Laboratoire d'Analyse et Modélisation de Systèmes pour
l'Aide à la Décision
CNRS UMR 7024
CAHIER DU LAMSADE
187
Janvier 2002
Polynomial approximation and graph coloring
Résumé i
Abstra t i
1 Introdu tion 1
I Graph- oloring in standard approximation 3 2 How an one always olor a graph with
∆(G)
olors? 42.1 Anon- onstru tive theorem andits ulterior onstru tive proof . . . 4
2.1.1 Coloring 3- onne tedgraphs . . . 4
2.1.2 Coloring onne ted graphs ontaining verti esof degreeless than
∆(G)
. 5 2.1.3 Theoverallalgorithm . . . 52.2 A
∆(G)/3
approximationratio for graph- oloring . . . 63 Towards ratios of
o(∆(G))
6 3.1 Asemidenite relaxationfor graph- oloring . . . 73.2 Therounding phase . . . 8
4 The greedy oloring-algorithm 9 4.1 Anex avations hemafor graph- oloring . . . 9
4.2 Instantiating INDEPENDENT_SETbythe greedy algorithm . . . 10
5 Improving the ratio for graph- oloring 11 5.1 Coloring
k
- olorable graphs . . . 115.2 Expanding k_COLOR to runfor all graphs . . . 11
5.3 Thewholeimprovement . . . 12
6 A better approximation ratiofor oloring 13 6.1 Improving GREEDY_Cin
k
- olorable graphs . . . 136.2 Modifying algorithmk_COLORINGto runon all graphs . . . 14
6.3 Thewholeimprovement . . . 14
7 A still better approximation ratio for graph- oloring 15 7.1 Findinglarge independent sets ... . . 15
7.1.1 ...Byremoving liques. . . 16
7.1.2 ...In
k
- olorable graphs . . . 177.2 Theoverall oloring-algorithm . . . 18
8 A further improvement of the performan e guarantee for the approximation of graph- oloring 19 9 Inapproximability results 21 9.1 Negative results via graph-theoreti gap te hniques . . . 21
ber of unused olors 23 10A few words about the worst-value solution for graph- oloring 23 11A dierential-approximation algorithm based upon ane-equivalen e 24 12Frommat hing to 3 - independent sets 25 13Coloring graphs via set- overing 25 14Coloring, set- overing and semi-lo al optimization 26 15Unifying the abovealgorithms 28
15.1 Re overing algorithmD_COLOR4 . . . 29
15.2 Re overing algorithmD_COLOR3 . . . 29
15.3 Re overing algorithmD_COLOR2 . . . 30
15.4 Re overing algorithmD_COLOR1 . . . 30
16Using
∆
_COLOR to olor sparse graphs 30 17Negative results 30 18Final remarks 31 Referen es 33 A Proof of the results of se tion 2 36 A.1 Proofof lemma2 . . . 36A.2 Proofof lemma3 . . . 36
A.3 Proofof theorem2 . . . 36
A.3.1
∆(G[A]) < 3
. . . 36A.3.2
|A| < ∆(G) + 1
. . . 36A.3.3
|A| = ∆(G) + 1
. . . 36A.3.4
G[A]
is 3- onne ted . . . 36A.3.5
G[A]
is bi onne ted . . . 36Résumé
Considérons un graphe
G = (V, E)
d'ordren
. Le problème de la oloration minimum onsisteàattribuerd'unnombreminimumde ouleursauxsommetsdeV
defaçon à eque deuxsommetsadja entsnesoientpas oloriésave lamême ouleur.Ceproblèmeestparmi lespremiers démontrés intrinsèquement di iles et, par onséquent, il est très improbable qu'ilpuisse êtrerésolu optimalementparunalgorithme polynomial. Dans et arti le,nous faisons un tour d'horizon des prin ipaux algorithmes d'approximation ( eux pour lesquels des rapports d'approximation théoriques ont été étudiés) pour le problème de oloration etdis utonsleurs rapports d'approximation etleur omplexité. Enn,nous proposons une améliorationdurapportd'approximationpour eproblème.Mots- lé: graphe, oloration, omplexité, NP- omplet,algorithmed'approximation.
Polynomial approximation and graph- oloring
Abstra t
Consider a graph
G = (V, E)
ofordern
. Inminimumgraph- oloringproblemwetryto olorV
withasfew olorsaspossiblesothatnotwoadja entverti esre eivethesame olor. Thisproblemis amongtherstonesprovedtobeintra tableandhen e,itisveryunlikely that anoptimal polynomial-time algorithm ould everbedevisedfor it. Inthis paper,we surveythemain polynomial timeapproximationalgorithms(theones forwhi htheoreti al approximabilityboundshavebeenstudied)fortheminimumgraph- oloringandwedis uss their approximation performan e and their omplexity. Finally, we further improve the approximationratioforgraph- oloring.Consider a graph
G = (V, E)
of ordern
. In minimum graph- oloring problem (C), we wish to olorV
with as few olors as possible so that no two adja ent verti es re eive the same olor. Thisproblem was shownto be NP-hard in Karp'soriginal paper ([40℄),and remains NP- omplete even restri ted to graphsof onstant (independent onn
) hromati number at least 3 (formoreinformationsaboutthe omplexityofnumerousrestri tions,orgeneralizationsofC,the interestedreaderisalsorefereedto [35 ℄). The hromati number ofagraph,denotedbyχ(G)
,is the smallestnumber of olorsthat an feasibly olor itsverti es. AgraphG
is alledk
- olorable if its verti es an be legally olored byk
olors, in other words if its hromati number is at mostk
; itwill be alledk
- hromati ifk
is its hromati number.Sin e adja ent verti es are forbidden to be olored with the same olor, a feasible solution of C an be seen as a partition of
V
into vertex-sets su h that, for ea h one of these sets, no two of its verti es aremutually adja ent. Su h sets areusually alled independent sets. So, the optimal solutionof Cisa minimum- ardinality partition intoindependentsets.Another ombinatorialquantitydenedondis retestru turesthatwillbeusefulinthesequel is the one of set- overing. Given a family
S
of sets drawn from a ground setC
(satisfying∪
Si∈S
S
i
= C
),a set- overing isa familyS
′
⊆ S
su h that
∪
Si∈S
′
S
i
= C
.Both quantities, independent set andset- overing, give rise to two well-known NP-hard op-timization problems, namely the one of nding the maximum- ardinality independent set of a graph, denoted by
α(G)
, and the one of nding the minimum- ardinality set- overing of a set system. Inwhat follows wewill denote byIS the formerandbySCthe latter.Given the omputationalhardness of C, if onewishes to devise fastalgorithms, then he/she hastodeveloppolynomialtimemethodsproviding,forallinstan es,feasiblesolutions,thevalues of whi hare as lose as possible to the valueof the optimal ones. Thisiswhat in the literature is ommonly alled polynomial time approximation algorithms (PTAA).
