Laboratoire d'Analyse et Modélisation de Systèmes pour
l'Aide à la Décision
CNRS UMR 7024
CAHIER DU LAMSADE
170
Juin 2000
The probabilistic minimum vertex covering problem
Cécile Murat, Vangelis Th. Paschos
Résumé ii
Abstra t ii
1 Introdu tion 1
2 The strategies M1, M2 and M3 and a general preliminary result 2 2.1 Spe i ation of M1, M2and M3 . . . 2 2.2 Arst expressionfor the fun tionals . . . 2
3 The omplexity of
(H, C, M1, E(G
M1
C
), min)
34 The omplexity of
(H, C, M2, E(G
M2
C
), min)
45 The omplexity of
(H, C, M3, E(G
M3
C
), min)
45.1 Building
E(G
M3
C
)
. . . 4 5.2 BoundsforE(G
M3
C
)
. . . 6 5.3 Onthe quality ofthe boundsobtained . . . 76 Some simulation results 10
7 Con lusions 11
A Proof of expression (11) 13
Résumé
Une instan eduproblèmedela ouverture desommetsprobabilisteestune paire(
G =
(V, E), Pr
)obtenueenasso iantà haquesommetv
i
∈ V
uneprobabilité((
d'o uren e)) p
i
. Nous onsidéronsunestratégiedemodi ationMtransformantune ouverturedesommetsC
deG
enune ouverturede sommetsC
I
pourlesous-graphedeG
induit parl'ensemblede sommetsI
⊆ V
. L'obje tif, pourla ouverture de sommetsprobabiliste,est de déterminer une ouverturedesommetsdeG
minimisantlasomme,sur touslessous-ensemblesI
⊆ V
, des produits: probabilité deI
foisC
I
. Dans et arti le, nous étudions la omplexité de résolutionoptimaleduproblèmede ouverturedesommetprobabiliste.Mots- lé: problèmes ombinatoires, omplexité, ouverturedesommets,graphe
The probabilisti minimum vertex overing problem
Abstra t
An instan e of the probabilisti vertex- overing problem is a pair
(G = (V, E), Pr)
ob-tainedbyasso iatingwithea hvertexv
i
∈ V
ano urren e probabilityp
i
. We onsidera modi ationstrategy Mtransformingavertex overC
forG
intoavertex overC
I
for the subgraph ofG
indu edby avertex-setI
⊆ V
. Theobje tive for the probabilisti vertex- overingistodetermineavertex overofG
minimizingthesum,overallsubsetsI
⊆ V
,of theprodu ts: probabilityofI
timesC
I
. Inthispaper,westudythe omplexityofoptimally solvingprobabilisti vertex overing.Given a graph
G = (V, E)
, a vertex over is a setC
⊆ V
su h that for everyv
i
v
j
∈ E
eitherv
i
∈ C
, orv
j
∈ C
. The vertex- overing problem (VC) onsists in nding a vertex over of minimum size. A naturalgeneralization ofVC,denotedbyWVCinthe sequel,istheone where positive weights areasso iated with the verti es ofV
. The obje tive for WVC is to ompute a vertex over forwhi hthe sumofthe weightsofitsverti es isthesmallest over allvertex overs ofG
.Let
G
be the set of graphs. GivenG
∈ G
, denote byC
(G)
the set of vertex overs ofG
. For any subsetI
⊆ V
we denote byG[I]
the subgraph ofG
indu ed byI
. The probabilisti vertex- overing, denoted by PVC in what follows, is aquintuple(
H, C(G), M, E(G
M
C
), min)
su h that:H
is the set of instan es of PVC dened asH = {G = (G, Pr)}
, whereG
∈ G
andPr
is ann
-ve tor of vertex-probabilities;C
(G)
is the setof vertex oversofG
(C
(G) = C(G)
);M
is a modi ation strategy, i.e., an algorithm re eivingC
∈ C(G)
and aG[I]
,I
⊆ V
as inputs and modifyingC
inorder to produ e avertex over forG[I]
;E(G
M
C
)
is the fun tional, or obje tive value ofC
; onsiderG
∈ H
,C
∈ C(G)
,I
⊆ V
, and letp
i
= Pr[v
i
]
,v
i
∈ V
, be the probability that vertexv
i
is present; moreover, denote byC
M
I
the solution omputed byM
(C, G[I])
and by|C
M
I
|
the ardinality ofC
M
I
; nally, denotebyPr[I] =
Q
i∈I
p
i
Q
i∈V \I
(1
−p
i
)
the probabilitythatvertex-subsetI
ispresent; thenE(G
M
C
) =
P
I⊆V
Pr[I]
|C
I
M
|
.The omplexityofPVCisthe omplexityof omputing
C = argmin
ˆ
C∈C(
G
)
{E(G
M
C
)
}
. Using the terminology of [4, 5℄, we will all the solutionC
(in whi h the modi ation strategy isapplied) a priori solution.Apriorioptimization,i.e., sear hingforoptimalapriorisolutionsofprobabilisti ombinato-rialoptimizationproblems,hasbeenstudiedinrestri tedversionsofroutingandnetwork-design probabilisti minimizationproblems([1 ,2 ,3,4,5,6,7 ,8℄), denedon ompletegraphs. Finally, a priori optimization has been used in [9℄ to study a probabilisti maximization problem, the longestpath.
