THE HARMONIOUS CHROMATIC NUMBER OF BOUNDED DEGREE GRAPHS KEITH EDWARDS A A harmonious colouring of a simple graph G is a proper vertex colouring such that each pair of colours appears together on at most one edge. The harmonious chromatic number h(G) is the least number of colours in such a colouring. Let d be a fixed positive integer, and ε " 0. We show that there is a natural number M such that if G is any graph with m & M edges and maximum degree at most d, then the harmonious chromatic number h(G) satisfies (2m)"/# % h(G) % (1ε) (2m)"/#. 1. Introduction A harmonious colouring of a simple graph G is a proper vertex colouring such that each pair of colours appears together on at most one edge. Formally a harmonious colouring is a function c from a colour set C to the set V(G) of vertices of G such that for any edge e of G, with endpoints x, y say, c(x) 1 c( y), and for any pair of distinct edges e, e«, with endpoints x, y and x«, y« respectively, ²c(x), c( y)´ 1 ²c(x«), c( y«)´. The harmonious chromatic number h(G) is the least number of colours in such a colouring. A survey on harmonious colourings is given by B. Wilson in [9]. If we have a harmonious colouring of G with k colours, then since each pair of colours can appear on at most one edge, it is clear that the number of colour pairs, k namely , must be at least m, the number of edges of G. This motivates the 2 following definition. 01 D. Let m be a positive integer. Then we define Q(m) to be the least k positive integer k such that & m. 2 01 It is easily calculated that Q(m) ¯ 8"(1(8m1)"/#7, however for our purposes # it suffices to note that Q(m) is approximately (2m)"/#. We are interested in classes of graphs for which h(G) is close to (2m)"/#, in particular, those classes Γ for which h(G) ¯ (2m)"/#o(m"/#). It follows readily from the theorem of Wilson [10] on the packing of copies of a graph into a complete graph, that if Γ(K ) consists of graphs which are disjoint unions of a fixed graph K, then h(G) ¯ (2m)"/#o(m"/#) for G ` Γ(K ), and this extends easily Received 30 January 1995. 1991 Mathematics Subject Classification 05C15. J. London Math. Soc. (2) 55 (1997) 435–447 Downloaded from https:/www.cambridge.org/core. IP address: 88.99.165.207, on 12 Jul 2017 at 18:09:43, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms. https:/www.cambridge.org/core/product/ACFF871550A2451AA6A7474CEA7A1385 436 to classes with bounded component size. It is proved in [2] that the same is true for the class Td consisting of trees of maximum degree at most d, and Γd, γ consisting of graphs of maximum degree at most d and genus at most γ. In this paper we extend all these results and show that if Γd consists of all graphs of maximum degree at most d, then h(G) ¯ (2m)"/#o(m"/#) for G ` Γd. The method used is quite different from those used for the earlier results, and relies heavily on the powerful Pippenger–Spencer theorem on the chromatic index of uniform hypergraphs [8]. We also discuss the relationship of this result to Wilson’s Theorem, and offer an alternative interpretation based on identifying vertices of graphs. 2. Regular graphs In this section we show how to find asymptotically good colourings of regular graphs. We start with some preliminaries. D. An incomplete harmonious colouring of a graph G is a colouring of a subset X of the vertices of G, such that the colouring is a harmonious colouring of the graph induced by the vertices X. D. A partial harmonious colouring of a graph G is an incomplete harmonious colouring which satisfies the additional property that any two coloured vertices of G which have a common neighbour must have distinct colours. R. A partial harmonious colouring can always be extended to a harmonious colouring ; this is not always possible with an incomplete colouring. Hypergraphs We now give some hypergraph notation and state the theorem of Pippenger and Spencer. Let G be a hypergraph ; then G is k-uniform if every edge contains k vertices. If and w are vertices of G, then the degree of , denoted by degG(), is the number of edges containing , and the codegree of and w, denoted by codegG(, w), is the number of edges containing both and w. Let D(G) ¯ max degG(), ` v V(G) d(G) ¯ min degG(), ` v V(G) C(G) ¯ max v,w ` V(G),v1w codegG(, w). T (Pippenger–Spencer [8]). For eery k & 2 and δ " 0, there exist δ« " 0 and n such that if G is a k-uniform hypergraph on n(G) & n ertices satisfying ! ! d(G) & (1®δ«) D(G) and C(G) % δ«D(G), then the chromatic index χ(G) satisfies χ(G) % (1δ) D(G). Probability Estimates We now state two more theorems used to estimate probabilities. These can be found, for example, in [5]. T (Azuma’s bounded differences inequality). Let X , … , Xn be inde" pendent random ariables with Xk taking alues in a set Ak for each k. Suppose that the Downloaded from https:/www.cambridge.org/core. IP address: 88.99.165.207, on 12 Jul 2017 at 18:09:43, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms. https:/www.cambridge.org/core/product/ACFF871550A2451AA6A7474CEA7A1385 437 (measurable) function f : 0 Ak ! R satisfies r f(x)®f(x«)r % ck wheneer the ectors x and x« differ only in the kth co-ordinate. Let Y be the random ariable f [X , … , Xn]. " Then for any t " 0, we hae P [rY®Ex (Y )r & t] % 2 exp (®2t#}3 ck#). T. Let X , … , Xn be independent random ariables with 0 % Xk % 1 for " each k, let Xa ¯ (1}n) 3 Xk and let p ¯ Ex [Xa ]. Then for any 0 ! ε ! 