The goodness of a PTAA A is measured by the so- alled approximation ratio. Two types of ratios have been used until now: the one that, in some way, ompares the value
A(I)
of the approximate solution forI
, provided by A, with the valueOPT(I)
of the optimal one, and the one omparingA(I)
not only withOPT(I)
but also withWORST(I)
, the value of the worst feasible solution ofI
. Ea h one of these measures draws its proper working-framework into whi h every NP-hard problem an be analyzed and implies its proper approximation results. This is, as we shall see in the sequel, parti ularly true for C (in general, given an NP-hard problem, approximation results an be very dierent whenusing the one working-framework or the other). For reasonsofsimpli ity, wewill speakaboutρ
-framework to denotetheformerand aboutδ
-framework to denotethe latter. Informally:•
in theρ
-framework, the omparison betweenA(I)
andOPT(I)
is usually performed by answering the following question: what is the ratio betweenA(I)
andOPT(I)
?; this question indu es the so- alled approximationratio;•
in theδ
-framework, the omparison betweenA(I)
,OPT(I)
andWORST(I)
is performed by answering the question: what is the leastǫ
for whi h A is to an extent of(1 − ǫ)
like the worst algorithm, and to an extent ofǫ
like the ideal one?; this question indu es the dierential-approximation ratio.LetA be aPTAA for an NP-hardproblem
Π
, letI
be the set ofinstan es ofΠ
, letA(I)
be the value of the solution provided by A on an instan eI ∈ I
,OPT(I)
be the value of the optimal solution forI
andWORST(I)
be the value of the worst solution ofI
. Moreover, letfor axed onstantM
,I
M
be the subset ofI
dened asI
M
= {I ∈ I : OPT(I) > M }
, and letσ(I)
be•
Approximation ratioρ
:the approximation ratio
ρ
A
(I)
of the algorithm A on an instan eI ∈ I
is dened asρ
A
(I) =
A(I)
OPT(I)
the approximation ratio
ρ
A
of A forΠ
isdened asρ
A
=
sup {r : ρ
A
(I) > r, ∀I ∈ I} Π
a maximizationprobleminf {r : ρ
A
(I) < r, ∀I ∈ I} Π
a minimizationproblem the asymptoti approximationratioρ
∞
A
of A forΠ
isdened asρ
∞
A
=
sup {r : ∃M, ρ
A
(I) > r, ∀I ∈ I
M
} Π
a maximization probleminf {r : ∃M, ρ
A
(I) 6 r, ∀I ∈ I
M
} Π
a minimization problem the approximation ratioρ
Π
forΠ
isdened asρ
Π
=
max {ρ
A
:
Aa PTAAforΠ} Π
a maximization problemmin {ρ
A
:
Aa PTAAforΠ} Π
a minimizationproblem•
Dierential-approximation ratioδ
:the dierential-approximation ratio
δ
A
(I)
of the algorithm A on an instan eI ∈ I
is dened asδ
A
(I) =
WORST(I) − A(I)
WORST(I) − OPT(I)
the dierential-approximation ratio
δ
A
of A forΠ
isdened asδ
A
= sup {r : δ
A
(I) > r, I ∈ I}
the asymptoti dierential-approximation ratio
δ
∞
A
of A forΠ
is dened asδ
∞
A
= lim
k→∞
sup
I
σ(I)>k
WORST(I) − A(I)
WORST(I) − OPT(I)
the dierential approximation ratio
δ
Π
forΠ
isdened asδ
Π
= max {ρ
A
:
Aa PTAA forΠ}
Let us note that the best expe ted ase for the behavior of an approximation algorithm is that it ould guarantee an approximation ratio tending to 1. More formally, the ideal would be that we had a PTAA A re eiving as inputs an instan e
I
of an NP-hard problem and a xed onstantǫ
and guaranteeing approximation ratio (resp., dierential-approximation ratio)1 + ǫ
or1 − ǫ
depending on if the problem at hand is a minimization or a maximization one (resp.,1 − ǫ
, for theδ
-framework), for everyǫ > 0
. In this ase, we have, in fa t, a se-quen e(A
ǫ
)
ǫ>0
of approximation algorithms having the desired approximation properties just des ribed. Su hasequen eis alled apolynomial timeapproximation s hema (PTAS)or,in theδ
-framework, dierential PTAS (DPTAS). Moreover, we an further lassifysu h s hemata fol-lowingthetime- omplexityofalgorithmA
ǫ
. WesospeakaboutPTAS(DPTAS)ifits omplexity isO(f (|I|)
Θ(1/ǫ)
)
is
O(g(1/ǫ)O(|I|
k
))
, where
|I|
isthe size ofI
,f
andg
polynomials not depending on|I|
andk
a xed onstant (not depending neither on|I|
, nor onǫ
). A PTAS, or FPTAS (resp., DPTAS, or FDPTAS) is alled asymptoti if, for every instan eI
, it guarantees approximation ratio1 + ǫ + c/OPT(I)
or1 − ǫ − c/OPT(I)
,depending onif the problem athand is aminimization or amaximizationone (resp.,1 − ǫ − c/(WORST(I) − OPT(I))
, fortheδ
-framework),for some xed onstantc
and for everyǫ > 0
.The approximation of C by e ient algorithms is a entral problem in omplexity theory. In
ρ
-framework, people knewquite early that a polynomial time approximation s hema annot exist for it, sin e alower bound 2 was proved for the ratio of every PTAA supposed solving C; this result, due to Gareyand Johnson, has been published in 1976. But even if the resear hers onje turedthatapproximationratiosgreater than2wereequally unlikely,su haresultwasnot formallyprodu edupto the early90'swhentheimpossibilityofapproximatingCbya onstant ratio approximation algorithm is proved byLund and Yannakakis. This result, motivates from then on, many resear hers for sear hing either for approximation algorithms with improved approximation ratios, or to strengthen the existing inapproximability results. On the other hand, inδ
-framework, C is better-approximable than in theρ
-one, sin e onstant dierential-approximationratio PTAAsexist sin e1994.This paper, even if it surveys both positive and negative approximation results 1
about C in both the approximation frameworks, is rather oriented towards the positive ones. For this reason,positiveresultsarepresented and ommentedindetailandtheunderlyingalgorithmsare spe iedinakindofpseudo-PASCAL,whiletheinapproximabilityonesaresimplymentioned. Beforepresentingpositiveresults,wegivetheintuition(orkey-idea)behindthem. Theobje tive of this survey is double: to present already known results and to introdu e some new ones. In order to be as short and easy to read as possible, only new results are proved, while already existingresults arestatedwithoutproofs. Also,inse tion3a oloringmethodisdis ussedusing a randomized algorithm that has been derandomized later by rather ompli ated te hniques. Sin e both random algorithm itself and its derandomization are long, and on the other hand, paper does not dealwith probabilisti methods,only the underlying idea is dis ussed,while we omitto spe ify the overallalgorithm.
Let us onsider asimple undire ted onne ted graph
G = (V, E)
of ordern
. Sometimes,we will denote by|G|
the order ofG
and byV (G)
its vertex-set. Given a setV
′
of verti es of a graph
G
,wewilldenotebyG[V
′
] = (V
′
, E(G[V
′
]))
the partialsubgraphof
G
indu edbyV
′
. For everyvertex
v ∈ V
,wedenotebyΓ
G
(v)
thesetofneighborsofv
(neighborhoodofv
)andbyd
◦
(v)
the quantity
|Γ
G
(v)|
, this quantity is lassi ally alled the degree ofv
;∆(G) = max
v∈V
{d
◦
(v)}
is the maximum degree of
G
. The notion of neighborhood an be extended to apply to a setV
′
⊆ V
;wedenotebyΓ
G
(V
′
)
thesetofneighborsoftheverti esof
V
′
,i.e.,
Γ
G
(V
′
) = ∪
v∈V
′
Γ
G
(v)
. Whenevernoambiguity ano urweusenotationΓ(v)
insteadofΓ
G
(v)
. Givenverti esv
andu
, we denotebyd(v, u)
their distan e,i.e., thelength (number ofedges) oftheshortest elementary path linkingv
andu
(sin eG
is undire tedd(v, u) = d(u, v)
). ByK
t
, we denote a omplete graphont
verti es. Alllogarithmsofthepaperaretothebase2unlessotherwisenoted. Finally, the following well-knownexpression ([10 ℄)linksχ(G)
andα(G)
and willbe usedlater:α(G)χ(G) > |G|
(1)1
For anNP-hard minimization (resp.,maximization) problem
Π
, an approximationresultis alledpositive if, givenan approximationalgorithmA, it provides an upper(resp., lower) bound onthe approximationratio guaranteedby AwhenusedtosolveΠ
;itis allednegative orinapproximability resultwhenitprovidesalower (resp.,upper)boundfortheratioofanalgorithmorofawhole lassofapproximationalgorithmsforΠ
Graph- oloring in standard approximation 2 How an one always olor a graph with
∆(G)
olors?We rst re all some standard graph-theoreti on epts used in what follows in this se tion (for more details one an be referred to [10 , 15℄). Consider a onne ted graph
G = (V, E)
. A subsetA ⊆ V
is alled arti ulation set ifG[V \ A]
is not onne ted; an arti ulation set of size 1 will be alled arti ulation point. AgraphG
isk
- onne ted i|G| > k + 1
andi itdoesnot ontain arti ulation set of size less thank
. As a onsequen e,G
is bi onne ted i|G| > 3
and it does not admitarti ulation points(i.e., for everypair(x, y) ∈ V × V
, there exist twovertex-disjoint elementary paths betweenx
andy
). A blo inG
is a setA ⊆ V
su h thatG[A]
is onne ted, without arti ulation points, and maximal (i.e., for anyv ∈ V \ A
,G[A ∪ {v}]
is either not onne ted, orit has at least an arti ulationpoint). Thenif|A| > 2
,G[A]
is bi onne ted, while, if|A| = 2
, it is an isthmus, i.e., an edge whose removal in reases the number of the onne ted omponentsof thegraph. Obviously, the setofblo sofG
oversV
. A bloB
is alled extremal ifit ontains avertexz
su hthatforanyotherbloB
′
,either
B
andB
′
arevertex-disjoint, or
z
istheir only ommonvertex.2.1 A non- onstru tive theorem and its ulterior onstru tive proof
Theorem 1 below (known also as Brook's theorem and non- onstru tively proved in [17 ℄) was onstru tively proved by Lovász in [42 ℄ in a rather ondensed way. We give in this se tion a personal interpretationof this result. The proofs aregiven in the appendix.
