus, a partial ordering relation on the set of k-subsets of E is established. Figure 6.10 shows the associated Hasse diagram for E = {1, . . . , 6} and k = 3. Again, it is easy to see that the convexity requirement of Theorem 6.1.6 is satisfied; moreover, X ∧ Y ⊆ X ∪ Y , since X ∧ Y consists of the k smallest elements of X ∪Y . Therefore, the results of Section 6.1 can be applied to k-out-of-n success or failure systems whose k-subsets are ordered as in (6.21). We conclude that for fixed k, the time and space complexity of the pseudopolynomial algorithm (Algorithm II), when applied to the k-subsets of an n-set, are O(n2k+1 ),

6

68

126

Reliability Applications

136

146

156

236

246

256

346

125

135

145

235

245

345

124

134

356

456

234

123

Figure 6.10: Hasse diagram for E = {1, . . . , 6} and k = 3.

resp. O(nk ). A more efficient method is the generating function method of Barlow and Heidtmann [BH84], whose time complexity is linear in n if k is fixed. For k-out-of-n success or failure systems Σ = (E, φ) we subsequently consider the number of terms in the improved inclusion-exclusion identity associated with the abstract tube of Theorem 6.1.3, that is, we compute the number of nonempty chains in the partially ordered set of k-subsets of E. Theorem 6.2.3 [Doh98] Let E be a finite set of cardinality n, whose k-subsets are ordered as in (6.21). Then, the number of non-empty chains in the poset of k-subsets of E is equal to 2fk (n − k) − 1, where fk (0) := 1 and (6.22)

 t−1   t−s+k−1 fk (s) fk (t) := 1 + k−1 s=0

(t = 1, . . . , n − k) .

Proof. For any k-subset I of E, let c(I) denote the number of chains extending upward from I. Then, the total number of chains is 2c(ˆ0) where ˆ0 = {1, . . . , k} denotes the minimum of the poset. We show that c(ˆ0) = fk (n − k). More generally, by induction on t we prove h(I) = n − k − t ⇒ c(I) = fk (t),

(6.23)

where h(I) denotes the height of I in the poset of k-subsets of E. For t = 0 this is obvious, since n − k is the maximum height of the poset. Now assume that h(I) = n − k − t for some t > 0. Then, by the induction hypothesis, c(I) = 1 +

t−1 



s=0

J≥I h(J)=n−k−s

c(J) = 1 +

t−1 



s=0

J≥I h(J)=n−k−s

fk (s) = 1 +

t−1  s=0

Ns (I)fk (s)

6.2 Special reliability systems

69

where Ns(I) := |{J  ≥ I | h(J) = n − k − s}| (s = 0, . . . , t). It is easy to see that Ns (I) = t−s+k−1 (hint: use downward induction on s), and therefore, k−1  t−1   t−s+k−1 fk (s) = fk (t) , c(I) = 1 + k−1 s=0 which settles (6.23). Thus, the proof is complete. 2 In order to compare the number of terms in the improved inclusion-exclusion identity with the number of terms in the classical inclusion-exclusion expansion for fixed k and increasing n, it seems reasonable to consider the ratio

k (n) :=

(6.24)

2fk (n − k) − 1 , n 2( k ) − 1

with fk (t) as in (6.22). It should be clear, though, that the classical inclusionexclusion method is not an effective method for analyzing k-out-of-n systems. The following numerical values are computed via (6.22) and (6.24):

2 (6) ≈ 7.0 × 10−3 ,

3 (6) ≈ 1.7 × 10−4 ,

4 (6) ≈ 1.8 × 10−3 ,

2 (8) ≈ 1.0 × 10−5 ,

3 (8) ≈ 6.0 × 10−14 ,

4 (8) ≈ 2.1 × 10−18 ,

2 (10) ≈ 9.0 × 10−10 ,

3 (10) ≈ 7.6 × 10−32 ,

4 (10) ≈ 5.9 × 10−59 .

This suggests that k (n) tends to 0 as n → ∞. In fact, this is the statement of the following theorem. Theorem 6.2.4 [Doh98] For any k > 1, limn→∞ k (n) = 0. Proof. By Theorem 6.2.3 and since fk (t) ≤ 1 +

t−1 

t−s+k−1 k−1

k t−s fk (s)

≤ k t−s we immediately find that (t = 0, . . . , n − k) ,

s=0

and hence, fk (t) ≤ 1 + k

t−1 

(2k)s = 1 + k

s=0

1 − (2k)t 1 − 2k

(t = 0, . . . , n − k) .

Hence, for fixed k, there are constants c and d depending only on k such that

k (n) ≤ c

k (2k)n ∼ c2dn−n , n ) ( k 2

which immediately implies the statement of the theorem. 2

70

6

Reliability Applications

Remark. An explicit solution for fk (t) for general ∞k is not known. However, we can compute the generating function Fk (z) := t=0 fk (t)xt . By (6.22),
  ∞ t−1    t−s+k−1 1+ fk (s) z t Fk (z) = k − 1 t=0 s=0  ∞ t−1   t−s+k−1 1 + fk (s) z t = 1 − z t=0 s=0 k−1  t  ∞   t−s+k−1 1 + fk (s) z t − Fk (z), = 1 − z t=0 s=0 k−1 whence




  ∞   1 t+k−1 t + Fk (z) z 1−z k−1 t=0   1 1 1 + Fk (z) . = 2 1−z (1 − z)k

1 Fk (z) = 2

Solving for Fk (z) we obtain Fk (z) =

(1 − z)k−1 . 2(1 − z)k − 1

For instance, for k = 3 we obtain the series expansion F3 (z) = 1 + 4z + 19z 2 + 92z 3 + 446z 4 + 2162z 5 + 10480z 6 + 50800z 7 + · · · From f3 (3) = 92 we conclude that the semilattice depicted in Figure 6.10 contains 2f3 (3) − 1 = 183 non-empty chains. So we again have 3 (6) ≈ 1.7 × 10−4 . Note. The derivation of the generating function is due to Peter Tittmann (Mittweida), whom I would like to thank very much for his worthy contribution.

6.2.3

Consecutive k-out-of-n systems

Consecutive k-out-of-n success (resp. failure) systems operate (resp. fail) whenever k or more consecutive components operate (resp. fail). Again, it is assumed that the components fail randomly and independently with known probabilities. Systems of this type were first considered by Kontoleon [Kon80]; the nomenclature goes back to Chiang and Niu [CN81], who also provide several applications of this model. For an account of consecutive systems, we recommend the survey paper of Papastavridis and Koutras [PK93]. A formal definition follows. Definition 6.2.5 Let k, n ∈ , 1 ≤ k ≤ n. A consecutive k-out-of-n success (resp. failure) system is a coherent binary system Σ = (E, φ) where E is a linearly ordered finite set of size n and where for any subset X of E, φ(X) = 1 (resp. φ(E \ X) = 0) if and only if X contains at least k consecutive elements of E.

6.2 Special reliability systems

71

As illustrated by the following example, consecutive k-out-of-n failure systems serve as a model for a particular type of communications networks. Example 6.2.6 In the network of Figure 6.11 the nodes 1–7 are assumed to fail randomly and independently with known probabilities whereas the edges and the two terminal nodes s and t are perfectly reliable. Evidently, in this network a message can pass from s to t unless four consecutive nodes among 1–7 simultaneously fail. Thus, we are faced with a 4-out-of-7 failure system.

s

1

2

3

4

5

6

7

t

Figure 6.11: A consecutive 4-out-of-7 failure network. In general, X is a minpath (resp. mincut) of a consecutive k-out-of-n success (resp. failure) system Σ = (E, φ) if and only if X is a consecutive subset of E containing exactly k elements. The following partial ordering relation on the set of consecutive k-subsets of E is adopted from Shier [Shi88, Shi91]: For any consecutive k-subsets X and Y of a linearly ordered set E define (6.25)

X≤Y

:⇔

min X ≤ min Y .

Thus, a partial (in fact: linear) ordering relation on the set of minpaths (resp. mincuts) of a consecutive k-out-of-n success (resp. failure) system is given, which satisfies the requirements of Theorem 6.1.6: If X and Y are two consecutive ksubsets of E, then X ∧ Y = X or X ∧ Y = Y and hence, X ∧ Y ⊆ X ∪ Y . If X ≤ Y ≤ Z are three consecutive k-subsets of E and e ∈ X ∩ Z, then min Y ≤ min Z ≤ e ≤ max X ≤ max Y and hence e ∈ Y . Therefore, the requirements of Theorem 6.1.6 are satisfied and thus Algorithm II can be applied. Since the number of minpaths (resp. mincuts) is n − k + 1, the algorithm has a space complexity of O(n) and a time complexity of O(n3 ). In contrast, the time complexity of the classical inclusion-exclusion method for that problem is O(n2n ). Note, however, that the classical method can be generalized to dependent component failures, whereas Algorithm II strongly relies on the independence assumption. Similar remarks as for the classical inclusionexclusion method applies to the inclusion-exclusion expansion of Kossow and Preuss [KP89], which involves only O(n4 ) non-cancelling terms, and which is best possible among all inclusion-exclusion expansions for the reliability of a consecutive k-out-of-n system with unequal component reliabilities. For k ≥ n/2 the following theorem provides an improved inclusion-exclusion expansion that contains only O(n2 ) terms none of which cancel. Thus, the expansion is best possible among all inclusion-exclusion expansions for this restricted case.

