Parallel and Online

Magns M Halldrsson

Science Institute

University of Iceland



IS Reykjavik Iceland

Abstract

We discover a surprising connection b etween graph coloring in two orthogonal paradigms

parallel and online computing We present a randomized online coloring algorithm with

p

a p erformance ratio of O n log n an improvement of log n factor over the previous b est

known algorithm of Vishwanathan Also from the same principles we construct a parallel

coloring algorithm with the same p erformance ratio for the rst such result As a byproduct

we obtain a parallel approximation for the indep endent set problem

Introduction

At the heart of most partition and allo cation problems is a graph coloring question The graph

coloring problem involves assigning values or colors to the vertices of a graph so that adjacent

vertices are assigned distinct colors with the ob jective of minimizing the number of colors used

We seek graph coloring algorithms that are b oth eective and ecient Eciency means

among other things using computation time b ounded by a p olynomial in the size of the input

Eectiveness means that the quality of the coloring is never much worse than the optimal

coloring it is measured in terms of the performance ratio of the algorithm dened as the ratio

of the number of colors used to the minimum number of colors needed maximized over all

inputs In this pap er we present a reasonably eective and ecient graph coloring algorithm

that yields b oth parallel and online algorithms

A parallel algorithm is said to b e ecient if it uses time p olylogarithmic and pro cessors

p olynomial in the size of the input The coloring problem is of central imp ortance in parallel and

distributed computation as a means of scheduling nonconicting concurrent tasks Previous

parallel algorithm have attained the b ound of Bro okss theorem coloring of planar

graphs and some other sp ecial cases but no nontrivial p erformance guarantee on general

k k

graph was known We present a parallel algorithm that uses at most n colors on

k colorable graphs for a p erformance ratio of O n log n where n is the number of vertices

An algorithm is said to b e online if the color assigned to i is indep endent of ver

tices with index greater than i This corresp onds to various dynamic situations such as the

assignments of frequencies to mobile users The online graph coloring problem has b een widely

studied and for various other reasons as well In particular an algorithm that improves



the trivial p erformance ratio to O n log n was given by Lovsz Saks and Trotter and

p

p

n on colorable graph and O n log n ratio on a randomized algorithm with ratio of O



email mmhrhihiis Research partly p erformed at Rutgers University Tokyo Institute of Technology and

Japan Advanced Institute of Science and Technology Hokuriku and supp orted there by a DIMACS Graduate

Fellowship a Toshiba Endowed Chair and a PFU Endowed Chair resp ectively

general graphs was given by Vishwanathan We improve here the randomized algorithm of

for a p erformance ratio of O n log n

We rst present in Section a sequential graph coloring algorithm with a p erformance ratio

of O n log n and use it to derive b oth a parallel algorithm in Section and a randomized on

line algorithm in Section with the same p erformance ratio The randomized algorithm is an

p

log n improvement of an algorithm of Vishwanathan and the p erformance ratio of O n

compares favorably with the lower b ound for randomized online coloring of n log n

The p erformance ratio of the parallel algorithm is the rst of its kind It also compares favor

ably with the b est p erformance ratio known for sequential coloring of O nlog log n log n

and the b est announced lower b ound known for that quantity assuming P NP of

n for each Finally using the parallel coloring algorithm we obtain in Section a

O nlog log n log n p erformance ratio for parallel indep endent set problem This also com

pares favorably to the sequential upp er b ound of O n log n and the announced lower b ound

of n for each See Table for comparisons

Online Parallel Sequential

Deterministic Randomized

p



log n O nlg lg n log n Known upp er b ound O n log n O n

Here O n log n O n log n

Known lower b ounds n log n n log n n n

Table New and known results on approximate graph coloring

This connection b etween apparently unrelated algorithmic paradigms is quite surprising as

the two paradigms we study are seemingly orthogonal and incompatible Parallel computation

is inherently oline and online computation inherently sequential Nevertheless there are

certain strong similarities in particular b oth require a lo ose coupling among subproblems