One anremarkthatin theabove denitionofPVC, M ispartoftheinstan eandthis seems somewhat unusual withrespe ttostandard omplexitytheorywhereno algorithm intervenes in the denition of a problem. But M is absolutely not an algorithm for PVC, in the sense that it does not ompute
C
∈ C(G)
. It simply tsC
(no matter how it has been omputed) toI
.Moreover,letusnotethatwhen hanging Mone hangesthedenitionofPVCitself. Stri tly speaking, the PVC-variants indu ed by the quintuples
Q1 = (
H, C, M
1
, E(G
M
1
C
), min)
andQ2 =
(
H, C, M
2
, E(G
M
C
2
), min)
aretwo distin t probabilisti ombinatorial optimization problems. Weusethreemodi ationstrategiesM1,M2andM3andstudytheir orrespondingfun tionals. For M1 and M2, we produ e expli it expressions for their fun tionals, expressions allowing us to ompletely hara terize the solutions minimizing these fun tionals. On the ontrary, the expressionforthefun tionalasso iatedwithM3isnotquiteexpli itinordertoallowa hievement of similar results. Therefore, we give boundsfor this fun tional and study their quality. Let us re allthat the threestrategies introdu e in fa tthree distin t probabilisti PVC-variants.In what follows, given a graph
G = (V, E)
of ordern
, we denote byC
a vertex over ofG
, byτ
its ardinality and byC
ˆ
i
,i = 1, 2, 3
, the optimal PVC-solutions asso iated withM1
,M2
andM3
, respe tively. Moreover, byΓ(v
i
)
,i = 1, . . . , n
, we denote the set of neighbors of thevertex
v
i
(thedegree ofv
i
). Wewill setδ = min
v
i
∈V
{|Γ(v
i
)
|}
and∆ = max
v
i
∈V
{|Γ(v
i
)
|}
. Given avertex overC
ofG
andasubsetV
′
⊆ V
,we willsetC[V
′
] = C
∩ V
′
and¯
C[V
′
] = (V
\ C) ∩ V
′
. Finally, when dealing with WVC,we will denotebyw
i
the weight ofv
i
∈ V
.2 The strategies M1, M2 and M3 and a general preliminaryresult 2.1 Spe i ation of M1, M2 and M3
Given a vertex over
C
and a vertex-subsetI
⊆ V
,M 1
onsists in simply moving verti es ofC
\ I
(theabsent verti es ofC
) out ofC
andin retainingC[I] = C
∩ I
asvertex over ofG[I]
. BEGIN (*M1(C,G[I℄)*)C
[I]
← C ∩ I
; OUTPUT C[I℄; END. (*M1(C,G[I℄)*)Itiseasy to seethat
C[I]
onstitutesa vertex over forG[I]
. Infa t,in the opposite ase, there would be at leastan edgev
i
v
j
ofG[I]
for whi h neitherv
i
, norv
j
would belong toG[I]
. But, sin ev
i
v
j
∈ E
andC
is a vertex over forG
, at least one ofv
i
,v
j
, sayv
i
, belongs toC
and, onsequently, toC[I] = C
∩ I
. Thereforev
i
ispartofthe vertex over forG[I]
, ontradi ting so the statement thatedgev
i
v
j
isnot overedinG[I]
.The se ond strategy studied in this paper is a slight improvement of M1 sin e one removes isolated verti es from
C[I]
.BEGIN (*M2(C,G[I℄)*)
C
[I]
← C ∩ I
;R
← {v
i
∈ C[I] : Γ(v
i
) =
∅}
; OUTPUTC2
[I]
← C[I] \ R
; END. (*M2(C,G[I℄)*)Finally, strategy M3is a further improvement of M2sin e itremoves from
C2[I]
verti es all the neighbors of whi h belong toC2[I]
. In the following spe i ation of M3, algorithm SORT sorts the verti es ofC2[I]
in in reasing order with respe tto theirdegrees.BEGIN (*M3(C,G[I℄)*)
C2
[I]
← M2(C, G[I])
;C3
[I]
← SORT(C2[I])
; FORi
← 1
TO |C3[I℄| DOIF
∀v
j
∈ Γ(v
i
), v
j
∈ C3[I]
THENC3
[I]
← C3[I] \ {v
i
}
FI ODOUTPUT C3[I℄; END. (*M3(C,G[I℄)*)
Itiseasytoseethatthevertex over
C3[I]
omputedbystrategyM3isminimal(forthein lusion) forG[I]
.2.2 A rst expression for the fun tionals
Consider any strategy that starting from
C[I]
redu es it by removing some of its verti es (if possible) in order to obtain smaller feasible vertex overs. Clearly, M1, M2 and M3 are su h strategies. Then, the following proposition provides a rst expression for fun tionalsE(G
M1
C
)
,E(G
M2
C
)
andE(G
M3
C
)
.Proposition 1. Considera vertex over
C
ofG
andstrategies M1,M2andM3. Withea hvertexv
i
∈ V
asso iate a probabilityp
i
and a random variableX
Mk,C
i
,k = 1, 2, 3
, dened, for everyI
⊆ V
,byX
i
Mk,C
=
1 v
i
∈ C
Mk
I
0
otherwise Then,E(G
Mk
C
) =
P
v
i
∈C
Pr[X
Mk,C
i
= 1]
.Proof. Bythe denition of
X
Mk,C
i
we have|C
Mk
I
| =
P
n
i=1
X
Mk,C
i
. So,E G
Mk
C
=
X
I⊆V
Pr[I]
C
I
M
=
X
I⊆V
Pr[I]
n
X
i=1
X
i
Mk,C
=
n
X
i=1
X
I⊆V
Pr[I]X
i
Mk,C
=
n
X
i=1
E
X
i
Mk,C
=
n
X
i=1
Pr
h
X
i
Mk,C
= 1
i
=
n
X
i=1
Pr
h
X
i
Mk,C
= 1
i
1
{v
i
∈C}
+ 1
{v
i
∈C}
/
=
n
X
i=1
Pr
h
X
i
Mk,C
= 1
i
1
{v
i
∈C}
+
n
X
i=1
Pr
h
X
i
Mk,C
= 1
i
1
{v
i
∈C}
/
.
