Structure functions and minimal path sets

Jean-Luc Marichal
Revised version, September 30, 2015

Abstract be written in the multilinear form

arXiv:1401.0803v2 [stat.AP] 21 Jun 2016

(x) = (x1 , . . . , xn )
In this short note we give and discuss a general multilinear X Y Y
expression of the structure function of an arbitrary semi- = (A) xi (1 xi ). (1)
coherent system in terms of its minimal path and cut sets. AC iA iC\A
We also examine the link between the number of minimal
path and cut sets consisting of 1 or 2 components and the Since the coefficients in this form are exactly the struc-
concept of structure signature of the system. ture function values, we will refer to this form as the self-
descriptive form of the structure function. By expanding
Keywords: System reliability, semicoherent system, the second product in Eq. (1) and then collecting terms, we
structure function, reliability function, minimal path and obtain the simple form of the structure function, namely
cut sets. X Y
(x) = d(A) xi , (2)
AC iA

where the link between the new coefficients d(A) and the
Notation values (A), which can be obtained from the Mobius in-
version theorem, is given through the following linear con-
[n] set {1, . . . , n} version formulas (see, e.g., [15, p. 31])
C set of components of the system
X
(x) structure function of the system (A) = d(B)
D (x) dual structure function BA
h(p) reliability function of the system
s signature of the system and X
sk k-th coordinate of s d(A) = (1)|A||B| (B) .
BA

Recall that a path set of the system is a component

subset P C such that (P ) = 1. A path set P of the
1 Introduction system is said to be minimal if (P ) = 0 for every P P .
It is well known [3, Ch. 2] that if P1 , . . . , Pr denote the
Consider an n-component system (C, ), where C is the set minimal path sets of the system, then
[n] = {1, . . . , n} of its components and : {0, 1}n {0, 1}
is its structure function which expresses the state of the a Y Y  Y 
(x) = xi = 1 1 xi , (3)
system in terms of the states of its components. We assume
j[r] iPj j[r] iPj
that the system is semicoherent, which means that the
structure function is nondecreasing in each variable and where [r] = {1, . . . , r} and is the coproduct operation
satisfies the conditions (0, . . . , 0) = 0 and (1, . . . , 1) = 1. (i.e., the dual of the product operation) defined by i xi =
Throughout we identify Boolean n-vectors x {0, 1}n 1 i (1 xi ).
and subsets A [n] in the usual way, that is, by set-
Example 1. Consider the bridge structure as indicated in
ting xi = 1 if and only if i A. This identification en-
Figure 1. This structure is characterized by four minimal
ables us to use the same symbol to denote both a func-
n path sets, namely P1 = {1, 4}, P2 = {2, 5}, P3 = {1, 3, 5},
tion f : {0, 1} R and the corresponding set function
[n] and P4 = {2, 3, 4}. Equation (3) then shows that the struc-
f: 2 R interchangeably. For instance, we write
ture function is given by
(0, . . . , 0) = () and (1, . . . , 1) = (C).
As a Boolean function, the structure function can always (x1 , . . . , x5 ) = x1 x4 x2 x5 x1 x3 x5 x2 x3 x4 . (4)

Jean-Luc Marichal is with the Mathematics Research Unit, The simple form of the structure function can be easily
FSTC, University of Luxembourg, 6, rue Coudenhove-Kalergi, L- computed by expanding the coproducts in (4) and simpli-
1359 Luxembourg, Luxembourg. Email: jean-luc.marichal[at]uni.lu fying the resulting algebraic expression using x2i = xi . We

1

then obtain
(x1 , . . . , x5 )
= 1 (1 x1 x4 )(1 x2 x5 )(1 x1 x3 x5 )(1 x2 x3 x4 )
= x1 x4 + x2 x5 + x1 x3 x5 + x2 x3 x4 x1 x2 x3 x4
x1 x2 x3 x5 x1 x2 x4 x5 x1 x3 x4 x5 x2 x3 x4 x5
+ 2 x1 x2 x3 x4 x5 ,
Q
iA xi in the simple form of D . By using the dual struc-
ture function we can easily derive various useful forms of
the structure function and its multilinear extension (see
also Grabisch et al. [6]). Table 1 summarizes the best
known of these forms (in addition to the minimal path set
representation given in Eq. (3)).