Parallel algorithms b enet greatly from lo cality which can b e thought of as a case of limited

information similar to the online regime Online algorithms know more ab out certain other

subproblems but with less control over what they learn

Approximate Coloring Algorithms

In this section we present and analyze the underlying sequential oline coloring algorithm that

forms the basis of the parallel and online variants

Let us rst review what we have learned from previous metho ds First consider an algorithm

due to Wigderson Given a threecolorable graph on n vertices we lo ok for a vertex of degree

p

n The neighborho o d of that vertex must b e twocolorable so we color these neighbors at least

with two fresh new colors in linear time and remove them from the graph We rep eat this until

p p

n for at most n iterations The remaining graph can easily b e maximum degree falls b elow

p p

n colors for a combined total of n colors This principle colored using max degree

can b e generalized to graphs of larger chromatic number where the key observation is that the

neighborho o d of any no de in a k colorable graph must b e k colorable This allows us the

break the problem into a collection of somewhat easier subproblems

The problem with this approach is that it lls the color classes in sequence We need a

metho d that accumulates many color classes concurrently rather than searching for individual

indep endent sets We extract an idea from the online algorithm of Vishwanathan

A simple greedy algorithm sometimes referred to as First Fit works as follows The vertices

are pro cessed in sequence according to some arbitrary order and a vertex is assigned to the

rst color to which none of the preceding neighbors of the vertex have b een assigned If we are

given a limited number of colors it may not b e p ossible to color all the vertices The resulting

coloring of a subset of the vertices is called a maximal partial coloring of the graph

A maximal partial coloring is an assignment of some of the no des into xed number of

greedy color classes so that each remaining no de is adjacent to some no de in each greedy

class The imp ortant p oint is that each color class partitions the remaining no des into subsets

of the neighborho o ds of the no des in the color class If we choose a small class the remaining

uncolored no des induce relatively few subproblems The total number of colors used is then the

number of greedy color classes plus the sum of the number of colors used in the subproblems

Such a maximal partial coloring is easy to achieve sequentially via the First Fit algorithm which

assigns a vertex to the rst compatible color class if one exists

KColorGk

b egin

if k BipartiteColorG

else if k log n Assign each vertex a dierent color

else

k k k

s sn k nk

ResidueNo des MaximalPartialColorGs

Find the smallest greedy color class and let w w b e its no des

p

Partition the ResidueNo des into R R such that no des in R are adjacent to w

p i i

for i to p

ColorR k

i

end

Figure An approximation algorithm for k colorable graphs

We state the main algorithm in Figure The algorithm nds a maximal partial coloring

partitions the remaining no des around the smallest color class and recurses on those subprob

lems with the recursion b ottoming out at bipartite graphs When the chromatic number is to o

large for us to handle with resp ect to the size of the graph we settle on the trivial coloring of

one color p er vertex

Theorem The number of colors used by Color on k colorable graphs is at most f k

log k

k k

n where f k O

k

Proof Let An k denote the maximum number of colors used by the algorithm to color any

k k k

k colorable graph on n vertices Let f k denote k k and let

k k

An k denote f k n Our prop osition is that An k An k for all p ositive

integers n k k

The pro of pro ceeds by induction on k with the base case k established by an optimal

bipartite algorithm Hence assume from here on that An k An k for all p ositive

integers n

The number of colors used by our algorithm is the number of greedy colors s sn k

plus the number of colors used in the t subproblems each on m vertices To maintain sanity

j

and simplicity of presentation we ignore all ceilings Note that An k k sn k

so our main task is to show that the number of colors used in the subproblems b e at most

k k

k sn k k n Intuitively we nd that the worst case o ccurs when the

subproblems are all of equal size and then derive the worstcase number of subproblems

If g is the total number of vertices in the greedy color classes then the number of vertices

in the smallest class and thus the number of subproblems is at most g s By the inductive

assumption we can assume that the number of colors used in the subproblems is b ounded by

A Thus

g s

X

Am k An k s

j

j

Since the approximation function is convex the sum is maximized when the subproblem sizes

are all equal

P

X

m

g

j

j

An k s maxf m g A k j n g

j

g

s g s

j

Substitution and simplication gives

n g k

k k k k

An k s maxfg g s

k k

g

g

k

ng

k k k k

attains a maximum value of k nk The function f g g

k

g

conveniently simplifying the formula By straightforward algebraic manipulation we nd

that

k k

An k s n s An k

Our algorithm assumes that the chromatic number of the graph is given as input To

circumvent that we can run the algorithm for all natural numbers less than n and retain the

b est result A more ecient binary search approach add only a log k O log log n factor to the

Start with the range n as the algorithm is guaranteed to fail for k and

succeed for k n and run the algorithm for the midvalue of the interval If it succeeds run it

on the lower half if it fails run it on the upp er half At the end we have determined a value k

for which the algorithm succeeds while failing for k Then k must b e a lower b ound on

the chromatic number of the graph and our p erformance no worse than if this was given

