Laboratoire d'Analyse et Modélisation de Systèmes pour
l'Aide à la Décision
CNRS UMR 7024
CAHIER DU LAMSADE
160
Janvier 1999
Approximation of weighted hereditary induced
subgraph maximization problems
Résumé ii
Abstra t ii
1 Preliminaries 1
2 The generi result 1 3 Important orollaries of Theorem 1 4 3.1 Maximum-weight independent setand maximum-weight lique . . . 4 3.2 Maximum-weight
ℓ
- olorable indu edsubgraph . . . 5 4 About ontinuous redu tions between weighted and unweighted versions ofre her he de sous-graphes induits pondérés
Résumé
Le but entral de et arti le est l'étude de l'approximation polynomiale de problèmes héréditaires de maximisation portant sur la re her he de sous-graphes induits pondérés. Aprèsavoirdonnéunrésultatgénérald'approximationpour ette lassedeproblèmes,nous montronsque elui- ia plusieurs orollairesimportants, parexemple, il améliored'un fa -teurd'
O(log log n)
lemeilleurrapportd'approximation onnupourlestablemaximum pon-déré.Mots- lé: problèmes ombinatoires, omplexité, algorithmepolynomiald'approximation, NP- omplétude,stable,problèmehéréditaire.
Approximation of weighted hereditary indu ed-subgraph maximization problems
Abstra t
Themainpurposeofthispaperistostudypolynomialapproximationofweighted hered-itaryindu edsubgraphproblems. Aftergivingageneralapproximationresultforthis lass ofproblems, weshowthat theresult hasmany important orollariessin e, forinstan e, it allowsimprovementofthebest-knownapproximationratioforweightedindependentsetby afa torof
O(log log n)
.Keywords: ombinatorial problems, omputational omplexity,polynomial-time approxi-mationalgorithm, NP- ompleteness,independentset, hereditaryproblem.
We onsider NP- omplete optimization graph-problems
Π
where the obje tive is to nd a maximum-order indu ed subgraph1
satisfying a non-trivial hereditary property
π
. LetG
be the lass of all the graphs; agraph-propertyπ
is a mapping fromG
to{0, 1}
, i.e., for aG ∈ G
,π(G) = 1
iG
satisesπ
andπ(G) = 0
, otherwise;π
ishereditary ifwhenever itissatised by agraph itisalso satised byevery one ofits indu ed subgraphs;it isnon-trivial ifitis truefor innitelymany graphsand false for innitelymany ones([2℄). The propertyπ
: isa lique, orπ
: isanindependent set aretypi alexamplesofsu hnon-trivial hereditaryproperties andthe maximum lique-problemor themaximum independent setarein their turntypi alexamplesof hereditaryindu edsubgraphproblems. Ageneralisation ofthe lassofproblemsjustintrodu ed isthe one where we onsiderthat positive weights areasso iatedwith the verti esof the input-graphs. GivenagraphG
,the obje tiveofaweightedindu ed subgraphproblemisto determine an indu edsubgraphG
∗
of
G
su h thatG
∗
satises
π
and,moreover, the sumof the weightsof the verti es ofG
∗
mustbethe largest possible.
Remark 1. By denitionof
π
(asa mapping fromthe lassof graphsto{0, 1}
), itis obvious thatπ
is ontext-free;inotherwords, ifagraphG
satisesπ
,thenG
satisesπ
intoeverygraph whoseG
is anindu ed subgraph.Given a polynomial-time approximation algorithm (PTAA) A that solves a maximization NP- omplete problem
Π
(we denote byWΠ
the weighted version ofΠ
), the quality of its approxi-mation behaviour is expressed by the ratioρ
of the value (size, or weight, or whatever) of the solution found by the algorithm to the optimal (obje tive) value ofΠ
; the smallest su h ratio over all graphs is the approximation ratioρ
A
of the algorithm; the best-known ratioρ
Π
over allPTAA'ssolvingΠ
istheapproximationratioforΠ
. Instan e-dependingapproximationratios for NP- omplete graph-problems are ommonlyexpressed bymeans of(one or more) the three standardgraph-parameters, namelythe ordern
ofthe inputgraph, themaximum vertex-degree and the average vertex-degree of the input graph. In what follows, we will denote byρ
Π
(n)
(resp.,ρ
WΠ
(n)
) the best-knownn
-depending approximation ratio forΠ
(resp., forWΠ
). Note thatsin e we dealwith maximization problems ratioρ
isnon-in reasing inn
.Given a vertex-weighted graph
G = (V, E, ~
w)
of ordern
with weight-ve torw
~
, we denote byw
i
the weight of avertexv
i
∈ V
, and byw
max
(G)
andw
min
(G)
the largest and the smallest vertex-weights, respe tively. We denote byβ
w
(G)
and byβ
′
w
(G)
the weight of a bestWΠ
-solution and the weight of a maximal2
Π
-solution ofG
, respe tively. For aV
′
⊆ V
, we denote byG[V
′
]
the subgraphof
G
indu ed byV
′
and by
n
′
its order. Given a subgraph
G[V
(k)
]
ofG
, we denote byS
∗
(G[V
(k)
])
a maximum-sizeΠ
-solution ofG[V
(k)
]
; also, given any
Π
-solutionS
′
of asubgraphG
′
ofG
, wedenote byw(S
′
)
the weight ofS
′
. 2 The generi resultThisse tionis mainly devotedto the proof ofthe following theorem.
