You are on page 1of 55

Graph Theory

Prof. A.G. Thomason Michaelmas 2005

L T Xed by Sebastian Pancratz E

A

ii
These notes are based on a course of lectures given by Prof. A.G. Thomason in Part II of the Mathematical Tripos at the University of Cambridge in the academic year 20052006. These notes have not been checked by Prof. A.G. Thomason and should not be regarded as ocial notes for the course. In particular, the responsibility for any errors is mine  please email Sebastian Pancratz (sfp25) with any comments or corrections.

Contents

1 Basic Denitions and Properties 2 Matchings and Connectivity 3 Extremal Graph Theory 4 Colouring 5 Ramsey Theory 6 Probabilistic Methods 7 Eigenvalue Methods

1 9 15 23 35 39 47

Chapter 1 Basic Denitions and Properties
Denition. A graph is a pair G = (V, E) where E ⊂ V (2) = {{x, y} ⊂ V : x = y}. Example. V = {a, b, c, d}, E = {{a, b}, {a, c}, {a, d}, {b, c}}. Informally, this graph can
be represented as follows.

d

a

•           • • •

c

b

Denition.
set of

The set V is the set of vertices of G, also denoted V (G). The set E is the edges of G, also denoted E(G). All graphs considered here are nite. |V | is called the order of G, often written |G|. |E| is called the size of G, often written e(G). Our graphs have no multiple edges, but some authors allow these and call our graphs

simple.

y • yy

yyy yyy yyy y •

Denition.
such that

Two graphs

G, H

are

isomorphic
n

if there is a bijection

ψ : V (G) → V (H)

ab ∈ E(G)
(i)

if and only if

ψ(a)ψ(b) ∈ E(H).
vertices and no edges, called the empty graph of with

Example.
order (ii) (iii) (iv)

n,

e.g.

En , the E3 .

graph with

Kn , the complete graph of order n Pn , the path of order n, e.g. P5 . Cn , the cycle of order n, e.g. C5 .

E(Kn ) = V (2) ,

e.g.

K4 .

• ÔXX ÔÔ XXX Ô ÔÔ l‚‚ XXX • X ÔÔ lll ‚‚‚‚‚ Ôl l •l •

r • vv rrr rr vv v rr vv rr vv rr v vv A •A • ! AA !! AA ! ! AA !! AA !! AA !! !
• •

• kkG kkk GGG kkk GG • • oo  Gooooo •   

is minimal connected. the addition of any new edge creates a cycle. uv forms a cycle in G.2. [(ii) =⇒ (i)] Suppose G is connected and has a cycle C .1. Note that the only induced subgraphs of Kn are Kk for k ≤ n. If uv ∈ E(G) then G + uv has a cycle containing uv . T is a tree. Let x. so G − uv is connected. i. the removal of any edge destroys connectivity. G is a disjoint union of its components. If uv ∈ E(G) there is a u−v path in G which with uv forms =⇒ (i)] G is maximal acyclic. v ∈ V (G) there is a of G is a maximal connected subgraph. So if G is acyclic and connected. that is. is maximal acyclic. By Theorem 1. a tree. Hence G is connected. i.e. and it is necessarily induced. A Denition. Proof. Conversely. so G has a u − v path. let Since T is connected and spanning. subgraph A graph T such that G is connected if and only if V (T ) = V (G) and T is a tree. T be a minimal connected spanning subgraph. if G is connected. tree is a connected acyclic graph. every such graph is Kn . it has a spanning tree. say. Denition. Then P − uv together with C − uv still contains an x − y path. .1.) Kn . Let uv be an edge of the cycle. n is a subgraph of Every graph of order at most isomorphic to a subgraph of graphs. A forest is an acyclic graph. (i) (ii) (iii) The following are equivalent. [(i) a =⇒ (iii)] G is cycle in G + uv . y ∈ G be vertices such that some x − y path P in G contains uv . but we will not distinguish between distinct isomorphic H is an induced subgraph of G if V (H) ⊂ V (G) and E(H) = V (2) (H) ∩ E(G). The components of a forest are trees. A G from u to v (called a component u. Hence if G is minimal connected then G is acyclic (and connected).2 Basic Denitions and Properties Note that paths and cycles do not allow repetitions of vertices. Then there is a Suppose u−v path P in G − uv which with uv ∈ E(G) and G − uv is still connected. G is connected. Denition. a Proof. H is a subgraph of G if H is a graph with V (H) ⊂ V (G) and E(H) ⊂ E(G). G G G is a tree. (Formally. [(iii) Corollary 1. it is minimal connected. Denition. 6 W• • W 666 WW6 ‰ • • ‰‰F FF FF FF xx FF xx FF xx cx x • c c  cc   • • • cc & • • c && o G o & •Go G c •  • c xx • cc• x • xx Theorem 1. [(i) =⇒ (ii)] G is connected. If W ⊂ V (G) we write G[W ] for the induced subgraph with vertex set W . path in A graph G is connected u−v if for every pair of vertices path).e. and acyclic by assumption.

else there exists xk u ∈ E(G) where u = xk−1 and by maximality of the length of the path. the maximum degree is ∆(G). Note that this is the best possible bound by considering a path. How many graphs of order gure illustrates the case there? Take the vertex set [n] = {1.3 neighbours of a vertex v is denoted Γ(v) = {w ∈ V (G) : vw ∈ E(G)}.5.4. Denition. so k = 1. The next 1 • 2 • • 3 2 • • ÔÔ ÔÔ ÔÔ ÔÔ Ô 1 1 X •X • 3 2 • XX XX XX X 1 • • 3 2 • • 3 . . of order is acyclic.1. A leaf is a vertex of degree one. By Theorem 1. (i) =⇒ (iii)] This follows from the denition and Corollary 1. Lemma 1. T be a tree and let x1 be a vertex. Let A tree of order at least 2 has at least 2 leaves. . (i) (ii) (iii) The following are equivalent. n≤2 are trivial. n are n = 3. The graph T − v has order n − 1. an edge vw with w ∈ Γ(v) is incident to v . xk be a path of maximal length. of The minimum degree G is δ(G). d(v) = 2e(G) v∈V Denition. . G is a tree. G. G G G is a tree of order n. Theorem 1. (The cases T be a tree n ≥ 3. Since T is connected. Proof. n − 1. [(ii) [(iii) =⇒ (i)] G contains a spanning tree T. u = xi for some i ≤ k − 2.) So T − v n − 2 by induction. Proof. Let x1 x2 . n and size is connected. A vertex w ∈ Γ(v) is adjacent to v . . Then xk is a leaf.5. contradicting that T is acyclic.) Let We prove this by induction on n. we have δ(T ) ≥ 1 (since |G| ≥ 2).6. . then v = x or v = y . which we choose to be a leaf if there is one. so G = G. The degrees of G written in some order form a degree sequence. . T acyclic. (If is a tree of order v lies on an n − 1.5. acyclic graph e(T ) = n − 1. [(i) =⇒ (ii). and connected.3 (Handshaking Lemma). The degree of v is d(v) = |Γ(v)|. so has size v . The set of Denition. Corollary 1. By Corollary 1. By Theorem 1. Now =⇒ (i)] Add edges to get a maximum e(G ) = n − 1. of order A tree of order n has size n − 1. has a leaf Corollary 1. is x − y path in T . of order n and size n − 1. Proof. n}. so T = G.4. So we can assume x1 is a leaf to obtain a second leaf xk .

n To count unlabelled graphs. .5.9. gf = ι. Remove the leaf. Given T .10. the result is a tree. → Strings] Choose the smallest labelled leaf. if b ∈ {1. .8. n} is minimal not in a. we obtain the string 3. join the two last available vertices.11. lemma. . for an action of a group 2( 2 ) .4. 3 4 4 11 5 34 order no. hence f is bijective. an−2 ). Amusingly. . of xed points of g ).5. Proof (Prüfer).8. .5.4 • ÔXX ÔÔ XXX Ô XX ÔÔ XX ÔÔ Ô 1 1 1 X •X Basic Denitions and Properties 2 • • 3 2 • • ÔÔ ÔÔ ÔÔ ÔÔ Ô • 3 2 • XX XX XX X • 3 2 • • ÔXX ÔÔ XXX Ô XX ÔÔ XX ÔÔ Ô 1 • 3 The number of labelled graphs.7 (Caley). G-orbits) × |G| = g∈G (no. Since f (T − b) = (a2 .6. . Continuiung the previous example.7. an ). and nally join 3 and 11.6. of (unlabelled) graphs How many trees of order n are there? 2 1 3 3 4 16 5 125 order no. .8.8. Repeat until one edge is left. Observe that this process produces an acyclic graph of size so by Corollary 1. . (d(v) − 1) = v v d(v) − n = 2(n − 1) − n = n − 2. g : Strings → Trees. When nished. (no. where we can distinguish vertices by name.2. is belled graphs under the action of some permutation group. let Remark.8.3.3 and mark the vertices 1.2.3. This states that. . an ) = T − b. [Strings → Trees] Note that a vertex v appears d(v) − 1 times. . 7 c •c 10 cc cc cc c •  5    • 1 D •D • 11 DD DD D3 • ! !! !! !! ! •   i 6  i•  i ii Q • Qi 8 QQ QQ QQ • •9 4 2 Deleting 1. Delete the rst number and repeat. and b is the smallest labelled leaf of g(a). .9. f : Trees → Strings. .8.7. . . . By removing the smallest leaf. we similarly obtain f g = ι so f is surjective.3. Mark it unavailable and join it to the rst number in the sequence. Write down its neighbour.5. More formally. so g(a) = T . . . we start with the string 3. by induction g(a2 . Consider the following tree as an example. hence f is injective. then g(a) has an edge ba1 .5 as unavailable.8. n − 1.11.6. we need to count the number of orbits in the set of laWe can use Burnside's of G.5.4. We construct a bijection between the set of labelled trees and the set of n−2 with alphabet [n]. strings of length [Trees There are nn−2 labelled trees of order n. Take the least available  vertex not in the sequence. let f (T ) = a = (a1 . of (labelled) trees Theorem 1.10.8.