1, we hae P[Xa ®p & εp] % exp (®"ε#np), $ P[Xa ®p %®εp] % exp (®"ε#np). # C 2.1. Let X , … , Xn be independent random ariables with P[Xk ¯ 1] " ¯ p and P[Xk ¯ 0] ¯ 1®p for each k. Then for any 0 ! ε ! 1, we hae P[r3 Xk®npr & εnp] % 2 exp (®"ε#np). $ Main lemma We now introduce some notation and state and prove the main technical lemma. N. Let G be a graph with vertex set V and edge set E. Let X, Y and Z be subsets of V. We define the set of neighbours of X, denoted by N(X ), by N(X ) ¯ ² ` V r (, w) ` E for some w ` X ´. Define the edge set induced by X, denoted by E(X ), by E(X ) ¯ ²(, w) ` E r , w ` X ´. Define the edge set induced by X and Y, denoted by E(X, Y ), by E(X, Y ) ¯ ²(, w) ` E r ` X, w ` Y ´. Finally define the neighbourhood set of X and Y in Z, denoted by NZ(X, Y ), by NZ(X, Y ) ¯ N(X )fN(Y )fZ. N. Statements below of the form xj ¯ f(n)O(z(n)) for j ¯ 1, … , J should be taken to be uniform over j, that is, to mean that there is a constant α and an integer N such that rxj®f(n)r % αz(n) for all n & N and for all j ¯ 1, … , J. L 2.2. Let d be a fixed integer, and let η " 0. Then there is an integer N such ! that if G is a d-regular graph with n ertices and m (¯ "dn) edges, where n & N , then # ! we can find an incomplete harmonious colouring of G with at most (1η) (2m)"/# colours such that at least (1®η) n of the ertices are coloured. Proof. Let δ ¯ η}6, and let δ« be the corresponding real number, depending on δ, given by the Pippenger–Spencer Theorem. We now choose an integer A satisfying A & max ²16d %, d #}(δ«)#, 6}η, 9}η#, 2}δ«´. We proceed by a sequence of stages. At each stage we shall colour approximately n}A of the vertices. We shall stop after Stage i when A®i ¯ 8A"/#7 (thus leaving some vertices uncoloured). Hence we shall always have A®i & A"/#. Let k ¯ (2m)"/#, and V be the vertex set of G. We will ensure that at the start of Stage i1, the following conditions hold : 1. V is partitioned as V ¯ C e…eCieZ, where rZ r ¯ (A®i) n}AO(n$/%), and " for each j ¯ 1, … , i, we have rCjr ¯ n}AO(n$/%) ; Downloaded from https:/www.cambridge.org/core. IP address: 88.99.165.207, on 12 Jul 2017 at 18:09:43, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms. https:/www.cambridge.org/core/product/ACFF871550A2451AA6A7474CEA7A1385 438 2. 3. 4. 5. 6. 7. each Cj is partitioned into colour classes Cj, , … , Cj,t , where tj ¯ k}AO(n$/)) ; " j rE(Cj)r ¯ 0 for each j ; rE(Z )r ¯ (A®i) (A®i®1) m}A(A®1)O(n$/%) ; rE(Cj,r, Z )r ¯ (A®i) k}(A®1)O(n$/)) for each j, r ; rNZ(Cj,r, Cj «)r % kd}AO(n$/)) for each j, r, and j « " j ; for each j, no two distinct elements of Cj have a common neighbour in Z. Note that before stage 1, that is, when i ¯ 0, all the above hold trivially. Stage i1. At stage i1 we first choose a certain subset W of Z. Let Z « ¯ Z cW. Then W and Z « will have the following properties : 1. rW r ¯ n}AO(n$/%) and rZ «r ¯ (A®i®1) n}AO(n$/%) ; 2. rE(Cj,r, W )r ¯ k}(A®1)O(n$/)) for each j, r ; 3. rE(W )r ¯ 0 ; 4. rE(W, Z «)r ¯ 2(A®i®1) m}A(A®1)O(n$/%) ; 5. rE(Z «)r ¯ (A®i®1) (A®i®2) m}A(A®1)O(n$/%) ; 6. rNZ «(Cj,r, W )r % kd}AO(n$/)) for each j, r ; 7. rNW(Cj,r, Cj «)r % kd}A$/#O(n$/)) for each j, r, and j « " j. To construct W, we first choose a subset B of Z by including each element, independently, with probability pi, where pi is chosen so that pi(1®pi)d # ¯ 1}(A®i). Note that this will always be possible since 1}(A®i) % 1}A"/# % 1}4d #, while an easy calculation shows that f(x) ¯ x(1®x)d # has maximum value at least 1}4d # on (0, 1). For each z ` Z, let the random variable Xz be the indicator function of the event [z ` B]. Thus P[Xz ¯ 1] ¯ pi. We construct W from B as follows : choose an arbitrary cyclic ordering of the vertices of Z. For each vertex of Z, construct the set T() by first including in T() all vertices of Z which are at distances 1 or 2 from in the graph induced by the vertices of Z. Then clearly rT()r % d #. To ensure that each T() has exactly d # elements, add to T() the first d #®rT()r vertices after in the cyclic ordering which are not already in T(). Note that for any « in