which reveals the coefficients d(A) of the simple form of 1. Self-descriptive form
the structure function. X Y Y
(A) xi (1 xi )
AC iA iC\A


1 4 2. Dual self-descriptive form

r
3
r X Y Y
1 D (A) xi (1 xi )
2 5
AC iC\A iC


3. Simple form
Figure 1: Bridge structure
X Y
d(A) xi
Example 1 illustrates the important fact that the sim- AC iA
ple form (2) of any structure function can be expressed in
terms of the minimal path sets of the system simply by 4. Dual simple form
expanding the coproduct in (3) and then simplifying the
X a
resulting polynomial expression (using x2i = xi ) until it dD (A) xi
becomes multilinear. It seems, however, that such a gen- AC iA
eral expression for the structure function is unknown in
the literature. 5. Disjunctive normal form
In Section 2 of this note we yield an expression of the a Y
simple form of the structure function in terms of the mini- xi
mal path sets. The derivation of this expression is inspired AC iA
from the exact computation of the reliability function of (A)=1

the system by means of the inclusion-exclusion principle.

We also provide the dual version of this expression in terms 6. Conjunctive normal form
of the minimal cut sets and discuss some interesting con- Y a
sequences of these expressions. In Section 3 we show that xi
the number of minimal path and cut sets consisting of 1 AC iA
D (A)=1
or 2 components can be computed easily from the concept
of structure signature of the system.

Table 1: Various forms of the structure function (x) and

2 Structure functions and minimal b
its multilinear extension (x)
path and cut sets
The concept of multilinear extension of the structure
By extending formally the structure function to the hyper- function has the following important interpretation in reli-
cube [0, 1]n by linear interpolation, we define its multilinear ability theory. When the components are statistically inde-
extension (a concept introduced in game theory by Owen pendent, the function b is nothing other than the reliability
[14]) as the multilinear polynomial function b : [0, 1]n function h : [0, 1]n [0, 1], which gives the reliability
[0, 1] defined by X Y Y
X Y Y h(p) = h(p1 , . . . , pn ) = (A) pi (1 pi )
b
(x) = (A) xi (1 xi ). AC iA iC\A
AC iA iC\A
of the system in terms of the reliabilities p1 , . . . , pn of the
Let D : {0, 1}n {0, 1} be the dual structure func- components (see, e.g., [3, Ch. 2]).
tion defined as D (x) = 1 (1 x), where 1 stands for The exact computation of the system reliability h(p) in
the n-vector (1, . . . , 1), and let dD (A) be the coefficient of terms of the minimal path sets P1 , . . . , Pr is usually done