classes. Proof. A graph is r-partite edge lies within a class. by Vi . Trace back paths of length i from v0 . the result is clear. Remark. Denition. Y = i odd Vi gives a bipartition. A graph is planar if it can be drawn in the plane without edges crossing.e.e. The rst time they meet together with the edge uv we obtain Since G has no odd cycle. bipartite graphs. A graph is bipartite if and only if it has no odd cycles. the length … r ‡ i • ‡‡‡ • ………… nnn rrrr‡‡‡‡‡‡‡ …iiiii ‡‡‡‡iiiiii……………… nnn rr n • nn rr rrr • rr nnn ddddd rr dd rrr d nnn dddddd d € d d n • €  d rr • rrrr v0 €€€€  ddddddddddd r rr rrr d €€€ • r i • €€€ iiii €€€ iiii € i i • •i V0 V1 V2 . G is not the empty graph. edges lie within classes denition of Vi and between Vi and Vi+1 but nowehere else. The conditions imply (since the order of taking in sub-eulerian tours when rst encountered. We prove the converse by induction on the δ(G) ≥ 2.5 Denition. Bipartitite if its vertex set can be partitioned into means r classes so no 2-partite.. and d(u. each having an eulerian tour by induction. v) is the distance from u to v . we can characterise Theorem 1. eulerian tour 0) of a graph is a walk along edges covering each edge exactly once and returning to the start. even. Go around C . . w) = i} the shortest u − v path..4 shows that there is a cycle C G is at least 2 and G is connected. Remark. Remarkably. A graph is vertices with degree eulerian if it has no isolated vertices (i. we may assume let of G is connected by considering components. The result is where trivial for the empty graph.8. i. If a graph G is bipartite the result is clear since cycles alternate between the two Conversely. G is eulerian if and only if |G| > 1. if G is acyclic there exists a leaf. so Theorem 1. uv say. v to Suppose there is an edge inside some each of u and an odd cycle. So X = i even Vi . This gives an algorithm for An 2-colouring graphs with no odd cycle. ‰‰‰‰‰‰ ‰‰‰‰‰‰ • eeee eeeeee ee • rr ‰‰‰‰‰‰ rrrr ‰ rrr‰‰‰‰‰ • rrr rr Vn In general. require the existence of exactly two vertices of odd degree. Vi . G is connected and all degrees are Proof. Denition. and has an eulerian tour. G − E(C) has all components isolated vertices or connected even degree components. so suppose Vi = {w ∈ G : d(v0 . G. when starting and nishing at distinct vertices. Analogeously. Theorem 1. Pick a vertex v0 . contradicting δ(G) ≥ 2).9. each Vi contains no edges. size of If G is eulerian.

Prove this by induction on minimal connected and have size m ≥ n. pick an edge e in some since e lies between two faces. Let G be a graph with d(v) even for all v ∈ V (G). not F on the other. components. all edges of (Euler) H are in cycles. Then E(G) can be partioned into cycles. so f = 1. so v in a dH (v) small loop.6 There are no essential diculties of analysis here. Let H F be the spanning subgraph of G consisting of all edges h with F on one side. i.11. a vertex of degree If there is no cycle then G contradiction. then e is not a bridge and lies G is a bridgeless plane graph i ifi = 2m. Connectivity implies m ≥ n − 1 as trees are n − 1.12 faces then . Note here that the ∞-face has a well-dened length for a bridgeless graph. we take the Jordan Indeed. m. A face is a simply- connected region. G − e has n vertices. Proof. Example. Going around a vertex the same number of times. contains a tree and hence a leaf. (See Example Sheet 1. the edge does not lie in a cycle. m − 1 edges and f − 1 faces Proof. A plane graph is a planar graph drawn in the plane.10. 1. • QQ  QQQ 4  2 Q  1rrvvv QQQ •  rr 3 vvvQQ v rrr  • • K4 has four faces. Thus if with fi faces of length i. Observe that if in a plane graph then it borders two distinct faces. .e. including the ∞-face. If G is a connected plane graph of order n and size m with f n − m + f = 2. By Lemma 1. we observe that we enter and leave is even. If G contains a cycle we can remove it and note that the all degrees remain even. F be one of the faces with e in its boundary. Proof. Then e is in a boundary of two faces if and only if there exist a cycle C in G containing e. Theorem 1. Denition. If m = n − 1 then G is a tree. If cycle. The drawing of C is a simple polygonal closed curve.10.) • cc  cc  cc  c cc  ccc cc  c   • • c •c • QQ  QQQ  QQ  Q •  rrrvvvvvQQ  vv Q rrr r  • • Denition. it can be shown that any such drawing can be done with straight line edges. separating the plane. Lemma 1. Let Let e be an edge of a plane graph G. Basic Denitions and Properties In particular. Curve Theorem for granted in this course. Lemma 1. A bridge in a graph is an edge whose removal increases the number of Equivalently. We may ignore isolated vertices and assume d(v) ≥ 2.

G − ab has two vertex- disjoint subgraphs satisfying the formula by induction. If G contains a bridge g . this holds for any planar graph G with n ≥ 3. n satises e(G) ≤ 3n − 6.3 ) = 9 > 2 (6 − 2) g(K3. no subdivision of K5 is planar.3 is non-planar. Draw Proof. The complete bipartite graph Kp. ≥ 2. so K3. The girth g(G) of a graph G is the length of the shortest cycle. Denition.12. ♠ h qqq qhq hhhh q hhhhqqqq qqq hhh qqq hhhh qqqq qqq hhhhhhhh qq h qqq qqq hhhh qqq h h cqq h cqq h h qh q c c c c  c  c  c G W E Observe that having too many edges is not the only reason graphs fail to be planar. or ∞ is G Let Theorem 1. vertices in the other.3 ) = 4. as can be seen on replacing edges by disjoing . In particular. a planar graph of order n and girth g .7 Denition.13.q pq as is bipartite with p vertices in one class. either by induction or by adding edges till G is bridgeless connected. i. Then e(G) = e(G1 ) + e(G2 ) + 1 g ≤ [(n1 − 2) + (n2 − 2)] + 1 g−2 g g (n − 2) + 1 − 2 = g−2 g−2 g ≤ (n − 2). g−2 Example. For example. q Note e(K5 ) = 10 > 3(5 − 2). Adding edges leads to problems since we may decrease we have solved the bridgeless case. G in the plane. so K5 is non-planar. Then G be a connected bridgeless planar graph of order e(G) ≤ g g−2 (n − 2). Then 2m = i ifi ≥ g i fi = gf . 4 e(K3. if G is a bridgeless connected planar graph then e ≤ 3n − 6. Example. m(g − 2) ≤ g(n − 2).e. n−m+ 2m g Remark. is acyclic. In particular. By Theorem 1. note that ab. Instead. and all possible edges between them.

.. c  . .8 A • !r vv A rr vv !! AA •rrr •r vv r • rr vv !! AA AA ! vv •rr ! vr AA rv • r • AA ! vv !! AA rr ! AA vvvv !! rr ! AA ! rr v A ! AA !! rrr vvvvAA !!! r AA !! vvv rr AA !! rr A ! AA !! vvv rr ! A! vv ! • • Basic Denitions and Properties paths. . . .. . .. ..... •           • • . .. . . . . . . Proof omitted. . . Given a plane graph Denition. but these may satisfy the conclusion of Theorem 1.... Place a dual vertex inside each face. . BB . . .. . 11 edges. . cc „„„ .. . .. . . . . . .. . . .. . . . . ♦. ... .. . But if G is a 3-connected simple graph then so is G∗ .c . . . for each edge. Remark. Theorem 1... ... .. . .. . .. . 6 faces and the dual has 6 vertices. BB . ... . . A graph is planar if and only if it contains no K5 K3. . . .. .  . . . . •           • • . . . BB . . .. BB .... . cc . . . .....14 subdivision of (Kuratowski. . .. c BB . • c „„„ ... BB . . . .. the original graph has 7 vertices. .. . . B . .• . . • • ..... .  B .. . . .. . ..♦. .3 . . Example.. . . . the dual might not be simple. . . . ...... . . . and. .. . G. In this example. we can construct the dual graph G∗ . . . . . .13 with suciently long paths... . ccB  . . draw a dual edge joining the corresponding dual vertices. . ... . Remarkably.♦ . . c„ „ .... •B . . . .. plane graphs can be characterised. . . . • . . .. .. . .. .. •.. Note that usually the dual of the dual is the original graph. . ... . Proof. ♦. .. . .„„„„ . . . . 7 faces. . . . . ♦ B. . . . .. . .. . . .. ♦ . cc. . . 1930) or of ..... ... . 11 edges.. . ... If a graph is not 3-connected (see Chapter 2). „„„„ cc . . . .. . . . . cc . . . ♦ ♦ . . ... ... . .... . c B . . . .. ... . ... ... . .. . „„„B .  .