Corollary The performance ratio of Color is O n log n

Proof The ratio of the size of the approximation to the chromatic number of the graph is

O n which is maximized when log n

Parallel Coloring Algorithm

In this section we discuss how to parallelize the steps of the preceding coloring algorithm while

maintaining an identical p erformance

The main algorithm can b e parallelized easily by executing the lo op iterations in parallel

The bipartite graphs can b e colored exactly by nding a forest of ro oted spanning trees using

the Euler tour technique see computing the depth of each no de using list ranking and

assigning no des of o dd depth one color and no des of even depth the other color

The hard part is nding a maximal partial coloring A construction of Luby reduces it

to the maximal indep endent set problem which has known NC solutions

0

We create a new graph G that consists of copies of the original graph with vertices in

dierent copies adjacent i they corresp ond to the same original vertex Consider a maximal

0

indep endent set I in G and let C b e the set of vertices in I b elonging to the ith copy of G

i

0

in G Note that no two vertices in I corresp ond to the same vertex in G thus I and C map

i

0 0 0

of V Moreover each vertex in V G I is adjacent to some onetoone to subsets I and C

i

MaximalPartialColorGx

b egin

0

Construct graph G on n x vertices

j

0

V G fv v V G j ng

i

i

j

0 z

i s j z v v E Gg E G fv v

i y

y

i

0

I MISG

i

C I G i x

i

x

C ResidueNo des G

i

i

return fC g ResidueNo des

i

end

Figure Maximal partial coloring algorithm in parallel

0 0

form as otherwise its addition to I would still b e indep endent Hence the C vertex in each C

i i

a maximal partial coloring of G

The complexity of the maximal partial coloring algorithm is dominated by the complexity

of the maximal indep endent set algorithm used The faster algorithm known for the latter

problem runs in O log jV j time due to Luby A pro cessorecient algorithm of Goldb erg

and Sp encer runs in O log jV j time with jV j jE j log jV j pro cessors The number of

0 k k

vertices of the comp osed graph G is s n where s sn k nk is the

s

number of greedy bins allo cated The number of edges is then sjE j n sjE j ns and

thus the pro cessor complexity of our algorithm would b e sjE j ns log n n

The complexity of the main algorithm is also dominated by k recursion levels of calls to the

maximal partial coloring subroutine adding a factor of k O log n to the time complexity

Online Coloring Algorithm

We show in this section how to convert Color into an online algorithm

The online coloring problem is dened as follows An adversary selects a graph and an

ordering of the vertices It feeds us one no de at a time along with its edges to the no des already

received We are to irrevocably assign that no de a color consistent with the colors of the previous

no des We are allowed arbitrary amount of time but we are not given any further information

ab out the graph including its size and chromatic number

In order to deter the adversary from foreseeing our every move randomization frequently

proves useful An adversary is said to b e oblivious if it picks the graph and the ordering b efore

we make any choices ie without knowledge of the values of the random bits rather than

constructing the graph adaptively according to our choices We measure the exp ected number

of colors used by a randomized algorithm on a given graph where the exp ectation ranges over the

values of the random bits The p erformance ratio is then dened as the ratio of this exp ectation

to the chromatic number maximized over all graphs

For some online problems randomized algorithm against adaptive adversary form an inter

mediate mo del b etween the deterministic and the randomized oblivious cases For the graph

coloring problem however no results are known that are sp ecic to this mo del

In order to convert the oline algorithm into an online one several changes need to b e

made and hurdles to b e overcome The greatest change is that we must pro cess the no des fully

in sequence rather than pro cessing the chromatic levels in sequence We must also overcome

three diculties the size of partitioning classes will not b e known b efore they must b e chosen

the size of the input will not b e known and the chromatic number will b e unknown

The rst hurdle is that a partitioning class must b e selected in advance b efore we know its

eventual size The concern is that if the size will b e large many subproblems will b e spanned

The solution is to use randomization pick a uniformly random class as a partitioning class