Theorem 1. ([17 ℄)If
G
is onne ted with∆(G) > 3
and ifit ontains no subgraphK
∆(G)+1
, then it is∆(G)
- olorable.Consider a graph
G
, and order its verti es, sayx
1
, x
2
, . . . , x
n
. One an olor them one-by-one in the following way: olorx
1
with1; then, olorx
2
with1, ifx
1
x
2
∈ E
/
, with2 otherwise and ontinue oloringea hvertexwiththesmallest olorit anbeassignedatthatstage. Denotethis algorithm by LEGAL_C and suppose it is alled withinputsG
andx
1
, x
2
, . . . , x
n
. The following easy lemmaholds for LEGAL_C.Lemma 1. LEGAL_C olors the verti es of any graph
G
with at most∆(G) + 1
olors, in polynomial time.Proof. Sin e
∆(G)
isthe maximumgraph-degree,∆(G) + 1
olorsaresu ient tolegally olor anyvertexofG
andall its neighbors.The key-ideaofthe algorithmof[42℄ isto onstru tgood orderingsofthe verti es of
G
, so thatLEGAL_C olorsthem withnomore∆(G)
olors.2.1.1 Coloring 3- onne ted graphs
Consideragraph
G = (V, E)
su hthat: itis onne ted,∆(G) > 3
,itdoesnot ontainaK
∆(G)+1
anditadmitsatleastapairofverti es,the removalofwhi hdoesnotdis onne tG
. Notethata 3- onne tedgraph verifyingthe onditions of theorem1 an besu h agraph. Also onsider the followingalgorithm3C_COLOR.We onsiderin whatfollowstwotypesofexe utionfor3C_COLOR. The rst one is parameterized by a graphG
, when it is exe uted starting from line (1). The se ond one is parameterized by a graphG
and two verti esv
andu
, when we suppose that lines(1)and(2)areomittedanditsexe utionstartsfromline(3). Thisisthe aseofexe ution of 3C_COLORin the bodyofalgorithm∆
_COLORin paragraph 2.1.3.(1) let
{1, . . . , ∆(G)}
be the set of olors to be used; (2) hoose verti es v and u su h thatd(v, u) = 2
and
G
′
← G[V \ {v, u}]
remains onne ted; (3) let w be a vertex adja ent to both v and u; (4) arrange the verti es of
G
′
in a sequen e
(w = x
1
, x
2
, . . . , x
n−2
)
su h that ea h vertexx
i
, i > 2
is adja ent to a vertexx
j
, j 6 i
;*this an be done breadth-first-sear h (BFS,[1℄) starting from w* (5) olor v and u with the olor 1;
(6) OUTPUT
{1} ∪ LEGAL
_C(G
′
, x
n−2
, x
n−3
, . . . , x
1
)
; END. *3C_COLOR*Lemma 2. Let
G
be a 3- onne ted graph not ontainingaK
∆(G)+1
. Then 3C_COLOR omputes a legal∆(G)
- oloring forG
in polynomial time.2.1.2 Coloring onne ted graphs ontaining verti es of degree less than
∆(G)
We now give another algorithm assigning at most∆(G)
olors to the verti es of a onne ted graph withat least one vertexof degree< ∆(G)
.BEGIN *1V_COLOR*
(1) let
{1, . . . , ∆(G)}
be the set of olors to be used; (2) hoose a vertex v su h thatd
◦
(v) < ∆(G)
;
(3) arrange the verti es of
G
in a sequen e(v = x
1
, x
2
, . . . , x
n
)
su h that ea h vertexx
i
, i > 2
is adja ent to a vertexx
j
, j 6 i
;*this an be done by BFS starting from v* (4) OUTPUT LEGAL_C(
G, x
n
, x
n−1
, . . . , x
1
);END. *1V_COLOR*
Lemma 3. Let
G
be a bi onne ted graph, not ontaining aK
∆(G)+1
, with at least one vertex of degree< ∆(G)
. Then 3C_COLOR omputes a legal∆(G)
- oloring forG
inpolynomial time. 2.1.3 The overall algorithmGiven a oloring,wewill all olor-permutation theex hangeofa olor
i
toj
andofj
toi
onall verti es olored withthese olors (whenj
has not been used,then olor-permutation betweeni
andj
be omesinusingj
instead ofi
). Wearewell-preparednowto spe ify analgorithmlegally oloringthe verti esof anygraphG
byusingat most∆(G)
olors.BEGIN *
∆
_COLOR*(1) ompute all the blo s of G; (2) FORevery blo A of G DO (3) IF
∆(G[A]) < 3
THEN(4) order arbitrarily the verti es of A, say
x
1
, x
2
, x
3
; (5) LEGAL_C(G[A], x
1
, x
2
, x
3
);(6) GOTO (2); (7) FI
(8) CASE |A| DO
(9)
|A| < ∆(G) + 1
: olor A with olors 1,...,∆(G)
and GOTO (2); (10)|A| = ∆(G) + 1
: 1V_COLOR(G[A℄) and GOTO (2);(13) IFG[A℄ is bi onne ted THEN
(14) hoose arbitrarily a vertex x with
d
◦
(x) > 3
; (15) IF
G[A \ {x}]
is bi onne ted THEN(16) let
y
∈ A
be su h that d(x,y) = 2; (17) 3C_COLOR(G[A℄,x,y) and GOTO (2); (18) FI(19) IF
G[A \ {x}]
is not bi onne ted THEN (20) letB
1
andB
2
two extremal blo s of G; (21) letz
1
∈ B
1
andz
2
∈ B
2
adja ent to x; (22) 3C_COLOR(G[A℄,z
1
, z
2
);(23) FI (24) FI (25) OD
(26) OD
(27) order the blo s in su h a way that ea h blo has at most a vertex in ommon with the blo s pre eding it in the ordering;
(29) OUTPUT the union of the olors used by permuting them if ne essary; END. *
∆
_COLOR*Theorem 2. Let
G
bea onne tedgraphwith∆(G) > 3
andnot ontainingsubgraphK
∆(G)+1
. Then, algorithm∆
_COLORlegally olorsthe verti esofG
with atleast∆(G)
olors,inpolynomial time.2.2 A
∆(G)/3
approximation ratio for graph- oloringSuppose now that 1_COLOR is an algorithmre eiving an input-graph
G
and de iding ifG
is an independent setand, if yes, oloring its verti es with one olor. Obviously, this an be done in polynomial time. Also,let2_COLOR be anotheralgorithmde iding ifG
isbipartite. This an be done by starting with two olors and by oloring a vertex with one of them and its neighbors withtheotherone. Ifattheend alltheverti esofG
arelegally olored,then itisbipartite and, moreover, a2- oloring is dis overed. Finally, onsiderthe following PTAA for C.BEGIN *APPROX_COLOR* 1_COLOR(G); 2_COLOR(G);
∆
_COLOR(G);OUTPUT the smallest of the legal olorings produ ed; END. *APPROX_COLOR*
Theorem 3.