Z, we have that « is in T() for at most 2d # vertices . Now to obtain W, let R ¯ ² r T()fB´ 1 W. Then let W ¯ BcR. It is clear that P[ ` W ] ¯ pi(1®pi)d # ¯ 1}(A®i) for any ` Z. We can establish, using Azuma’s bounded differences inequality, that with probability very close to 1, the set W will have each of the properties above. We do this as follows. Let X be any subset of Z or of E(Z ), and suppose that rX r ¯ λO(λ$/%). Now suppose that f is a function from subsets of Z to subsets of Z or E(Z ), and satisfies the following conditions : (i) If S, S « are subsets of Z such that rS∆S «r % 1, then r f(S ) ∆f(S «)r % d and (ii) the event [x ` f(T )] depends on at most d of the events [z ` T ]. Let Y be the random variable rXff(W )r. Then we claim that P[Y ¯ Ex (Y )O(λ$/%)] & 1®2 exp (®Ω(λ"/#)). To see this, first note that the random variables Xz are independent, and the random variable Y is functionally dependent on them. Now a change in the value of one Xz, corresponding to the removal from or addition to B of some vertex , results in a change in W of at most 2d # elements. Hence the resulting change in f(W ) is of at most 2d $ elements. Now for any x ` X, let Yx ¯ I[x ` f(W)], so that Y ¯ 3x ` X Yx. We know that the event [z ` W ] depends on the event [ ` B] only if ¯ z or if ` T(z). Hence [z ` W ] depends Downloaded from https:/www.cambridge.org/core. IP address: 88.99.165.207, on 12 Jul 2017 at 18:09:43, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms. https:/www.cambridge.org/core/product/ACFF871550A2451AA6A7474CEA7A1385 439 on at most d #1 of the events [ ` B]. It follows that [x ` f(W )] depends on at most d(d #1) of the events [ ` B], and so Y depends on at most rX r d(d #1) of the random variables Xv. Furthermore, for those on which it may depend, the constant cv of Azuma’s inequality is at most 2d $. Hence finally we have, by Azuma’s inequality, that P[rY®Ex (Y )r & λ$/%] % 2 exp (®2λ$/#}rX r d(d #1) 4d ') % 2 exp (®Ω(λ"/#)). In a similar way we can show that if rX r % λO(λ$/%), and Y ¯ rXff(W )r then P[Y % Ex (Y )O(λ$/%)] & 1®2 exp (®Ω(λ"/#)). Note that Ex (Y ) ¯ 3x ` X P[x ` f(W )]. We can now show that W has the required properties with high probability, as follows. 1. rW r ¯ n}AO(n$/%) and rZ« r ¯ (A®i®1) n}AO(n$/%). For W, take X ¯ Z, so that λ ¯ (A®i) n}A, and let f(S ) ¯ S. Then Y ¯ rW r and P[x ` f(W )] ¯ 1}(A®i), so Ex (Y ) ¯ n}AO(n$/%). Hence rW r ¯ n}AO(n$/%) with probability at least 1®2 exp (®Ω(n"/#)). To deal with rZ «r, we simply note that rZ «r ¯ rZ r®rW r. 2. rE(Cj,r, W )r ¯ k}(A®1)O(n$/)). Note that since any element of Z has at most one neighbour in Cj,r, then rE(Cj,r, Z )r ¯ rN(Cj,r)fZ r. So take X ¯ N(Cj,r)fZ, so that λ ¯ (A®i) k}(A®1), and let f(S ) ¯ S. Then Y ¯ rXfW r ¯ rE(Cj,r, W )r. Now P[x ` f(W )] ¯ 1}(A®i), so Ex (Y ) ¯ k}(A®1)O(n$/)). Hence Y ¯ k}(A®1)O(n$/)) with probability at least 1®2 exp (®Ω(k"/#)). 3. rE(W )r ¯ 0. This follows immediately from the construction of W. 4. rE(W, Z «)r ¯ 2(A®i®1) m}A(A®1)O(n$/%). Let X ¯ rE(Z )r, so that λ ¯ (A®i) (A®i®1) m}A(A®1), and let f(S ) ¯ E(S, Z®S ). Thus f(W ) ¯ E(W, Z «), so that Y ¯ rE(W, Z «)r. Now let (u, ) be any edge of E(Z ). Then by definition of W, if u is in W then is in Z «, and vice versa, hence the probability that the edge is in E(W, Z «) is 2}(A®i). Hence P[x ` f(W )] ¯ 2}(A®i). Thus Ex (Y ) ¯ 2(A®i®1) m}A(A®1), and so Y ¯ 2(A®i®1) m}A(A®1)O(n$/%) with probability at least 1®2 exp (®Ω(n"/#)). 5. rE(Z «)r ¯ (A®i®1) (A®i®2) m}A(A®1)O(n$/%). This follows from the fact that rE(Z «)r ¯ rE(Z )r®rE(W, Z «)r. 6. rNZ «(Cj,r, W)r % kd}AO(n$/)). We have NZ «(Cj,r, W ) ¯ N(Cj,r)fN(W )fZ « ¯ N(Cj,r)fZfN(W ). So take X ¯ N(Cj,r)fZ. Now each element of Z has at most one neighbour in Cj,r, hence we have λ ¯ rN(Cj,r)fZ r ¯ rE(Cj,r, Z )r ¯ (A®i) k}(A®1)O(n$/)). Let f(S ) ¯ N(S ). Then Y ¯ rNZ «(Cj,r, W )r, and P[x ` f(W )] % (d®1)}(A®i). Hence Ex (Y ) % (d®1) k}(A®1) ! dk}AO(n$/)), whence Y % kd}AO(n$/)) with probability at least 1®2 exp (®Ω(k"/#)). 7. rNW(Cj,r, Cj «)r % kd}A$/#O(n$/)). Take X ¯ NZ(Cj,r, Cj «), so that λ ¯ dk}A, and let f(S ) ¯ S. Then Y ¯ rNW(Cj,r, Cj «)r, and P[x ` f(W )] ¯ 1}(A®i). Then Ex (Y ) ¯ kd}A(A®i)(n$/)), whence Y % kd}A$/#O(n$/)) with probability at least 1®2 exp (®Ω(k"/#)). Downloaded from https:/www.cambridge.org/core. IP address: 88.99.165.207, on 12 Jul 2017 at 18:09:43, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms. https:/www.cambridge.org/core/product/ACFF871550A2451AA6A7474CEA7A1385 440 It is clear that with probability near 1, provided that n is large enough, W has all the above properties, hence we can certainly choose a W which does have all of these properties. Now in order to apply the Pippenger–Spencer Theorem, we need to give some