6

72

Reliability Applications

Theorem 6.2.7 [Doh03] Let Σ be a consecutive k-out-of-n success system whose component reliabilities are given by the vector p = (p1 , . . . , pn ). If k ≥ n/2, then (6.26)

RelΣ (p) =

n−k+1  i+k−1 ( i=1

pj −

j=i

n−k  i+k (

pj .

i=1 j=i

Proof. For i = 1, . . . , n−k +1 let Ai be the event that components i, . . . , i+k −1 operate. Then, RelΣ (p) = Pr(A1 ∪ · · · ∪ An−k+1 ). Since k ≥ n/2, Ax ∩ Ay ⊆ Az for x, y = 1, . . . , n − k + 1 and any z between x and y. By Corollary 4.1.12, RelΣ (p) =

n−k+1 

P (Ai ) −

i=1

where P (Ai ) =

)i+k−1 j=i

n−k 

P (Ai ∩ Ai+1 ) ,

i=1

pj and P (Ai ∩ Ai+1 ) =

)i+k j=i

pj . 2

Remark. Note that (6.26) can also be written as (6.27)

RelΣ (p) =

n−k  i=1

(1 − pi+k )

i+k−1 ( j=i

pj +

n (

pj ,

j=n−k+1

and that both identities can be generalized to dependent component failures. For independent components, the expression in (6.27) can be computed in O(n) steps by means of the following algorithm. Algorithm III Reliability of a consecutive k-out-of-n system where k ≥ n/2 Require: Components 1, . . . , n work independently with probability p1 , . . . , pn Ensure: acc gives the probability that at least k consecutive components work 1: acc ← 0 2: h ← p1 . . . pk 3: for i = 1 to n − k do 4: acc ← acc + (1 − pi+k )h 5: h ← hpi+k /pi 6: end for 7: acc ← acc + h In the case of equal component reliabilities we even obtain a closed formula: Corollary 6.2.8 [Doh03] Let Σ = (E, φ) be a consecutive k-out-of-n success system with component reliabilities given by p = (p, . . . , p). If k ≥ n/2, then RelΣ (p) = (n − k + 1)pk − (n − k)pk+1 = pk [(n − k)(1 − p) + 1] . Proof. Corollary 6.2.8 is an immediate consequence of Theorem 6.2.7. 2

6.2 Special reliability systems

73

Remarks. Theorem 6.2.7 and Corollary 6.2.8 can equivalently be formulated for consecutive k-out-of-n failure systems by replacing the preceding identities with n−k+1  i+k−1 (

1 − RelΣ (p) =

i=1

j=i

qj −

n−k  i+k ( i=1 j=i

qj =

n−k 

(1 − qi+k )

i=1

i+k−1 (

qj +

j=i

n (

qj

j=n−k+1

in the case of unequal component failure probabilities qj = 1 − pj , and 1 − RelΣ (p) = (n − k + 1)q k − (n − k)q k+1 = q k [(n − k)(1 − q) + 1] in the case where all component failure probabilities are equal to q = 1 − p. This latter identity was first proved by Shanthikumar [Sha82] without making use of the inclusion-exclusion principle. It should be noted at this point that the algorithms of Shanthikumar [Sha82] and Hwang [Hwa82] are the most efficient algorithms for computing the reliability of a consecutive k-out-of-n system. Note, however, that these most efficient algorithms strongly rely on the assumption that the components of the system fail in a statistically independent fashion.

Example 6.2.9 As in Example 6.2.6 we consider the consecutive 4-out-of-7 failure system associated with the network in Figure 6.11. In view of the preceding remarks the reliability of this consecutive 4-out-of-7 failure system is 1−q1 q2 q3 q4 −q2 q3 q4 q5 −q3 q4 q5 q6 −q4 q5 q6 q7 +q1 q2 q3 q4 q5 +q2 q3 q4 q5 q6 +q3 q4 q5 q6 q7 = 1 − (1 − q5 )q1 q2 q3 q4 − (1 − q6 )q2 q3 q4 q5 − (1 − q7 )q3 q4 q5 q6 − q4 q5 q6 q7 , where qi denotes the failure probability of node i (i = 1, . . . , 7). In particular, if the qi ’s are all equal to q, then the reliability is easily seen to be 1 − 4q 4 + 3q 5 . We close this section with a generalization of Theorem 6.2.7, to which the remarks after Corollary 6.2.8 likewise apply. Theorem 6.2.10 Let t ∈ and n1 , . . . , nt , k1 , . . . , kt ∈ such that n1 < · · · < nt ≤ n1 +k1 < · · · < nt +kt . Let Σ be a coherent binary system with components n1 , . . . , nt + kt − 1, minpaths {n1 , . . . , n1 + k1 − 1}, . . . , {nt , . . . , nt + kt − 1} and component reliabilities given by p = (pn1 , . . . , pnt +kt −1 ). Then, RelΣ (p) equals   ni+1 +ki+1 −1 ni +k t ni +k t−1 ni+1 +k t−1  (i −1  (i+1 −1  ( (i −1 nt +k (t −1 1 − pj − pj = pj  pj + pj . i=1

j=ni

i=1

j=ni

i=1

j=ni +ki

j=ni

j=nt

Proof. For i = 1, . . . , t let Ai be the event that components ni , . . . , ni + ki − 1 operate. It follows that RelΣ (p) = P (A1 ∪ · · · ∪ At ) and, by the requirements of the theorem, Ax ∩ Ay ⊆ Az for x, y = 1, . . . , t and any z between x and y.

74

6

Reliability Applications

Therefore, the same argument as in the proof of Theorem 6.2.7 reveals that RelΣ (p) =

t 

P (Ai ) −

i=1

=

=

P (Ai ∩ Ai+1 ) =

i=1 t−1 ni+1 +k  (i+1 −1

i=1

i=1

pj −

j=ni

t−1 ni+1 +k  (i+1 −1 pj − pj

t ni +k  (i −1 i=1

t−1 ni +k  (i −1

pj +

j=ni

i=1

j=ni

nt +k (t −1

j=ni

pj

j=nt

nt +k t−1 ni +k t−1 ni +k  (i −1  (i −1 ni+1 +k (i+1 −1 (t −1 pj − pj pj + pj i=1

=

t−1 

j=ni

t−1 



1 −

i=1 ni+1 +ki+1 −1

(



pj 

j=ni +ki

i=1

j=ni +ki

j=ni

j=nt

ni +k (i −1

nt +k (t −1

j=ni

j=nt

pj +

pj .

2

Example 6.2.11 Consider the network in Figure 6.12, where nodes 1–7 are assumed to fail randomly and independently with probabilities q1 , . . . , q7 and all other nodes and edges are perfectly reliable. Again, we are interested in the probability that a message can pass from s to t along a path of operating nodes. Thus, an appropriate model is a coherent binary system having components 1– 7 and mincuts {1, 2, 3}, {2, 3, 4}, {3, 4, 5, 6}, {4, 5, 6, 7}. By applying the mincut analogue of Theorem 6.2.10 the reliability of this system is easily seen to be 1 − q1 q2 q3 − q2 q3 q4 − q3 q4 q5 q6 − q4 q5 q6 q7 + q1 q2 q3 q4 + q2 q3 q4 q5 q6 + q3 q4 q5 q6 q7 = 1 − (1 − q4 )q1 q2 q3 − (1 − q5 q6 )q2 q3 q4 − (1 − q7 )q3 q4 q5 q6 − q4 q5 q6 q7 , which equals 1 − 2q 3 − q 4 + 2q 5 if all node failure probabilities are equal to q.

s

1

2

3

4

5

6

7

t

Figure 6.12: A consecutive network.

Remark. The inclusion-exclusion identity of Theorem 6.2.10 contains only noncancelling terms. The coefficient of each such term is either +1 or −1. Due to Shier and McIlwain [SM91], this ±1 property holds for any coherent binary system whose minpaths (resp. mincuts) are consecutive sets of components. This strongly generalizes a corresponding result of Kossow and Preuss [KP89], who showed that this ±1 property holds for any consecutive k-out-of-n system.

6.3

6.3

Reliability covering problems

75

Reliability covering problems

Reliability covering problems were introduced by Ball, Provan and Shier [BPS91, Shi91] as an abstract model for several types of reliability problems such as evaluating the probability that in a mass transit system with reliable stops and unreliable routes each stop is served by a route. Further examples include determining the reliability of flight schedules for aircraft [BPS91] and of maintaining continuous surveillance of a critical segment of a country’s border [Shi91].

6.3.1

The hypergraph model

As in [BPS91, Shi91] we use the terminology of hypergraphs: Definition 6.3.1 A hypergraph is a couple H = (V, ) where V is a finite set and  is a set of non-empty subsets of V . The elements of V resp.

 are the vertices resp. edges of H. A covering of V is a subset  of  such that  = V . In case of a mass transit system, the vertices and edges of the hypergraph correspond to the stops and routes of the system, respectively, while the coverings correspond to those sets of routes such that each stop is served by a route. Throughout, we make the assumption that the vertices of H are perfectly reliable, whereas the edges fail randomly and independently with known probabilities, which are given by a vector p = (pE )E∈ ∈ [0, 1] . The general objective is to determine or compute bounds on the probability that the operating edges of H cover the whole vertex-set of H, which is abbreviated to Cov(H; p). Besides their practical applicability, reliability covering problems are also interesting from a theoretical point of view. Namely, any coherent binary system gives rise to an equivalent reliability covering problem, and vice versa [BPS91, Shi91]. In this way, the results of the preceding sections (including those on network reliability) can be reformulated as reliability covering problems. One direction of this equivalence is used in deriving the following results on reduced hypergraphs, i.e., hypergraphs of type H = (V, ) where the sets (v) := {E ∈  | v ∈ E} (v ∈ V ) are all distinct. Note that restricting to reduced hypergraphs does not cause any loss of generality, since by deleting vertices any hypergraph H can be transformed efficiently into a reduced hypergraph R(H) such that Cov(H; p) = Cov(R(H); p). Now, using the above sets, the coverage probability can be expressed as     (6.28) Cov(H; p) = 1 − P  {E fails} . v∈V E∈(v)

6.3.2

Abstract tubes and polynomial algorithms

In connection with Theorem 3.1.9 the first part of the following theorem, which is implicit in [Doh99b], yields improved inclusion-exclusion identities and Bonferroni inequalities for the last term in (6.28) and thus for the coverage probability

76

6

Reliability Applications

Cov(H; p). We do not mention the improved inclusion-exclusion identities explicitly, since they follow immediately from the corresponding inequalities. As in the preceding sections, we use (V ) to denote the order complex of V . Theorem 6.3.2 [Doh99b] Let H = (V, ) be a reduced hypergraph whose edges fail randomly and independently and whose vertex-set is given the structure of a lower semilattice such that the complement of each edge is infimum-closed. Then, ,   {E fails} , (V ) E∈(v)

v∈V

is an abstract tube. In particular, for any p = (pE )E∈ ∈ [0, 1] and any r ∈ , (6.29)

Cov(H; p) ≤



(−1)|I|



Cov(H; p) ≥



qE

(r even),

qE

(r odd),

E∈ E∩I=∅

I⊆V,|I|≤r I is a chain

(6.30)

(

(−1)|I|

I⊆V,|I|≤r I is a chain

(



E∈ E∩I=∅

where qE = 1 − pE for any E ∈ . Proof. Consider the coherent binary system Σ = (, φ) where for any subset  of , φ( \ ) = 0 if and only if  ⊇ (v) for some v ∈ V . Clearly := {(v) | v ∈ V } is an extended set of mincuts of this system. Now, by defining (v) ≤ (w) :⇔ v ≤ w for any v, w ∈ V , a partial ordering relation on is established, which satisfies (v) ∧ (w) = (v ∧ w) for any v, w ∈ V . Since the complement of each edge is infimum-closed, (v ∧ w) ⊆ (v) ∪ (w) and therefore, (v) ∧ (w) ⊆ (v) ∪ (w) for any v, w ∈ V . Hence, the requirements of the extended mincut version of Theorem 6.1.3 are satisfied, and thus the first part of the theorem follows. The second part follows from Theorem 3.1.9. 2 Remarks. Theorem 6.3.2 can be generalized to reduced hypergraphs with statistically dependent edge failures: Simply replace the product in (6.29) and (6.30) with the probability that all edges having a non-empty intersection with I fail. Note that the requirements of Theorem 6.3.2 (as well as those of the next theorem) are satisfied if H = ({1, . . . , n}, ) where  ⊆ {{k, . . . , l} | 1 ≤ k ≤ l ≤ n} and n ∈ . For an application of this type of hypergraph, see Shier [Shi91]. Theorem 6.3.3 [BPS91, Shi91] Let H = (V, ) be a reduced hypergraph whose edges fail randomly and independently according to some vector q = (qE )E∈ of edge failure probabilities and whose vertex-set is a lower semilattice such that each edge is convex and the complement of each edge is infimum-closed. Then,  Cov(H; p) = 1 − (6.31) Λ(v, q) , v∈V

6.3

Reliability covering problems

77

where p = 1 − q and where Λ is defined by the following recursive scheme: (6.32)

Λ(v, q) :=

( E∈(v)

qE −



Λ(w, q)

(  

qE .