e.Chapter 2 Matchings and Connectivity Denition. For A ⊂ X dene Γ(A) = x∈A Γ(x). i. by induction. edges. to G be a bipartite graph with vertex classes X. i. |X|. If for every G = G − xy together with xy gives a and in . every set of k men needs to know at least k women. The matching in G. and if following gure illustrates cases in which this fails. where n-regular means every vertex has degree n. Y . The a set of men. a set of women. X Y xy ∈ E(G) P •P PP PP hh PPP • • hh P hh PP hh PP h • • • • P •P •  PP  PP   h P • h PP • hh P  h PP   hhhh   PP  • •      • • P •P PP  PP   h P • • h PP z hh P z  hhz zz P  PPP zz hhh zz  •  •     • • • Then Theorem 2. We need |Y | ≥ |X| and that every man knows a woman.1 (Hall). A matching from X |X| independent. In general. a matching from Let G be a bipartite graph with vertex classes X. ∅=A The necessity is clear. pairwise non-incident. We prove suciency by induction on X we have |Γ(A)| > |A|. Clearly we need |Γ(A)| ≥ |A| for all A ⊂ X . Y . If |X| = |Y |. also called a 1-factor. Finding a matching from X to Y is to nd wives for all men. then pick any edge xy ∈ E(G) Hall's condition holds. Let Y is a set of • • • ——————————— — • —————— • “ • “““““““““““““““““ • ” • ””””””””””””””””” • X Consider Y x can marry y . a 1-regular spanning subgraph.e. G has X to Y if and only if ∀A ⊂ X |Γ(A)| ≥ |A| (HC) Proof 1. this is also a matching from Y to X .

So (HC) holds in G2 . Here we have used (HC) on second inequality. hence by (HC) in G |Γ1 (A)| = |Γ(A)| ≥ |A|. bi = Γ(Ai ) and Γ(Ai ∪ {a}) = Γ(Ai ) ∪ {bi }. G1 . Otherwise. For so (HC) holds in A ⊂ X − B. there exists a critical set ∅ = B G1 = G[B ∪ Γ(B)]. violating (HC). a ∈ X joined to b1 . Consider a minimal (with respect to removing edges) subgraph in which (HC) d(x) = 1 for all x ∈ X then what is left is a matching. for i = 1. G1 B X such that |Γ(B)| = |B|. Thus |Γ(A1 ∪ A2 ∪ {a})| = |Γ(A1 ∪ A2 )| = |Γ(A1 ) ∪ Γ(A2 )| = |Γ(A1 )| + |Γ(A2 )| − |Γ(A1 ) ∩ Γ(A2 )| ≤ |Γ(A1 )| + |Γ(A2 )| − |Γ(A1 ∩ A2 )| = |A1 | + |A2 | − |Γ(A1 ∩ A2 )| ≤ |A1 | + |A2 | − |A1 ∩ A2 | = |A1 ∪ A2 | = |A1 ∪ A2 ∪ {a}| − 1.10 matching in Let Matchings and Connectivity G. G2 = G[(X − B) ∪ (Y − Γ(B))]. A2 ⊂ X − {a} such that |Γ(Ai )| = |Ai |. Proof 2. there exists A1 • hhhh hhhh b2 hhhh † h†h a • h††† †††† †††† †† b • 1 Γ(A1 ) Hence Γ(A1 ∪ A2 ∪ {a}) = Γ(A1 ∪ A2 ). 2. If in the resulting graph If not. • • • ••••••••••••••••••• –––– • – • ––––––––––––––––––– • ––––––––––– mmmm – • –––– • m n mmm nnn mmmnnnn mm n mmm nnn mmmnnnn m m • mnnnn • ’’ n • n’’’’’’’’’’’’’ • ’’’’’’ • • where Γ(B) G2 X−B Y −Γ(B) For A ⊂ B. we have Γ1 (A) = Γ(A). A1 ∩ A2 to establish the . we have |Γ2 (A)| = |Γ(A ∪ B)| − |Γ(B)| = |Γ(A ∪ B)| − |B| ≥ |A ∪ B| − |B| = |A| using (HC) in G to establish the inequality. still holds. Thus G1 . b2 ∈ Y and sets A1 . hence so does G. G 2 both have matchings. Γ1 (A) are the neighbours of A in G1 . a contradiction.

. X = Proof. G) = min{|S| : ∃S ⊂ V (G) − {a. is G is not complete. Yn } and edges from A graph Yi ∈ X to y∈Y if if y ∈ Yi . . Necessity is clear. distinct repre- Corollary 2. and Now apply Theorem 2. b. . . b. i. Tutte's theorem gives a necessary and sucient condition for a necessarily bipartite. Y joined to all members of exists |X| − d d new members of X. (See Example Sheet 1. Yn of subsets of a set Y . b. wives to the original men. . . ab ∈ E(G). κ(G) ≤ λ(G) ≤ δ(G) For the rst inequality. a−b path in G − S}. Now remove the d new vertices. Let independent edges in Introduce G if and only G be as above and d ∈ N. . We can give every man for all A ⊂ X. In the Now remove the clones and give their new society (HC) holds. a. 1-factor in a general. Proof. (Defect form) . b. . so there is a matching. is k -connected |G| > k to be is G−S Denition.11 Corollary 2. . 2 In the new graph (HC) holds. To prove suciency. Replace each man by 2 clones of himself. b} Clearly. G) F ⊂ E(G) is if G is not complete. G) a.e. There if |Γ(A)| ≥ |A| − d for all A ⊂ X . is disconnected}. yn } with Y1 . not set of Denition. no special care is required for complete graphs. analogues where a set κ(G) = minab∈E κ(a. sentatives Given a collection is a set {y1 . so marry o all men. Dene the vertex connectivity κ(G) = max{k : G If k -connected}.2 Proof. . Given κ(G) = min{|S| : ∃S ⊂ V (G) G − S the Denition. . For the second inequality. The following holds.4. wives if and only if Corollary 2. S ⊂ V (G) with G |S| < k . graph. is connected for every set Denition. There are edge connectivity removed. There is a set of distinct representatives if and only if ∀S ⊂ [n] i∈S Yi ≥ |S|. a yi ∈ Yi and yi = yj for i = j .b with no a−b path in G − F} Note that in the case of edge connectivity. b ∈ V (G).) . G) = min{|F | : ∃F ⊂ E(G) λ(G) = min λ(a. local connectivity with no κ(a. is disconnected} λ(G) = min{|F | : ∃F ⊂ E(G) G − F λ(a. remove δ(G) edges from a vertex of least degree.1. . with the same neighbours. remove one endvertex for each edge in our set of whose removal disconnects λ(G) edges G.3 |Γ(A)| ≥ 2|A| (Polygamous version) . construct a bipartite graph with {Y1 .

e. there is a set of a from b. then S = Γ(a) or be the graph obtained by contracting the component S = Γ(b). with ab ∈ E(G). b. and let such that Proof. G). κ(a. to remove at least one vertex from each path. e not meeting G has minimal S ⊂ V − {a. let Ga of G − S containing a to a single a∗ . derived from We use the notion of graph contraction. |Ga | < |G|. . Suppose not. Y be disjoint subsets of V (G) k completely vertex disjoint X − Y paths. e ∈ E(G). we need κ(a. Claim (ii): If vertex and a set of k vertex-disjoint a−b paths in G/e would yield a set in G. Thus there exists a k − b paths in Ga . If Corollary 2. b. b ∈ V (G) paths. |Y | ≥ k .6. let X. G Ga a • ‡ ggggg S ‡‡‡‡‡‡‡‡‡ ggggg ‡‡‡‡‡ ‡‡‡‡‡ ggggg ggggg • b a∗ jjj S jjjj „„„„ jjjj • „„„„ „„„ ‡‡‡‡‡ ‡‡‡‡‡ ggggg ggggg • b S = Γ(a). cutset to be any set order. since to separate Clearly. b} with Claim (i): Every edge a or b lies inside a minimal cutset. If Then there exists a set of vertex-disjoint a−b Denition. Let a. S is a minimal cutset. then there exists an edge (in the cutset) lying neither inside Γ(a) nor in Γ(b). b. But if Γ(a) = Γ(b) we immediately have k a − b paths. b. |S| = k G. paths would yield k vertex-disjoint a − b paths in G. For otherwise.12 Denition. b be a minimal counterexample. Let Then there exists a set of κ(G) ≥ k .5 κ(a. Likewise dene Gb and nd a set of k a − b∗ paths in Gb . the size of any such set is at most paths. i. Since set of These a∗ Γ(a) = Γ(b) then |Γ(a) ∩ Γ(b)| < k . with |X|. G) Theorem 2. G/e) ≥ k contradiction. b. Γ(u) ∪ Γ(v). G) and dene a minimal G − S has no a − b path. contradicting the claims. a. Remarkably. Let k = κ(a. κ(a. v of e introducting a new vertex x joined to G e v •55 •H WWW '' H H  55 WW ' H  55 WW '' HHH  5 WW ' HH 55  WW'' H  ' • • • • u G/e x Bc • BBccc  B c  B c  BB cc  BB ccc   cc B   • • • • A contraction of G is obtained by a sequence of these operations. Clearly κ(a∗ . path are A set of Matchings and Connectivity a−b paths is vertex disjoint if the only vertices in more than one a. Ga ) ≥ k . G) (Menger) . the graph G/e and G by contracting e is obtained by removing both endvertices u. For otherwise. b. b.