temporary ‘ colours ’ to any vertex in Z « which is the neighbour of a vertex in W. To avoid confusion with the real colours, we shall refer to these temporary colours as ‘ shades ’. Let S ¯ N(W )fZ «. Note that, since no vertex in Z has two neighbours in W, we have rS r ¯ rE(W, Z «)r ¯ (A®i®1) m}A(A®1)O(n$/%). We first partition S as follows : form a graph Γ with vertex set S by joining any vertices which have a common neighbour in W. Since each vertex of W has degree at most d, it follows that Γ has maximum degree at most d®1, and consequently the vertices can be partitioned into d independent (in Γ) sets S , … , Sd. Let rSjr ¯ sj n. By moving some vertices if " necessary, we can assume that sj % dsj « for any j, j «, for if sj " dsj « some vertex of Sj has no neighbour (in Γ) in S j! and so can be moved to S !j . Note that no two vertices of any Sj have a common neighbour in W. Now let sj A}d ¯ tj. Note that since sj % dsj « for any j, j «, and 3 sj is about (A®i®1) d}2A(A®1), then the numbers sj are bounded above and below by constants independent of j. Thus the same is true for the tj. We shall assign each element of Sj independently to one of tj k shades, each with probability 1}tj k. (The sets of shades for distinct Sj are disjoint.) Thus, for each j, we can define sj n random variables Y j , Y j , … , Y js n such that P[Y jr ¯ α] ¯ 1}tj k for α ¯ 1, … , tj k. Furthermore " # j all these random variables are independent. Now let C be some particular shade used on Sj, and define the random variables Z jr by Z jr ¯ 1 if Y jr ¯ C and Z jr ¯ 0 otherwise. Then clearly all the Z jr are independent. It follows that Ex (3r Z jr) ¯ sj n}tj k ¯ nd}Ak ¯ k}A and by Corollary 2.1 P [r3 Z jr®k}Ar & ηk}A] % 2 exp (®"η#k}A). $ Hence taking η ¯ 1}n"/), we obtain P [r3 Z jr®k}Ar & O(n$/))] % 2 exp (®Ω(n"/%)). Thus we know that with high probability, each of the shades is assigned to k}AO(n$/)) vertices. Now consider two of these shades, say C and C «, used on sets Sj and Sj « respectively. We wish to determine the number N of vertices in W which are joined to both C and C «. If j ¯ j « then there are none, so assume that j 1 j «. Now each vertex of W has at most one neighbour in each of Sj and Sj « ; furthermore for distinct elements of W these sets of neighbours are disjoint. For each vertex w of W, define a random variable Xw by setting Xw ¯ 1 if w has a neighbour in both Sj and Sj « and they are shaded with shades C and C « respectively, and Xw ¯ 0 otherwise. Then the random variables Xw are independent, and for each P[Xw ¯ 1] ¯ 1}tj tj « k# % c}n for some positive constant c independent of j, j «. Then P[3 Xw & t] % 0rWt r1 (c}n) . t Using the (very crude) estimate rW r ! n, we obtain P[3 Xw & t] % (nt}t!) (c}n)t ¯ ct}t!. Taking t ¯ 8n"/%7, for example, we find that P[N & n"/%] % cn"/%}n"/%!. Similarly, if we consider a new shade C and a colour Cj,r we can show that the number of common Downloaded from https:/www.cambridge.org/core. IP address: 88.99.165.207, on 12 Jul 2017 at 18:09:43, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms. https:/www.cambridge.org/core/product/ACFF871550A2451AA6A7474CEA7A1385 441 neighbours of these in W is O(n"/%) with probability very near to 1. It follows that the choice of shades for the vertices in S can be made so that (i) each shade is used on k}AO(n$/)) vertices, and (ii) for any shade C, and colour or shade C «, we have NW(C, C «) ¯ O(n"/%). We now construct a d-uniform hypergraph H. The vertex set V(H ) is the set of colours and shades. The edges are defined as follows. For each vertex w in W, the set of colours and shades of the neighbours of w is an edge e(w). Note that these colours and shades are all distinct, so that H is d-uniform as required. We shall call w the parent vertex of the edge. We have to consider the degrees and co-degrees of vertices in H. Firstly, the degree of a vertex C in H (that is, a colour or shade C in G) is the number of vertices of W which are adjacent to a vertex of C, which equals rE(C, W )r since each vertex in W is adjacent to at most one in C. Now we have rE(C, W )r ¯ k}(A®1)O(n$/)) or k}AO(n$/)) for any colour or shade C. Hence we have d(H )}D(H ) ¯ (k}AO(n$/)))}(k}(A®1)O(n$/))) ¯ (A®1)}AO(n−"/)) ¯ 1®1}AO(n−"/)). Hence since 1}A % δ«}2, provided that n is large enough we have d(H ) & (1®δ«) D(H ). Now for any vertices C, C « of H, we have codeg (C, C «) ¯ rNW(C, C «)r % kd}A$/#O(n$/)) for any C, C «. Since A & d #}(δ«)#, we have d % δ«A"/#, and so C(H ) % kd}A$/#O(n$/)) % kδ«}AO(n$/)) ¯ δ«k}(A®1)®δ«k}A(A®1)O(n$/)) % δ«D(H ), provided that n is large enough ; hence by the Pippenger–Spencer Theorem [8], we have χ(H ) % (1δ) D(H ). Thus we obtain a colouring of the edges of the hypergraph H with at