But, if
v
i
∈ C
/
, then,∀I ⊂ V
su h thatv
i
∈ I
,X
Mk,C
i
= 0
. So,Pr[X
Mk,C
i
= 1] = 0
,∀v
i
∈ C
/
. Consequently,E(G
Mk
C
) =
P
n
i=1
Pr[X
Mk,C
i
= 1]1
{v
i
∈C}
=
P
v
i
∈C
Pr[X
Mk,C
i
= 1]
.3 The omplexity of
(H, C, M1, E(G
M1
C
), min)
Thefollowing theorem an bededu ed fromproposition 1. Theorem 1 .E(G
M1
C
) =
P
v
i
∈C
p
i
andC
ˆ
1
= argmin
C∈C(G)
{
P
v
i
∈C
p
i
}
. In other wordsC
ˆ
1
is a minimum-weight vertex over ofG
where the verti es ofV
are weighted by their orresponding presen e-probabilities. Inthe asewhereallvertex-probabilitiesareidenti al,E(G
M1
C
) = pτ
andˆ
C
1
isa minimum vertex over ofG
. Consequently,(
H, C, M1, E(G
M1
C
), min)
isNP-hard.Proof. By strategy M1, if
v
i
∈ C
, thenv
i
∈ C
M1
I
, for all subgraphsG[I]
su h thatv
i
∈ I
. Consequently,Pr[X
M1,C
i
= 1] = p
i
,∀v
i
∈ C
andthe resultfollows from proposition1.Let us now onsider
G
with its verti es weighted by their orresponding probabilities and denote the soobtained instan e of WVCbyG
w
. The total weight of every vertex over ofG
w
isP
v
i
∈C
p
i
and the optimal weighted vertex over ofG
w
is the one for whi h the sum of the weightsof itsverti esisthesmallest over allvertex oversofG
w
. Su havertex over minimizes alsoE(G
M1
C
)
and, onsequently onstitutes anoptimal solution for(
H, C, M1, E(G
M1
C
), min)
. Ifp
i
= p
,1 6 i 6 n
, thenP
v
i
∈C
p
i
=
P
v
i
∈C
p = pτ
and, onsequently, the vertex over minimizing this last expressionisthe one minimizingτ
, a minimum vertex over ofG
.Expression
P
v
i
∈C
p
i
(theobje tivevalueoftheproblem(
H, C, M1, E(G
M1
C
), min)
) anbe om-putedinO(n)
,therefore,(
H, C, M1, E(G
M1
C
), min)
∈ NP
. Furthermore,bytheisomorphybetween thisproblemandWVC,the formerproblemishard forNPandthis ompletesthe proofof the-orem 1.Let us revisit ase
p
i
= p
,1 6 i 6 n
. Sin eC
M
I
= C[I]
and0 6
|C[I]| 6 τ
, we get:|C
M
I
| = |C[I]|
P
τ
i=1
1
{|C[I]|=i}
and the fun tionalforM1
an be written asE G
M1
C
=
X
I⊆V
Pr[I]
|C[I]|
τ
X
i=1
1
{|C[I]|=i}
=
τ
X
i=1
i
X
I⊆V
Pr[I]1
{|C[I]|=i}
=
τ
X
i=1
i
τ
i
p
i
(1
− p)
τ −i
n−τ
X
j=0
n − τ
j
p
j
(1
− p)
n−τ −j
= pτ
where in thelast summation we ount allthe sub-graphs
G[I]
su hthat|C[I]| = i
, weadd their probabilities and we take into a ount thatP
n−τ
j=0
C
j
n−τ
p
j
(1
− p)
n−τ −j
= 1
.The above an be generalized in order to ompute every moment of any order for
M1
. For instan e,E((C
M
I
)
2
) =
P
I⊆V
Pr[I](
|C
I
M
|)
2
=
P
τ
i=1
i
2
C
i
τ
p
i
(1
− p)
τ −i
= pτ (pτ + 1
− p)
and, onse-quently,Var G
M1
C
= E
G
M1
C
2
− E G
M1
C
2
= τ p(1
− p).
So,for
M1
andfor the aseofidenti alvertex-probabilities,the randomvariablerepresentingthe sizeτ
follows abinomial lawwith parametersτ
andp
.4 The omplexity of
(H, C, M2, E(G
M2
C
), min)
We re all that strategy M2 onsists in removing the isolated verti es from
C[I]
. Then, the following theoremholds.Theorem 2.
E G
M2
C
=
X
v
i
∈C
p
i
1
−
Y
v
j
∈Γ(v
i
)
(1
− p
j
)
and
C
ˆ
2
is a minimum-weightvertex over ofG
w
where, for everyv
i
∈ V
,w
i
= p
i
1
−
Y
v
j
∈Γ(v
i
)
(1
− p
j
)
.
Consequently,(
H, C, M2, E(G
M2
C
), min)
isNP-hard.Proof. Remarkrst that
C
M2
I
= C2[I] = C[I]
\ (∪
{v
i
∈C:(Γ(v
i
)∩I)=∅}
v
i
)
. Starting fromproposition 1we getE G
M2
C
=
X
v
i
∈C
Pr
h
X
i
M2,C
= 1
i
=
X
v
i
∈C
X
I⊆V
vi∈I
Pr[I]1
{v
i
∈C2[I]}
=
X
v
i
∈C
X
I⊆V
vi∈I
Pr[I] 1
− 1
{v
i
∈C2[I]}
/
.
But, for any
v
i
∈ C
,v
i
∈ C2[I] ⇐⇒ I ∩ Γ(v
/
i
) =
∅
; onsequently,E G
M2
C
=
X
v
i
∈C
X
I⊆V
vi∈I
Pr[I]
−
X
v
i
∈C
X
I⊆V
vi∈I
Pr[I]1
{I∩Γ(v
i
)=∅}
=
X
v
i
∈C
p
i
1
−
Y
v
j
∈Γ(v
i
)
(1
− p
j
)
.
Itiseasyto seethat
E G
M2
C
anbe omputed inO(n
2
)
; onsequently(
H, C, M2, E(G
M2
C
), min)
∈
NP
. Witha reasoning ompletely similarto the oneof theorem 1 one an immediatelydedu e thatC
ˆ
2
= argmin
C∈C(G)
{
P
v
i
∈C
p
i
(1
−
Q
v
j
∈Γ(v
i
)
(1
− p
j
))
}
, i.e., aminimum-weight vertex over of
G
w
where,forv
i
∈ V
,w
i
= p
i
(1
−
Q
v
j
∈Γ(v
i
)
(1
− p
j
))
. 5 The omplexity of(H, C, M3, E(G
M3
C
), min)
5.1 BuildingE(G
M3
C
)
Re allthatM3 onsistsin removing from
C[I]
boththe isolatedverti es and the verti es all the neighborsofwhi hbelonginC[I]
. Consequently, thesolutionC3[I]
omputedbyM3isminimal.E G
M3
C
=
X
vi∈C
p
i
1
−
Y
vj
∈Γ(vi)
(1
− p
j
)
−
X
vi∈C
X
I⊆V
Pr[I]1{
vi
∈C
/
M3
I
}1
{I∩C[Γ(vi)]6=∅}
1{
I∩ ¯
C[Γ(vi)]=∅
}
(1)
Proof. Byproposition 1,
E(G
M3
C
) =
P
v
i
∈C
Pr[X
M3,C
i
= 1]
. Moreover,Pr
h
X
i
M3,C
= 1
i
=
X
I⊆V
Pr[I]1{
v
i
∈C
I
M3
}.