Since the adversary must choose the ordering of the graph in advance the exp ected number of

no des in the class will b e suciently small

The second problem is estimating the size of the input and recursively the sizes of the

subproblems For this we apply a well known trick in the online trade see eg the

doubling strategy which works as follows Assume to b egin with that the size of the input n

i

equals for i initially equal When the actual number of no des exceeds this estimate we

increment the estimate to the next p ower of retire the old color classes and start afresh with a

new set of color classes We continue this way and when the input nally ends our estimate of n

is never o by more than a factor of Since the function giving our approximation p erformance

is necessarily convex it follows that the overhead caused by this strategy is no more than a small

constant

It is imp ortant to notice that we can apply the same strategy recursively for the subproblems

as well From the p ersp ective of an online algorithm the subgraph induced by the no des in

a residue class ie in a subproblem is structurally no dierent from the original problem

except for the fact that the chromatic number is one less We shall therefore allo cate sets of

color classes to this subproblem completely indep endent of the other subproblems or the parent

problem This strategy is what primarily distinguishes our algorithm from Vishwanathans

The third problem not knowing the chromatic number in advance can b e handled simi

larly To b egin with we assume the graph is colorable and as our assumptions are shattered

usually by the bipartite algorithm stumbling up on a nonbipartite subgraph we increment our

estimate by one and start from scratch with a fresh set of colors

Finally we need to nd online versions of the subroutines called by algorithm KColor

As stated in there is a simple algorithm for online coloring using no more

than log n colors And a FirstFit algorithm that is restricted to using only the greedy color

classes actually available is p erfect for p erforming a maximal partial coloring

The resulting algorithm is shown in Figures

OnlineColor G

Online color a graph G

b egin

k

R oot NewColoring k

foreach v V G in sequence do

R esul t KColorOnlineR oot v

if R esul t F ail ur e then

k k

R oot NewColoringTree k

end

Figure Online Graph Coloring Algorithm

The online algorithm colors the graph by constructing and maintaining a coloring tree This

tree contains each previous vertex in a greedy color class at some no de in the tree Each no de

in the tree contains the following static information n the capacity of the subtree k the

colorability of the subtree s sn k the number of greedy bins and par t the index of the

1

Static in the sense that the values are given at the creation of the subtree ro oted by the no de

KColorOnline T v

Reentrant procedure

Assign a color to the vertex v using coloring tree T

Return success or failure

b egin

with T do

if k then

return BipartiteColorT v

if N n then

T NewColoringTreen k

N N

for j to s do

if v is nonadjacent to all vertices in P j then

Assign v the color P j or C O LO R v P j

return Success

v must now b e adjacent to some no de in each color class

Nondeterministically cho ose a neighb or w of v in partitioning class P par t

if R w is uninitialized then

R w NewColoringTree k

return KColorOnlineR w v

end

Figure A randomized online coloring algorithm

NewColoringTree n k

Create and return a new coloring tree with capacity n and chromatic upp er b ound k

b egin

with T do

n n

k k

c k k

s sn k nk

P s A new set of colors

par t Random integer b etween and s

N

return T

end

Figure Create and initialize a new coloring tree

partitioning bin Each no de also contains the following variables N the current number of

vertices in the subtree P s the greedy bins the bins themselves are static their contents

are dynamic and R a residue tree for each vertex assigned to the partitioning bin P par t

w

Finally each no de implicitly contains a host of cuto subtrees

Here c is a constant whose optimal value is determined in the analysis

When a vertex v arrives the algorithm will assign it to one of the static greedy bins at

the ro ot whenever p ossible If this fails we choose any neighbor w of v in the partitioning bin

P par t The vertex is the sent recursively to the subproblem corresp onding to the subtree R w

which is constructed if necessary This continues until the colorability index of the tree is at

most two in which case we know how to online color the no de eciently

Two invariants must b e maintained in each subtree of the coloring tree

N n That is the number of vertices in the subtree at any time cannot exceed the

capacity of the subtree

The colorability index for the subtree ro oted at no de v must b e at least one greater than

the colorability index for the subtree ro oted at some child of v This insures that the

colorability index stays an upp er b ound on the chromatic number of the graph

The former invariant is maintained by KColorOnline by creating a new subtree with double the

capacity whenever the vertex count exceeds capacity The old greedy bins color classes are

then set aside and never used again and the vertices they contain do not factor into the coloring

assignment other than that their number is still a part of the content of the subtree