ρ
APPROX
_
COLOR
6
∆(G)/3
. 3 Towards ratios ofo(∆(G))
Awell-knownmethodfor solvingnumerousNP-hardproblemsapproximately usesthe following s hema, denoted bySCHEMA in the sequel:
[Phase1℄ optimallysolve the linear-programmingrelaxationof theproblem (this an bedone in polynomial time([5 ℄));
[Phase2℄ round this solution to a feasible (but approximate) solution for the original (integer programming)problem.
the so- alled semidenite programming relaxations ([3 , 30 ℄). As opposed to the formers, the latters oer rounding te hniques leading to feasible integer solutions thatare guaranteed to be withinaspe iedfra tion tothe optimalones. Semideniteprogrammingin omputing approx-imatesolutionsfor ombinatorialproblems isoriginallyusedbyGoemans andWilliamson([29 ℄) formaximum utandmaximum2-satisabilityproblems. Basedupontheworkof[29℄,Kargeret al.([39 ℄)devisearandomizedapproximationalgorithmfor
C
a hievingapproximationratioo(∆)
. Theirmethod an be shortlydes ribed as follows.3.1 A semidenite relaxationfor graph- oloring
Consider a
k
- olorable graph. Instead of assigning integers (or olors) to the verti es of the graph, a unitve tor~v
i
∈ R
n
isassigned to anyvertex
v
i
∈ V
,i = 1, . . . , n
. In order to apture thefa tthatadja ent verti esareassignedwithdierent olors,theve tors assignedto adja ent verti es have to be dierent in some spe i (but natural) way. This requirement leads to the ve tork
- oloring.Given a graph
G = (V, E)
of ordern
and a realk > 1
, a ve tork
- oloring ofG
is an assignmentofunitve tors~v
i
∈ R
n
toanyvertex
v
i
∈ V
,su hthatforanytwoadja entverti esv
i
andv
j
the dot produ t of~v
i
and~v
j
satiseshv
i
, v
j
i 6 −1/(k − 1)
(in other words, the angle between the ve tors orresponding toadja ent verti esmustbe su iently large).The so-dened ve tor
k
- oloring is seen in [39 ℄ as a kind of relaxation for C that plays the role that a hypotheti al fra tionalk
- oloring would play if one used a onventional linear-programmingrelaxation for the problem. Thisis justied bythe following lemma.Lemma 4. ([39℄)
•
Everyk
- olorable graph has a ve tork
- oloring.•
Moreover, for all positiveintegersk
andn
withk 6 n + 1
,there existk
unit ve torsofR
n
su h that the dot produ tof any distin t pair is
−1/(k − 1)
.Remarkthatthe
k
ve tors ofthe se onditemoflemma4fulll thespe i ation oftheve tork
- oloring(theemphasized propositionjustabove). Furthermore,one an immediatelybije tively mapthese ve tors tok
distin t olorsin order to produ e ak
- oloring of agraphG
ofordern
.Following [Phase1℄ of SCHEMA, one has now to determine a ve tor
k
- oloring ofG
. This an be done using the following auxiliary problem. Given a graphG = (V, E)
of ordern
, a matrixk
- oloring ofG
is ann × n
symmetri positive semidenite matrixM
withm
ii
= 1
andm
ij
6
−1/(k − 1)
forv
i
v
j
∈ E
. The key-point of the relationship between matrix and ve tor olorings ofa graph isgiven by the following lemma.Lemma 5. ([39℄)
•
A graph has a ve tork
- oloring ifand only if it has a matrixk
- oloring.•
If a graph has a matrixk
- oloring, then a ve tor(k + ǫ)
- oloring an be onstru ted from this matrix oloring in time polynomial inn
andinlog(ǫ
−1
)
.
•
Conversely, if a graphG
has a ve tork
- oloring, then a matrix(k + ǫ)
- oloring ofG
an be onstru ted in time polynomial inn
and inlog(ǫ
−1
)
.
We have supposedat the beginning of the urrent paragraph that
G
isk
- olorable. Therefore, by the se ond item of lemma 4, there exists a ve tork
- oloring forG
and, by the rst itemof lemma 5, there exists a matrix
k
- oloring ofG
. This matrix oloring an be onstru ted by solving the following semideniteoptimization problem:SDP =
min σ
M = {m
ij
}
positive semidenitem
ij
6
σ
forv
i
v
j
∈ E
m
ij
= m
ji
m
ii
= 1
Remarkthat,bydenitionofthematrix oloringgiven justabovelemma5,the solutionofSDP spe ies the entries of a matrix
k
- oloringM
. Sin eG
has a ve tor (and, by the rst item of lemma 5, a matrix) oloring, there exists a solution to SDP withσ
∗
= −1/(k − 1)
. If one uses a linear-programmingmethod, she/he is ableto determine a feasible solution ofSDP with
σ 6 −1(k + ǫ − 1)
, for some arefully hosenǫ ∈ R
. This solution, by the denition of the matrix oloring, spe ies a matrix(k + ǫ)
- oloring ofG
. This, by the se ond item of lemma 5, an produ e a ve tor(k + 2ǫ)
- oloring ofG
. In order to simplify presentation, the errorǫ
an be ignored sin e it an be made very small as to be irrelevant to the analysis in [39 ℄. So, the following propositionholds.Proposition 1. ([39 ℄)Givena
k
- olorable graphG
ave tork
- oloringofG
anbe determined in polynomial time.3.2 The rounding phase
On e[Phase1℄a omplished, onehastopro ess[Phase2℄thatimpliestheroundingoftherelaxed solutiontoafeasibleintegerone. Theoriginalroundingte hniqueproposed, alledsemi- oloring in [39℄, is randomized and produ es an assignment of olors with relatively few identi ally olored adja ent verti es. This semi- oloring is then transformed into a legal graph- oloring. Moreformally, a
k
-semi- oloring ofa graphG
isan assignment ofk
olorsto atleast halfof its verti es su h that no two adja ent verti es re eive the same olor. Dealing withsemi- olorings, the following holds.Lemma 6. ([39 ℄) If an algorithm SEMICOLOR an
k
p
-semi- olor any subgraph of orderp
of a graphG
in randomized polynomial time (wherek
p
in reases withp
), then SEMICOLOR an polynomially olor the wholegraphG
withO(k
n
log n)
olors.In other words, a semi- oloring of
G
an be transformed into a legal oloring by losing only a logarithmi fa tor with respe t to the olors used for the semi- oloring. The randomized algorithm transforming ve tor olorings into semi- olorings is not given here. The interested reader an be referred in [39 ℄. Inany ase, the following an be shown.Proposition 2 . ([39℄) For every integer fun tion
k = k(n)
, a ve tork
- olorable graphG
an be semi- olored with atmostO(∆(G)
1−(2/k)
plog ∆(G))
olors in randomized polynomial time. Combination of lemma 6 withthe rst item of lemma 4 and with propositions 1 and 2 implies the following.
Proposition 3. ([39 ℄) Any
k
- olorable graphG
of ordern
an be olored in randomized polynomial time with atmostO(∆(G)
1−(2/k)
plog ∆(G) log n)
olors.
Two last points remain to be settled: (i) an the randomized method proposed be transformed into a deterministi one, and (ii) how one an make this method to run for any graph (re all thatuntil now the graph issupposed
k
- olorable2 )? 2
nomial derandomization ofthe algorithms in [39 ℄ a hieving the same approximation guarantees thanthe original(random)ones. Ontheotherhand,forpoint(ii),the followingpro edure,that will be usedlater also, an be used. Sin e the omplexity of the overall
k
- oloring algorithmis polynomial ink
, one an run it for allk ∈ {2, . . . n}
in this order (multiplying so its worst- ase omplexitybyn
), and stop it for the rstk
for whi h a feasible solution is produ ed. Asit is pointed out in [51 ℄, this thought pro ess an be implemented by divide-and- onquer in su h a waythatthe whole omplexityofthederived algorithmwillbemultiplied notbyk
but bylog k
. So, in the light of the above settlements, proposition 3 holds for any graph and the algorithm laimed runsin deterministi polynomial time.Finally,runbothalgorithm
∆
_COLORandthe s hemadealtin the urrent se tion (parameter-ized by the algorithms of [39 ℄, derandomized bythe method of [44 ℄, and extended so as it runs for any graph bythe remark settling point (ii)above),and take nallythe bestof the solutions produ ed. Then, the following on ludingtheorem holds.Theorem 4. C an be approximately solved in polynomial time and in any graph
G
within approximation ratioρ
C
6
min
(
∆(G)
χ(G)
, O
∆(G)
1−
2
k
plog ∆(G) log n
χ(G)
!)
.
4 The greedy oloring-algorithm
4.1 An ex avation s hema for graph- oloring
When one triesto approximately solve C, the rst algorithm omingin mind isthe greedy one des ribedbythefollowingex avations hema exe utedwithparameters
G
andanIS-algorithm INDEPENDENT_SET.BEGIN *EXCAVATION* REPEAT
S
← INDEPENDENT
_SET(G)
;olor the verti es of S with the same not already used olor; remove S from G;
UNTIL G be omes empty;
OUTPUT X the set of used olors; END. *EXCAVATION*
AlgorithmEXCAVATIONhastwointerestingproperties,expressedbyitems1and2ofproposition4 below. Item1isproved in[31 ℄,but hasbeenimpli itlyusedin [11 ,37 ,51℄andexpli itlyin[31 ℄. Item 2 is proved in [36 ℄ for the ase where
S
is a maximumindependent set and is generalized in [2 ℄ forthe ase whereS
is anyindependent set. Here itismainly usedin se tion 8.Proposition 4.