Theorem 1. Consider a hereditary property
π
, an indu ed subgraph problemΠ
stated with respe t toπ
and the weighted versionWΠ
ofΠ
(we suppose that weights are positive). Let1
Inotherwords,afeasiblesolutionof
Π
isasubsetoftheverti esoftheinput-graph. 2Throughoutthepaperweusethenotationmaximalsettodenoteasetwhi hismaximalforthein lusion withrespe ttoagivenproperty
π
;inotherwords,asetS
ofitemsismaximalwithrespe ttoπ
if(a)S
fullsπ
and(b)ifweaddanyiteminS
,thentheresultingsetdoesnotfullπ
;asetismaximumwithrespe ttoπ
if itisthelargestbetweenthemaximalsets(withrespe ttoπ
).M > 2
be an integer andletV
(i)
= {v
j
∈ V : w
max
(G)/M
i
< w
j
≤ w
max
(G)/M
i−1
}
,i = 1, . . .
Finally, letx = sup{ℓ : β
w
(G[∪
1≤i≤ℓ
V
(i)
]) < β
w
(G)/2, β
w
(G[∪
1≤i≤ℓ+1
V
(i)
]) ≥ β
w
(G)/2}
. Then,ρ
WΠ
(n) ≥ max
M
x
2n
, min
M − 2
2M
2
x
ρ
Π
(n),
1
M
2
ρ
Π
(n)
.
Proof: For the sequel, we set
G
x
= G[∪
1≤i≤x
V
(i)
]
,
G
x+1
= G[∪
1≤i≤x+1
V
(i)
]
and
G
d
= G[V \
∪
1≤i≤x
V
(i)
]
(of ourse,β
w
(G
d
) ≥ β
w
(G)/2
).We rst remove verti es
v
k
su h that the graph({v
k
}, ∅)
does not satisfyπ
. Then, the following lemmaholds.Lemma 1. There exists a PTAA for
WΠ
a hieving approximation ratioM
x
/(2n)
. Proof: The algorithm laimed onsists of simplytaking
v
∗
= argmax
v
i
∈V
{w
i
}
asWΠ
-solution andthenaddingverti esinsu hawaythatthesolutionnallyobtainedremainsfeasible3 for
G
. Moreover, notethat, sin everti esv
k
su hthatthe graph({v
k
}, ∅)
doesnot satisfyπ
havebeen removed, the graph({v
∗
}, ∅)
is already a feasible solution. Consequently
β
′
w
(G) ≥ w
max
(G)
,β
w
(G) ≤ 2β
w
(G
d
) ≤ 2|S
∗
(G
d
)|w
max
(G
d
) ≤ 2|S
∗
(G
d
)|w
max
(G)/M
x
. Consequently,ρ(n) =
β
′
w
(G)
β
w
(G)
≥
M
x
2|S
∗
(G
d
)|
≥
M
x
2n
(1) q.e.d.Remark 2. For every
i ≥ 1
, the weight of anyΠ
-solutionS
(i)
of
G[V
(i)
]
lies in the interval
[|S
(i)
|w
max
(G)/M
i
, |S
(i)
|w
max
(G)/M
i−1
]
. It is at least|S
(i)
|w
min
(G[V
(i)
]) ≥ |S
(i)
|w
max
(G)/M
i
and atmost
|S
(i)
|w
max
(G[V
(i)
]) ≤ |S
(i)
|w
max
(G)/M
i−1
.For the rest of the proof, we distinguish two ases with respe t to
β
w
(G
x
)
, namelyβ
w
(G)/2 ≥
β
w
(G
x
) ≥ ((M − 2)/2M )β
w
(G)
andβ
w
(G
x
) ≤ ((M − 2)/2M )β
w
(G)
, respe tively.We rst onsider the ase
β
w
(G)/2 > β
w
(G
x
) ≥ ((M − 2)/2M )β
w
(G)
. Then the following lemmaholds.Lemma 2. Let
β
w
(G)/2 ≥ β
w
(G
x
) > ((M − 2)/2M )β
w
(G)
,p = argmax