E∈ (v) E∈ / (w)

w 0 for any planar graph G. To state the main result of this section, a further definition is needed. Definition 7.2.2 Let G be a graph whose edge-set is endowed with a linear ordering relation. A broken circuit of G is obtained from the edge-set of a cycle of G by removing its maximum edge. The broken circuit complex of G, abbreviated to BC(G), is the abstract simplicial complex consisting of all nonempty subsets of the edge-set of G that do not include any broken circuit of G as a subset. The definition of a broken circuit goes back to Whitney [Whi32], while the broken circuit complex is due to Wilf [Wil76] (see also [Bry77, BO81, BZ91]). Example 7.2.3 Consider the graph in Figure 7.1, whose edge-set is linearly ordered according to the labelling of the edges. Obviously, the broken circuits are {1, 2}, {1, 2, 4} and {3, 4}, whence the broken circuit complex is equal to {{1}, {2}, {3}, {4}, {5}, {1, 3}, {1, 4}, {1, 5}, {2, 3}, {2, 4}, {2, 5}, {3, 5}, {4, 5}, {1, 3, 5}, {1, 4, 5}, {2, 3, 5}, {2, 4, 5}}.

We now state the main result of this section, which is an abstract tube generalization of Whitney’s broken circuit theorem [Whi32]. For an alternative proof of the associated inequalities, the reader is referred to [Doh99e].

7.2 Chromatic polynomials and broken circuits

1

85

2 3

4

5

Figure 7.1: A graph with labelled edges.

Theorem 7.2.4 [Doh99d] Let G be a graph whose edge-set is endowed with a linear ordering relation. Then, for any λ ∈ ,    (7.1) {f : V (G) → {1, . . . , λ} | f (v) = f (w)} {v,w}∈E(G) , BC(G) is an abstract tube, and for any r ∈ (7.2) (7.3)

PG (λ) ≥ PG (λ) ≤

r  k=0 r 

the following inequalities hold:

(−1)k bk (G) λn(G)−k

(r odd),

(−1)k bk (G) λn(G)−k

(r even),

k=0

where b0 (G) = 1 and bk (G), k > 0, counts the faces of cardinality k (dimension k − 1) in BC(G). Proof. Define  as the set of broken circuits of G, and for any edge e of G define Ae := {f : V (G) → {1, . . . , λ} | f (v) = f (w)} ,

e = {v, w}.

Then, in the sense of Definition 4.1.18, (E, ) = BC(G). Now, by applying Corollary 4.1.19 (with E instead of V ) we immediately find that (7.1) is an abstract tube. From this and Theorem 3.1.9 we conclude that   r      n(G) k PG (λ) ≥ λ (7.4) + (−1) Ai  (r odd),    I∈BC(G) k=1

(7.5)

PG (λ)

≤

|I|=k

i∈I

  r      n(G) k λ + (−1) Ai     I∈BC(G) k=1

|I|=k

(r even).

i∈I

Since the edge-subgraph G[I] is cycle-free for any I ∈ BC(G), it follows that (7.6)

m(G[I]) − n(G[I]) + c(G[I]) = 0

7

86

Combinatorial Applications and Related Topics

and hence,       Ai  = λn(G)−n(G[I]) λc(G[I]) = λn(G)−m(G[I]) = λn(G)−|I| .    i∈I

By putting this into (7.4) and (7.5) and collecting powers of λ, the inequalities (7.2) and (7.3) are proved. 2 As a corollary we deduce Whitney’s broken circuit theorem [Whi32]. Corollary 7.2.5 [Whi32] Let G be a graph whose edge-set is endowed with a linear ordering relation. Then, for any λ ∈ , 

n(G)

(7.7)

PG (λ) =

(−1)k bk (G) λn(G)−k ,

k=0

where bk (G) is defined as above. Proof. Corollary 7.2.5 is an immediate consequence of Theorem 7.2.4. 2 As a further corollary we conclude that in absolute value each coefficient of the chromatic polynomial is at most the sum of its left and right neighbour: Corollary 7.2.6 Let the chromatic polynomial of G be given by (7.7). Then, bk (G) ≤ bk−1 (G) + bk+1 (G)

(k = 1, . . . , n(G) − 1).

Proof. We only consider the case where k is odd. Then, by applying Theorem 7.2.4 with λ = 1 and r = k − 2 resp. r = k + 1 we get k−2 

(−1)i bi (G) ≤

i=0

k+1 

(−1)i bi (G) ,

i=0

which gives the result. 2

7.2.2

The generalized chromatic polynomial

In the sequel, we consider a new two-variable generalization of the chromatic polynomial and establish a corresponding generalization of Theorem 7.2.4. As shown in [DPT03], this new polynomial simultaneously generalizes the chromatic polynomial, the independence polynomial and the matching polynomial, and is closely related to Stanley’s chromatic symmetric function [Sta99]. Definition 7.2.7 [DPT03] Let G = (V, E) be a graph. For any λ ∈ and µ ∈ {0, . . . , λ} we use PG (λ, µ) to denote the number of all mappings f : V → {1, . . . , λ} such that f (v) = f (w) or f (v) = f (w) > µ for any {v, w} ∈ E.

7.2 Chromatic polynomials and broken circuits

87

Remark. Evidently, PG (λ, 0) = λn(G) and PG (λ, λ) = PG (λ). Moreover, PG (2, 1) gives the number of independent sets in G. Recall that a subset W ⊆ V (G) is independent if the vertex-induced subgraph G[W ] has no edges. Theorem 7.2.8 Let G be a graph whose edge-set is endowed with a linear ordering relation, and let λ ∈ and µ ∈ {0, . . . , λ}. Then    {f : V (G) → {1, . . . , λ} | f (v) = f (w) ≤ µ} {v,w}∈E(G) , BC(G) is an abstract tube, and for any r ∈ (7.8)

PG (λ, µ) ≥

r  k 

the following inequalities hold:

(−1)k bk,l (G) λn(G)−k−l µl

(r odd),

(−1)k bk,l (G) λn(G)−k−l µl

(r even),

k=0 l=0

(7.9)

PG (λ, µ) ≤

r  k  k=0 l=0

where b0,0 (G) = 1 and bk,l (G), k > 0, counts all faces I of cardinality k (dimension k − 1) in BC(G) such that the edge-subgraph G[I] has l connected components. Proof. Define  as the set of broken circuits of G, and for any edge e of G define Ae := {f : V (G) → {1, . . . , λ} | f (v) = f (w) ≤ µ} ,

e = {v, w}.

As in the proof of Theorem 7.2.4, the first part is an immediate consequence of Corollary 4.1.19. In combination with Theorem 3.1.9 the first part gives   r      (7.10) PG (λ, µ) ≥ λn(G) + (−1)k Ai  (r odd),    I∈BC(G) k=1

(7.11)

PG (λ, µ)

≤ λn(G)

|I|=k

i∈I

  r      + (−1)k Ai     I∈BC(G) k=1

|I|=k

(r even).

i∈I

Since G[I] is cycle-free for any I ∈ BC(G), we again have (7.6) and hence,       Ai  = λn(G)−n(G[I]) µc(G[I]) = λn(G)−m(G[I])−c(G[I])µc(G[I]) .    i∈I

Now, by putting this expression into (7.10) and (7.11) and taking account of c(G[I]) ≤ m(G[I]) = |I|, the inequalities (7.8) and (7.9) are proved. 2

88

7

Combinatorial Applications and Related Topics

Corollary 7.2.9 [DPT03] Let G be a graph whose edge-set is endowed with a linear ordering relation. Then, for any λ ∈ and any µ ∈ {0, . . . , λ}, k  

m(G)

(7.12)

PG (λ, µ) =

(−1)k bk,l (G) λn(G)−k−l µl ,

k=0 l=0

where b0,0 (G) = 1 and bk,l (G), k > 0, counts all faces I of cardinality k (dimension k − 1) in BC(G) such that G[I] has l connected components. Proof. Corollary 7.2.9 is an immediate consequence of Theorem 7.2.8. 2 Remarks. Some remarks on Theorem 7.2.8 and Corollary 7.2.9 follow. 1. The preceding equation (7.12) can equivalently be stated as   PG (λ, µ) = λn(G) Q G; −λ−1 , µλ−1 , where Q is defined by Q(G; x, y) =



bk,l (G) xk y l .

k,l

Thus, Q(G; −λ−1 , µλ−1 ) expresses the probability that a λ-coloring of G, which is chosen uniformly at random, is admissible in the sense of Definition 7.2.7. 2. Statements on the number of independent sets are obtained from the preceding results by putting λ = 2 and µ = 1. The corresponding abstract tube is    {W | W ⊆ V (G), e ⊆ W } e∈E(G) , BC(G) . 3. We further remark that non-isomorphic trees on the same number of vertices may have different polynomials in λ and µ. This contrasts the situation for the usual chromatic polynomial, which equals λ(λ − 1)n−1 for all trees on n vertices. Consider for instance a path G and a star G , both on four vertices. Then, b0,0 (G) = 1,

b2,2 (G) = 1,

b0,0 (G ) = 1,

b2,2 (G ) = 0,

b1,0 (G) = 0, b1,1 (G) = 3,

b3,0 (G) = 0, b3,1 (G) = 1,

b1,0 (G ) = 0, b1,1 (G ) = 3,

b3,0 (G ) = 0, b3,1 (G ) = 1,

b2,0 (G) = 0, b2,1 (G) = 2,

b3,2 (G) = 0, b3,3 (G) = 0,

b2,0 (G ) = 0, b2,1 (G ) = 3,

b3,2 (G ) = 0, b3,3 (G ) = 0.