most (1δ) D(H ) colours, so that any two non-disjoint edges receive different colours. We now transfer the colour of each edge e(w) to its parent vertex. Notice that if u and are two vertices of W with the same colour, then there cannot be vertices x and y, with (u, x) and (, y) both edges of G, and x and y having the same colour Cj,r. For then Cj,r ` e(u)fe(), and so u and receive distinct colours, a contradiction. Thus, we have, roughly speaking, extended the harmonious colouring to W using approximately k}A new colours. We have to recolour a few of the newly coloured vertices to set up the correct conditions for the next stage. We have that rE(W, Z «)r ¯ 2(A®i®1) m}A(A®1)O(n$/%). This is equal to the sum over the colours C used for W of rE(C, Z «)r. Also we know that each shade is used on k}AO(n$/)) vertices, hence the number of shades is [2(A®i®1) m}A(A®1)O(n$/%)]}[k}AO(n$/))] which is (A®i®1) k}(A®1)O(n$/)). Since rE(C, Z «)r is bounded by the number of shades, we have rE(C, Z «)r ¯ (A®i®1) k}(A®1)O(n$/)). Now let a be the maximum value of rE(C, Z «)r, and let b ¯ :rE(W, Z «)r}a9. Then b ¯ k}AO(n$/)). Now discard Downloaded from https:/www.cambridge.org/core. IP address: 88.99.165.207, on 12 Jul 2017 at 18:09:43, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms. https:/www.cambridge.org/core/product/ACFF871550A2451AA6A7474CEA7A1385 442 some of the colours of W to leave exactly b. We then recolour these vertices as follows : consider them in some sequence, and for each, colour it with the colour C for which rE(C, Z «)r is least. Since ab % rE(W, Z «)r, we know that the maximum value of rE(C, Z «)r is now greater than a, and it follows that the difference between the minimum and maximum values can be at most d. Suppose the maximum value is now a«. Then ba« & rE(W, Z «)r & b(a«®d ) and it follows easily that a« ¯ (A®i®1) k}(A®1)O(n$/)) from which rE(C, Z «)r ¯ (A®i®1) k}(A®1)O(n$/)). We now set Ci+ ¯ W, ti+ ¯ b and the sets Ci+ , … , " " "" Ci+ , t equal to the colour classes of the new colours. " i+" This completes step i1. All the necessary conditions are satisfied. It is clear that we can successively complete stage i for i ¯ 1, … , A®8A"/#7. Now the number of colours used at stage i is k}AO(n$/)), hence in total we have used at most kO(n$/)), which is at most (1η) (2m)"/# provided that n is large enough. Next we have to consider the vertices recoloured at each stage, after the colouring of the hypergraph H. It is these vertices only which may cause the colouring not to be harmonious. The number of colours used to colour H is at most (1δ) D(H ), where D(H ) ¯ k}(A®1)O(n$/)). Hence the number of colours discarded at stage i is D(H )®b, which is (1δ) k}(A®1)®k}AO(n$/)) ¯ δk}(A®1)k}A(A®1)O(n$/)). Now any colour used on H can be used on at most V(H )}d edges, and rV(H )r ¯ ik}(A®1)(A®i®1) k}(A®1)O(n$/)) ¯ kO(n$/)). Hence each discarded colour is used on at most k}dO(n$/)) vertices, so the total number of vertices recoloured at stage i is at most δk#}d(A®1)k#}dA(A®1)O(n(/)) ¯ δn}(A®1)n}A(A®1)O(n(/)). Thus, since there are fewer than A®1 stages, in all at most δnn}AO(n(/)) vertices are recoloured. Since δ ¯ η}6, and A & 6}η, this is less than ηn}2, provided that n is large enough. We discard the colours of these vertices, so that these vertices become uncoloured. Finally we have to consider the vertices which are never coloured at all. After the last stage i, where A®i ¯ 8A"/#7, we have rZ r ¯ (A®i) n}AO(n$/%) ¯ n}A"/#O(n$/%). Since A"/# & 3}η, this is less than ηn}2, provided that n is large enough. Hence we have coloured at least (1®η) n of the vertices with at most (1η) (2m)"/# colours, as required. The next lemma shows how to extend the incomplete colouring to the whole graph without using too many extra colours. Essentially this lemma appeared in [2, Theorem 3.5] and [1, Theorem 2.1] ; we restate it here for convenience as a stand-alone lemma. L 2.3. Let d be a fixed integer, and let η " 0. Suppose that G is a graph with maximum degree at most d, n ertices and m edges, and G has an incomplete harmonious colouring with k colours such that at least (1®η) n of the ertices are coloured. Then h(G) % k12d #η"/#n"/#, proided that n is large enough. Downloaded from https:/www.cambridge.org/core. IP address: 88.99.165.207, on 12 Jul 2017 at 18:09:43, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms. https:/www.cambridge.org/core/product/ACFF871550A2451AA6A7474CEA7A1385 443 Proof. Denote the set of uncoloured vertices by S, so that rS r % ηn. Now set N(S ) ¯ ² r (, x) ` E for some x ` S ´cS, N #(S ) ¯ ² r (, x) ` E for some x ` N(S )´c(SeN(S )). Note that rN(S )r % drS r % dηn, and rN #(S )r % (d®1) rN(S )r % d #ηn. We now extend the harmonious colouring to S in