Note that
Γ(v
i
) = C[Γ(v
i
)]
∪ ¯
C[Γ(v
i
)]
and that ifv
i
∈ C
, thenC[Γ(v
¯
i
)]
6= ∅
(be auseC
is minimal).Letnow
v
i
∈ C
andI
i
=
{I ⊆ V : v
i
∈ I}
. Thesetof graphsG[I]
,I
∈ I
i
an be partitioned in the following four subsets,a graphG[I]
belonging to onlyone of them:(i) graphssu hthat
I
∩ Γ(v
i
) =
∅
;(ii) graphssu hthat
I
∩ C[Γ(v
i
)]
6= ∅
andI
∩ ¯
C[Γ(v
i
)]
6= ∅
; (iii) graphssu hthatI
∩ C[Γ(v
i
)]
6= ∅
andI
∩ ¯
C[Γ(v
i
)] =
∅
; (iv) graphssu hthatI
∩ C[Γ(v
i
)] =
∅
andI
∩ ¯
C[Γ(v
i
)]
6= ∅
. Letus analysestrategy M3in ases (i)to (iv).In ase (i) sin e
v
i
is isolated,M3will remove itfromC[I]
. In ases (ii) and (iv)v
i
will be in luded inC
M3
I
(sin eI
∩ ¯
C[Γ(v
i
)]
6= ∅
,∃v
j
∈ I ∩ Γ(v
i
)
su h thatv
j
∈ C
/
).In ase (iii) thefa tthat
v
i
will be in ludedin, orwill be removed from, the solutiondepends on the order on whi h it will be treated by M3 and on wether or not all its neighbors are partofC[I]
.Inall,
• v
i
∈ C
/
I
M3
forall graphs verifying ase(i) andfor a partofgraphs verifying ase(iii);• v
i
∈ C
I
M3
for all graphs verifying ases (ii) and (iv) and for the rest of graphs verifying ase(iii). We sohave:Pr[X
i
M3,C
= 1] =
X
I⊆V
Pr[I]1{
vi∈C
M3
I
} =
X
I⊆V
Pr[I]1{
vi∈C
M3
I
}
1
{I∩C[Γ(vi)]6=∅}
1{
I∩ ¯
C[Γ(vi)]6=∅
}
+ 1
{I∩C[Γ(vi)]=∅}
1{
I∩ ¯
C[Γ(vi)]6=∅
} + 1
{I∩C[Γ(vi)]6=∅}
1{
I∩ ¯
C[Γ(vi)]=∅
}
=
X
I⊆V
Pr[I]1{
vi∈C
M3
I
}
1{
I∩ ¯
C[Γ(vi)]6=∅
} + 1
{I∩C[Γ(vi)]6=∅}
1{
I∩ ¯
C[Γ(vi)]=∅
}
=
X
I⊆V
Pr[I]1{
vi∈C
M3
I
}1{
I∩ ¯
C[Γ(vi)]6=∅
}
+
X
I⊆V
Pr[I]1{
vi∈C
M3
I
}1
{I∩C[Γ(vi)]6=∅}
1{
I∩ ¯
C[Γ(vi)]=∅
}
=
X
I⊆V
Pr[I]1{
vi∈C
M3
I
}
1
− 1{
I∩ ¯
C[Γ(vi)]=∅
}
+
X
I⊆V
Pr[I]1{
vi∈C
M3
I
}1
{I∩C[Γ(vi)]6=∅}
1{
I∩ ¯
C[Γ(vi)]=∅
}
= p
i
1
−
Y
vj
∈ ¯
C[Γ(vi)]
(1
− p
j
)
+
X
I⊆V
Pr[I]1{
vi∈C
M3
I
}1
{I∩C[Γ(vi)]6=∅}
1{
I∩ ¯
C[Γ(vi)]=∅
}
Pr[X
i
M3,C
= 0] =
X
I⊆V
vi∈I
Pr[I]1{
vi
∈C
/
M3
I
} +
X
I⊆V
vi /
∈I
Pr[I]1{
vi
∈C
/
M3
I
}
=
X
I⊆V
vi∈I
Pr[I]1{
vi
∈C
/
M3
I
}
1
{I∩Γ(vi)=∅}
+ 1
{I∩C[Γ(vi)]6=∅}
1{
I∩ ¯
C[Γ(vi)]=∅
}
+ 1
− p
i
=
X
I⊆V
vi∈I
Pr[I]1{
vi
∈C
/
M3
I
}1
{I∩Γ(vi)=∅}
+ 1
− p
i
+
X
I⊆V
vi∈I
Pr[I]1{
vi
∈C
/
M3
I
}1
{I∩C[Γ(vi)]6=∅}
1{
I∩ ¯
C[Γ(vi)]=∅
}
= p
i
Y
vj
∈Γ(vi)
(1
− p
j
) + 1
− p
i
+
X
I⊆V
vi∈I
Pr[I]1{
vi
∈C
/
M3
I
}1
{I∩C[Γ(vi)]6=∅}
1{
I∩ ¯
C[Γ(vi)]=∅
}
It is easy to see, after some easy but tedious algebra, that
Pr[X
M3,C
i
= 1] + Pr[X
M3,C
i
= 0] = 1
. Then, the resultof proposition2 isget usingE G
M3
C
=
X
v
i
∈C
Pr
h
X
i
M3,C
= 1
i
=
X
v
i
∈C
1
− Pr
h
X
i
M3,C
= 0
i
.