so the desired paths exist by Menger's theorem. f are incident edges of G. b . b. L(G) Either we mimic the proof of the vertex form. trivially. Add new vertices x joined to all vertices in X. whose vertex set is the set of edges of G. G) vertex-disjoint a of λ(a. H) = λ(a. or we construct the of line graph if G.) to all vertices in x G. b ∈ V (G) then there exists a set λ(a. If a. to ∗ form G . G). no larger such set exists.5. and an a∗ − b∗ path in H gives a ∗ ∗ set of edges containing an a − b path in G. G) edge-disjoint a − b paths in G. y. so there ∗ − b∗ paths in H by Theorem 2. so there is a set is a set of λ(a. Then κ(x. We have κ(a . b. y joined to all vertices in Y. ((HC) fails if and only if there xists A ⊂ X such that |A| > |Γ(A)| if and only if we can remove (X − A) ∪ Γ(A) of size less than X .7 of (Edge form of Menger's theorem) . path in a∗ − b∗ path in H . Theorem 2. y. G∗ ) ≥ k. G) edge-disjoint a−b paths. Introduce new vertices X and y joined to all vertices in Y to obtain G from Then κ(x. b. and introduce b likewise. G  c  ccc cc   c  c •c •  cc  cc  cc  c • Now form to edges Note an • cc L(G) c •c • cc  cc  cc  c • cc  ccc  cc  cc   • • H from L(G) by joining a new vertex a∗ to all vertices of L(G) that correspond ∗ of G meeting a. Note that. where ef ∈ E(L(G)) e. a−b G gives an Remark (Menger's theorem implies Hall's theorem) . joined .13 Proof. G ) = |X| if and only if (HC) holds in G. b. Proof.

.

e. v1 vl ∈ E(G). Hamiltonian cycle A in a graph is a spanning cycle. |S| = d(v1 ). . so neighbours of v1 . Note that all l < n a path of length l + 1.) • • Kn−1 Q: How large must A: We need δ(G) n 2. . • v1 • v2 • • • • • • • vj • • • vl Let S = {i : v1 vi ∈ E(G)}. T = {i : vl vi−1 ∈ E(G)}. for if j ∈ S ∩ T then v1 v2 . . Hence l − 1 ≥ |S ∪ T | = |S| + |T | = d(v1 ) + d(vl ) ≥ k. since any two non-adjacent Suppose that vertices have a common neighbour. . vl lie in P . vj would be an l -cycle. because if P = v1 v2 . If k < n and G is connected then G has a path of length k . by maximality of P . otherwise we are done. vl be a l = n this would path of maximum length.1. . Let has no G has no Hamiltonian cycle. G Observe that if k = n then G must be connected. . i. |T | = d(vl ). for l-cycle. . vj−1 vl vl−1 . it meets every vertex exactly once.Chapter 3 Extremal Graph Theory Denition. Now S ∪ T ⊂ {2. Proof. . Notice that be a Hamiltonian cycle and if by connectivity there would be d(v1 ) + d(vl ) ≥ k . (See Example Sheet 2. If k = n then G has a Hamiltonian cycle. y satises d(x) + d(y) ≥ k . Let Theorem 3. . be to guarantee a Hamiltonian cycle? δ(G) ≥ Kn • Kn G be a graph of order n ≥ 3 such that every pair of non-adjacent vertices x. l} and S ∩ T = ∅. Q: How many edges are needed to guarantee a Hamiltonian cycle? A: We need more than n 2 − (n − 2) edges. . In particular. .

Then k−1 (n − 1) 2 e(G) ≤ e(G − x) + Denition. It is called the r -partite Turán graph of order n. dene ex(n. If Extremal Graph Theory k < n. The result is easily true for n ≤ k. k|n and G is a disjoint of copies of Kk . if disconnected. remove the vertices at the bottom for (∗∗). connected. By induction on n. G G is Proof. so e(G − x) < by the inductive hypothesis. G is δ(G) ≤ k−1 2 by Theorem 3. changing to class sizes |X| − 1 and |Y | + 1 gains us −|Y | + (|X| − 1) > 0 edges. F ) = max{e(G) : |G| = n. Then Theorem 3.3. Moreover. we are home at once by the hypothesis applied to each component. remainder 3. If G has order n and δ(G) ≥ n then G has a Hamiltonian cycle. Let G be a graph of order n with no path of length k. Y with |X| ≥ |Y | + 2.1. 2 k Remark. at least k . imum size and of order The value of tr (n) can be written explicity in terms of the remainder after n divided by r but this is awkward to work with. It is denoted by Tr (n) and its size is tr (n). Remove the vertex at the top for (∗). Moreover. there is a unique r-partite graph of order n and n n maximum size. If is connected. When working with use some observations about it derived from the Consider the case tr (n). equality holds only if e(G) ≤ k−1 2 n. F ⊂ G}.1 is best possible. if we have two X. • • • • • • . the longest path has length l = k + 1. In general. Since Kk is not a subgraph of G. Let x have degree at most k−1 2 . Note if 2 | n and 2 | n − 1 then Theorem 3.2 (Dirac). Q: How many edges are there in a Observe that classes Kr+1 -free graph? To obtain an r-partite graphs contain no Kr+1 . it is structure of Tr (n). it should be complete r-partite. If so we have a Hamiltonian cycle. r-partitite graph of max- n. k−1 k−1 < n. then P has length Corollary 3. more convenient to r = 5. Hence. •C Cdd ! ! CCddd  !! C d   ! CC ddd dd CC  !!!  dd CC !  dd  CC !! dd  !  dd C !!  Kk/2 Kk/2 Kk/2 Kk/2 In this example. this is impossible.16 k = n. 2 2 Given a xed graph F. The classes have size r or r and it is complete r -partite.

|G| = n and 1 e(G) ≥ tr (n) + 2 M where (†) M is the number of vertices of minimum degree in Tr (n). The case G. remove edges to obtain G be a vertex of minimum degree in n ≤ r is trivial as then Tr (n) = Kn . we get Tr (n) lie in the largest classes. note r + (n − r)(r − 1) + tr (n − r) 2 (∗∗) δ(Tr (n)) If then ∆(Tr (n)) ≤ δ(Tr (n)) + 1. so if we have a graph with |G| = n and δ(G) > e(G) > tr (n). we have the following. given By induction on n. Let x G . Each vertex of G has at r − 1 neighbours in Kr . throughout. observe that vertices of least degree in remove one such vertex.e. miss dierent vertices of K . r + (n − r)(r − 1) + e(G − K).4 G = Tr (n). hence (∗) and noting we have now a tr (n) − δ(Tr (n)) = tr (n − 1) Moreover. In general. . Since Tr (n) is maximal Kr+1 -free. δ(G ) ≤ δ(Tr (n)). More explicitly. so G = Tr (n). Since vertices in dierent classes of G − K . Then G − x has order n − 1. (Turán. By the induction hypothesis. K say. by comparing the degree sequence of G δ(G) > δ(Tr (n)) then with that of the Turán graph. so every vertex of e(G − K) ≤ tr (n − r). Q: Can we do better than Tr (n) and still be Kr+1 -free? of order Theorem 3. Let G be Kr+1 -free n with e(G) ≥ tr (n). n. most maximal e(G ) ≤ By (∗∗). But e(G ) = tr (n). By (†). 2 so G − K = Tr (n − r) by induction and equality holds G − K is joined to all vertices but one of K . Certainly G contains some Kr . e(G ) = tr (n). Then Proof 1. removing one vertex from each class (i. 1941) . G = G = Tr (n). G is r -partite. is Kr+1 -free and with e(G − x) = e(G ) − δ(G ) = tr (n) − δ(G ) ≥ tr (n) − δ(Tr (n)) = tr (n − 1) using (∗). Obtain G from G by adding edges until the graph is Kr+1 -free.e. a Tr (n − r) in which each vertex has r−1 neighbours in Kr ) Kr . Tr (n − r). and if we Tr (n − 1). i. we have G = Tr (n) and G = G . tr (n) = Finally. G − x = Tr (n − 1). so By induction on Proof 2. Since there must be some class of G in which x has no neighbour (otherwise Kr+1 ⊂ G ).17 First.