three stages. Stage 1. We ensure that no colour occurs more than 8η"/#n"/#7 times on N #(S ), as follows. If colour c occurs more than 8η"/#n"/#7 times then pick 8η"/#n"/#7 of the vertices with colour c and recolour them with a new colour not previously used for any vertex. Repeat this until no colour occurs more than 8η"/#n"/#7 times. Now since each new colour occurs exactly 8η"/#n"/#7 times, and rN #(S )r % d #ηn, then we have used at most d #η"/#n"/# new colours. Stage 2. Now recolour N(S ) with new colours not previously used. We shall ensure that after this has been done, we have a partial harmonious colouring of G so that in particular, for each ` S, no two neighbours of have the same colour. We colour N(S ) sequentially. Order the vertices arbitrarily. We ensure that no colour set on N(S ) ever has size greater than 8η"/#n"/#7. Thus when we colour an element ` N(S ) some colours are unavailable. (a) The colours on the coloured neighbours of are unavailable. This excludes at most d colours. (b) If w ` N #(S )eN(S ) is adjacent to and is already coloured, and w« ` N #(S )eN(S ) is the same colour as w, then we cannot use the colour of any neighbour of w«. There are at most d choices for w, then at most 8η"/#n"/#7 for w«, and then at most d for a neighbour of w«. Hence this excludes at most d #8η"/#n"/#7 colours. (c) We must ensure that no uncoloured vertex in N(S ) and no vertex in S gets two neighbours in N(S ) of the same colour. This excludes at most d(d®1) colours. (d) We must exclude any colour which has already occurred 8η"/#n"/#7 times on N(S ) ; there are at most rN(S )r}8η"/#n"/#7 such colours. So we exclude at most dη"/#n"/#. Hence in total at most d #8η"/#n"/#7dη"/#n"/#d(d®1)d colours are excluded, so we can colour N(S ) with at most d #8η"/#n"/#7dη"/#n"/#d #1 colours. Stage 3. We colour the elements of S sequentially, with brand new colours, never allowing any colour set to become larger than 8η"/#n"/#7. For reasons as in (a) to (d) above, we have to exclude at most d #8η"/#n"/#7η"/#n"/#d # colours when we colour each vertex, so we can complete the colouring with at most d #8η"/#n"/#7 η"/#n"/#d #1 new colours. Hence h(G) % k3d #8η"/#n"/#7(d1) η"/#n"/#2d #2. It follows that if η"/#n"/# & 1, we have h(G) % k12d #η"/#n"/#. We can now give the main theorem for regular graphs. T 2.4. Let d be a fixed integer, and let ε " 0. Then there is an integer N(d, ε) such that if G is a d-regular graph with n ertices and m (¯ "dn) edges, where # m & N, then h(G) % (1ε) (2m)"/#. Downloaded from https:/www.cambridge.org/core. IP address: 88.99.165.207, on 12 Jul 2017 at 18:09:43, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms. https:/www.cambridge.org/core/product/ACFF871550A2451AA6A7474CEA7A1385 444 Proof. Choose η so that η % ε}2 amd η % ε#}576d $. Then by Lemma 2.2, provided that m is large enough, we can find an incomplete harmonious colouring of G with at most (1η) (2m)"/# colours, and with at most ηn vertices uncoloured. Then by Lemma 2.3, provided that m, and hence n, is large enough, we can complete the colouring with the addition of at most 12d #η"/#n"/# new colours. Hence h(G) % (1η) (2m)"/#12d #η"/#n"/# % (1ε}2) (2m)"/#12d #εn"/#}24d $/# % (1ε}2) (2m)"/#ε(2m)"/#}2 % (1ε) (2m)"/#, as required. It remains only to choose N(d, ε) appropriately. 3. Bounded degree graphs We now extend the results of the previous section to graphs of bounded degree. To do this we show how such a graph G can be turned into a regular graph H without either decreasing the harmonious chromatic number or increasing the number of edges too much. Several authors have proved results of the form : for any graph G, with n vertices and maximum degree at most d, we have h(G ) % Cd n"/#, where Cd is some constant depending on d (see for example [2, 3, 4, 6]). For our purposes the most convenient (though not the strongest) is the following result of McDiarmid and Luo [6]. T. For any nontriial graph G with maximum degree d, we hae h(G ) % 2d(n®1)"/#. T 3.1. Let d be a fixed integer, and let ε " 0. Then there is an integer M such that if G is a graph with maximum degree at most d, and G has m edges, ! where m & M , then ! h(G ) % (1ε) (2m)"/#. Proof. Isolated vertices make no difference to the hypothesis or conclusion, so assume that G has no isolated vertices. Let the number of vertices of G be n, and let k ¯ 2dn"/#. Note that m & n}2. From the above theorem of McDiarmid and Luo, we can find a harmonious colouring c : V(G ) ! ²1, … , k´. Define the degree sum σ(i) of a colour i to be 3c(v)=i d(), where as usual d() is the degree of in G. By recolouring some vertices if necessary, we can ensure that for any colours i, j, we have σ(i) % dσ( j). For suppose that σ(i) " dσ( j). Let the set of