1≤i≤x
{β
w
(G[V
(i)
])}
,|V
(p)
| = n
(p)
and
S
∗
(G[V
(p)
])
be themaximum-size
Π
-solutioninG[V
(p)
]
. IfthereexistsaPTAA for
Π
providing a maximal solutionS
(p)
su h that
|S
(p)
| ≥ ρ
Π
(n
(p)
)|S
∗
(G[V
(p)
])|
, then one an solveWΠ
inG
, in polynomialtime, withinratio((M − 2)/2xM
2
)ρ
Π
(n)
. Proof: Obviously,β
w
(G
x
) ≤ xβ
w
(G[V
(p)
])
andβ
w
(G
x
) ≤ x|S
∗
(G[V
(p)
])|w
max
(G)/M
p−1
(by Remark2). So,β
w
(G) ≤ (2M /(M − 2))x|S
∗
(G[V
(p)
])|(w
max
(G)/M
p−1
)
.On the other hand, appli ation of a PTAA guaranteeing approximation ratio
ρ
Π
(n) < 1
forΠ
inG[V
(p)
]
onstru ts a solutionS
(p)
ofWΠ
of weight at least|S
(p)
|w
min
(G[V
(p)
]) ≥
|S
(p)
|w
max
(G)/M
p
. Note that (byRemark 1)S
(p)
is
Π
-feasible forG
. Moreover, starting from this solution, one an greedily augment it in order to nally produ e a maximalWΠ
-solution forG
. This nalsolutionveriesβ
′
w
(G) ≥ |S
(p)
|w
max
(G)/M
p
. Combination ofthe above expressions forβ
′
w
(G)
andβ
w
(G)
yieldsρ(n) =
β
′
w
(G)
β
w
(G)
≥
M − 2
2xM
2
|S
(p)
|
|S
∗
(G[V
(p)
])|
!
≥
M − 2
2xM
2
ρ
Π
(n
(p)
) ≥
M − 2
2xM
2
ρ
Π
(n)
(2) and this on ludes the proof of the aseβ
w
(G)/2 > β
w
(G
x
) ≥ ((M − 2)/2M )β
w
(G)
and of the lemma.Lemma 3. Let
β
w
(G
x
) ≤ ((M − 2)/2M )β
w
(G)
andS
∗
(G[V
(x+1)
])
be a maximum-size
Π
-solution ofG[V
(x+1)
]
. If there exists a PTAA for
Π
providing a maximalΠ
-solutionS
(x+1)
of ardinality at least
ρ
Π
(n
(x+1)
)
-times the ardinality of
S
∗
(G[V
(x+1)
])
, then one an solve
WΠ
inG
, in polynomialtime, within ratio(1/M
2
)ρ
Π
(n)
.Proof: Notethat sin e
x
is the largestℓ
for whi hβ
w
(G[∪
1≤i≤ℓ
V
(i)
]) < β
w
(G)/2
, setV
(x+1)
is not empty. LetS
opt
(G
x+1
)
be an optimal WΠ
-solution inG
x+1
(i.e.,β
w
(G
x+1
) = w(S
opt
(G
x+1
))
); letS(G
x
) = S
opt
(G
x+1
) ∩ V (G
x
)
(whereV (G
x
)
denotes the vertex-set ofG
x
) andS(G[V
(x+1)
]) =
S
opt
(G
x+1
)∩V
(x+1)
(inotherwords,{S(G
x
), S(G[V
(x+1)
])}
isapartitionof
S
opt
(G
x+1
)
). Sin eπ
is hereditary, setsS(G
x
)
andS(G[V
(x+1)
])
, being subsets of
S
opt
(G
x+1
)
, also verifyπ
(and, onsequently theyarefeasibleWΠ
-solutionsforG
x
andG[V
(x+1)
]
,respe tively). Wethen have:
w(S(G
x
)) ≤ β
w
(G
x
)
w(S(G[V
(x+1)
])) ≤ β
w
(G[V
(x+1)
])
β
w
(G
x+1
) = w(S(G
x
)) + w(S(G[V
(x+1)
])) ≤ β
w
(G
x
) + β
w
(G[V
(x+1)
])
≤
M − 2
2M
β
w
(G) + β
w
(G[V
(x+1)
])
β
w
(G
x+1
) ≥
β
w
(G)
2
.