This is a set of Γ(x). The existence of this graph is equivalent to the existence of a projective plane of order where p. and vice versa by arguing with X. 2) ≤ 1 n(1 + 2 4n − 3) and equality holds for innitely many Hence n. even approximately. Theorem 3. from (A) with t = 2. |X| = |Y | = n. and each two lines have exactly one common point. t) with bipartition X. together with n subsets + 1).e. Proof. dn . and. If T ⊂Y di t = |S| ≤ (t − 1) w ≥ t − 1. . The value of z(n. T ) in S . . . points (p + 1). Let the vertices in X have degrees d1 . that contains no complete Kt. Let nd = i=1 di = z(n. Let S = {(x. n. i. Y transposed. the vertices of Y . as a power of Let G be a maximal n × n Kt. t). the sets each line has the same number of n points.t -free graph of size z(n.t ? This is known as the problem of Zarankiewicz. from (B ). there are dti pairs (x. of points called lines. t) ≤ (t − 1)1/t (n − t + 1)n1−1/t + (t − 1)n = O(n2−1/t ) if t is xed and n → ∞. √ d ≤ 1 (1+ 4n − 3).6. ≤ d(d − 1)(d − 2) · · · (d − t + 1) t−1 ≤ n(n − 1)(n − 2) · · · (n − t + 1) n √ (∗) Theorem 3. . . 2 To obtain equality. T ⊂ Y If such that |T | = t ∧ T ⊂ Γ(x)}. Thus n i=1 Since the polynomial x. d t t n ≤ i=1 di t n t (B ) d−t+1 n−t+1 The result follows. with n vertices in each class. Thus. Remark. n. Even the proper rate of growth. Denote the maximum size by z(n.18 Extremal Graph Theory It is natural to ask the bipartite analogue of the previous question. . each pair of points is in exactly one Each point is in the same number of lines (p common line. note by maximality of G n that di ≥ t − 1 for all i = 1. . T ) in S with this with |T | = t there are at most t − 1 x's with (x. The following simple idea is thought to be accurate.5. which hence must be an integer. z(n. . we need all above inequalities to holds exactly. t). The above proof shows d(d−1) ≤ n−1 by (∗) for t = 2. all degrees in X are equal to d. It is known that there exists a projective plane for every prime power order. . is unknown. What is the maximum size of an bipartite subgraph n×n bipartite graph. t) is unkown. every two vertices of Y have exactly one common neighbour in X . ≤ (t − 1) n t (A) w t in w is convex if n Therefore. T ) : x ∈ X. Y . Proof. choice of x ∈ X has degree di . n= p2 + p + 1. z(n.

2 ex(n. the following image shows the Fano plane.7. K2. see Example Sheet 2. there is no nice exact description of Usually. but sometimes things get nicer for n. F ) in general. G ⊃ K2. where d is the average degree. Kr+1 ) = tr (n) ∼ (1 − 1 ) r This is clear by noting that any n 2 r−1 r x ∈ Tr (n) is joined to a share of approximately vertices. ex(n. 8 • •8 88 n $$ $ nn • €€€ 88 $$ €€€nnn8n $ nnn€€ 8 W •W • €€8€ Õ $$ WW nn$nnn 88 €€€ÕÕÕ n $ W € nnW 88 ÕÕ €€€ nnn WW$$ • • WW ÕÕ 8 $$ W ÕÕ88 $ W Õ $$ WWW ÕÕÕ 888 • • $ WW ÕÕ 88 $$ W ÕW $ • • ÕÕ WW ÕÕ € • €€ • • It is known that there are no such planes of order 6 (easy) or 10 (hard). larger Not surprisingly. the case of n small relative to |F | is a mess. P ) = t2 (n) + n − 2 for large n P and where P is the Petersen graph. |S| = t. and both F are shown below. x ∈ S . • II  III  II  I  I • w rI •Iwwww rrr III II w   rrrrwww II II  r  w I  rr III  wwwwII rrr w   r • • • This gives rise to the Heawood graph. . K3 ) = n2 4 ex(n. For the next examples. ex(n. x ∈ V (G) and covering As before we count vertices We obtain S ⊂ V (G) with v∈V d(v) t ≤ (t − 1) n t e(G) = 1 dn.2 ) ≤ n (1 + 4 |G| = n. C5 ) = t2 (n) ex(n. We have the following examples. the projective plane of order 2. For example.19 Example. n d t ≤ √ 4n − 3). sets Suppose ex(n.2 . i. F ) = n2 for n≥6 + 2 for n ≥ 5 4 ex(n. The result follows by observing Remark.e. Proof. Theorem 3.

1 Since each vertex of K has degree at least (1 − r + ε)n. Then. Let and if n is suciently with |G| = n and Kr+1 (1) = Kr+1 δ(Tr (n)) ≈ (1 − 1 )n. every graph G + ε)n contains Kr+1 (t). ε t (i) This is trivial if r = 1. (Also note that the case r = 1 2t Let T = εr . this gives ≥ εn 3 . T. r r. ε) ≥ 6rT . δ(G) ≥ (1 − Note that large (i. t. and (iii) n1 (r.5. provided n ≥ T. t. K) for the number of edges between G − K and K .8. We proceed by induction on r=1 and the general case r > 1 simultaneously. (This part uses induction. class. so |U | ≥ εrn 3 .) G − K contains a large set of vertices U . ε) such that if n > n1 ). can we nd ex(n. n1 (r. . F )/ n 2 ? to be the complete Denition. we have if G ⊃ Kr (T ) |K|[(1 − 1 r + ε)n − |K|] ≤ e(G − K. r-partite graph with t vertices in each The following lemma is the heart of the matter. 1 ε Let U be the vertices in G − K having at least (1 − r + 2 )|K| neighbours in K .) G contains Kr (T ). t ≥ 1 be integers and let ε > 0 be real. (ii) r 3t n1 (r. t. K) ≤ |U ||K| + (n − |U |)(1 − 1 r ε + 2 )|K|.e.20 r • vv rrr rr vv rr vv rr vv rr vv AA • v !A rr vv ! A v……… AA … • • ii ! AA ……rr !! AA viiii !!! • r • AAvvv ! rr AA ! rr vv !! AA r !! vvv rrAAA !! AA rT !! v !! • •T AA ØØv TT !! AA ØØ TT !! ØØ ! • • For general graphs Extremal Graph Theory t • tt tt tt tt tt tt tt tt tttt • ttt tt tttt tt tt t t tt tt t tt • • • F. Lemma 3. call it K . Dene Kr (t) Kr (t) = Tr (rt). (i) (ii) (iii) can be derived from Theorem 3. r |U | r |K| ≤ εn 6 if n is large. by the induction hypothesis. ε) ≥ T and n1 (r. ε) ≥ εr T .e. there exists 1 r r. 1/r(r − 1)) for r ≥ 2. This gives Kr+1 (t). t. If r>1 + and since δ(G) ≥ (1 − we have that 1 r−1 1 r(r−1) )n (ii) n is large. writing e(G − K. F ) approximately? Can we nd limn→∞ ex(n. For exactness. i. We proceed in three simple steps. Many vertices in U (certainly at least t) are joined to the same t in K . t. so εn 2 Since ε − |K| ≤ |U |( 1 − 2 ). we shall n1 (1. each joined to at least t in K. proving the base case Proof. ε) ≥ n1 (r − 1. each class of each class of It will be evident that each step holds if use only (i) n is suciently large.

8 shows Kr+1 (t) ⊂ H . and the vertex in Gj but not in Gj−1 has 1 ε degree at most (1 − r + 2 )j in Gj . we 2 s+1 also require ≥ 2n which is possible since the left-hand side is of order n2 . such that Denition. ex(n. If n n0 (r. given Theorem 3. if |G| > n0 (r. Then if n0 > 2ε−1/2 n1 (r. there t 3 ≥ t T exist t vertices in U for which the same choice was made. δ(H) ≥ (1 − 1 + r ε ε )|H|. t. Then we can construct a sequence of graphs G = Gn ⊃ Gn−1 ⊃ Gn−2 ⊃ · · · ⊃ Gs where |Gj | = j . 2 ). tr (n) n 2 and lim inf n→∞ Conversely. Let r + 1 = χ(F ). F ) ≥ tr (n). 1946) is suciently large (i. lim sup n→∞ e(G) ≥ (1 − + ε) |G| 2 then. This is a contradiction since |Gs | = s. Then F ⊂ Tr (n). Since there are only T r T r εrn possible choices for Kr (t).e. 2 It is enough to show that G contains a large subgraph H with H exists. r 1 r ε > 0. i. t.e. To be precise. F ) n 2 ex(n. The chromatic number χ(F ) of a graph F is the smallest F is Corollary 3. Thus. 2 ε we have |H| > n1 (r. χ(F ) − 1 whence Proof. Proof. ε) G ⊃ Kr+1 (|F |) ⊃ F . n→∞ lim ex(n. For technical reasons.9. neighbours in we can pick a Theorem 3. we nd H with |H| > S = ε1/2 n . Therefore. t. . ≥ lim n→∞ 1 =1− . and hence to at least t in each class of K . 2 . for each u ∈ U . ) and Lemma 3. |F |. and since |U | ≥ if n is large.9 with (Erd®sStone. we have Kr+1 (t). for all ε > 0.10.21 (iii) Each vertex of U has at least (1 − 1 r ε + 2 )|K| = (1 − 1 r ε + 2 )rT = (r − 1)T + εrT 2 ≥ (r − 1)T + t K . t ≥ 1 be integers and ε > 0 be real. k -partite. F ) n 2 ≤ (1 − 1 r + ε). by ex(n. Kr (t) ⊂ K that is in the neighbourhood of u. F ) n 2 =1− so 1 . there exists |G| = n and e(G) ≥ (1 − 1 r Let r. Then Suppose no such e(Gs ) ≥ (1 − 1 r + ε) n − 2 n (1 − j=s+1 ε 2 1 r ε + 2 )j s+1 ε + = (1 − 1 + 2 ) r 2 εn2 s > > 4 2 for suciently large − (1 − 1 r ε + 2 )n n. ε) such that if n > n0 ) then every graph G + ε) n contains Kr+1 (t). so e(Gs ) ≤ k s 2 .