Putting these values into (7.12) we obtain PG (λ, µ) = λ4 − 3λ2 µ + 2λµ + µ2 − µ, PG (λ, µ) = λ4 − 3λ2 µ + 3λµ − µ. Evidently, these two polynomials differ unless µ = λ or µ = 0. Consequently, the generalized chromatic polynomial PG (λ, µ) is not an evaluation of the Tutte polynomial [Tut47, Tut54], which is the same for all trees on a given number of vertices. We shall be concerned with the Tutte polynomial in Section 7.4.

7.3 Sums over partially ordered sets

7.3

89

Sums over partially ordered sets

In this section we strongly generalize the improved inclusion-exclusion identity associated with the abstract tube of Theorem 4.2.3. The result is a general theorem on sums over partially ordered sets, which has applications not only to the inclusion-exclusion principle, but as we will see in Section 7.4 and Section 7.5, also to matroid polynomials, Euler characteristics and M¨ obius functions.

7.3.1

A general theorem on sums

The following definition generalizes the concept of a kernel operator resp. closure operator from power set lattices to arbitrary partially ordered sets. Definition 7.3.1 Let P be a partially ordered set. A mapping k : P → P is a kernel operator if for any x, y ∈ P , (i) (ii) (iii)

k(x) ≤ x (intensionality), x ≤ y ⇒ k(x) ≤ k(y) (monotonicity), k(k(x)) = k(x) (idempotence).

Dually, a mapping c : P → P is a closure operator if for any x, y ∈ P , (i) (ii) (iii)

x ≤ c(x)

(extensionality),

x ≤ y ⇒ c(x) ≤ c(y) (monotonicity), c(c(x)) = c(x) (idempotence).

An element x ∈ P is called k-open if k(x) = x and c-closed if c(x) = x. Kernel and closure operators for partially ordered sets were introduced by Ward [War42] and extensively studied by Rota [Rot64]; see also Aigner [Aig79]. The main result of this section is the following: Theorem 7.3.2 [Doh99f] Let P be an upper-finite partially ordered set, and let f and g be mappings from P into an abelian group such that f (x) = y≥x g(y) for any x ∈ P . Furthermore, let k : P → P be a kernel operator, and let x0 be a k-open element of P such that f (x) = 0 for any k-open x > x0 . Then,  g(y) . f (x0 ) = y : k(y)=x0

The following proof of Theorem 7.3.2 is due to an anonymous referee. For the author’s original proof, see [Doh99f]. Proof. It suffices to show that y : k(y)>x0 g(y) = 0. If x0 is maximal in P , then the statement holds. We proceed by downward induction on x0 . In this way,     g(y) = g(y) = f (x) = 0 , y : k(y)>x0

x>x0 x k-open

y : k(y)=x

x>x0 x k-open

7

90

Combinatorial Applications and Related Topics

where the second equality comes from the induction hypothesis and the third from the hypothesis of the theorem. 2 Dualizing Theorem 7.3.2 we obtain Theorem 7.3.3 [Doh99f] Let P be a lower-finite partially ordered set, and let f and g be mappings from P into an abelian group such that f (x) = y≤x g(y) for any x ∈ P . Furthermore, let c : P → P be a closure operator and x0 a c-closed element of P such that f (x) = 0 for any c-closed x < x0 . Then,  g(y) . f (x0 ) = y : c(y)=x0

From the preceding theorem we now deduce a prominent result of Rota [Rot64] on the M¨ obius function of a poset (see Definition 6.1.12), which in turn specializes to Weisner’s theorem [Wei35]. The proof is adopted from [Doh99f]. Corollary 7.3.4 [Rot64] Let P be a lower-finite partially ordered set with least element ˆ 0 and c : P → P a closure operator where c(ˆ0) > ˆ0. Then, for all x0 ∈ P ,  µP (y) = 0 . y : c(y)=x0

Proof. For any x ∈ P define f (x) := y≤x µP (y). There is nothing to prove if x0 is not c-closed. Otherwise, x0 = ˆ 0 and hence by (6.3), f (x0 ) = 0. Likewise, one finds that f (x) = 0 for any c-closed x < x0 . Now apply Theorem 7.3.3. 2 The preceding corollary implies one of the most important classical results of enumerative combinatorics, which is known as Weisner’s theorem [Wei35]. Corollary 7.3.5 [Wei35] Let P be a lower-finite partially ordered set with least element ˆ 0. Then, for any a > ˆ 0 and all x0 ∈ P ,  µP (y) = 0 . y : y∨a=x0

Proof. Define c(y) := y ∨ a for any y ∈ P and apply Corollary 7.3.4. 2

7.3.2

Application to inclusion-exclusion

We proceed by showing how the inclusion-exclusion identity       (7.13) χ Av = (−1)|I|−1 χ Ai v∈V




I∈ ∗ (V ) k(I)=∅

i∈I

associated with the abstract tube of Theorem 4.2.3 may be deduced from Theorem 7.3.2. Recall that by the requirements of Theorem 4.2.3, {Av }v∈V is a finite

7.4 Matroid polynomials and the β invariant

91

family of sets,  and k is

a kernel operator on V (
(V ) w.r.t. Definition 7.3.1) such that x∈X Ax ⊆ v∈X / Av for any non-empty and k-open subset X of V . Proof of (7.13). For any subsets X and Y of V define   
   f (X) := (−1)|X| χ Ax ∩ Av ; g(Y ) := (−1)|Y | χ  Ay  x∈X

v ∈X /

y∈Y





v∈V

I:k(I)=∅

Av = Ω (the universal where, by convention, x∈∅ Ax = v∈V / set). By applying the principle of inclusion-exclusion we find that f (X) = Y ⊇X g(Y ) for any subset X of V . Now, by the requirements, f (X) = 0 for any non-empty and k-open subset X of V . Hence, by applying Theorem 7.3.2 with P =
(V ),       χ Av = (−1)|I| χ Ai , from which identity (7.13) follows since χ

7.4



i∈I





Av = 1 − χ

 Av . 2

Matroid polynomials and the β invariant

Our general result on sums over partially ordered sets, which we established in the previous section, is now applied to the Tutte polynomial, the characteristic polynomial and the β invariant of a matroid. In this way, some of the identities obtained for the chromatic polynomial in Section 7.2 are generalized to matroids.

7.4.1

The Tutte polynomial

In order to introduce the Tutte polynomial, we briefly review some basics of matroid theory. For a detailed exposition, the reader is referred to Welsh [Wel76]. Definition 7.4.1 A matroid is a pair M = (E, r) consisting of a finite set E and a -valued function r on the power set of E such that for any A, B ⊆ E, (i) (ii) (iii)

0 ≤ r(A) ≤ |A|, A ⊆ B ⇒ r(A) ≤ r(B), r(A ∪ B) + r(A ∩ B) ≤ r(A) + r(B).

Example 7.4.2 If G is a graph and r(I) = n(G[I]) − c(G[I]) for any I ⊆ E(G), then M (G) := (E(G), r) is a matroid, which is called the cycle matroid of G. Definition 7.4.3 The Tutte polynomial T (M ; a, b) of a matroid M = (E, r) is defined by  T (M ; a, b) := (a − 1)r(E)−r(I)(b − 1)|I|−r(I) , I⊆E

where a and b are independent variables.

92

7

Combinatorial Applications and Related Topics

Specializations of the Tutte polynomial T (M (G); a, b) count various objects associated with a graph G, e.g., subgraphs, spanning trees, acyclic orientations and proper λ-colorings. It is also related to the all-terminal reliability R(G). Namely, if G is a connected graph whose nodes are perfectly reliable and whose edges fail randomly and independently with equal probability q = 1 − p, then R(G) = q m(G)−n(G)+1 pn(G)+1 T (M (G); 1, q −1 ) . For further applications of the Tutte polynomial, we refer to [BO92, Wel93]. Definition 7.4.4 For any matroid M = (E, r) and any subset X of E, the contraction of X from M is defined by M/X := (E \ X, rX ) where rX (I) := r(X ∪ I) − r(X) for any I ⊆ E \ X. Our result on the Tutte polynomial is stated below. An equivalent formulation using union-closed sets instead of kernel operators is left to the reader. Theorem 7.4.5 [Doh99f] Let M = (E, r) be a matroid, k a kernel operator on E and a, b ∈  such that T (M/X; a, b) = 0 for any k-open X ∈
∗ (E). Then, 

T (M ; a, b) =

(a − 1)r(E)−r(I) (b − 1)|I|−r(I) .

I : k(I)=∅

Proof. For any subset X of E define f (X) :=



(a − 1)r(E)−r(I) (b − 1)|I|−r(I) .

I⊇X

Then, by Theorem 7.3.2, it suffices to prove that f (X) = 0 for any non-empty and k-open subset X of E. With X denoting the complement of X in E and h(X) := (a − 1)r(E)−r(X)−rX (X) (b − 1)|X|−r(X) we obtain f (X) =



(a − 1)r(E)−r(X∪I) (b − 1)|X∪I|−r(X∪I)

I⊆X

= h(X)



(a − 1)rX (X)−rX (I) (b − 1)|I|−rX (I)

I⊆X

= h(X) T (M/X; a, b) = 0 which gives the result. 2

7.4 Matroid polynomials and the β invariant

7.4.2

93

The characteristic polynomial

From Theorem 7.4.5 we now deduce a theorem due to Heron [Her72], which generalizes Whitney’s broken circuit theorem [Whi32] from graphs to matroids. Definition 7.4.6 The characteristic polynomial C(M ; λ) of a matroid M = (E, r) is defined by  C(M ; λ) := (−1)r(E) T (M ; 1 − λ, 0) = (7.14) (−1)|I| λr(E)−r(I) . I⊆E

A circuit of a matroid M = (E, r) is set C ∈
(E) such that r(C \ {c}) = |C| − 1 = r(C) for any c ∈ C. A loop is a circuit of cardinality 1. If E is linearly ordered and C a circuit of M , then C \ {max C} is called a broken circuit of M . ∗

Remark. By the classical inclusion-exclusion principle, the chromatic polynomial of a graph G is related to the characteristic polynomial of M (G) by PG (λ) = λc(G) C(M (G); λ) . We are now ready to state Heron’s broken circuit theorem [Her72]. The proof is adopted from [Doh00d]. Theorem 7.4.7 [Her72] Let M = (E, r) be a matroid, where E is endowed with a linear ordering relation. Then, 

r(E)

C(M ; λ) =

(−1)k bk (M ) λr(E)−k

k=0

where bk (M ) is the number of k-subsets of E which do not include a broken circuit of M as a subset. Proof. By Lemma 1.4 of Heron [Her72] we have (7.15)

C(M ; λ) = 0 if M contains a loop.