The latter invariant is maintained by OnlineColor again by cutting o the old tree and

creating a new tree with a colorability index one larger Such an action is triggered by the

failure of the online bipartite coloring algorithm at some leaf when some leaf b ecomes three

chromatic and the eect lters up the to the ro ot

Analysis

The analysis of the p erformance of the online algorithm resembles that of the parallel approxi

mation but is more involved The heart of the analysis is contained in the following result

k k k

Theorem KColorOnline uses at most O n log n expected number of col

ors on k colorable graphs with n vertices

Proof Let us rst dene some notation Let An k denote the maximum exp ected number

of colors the algorithm uses to color any k colorable graph on n vertices when n is known in

advance Let B n k similarly b e the maximum exp ected number of colors when n is not known

in advance Let c denote a realvalued constant to b e dened later We shall need the following

functions and constants

c

k k

c

k

k

k k k

sn k c n log n

k

An k k sn k

ck

B n k An k

B n k is the upp er b ound that we shall prove for B n k and similarly An k is the b ound

for An k sn k denotes the number of greedy bins allo cated as previously dened Default

logarithm is base

Given the ab ove notation the prop osition of the theorem can b e restated as

B n k B n k for all n k k

The pro of pro ceeds by induction on k simultaneously for A and B For the base case observe

that B n An log n An B n To prove the inductive step we rst show

that An k An k for any p ositive integer n

The number An k of bins used during the time that the size estimate n remains unchanged

equals the number of greedy bins plus at most the average cost of the subproblems over the s

dierent choices of a partitioning bin The latter amounts to at most

n

s

i

X X

i

B m k

j

s

j i

i

where n is the number of elements in greedy bin i and m the size of the subgraph Let g

i

j

denote the total number of elements placed in greedy bins By the inductive assumption and

the convexity of B is at most

n g

maxfg B s k g

g

s g

We now plug in the value of B obtaining

n g

k k c k k

max fg g c k s log n

k

g

g

As in Theorem the maximum of the inner function simplies the formula leading to

An k s k sn k An k

Hence An k An k for any n We now show that this implies that the same holds for

B and B from which the theorem follows

In general we do not know the number of no des in advance at any level in the recursion

The algorithm xes a value for n and when that count has b een exceeded it doubles its exp ec

tation for the total number of vertices We obtain d dlog ne subproblems coloring the no des

d d d

fg f g f g f ng resp ectively Let r n denote the total number

d

of no des colored by the end of the last doubling step ie r n

The total number of bins used is at most the sum of those used on each of the sizeb ounded

subproblems the residue bins used on the last r n no des and the greedy bins allo cated in

the last step That is

d

X

r n

An k k A r n k A B n k

i

k

i

The rst term evaluates to

d

X

ik k k k

Ar n k Ar n k

i

Maximizing the rst two terms over the choices for r using standard calculus yields

k k k k

An k B n k

k

Insp ection shows that this is at most

k

B n k An k

Thus we set c and the theorem follows

Theorem Our online algorithm uses at most O n log n colors without

a priori know ledge of either n or the chromatic number G

Proof We maintain a lower b ounded estimate k on the chromatic number of the graph This

estimate is incremented only when the current graph has b een proven to b e non k colorable

Consider the no des colored while the estimate was i If we denote their count by N we know

i

that OnlineColor used no more than B N i colors to color those no des This can b e b ounded

i

from ab ove by B N k which is a convex function Thus the sum of these values i k

i

is maximized when all N are equal for a total of

i

n

k k k k

k k B n k n log n k B

k

number of colors

We obtain the same p erformance ratio as in the oline case with a similar pro of

Corollary Randomized online coloring is O n log n approximable

Indep endent Sets in Parallel

The parallel coloring algorithm can also b e used to nd cliques or indep endent sets in parallel

k k

via an observation due to Wigderson Recall that the algorithm used at most n

colors where the graph was provably not k colorable since a k chromatic subgraph had b een

found This subgraph is formed by the pivot no des w at each of the k recursive levels

i

along with a nonbipartite subgraph i e an o dd at the end of the recursion These pivot

no des are by denition mutually adjacent as well as adjacent to the o dd cycle thus pro ducing

a k Since the clique number of the graph is at most the chromatic number which is at

k k

most the number of colors we used this yields a n k approximation of the maximum

clique for an O n log n p erformance ratio Finally recall that we can obtain an indep endent

set algorithm with the same p erformance by applying our algorithm to the

Notice that the algorithm runs in p olylogarithmic time for all interesting values of k i e

k log n

Another source of a parallel indep endent set algorithm can b e found in the rst step of