2
by means of the inclusion-exclusion method (see, e.g., [1, Clearly, Eq. (8) still holds on [0, 1]n if we replace the
Sect. 6.2] and [3, Ch. 2]). In this section we recall this structure function with its multilinear extension. In partic-
method and show how we can adapt it to derive a concise ular, when the components are statistically independent,
expression of the simple form of the structure function in we see that (7) immediately follows from (8).
terms of the minimal path sets of the system.
For every j [r] = {1, . . . , r}, let Ej be the event that Example 2. Consider a 4-component system defined by
all components in the minimal path set Pj work. Then, the three minimal path sets P1 = {1, 2}, P2 = {2, 3},
using the inclusion-exclusion formula for probabilities, we and P3 = {3, 4}. The constituting elements of the sum in
obtain S
 [  B (1)|B|1 jB Pj
h(p) = Pr Ej {1} 1 {1, 2}
j[r] {2} 1 {2, 3}
X \ 
|B|1 {3} 1 {3, 4}
= (1) Pr Ej . (5)
{1, 2} 1 {1, 2, 3}
6=B[r] jB
{1, 3} 1 {1, 2, 3, 4}
Let Di denote the event that component i works. Then, we {2, 3} 1 {2, 3, 4}
have pi = Pr(Di ) and, using the independence assumption, {1, 2, 3} 1 {1, 2, 3, 4}
we have
 \  Y
Pr(Ej ) = Pr Di = pi Table 2: Example 2
iPj iPj
Eq. (8) are gathered in Table 2. Summing up the monomi-
and, more generally, als defined by the subsets given in the third column, each
\  \ \ 
multiplied by the corresponding number (+1 or 1) from
Pr Ej = Pr Di the second column, by (8) we obtain
jB jB iPj
 \  Y (x1 , x2 , x3 , x4 ) = x1 x2 + x2 x3 + x3 x4 x1 x2 x3 x2 x3 x4 ,
= Pr Di = pi . (6)
S S
i jB Pj i jB Pj which is the simple form of the structure function.
Substituting (6) in (5), we obtain the following multilinear Interestingly, Theorem 1 enables us to identify the min-
expression of h(p) in terms of the minimal path sets of the imal path sets of the system from the simple form of the
system structure function by quick inspection. We state this result
X Y
h(p) = (1)|B|1 pi , (7) in the following immediate corollary.
S
6=B[r] i Pj
jB
Corollary 2. The minimal path sets P1 , . . . , Pr are ex-
or equivalently, actly the minimal elements (with respect to inclusion) of
X Y X Y the family of subsets defined by the monomials (or equiv-
h(p) = pi pi alently, the monomials with coefficient +1) in the simple
j[r] iPj {j,k}[r] iPj Pk
X Y form of the structure function.
+ pi
{j,k,l}[r] iPj Pk Pl
Corollary 2 enables us to reconstruct the minimal path
set representation (3) of the structure function from its
We now show that a similar formula can be obtained for simple form. Considering for instance the simple form of
the structure function without an appeal to the indepen- the bridge structure function as described in Example 1,
dence assumption on the system components. Actually, by Corollary 2 we see that the corresponding minimal path
our result and its proof are purely combinatorial and does sets are P1 = {1, 4}, P2 = {2, 5}, P3 = {1, 3, 5}, and
not need any stochastic setting. P4 = {2, 3, 4}. We then immediately retrieve Eq. (4).
Theorem 1. If P1 , . . . , Pr denote the minimal path sets Theorem 1 has the following additional consequence.
of the system, then Recall that a formation of a subset A of C is a collec-
X Y tion of minimal path sets whose union is A. A formation
(x) = (1)|B|1 xi , (8) of A is said to be odd (resp. even) if it is the union of an
S
6=B[r] i jB Pj odd (resp. even) number of minimal path sets. Note that a
or equivalently, particular formation can be both odd and even. By equat-
X Y X Y ing the corresponding terms in (2) and (8), we obtain the
(x) = xi xi following identity
j[r] iPj {j,k}[r] iPj Pk X
X Y d(A) = (1)|B|1 .
+ xi
{j,k,l}[r] iPj Pk Pl S6=B[r]
jB Pj =A