1. Anyiterativeappli ation ofalgorithm INDEPENDENT_SET that omputes an independent set of size
ι
k
(G) = Θ(n
1/t
)
,
t > 1
,in ak
- olorable graphG
of ordern
,produ es a oloringX
verifying|X| 6 2n/ι
k
(G)
.2.
ρ
EXCAVATION
6
ln n/ρ
INDEPENDENT
_SET
An interesting polynomial instantiation of the s hema des ribed above, where the independent setisfoundviaagreedyalgorithm,hasbeen studiedbyJohnsonin [37 ℄. Wegiveinwhatfollows an outlinefor both the greedy IS-algorithm andthe greedy C-algorithmthat ensues.
BEGIN *GREEDY_IS*
S
← ∅
; REPEATv
← argmin
v
i∈V
{d
◦
(v
i
)}
;S
← S ∪ {v}
;V
← V \ ({v} ∪ Γ(v))
;G
← G[V]
; UNTILV
= ∅
; OUTPUT S; END. *GREEDY_IS* BEGIN *GREEDY_C* REPEATS
← GREEDY
_IS(G)
;olor the verti es of S with the same non already used olor;
V
← V \ S
;G
← G[V]
; UNTILV
= ∅
;OUTPUT the set
X
J
of used olors; END. *GREEDY_C*The omplexity of algorithm GREEDY_IS is
O(|E|)
and will be alled at mostn
times within the REPEATloop of algorithmGREEDY_C,ea h su h all stri tly de reasingV
. Hen e, the overall worst- ase omplexityof GREEDY_CisO(n|E|)
.Let us denote by
V
(ℓ)
the vertex-set of the urrent, surviving, graph at the beginning of the
ℓ
thiterationofGREEDY_C.Thekey-pointforthestudyofitsapproximationperforman eisthe following. Ifagraphisk
- olorable,then,atea hiterationℓ
of GREEDY_C,d
◦
(v
j
) 6 |V | − ⌈|V |/k⌉
, wherev
j
is the minimum- urrent-degree vertex hosen by GREEDY_IS. Therefore, the verti es oloredduringℓ
thiteration(i.e., the onstru tedmaximalindependent setduringthisiteration) will be ofsize atleastlog
k
|V |
. Thisleads tothe following lemma,provedbyWigderson in [51 ℄. Lemma 7. ([51℄) Algorithm GREEDY_C olors anyk
- olorable graphG
with|X
J
| 6 3n/ log
k
n
olors.
Bylemma7,theapproximationratioofalgorithmGREEDY_Cina
χ(G)
- hromati graphbe omes:ρ
GREEDY
_C
6
|X
J
|
χ(G)
6
1
χ(G)
3n log χ(G)
log n
6
n
log n
(2)leading so to the following on ludingtheorem. Theorem 5. ([37℄)
ρ
GREEDY
_
C
6
n/log n
.TheimprovementoftheapproximationratioofChasremainedopenfor9yearsuntil1982when Wigdersonhasshownin[51 ℄howtoobtainbetterperforman eguarantees. Hismethod,outlined in what follows, isbaseduponthe following observations:
1. the neighborhood of any vertexin a
k
- olorable graph is(k − 1)
- olorable; 2. 2- oloring ispolynomial (see se tion 2).These observations (the se ond being a termination ondition) lead, as we will see, to a ni e re ursive approximationstrategy for C.
5.1 Coloring
k
- olorable graphsWe rst present the following pro edure, alled withparameters
k
andG
andi
(wherei
means thatG
will be olored with olorsi, i + 1, . . .
), that olors ak
- olorable graph with relatively few olors. Set,fork = 2, 3, . . .
,f
k
(n) = n
{1−(1/(k−1))}
. BEGIN *k_COLOR*
(1) CASEk DO
(2) k = 2: OUTPUT 2_COLOR(G);
*verti es of G will be olored with olors i and i+1* (3)
k > log|G|
: olor V(G) with a distin t olor per vertex;*verti es of G are assigned olors i, i+1,...,i+|G|-1* (4)
k 6 log|G|
: WHILE∆(G) > ⌈f
k
(|G|)⌉
DO(5) let v be su h that
d
◦
(v) = ∆(G)
;
(6)
H
← G[Γ
G
(v)]
;(7)
{i, i + 1, . . . , i + j − 1} ← k
_COLOR(k − 1, H, i)
; (8) olor v with i+j;(9)
i
← i + j
; (10)G
← G[V \ (Γ
G
(v) ∪ {v})]
; (11) OD (12){i, i + 1, . . . , } ← ∆
_COLOR(G)
; (13) OUTPUTX
k
_C
the set of used olors; (14) OD
END. *k_COLOR*
In the initial graph, algorithm is alled as k_COLOR(k,G,1). Line (12) of k_COLOR will be exe uted on graphs with
∆(G) < ⌈f
k
(|G|)⌉
, using so less than⌈f
k
(|G|)⌉
unused olors. The algorithm alled in line (2) is the one de iding if a graph is bipartite, and if yes, omputing a 2- oloring of itsverti es.Lemma 8. ([51 ℄) k_COLOR assigns, in
O(k(n + |E|))
, the verti es of anyk
- olorable graph with at most2k⌈f
k
(n)⌉ = 2k⌈n
(1−(1/(k−1)))
⌉
olors. 5.2 Expanding k_COLOR to run for all graphs
Aswehave already mentioned at the end of paragraph 3.2, one an expand algorithmk_COLOR (destinated to olor
k
- olorable graphs) to run for any graph. Re all that she/he only has to run itfor allk ∈ {2, . . . n}
in this order, andstop itfor the rstk
for whi h pro edure k_COLOR produ esa feasiblesolution. This an beimplemented as follows.(1) FOR
p
← 1
TO⌈logn⌉
DO (2)X
← k
_COLOR(2
p
, G, 1)
;
(3) IF X is feasible THEN
p
0
← p
and GOTO (5) FI; (4) OD(5) applybinary sear h in
{2
p0−1
+ 1, . . . 2
p0
}
to find the least
k
← k
0
for whi h k_COLOR produ es a legal oloring;(6) OUTPUT
X
E
← k
_COLOR(k
0
, G, 1)
; END. *Ek_COLOR*Sin e, givena graph
G
:(i) theexe ution ofline(2)ofalgorithmEk_COLORprodu esalegal oloringforall
k > χ(G)
, (ii)G
is alwaysχ(G)
- olorable,and(iii)
k
0
isthe smallestk
for whi h the exe ution ofline (5) willprodu e a feasible oloring, the following lemmaholds.Lemma 9. ([51℄)
k
0
6
χ(G)
.Combining lemmas 8 and 9, we get the following on luding theorem for the approximation performan e of algorithmEk_COLOR.