It follows from the above expressions that
β
w
(G[V
(x+1)
]) ≥ β
w
(G)/M
and this together with Remark2 yield, aftersome easy algebra,β
w
(G) ≤ |S
∗
(G[V
(x+1)
])|w
max
(G)/M
x−1
.Aspreviously,supposethataPTAAprovides,in polynomialtime,amaximalsolution
S
(x+1)
for
Π
inG[V
(x+1)
]
, the ardinalityofwhi hisatleast
ρ
Π
(n
(x+1)
)|S
∗
(G[V
(x+1)
])|
. Then, one an greedily augment this solution in order to produ e a maximal
Π
-solution forG
of total weightβ
′
w
(G) ≥ |S
(x+1)
|w
max
(G)/M
x+1
. Combination ofexpressions forβ
′
w
(G)
andβ
w
(G)
yieldsρ(n) =
β
′
w
(G)
β
w
(G)
≥
1
M
2
|S
(x+1)
|
|S
∗
(G[V
(x+1)
])|
!
≥
1
M
2
ρ
Π
(n
(x+1)
) ≥
1
M
2
ρ
Π
(n)
(3) and this on ludes the proofof the aseβ
w
(G
x
) ≤ ((M − 2)/2M )β
w
(G)
andof the lemma. Remark 3. For the ase wherex = 0
, i.e.,β
w
(G[V
(1)
]) ≥ β
w
(G)/2
, arguments similar to the ones of the proof of Lemma 3 lead to an approximation ratioρ
WΠ
(n) = β
′
w
(G)/β
w
(G) ≥
ρ
Π
(n)/2M
, better than the oneof expression(3).Consider now the following algorithm where we take up the ideas of Lemmata 1, 2 and 3 and where, for a graph
G
′
, we denote by
A(G
′
)
the solution-set provided by the exe ution of the
Π
-PTAA A onthe unweighted versionof agraphG
′
. BEGIN (*WA*) fix a onstantM
> 2
; partition V in setsV
(i)
← {v
k
: w
max
/M
i
< w
k
≤ w
max
/M
i
−1
}
;S
(0)
← {argmax
v
i
∈V
{w
i
}}
; OUTPUTargmax{w(S
(0)
), w(A(G[V
(i)
])), i = 1, . . .}
; END. (*WA*)Revisitexpressions (1), (2)and (3). It is easy to see that, sin e
M
is xed, worst ases for the respe tive ratiosarerepresented byexpressions (1)and (2);so,ρ
WΠ
(n) ≥ ρ
WA
(n) ≥ max
M
x
2n
, min
M − 2
2M
2
x
ρ
Π
(n),
1
M
2
ρ
Π
(n)
(4)areexponential in
n
.The proof of Theorem 1 an be seen as a redu tion transforming every PTAA A a hieving ratio
ρ
A
(n)
forΠ
into a PTAA for WΠ
a hieving, at worst, ratioe = Ω(ρ
A
(n)
/x)
. Quantity1/x
is usually alled expansion of the redu tion (see [7℄). By expression (1) and by the fa t that the approximation ratio of any PTAA for WΠ
mustbe lessthan 1 (WΠ
being a maximization problem),x = O(log
M
n)
(in fa t, as we will see in the next se tion, it an be mu h smaller depending on the form ofρ
Π
(n)
). But even su h an expansion is a meaningful improvement with respe t to expansions indu ed by redu tions between weighted no-weighted versions of problems whi h do not admit onstant-ratio PTAA's. For example, for the ase of IS, if we take into a ount thatnoPTAA an guaranteexed onstant ratio forit, redu tionsof [7℄have expansion1/n
2
.