For any X ⊆ E define k(X) :=

 {C ⊆ X | C is a broken circuit of M }.

Then, for any non-empty and k-open X ⊆ E there is some e > max X such that X ∪ {e} includes a circuit of M . Therefore, r(X ∪ {e}) = r(X), or equivalently, rX ({e}) = 0. From this we conclude that e is a loop of M/X and hence by (7.15), C(M/X; λ) = 0. By applying Theorem 7.4.5, using identity (7.14) and the fact that r(I) = |I| for any I including no broken circuit of M , we obtain  C(M ; λ) = (−1)|I| λr(E)−|I| , I⊆E I⊇X(∀X∈

)

where  is the set of broken circuits of M . 2

7

94

7.4.3

Combinatorial Applications and Related Topics

The β invariant

Results similar to Theorem 7.4.5 and Theorem 7.4.7 can also be established for Crapo’s β invariant [Cra67], which among other things indicates whether M is connected and whether M is the cycle matroid of a series-parallel network. Definition 7.4.8 The β invariant of a matroid M = (E, r) is defined by  (7.16) β(M ) := (−1)r(E) (−1)|I| r(I) . I⊆E

A matroid M = (E, r) is called disconnected if there is a pair of distinct elements of E that are not jointly contained by a circuit of M . Remark. By Lemma II of Crapo [Cra67], β(M ) = 0 if M is disconnected or a loop. This will be used in the proof of the following theorem. Theorem 7.4.9 [Doh99f] Let M = (E, r) be a matroid, and let k be a kernel operator on E such that E is not k-open and M/X is disconnected or a loop for any non-empty and k-open subset X of E. Then,  β(M ) = (−1)r(E) (−1)|I| r(I) . I : k(I)=∅

Proof. By the assumptions of the theorem and the preceding remark, β(M/X) = 0 whenever X ⊆ E is non-empty and k-open. For any subset X of E define  f (X) := (−1)|I| r(I) . I⊇X

Then, by Theorem 7.3.2 it suffices to show that f (X) = 0 for any non-empty and k-open subset X of E. By the assumptions, X is non-empty and therefore,  .  . (−1)|X∪I| r(X ∪ I) = (−1)|I| (rX (I) + r(X)) = β(M/X) = 0, f (X) = I⊆X

I⊆X

. where a = b means that a = b or a = −b. 2 As a corollary we obtain the following well-known result: Corollary 7.4.10 Let M = (E, r) be a matroid, where E is endowed with a linear ordering relation. Then, β(M ) = (−1)

r(E)

|E| 

(−1)k k bk (M )

k=1

where again bk (M ) is the number of k-subsets of E including no broken circuit. Proof. Corollary 7.4.10 follows from Theorem 7.4.9 in the same way as Theorem 7.4.7 follows from Theorem 7.4.5. 2

7.5 Euler characteristics and M¨obius functions

7.5

95

Euler characteristics and M¨ obius functions

In this section, our general result on sums over partially ordered sets, which we established in Theorem 7.3.2, is applied to the Euler characteristic of an abstract simplicial complex and to the M¨obius function of a finite lattice.

7.5.1

Euler characteristics

Our result on the Euler characteristic requires the following definition. Definition 7.5.1 For any abstract simplicial complex and any simplex X ∈ the link /X is the abstract simplicial complex defined by /X := {I ∈ | I ∩ X = ∅, I ∪ X ∈ } . Theorem 7.5.2 Let k be a kernel operator on the vertex-set of an abstract simplicial complex such that γ( /X) = 1 for any k-open X ∈ . Then, γ( ) = γ ({I ∈ | k(I) = ∅}) . Proof. Let

+

:=

f (X) :=

∪ {∅}, and for any subset X of the vertex-set of 

g(Y ), where

g(Y ) :=

Y ⊇X

(−1)|Y |−1 0

define

if Y ∈ + , otherwise.

Then, f (X) = 0 for any non-empty X ⊆ Vert( ) which is not a simplex of , and f (X) = (−1)|X|−1 +



(−1)|Y |−1 = (−1)|X|−1 + (−1)|X| γ( /X) = 0

Y∈ Y ⊃X

for any k-open simplex X ∈ . Hence, by applying Theorem 7.3.2,  I∈

+

(−1)|I|−1 =



(−1)|I|−1 .

I∈ + k(I)=∅

Now, by subtracting the term for I = ∅ from both sides the result follows. 2 In view of Theorem 7.5.2 we pose the following conjecture. For the notion of homotopy equivalence, the reader is referred to the textbook of Harzheim [Har78]. Conjecture 7.5.3 Let k be a kernel operator on the vertex-set of an abstract simplicial complex such that /X is contractible for any k-open simplex X ∈ . Then, the complexes and {I ∈ | k(I) = ∅} are homotopy equivalent.

96

7.5.2

7

Combinatorial Applications and Related Topics

M¨ obius functions

Our result on M¨ obius functions, stated as Theorem 7.5.8 below, is an extension of Rota’s crosscut theorem [Rot64] as well as a generalization of a result due to Blass and Sagan [BS97]. We need the following notions due to Rota [Rot64]. Definition 7.5.4 Let L = [ˆ 0, ˆ 1] be a finite lattice. L is called non-trivial if ˆ ˆ L \ {0, 1} = ∅. An element a ∈ L is an atom of L if a > ˆ0 and no a ∈ L satisfies a > a > ˆ 0. The set of atoms of L is denoted by A(L). A crosscut of L is an antichain C ⊆ L\{ˆ 0, ˆ 1} having a non-empty intersection with any maximal chain ˆ ˆ from 0 to 1 in L. The crosscut complex of L associated with C is defined by    Γ(L, C) := I ∈
∗ (C)  I > ˆ0 or I < ˆ1 and is easily seen to be an abstract simplicial complex. For any abstract simplicial complex , we refer to γ˜ ( ) := γ( ) − 1 as the reduced Euler characteristic of . Example 7.5.5 A(L) is a crosscut for any non-trivial finite lattice L. The following proposition, which is stated without proof, is known as Rota’s crosscut theorem [Rot64]. A proof of Rota’s crosscut theorem can also be found in the textbook of Aigner [Aig79]. In the following, we write µ(L) instead of µL (ˆ 1), where µL denotes the M¨obius function of L (see Definition 6.1.12). Proposition 7.5.6 [Rot64] Let L = [ˆ 0, ˆ 1] be a non-trivial finite lattice, and let C be a crosscut of L. Then, µ(L) equals the reduced Euler characteristic of the crosscut complex of L associated with C, or equivalently,  (7.17) µ(L) = (−1)|I| . ∗
Ï(C) ÎI∈ I=ˆ 0, I=ˆ 1 Example 7.5.7 Let L be the lattice shown in Figure 7.2. Evidently, C = {a, b, c, d, e} is a crosscut of L. The associated crosscut complex Γ(L, C) equals {{a}, {b}, {c}, {d}, {e}, {a, b}, {a, c}, {a, d}, {b, c}, {b, d}, {b, e}, {c, d}, {c, e}, {d, e}, {a, b, c}, {a, b, d}, {a, c, d}, {b, c, d}, {b, c, e}, {b, d, e}, {c, d, e}, {a, b, c, d}, {b, c, d, e}} and contains 23 simplices each of which contributes to the reduced Euler characteristic γ˜ (Γ(L, C)) = µ(L). Alternatively, we can compute µ(L) via (7.17), which gives a sum involving only eight terms. Anyway, µ(L) = 0. Our result on M¨ obius functions shows that µ(L) may be considered as the reduced Euler characteristic of a subcomplex of Γ(L, C).

7.5 Euler characteristics and M¨obius functions

97

1

a

b

c

d

e

0

Figure 7.2: A lattice with crosscut {a, b, c, d, e}.

ˆ 1] ˆ be a non-trivial finite lattice, and let C be a Theorem 7.5.8 Let L = [0, crosscut of L, which is endowed with a partial order which is denoted by  to distinguish it from the partial order ≤ in L. Furthermore, let k be a kernel operator on C such that for any non-empty and  k-open subset X of C and any x ∈ X there is some c ∈ C satisfying c  x and X < c < X. Then, (7.18)

µ(L) = γ˜ ({I ∈ Γ(L, C) | k(I) = ∅})

or equivalently, (7.19)

µ(L) =

 Î
Ï

(−1)|I| .

I∈ ∗ (C) I=ˆ 0, I=ˆ 1 k(I)=∅

Proof. By Theorem 7.5.2 and Proposition 7.5.6, (7.18) is proved if (7.20)

γ (Γ(L, C)/X) = 1

for any k-open simplex X ∈ Γ(L, C). We first observe that    (7.21) Γ(L, C)/X = I ∈
∗ (C \ X)  (I ∪ X) > ˆ0 or (I ∪ X) < ˆ1 .  Now, let cX be -minimal | X < c < X}. Then, cX ∈ / X since  in {c ∈ C otherwise c  cX and X < c < X for some c ∈ C, contradicting the minimality of cX . Now, by (7.21) and since X < cX < X and cX ∈ / X, it

98

7

Combinatorial Applications and Related Topics

follows that Y → Y
{cX } is a sign-reversing involution on Γ(L, C)/X ∪ {∅}, whence (7.20) and thus (7.18) is shown. To establish (7.19) we first note that  (7.22) γ˜ ({I ∈ Γ(L, C)|k(I) = ∅}) = γ˜({I ∈
∗ (C)|k(I) = ∅}) + (−1)|I| . ∗
Ï(C) ÎI∈ I=ˆ 0, I=ˆ 1 k(I)=∅

The assumptions of the theorem imply that C is not k-open, whence =
∗ (C) satisfies the requirements of Theorem 7.5.2. Thus, by applying Theorem 7.5.2, γ˜({I ∈
∗ (C) | k(I) = ∅}) = γ˜(
∗ (C)) = 0 .