Theorem 6. ([51 ℄) Ek_COLOR olors, in
O((n + |E|)χ(G) log χ(G))
, the verti es ofG
with at most2χ(G)⌈n
(1−(1/(χ(G)−1)))
⌉
olors. 5.3 The whole improvement Bytheorem6,
ρ
Ek
_COLOR
6
2⌈n
(1−(1/(χ(G)−1)))
⌉
andfun tionf (x) = 2⌈n
(1−(1/(x−1)))
⌉
isin reasing inx
. On the other hand, by lemma 7and expression(2),ρ
GREEDY
_
C
6
3n log χ(G)/(χ(G) log n)
and fun tiong(x) = 3n log x/(x log n)
is de reasing inx
. Let us ombine the two algorithms to produ e the following nalalgorithm.BEGIN *W_COLOR*
X
E
← Ek
_COLOR(G)
;X
J
← GREEDY
_C(G)
;OUTPUT
X
W
= argmin{|X
E
|, |X
J
|}
; END. *W_COLOR*Of oursealgorithmW_COLOR omposedbypolynomial omponent-algorithmsisalsopolynomial. A little algebra shows that the interse tion point of the urves
f (x)
andg(x)
is in the neighborhood ofx = ⌈n log log n/2 log n⌉
. Algorithm Ek_COLOR is superior to GREEDY_C forχ(G) 6 log n/ log log n
while, forχ(G) > log n/ log log n
GREEDY_C is more performant than Ek_COLOR.Forχ(G) = log n/ log log n
, bothρ
Ek
_
COLOR
and
ρ
GREEDY
_C
areof
O(n log
2
log n/log
2
n)
and the following theorem on ludesthe se tion. Theorem 7. ([51℄)
ρ
W
_
COLOR
6
3n log
2
log n/ log
2
n
Seven years later, Bergerand Rompel, using thought pro esses quite lose to the onesadopted in [51 ℄ (i.e., oloring rst
k
- olorable graphs, extending the result to work for every graph, the hromati number of whi h belongs to a ertain interval of values, and next omposing the algorithm produ ed with another well-working in graphs with hromati -number values out of the interval onsidered), perform, in [11 ℄, a further notable improvement of the approximation ratio ofC.The key-observation in [11 ℄ is a kind of renement of the respe tive observation of [37 ℄. Re all that, as we have seen in se tion 4, Johnson observed that if
G
isk
- olorable, then there exists an independent set of size at least⌈|V |/k⌉
and, onsequently, any nodev
in this set hasd
◦
(v) 6 |V | − ⌈|V |/k⌉
; this observation enabled him, using algorithm GREEDY_IS, to nd an independent set
S
ofO(log
k
|V |)
small-degree verti es to whi h gave the same olor. The renement lying at the heart of the method proposed in [11 ℄ is that anysubsetS
′
of
S
veries|Γ(S
′
)| 6 |V | − ⌈|V |/k⌉
. Thisallows them to somewhatmodify algorithmGREEDY_IS to hoose at ea hstepa set(insteadof one)ofO(log
k
|V |)
small-degreeverti esthatareindependent and havesmall neighborhood. Theyareso abletolegally olorO[(log
k
|V |)
2
]
verti eswiththe same olor.LetusnowoutlinehowBergerandRompelprodu e feasible oloringsfor
k
- olorable graphs. We onsider, without lossof generality, thatthe olors aredrawnfromthe set{1, 2, . . .}
. 6.1 Improving GREEDY_C ink
- olorable graphsConsiderthefollowingalgorithmk_COLORINGparameterizedbyaninteger
k
,axedα > 0
anda graphG
. Therst onditionin line(10)re eivesTRUEif the setS
isanindependent set, FALSE otherwise.BEGIN *k_COLORING* (1) IF
k
> min{n
1/2
, n
α
}
THEN OUTPUT GREEDY_C(G) FI (2)
m
← ⌊αlog
k
n⌋
; (3)c
← 1
; (4)UNCOLORED
← V
; (5) WHILE|UNCOLORED| > km
DO (6)U
← UNCOLORED
; (7) WHILE|U| > km
DO (8)H
← G[U]
;(9) FOR
S
⊆ H
su h that |S| = m DO(10) IF S AND
|Γ
H
(S)| 6 |U| − (|U|/k)
THEN GOTO (13) FI (11) OD(12) OUTPUT G not k- olorable; (13) olor S with olor ;
(14)
U
← U \ [S ∪ Γ
H
(S)]
;(15)
UNCOLORED
← UNCOLORED \ S
; (16) OD(17)
c
← c + 1
; (18) OD(19) olorthe UNCOLORED-verti es with olors , ..., +|UNCOLORED|-1; *assigns an unused olor per UNCOLORED-vertex*
partition the verti es of U into
ℓ = ⌊|U|/km⌋
setsB
i
,i
= 1, . . . , ℓ
, su h thatB
j
,j
= 1, . . . , ℓ − 1
, ontains km elements andB
ℓ
ontains the remaining ones (km 6
|B
ℓ
| < 2km
); FORj
← 1
T0ℓ
DO FORS
⊆ B
j
su h that|S| = m
DO GOTO line (10);Bythe pigeonholeprin iple,atleast oneofthesets
B
i
will ontainasetS
ofsizem
andthisset an be found by exhaustively sear hingtheC
kα log
k
n
α log
k
n
= O(n)
subsets ofea h
B
i
. Consequently, implementation of line(9) of algorithmk_COLORING an be done in polynomial time.Theorem 8. ([11 ℄) For any
α > 0
, algorithm k_COLORING olors, inO[n
3+3α
/(k log
k
n)]
, the verti es of anyk
- olorable graph with2n/(α log
2
k
n) + O(n/ log
3
k
n)
olors. 6.2 Modifying algorithm k_COLORING torun on all graphsThe following algorithm (parameterized by
α
andG
) modies, in the spirit of [51 ℄, algorithm k_COLORINGto workon anygraph. Algorithms1_COLOR and2_COLOR alled byk_COLORINGare as in se tion2.BEGIN *Ek_COLORING*
(1) IF 1_COLOR(G) feasible THEN OUTPUT 1_COLOR(G) FI (2) IF 2_COLOR(G) feasible THEN OUTPUT 2_COLOR(G) FI (3) FOR
p
← 1
TO⌈logn⌉
DO(4)
X
← k
_COLORING(2
p
, α, G)
;
(5) IF X is feasible THEN
p
0
← p
and GOTO (7) FI; (6) OD(7) applybinary sear h in
{2
p
0−1
+ 1, . . . 2
p
0
}
to find the least
k
← k
0
for whi h k_COLORING produ es a legal oloring;(8) OUTPUT
X
E
← k
_COLORING(k
0
, α, G)
; END. *Ek_COLORING*The omplexity of algorithm Ek_COLORING is
O(n
4+3α
/ log n)
(it alls at most
n − 2
times k_COLORING). Moreover, with the same arguments as the ones for lemma 9,k
0
6
χ(G)
. So, the oloringprodu ed satises|X
E
| 6
2n
α log
2
χ(G)
n
+ O
n
log
3
χ(G)
n
!
(3) 6.3 The whole improvementRe all that
ρ
Ek
_
COLOR
6
2⌈n
(1−(1/(χ(G)−1)))
⌉
and this bound is in reasing in
χ(G)
. On the other hand, by theorem 8 and expression (3),ρ
Ek
_
COLORING
6
2n/(αχ(G) log
2
χ(G)
n) +
o(n/ log
2
χ(G)
n)
and this last bound is de reasing inχ(G)
. Forχ(G) = O(⌈log n/ log log n⌉)
, both bounds are at mostO(n log
3
log n/ log
3
n)
. The following algorithm, BR_COLOR, ombines algorithms Ek_COLOR and Ek_COLORING into an algorithmi s hema for C. Its running time is of
O(max{n
(4+(21/η))
/ log n, (n + |E|)χ(G) log χ(G)})
([11 ℄). BEGIN *BR_COLOR* fix a large
η > 0
; OUTPUTX
E
← Ek
_COLORING(7/η, G)
; OUTPUTX
′
E
← Ek
_COLOR(G)
; OUTPUTargmin{|X
E
|, |X
′
E
|}
; END. *BR_COLOR*Theorem 9. ([11℄)
ρ
BR
_
COLOR
6
ηn log
3
log n/ log
3
n
.
Let us observe that algorithmBR_COLOR has a fairly serious drawba k sin e its exe ution time requirements(beingpolynomialin
n
for xedη
)isexponential inη
. Thismeans thatthebetter the approximation ratioa hieved,the higher its exe ution time.7 A still better approximation ratio for graph- oloring
In1993,afurtherimprovementof oloring'sapproximationratiohasbeenpublishedby Halldórs-son in [31 ℄. Thespirit ofthis work isquitesimilar to the previousones(ex ept the one of[42 ℄), i.e., one olors a graph by ex avating independent sets, but the used IS-algorithms are quite dierent from the greedy one used until then. It is well-known to people working on the de-sign of approximation algorithms that very frequently the e ien y of an algorithm strongly depends on the value of the optimal solution of an instan e. Some algorithms work well for smalloptimal values,whilesome otheronesworkbetteron instan es withlargeoptimal values. The key idea of [31 ℄ is to ombine into an ex avation s hema two IS-algorithms, one of them (OPT-CLIQUE_REMOVAL) behaving e iently in graphs with small hromati numbers, while the otherone(COLOR_IS) behavingwellingraphswithlarge hromati numbers. Simultaneous run-ning, atea hiteration ofthe ex avations hema,of both algorithms and oloring largestamong the independent sets omputed with an unused olor leads to an improved ratio for any value of
χ(G)
.7.1 Finding large independent sets ...