5.2 Bounds for
E(G
M3
C
)
Letusrstnotethatfromtheorems1,2andproposition2,wehave
E(G
M3
C
) 6 E(G
M2
C
) 6 E(G
M1
C
)
. Sin etheresultofproposition2doesnotallowthea hievementofapre ise hara terizationofC
ˆ
3
, we givein the following theoremboundsforE(G
M3
C
)
. Theorem 3.X
v
i
∈C
p
i
1
−
Y
v
j
∈ ¯
C[Γ(v
i
)]
(1
− p
j
)
6
E G
M3
C
6
X
v
i
∈C
p
i
1
−
Y
v
j
∈Γ(v
i
)
(1
− p
j
)
.
Ifp
i
= p
,∀v
i
∈ V
andδ
C
= min
v
i
∈C
{|Γ(v
i
)
|}
, thenp
τ
− max
n
τ (1
− p)
∆
, (1
− p)
δ
C
o
>
E G
M3
C
>
p
2
τ.
The upper bound for
E(G
M3
C
)
is attained for the bipartite graphs when onsidering the one of the olor lasses as a priori solution.Proof. The upper bound is obvious sin e it is nothing else than the expression of theorem 2 for
E(G
M2
C
)
. LetA
i
=
X
I⊆V
vi∈I
Pr[I]1{
v
i
∈C
/
I
M3
}1
{I∩C[Γ(v
i
)]6=∅}
1{
I∩ ¯
C[Γ(v
i
)]=∅
}
A
i
6
X
I⊆V
vi∈I
Pr[I]1
{I∩C[Γ(v
i
)]6=∅}
1{
I∩ ¯
C[Γ(v
i
)]=∅
}
6
X
I⊆V
vi∈I
Pr[I] 1
− 1
{I∩C[Γ(v
i
)]=∅}
1{
I∩ ¯
C[Γ(v
i
)]=∅
}
6
p
i
Y
v
j
∈ ¯
C[Γ(v
i
)]
(1
− p
j
)
−
Y
v
j
∈ ¯
C[Γ(v
i
)]
(1
− p
j
)
Y
v
j
∈C[Γ(v
i
)]
(1
− p
j
)
6
p
i
Y
v
j
∈ ¯
C[Γ(v
i
)]
(1
− p
j
)
−
Y
v
j
∈Γ(v
i
)
(1
− p
j
)
.
Combiningtheexpressionfor
A
i
abovewiththeexpressionforE(G
M3
C
)
ofproposition2,weeasily getthe lower bound laimed.Let usnowsupposethat
p
i
= p
,∀v
i
∈ V
. Then the expressionforE(G
M3
C
)
be omesE G
M3
C
= pτ − p
X
v
i
∈C
(1
− p)
|Γ(v
i
)|
−
X
v
i
∈C
A
i
with0
≤ A
i
≤ p((1 − p)
| ¯
C[Γ(v
i
)]|
− (1 − p)
|Γ(v
i
)|
)
. Sin eA
i
>
0
we haveE G
M3
C
6
pτ
− p
X
v
i
∈C
(1
− p)
|Γ(v
i
)|
.
Using,
∀i
,|Γ(v
i
)
| ≤ ∆
,wegetE G
M3
C
6
p τ
− τ(1 − p)
∆
.
(3) Remarkthatp
X
v
i
∈C
(1
− p)
|Γ(v
i
)|
= p(1
− p)
δ
C
X
v
i
∈C
(1
− p)
|Γ(v
i
)|−δ
C
X
v
i
∈C
(1
− p)
|Γ(v
i
)|−δ
C
>
1
be ausethere existsat leastavertex
v
i
0
∈ C
with|Γ(v
i
0
)
| = δ
C
. Consequently,E G
M3
C
6
p
τ
− (1 − p)
δ
C
.
(4)Combining expressions (3)and(4) oneimmediately obtains the upper bound laimed. Foridenti alvertex-probabilities,thelowerboundfor
E(G
M3
C
)
be omespτ (1
−(1−p)
| ¯
C[Γ(v
i
)]|
)
. Sin eC
is minimal,¯
C[Γ(v
i
)]
6= ∅
, i.e.,| ¯
C[Γ(v
i
)]
| > 1
. Using the latter inequality in the former one we getE(G
M3
C
) > pτ (1
− (1 − p)) = p
2
τ
, proving sothe lower bound laimed.Considernowabipartitegraph
B = (V
1
, V
2
, E)
andoneofits olor lasses,say olor lassV
1
, as a priori solution. Remark that,∀v
i
∈ V
1
,Γ(v
i
) = ¯
V
1
[Γ(v
i
)]
∈ V
2
. Consequently, the upper and lower boundsforE(B
M3
V
1
)
oin ide.5.3 On the quality of the bounds obtained We prove in this se tionthatthe boundsof
E(G
M3
C
)
obtained in theorem3are quitetight . LetG = (V, E)
be a graph onsisting of a liqueK
ℓ
(onℓ
verti es) and of an independent setS
onσ
verti es; moreover onsider that any vertex ofK
ℓ
is linked to anyvertexofS
. Letus suppose thatall the verti es of
G
have the same probabilityp
and onsiderC = V (K
ℓ
)
, the vertex-set ofK
ℓ
, asa priorisolution. Finallysetn = ℓ + σ
.For
v
i
∈ C
wehaveC[Γ(v
i
)] = V (K
ℓ
)
\ {v
i
}
andC[Γ(v
¯
i
)] = S
. Then, expression(1)be omesE G
M3
C
=
X
v
i
∈C
p
−
X
v
i
∈C
p
Y
v
j
∈Γ(v
i
)
(1
− p) −
X
v
i
∈C
X
I⊆V
Pr[I]1{
v
i
∈C
/
I
M3
}1
{I∩(V (K
ℓ
)\{v
i
})6=∅}
1
{I∩S=∅}
(5)Let
V (K
ℓ
) =
{v
1
, v
2
, v
3
, . . . , v
ℓ
}
(the degrees of the verti es ofV (K
ℓ
)
are all equal) and letS =
{v
ℓ+1
, v
ℓ+2
, v
ℓ+3
, . . . , v
n
}
. Instepi
, strategy M3tests ifvertexv
i
an be removed.Revisit now the third term of expression (5) (the double sum). This term deals with in-stan es
G[I]
su hthatI
⊆ V (K
ℓ
)
\ {v
i
}
, in otherwords, instan eswhere all present verti es are elementsofV (K
ℓ
)
, i.e., they arealsopartofC
. Sin ev
i
∈ V (K
ℓ
)
,Γ(v
i
)
∩ I ⊆ C ∩ I
andvertexv
i
∈ C
an be removed only if, it isnot isolated and forj < i
,v
j
∈ I
/
(be ause,in the opposite ase,v
j
would be removed at stepj < i
, in other words, beforev
i
and, sin ev
i
v
j
∈ E(K
ℓ
)
,v
i
shouldnotberemoved). Consequently,1
{v
i
∈C
/
I
M3
}
= 1
{I∩{v
1
,v
2
,v
3
,...,v
i−1
}=∅}
andforaxedindexi
, expression(2) be omesA
i
=
X
I⊆V
Pr[I]1
{I∩{v
1
,v
2
,v
3
,...,v
i−1
}=∅}
1
{I∩V (K
ℓ
)\{v
i
}6=∅}
1
{I∩S=∅}
=
X
I⊆V
Pr[I]1
I
∩vi=vi
I∩
{
v1,v2,v3,...,vi−1
}
=∅
I∩
{
vi+1,vi+2,vi+3,...,vℓ
}
6=∅
1
{I∩S=∅}
=
i−1
Y
j=1
(1
− p)p
1
−
ℓ
Y
j=i+1
(1
− p)
n
Y
j=ℓ+1
(1
− p)
= (1
− p)
i−1
p
1
− (1 − p)
ℓ−(i+1)+1
(1
− p)
n−(ℓ+1)+1
= (1
− p)
i−1
p
1
− (1 − p)
ℓ−i
(1
− p)
n−ℓ
.