3
From this identity, we immediately retrieve the important which holds regardless of the distribution of the component
fact (see, e.g., Barlow and Iyer [2]) that the coefficient d(A) lifetimes, has been obtained for instance in [4, Eq. (3.1)],
is exactly the number of odd formations of A minus the [12, Eq. (2.2)], and [13, Eq. (3.4)]. The corresponding dual
number of even formations of A. version can be easily derived by considering parallel sys-
A dual argument enables us to yield an expression of tems and minimal cut sets; see for instance [13, Eq. (3.5)].
the simple form of the structure function in terms of the For every t > 0 we have
minimal cut sets of the system. Recall that a subset K of  
X
C is a cut set of the system if (C \ K) = 0. It is minimal Pr(T 6 t) = (1)|B|1 Pr max T 6 t .
S i
if (C \ K ) = 1 for every K K. If K1 , . . . , Ks denote 6=B[s]
i jB Kj
the minimal cut sets of the system, then
Y a Y Y 
3 Minimal path and cut sets of
(x) = xi = 1 (1 xi ) .
j[s] iKj j[s] iKj small sizes
Starting from the well-known fact that the minimal cut By identifying the variables x1 , . . . , xn in the multilinear
sets of the system are the minimal path sets of the dual, extension (x)
b of the structure function, we define its diag-
and vice versa, from Theorem 1 and Corollary 2 we imme- onal section (x,b . . . , x), which will be simply denoted by
diately derive the following dual versions. b
(x). From the simple form (2) of the structure function,
Theorem 3. If K1 , . . . , Ks denote the minimal cut sets of we immediately obtain the polynomial function
the system, then n
X
X Y b
(x) = dk xk ,
D (x) = (1)|B|1 xi . (9)
S k=1
6=B[s] i jB Kj
where X
Corollary 4. The minimal cut sets K1 , . . . , Ks are ex- dk = d(A) .
actly the minimal elements (with respect to inclusion) of AC
the family of subsets defined by the monomials (or equiv- |A|=k
alently, the monomials with coefficient +1) in the simple For instance, considering the bridge structure function de-
form of the dual structure function. fined in Example 1, we obtain (x) b = 2x2 +2x3 5x4 +2x5 .
Locks [7] described a method for generating all mini- By definition, the diagonal section of the multilinear ex-
mal cut sets from the set of minimal path sets (and vice tension of the structure function is also the one-variable
versa) using Boolean algebra. Interestingly, an alternative reliability function h : [0, 1] [0, 1] which gives the sys-
method simply consists in applying Corollary 4 to the dual tem reliability h(p) = h(p, . . . , p) of the system whenever
structure function the components are statistically independent and have the
Y a same reliability p.
D (x) = 1 (1 x) = xi . Using Theorem 1, we can easily express the function
j[r] iPj b
(x) in terms of the minimal path sets. We simply have
Consider for instance the structure function defined in Ex- X S
b
(x) = (1)|B|1 x| jB Pj |
ample 2. The dual structure function is given by 6=B[r]

D (x) = (x1 x2 )(x2 x3 )(x3 x4 )

b
and the coefficient dk of xk in (x) is then given by
= x1 x3 + x2 x3 + x2 x4 x1 x2 x3 x2 x3 x4 .
X
Corollary 4 then immediately yields the minimal cut sets dk = (1)|B|1 . (10)
of the system, namely K1 = {1, 3}, K2 = {2, 3}, and K3 = S B[r]
| jB Pj |=k
{2, 4}.
Remark 1. It is noteworthy that from (5) and (6) we can Dually, the coefficient dD of xk in bD (x) is given by
k
immediately derive a representation of the system relia-
X
bility function in terms of the reliability functions of the dD (1)|B|1 .
k = (11)
series systems defined from the unions of minimal path B[s]
S
sets. More precisely, for every t > 0 we have | jB Kj |=k

X  
Pr(T > t) = (1)|B|1
Pr min Ti > t , For every k [n], let k (resp. k ) denote the number of
S
6=B[r]
i jB Pj minimal path (resp. cut) sets of size k of the system. The
knowledge of these numbers for small k may be relevant
where T and Ti denote the lifetime of the system and the when analyzing the reliability of the system. For instance,
lifetime of component i, respectively. This representation, if the system has no minimal cut set of size 1, it may be

4
informative to count the number 2 of minimal cut sets of lifetimes. He originally defined sk as the probability that
size 2 and so forth. the k-th component
Pn failure causes the system to fail (hence
The following proposition shows that 1 and 2 (resp. the property k=1 sk = 1). More recently, Boland [5]
1 and 2 ) can be computed directly from the coefficients showed that this probability can be explicitly given by
d1 and d2 (resp. dD D
1 and d2 ), and vice versa. (13). The expression given in (14) was derived later in [8,
d1

Cor. 12] and [10, Prop. 3] (see also [11] for a preliminary
Proposition 5. We have 1 = d1 , 1 = dD 1 , 2 = 2 + work). Thus defined, the structure signature depends only
D

d2 , and 2 = d21 + dD2 . on the structure function and can actually be considered
for any system, without any assumption on the distribu-
The following example shows that, in general, for k > 3
tion of the component lifetimes.
neither k nor k can be determined only from the coeffi-
b Combining this concept with Proposition 5 shows that
cients of (x) and bD (x).
1 and 2 (resp. 1 and 2 ) can be computed directly
Example 3. Consider the structure functions from sn and sn1 (resp. s1 and s2 ), and vice versa. The
conversion formulas are given in the following proposition.
1 (x) = x1 x2 x3 x4
Proposition 6. We have
= x1 x2 + x3 x4 x1 x2 x3 x4 (12)
1 = d1 = nsn , (15)
and
1 = dD
1 = ns1 , (16)
d1

nsn



2 (x) = x1 x2 x1 x3 x2 x3 x4 2 = 2 + d2 = 2 + n2 (sn1 sn ) , (17)
dD




= x1 x2 + x1 x3 + x2 x3 x4 x1 x2 x3 x1 x2 x3 x4 . 2 = D ns1
+ n2 (s2 s1 ) .
2 + d2 = (18)
1
2