Werstbrieydes ribethe twoIS-algorithms,OPT-CLIQUE_REMOVALandCOLOR_IS,usedin[31 ℄ for ex avating independent sets.
The ideaofusing theformerisbased upon thefa tthatindependent sets in graphswithout liques or with small liques are larger than independent sets in general graphs. Furthermore satisfa torily large independent sets (a hieving improved approximation ratios) an be polyno-mially found there (see for example [49 , 50 ℄, where the author deals with triangle-free graphs). Hen e,givenagraph
G
,one anredu eitbyremoving liquesofa ertainsizeℓ
. Inthesurviving graph (that isℓ
- lique-free), in parti ular if its size remains large (in some sense), she/he an apply some e ient algorithm omputing a large independent set. Ex avation of su h large independent setsispossibleas longasthe initial andthe onse utive(surviving) graphs ontain few disjoint liques. Of ourse, if large independent sets are ex avated, by item 1 of proposi-tion4,thegraph anbe oloredwithrelativelyfew olors. Ontheotherhand,iftheinitialgraph ontains manydisjoint liques, thenα(G)
must be small and, byexpression (1),χ(G)
must be large. Inboth ases the approximationratio for C an so be improved.AlgorithmCOLOR_ISoriginallyoperatesin
k
- olorable graphs. Informally,itre ursivelynds an independent setS
of a ertain size and takes the union ofS
with the resultof its re ursive running on the graphG[S ∩ Γ(S)]
. This is done as long as the order of the surviving graph ex eedsaxedthresholdt
. Assoonas the orderofthe survivinggraph be omessmallerthant
, OPT-CLIQUE_REMOVAL is alled. The nal independent set is the union of the independent sets re ursively omputed by COLOR_IS together with the one omputed by OPT-CLIQUE_REMOVAL. RunningCOLOR_ISforanyvalueofk
andretainingthebestresult,one anobtainanindependent set at least as good as the result of the exe utionCOLOR
_IS(χ, G)
, and then using item 1 of proposition4, one an hope to ompute a small oloringforG
.We rst present an intermediate algorithm, originally devised in [16 ℄, and used, dire tly or undire tly, by both OPT-CLIQUE_REMOVAL and, COLOR_IS. It is based upon the following result due to Ramsey.
Theorem 10. For any pair
(s, t)
of integers, there is an integern
for whi h every graph of ordern
ontains either a liqueK
s
,or an independent setS
of sizet
.Ifwe denoteby
R(s, t)
theminimalvalueofn
forwhi h theabovetheoremholds, then the best knownupperboundforR(s, t)
,dueto ErdösandSzekeres([25℄),isR(s, t) 6 C
s−t+2
t−1
. LetussetC
t−1
s−t+2
= r(s, t)
for all positive integerss
andt
(by onvention, if one ofs
andt
isnegative or zero, thenC
s−t+2
t−1
= 1
),and lett
s
(n) = min{t : r(s, t) > n}
.Algorithm RAMSEY (parameterized by a graph
G
and an integers
) developed in [16 ℄ and stronglyinspiredbytheproofoftheupperboundforR(s, t)
([25℄)nds,intimeO(|E|)
, eithera liqueK
of orders
, or anindependent setS
of sizet
. Moreover,r(s, t) > n
ands.t > c(log n)
2
, for some onstant
c
.BEGIN *RAMSEY*
S
← ∅
;K
← ∅
; WHILE|V(G)| > 1
DO hoosev
∈ V(G)
;t
← t
s
(n)
; IF|Γ(v)| > r(s − 1, t)
THENK
← K ∪ {v}
;G
← G[Γ(v)]
;s
← s − 1
; ELSES
← S ∪ {v}
;G
← G[V \ ({v} ∪ Γ(v))]
; FI OD OUTPUTK
← K ∪ V(G)
; OUTPUTS
← S ∪ V(G)
; END. *RAMSEY*It is next ombined witha greedy me hanism of lique removaland the following IS-algorithm (parameterized byagraph
G
andan integers
) isderivedin [16 ℄.BEGIN *CLIQUE_REMOVAL*
(S, K) ←
RAMSEY(G, s)
; WHILE|K| > s
DOG
← G[V \ K]
;(S, K) ←
RAMSEY(G, s)
; OD OUTPUT S; END. *CLIQUE_REMOVAL*Inorderthatthe nal oloring-algorithmisthe bestpossible,oneneedsto ndthe valueof
s
for whi htheapproximationratioofalgorithmCLIQUE_REMOVAListhebestpossible(inotherwords, the independentset omputed isthe largestpossible). Then,one anfollowthethought pro ess originally proposed in [51 ℄, i.e., one an use binary sear h to dene a goods
. This leads to the following nal version of CLIQUE_REMOVAL, alled OPT-CLIQUE_REMOVAL and parameterizedby a graph
G
, of omplexityO(|E|n log n)
, where we denote byρ
i
the approximation ratio of CLIQUE_REMOVAL(G,i). BEGIN *OPT-CLIQUE_REMOVAL* guess an s;S
← CLIQUE
_REMOVAL(G, s)
; IFρ
s
> (n/s
2
)
(1/s)
THEN RETURN S FIREPEATrun CLIQUE_REMOVAL(G,p) with
p
= 2
ℓ
,
ℓ = ⌈logk⌉, . . .
; UNTIL the firstℓ ← ℓ
0
for whi hρ
p
> (n/p
2
)
(1/p)
; REPEAT apply binary sear h in
{2
ℓ0−1
+ 1, . . . 2
ℓ0
}
; UNTIL the least
p
← p
0
for whi hρ
p
O
> (n/p
2
0
)
(1/p0)
; OUTPUTS
← CLIQUE
_REMOVAL(G, p
0
)
;END. *OPT-CLIQUE_REMOVAL*
Lemma 10. ([31 ℄) Running algorithm OPT-CLIQUE_REMOVAL on a graph
G
of ordern
withα(G) > tn
,t > 1/ log n
, returns an independent setof size atleaste
−1
n
t
/t
. 7.1.2 ...In
k
- olorable graphsLetus now onsiderthe se ond IS-algorithm(parameterized byaninteger
k
and byagraphG
) presented in [31 ℄ and running ink
- olorable graphs. As for algorithmk_COLORING in se tion 6, the rst onditionin line (3)re eivesTRUEif the setS
is anindependent set, FALSE otherwise. BEGIN *k-COLOR_IS*(1) IF
|V| = 1
THEN OUTPUT V FI; (2) FORS
⊆ V
su h that|S| = log
k
n
DO (3) IF S THEN(4) IF
|G[V \ (S ∪ Γ(S))]| > nlogn/(2kloglogn)
(5) THENOUTPUT
S
∪
k-COLOR_IS(k, G[V \ (S ∪ Γ(S))])
; (6) ELSEI
←
OPT-CLIQUE_REMOVAL(G[V \ (S ∪ Γ(S))]) ∪ S
;(7) IF
|I| > (logn)
3
/(6loglogn)
THEN OUTPUT
S
∪ I
FI; (8) FI(9) FI (10) OD
(11) OUTPUT G not k- olorable; END. *k-COLOR_IS*
Notethatremark1holdsalsoforalgorithmk-COLOR_IS.Consequently, the exe utionoftheFOR loop ofline (2) isperformedin polynomial time.
We an modify algorithm k-COLOR_IS to work for all graphs (i.e., for any
k
) produ ing so the following IS-algorithm (parameterized byG
) of omplexityO(n|E|χ(G))
. The IF- ondition in this algorithmre eives TRUEifS
k
isindependent.BEGIN *COLOR_IS*
FOR
k
← 1
TO n DOS
k
←
k-COLOR_IS(k, G)
; IFS
k
THEN storeS
k
FIOUTPUT the best among the
S
k
's stored; ODEND. *COLOR_IS*
Lemma 11. ([31 ℄) The appli ation of algorithm COLOR_IS in
G
produ es an independent setWe areready nowto outlinethe overallC-algorithmof [31 ℄,the worst- ase omplexityofwhi h is
O(n
2
|E|)
. BEGIN *H_COLOR* REPEATSCR
←
OPT-CLIQUE_REMOVAL(G)
;SCS
← COLOR
_IS(G)
;S
← argmax{|SCR|, |SCS|}
;olor the verti es of S with the same unused olor;
V
← V \ S
;G
← G[V \ S]
; UNTILV
= ∅
;OUTPUT the set of the olors used; END. *H_COLOR*
Suppose rst
χ(G) 6 log n/(2 log log n)
. Then, by lemma 10, algorithm OPT-CLIQUE_REMOVAL guarantees an independent setof size|SCR| > e
−1
n
1
χ(G)
1
χ(G)
= e
−1
χ(G)n
χ(G)
1
.