We sohave,X
v
i
∈C
A
i
=
ℓ
X
i=1
p
(1
− p)
n+i−ℓ−1
− (1 − p)
n−1
= p(1
− p)
n−ℓ−1
ℓ
X
i=1
(1
− p)
i
− pℓ(1 − p)
n−1
= (1
− p)
n−ℓ−1
1
− p − (1 − p)
ℓ+1
− pℓ(1 − p)
n−1
= (1
− p)
n−ℓ
− (1 − p)
n
− pℓ(1 − p)
n−1
(6) Remark that omputation ofA
i
's in the expression above is polynomial and, onsequently, the whole omputation forE(G
M3
C
)
is alsopolynomial. Combining nowexpressions (5)and (6),wegetE G
M3
C
=
ℓ
X
i=1
p
−
ℓ
X
i=1
p
Y
vj
∈Γ(vi)
(1
− p) − (1 − p)
n−ℓ
− (1 − p)
n
− pℓ(1 − p)
n−1
=
ℓp
− ℓp(1 − p)
n−1
− (1 − p)
n−ℓ
+ (1
− p)
n
+ pℓ(1
− p)
n−1
=
ℓp
− (1 − p)
n−ℓ
1
− (1 − p)
ℓ
.
e(ℓ, p, n)
def
=
ℓp
− (1 − p)
n−ℓ
1
− (1 − p)
ℓ
(7)b(ℓ, p, n)
def
=
ℓp
1
− (1 − p)
n−ℓ
(8)B(ℓ, p, n)
def
=
ℓp 1
− (1 − p)
n−1
(9) and remark thatexpressions (8)and (9) are, respe tively, the lower and upper boundsof theo-rem 3for
E(G
M3
C
)
.Wenowstudy the dieren e
(B
− e)(ℓ, p, n)
(expressions(7) and(9)). We have,(B
− e)(ℓ, p, n) = (1 − p)
n−ℓ
− (1 − p)
n
− ℓp(1 − p)
n−1
6
1
(10) Startingfrom expression(10), we prove in appendix thatthe following hold:0 6 D(ℓ, n)
def
= (B
− e)(ℓ, p, n) 6
n − ℓ
n
− 1
n−ℓ
ℓ−1
−
n − ℓ
n
− 1
n−1
ℓ−1
(11)Notethatthe upper boundofexpression(11)isquitetight for
D(ℓ, n)
. Letusnowestimatethis boundfor several values ofℓ
:•
ifℓ = 1
, thenD(ℓ, n) = 0
;•
if1 < ℓ ≪ n
(e.g.,ℓ = o(n)
), thenlim
n→∞
D(ℓ, n) = 0
;•
ifℓ = λn
, for axedλ < 1
, thenlim
n→∞
D(λn, n) = e
1−λ
λ
ln(1−λ)
− e
ln(1−λ)
λ
and thislimit's valueis axedpositive onstant;
•
if,nally,ℓ = n
− 1
, thenlim
n→∞
D(n
− 1, n) = 1
. Letus nowstudy the dieren e(e
− b)(ℓ, p, n)
. We rsthave(e
− b)(ℓ, p, n) = ℓp(1 − p)
n−ℓ
− (1 − p)
n−ℓ
+ (1
− p)
n
6
(ℓ
− 1)
(1
− p)
n−ℓ
− (1 − p)
n−ℓ+1
(12) one an renethe bound of expression(12)obtaining (seein the appendix for the proof)0 6 d(ℓ, n)
def
= (e
− b)(ℓ, p, n) 6 (ℓ − 1)
n
− ℓ
n
− ℓ + 1
n−ℓ
1
−
n
− ℓ
n
− ℓ + 1
6
(ℓ
− 1)e
(n−ℓ) ln
(
1−
n−ℓ+1
1
)
1
n
− ℓ + 1
(13) Aspreviously, we estimatethis bound for severalvalues ofℓ
:•
ifℓ = 1
, thend(ℓ, n) = 0
;•
ifℓ = o(n)
, thenlim
n→∞
d(ℓ, n) = 0
;•
ifℓ = λn
, for axedλ < 1
, thend(λn, n)
≤
λn
− 1
(1
− λ)n + 1
e
n(1−λ) ln
1−
1
(1−λ)n+1
;
λ
1
− λ
e
−1
p = 0.1
p = 0.5
p = 0.9
p = 1/n
p = 1/
√
n
p = 1/ ln n
E(G
M3
C
∗
)/E(G
M1
C
∗
)
0.34 0.82 0.98 0.26 0.62 0.74n = 14
E(G
M3
C
∗
)/E(G
M2
C
∗
)
0.73 0.88 0.98 0.73 0.79 0.84E(G
opt
)/E(G
M3
C
∗
)
0.97 0.96 0.99 0.98 0.95 0.95E(G
M3
C
∗
)/E(G
M1
C
∗
)
0.34 0.83 0.97 0.25 0.62 0.73n = 15
E(G
M3
C
∗
)/E(G
M2
C
∗
)
0.73 0.89 0.98 0.72 0.79 0.83E(G
opt
)/E(G
M3
C
∗
)
0.97 0.95 0.99 0.97 0.95 0.95E(G
M3
C
∗
)/E(G
M1
C
∗
)
0.35 0.83 0.98 0.25 0.62 0.73n = 16
E(G
M3
C
∗
)/E(G
M2
C
∗
)
0.72 0.88 0.98 0.71 0.78 0.83E(G
opt
)/E(G
M3
C
∗
)
0.96 0.95 0.99 0.97 0.94 0.94Table1: Relations between fun tionals.