3 Important orollaries of Theorem 1
3.1 Maximum-weight independent set and maximum-weight lique
Consider now one of the most lassi al NP- omplete problems, the maximum independent set. Given agraph
G = (V, E)
, anindependent setisasubsetV
′
⊆ V
su hthatwhenever
{v
i
, v
j
} ⊆
V
′
,v
i
v
j
∈ E
/
, and the maximum independent set problem (IS) is to nd an independent set of maximum size. As we have already noted in se tion 1, property is an independent set is hereditary (the subset of an independent set is an independent set); moreover, it is non-trivial (notallsetsV
′
⊆ V
areindependent). Averywell-knownandlargelystudiedgeneralizationofIS is its weighted version (denoted by WIS); here, we asso iate positive weights with the verti es of
G
andthe obje tive be omesto maximize the sum ofthe weights ofan independent set.In terms of
n
, the best known approximation ratio for IS is, to our knowledge, a hieved by the IS-PTAA of Boppana and Halldórsson ([1 ℄):ρ
IS
(n) = Θ(log
2
n/n)
. Embedding it in expression(2), we getρ(n) ≥
M − 2
2xM
2
Θ
log
2
n
n
!
= Θ
log
2
n
nx
!
.
(5)For instan e, if
x ≤ log log n
, expression (5) indu esρ(n) ≥ Θ(log
2
n/(n log log n))
, while, if
x ≥ log log n
, then expression (1) guaranteesρ(n) ≥ Θ((log n)
log M
/n)
; on the other hand, expression (3)always guarantees
ρ(n) ≥ Θ(log
2
n/n)
. So, byexpression (4), the approximation ratioguaranteedbyalgorithm WA(supposing
M > 8
) alledwhenusedtosolveWISis,atworst,ρ
WA
(n) ≥ Θ
log
2
n
n log log n
!
.
(6)Expression(6)and the dis ussionaboveintrodu e the following orollary. Corollary 1.
ρ
WIS
(n) = Ω(log
2
n/(n log log n))
.
Theaboveresultimprovesbyafa tor
O(log log n)
the best-knownapproximationratio(fun tion ofn
) for WIS(Θ(log
2
n/(n(log log n)
2
))
, dueto Halldórsson ([5℄)).
Finally,letusnotethatthesameapproximationresultholdsforanotherfamousNP- omplete problem, the maximum weighted lique(see [2℄for the statement ofthis problem).
3.2 Maximum-weight
ℓ
- olorable indu ed subgraphGiven a graph
G = (V, E)
and a positive onstantℓ
, the problem of the maximumℓ
- olorable indu ed subgraph (denotedbyCℓ
)istondamaximum-orderindu edsubgraphG
′
= G[V
′
]
ofG
(V
′
⊆ V
)su h thatG
′
is
ℓ
- olorable (i.e., thereexistsa oloring forG
′
of ardinality atmost
ℓ
). IntheweightedversionofCℓ
,denotedbyWCℓ
,positiveweightsareasso iatedwiththe verti es ofG
andtheobje tivebe omestondthemaximum-weightℓ
- olorableindu edsubgraph. On e more,propertyisℓ
- olorable ishereditary(ifthe verti es of agraphG
an be feasibly olored byat mostℓ
olors, then every subgraphofG
indu edbyasubset of itsverti es an be olored by atmostℓ
olors)and non-trivial (given anumberℓ
every graph isnotℓ
- olorable).The best-knownratiosfor C
ℓ
arethe onesof[4℄:Θ((log n)
2
/n)
(asfun tion of
n
) andof [3℄:min{κ/µ, Θ((log log ∆)/∆)}
, for every xedκ ∈ IN
+
, where
∆
,µ
are the maximum and the average degrees of the graph, respe tively (as degree-fun tion). For WCℓ
, no approximation resultis, to our knowledge, mentioned until now.Appli ation ofTheorem 1,usingthealgorithm of[4℄ asapproximationalgorithm for C
ℓ
(the algorithm A in algorithm WAof se tion2), leadsto the following orollary.Corollary 2.
ρ
WCℓ
= Ω(log
2
n/(n log log n))
.