Byitem1ofproposition4,anex avations hemausingOPT-CLIQUE_REMOVAL(algorithmH_COLOR isinstantiation ofsu ha s hema) omputesa oloring ofsize
|X| = O
n
e
−1
χ(G)n
χ(G)
1
!
= O
n
1−
1
χ(G)
χ(G)
!
(4) On the other hand, ifχ(G) > log n/(2 log log n)
, then by lemma 11, algorithm COLOR_IS guar-anteesan independent set ofsize|SCS| >
log
2
n
log χ(G)
1
2 max
n
log
2χ(G) log log n
log n
, 1
o .
By item1of proposition4,algorithmH_COLOR (using COLOR_IS) willprodu e a oloringwith
|X| = O
n log χ(G) max
n
log
2χ(G) log log n
log n
, 1
o
log
2
n
(5)The orrespondingratiosaretheright-handsidesofexpressions(4)and(5)dividedby
χ(G)
,the formerbeing de reasing and the se ond in reasing inχ(G)
. Forχ(G) = log n/(2 log log n)
, the values of the two ratiosare both ofO(n log
2
log n/ log
3
n)
and the following theorem on ludes the se tion.
Theorem 11. ([31 ℄)
ρ
H
_
COLOR
= O(n log
2
log n/ log
3
n)
tion of graph- oloring
RevisitforawhilealgorithmOPT-CLIQUE_REMOVALofse tion7. In[23 ℄,thefollowingproposition isproved.
Proposition 5. ([23 ℄) Forevery fun tion
ℓ
su hthat,∀x > 0
,0 < ℓ(x) ≤ log log x
, thereexist onstantsκ
andK
su hthat algorithmOPT-CLIQUE_REMOVAL omputes,forevery graphG
of or-dern > κ
,anindependentsetS
su hthatifα(G) > ℓ(n)n log log n/ log n
,then|S| > K log
ℓ(n)
n
. In what follows, we denote by EXHAUST, an exhaustive-sear h algorithm for C. Without loss of generalitywe supposethatverti esare olored with
1, 2, . . .
Moreover, letK
andκ
be as in the quoted propositionabove.BEGIN *COLOR*
(1) IF
n 6
κ
THEN OUTPUT EXHAUST(G) FI (2)S
←
OPT-CLIQUE_REMOVAL(G); (3)i
← 1
; (4)^
X
← ∅
; (5)V(^
G) ← ∅
; (6) WHILE|S| > Klog
ℓ
n
DO(7) olor S with olor i; (8)
V(^
G) ← V(^
G) ∪ S
; (9)^
G
← G[V(^
G)]
; (10)^
X
← ^
X
∪ {i}
; (11)i
← i + 1
; (12)G
← G[V \ S]
; (13) IFG
= ∅
THEN OUTPUT^
X
FI (14)S
←
OPT-CLIQUE_REMOVAL(G); (15) OD (16)~
X
← ∆
_COLOR(G)
; (17) OUTPUTX
← ^
X
∪ ~
X
; END. *COLOR* Theorem 12.ρ
C
6
max
2n
k log
ℓ(n)−1
n
,
2∆(G)ℓ(n) log log n
log n
.
Proof. Obviously,ifline(1)ofalgorithmCOLORisexe uted, thenitreturnsaminimum oloring for
G
in polynomial time. The WHILE-loop of the algorithm above (lines (6) to (15)) is an appli ation of EXCAVATION(G,OPT-CLIQUE_REMOVAL). Observe also that, for everyiterationi
of the WHILE-loop,if we denotebyG
i
the graphG[V \ V ( ˆ
G)]
(G
1
= G
),thenρ
OPT-CLIQUE_REMOVAL(G
i
) >
K log
ℓ(|Gi|)
|G
i
|
|G
i
|
(6) Thenbyitem2 ofproposition 4and by expression(6):
ρ
EXCAVATION
ˆ
G
6
ln
G
ˆ
K log
ℓ
(|
G
ˆ
|)
|
G
ˆ
|
|
G
ˆ
|
=
log
G
ˆ
log eK log
ℓ
(|
G
ˆ
|)
|
G
ˆ
|
|
G
ˆ
|
6
G
ˆ
k log
ℓ
(|
G
ˆ
|)
−1
ˆ
G
(7)Set
˜
G = G[V \ V ( ˆ
G)]
(in other words,˜
G
is the subgraph ofG
input of algorithm∆
_COLOR in line(16)). Then,byproposition5andexpression(1),appli ationof∆
_COLORinG
˜
will ompute a setX
˜
of olorsverifyingρ
∆
_COLOR
˜
G
=
˜
X
χ
˜
G
6
∆
˜
G
log
|
G
˜
|
ℓ
(|
G
˜
|)
log log
|
G
˜
|
=
∆
˜
G
ℓ
˜
G
log log
˜
G
log
˜
G
(8) Thefollowing holds for the setX
of olors omputed byalgorithmCOLOR:|X| =
X
ˆ
+
X
˜
6
ρ
EXCAVATION
ˆ
G
χ
ˆ
G
+ ρ
∆
_COLOR
˜
G
χ
˜
G
6
max
n
ρ
EXCAVATION
ˆ
G
, ρ
∆
_COLOR
˜
G
o
χ
ˆ
G
+ χ
˜
G
(9) Obviously, bothχ( ˆ
G)
andχ( ˜
G)
are smaller thanχ(G)
. So, using expressions (7), (8 and (9), one getsρ
COLOR
=
|X|
χ(G)
6
max
2
G
ˆ
k log
ℓ
(|
G
ˆ
|)
−1
G
ˆ
,
2∆
˜
G
ℓ
G
ˆ
log log
G
˜
log
G
˜
6
max
2n
k log
ℓ(n)−1
n
,
2∆(G)ℓ(n) log log n
log n
(10) This ompletesthe proof ofthe theorem.As we have already seen, in terms of
n
the best known approximation ratio for C isO(n log
2
log n/ log
3
n)
(se tion 7) and the very tight analysisof [31 ℄ does not allow improve-ment of this ratio even in parti ular lasses of graphs. Let
ℓ > 5
and suppose rst that the maximum in expression (10) is realized by the termO(n/ log
ℓ−1
n)
. Then, for graphs with
∆(G) = O(n/ log
ℓ−2
n)
, CisapproximablewithinratioO(n/ log
ℓ−1
n)
andtheorem12improves the ratio of[31 ℄ byafa tor
Ω(log
2
log n log
ℓ−4
n)
.
Revisit nowthe ratioin theorem4and remark thatthe rst fun tion isde reasing in
χ(G)
, whilethe se ond one isin reasing forχ(G) 6 2 log ∆(G)
and de reasing forχ(G) > 2 log ∆(G)
.•
Ifχ(G) 6 2 log ∆(G)
, the ratio expression in theorem 4 attains its minimumvalue when the two terms are equal, in other words when(∆(G))
2/χ(G)
= Θ(plog ∆(G) log n)
, i.e., when
χ(G) = Θ(log ∆(G)/ log log n)
. Inthis ase, the minimumandthereforethe valueof the ratio guaranteed bythe expression in theorem4 be omeO(∆(G) log log n/ log ∆(G))
. On the other hand, by theorem 12, when∆(G) = Ω(n/(log
ℓ−2
n log log n))
, the ratio
O(∆(G) log log n/ log n)
is always a hieved independently on the values ofχ(G)
. Hen e, for∆(G) > n/ log
ℓ−2
n
and
χ(G) 6 2 log ∆(G)
,the resultof [39℄(se ond term ofthe ratio in theorem4) isimprovedbya fa torlog n/ log ∆(G)
.•
Ifχ(G) > 2 log ∆(G)
(and∆(G) > n/ log
ℓ−2
n
), then the ratio of theorem 4 is bounded above by
O((∆(G))
1−(2/ log ∆(G))
log n/plog ∆(G))
, while the ratio of COLOR remains bounded above by
O(∆(G) log log n/ log n)
. Therefore, in this ase also,algorithm COLOR dominates the one oftheorem4.Corollary 1.
•
Forℓ > 1
, C is approximable withinO(n/log
ℓ−1
n)
in graphs with maximum degree at most