p = 0.1
p = 0.5
p = 0.9
p = 1/n
p = 1/
√
n
p = 1/ ln n
n = 14
B
3
ℓ
/E(G
M3
C
∗
)
0.34 0.62 0.92 0.32 0.45 0.53B
u
3
/E(G
M3
C
∗
)
1.67 1.18 1.02 1.71 1.42 1.29n = 15
B
ℓ
3
/E(G
M3
C
∗
)
0.33 0.61 0.92 0.31 0.44 0.52B
u
3
/E(G
M3
C
∗
)
1.68 1.18 1.02 1.74 1.42 1.29n = 16
B
ℓ
3
/E(G
M3
C
∗
)
0.32 0.61 0.92 0.29 0.43 0.51B
u
3
/E(G
M3
C
∗
)
1.69 1.17 1.02 1.77 1.43 1.29Table 2: On the qualityof the boundsfor
E(G
M3
C
∗
)
.•
if,nally,ℓ = n
− 1
, thend(n
− 1, n) ∼ n/4
. One an remark that forℓ = o(n)
,E(G
M3
C
)
tends (forn
→ ∞
) to bothb(ℓ, n, p)
andB(ℓ, n, p)
. Thisisdue to the fa tthato(n)/ lim
n→∞
(B
− b)(ℓ, n, p) ; 0
.Weseefromtheabovedis ussionthatforthegraph andthe apriorisolution onsideredand for
n
→ ∞
, the distan e ofE(G
M3
C
)
fromthe bounds given in theorem 3 an take an innity of valuesbeingeither arbitrarily loseto,or arbitrarilyfarfromthem. Consequently,itseemsvery unlikely thatthe bounds omputed ouldbesubstantiallyimproved.6 Some simulation results
Limited omputational experiments have been ondu ted in order to study the realisti per-forman es of the three strategies proposed. We have randomly generated graphs of orders
n
∈ {14, 15, 16}
, fortygraphsperorder,inthefollowingway: forea hgraph,anumberp
E
∈ [0, 1]
is randomly drawn; next,for ea h edgee
, a randomly hosen probabilityp
e
is asso iated with, and, ifp
e
≤ p
E
, thene
is retained as an edge in the graph under onstru tion. For all the generatedinstan es,theminimum vertex-degreevariesbetween 0(graphnon- onne ted)and14, while the maximum one varies between 1 and 15. The ratio|E|/n
varied between 0.06and 7.3 for all the graphs generated. Thea priorisolution onsidered wasthe minimum vertex over of ea hgraph, denotedbyC
∗
. We have supposedidenti alvertex-probabilitiesand have tested six probability-values: 0.1, 0.5,0.9,
1/ ln n
,1/
√
n
and1/n
. We have omputed quantitiesE(G
M1
C
∗
)
,E(G
M2
C
∗)
,E(G
M3
C
∗)
andE(G
opt
)
. The last quantityis themathemati al expe tationofathought pro ess onsistinginndingtheoptimalvertex over
p = 0.1
p = 0.5
p = 0.9
p = 1/n
p = 1/
√
n
p = 1/ ln n
σ
M1
C
∗/E(G
M1
C
∗)
1.08 0.36 0.12 1.30 0.60 0.46n = 14
σ
M3
C
∗/E(G
M3
C
∗)
2.26 0.46 0.13 3.08 0.90 0.63σ
opt
/E(G
opt
)
2.23 0.46 0.14 3.05 0.89 0.63
σ
M1
C
∗/E(G
M1
C
∗)
1.03 0.34 0.11 1.29 0.58 0.45n = 15
σ
M3
C
∗/E(G
M3
C
∗)
2.13 0.44 0.13 3.08 0.88 0.61σ
opt
/E(G
opt
)
2.10 0.44 0.13 3.05 0.87 0.61σ
M1
C
∗/E(G
M1
C
∗)
0.99 0.33 0.11 1.28 0.57 0.44n = 16
σ
M3
C
∗/E(G
M3
C
∗)
2.03 0.42 0.12 3.13 0.87 0.61σ
opt
/E(G
opt
)
2.01 0.42 0.13 3.09 0.86 0.60
Table3: Standarddeviations for M1,M3 andreoptimization.
for any subgraph of
G
, i.e.,2
n
exe utions of bran h-and-bound. This is the reason for having restri tedourstudiesinsmallgraphs. Inthetable1wepresentthemeanratiosforthefun tionals of the dierent strategies. What we an see from this table is that the ratio
E(G
opt
)/E(G
M3
C
∗
)
variesbetween0.94and0.99. Inotherwords,
C
∗
seemstobevery loseto
C
ˆ
3
. Intheoppositethe omputationstimesforE(G
M3
C
∗)
arequitelarge,mu hmore largethanthe onesfor bothE(G
M1
C
∗)
andE(G
M2
C
∗)
, whose the ratios withE(G
M3
C
∗)
arerelatively small. Another remark we an make from theresults of table1 is thatthe larger the vertex-probabilities, the larger ( loser to 1) the ratios.Wealsohave onsideredthebounds
B
3
ℓ
andB
3
u
forE(G
M3
C
)
inthelastexpressionoftheorem3 and omputed their ratioswithE(G
M3
C
∗)
. The results are presented in table 2. The mean value of ratioB
3
ℓ
/E(G
M3
C
∗)
is 0.52, while forB
3
u
/E(G
M3
C
∗)
this value is 1.39. So, although these values are fairly lose to 1, it seems to us that the bounds onsidered have to be improved. Also one an notethaton e more the larger the vertex-probabilities, the loserto 1 the ratios.Finally,forstrategiesM1,M3wehave omputedtheirstandarddeviationsdened,for
k = 1, 3
, asσ
Mk
C
∗
= [
P
I⊆V
Pr[I](
|C
I
Mk
|)
2
− (
P
I⊆V
Pr[I]
|C
I
Mk
|)
2
]
1/2
and ompared theexpe tationswith the orresponding deviations. We also have done soforσ
opt
and
E(G
opt
)
. The relative results are presented in table 3. As one an see entries of this table show important dispersions of the fun tionals around the orresponding mathemati al expe tations.