4 About ontinuous redu tions between weighted and unweighted versions of independent set
In what follows, we use the term ontinuous redu tion, introdu ed by Simon ([7℄), to denote the approximation-ratio-preserving redu tions. Informally, given two maximization problems
Π
andΠ
′
, a redu tion
Π ∝ Π
′
is alled ontinuous with an expansion of
e
ifea h polynomial-timeρ
′
-approximation algorithm
A
′
for
Π
′
an be onverted into a polynomial-time
ρ
-approximation algorithmA
forΠ
su h thatρ ≥ eρ
′
. The orresponding notation is
Π
e
=⇒ Π
′
. It is very well-knownandlargelynoteduntilnowbymanyauthorsthatthe on eptof ontinuousredu tionsis strongly based upon the very restri tive hypothesisthat the approximation ratioρ
′
of
A
′
hasto beaxed onstant. Sin eISdoesnotadmit su halgorithms(in 1996ithasbeen provedthatit ishardtoapproximateISwithin
1/n
1−ǫ
,forany
ǫ > 0
,[6 ℄), ontinuousredu tionsproposeduntil nowbetweenISandWISfailtoprodu eWIS-ratiosasgoodastheonesknownforIS.Infa t,the bestknownratiosforISareΘ(log
2
n/n)
(intermsof
n
,[1℄)andmin{κ/µ, Θ((log log ∆)/∆)}
, for everyxedκ ∈ IN
+
, where
∆
,µ
arethemaximumandthe averagedegreesofthe graph, respe -tively([3 ℄). Ontheotherhand,thebest orrespondingratiosforWISareΘ(log
2
n/(n log log n))
(proved in this paper) and
3/(∆ + 2)
(see [5 ℄ for this last ratio), respe tively. We think that devising a ontinuous redu tion of onstant expansion between IS and WIS, or proving that su h a redu tiondoes not exist,is a very interesting open problem, the solution ofwhi h is far from beingobvious. In what follows, we give some easy ases of su h redu tions for restri tive WIS- lasses re ognizable in polynomial time. We denote byα
w
(G)
and byα
′
w
(G)
the weight of a maximum-weight independent set and the weight of an approximately maximum-weight independent set ofG
, respe tively.Theorem 2. Consider a xed onstant
ℓ
andgraphsG = (V, E)
of ordern
where atleast one of the following onditions isveried:1. there exist atmost
ℓ
distin t weights, 2.w
max
(G)/w
min
(G) ≤ ℓ
,4.
w
max
(G) ≥ nw
′
, where by
w
′
we denote the se ond largest vertex weight.
For all these graph-families, there exist ontinuous redu tions of onstant expansion between IS andWIS.Consequently,in allthe above families of graphs,WIS an be approximated in polyno-mial timewithin
ρ
WIS
= ρ
IS
≥ max
(
Θ
log
2
n
n
!
, min
κ
µ
, Θ
log log ∆
∆
)
for every xed
κ ∈ IN
+
.
Proof: We onsider the
ℓ
subgraphs,G
(1)
, . . . , G
(ℓ)
, of
G
, indu ed by the verti es of the same weight. Letp = argmax
1≤i≤ℓ
{α
w
(G
(i)
)}
andw
p
be the weight of the verti es ofG
(p)
; then,
α
w
(G) ≤ ℓα
w
(G
(p)
) = ℓw
p
|S
∗
(G
(p)
)|
. On the other hand, one an easily takeα
′
w
(G) =
max
1≤i≤ℓ
{w
i
|S
(i)
|}
, whereS
(i)
is the largest among the independent sets provided by the al-gorithms of [1℄ and of [3℄ for IS. It is easy to see that, in this ase,
ρ
WIS
= α
′
w
(G)/α
w
(G) ≥
w
p
|S
(p)
|/(ℓw
p
|S
∗
(G
(p)
)|) ≥ ρ
IS
/ℓ
and thatthis ratio veriesthe nal statement of the theorem, on ludingsothe proof ofitem 1.For item 2, it su es to remove weights from the verti es of graph and to solve IS in
G
. LetS(G)
be thelargestamong theindependent sets provided bythe algorithmsof[1 ℄and of[3℄ for ISandS
∗
(G)
bethe optimal one in
G
. Then, itiseasy to seethatρ
WIS
= α
′
w
(G)/α
w
(G) ≥
w
min
(G)|S(G)|/(w
max
(G)|S
∗
(G)|) ≥ ρ