.

G is the smallest k such that G can be coloured with k colours. Indeed. It assigns the least colour to a vertex that is not used on its already coloured neighbours. It is important to realise that the number of colours used depends on the ordering.1. let j = max{i : vi ∈ H}. such that A (vertex) colouring if of a graph G with k colours is a map c : V (G) → [k] c(u) = c(v) uv ∈ E(G). Taking any vertex vn with G is connected and not regular. if χ(G) is equal to the maximum of the chromatic numbers of its components. It looks like Corollary 4. Then each at most vj 1+d has at most d neighbours vi with colours when run on this ordering d ≤ maxH δ(H). denoted χ(G). then χ(G) is the maximum chromatic numbers of the pieces meeting at x. the maximum taken over all subgraphs of G. Proof. κ(G − x) = 0. then χ(G) ≤ d(vn ) < ∆. Unlike the case k = 2. vn be a vertex of minimum degree in G. i. etc. Let χ(G) ≤ 1 + maxH δ(H). we can apply the greedy algorithm with the above ordering. then we can order the remaining vertices s. As a consequence. let vn−1 be a vertex of minimum degree in Hn−1 := G[V (G) − {vn }]. so δ(H) ≤ δ(Hj ) ≤ d since vj is of minimum degree in Hj and vj ∈ H . given H ⊂ G. distance from G is connected and vn ∈ V (G). each has at least one later neighbour. κ(G) = 1 and x is a cutvertex. Let d = maxj δ(Hj ). . algorithm never uses more than Clearly. of the Also note that. if ∆(G). there is no known good way of nding χ(G). taking Hn = G. Note that equality holds if G is complete.2. let vn−2 be a vertex of minimum degree in Hn−2 := G[V (G) − {vn . Theorem 4. the greedy 1+∆ colours. For. there is no nice characterisation of k colourable graphs for k ≥ 3. χ(G) ≤ ∆(G) + 1.Chapter 4 Colouring Denition.t. e. The chromatic number of The greedy algorithm runs through the vertices of a graph using some pre-arranged order. Likewise. Then H ⊂ Hj . by decreasing vn . Remark.g. in fact.e. Then the greedy algorithm uses of the vertices. c(vj ) = min(N − {c(vi ) : i < j ∧ vi vj ∈ E(G)}). vn−1 }]. but in fact equality holds. If We make the following observation. i < j .

vn }. two blocks pairwise intersect in at most one Theorem 4. 2. v3 . A Colouring block of a graph G is a maximal 2-connected subgraph. .t. v2 . ¯ by choosing any of colours for each of the (ii) For a tree n vertices. and note all other edges lie in blocks. We can do this because of (iii). take G = K∆+1 . vn κ(G) = 2. If Also. all vertices of G. Given these vertices. clique number of G is ω(G) = max{r : G ⊃ Kr }. v2 be two such vertices in dierent endblocks. i. (Suppose all neighbours of vn are adjacent. each block needs at most such that (i) (ii) (iii) ∆ colours. so does G. since c(v1 ) = c(v2 ) = 1 by (i). v1 . and the theorem is easily may assume ∆ ≥ 3. v2 of vn . |G| α(G) Denition. Then 3 ≤ j ≤ n−1 To nd by the preceding observation. v2 ∈ Γ(vn ). We end up with v1 . Recall that a bridge is an edge in no cycle. Let us use the greedy algorithm. with colours Given a graph G. so We have the following trivial lower bound for max ω(G). v2 . We assume G is connected. Denition. vn−1 . x ¯ Kn is the complement of Kn then pKn (x) = xn . Also c(vj ) ≤ ∆ for vn has two neighbours of the ∆(G) and since same colour by (ii). vn κ(G) ≥ 3. let pG (x) be dened as the number of ways to colour 1. and we most ∆ colours. . Then every endblock of G − vn contains vn . ≤ χ(G). Then Let G be a connected graph with χ(G) = ∆(G) + 1. . pick vn−2 joined to one of {vn−1 . vn v1 v2 ∈ E(G). take vn s. by induction on the order of the graph. . we have c(vn ) ≤ ∆. where every vertex vj for 3 ≤ j ≤ n − 1 has a neighbour vi with i > j . v2 } is since connected. i.) if of degree In the case a non-cutvertex joined to v1 . contradiction. not complete. pT (x) = x(x − 1)n−1 . vn } etc. . Also. veried.e.e. . then K∆+1 ⊂ G. G is a tree of blocks and bridges. v2 . ¯ α(G) = ω(G). we can take two non-adjacent neighbours v1 . T . G is complete or an odd cycle. κ(G − vn ) = 1. vn . We shall nd vertices v1 . G dent set. In particular. a set of vertices with no edges between them. . vertex and there exists at least two endblocks. G − {v1 .3 (Brooks' theorem). (i) If Example. x. hence G = K∆+1 . pick n−1 vn−1 joined to vn (there will be at least one vj with 3 ≤ j ≤ ∆ ≥ 3 implies |G| ≥ 4). Let v1 . . pick vn−3 joined to one of {vn−2 . ∆ = 2 then G is a path or a cycle. The The Denition. independence number α(G) of G is the maximum size of any indepenχ(G).24 Denition. . but G is connected and ∆ is maximal. so we colour with at κ(G) = 1 then no block is K∆+1 and since. if Proof.

. Remark. Remark. where n = |G|. pG (x) = pG−e (x) − pG/e (x). and χ (G) ≥ ∆(G) χ (G) ≤ 1 + ∆(L(G)) ≤ 1 + (2∆(G) − 2) = 2∆(G) − 1. e ∈ E(G). But edge-colourings enjoy special properties that merit attention. index Clearly The minimum number of colours needed to edge-colour G is the chromatic denoted χ (G). aj ≥ 0 for all pG (x) = xn − an−1 xn−1 + · · · + (−1)n a0 . we have pKn (x) = x(x − 1)(x − 2) · · · (x − n + 1) for x ≥ n as we have x choices for the rst vertex. In particular. etc.6. and min{j : aj = 0} = k where k is the number of components of G. pG (x) is derived given Theorem 4. If 1 ≤ x ≤ n − 1 then pKn (x) = 0. see Example Sheet 3. for example. c : E(G) → [k] where c(e) = Note an edge-colouring of G is a vertex-colouring of the line graph L(G).) Corollary 4. c(f ) Note that A G is not specied by pG (x) up to isomorphism. Let G be a bipartite multigraph. Denition.3. j . an−1 = e(G). f share an endvertex. g37 • ggggg %%))3377 ggggg g %% ) 33 77 • & ) 7   & %% % )) 333 77  &&  % )) 3 7   &&   & %% )) 3  ) ∆−1 By Brooks' theorem 4.4 and induction. ∆−1 if χ (G) ≤ 2∆(G) − 2 G is connected and not an odd cycle or Theorem 4. But the colourings of colourings of G are these colourings of G − e where the ends of e get dierent G − e where ends of e have the same colour are precisely G/e. This is left as an exercise via Theorem 4. an edge.4 and by induc- Observe. For all edges Theorem 4. that tion on C But nearly all other information about |E(G)|. (Although this can be seen directly as well. Proof. x − 1 choices for the second vertex. if the two k -edge-colouring of the graph G edges e.4. Proof.25 (iii) For a complete graph. Then χ (G) = ∆(G). pG (x) = pG (x) pC (x) where C runs over the components of G.5. is a map Denition. The colourings of colours. is a polynomial called the chromatic polynomial.

This is clearly the best bound possible. Proof. Note χ (Kn ) = n only n≥3 odd. Make G − {u. xyk ck = cj ci available at yi is xyi+1 . the following graph has K3 . joining ∆-regular. If not. xy1 . Remove them from G and get a (∆ − 1)-regular graph. Pick an edge uv of multiplicity m ≥ 1. Some colour. For example. G has a colouring with ∆ + 1 colours leaving one edge xy0 be the uncoloured red • y3 € •U € #U € €€ U ## UUU €€€€€ yellow # €€ UU ## UUgreen €€€€ # • UU # y2 UU ## U ## • • y0 (green ) Construct a sequence of edges the colour of x (red ) y1 (yellow ) such that a colour xy0 . Let edge. 1965). . e. Thus the red edges together with Note Theorem 4. For for K3 . It is also true if ∆ = 0. . . We must stop either (i) because ck does not appear at x or (ii) because for some 0 ≤ j < k. does not appear on a new uv (one copy) form a 1-factor in G.g. and even ∆ = 6 but χ = 9. red say. Note that at every vertex at least one colour is unused.6 fails for non-bipartite graphs. v} ∆-regular by adding ∆ − m edges between Γ(u) and Γ(v). edge. but meet every vertex. ∆-regular multigraphs by induction on |G| + ∆(G). between Γ(u) − v and Γ(v) − u. Here is a direct proof.e. . then it has a colouring with every edge coloured. Do this maximally with distinct ci .7 (Vizing. observe we may assume graph formed from G is G and a copy G of G. Note it |G| = 2. replace G by the v ∈ G to v ∈ G by ∆ − d(v) edges. It is enough to show that if uncoloured. graphs. χ = 3. so we may assume ∆ ≥ 1. more so for multi- • I •  III  II II   II  II   • Theorem 4. Colour this by induction. . χ (K2n ) = 2n − 1 according to Bollobás. Colour one uv -edge red. then the red edges do not meet each other. Then ∆(G) ≤ χ (G) ≤ ∆(G) + 1. Let G be a graph. First. i. Colouring This can be proved by applying Hall's theorem (see Example Sheet 3). We prove the theorem for is true if u cc WW R •R Γ(v)−u • c  WW ÕÕ RRc   WW Rccc ÕÕ  RRW c  ÕÕ cc Õ RWWc RWWÕc RRÕW ÕÕWcc  Õ R  Õ  R WW cc  ÕÕÕ RRR WW c Õ  RR WW ÕÕ  RR  R  v Γ(u)−v Colour this multigraph with ∆ colours. ∆ = 2.26 Proof.