7 Con lusions
We have rstgiven aformaldenition for probabilisti vertex over andnext we haveanalyzed three modi ation strategies and studied the omplexities for the orresponding probabilisti vertex overing problems. Moreover we have shown thattwo of the three variants studied have natural interpretations asweighted vertex over versions.
The framework of our paper diersfor the one [1, 2, 3,4, 5 ,6, 7, 8℄where besides the fa t thatalltheproblems onsidered(even minimum spanningtreeand shortestpath)aredened in ompletegraphs, asinglemodi ationstrategy onsistingin droppingabsentverti es out ofthe apriori solutions (asour M1) isstudied.
Dealingwithourresults,furtherworkhastobemadeforstrategyM3inorderto(a) ompletely hara terize
E(G
M3
C
)
and(b)determineC
ˆ
3
andthe omplexityofits omputation. Wethink that if one an answer point (a), then one will be able with very little more work to answer also point(b).Referen es
[1℄ I.Averbakh,O.Berman,andD.Sim hi-Levi.Probabilisti apriorirouting-lo ationproblems. Naval Res. Logisti s, 41:973989,1994.
[2℄ D. J.Bertsimas. Onprobabilisti traveling salesman fa ility lo ationproblems. T ransporta-tion S i.,3:184191, 1989.
[3℄ D. J. Bertsimas. The probabilisti minimum spanning tree problem. Networks, 20:245275, 1990.
[4℄ D.J.Bertsimas,P.Jaillet,andA.Odoni.Apriorioptimization.Oper.Res.,38(6):10191033, 1990.
[5℄ P. Jaillet. Probabilisti traveling salesman problem. Te hni al Report 185, Operations Re-sear h Center, MIT,Cambridge Mass.,USA, 1985.
[6℄ P. Jaillet. Apriorisolutionofa traveling salesmanproblemin whi harandom subset ofthe ustomers arevisited. Oper. Res.,36:929936, 1988.
[7℄ P. Jaillet. Shortest path problemswith node failures. Networks, 22:589605,1992.
[8℄ P.JailletandA.Odoni.Theprobabilisti vehi leroutingproblem. InB.L.GoldenandA.A. Assad, editors, Vehi le routing: methods andstudies. NorthHolland, Amsterdam,1988. [9℄ C. MuratandV.T.Pas hos. Theprobabilisti longestpath problem. Networks, 33:207219,
Consider expression(10) andset
x = 1
− p
. Then, expression(10) be omes(B
− e)(ℓ, x, n) = x
n−ℓ
− x
n
− ℓ(1 − x)x
n−1
= x
n−ℓ
+ (ℓ
− 1)x
n
− ℓx
n−1
= x
n−ℓ
(1 + (ℓ
− 1)x
ℓ
− ℓx
ℓ−1
) 6 x
n−ℓ
− x
n−1
6
1
(A.1) Setf (x) = 1 + (ℓ
− 1)x
ℓ
− ℓx
ℓ−1
. Then,f
′
(x) = ℓ(ℓ
− 1)(x
ℓ−1
− x
ℓ−2
) < 0
. Consequently,f (x)
is de reasing inx
∈ [0, 1]
, thereforef (x) > f (1) = 0
and the lower bound of expression(11) is proved.Revisitexpression(A.1)andset
g(x) = x
n−ℓ
− x
n−1
in[0,1℄. Firstremarkthat
g(0) = g(1) =
0
. Moreover,g
′
(x) = (n
− ℓ)x
n−ℓ−1
− (n − 1)x
n−2
andg
′
(x) = 0
⇐⇒ x
0
=
n − ℓ
n
− 1
ℓ−1
1
.
So,
g(x)
in reases, with respe ttox
,in[0, x
0
)
whileitis de reasing inx
in(x
0
, 1]
. Hen e,(B
− e)(ℓ, x, n) 6 g(x) 6 g(x
0
) =
n − ℓ
n
− 1
n−ℓ
ℓ−1
−
n − ℓ
n
− 1
n−1
ℓ−1
and the upper boundof expression(11) isalso proved.
B Proof of expression (13)
Revisitexpression(12) and,aspreviously, set
x = 1
− p
. Then, expression(12)be omes(e
− b)(ℓ, x, n) = ℓ(1 − x)x
n−ℓ
− x
n−ℓ
+ x
n
= (ℓ
− 1)x
n−ℓ
− ℓx
n−ℓ+1
+ x
n
= x
n−ℓ
(x
ℓ
− ℓx + (ℓ − 1)) 6 (ℓ − 1)
x
n−ℓ
− x
n−ℓ+1
(B.1) Seth(x) = x
ℓ
− ℓx + (ℓ − 1)
. Then,h
′
(x) = ℓ(x
ℓ−1
− 1) < 0
, so,h(x)
isde reasing inx
∈ [0, 1]
and, onsequently,h(x) > h(1) = 0
, proving sothe lower bound ofexpression (13).Revisitexpression (B.1)andset
ϕ(x) = x
n−ℓ
− x
n−ℓ+1
. Then,ϕ
′
(x) = (n
− ℓ)x
n−ℓ−1
− (n −
ℓ + 1)x
n−ℓ
andϕ
′
(x) = 0
⇐⇒ x
0
=
n
− ℓ
n
− ℓ + 1
.
So,