Berger and Romp els coloring algorithm It essentially involves choosing p olynomial number

of random vertex subsets of size log n and testing for nonadjacency This can b e made deter

k

ministic via the pigeonhole principle without destroying the indep endence of the samples and

hence can b e parallelized trivially This metho d yields a n log n approximation

k

Notice that the two metho ds complement each other since ourWigdersons metho d p erforms

b est for small values of k while the ab ove random sampling p erforms relatively b etter for larger

values of k If we run b oth metho ds and retain the b etter result we get an algorithm with a

p erformance ratio of

k k

n n

max min

k

k log n

k

which is minimized when k log n log log n as O nlog log n log n

Graphs with large indep endent sets

We shall show how given a graph with an indep endence number at least n we can nd

p

n in p olylogarithmic time an indep endent set of size

The metho d starts by removing a maximal collection of disjoint triangles cliques of size

from the graph Such a collection can b e obtained as a maximal indep endent set in the

graph whose vertices are all the triangles in the original graph with edges b etween intersecting

triangles Since at most one no de in each triangle can b e a member of a given indep endent set

at least n n no des remain after the triangle removal

TriangleFreeIS G

b egin

I MaximalIndep endentSetG w w w

t

N V G

for i to t do

N N fv V G v w E Gg

i i i

N fv V G v w E Gg N

i i i

return max N fw g N fw w w g N fw w w g

t t t

end

Figure Algorithm for nding indep endent sets in triangle free graphs

Given a trianglefree graph we apply the algorithm of Figure which is based on a sequential

algorithm of The algorithm nds a maximal indep endent set with an ordering of the vertices

N of any prex of the vertices as well as in the set and nds the common nonneighborho o ds

i

the complementary neighborho o ds N Each neighborho o d must b e indep endent since the graph

i

contains no triangles and is also contained in an earlier nonneighborho o d The nal result is

the largest of these indep endent sets considered and its size must b e at least the smallest t such

p

t

that n or approximately n

N is a prex computation that can b e p erformed Finding the recursive nonneighborho o ds

i

in logarithmic time with linear number of pro cessors The complexity is therefore asymptotically

equivalent to that of the MIS algorithm used

k

The ab ove construction can b e generalized to nd indep endent sets of size n in

k clique free graphs providing a constructive parallel pro of of the upp er b ound on Ramsey

numbers by Erds and Szekeres That also allows us to nd such indep endent sets in graphs

with indep endence number greater than nk for k but alas the pro cessor complexity of

k

removing the k cliques grows as fast as n

Acknowledgements

I would like to thank Ravi Boppana and Sundar Vishwanathan for helpful discussions and the

two anonymous referees for comments that improved the presentation I also thank Sundar for

reinventing Lubys construction for me

References

D Bean Eective coloration J Symbolic Logic

B Berger and J Romp el A b etter p erformance guarantee for approximate graph coloring

Algorithmica

R B Boppana and M M Halldrsson Approximating maximum indep endent sets by

excluding subgraphs BIT June

P Erds and G Szekeres A combinatorial problem in geometry Compositio Math

U Feige and J Kilian Zero knowledge and the chromatic number In Proc of Conference

on Computational Complexity June

A V Goldb erg S A Plotkin and G E Shannon Parallel symmetrybreaking in sparse

graphs SIAM J Disc Math Nov

M Goldb erg and T Sp encer Constructing a maximal indep endent set in parallel SIAM

J Disc Math Aug

A Gyrfs and J Lehel Online and rst t colorings of graphs J

P Ha jnal and E Szemerdi Bro oks coloring in parallel SIAM J Disc Math

M M Halldrsson A still b etter p erformance guarantee for approximate graph coloring

Inform Process Lett January

M M Halldrsson and M Szegedy Lower b ounds for online graph coloring Theoretical

Comput Sci Aug

J Hstad Clique is hard to approximate within n Manuscript

R Karp and V Ramachandran A survey of parallel algorithms for sharedmemory ma

chines In J van Leeuwen editor Handbook of Theoretical Computer Science volume A

chapter pages Elsevier Science Publishers BV

H A Kierstead and W T Trotter An extremal problem in recursive combinatorics In

Proc th Southeastern Conf on Combinatorics Graph Theory and Computing Congres

sus Numerantium XXXIII pages

L Lovsz M Saks and W T Trotter An online graph coloring algorithm with sublinear

p erformance ratio Discrete Math

M Luby A simple parallel algorithm for the maximal indep endent set problem SIAM J

Comput

S Vishwanathan Randomized online graph coloring J Algorithms Dec

A Wigderson Personal communications

A Wigderson Improving the p erformance guarantee for approximate graph coloring

J ACM