χl (G) = χ (G) = ∆(G) G (Galvin. Amazingly.3 • vrr vv rrr rr vv rr vv rr vv r 1.2 1. It is conjectured χl (G) = χ (G) for all graphs G. Swap colours in the Recolour xyi with ci . yj consists of paths and cycles. 0 ≤ i ≤ k . χ(G) ≤ 5. χ. Let G be a minimal counterexample. Let G be a planar graph. χl (G) ≤ 1 + maxH δ(H) holds by the previous proof. drawn in the plane. Now let c be a colour not used x x and let H be the subgraph consisting of edges coloured c and ck .2 •Gc 1. Since e(G) ≤ 3|G| − 6. j ≤ i < k . χl G (e. In case (ii). χl (K3. Five Colour Theorem) .1. it follows that δ(G) ≤ 5. yj . We can swap c and If ck in any component of and still have a proper colouring. recolour xyi with ci . Thus χ(G) ≤ 6 by Theorem 4. even for a bipartite χl (G) ≥ χ(G). Suppose otherwise and let be a vertex with Proof. Colour G−v with 5 colours. swap c and ck in the component containing x. so colour xyj with ck = cj .2 1.3 • 2.3 • • c GG cc   GG cc  GGccc    GG  ccc   c  • c GG • cc  G   GG cc  cc   GG  cccGG  cc G   1. so x. and we are done. and yj lie in dierent components of H. H⊂G this.3 ) = 3). then χ(G) for G planar. But yk is in H at yk . dH (yk ) ≤ 1. We can improve Theorem 4. We can dene for bipartite graphs Consider χl (G) analogously for edge colourings. 1994). 0 ≤ i < j H and uncolour xyj . recolour at xyi .2 c G 1. δ(H) ≤ 5. In general. Also.27 In case (i). But ∆(H) ≤ 2. Dene χl (G) = min{k : ∃ Clearly graph list colouring whenever |L(v)| ≥ k ∀v ∈ G}. so H dH (x).2 • 1. if H is planar. Then v must have a . v d(v) ≤ 5.3 However.2 vv v A •A • ! AA !! AA ! AA !! AA !! AA !! !! • 1.g. dH (yj ).8 Then (Heawood. at A list colouring c(v) = c(u) v. and colour xyk with c. Thus component of x. This frees ck at x and leaves yj unaected. can be much bigger than 1. yk lie at the ends a dierent component of H from x and yj .3 2. so we are done. H . if uv ∈ E(G)) with G is a colouring c : V (G) → N (as usual c(v) ∈ L(v) where L(v) ⊂ N is a list of colours available For the usual colouring we have L(v) = [k]. Thus we may assume of paths. are in the same component of of a graph Denition. 1890.

Let χl (G) ≤ 5 for planar G. This makes u2 colour 4 and leaves colours at ui . let v1 x1 x2 . vj . we can get more for less. L(v1 ) = L(vp ). so colour G2 using (†). . So there exists a path coloured 2/4 from u2 to u4 . 1 ≤ i ≤ 5. let G1 . . vj w • ww vp • • v1 www G2 www www www www www G1 ww • v3 • vi • v2 If there is no chord. unchanged. . contradiction. .e. We prove the following by induction on |G|. Then G can be coloured. vp •  • v1  A $ ÓÓ G AA $ ÓÓ G x •  AA $$ ÓÓlllll k $lÓ l l  $Ó • v2  • v3 x2 x1dddd • A •A d $$ G •  A $ ÓG . where vi . let ui be the neighbour of v with i. Colours at vi . vi vj . i. Let |L(vi )| ≥ 3 for 2 ≤ i ≤ p − 1 and |L(v)| ≥ 5 (†) elsewhere. vp and vi vi+1 . . . subgraph induced 4 in the component at u2 . then swap 2 and But likewise there is a 3/5 coloured path from u3 to u5 . respectively. . u 4 2/4. vp and have triangular faces inside. are in dierent components of the u2 .28 neighbour of each colour. or else we could colour colour Colouring v too. . G1 can be coloured by (†). • jj„„„„„„ „„„„ 4 jjjj jjjj • • # ## ## u1 • • u2 ## u5  y ##  • yy # yyy  y y ## y • • yy jjjj yyy yyy 2 yjjjjjj v yyy • • 2 4 u4 u3 If by vertices of colours 2/4 coloured subgraph. i = 2. Let |L(v1 )| = |L(vp )| = 1. Then colour v with 2. contradicting planarity. vj are now forced. vj are not successive elements of the cycle. .9 (Thomasson. Theorem 4. 1993). Proof. G have an outer cycle v1 v2 . If there is a chord vi vj . G2 be the subgraphs with boundaries v1 v2 . xk v3 be the neighbours of v2 . . Remarkably.

Pick i. i. ………… ………… iiiiiii …T i •T Øi ØØ TTT Ø TT ØØ TT ØØ Ø / r • vv rrr rr vv ………vvv riii A •A • ! AA !! AA !! A !! TT • • Ø T ØØ Theorem 4.10 Proof. Label each dual edge with the It is a triangulated plane graph. Give an edge the colour which is G is bridgeless. and the paths used are still known as Kempe chains. 00. {00. There exist (non-trivial) planar graphs G with χl (G) = 5. colour of the original edge that it crosses. Suppose conversely from G 3-edge-coloured. It was proved by Kempe in 1879 for which he was made FRS. Since G was properly coloured. connected bridgeless plane multigraph. i. so colour v2 with it. 10 . and the same without on edges. returning to F0 . One of i or j is not used to colour v3 .e. 00 is on no edge.29 G = G − v2 . ulo (Tait. The faces of any map. 1880) . j ∈ L(v2 ) − L(v1 ). Pick a face F0 and.e. j}. the edges on each triangular face are 01. and give the result to F. is so the edge colouring is proper. sum to F0 . First observe that it is enough to colour cubic maps: see this either by triangulating the original graph or making the following replacement. 4-face-colourable if and only if it is 3-edge-colourable. We must show the labels around any cycle 00. 2. 1852) is to show χ(G) ≤ 4 G. F adding up the colours of edges when crossed. In 1890 Heawood found a mistake. But we must check This is equivalent to checking that. can be coloured with four colours so that no two contiguous faces have the same colour. 01. The Four Colour Problem is stated in dual form. if we go for a circular walk from the sum of edges crossed is Consider the dual. walk F0 to F. no two incident edges have the same colour. Let L (xm ) = L(xm ) − {i. for any other face F. 10. using L by (†). for planar graphs But the Four Colour Problem (Guthrie. The Four Colour Theorem holds if and only if χ (G) = 3 for every cubic bridgeless planar G. Note this gives adjacent faces dierent colours since the colour of is independent of the route chosen. Tait (1880) found a beautiful equivalent form of this. Suppose G is colours. Kempe's proof was similar to the proof of Theorem 4. 00 does not appear on an edge.8. Since 4-face-coloured. i. 11} with addition coordinatewise modthe sum of adjacent face We must show such a G 00 is We use colours Z2 × Z2 on faces. w • qqq www qqq c wwwww qqq Note a b a+c = b+c as a = b. Let Colour G Remark.e.

e. k≥7 is unknown. χ(G) ≥ k =⇒ G Kk . 46 (see Example Sheet 3).e. χ (G) = 3. is a sphere with some number (i) g≥0 of handles. the sum around any face is 00. = 0). Since a cubic graph has even order by the Handshaking Lemma. Note E = 2 − 2g .e. and indeed believed he had proved. closed. Tait further conjectured. Sanders. is simply connected satises the number of faces. announced a computer-based proof. Hence. i. faces inside cycle edges around face ≡ edges around cycle (mod 2). Few people have read it. a surface with an inside and outside. i. n−m+f = E where n is the order. face. In 1976. f is An orientable surface. Remark. . The k=6 (equivalent to the Four Colour Theorem). that every cubic plane bridgeless graph has a Hamiltonian cycle. Appel and Haken. This is easy for Haderinger (1943) conjectured only case known is k = 4. χ(G) ≥ 5 =⇒ Wagner (1935) proved that the Four Colour Theorem holds if and only if G K5 . 2 A surface (2-dimensional. colour cycle edges red and blue. Graphs on other surfaces Denition. Seymour.30 and Colouring 11. Robertson. Thomas gave a new simpler proof based on the same ideas. The But in 1946 Tutte found a counterexample of order smallest counterexample has order 38. In 1997. and colour the remaining edges green. i. g=1 yields the torus (E / O / / / O (ii) g=2 yields the double torus (E = −2). compact) has an Euler characteristic E ≤ m is the size and such that a graph drawn on this surface in such a way that each region. using ideas of Heesch. But if we have a cycle.