colours which occur adjacent to a vertex of colour j be S, so that rSr ¯ σ( j). Then the number of vertices of colour i is at least σ(i)}d " rS r, hence at least one vertex with colour i has no neighbour with a colour from S. Hence we can recolour this vertex with colour j. We repeat this process until σ(i) % dσ( j) for all pairs of colours. Let σmin be the minimum degree sum. Now observe that 3ki= σ(i) ¯ 2m and σ(i) % dσmin for each i, hence 2m % kdσmin ¯ " 2d #n"/#σmin % 2d #(2m)"/# σmin from which we obtain σmin & (2m)"/#}2d #. Now choose L to be the least even integer satisfying L & 16d}ε2d. We consider each colour class, and partition each into sets of vertices of total degree between L®2d and L®d. We may be left with a residue of total degree less than L®2d, in this case we add at most one of these extra vertices to each part. (There will certainly be enough parts provided that m & M say.) Thus we obtain partitions of each colour class, with each part of " Downloaded from https:/www.cambridge.org/core. IP address: 88.99.165.207, on 12 Jul 2017 at 18:09:43, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms. https:/www.cambridge.org/core/product/ACFF871550A2451AA6A7474CEA7A1385 445 total degree between L®2d and L. We now take each part, and identify its vertices to form a single vertex of degree between L®2d and L. It is easy to see that we do indeed obtain a graph G « by this process, that is, no loops or multiple edges are formed. Furthermore, any harmonious colouring of G « immediately induces one of G in the obvious way, hence h(G ) % h(G «). We now form from G « a graph H which is regular of degree L. We do this as follows. If G « has at least L vertices of degree less than L, add a new vertex and join it to L such vertices. Repeat this until the set U of vertices of G « with degree less than L satisfies rU r ! L. Now add L1 new vertices , , … , L to form a copy of KL+ " ! " disjoint from G «. Let R be the set of vertices u of U for which L®d(u) is odd, and let r ¯ rRr. Then r is necessarily even, for 3 L®d(u) ¯ 3 L®d() ¯ LrV r®2m. u`U v`V Since L is even, the result follows. Remove the edges ( , ), … ( , r/ ), and join ! " ! # ! to r}2 of the elements of R, and each of , … , r/ to one of the remaining elements " # of R. Now L®d(u) is even for each element of U. The vertices , … , L induce a copy " of KL, which contains L®1 disjoint matchings of size L}2. For each element u of U, select one of these matchings, remove (L®d(u))}2 of the edges, and join the endpoints of these removed edges to u. This forms a graph H regular of degree L. We must determine how many edges H has. Apart from the edges of KL+ , one end " of every new edge is a vertex of G «. Hence the number of new edges is at most L1 % 2d rV(G «)r0 . 0L1 1 2 2 1 3 L®dG «() v ` V(G «) But 2m & (L®2d ) rV(G «)r, hence since L & (16d}ε)2d, the number of new edges is at most L1 L1 4dm}(L®2d ) % εm}4 . 2 2 0 1 0 1 Then if m is large enough, say m & M it follows that rE(H )r % (1ε}2) m. Hence, # provided that m & N(L, ε}2), by Theorem 2.4 we have h(H ) % (1ε}2) (1ε}2)"/# (2m)"/# % (1ε) (2m)"/#. Since h(G ) % h(H ), it remains only to choose M ¯ max ²M , M , N(L, ε}2)´. This ! " # completes the proof. It is perhaps worth presenting an alternative formulation of this result. If we identify two distinct vertices , w of a simple graph G, then we shall obtain a loop if they are adjacent and a multiple edge if they have a common neighbour. However if and w are at distance at least 3 in G, then we shall still have a simple graph after identifying and w. Let us say that G can be collapsed to H if the graph H can be obtained from G by a sequence of identifications of pairs of vertices distance at least 3 apart. It is easy to see that h(G ) is the smallest number of vertices in a graph H to which G may be collapsed. Also, G can be collapsed to H if and only if H is a detachment of G (see for example [7]). Define the density of a graph G as rV(G )r rE(G )r} . Then Theorem 3.1 may be stated as follows. 2 0 1 Downloaded from https:/www.cambridge.org/core. IP address: 88.99.165.207, on 12 Jul 2017 at 18:09:43, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms. https:/www.cambridge.org/core/product/ACFF871550A2451AA6A7474CEA7A1385 446 T 3.2. Let d be a fixed integer, and let ε " 0. Let G be a graph with maximum degree at most d. Then if G has sufficiently many edges, it can be collapsed to a graph of density at least 1®ε. 4. Locally injectie harmonious colourings Recall that in the Introduction we defined, for any graph K, the class Γ(K ) of graphs consisting of a finite number of disjoint copies of K, and mentioned that, by the theorem of Wilson [10], we have h(G ) ¯ (2m)"/#o(m"/#) for G ` Γ(K ). (In fact Wilson’s