IS
/ℓ
. On e more the nal statement of the theorem is veried and the proofof item 2is on luded.The proof of item 3 isstandard and already used by many authors (see for example [7℄ for the proof of the result WIS(
k
)1−ǫ
=⇒
IS, where WIS(k
) denotes the version ofWIS where vertex-weights arebounded above byn
k
,
k ≥ 1
; this proof unfortunately worksonly for xed- onstant ratios). Given aninstan eG
w
ofWIS,one an onstru t aninstan eG
of ISbyrepla ing every vertexv
i
of weightw
i
by an independent setW
i
of sizew
i
; next, if two verti esv
i
andv
j
are linked by an edge inG
w
, one draws a omplete bipartite graph between the two independent setsW
i
andW
j
ofsizesw
i
andw
j
,respe tively(inthe asewhereweightsarerationalswesimply repla e ea h weight by itself multiplied by the LCM of the weight-denominators). Now, note thatifanalgorithm deliversa solutionforG
(instan eof IS),one an dire tlyobtain asolution ofG
w
by simply repla ing the independent sets of the former by the orresponding verti es of the latter, this operation leading to a solutionforG
w
with the same valueasthe ardinality of the solution forG
. Denoting byn
w
and∆
w
the order and the maximum vertex-degree ofG
w
, respe tively, and byn
and∆
the orresponding parameters ofG
, we haven ≤ w
max
(G)n
w
,∆ ≤ w
max
(G)∆
w
. It is easy to see that following the dis ussion just above and thanks to the fa tthatthe integralweightsarexed onstants,applyingthe algorithmsof[1℄and[3 ℄inG
and retaining the best of the omputed solutions, a hieves for WISinG
w
the same orders ofratios asfor ISinG
, on ludingsothe proof ofitem 3.For the proof of item 4, let us denote by
G
m
the subgraph ofG
indu ed by the verti es of weightw
max
(G)
and byS(G
m
)
the bestof the solutions provided whenrunning the algorithms of [1 ℄ and [3 ℄ onG
m
. Obviouslyα
′
w
(G) ≥ α
′
w
(G
m
) ≥ w
max
(G)|S(G
m
)|
. On the other hand,α
w
(G) ≤ α
w
(G
m
)+(n−|S
∗
(G
m
)|)w
′
≤ w
max
(G)|S
∗
(G
m
)|+(n−|S
∗
(G
m
)|)w
max
(G)/n
. Divising, member-by-member, the above expressions forα
′
w
(G)
andα
w
(G)
, we getρ
WIS
=
α
′
w
(G)
α
w
(G)
≥
|S(G
m
)|
|S
∗
(G
m
)| +
n−|S
∗
(G
m
)|
n
≥
|S(G
m
)|
|S
∗
(G
m
)| + 1
.
ase,
α
′
w
(G)/α
w
(G) ≥ |S
∗
(G
m
)|/(|S
∗
(G
m
)| + 1) ≥ 1/2
(sin e|S
∗
(G
m
)| ≥ 1
). Consequently, on e more, the nalstatement of the theorem isveried and this on ludesthe proof of item 4 and ofthe theorem.Referen es
[1℄ B. B. Boppana and M. M. Halldórsson. Approximating maximum independent sets by ex- luding subgraphs. BIT,32(2):180196,1992.
[2℄ P. Cres enzi and V. Kann. A ompendium of NP optimization problems. Available on www_address: http://www.nada.kth.se/~viggo/problemlist/ ompendium.html,1995. [3℄ M. Demangeand V.Th. Pas hos. Improved approximations for maximum independent set
via approximation hains. Appl. Math.Lett., 10(3):105110,1997.
[4℄ M. M. Halldórsson. A still better performan e guarantee for approximate graph oloring. Inform. Pro ess. Lett.,45(1):1923, 1993.
[5℄ M. M. Halldórsson. Approximations via partitionning. JAIST Resear h Report IS-RR-95-0003F,Japan Advan ed Institute ofS ien e and Te hnology, Japan, 1995.
[6℄ J. Håstad. Clique is hard to approximate within
n
1−ǫ
. In Pro . FOCS'96, pages 627636, 1996.
[7℄ H. U. Simon. On approximate solutions for ombinatorial optimization problems. SIAM J. Dis . Math.,3(2):294310, 1990.