(7.23)

In view of (7.22) and (7.23) the equivalence of (7.18) and (7.19) is obvious. 2 Corollary 7.5.9 Let L = [ˆ 0, ˆ 1] be a non-trivial finite lattice and C be a crosscut of L which is given a partial order denoted by  to distinguish it from the partial order ≤ in L. Let  consist of all non-empty subsets X of C  such that for any x ∈ X there is some c ∈ C satisfying c  x and X < c < X. Then, (7.24)

µ(L) = γ˜ ({I ∈ Γ(L, C) | I ⊇ X for any X ∈ })

or equivalently, (7.25)

µ(L) =

 Î
Ï

(−1)|I| .

I∈ ∗ (C) I=ˆ 0, I=ˆ 1 I⊇X(∀X∈ )



Proof. Since  turns out as union-closed, the corollary follows from Theorem 7.5.8 and the correspondence between kernel operators and union-closed sets. 2

Example 7.5.10 Let L and C be as in Example 7.5.7. By putting d  b, d  e, b  a and b  c a partial ordering relation  on C is defined, whose Hasse diagram is shown in Figure 7.3. In connection with this partial ordering relation,  = {{a, c}, {b, e}, {c, e}} satisfies the requirements of Corollary 7.5.9, and thus µ(L) can be expressed via (7.24) as the reduced Euler characteristic of the subcomplex {{a}, {b}, {c}, {d}, {e}, {a, b}, {a, d}, {b, c}, {b, d}, {c, d}, {d, e}, {a, b, d}, {b, c, d}} of Γ(L, C), which contains only 13 of the total 23 simplices of Γ(L, C). Moreover, for the present choice of  it turns out that the sum on the right-hand side of (7.25) contains only two of the eight terms that appear in the sum of (7.17).

Remarks. In view of Theorem 7.5.8 the corollary can be stated more generally by requiring  to be a union-closed set of non-empty subsets of the above type.

7.5 Euler characteristics and M¨obius functions a

99

c

b

e

d

Figure 7.3: Hasse diagram of {a, b, c, d, e} with respect to .

By putting C = A(L) Corollary 7.5.9 specializes to a result of Blass and Sagan [BS97], which they used in computing and combinatorially explaining the M¨obius function of various lattices and in generalizing Stanley’s well-known theorem [Sta72] that the characteristic polynomial of a semimodular supersolvable lattice factors over the integers. As pointed out by Blass and Sagan [BS97], their result generalizes a particular case of Rota’s broken circuit theorem [Rot64] as well as a prior generalization of that particular case due to Sagan [Sag95]. For a restatement of Rota’s result [Rot64], the following definitions are necessary: ˆ 1] ˆ be a finite lattice. A rank function of L is a Definition 7.5.11 Let L = [0, function r : L → ∪ {0} such that r(ˆ 0) = 0 and r(a) = r(b) + 1 whenever a is an immediate successor of b. It is well known and straightforward to prove that the rank function is unique if it exists. A geometric lattice is a finite lattice L whose rank function exists and satisfies r(a ∧ b) + r(a ∨ b) ≤ r(a) + r(b) for any a, b ∈ L. A subset I of A(L) is independent if r( I) = |I|, and dependent otherwise. Each minimal dependent subset of A(L) is called a circuit of L. Given a linear ordering relation on A(L), each set C \ {max C}, where C is a circuit of L, is referred to as a broken circuit of L. Note that this depends on the ordering of the atoms. Corollary 7.5.12 [Rot64] Let L = [ˆ 0, ˆ 1] be a non-trivial geometric lattice with rank function r. Furthermore, let the set of atoms A(L) be endowed with a linear ordering relation, and let consist of all broken circuits of L. Then,    ˆ µ(L) = (−1)r(1) ×  I ⊆ A(L)  I = ˆ 1 and I ⊇ B for any B ∈  . Proof. [BS97] (Sketch) Let C = A(L) and  as in Corollary 7.5.9. Then,  ⊇ and any X ∈  includes some B ∈ . Thus, for any I ⊆ A(L), I ⊇ X for any X ∈  if and only if I ⊇ B for any B ∈ . In this case, I is independent and hence r( I) = |I|. Thus, Corollary 7.5.12 follows from Corollary 7.5.9. 2

Bibliography [Aig79]

M. Aigner, Combinatorial Theory, Springer-Verlag, New York, 1979.

[AJ99]

T. Aven and U. Jensen, Stochastic Models in Reliability, SpringerVerlag, New York, 1999.

[Ale37]

P.S. Alexandroff, Diskrete R¨ aume, Mat. Sbornik (N.S.) 2 (1937), 501–518.

[Bal86]

M.O. Ball, Computational complexity of network reliability analysis: an overview, IEEE Trans. Reliab. 35 (1986), 230–239.

[BC84]

J.A. Buzacott and J.S.K. Chang, Cut-set intersections and node partitions, IEEE Trans. Reliab. 33 (1984), 385–389.

[BES61]

Z.W. Birnbaum, J.D. Esary, and S. Saunders, Multi-component systems and structures and their reliability, Technometrics 3 (1961), 55–77.

[BH84]

R.E. Barlow and K.D. Heidtmann, Computing k-out-of-n system reliability, IEEE Trans. Reliab. 33 (1984), 322–323.

[Bir12]

G.D. Birkhoff, A determinant formula for the number of ways of coloring a map, Ann. Math. 14 (1912), 42–46.

[BO81]

T. Brylawski and J. Oxley, The broken-circuit complex: its structure and factorizations, Europ. J. Comb. 2 (1981), 107–121.

[BO92]

T. Brylawski and J. Oxley, The Tutte polynomial and its applications, Matroid Applications (N. White, ed.), Encyclopedia of Mathematics and Its Applications, vol. 40, Cambridge University Press, 1992, pp. 123–225.

[Bon36]

C.E. Bonferroni, Teoria statistica delle classi e calcolo delle probabilit´ a, Pubblicazioni del R. Istituto Superiore di Scienze Economiche e Commerciali di Firenze 8 (1936), 1–62.

[Boo54]

G. Boole, An investigation of the laws of thought on which are founded the mathematical theories of logic and probabilities, Dover 1st printing, 1854.

K. Dohmen: LNM 1826, pp. 100–109, 2003. c Springer-Verlag Berlin Heidelberg 2003


Bibliography

101

[Bor48]

K. Borsuk, On the embedding of systems of compacta in simplicial complexes, Fund. Math. 35 (1948), 217–234.

[BP87]

M.O. Ball and J.S. Provan, Computing K-terminal reliability in time polynomial in the number of (s, K)-quasicuts, Fourth Army Conference on Applied Mathematics and Computing (Washington), Army Research Office, 1987, pp. 901–907.

[BP01]

J. Buksz´ar and A. Pr´ekopa, Probability bounds with cherry trees, Math. Oper. Res. 26 (2001), 174–192.

[BPS91]

M.O. Ball, J.S. Provan, and D.R. Shier, Reliability covering problems, Networks 21 (1991), 345–357.

[Bry77]

T. Brylawski, The broken-circuit complex, Trans. Amer. Math. Soc. 234 (1977), 417–433.

[BS97]

A. Blass and B.E. Sagan, M¨ obius functions of lattices, Adv. Math. 127 (1997), 94–123.

[Buk01]

J. Buksz´ ar, Upper bounds for the probability of a union by multitrees, Adv. Appl. Probab. 33 (2001), 437–452.

[Buz87]

J.A. Buzacott, Node partition formula for directed graph reliability, Networks 17 (1987), 227–240.

[BW85]

E.A. Bender and H.S. Wilf, A theoretical analysis of backtracking in the graph coloring problem, J. Algorithms 6 (1985), 275–282.

[Bye98]

O. Byer, Some new bounds for the maximum number of vertex colorings of a (v, e)-graph, J. Graph Theory 28 (1998), 115–128.

[BZ91]

A. Bj¨orner and G.M. Ziegler, Broken circuit complexes: factorizations and generalizations, J. Comb. Theory Ser. B (1991), 96–126.

[BZ92]

A. Bj¨orner and G.M. Ziegler, Introduction to greedoids, Matroid Applications (N. White, ed.), Encyclopedia of Mathematics and Its Applications, vol. 40, Cambridge University Press, 1992, pp. 284–357.

[CN81]

D.T. Chiang and S.C. Niu, Reliability of consecutive k-out-of-n:F system, IEEE Trans. Reliab. 30 (1981), 87–89.

[Col87]

C.J. Colbourn, The Combinatorics of Network Reliability, Oxford University Press, New York, 1987.

[Cra67]

H.H. Crapo, A higher invariant for matroids, J. Combin. Theory 2 (1967), 406–417.

[Doh98]

K. Dohmen, Inclusion-exclusion and network reliability, Electron. J. Combin. 5 (1998), research paper no. 36, Internet: http://www.combinatorics.org.

102

Bibliography

[Doh99a] K. Dohmen, An improvement of the inclusion-exclusion principle, Arch. Math. 72 (1999), 298–303. [Doh99b] K. Dohmen, Convex geometric inclusion-exclusion identities and Bonferroni inequalities with applications to system reliability analysis and reliability covering problems, Informatik-Bericht Nr. 132, Humboldt-Universit¨ at zu Berlin, 1999. [Doh99c] K. Dohmen, Improved Bonferroni inequalities for the reliability of a communication network, in: Safety and Reliability, Proceedings of the 10th European Conference on Safety and Reliability (ESREL), Munich-Garching, Germany, September 13–17, 1999 (G.I. Schu¨eller and P. Kafka, eds.), vol. I, Balkema Publishers, Rotterdam/Brookfield, 1999, pp. 67–71. [Doh99d] K. Dohmen, Improved inclusion-exclusion identities and inequalities based on a particular class of abstract tubes, Electron. J. Probab. 4 (1999), paper no. 5, Internet: http://www.math.washington.edu/~ejpecp. [Doh99e] K. Dohmen, On partial sums of chromatic polynomials, Ars. Combin. 52 (1999), 125–127. [Doh99f]

K. Dohmen, On sums over partially ordered sets, Electron. J. Combin. 6 (1999), research paper no. 34, Internet: http://www.combinatorics.org.