Theorem actually shows in some cases that h(G ) ¯ Q(m) ¯ 8"(1(8m1)"/#7.) These colourings also have the added property of being one-to-one # on each copy of K, that is, within any copy of K, each vertex has a distinct colour. In the situation considered in this paper of general bounded degree graphs, we clearly cannot insist that the harmonious colourings are one-to-one, however a natural generalisation of the property described above is that no two vertices joined by a short path should have the same colour. This motivates the following definition. D. Let r be a positive integer, and G a graph. A harmonious colouring of G is r-locally injectie if, whenever a pair , w of vertices of G is joined by a path of length at most r, then and w have distinct colours. R. By definition, any harmonious colouring is 2-locally injective. By making small changes to the proofs of Lemma 2.2 and Lemma 2.3, we can prove the following. T 4.1. Let d, r be fixed integers, and let ε " 0. Then there is an integer M ! such that if G is a graph with maximum degree at most d, and G has m edges, where m & M , then there is an r-locally injectie harmonious colouring of G with at most ! (1ε) (2m)"/# colours. In the case covered by Wilson’s Theorem above, we can take r to be one greater than the diameter of K to obtain a colouring which is one-to-one on each copy of K. Thus Theorem 4.1 gives a weak form of the conclusions of Wilson’s Theorem, but from considerably weaker hypotheses. 5. Concluding remarks Let us recall the classes Γd of graphs of maximum degree d, and Γ(K ) of graphs consisting of disjoint copies of a fixed graph K. We have seen that h(G ) ¯ (2m)"/#o(m"/#) for G ` Γd. There are two natural ways in which this might be strengthened, namely a strengthening of the conclusions or a weakening of the conditions. The first possibility is that the term o(m"/#) might be reduced, perhaps to a constant. Let K be a d-regular graph, then the elements of Γ(K ) are of course also d-regular. It follows from Wilson’s Theorem [10] that if G is a graph Γ(K ) with C m¯ edges, and d divides C®1, then h(G ) ¯ C ¯ Q(m), provided that m is large 2 01 Downloaded from https:/www.cambridge.org/core. IP address: 88.99.165.207, on 12 Jul 2017 at 18:09:43, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms. https:/www.cambridge.org/core/product/ACFF871550A2451AA6A7474CEA7A1385 447 enough. It seems conceivable that such a result could hold for d-regular graphs in general. More generally, it may be that for graphs in Γd, we have h(G ) % (2m)"/#Cd, where Cd is a constant. On the other hand, it is natural to ask if the condition of bounded degree is necessary. It appears that it may be, for Alon [3] proved by random methods that for 400 log (2n) % d % (n log (2n))"/#, there is a graph G with n vertices and maximum degree at most d for which h(G ) " dn"/# d "/#(2m)"/# & , / 100 (log n)" # 100 (log n)"/# so that h(G )}(2m)"/# ! 1 does not hold even for classes of graphs with quite slowly growing degree. References 1. K. J. E, ‘ The harmonious chromatic number of almost all trees ’, Combinatorics, Probability and Computing 4 (1995) 31–46. 2. K. J. E and C. J. H. MD, ‘ New upper bounds on harmonious colorings ’, J. Graph Theory 18 (1994) 257–267. 3. I. K Y. R, ‘ Bounds for the harmonious chromatic number of a graph ’, J. Graph Theory 18 (1994) 205–209. 4. Z. L, ‘ On an upper bound for the harmonious chromatic number of a graph ’, J. Graph Theory 15 (1991) 345–347. 5. C. J. H. MD, ‘ On the method of bounded differences ’, Sureys in combinatorics 1989 (ed. J. Siemons ; Cambridge University Press, 1989) 148–188. 6. C. J. H. MD and L X, ‘ Upper bounds for harmonious colourings ’, J. Graph Theory, 15 (1991) 629–636. 7. C. S. J. A. N-W, ‘ Detachments of graphs and generalised Euler trails ’, Sureys in combinatorics 1985 (ed. I. Anderson ; Cambridge University Press, 1985) 137–151. 8. N. P and J. S, ‘Asymptotic behavior of the chromatic index for hypergraphs ’, J. Comb. Theory Ser. A 51 (1989) 24–42. 9. B. W, ‘ Line distinguishing and harmonious colourings ’, Graph colourings, Pitman Research Notes in Mathematics 218 (eds., R. Nelson and R. J. Wilson ; Longman Scientific & Technical, Essex, 1990). 10. R. M. W, ‘ Decomposition of complete graphs into subgraphs isomorphic to a given graph ’, Proceedings of the 5th British Combinatorial Conference (Aberdeen, 1975), Congr. Numer. XV (1976) 647–659. Department of Mathematics and Computer Science University of Dundee Dundee DD1 4HN Downloaded from https:/www.cambridge.org/core. IP address: 88.99.165.207, on 12 Jul 2017 at 18:09:43, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms. https:/www.cambridge.org/core/product/ACFF871550A2451AA6A7474CEA7A1385
© Copyright 2026 Paperzz