We have b1 (x) = b2 (x) = 2x2 x4 and hence bD 1 (x) = For instance, consider again the structure function 1
1 1 (1 x) = 1 2 (1 x) = 2 (x). However, we clearly given in (12). We have b1 (x) = 2x2 x4 , bD
b b b D 2
1 (x) = 4x
3 4 D D
have 3 = 0 for function 1 and 3 = 1 for function 2 . 4x + x , and therefore d1 = d1 = 0, d2 = 2, and d2 = 4.
Using Proposition 6, we finally obtain 1 = 1 = 0, 2 =
Remark 2. Contrary to the number k , the number of 2, = 4, and s = (0, 2 , 1 , 0).
2 3 3
path sets (not necessarily minimal) of size k can always
b
be determined from the coefficients of (x). Indeed, this Example 4. Consider an n-component system having 2
number is given by (see, e.g., [10, Prop. 1]) minimal cut sets of size 2 and no cut set of size 1. By

(16) and (18) we necessarily have s1 = 0 and s2 = 2 / n2 .
X X k   This result was expected since sk is the probability that,
nj
(A) = dj . assuming that the component lifetimes are continuous and
kj
AC j=0 i.i.d. (and hence exchangeable), the system fails exactly at
|A|=k
the k-th component failure. Thus, s1 is clearly zero and
Dually, the number of cut sets of size k is given by s2 is the ratio of the number 2 of minimal
cut sets of
size 2 (favorable cases) over the number n2 of pairs of
X k
X nj 
components (possible cases).
D (A) = dDj .
j=0
kj Remark 3. It is noteworthy that, combining (10) with (14),
AC
|A|=k we obtain a simple way to compute the structure signature
of the system directly from the minimal path sets. Dually,
We end this section by giving expressions for the num-
combining (11) with the immediate formula (see also [10,
bers 1 , 2 , 1 , and 2 in terms of the structure signature
Sect. 3.5])
of the system. k k1


X
Recall that the structure signature of the system is the sk =
j1

dD
n j
n-vector s = (s1 , . . . , sn ) whose k-th coordinate is defined j=1 j
as
shows how we can compute the structure signature directly
X 1 X 1 from the minimal cut sets.
sk =

n (A)


n (A) , (13)
AC |A| AC |A| Remark 4. One can easily show that, when the compo-
|A|=nk+1 |A|=nk nents have exchangeable lifetimes, the system reliability
function can be expressed as
or equivalently,
Xn



X nk
nk+1
j1
Pr(T > t) = dk Pr min{T1 , . . . , Tk } > t .
sk = n

dj . (14) k=1
j=1 j
This result shows that the n-vector d = (d1 , . . . , dn ) can
This concept was introduced in 1985 by Samaniego [16] be interpreted as a signature vector, called minimal sig-
for systems whose components have continuous and i.i.d. nature in [13, Def. 4.1]. The structure signature s can

5
then be computed from this minimal signature by using References
(14) and vice versa (see, e.g., [10]). Dually, one can show
that [1] R. E. Barlow. Engineering reliability. ASA and SIAM,
Philadelphia, PA, 1998.
Xn


Pr(T 6 t) = dD
k Pr max{T1 , . . . , Tk } 6 t , [2] R. E. Barlow and S. Iyer. Computational complexity of
k=1 coherent systems and the reliability polynomial. Prob-
ability in the Engineering and Informational Sciences,
and this formula then shows that the coefficients
2:461469, 1988.
dD D
1 , . . . , dn can be used to define the maximal signature
[13, Def. 4.2], which also determines the structure signa- [3] R. E. Barlow and F. Proschan. Statistical theory of re-
ture s (as indicated in Remark 3) and the n-vector d (and liability and life testing. To Begin With, Silver Spring,
vice versa). MD, 1981.