[Doh00a] K. Dohmen, Improved Bonferroni inequalities via union-closed set systems, J. Combin. Theory Ser. A 92 (2000), 61–67. [Doh00b] K. Dohmen, Improved inclusion-exclusion identities via closure operators, Discrete Math. Theoret. Comput. Sci. 4 (2000), 61–66. [Doh00c] K. Dohmen, On closure operators and the sieve formula, Util. Math. 58 (2000), 203–207. [Doh00d] K. Dohmen, Some remarks on the sieve formula, the Tutte polynomial and Crapo’s beta invariant, Aequ. Math. 60 (2000), 108–115. [Doh01a] K. Dohmen, A note on inclusion-exclusion on semilattices, Util. Math. 59 (2001), 125–127. [Doh01b] K. Dohmen, A note on Narushima’s principle of inclusion-exclusion on partition lattices, Graphs Comb. 17 (2001), 607–610. [Doh01c] K. Dohmen, Improved Inclusion-Exclusion Identities and Bonferroni Inequalities with Reliability Applications, Habilitationsschrift, Humboldt-Universit¨ at zu Berlin, 2001. [Doh02a] K. Dohmen, A note on Zeilberger’s abstract lace expansion, Adv. Appl. Math. 28 (2002), 272–277.

Bibliography

103

[Doh02b] K. Dohmen, Bonferroni-type inequalities via chordal graphs, Combin. Probab. Comput. 11 (2002), 349–351. [Doh02c] K. Dohmen, Kernel operators and improved inclusion-exclusion bounds, Australas. J. Combin. 26 (2002), 219–224. [Doh03]

K. Dohmen, Improved inclusion-exclusion identities and Bonferroni inequalities with reliability applications, SIAM J. Discrete Math. 16 (2003), 156–171.

[DPT03]

K. Dohmen, A. P¨ onitz, and P. Tittmann, A new two-variable generalization of the chromatic polynomial, DMTCS, Discrete Math. Theoret. Comput. Sci. 6 (2003), 69–90.

[DS67]

D.A. Dawson and D. Sankoff, An inequality for probabilities, Proc. Amer. Math. Soc. 18 (1967), 504–507.

[Ede95]

H. Edelsbrunner, The union of balls and its dual shape, Discrete Comput. Geom. 13 (1995), 415–440.

[EJ85]

P.H. Edelman and R. Jamison, The theory of convex geometries, Geom. Dedicata 19 (1985), 247–270.

[EP63]

J.D. Esary and F. Proschan, Coherent structures of non-identical components, Technometrics 5 (1963), 191–209.

[ER97]

H. Edelsbrunner and E.A. Ramos, Inclusion-exclusion complexes for pseudodisk collections, Discrete Comput. Geom. 17 (1997), 287–306.

[ER00]

P.H. Edelman and V. Reiner, Counting the interior points of a point configuration, Discrete Comput. Geom. 23 (2000), 1–13.

[Ern82]

M. Ern´e, Einf¨ uhrung in die Ordnungstheorie, Bibliographisches Institut, Mannheim, 1982.

[FJ86]

M. Farber and R.E. Jamison, Convexity in graphs and hypergraphs, SIAM J. Alg. Disc. Meth. 7 (1986), 433–444.

[Gal66]

J. Galambos, On the sieve methods in probability theory I., Stud. Sci. Math. Hungar. 1 (1966), 39–50.

[Gal72]

J. Galambos, The role of graph theory in some sieve arguments of probability theory, in: Graph Theory Appl., Proc. Conf. Western Michigan Univ. 1972, Lecture Notes Math., vol. 303, 1972, pp. 81– 87.

[Gal75]

J. Galambos, Methods for proving Bonferroni type inequalities, J. London Math. Soc. 2 (1975), 561–564.

[Gal77]

J. Galambos, Bonferroni inequalities, Ann. Probab. 5 (1977), 577– 581.

104

Bibliography

[Gal94]

J. Galambos, Bonferroni-type inequalities in statistics: a survey, J. Appl. Statist. Science 1 (1994), 195–209.

[GJ79]

M.R. Garey and D.S. Johnson, Computers and Intractability: A Guide to the Theory of NP-Completeness, W.H. Freeman, New York, 1979.

[GM81]

A. Garsia and S. Milne, A Roger-Ramanujan bijection, J. Comb. Theory Ser. A 31 (1981), 289–339.

[GNW00] B. Giglio, D.Q. Naiman, and H.P. Wynn, Abstract tube theory and Gr¨ obner bases for improved inclusion-exclusion reliability bounds, Second International Conference on Mathematical Methods in Reliability — MMR 2000, Bordeaux, July 4–7, 2000, vol. 1, 2000, pp. 451– 454. [GNW02] B. Giglio, D.Q. Naiman, and H.P. Wynn, Gr¨ obner bases, abstract tubes, and inclusion-exclusion reliability bounds, IEEE Trans. Reliab. 58 (2002), 358–366. [Gor97]

G. Gordon, A β invariant for greedoids and antimatroids, Electron. J. Combin. 4 (1997), research paper no. 13, Internet: http://www.combinatorics.org.

[Gra93]

D.A. Grable, Sharpened Bonferroni inequalities, J. Comb. Theory Ser. B 57 (1993), 131–137.

[Gra94]

D.A. Grable, Hypergraphs and sharpened sieve inequalities, Discrete Math. 132 (1994), 75–82.

[Gra98]

G. Graetzer, General Lattice Theory, 2nd ed., Birkh¨ auser, Basel, 1998.

[GS96a]

J. Galambos and I. Simonelli, Bonferroni-type inequalities—recent advances, Applied Statistical Science, I, Proceedings of the 1st conference held in Lawrenceville, NJ, May 20–21, 1995 (M. Ahsanullah, ed.), Nova Science Publishers, Commack, NY, 1996, pp. 127–135.

[GS96b]

J. Galambos and I. Simonelli, Bonferroni-type Inequalities with Applications, Springer-Verlag, New York, 1996.

[GX90]

J. Galambos and Y. Xu, A new method for generating Bonferronitype inequalities by iteration, Math. Proc. Camb. Phil. Soc. 107 (1990), 601–607.

[Har78]

E. Harzheim, Einf¨ uhrung in die Kombinatorische Topologie, Wissenschaftliche Buchgesellschaft, Darmstadt, 1978.

[Her72]

A.P. Heron, Matroid polynomials, Combinatorics (D.J.A. Welsh and D.R. Woodall, eds.), vol. X, The Institute of Combinatorics and Its Applications, Southend-on-Sea, 1972, pp. 164–202.

Bibliography

105

[How81]

E. Howorka, A characterization of Ptolemaic graphs, J. Graph Theory 5 (1981), 323–331.

[Hun76]

D. Hunter, An upper bound for the probability of a union, J. Appl. Prob. 13 (1976), 597–603.

[Hwa82]

C.L. Hwang, Fast solutions for consecutive-k-out-of-n:F system, IEEE Trans. Reliab. 31 (1982), 447–448.

[Jor27]

Ch. Jordan, The foundations of the theory of probability, Mat. Phys. Lapok 34 (1927), 109–136.

[JW81]

R.E. Jamison-Waldner, Partition numbers for trees and ordered sets, Pacific J. Math. 96 (1981), 115–140.

[Kon80]

J.M. Kontoleon, Reliability determination of a r-successive-out-ofn:F system, IEEE Trans. Reliab. 29 (1980), 437.

[Kou68]

E.G. Kounias, Bounds for the probability of a union of events, with applications, Ann. Math. Statist. 39 (1968), 2154–2158.

[KP89]

A. Kossow and W. Preuss, Reliability of consecutive-k-out-of-n:F systems with nonidentical component reliabilities, IEEE Trans. Reliab. 38 (1989), 229–233.

[Kru56]

J.B. Kruskal, On the shortest spanning subtree of a graph and the traveling salesman problem, Proc. Amer. Math. Soc. 7 (1956), 48– 50.

[Kwe75a] S.M. Kwerel, Bounds on the probability of the union and intersection of m events, Adv. Appl. Probab. 7 (1975), 431–448. [Kwe75b] S.M. Kwerel, Most stringent bounds on aggregated probabilities of partially specified dependent probability systems, J. Amer. Statist. Assoc. 70 (1975), 472–479. [Laz90]

F. Lazebnik, New upper bounds for the greatest number of proper colorings of a (v, e)-graph, J. Graph Theory 14 (1990), 25–29.

[Loz92]

E.L. Lozinskii, Counting propositional models, Inform. Process. Lett. 41 (1992), 326–332.

[Man90]

Eckhard Manthei, Domination theory and network reliability analysis, J. Inf. Process. Cybern. 27 (1990), 129–139.

[Man91]

Eckhard Manthei, Kombinatorische Methoden der Zuverl¨ assigkeitsanalyse monotoner bin¨ arer Systeme, Habilitationsschrift, Bergakademie Freiberg, 1991.

[M˘ ar89]

E. M˘ arg˘aritescu, Improved Bonferroni-Galambos-type inequalities, An. Univ. Bucur., Mat. 38 (1989), 31–41.

106

Bibliography

[McK97]

T.A. McKee, Graph structure for inclusion-exclusion inequalities, Congr. Numer. 125 (1997), 5–10.

[McK98]

T.A. McKee, Graph structure for inclusion-exclusion equalities, Congr. Numer. 133 (1998), 121–126.

[MS85]

T.F. M´ ori and G.J. Sz´ekely, A note on the background of several Bonferroni-Galambos-type inequality, J. Appl. Probab. 22 (1985), 836–843.

[Nar74]

H. Narushima, Principle of inclusion-exclusion on semilattices, J. Combin. Theory Ser. A 17 (1974), 196–203.

[Nar77]

H. Narushima, Principle of Inclusion-Exclusion on Semilattices and Its Applications, Ph.D. thesis, Waseda Univ., 1977.

[Nar82]

H. Narushima, Principle of inclusion-exclusion on partially ordered sets, Discrete Math. 42 (1982), 243–250.

[NW92]

D.Q. Naiman and H.P. Wynn, Inclusion-exclusion-Bonferroni identities and inequalities for discrete tube-like problems via Euler characteristics, Ann. Statist. 20 (1992), 43–76.

[NW97]

D.Q. Naiman and H.P. Wynn, Abstract tubes, improved inclusionexclusion identities and inequalities and importance sampling, Ann. Statist. 25 (1997), 1954–1983.

[NW01]

D.Q. Naiman and H.P. Wynn, Improved inclusion-exclusion inequalities for simplex and orthant arrangements, JIPAM, J. Inequal. Pure Appl. Math. 2 (2001), Article No. 18, Internet: http://jipam.vu.edu.au.

[Pau86]

¨ P. Paule, Uber das Involutionsprinzip von Garsia und Milne, Bayreuther Math. Schr. 21 (1986), 295–319.

[PB84]

J.S. Provan and M.O. Ball, Computing network reliability in time polynomial in the number of cuts, Oper. Res. 32 (1984), 516–526.

[PK93]

S.G. Papastavridis and M.V. Koutras, Consecutive k-out-of-n systems, New Trends in System Reliability Evaluation (K.B. Misra, ed.), Elsevier Science Publishers, 1993, pp. 228–248.