[4] H. W. Block, Y. Li, and T. H. Savits. Initial and final

Acknowledgments behaviour of failure rate functions for mixtures and
systems. J. Appl. Prob., 40:721740, 2003.
This research is supported by the internal research project
[5] P. J. Boland. Signatures of indirect majority systems.
F1R-MTH-PUL-15MRO3 of the University of Luxem-
J. Appl. Prob., 38:597603, 2001.
bourg.
[6] M. Grabisch, J.-L. Marichal, and M. Roubens. Equiva-
lent representations of set functions. Math. Oper. Res.,
Appendix 25(2):157178, 2000.

Proof of Theorem 1. The proof relies on the classical poly- [7] M. O. Locks. Inverting and minimalizing path sets
nomial inclusion-exclusion identity and cut sets. IEEE Trans. Reliability, R-27(2):107
Y X Y 109, 1978.
1 (1 zj ) = (1)|B|1 zj ,
[8] J.-L. Marichal. Subsignatures of systems. J. Multivari-
j[r] 6=B[r] jB
ate Analysis, 124:226236, 2014.
Q
which holds for all z1 , . . . , zn R. Setting zj = iPj xi [9] J.-L. Marichal. Computing subsignatures of systems
in the latter identity and then combining the resulting for- with exchangeable component lifetimes. Stat. Prob.
mula with the right-hand expression in (3), we immediately Letters, 94:128134, 2014.
obtain
X Y Y [10] J.-L. Marichal. Algorithms and formulas for conver-
(x) = (1)|B|1 xi . sion between system signatures and reliability func-
6=B[r] jB iPj tions. J. Appl. Prob., 52(2):490507, 2015.

Formula (8) then follows by simplifying the latter expres- [11] J.-L. Marichal and P. Mathonet. Computing system
sion using x2i = xi . signatures through reliability functions. Stat. Prob.
Letters, 83(3):710717, 2013.
Proof of Proposition 5. On the one hand, setting k = 1 in [12] J. Navarro, Y. del Aguila, M. A. Sordo, and A. Suarez-
(10) and (11) shows that d1 = 1 and dD 1 = 1 . On the Llorens. Preservation of reliability classes under the
other hand, setting k = 2 in (10), we obtain formation of coherent systems. Appl. Stochastic Models
    Bus. Ind., 30:444454, 2014.
d2 =  i [r] : |Pi | = 2   {i, j} [r] : |Pi Pj | = 2  ,

[13] J. Navarro, J. M. Ruiz, and C. J. Sandoval. Proper-
that is d2 = 2 21 . Dually, we obtain dD 2 = 2 ties of coherent systems with dependent components.
1


2 .
Communications in StatisticsTheory and Methods,
36(1):175191, 2007.
Proof of Proposition 6. From Eq. (14) we
immediately de- [14] G. Owen. Multilinear extensions of games. Manage-
rive the equations d1 = nsn and d2 = n2 (sn1 sn ). Now,
ment Sci., 18:P64P79, 1972.
if sD = (sD D
1 , . . . , sn ) denotes the structure signature asso-
D
ciated with the dual structure function , then we have [15] K. G. Ramamurthy. Coherent structures and sim-
sDk = sn+1k for k = 1, . . . , n. Combining this observation ple games, volume 6 of Theory and Decision Library.
with the previous two equations,

we obtain immediately Series C: Game Theory, Mathematical Programming
n
dD D
1 = ns1 and d2 = 2 (s2 s1 ). We then conclude by and Operations Research. Kluwer Academic Publish-
Proposition 5. ers Group, Dordrecht, 1990.

6
[16] F. J. Samaniego. On closure of the IFR class under
formation of coherent systems. IEEE Trans. Reliability
Theory, 34:6972, 1985.

Jean-Luc Marichal is currently an associate professor in

the Mathematics Research Unit at the University of Lux-
embourg. He received his Ph.D. in Mathematics at the
University of Liege (Belgium) in 1998. His research area
mainly includes aggregation function theory, functional
equations, non-additive measures and integrals, conjoint
measurement theory, cooperative game theory, and system
reliability theory.