[Pr´e88]

A. Pr´ekopa, Boole-Bonferroni inequalities and linear programming, Oper. Res. 36 (1988), 145–162.

[Pr´e90]

A. Pr´ekopa, Sharp bounds on probabilities using linear programming, Oper. Res. 38 (1990), 227–239.

[Pr´e95]

A. Pr´ekopa, Stochastic Programming, Kluwer Academic Publishers, Dordrecht, 1995.

Bibliography

107

[Pri57]

R.C. Prim, Shortest connection networks and some generalizations, Bell. System Tech. J. 36 (1957), 1389–1401.

[PS96]

J.S. Provan and D.R. Shier, A paradigm for listing (s, t)-cuts in graphs, Algorithmica 15 (1996), 351–372.

[R´en61]

A. R´enyi, Egy ´ltal´ anos m´ odszer val´ osz´in¨ us´egsz´ am´it´ asi t´etelek bi´ zonyit´ as´ ara ´es annak n´eh´ any alkalmaz´ asa, Magyar Tud. Akad., Mat. Fiz. Tud. Oszt. Koezl. 11 (1961), 79–105.

[Rot64]

G.-C. Rota, On the foundations of combinatorial theory I. Theory of M¨ obius functions, Z. Wahrscheinlichkeitstheorie Verw. Gebiete 2 (1964), 340–368.

[Rot88]

J.J. Rotman, An Introduction to Algebraic Topology, SpringerVerlag, New York, 1988.

[RT75]

R.C. Read and R.E. Tarjan, Bounds on backtrack algorithms for listing cycles, paths and spanning trees, Networks 5 (1975), 237–252.

[Rus93]

A.M. Rushdi, Reliability of k-out-of-n systems, New Trends in System Reliability Evaluation (K.B. Misra, ed.), Elsevier Science Publishers, 1993, pp. 185–227.

[Sag95]

B.E. Sagan, A generalization of Rota’s NBC Theorem, Adv. Math. 111 (1995), 195–207.

[Sat82]

A. Satyanarayana, A unified formula for analysis of some network reliability problems, IEEE Trans. Reliab. 31 (1982), 23–32.

[Sen88]

E. Seneta, Degree, iteration and permutation in improving Bonferroni-type bounds, Austral. J. Statist. 30A (1988), 27–38.

[SH81]

A. Satyanarayana and J.N. Hagstrom, A new algorithm for the reliability analysis of multi-terminal networks, IEEE Trans. Reliab. 30 (1981), 325–334.

[Sha82]

J.G. Shanthikumar, Recursive algorithm to evaluate the reliability of a consecutive-k-out-of-n:F system, IEEE Trans. Reliab. 31 (1982), 442–443.

[Shi88]

D.R. Shier, Algebraic aspects of computing network reliability, Applications of Discrete Mathematics (R.D. Ringeisen and F.S. Roberts, eds.), SIAM, Philadelphia, 1988, pp. 135–147.

[Shi91]

D.R. Shier, Network Reliability and Algebraic Structures, Clarendon Press, Oxford, 1991.

[SM91]

D.R. Shier and M.H. McIlwain, Inclusion-exclusion, cancellation, and consecutive sets, Australas. J. Combin. 4 (1991), 121–142.

108

Bibliography

[SP78]

A. Satyanarayana and A. Prabhakar, New topological formula and rapid algorithm for reliability analysis of complex networks, IEEE Trans. Reliab. 27 (1978), 82–100.

[Spa66]

E. Spanier, Algebraic topology, McGraw Hill, New York, 1966.

[Sta72]

R.P. Stanley, Supersolvable lattices, Alg. Univ. 2 (1972), 197–217.

[Sta99]

R.P. Stanley, Enumerative Combinatorics, vol. 2, Cambridge University Press, Cambridge, 1999.

[Tak67]

L. Tak´ acs, On the method of inclusion and exclusion, J. Amer. Stat. Assoc. 62 (1967), 102–113.

[Tom86]

I. Tomescu, Hypertrees and Bonferroni inequalities, J. Combin. Theory Ser. B 41 (1986), 209–217.

[TSOA80] S. Tsukiyama, I. Shirakawa, H. Ozaki, and H. Ariyoshi, An algorithm to enumerate all cutsets of a graph in linear time per cutset, Journal of the ACM 27 (1980), 619–632. [Tut47]

W.T. Tutte, A ring in graph theory, Proc. Cambridge Phil. Soc. 43 (1947), 26–40.

[Tut54]

W.T. Tutte, A contribution to the theory of chromatic polynomials, Canadian J. Math. 6 (1954), 80–91.

[TX89]

X. Tan and Y. Xu, Some inequalities of Bonferroni-Galambos type, Statist. Probab. Lett. 8 (1989), 17–20.

[Val79]

L. Valiant, The complexity of enumeration and reliability problems, SIAM J. Comput. 8 (1979), 410–421.

[Vog99]

M. Vogel, Experimentelle Untersuchungen zu einem InklusionsExklusions-Algorithmus zur Bestimmung der zwei-terminalen Zuverl¨ assigkeit von Netzwerken, Studienarbeit (student’s work), Humboldt-Universit¨ at zu Berlin, 1999.

[War42]

M. Ward, The closure operators of a lattice, Ann. Math. II. Ser. 43 (1942), 191–196.

[Wei35]

L. Weisner, Abstract theory of inversion of finite series, Trans. Amer. Math. Soc. 38 (1935), 474–484.

[Wel76]

D.J.A. Welsh, Matroid Theory, Academic Press, New York, 1976.

[Wel93]

D.J.A. Welsh, Complexity: Knots, Colourings and Counting, LMS Lecture Note Series, vol. 186, Cambridge University Press, 1993.

[Whi32]

H. Whitney, A logical expansion in mathematics, Bull. Amer. Math. Soc. 38 (1932), 572–579.

Bibliography

109

[Wil76]

H.S. Wilf, Which polynomials are chromatic?, Proc. Colloquio Internazionale sulle Teorie Combinatorie (Roma,1973), Tomo I, Atti dei Convegni Lincei, No. 17, Accademia Naz. Lincei (Rome), 1976, pp. 247–256.

[Wil84]

H.S. Wilf, Backtrack: An O(1) average time algorithm for the graph coloring problem, Inf. Process. Lett. 18 (1984), 119–122.

[Wil86]

H.S. Wilf, Algorithms and Complexity, Prentice-Hall, Englewood Cliffs, New Jersey, 1986.

[Wor82]

K.J. Worsley, An improved Bonferroni inequality and applications, Biometrika 69 (1982), 297–302.

[Zei84]

D. Zeilberger, Garsia and Milne’s bijective proof of the inclusionexclusion principle, Discrete Math. 51 (1984), 109–110.

[Zei97]

D. Zeilberger, The abstract lace expansion, Adv. Appl. Math. 19 (1997), 355–359.

Author Index Alexandroff, 22 Ball, 44, 49, 56, 75, 77, 78 Barlow, 68 Bernoulli, 2 Birkhoff, 83, 84 Bj¨orner, 21 Blass, 99 Bonferroni, 2 Boole, 2 Borsuk, 15 Buksz´ar, 37 Buzacott, 56, 57 Colbourn, 47, 55 Crapo, 94 Da Silva, 2 Dawson, 64, 65, 79, 81 Edelman, 20, 21, 31, 32 Edelsbrunner, 14 Galambos, 2, 37, 38, 41, 42 Garsia, 33 Giglio, 81 Gordon, 32 Heidtmann, 68 Heron, 93 Howorka, 23 Hunter, 37, 39 Hwang, 73 Jamison, 20, 22, 31 Jordan, 2 Kossow, 71, 74 Kounias, 42 Kwerel, 42, 64, 65, 79, 81 Lawrence, 32 Lozinskii, 3 Manthei, 52, 53

McIlwain, 74 McKee, 37 Milne, 33 Montmort, 1 Naiman, 4, 9–18, 23, 81 Narushima, 4, 22, 25, 31, 82, 83 Poincar´e, 1 Pr´ekopa, 37 Prabhakar, 52 Preuss, 71, 74 Provan, 44, 49, 54, 56, 75–78 Pr´ekopa, 64 Ramos, 14 Read, 55 Reiner, 21, 32 Rota, 90, 96, 99 Sagan, 99 Sankoff, 64, 65, 79, 81 Satyanarayana, 52 Seneta, 41 Shanthikumar, 73 Shier, 4, 44, 47, 49–51, 54–56, 71, 74–78 Simonelli, 2, 37, 38, 41 Stanley, 86, 99 Sylvester, 1 Tak´ acs, 2 Tarjan, 55 Tittmann, 70 Tomescu, 37 Weisner, 90 Whitney, 4, 83, 84, 86 Wilf, 83, 84 Worsley, 37, 39 Wynn, 4, 9–18, 23, 81 Xu, 42 Zeilberger, 4, 28 Ziegler, 21

Subject Index abstract lace expansion, 28 abstract simplicial complex, 7 abstract tube, 9–18 all-terminal reliability, 54

importance sampling scheme, 13 inclusion-exclusion principle, 1

basis, 20 beta invariant, 94 betti numbers, 11 binomial moments, 64 broken circuit, 84, 93, 99 broken circuit complex, 84

lace map, 28

characteristic polynomial, 93 chordal graph sieve, 38 chromatic polynomial, 84 clique complex, 23 clique number, 5 closed, 20, 89 closure operator, 20, 89 coherent binary system, 47 contractible, 8 convex geometry, 20 convex polyhedron, 14 crosscut complex, 96

nerve, 10 network reliability, 53–66

Delauney complex, 16, 18 depth of an abstract tube, 13 domination, 31, 52 dual antimatroid, 20 dynamic programming, 45, 50 Euler characteristic, 7, 96 Euler-Poincar´e formula, 11 extended set of mincuts, 51 extended set of minpaths, 51 free set, 20

kernel operator, 25, 89

M¨ obius function, 52 matroid, 91 method of indicators, 1 mincut, 48 minpath, 48

open, 25, 89 order complex, 22 poset terminology, 6 pseudopolynomial algorithm, 51 quasi-cut, 57 recursive scheme, 44 reliability covering problems, 75 semilattice sieve, 22 source-to-T -terminal reliability, 54 source-to-terminal reliability, 54 subtube of an abstract tube, 9 system reliability, 48 Tutte polynomial, 91 two-terminal reliability, 54 unions of balls, 15 unions of spherical caps, 17 Voronoi subdivision, 16, 18

generalized chromatic polynomial, 86 graph terminology, 5 Helly number, 22 homology group, 11

weak abstract tube, 10