1890) . Then surface. We can draw K6 on the projective plane fact χ(G) = 6 is attained.11 tic (Heawood. so δ(G) ≤ 5 for every graph on χ(G) ≤ 6. so m = 3(n − E). so assume the surface having chromatic number k. so surface E. so in m ≤ 3n − 3. 2 Proof. Let G be a minimal graph on |G| ≥ k and δ(G) ≥ k − 1.31 There are also non-orientable surfaces.1. else we can k − 1 colours. We already considered E = 1. / O /  / / If we have a maximal planar graph on a surface of characteristic triangle. then χ(G) ≤ H(E) = 7+ √ 49 − 24E . one for each value of (i) The projective plane (E E ≤ 1. If G is a graph drawn on a surface of characteris- E ≤ 1. this (see Theorem 4. Hence every graph on the Consider the projective plane. = 1). satises m ≤ 3(n − E). Then remove a vertex and colour the graph with E ≤ 0. Example Sheet 3). Hence k − 1 ≤ δ(G) ≤ 2(3(|G| − E)) E =6−6 |G| k . By Theorem 4. / O /  o (ii) The Klein bottle (E o = 0). then every face is a 2m = 3f and n − m + f = E .

then equality holds above. By Brooks' theorem 4. II I I4  II II II II  III  III  III  I  I   III  III II  II  II  II  II  II  II  II II II2 II3 IIx I III II III III III III  II  III  II  II  II  II  II II  II  II  II  II   I I 6 I 0 I 1 I I  I I I I I  II II  III  III  III  III  III  II  II  II  II  II  II  II  II  II  II  II  II  II 4 II 5 II II I II I I I I I  I  I  I  I  I III  II  II  II  II  II  II II  II II  II  II  II    I I I I I I  . it looks like a planar triangulation. Locally. i. G = K7 . Let 7. 1 3 5 jj jjjj jjjj jjj 2 O jjjj O jjjj 1 jj j jj 0 jjjj jjjj jj 6 jjjj 6 jjjj jjjj 5 jj jjjj 4 / / jjjj 3 4 2 How about the Klein bottle? We must embed / / jj 5 jjjj jjjj j jjjj 3 4 4 K7 so that every face is a triangle. when H(0) = 7. G H(2) = 4. be minimal of chromatic number For the torus this is possible.3. we have the following.e. torus. or k 2 − 7k + 6E ≤ 0. Note δ(G) ≥ 6 but e(G) ≤ 3|G| so G is 6-regular. Tile the plane using the shaded quadrangle forming a I0 II 2 I3 I4 I5 I6 • II  III  III  III  III  III  II  II  II  II  II  II  II  II  II  II  II  II  I 5 II 2 II 3 I 4 II 0 II 1 • I I I I II  I  I  I  I  I III  II  II  II  II  II  II II II  II  II  II  II    I I 5 I 6 I 0 I 1 I 2 II II II II II II 4  II  II  II  II  II  II  II  II  II  II  II  II  II  II  II  II  II  II  II 4 II 5 II 6 I 0 II 3 I 2      III  II • II  II  II III  II  II  II  II  II  II II  II  II  II  II II    I 1 I 2 I 3 I 4 I I 0  6 • Dually. Therefore. the torus or Klein bottle. since 21 = m = 3(n − E). Remark. Can equality hold? Consider E = 0. equality is attainable if and only if K7 can be drawn on the surface.32 as Colouring E ≤ 0. if equality is possible.

this is equivalent to embedding KH(E) on S. By symmetry x = 4. . . By the same argument y is 4 or 5 but y is joined to 3 which joins x = 4.33 0 and numbers 1. Without loss of generality. . it was completed by Ringel and Youngs (1945).0 or 5. all 2s etc. we get a torus. the pattern is now determined: it is the previous pattern. Hence. so cannot be 2. the Klein bottle is the only exception to max{χ(G) : G For embeds on surface S of characteristic E} = H(E). 6 around it. x Amazingly. Heawood thought he had proved this. E ≤ 1. .. x meets 1 and 3. . Hence. But when we identify all 1s. so y = 5. K7 can not be embedded on a Klein bottle. so x is 4 or 6.

.

1935) . there are either at least a red or b blue. Amongst the a + b − 1 edges meeting v . t − 1). R(s. It is enough to verify the inequality. so this Kb either contains a red Ks . edges of In general. no other values are known. b = R(s. there is a monochromatic Ks . or it contains a blue Kt−1 which. a = R(s-1. forms a blue Kt . s). Theorem 5. t) ≤ s+t−2 s−1 . Let Remark.g. . So v is joined to a Kb by blue edges. if the Kn are coloured red and blue. (i) The proof shows that R(s. Now b = R(s. if it exists. 1930. Moreover. R(s. R(2) = 2. trivially. Then R(s) = R(s. t) ≤ R(s − 1. (Ramsey. Pick a vertex v . such that. r • vv rrr rr vv v rr vv rr vv r v v A •A • ! AA !! AA ! AA !! AA !! ! A !! • • Is it true if n is large enough and edges of Kn are coloured red and blue then there is a monochromatic K4 ? s∈N let Denition. if R(s) be the smallest n. equality can hold in the above proof only if there is an (a − 1)-regular (iii) We only know graph of order a + b − 1 which is false if a. there must be a red Ks or a blue Kt . 2) = s. n = a + b and colour Kn red and blue. e. such that. R(min{s. t}) ≤ R(s.1 Proof. Suppose the edges of This fails for K6 are coloured red and blue.Chapter 5 Ramsey Theory The simplest example is the pigeon-hole principle. Also R(2) = 2. t − 1). with v . R(s. t-1). if the edges of are coloured red and blue. For example. b both even. t) exists. R(3) = 6. R(3) = 6 Kn by the above. Erd®sSzekares. It is convenient to dene R(s. s). Denition. t). t) + R(s. K5 . (ii) This is not exact. Then there is a monochromatic K3 . R(s. t) to be the smallest n. if it exists. We may assume the latter. t}). R(4) = 18. t) ≤ R(max{s. t) = R(t.

. t) = R2 (s. . sk ) ≤ Rk (s1 − 1. without loss of generality assume the latter. . . . sk ). . or there is a KR(s1 . t) = R(s. k . . . . Given r ∈ N. . . . . s3 . t) are huge. t ≥ r. t) ≤ n. . s2 . . such that. . R(2) (s. colour (r−1) (a. if edges Ksi coloured n. . sk ). an innite . k ≥ 2. . . . Go colourblind. has vertex set [n] and Does there exist a smallest number red and blue then there is a red and R(1) (s. some Note Suppose we use Ramsey Theory k ≥ 2 colours. Since b = R (r) (r) r Ks or else a blue Kt−1 which with v forms a blue Kt . t) such that if we colour the edges of Kn (r) (r) Ks or a blue Kt ? For example. i. sk ) of Kn are coloured 1.2. .3 (Ramsey for r-sets). . whence we are done. b) this means we it in the same colour as we gave to Z ∪ {v}. . Hence in the original colouring there are b vertices such that every r -set formed from v and r − 1 (r) (s. . . Let Rk (s1 . Let the edges of K (r) be coloured red and blue and let v ∈ V (K r ). t − 1) and n = 1 + R(r−1) (a. s2 . Use induction on The proof shows Rk (s1 . Proof. . . 2. . . b) exist (r) (s. Note that the second bound is much better. Theorem 5. E) where E ⊂ V (r) = {Y ⊂ V : |Y | = r}. . obtained on R Rk (s1 . be healed. do we M ⊂ N with M (2) same colour? (r) The bounds What if we colour innite graphs? If we colour monochromatic set get an innite M. Example. so there is a Ks1 coloured 1 or a Ks2 coloured 2. . s2 ). if i. let Kn be coloured with k colours 1. . . let s1 . there is be the smallest a it exists. . . n = R(r) (s. . t) = s + t − 1. these vertices either contain a red of these b vertices is blue. t).s2 ) coloured 1/2. sk ) + · · · + Rk (s1 . s2 . s3 . sk ) ≤ Rk−1 (R(s1 . So r-uniform |E| = n . We can either generalise the proof for Theorem 5. Let Theorem 5. an r-uniform hypergraph 2-uniform is a pair (V. so that colours 1 and 2 are indistinguishable. k . Denition. Colour blindness shows (r) (s. . s2 − 1. Given an edge Z of Kn−1 .1 or proceed as follows. . a = R(r) (s − 1. . Since n − 1 = R (r−1) (r−1) get a red Ka or a blue Kb . The case k = 2 is Theorem 5. N(2) red and blue. t). (ii) The complete E = V (r) . sk ) exists with the natural denition. A modication of the proof of Theorem 5. sk ∈ N. .e. . sk ) + Rk (s1 . s.1. Then Rk (s1 . s3 . Either there is a Ksi coloured i for some i ≥ 3. In the latter case. t − 1). . . . then R n n (r−1) (r) (r−1) Consider the Kn−1 with vertex set V (Kn ) − {v}. R(s.1 gives Rk (s1 . . s2 ). sk − 1) − k + 2. In general. sk−1 . We show that if r ≥ 1. . . . . sk ) exists. . r hypergraph of order n is Kn (r) . b = R(r) (s. k . Proof. t) exists for all (r) which again is the pigeon-hole principle. . (i) A graph is a hypergraph. . t) R(r) (s. for i.36 Denition. Suppose n ≥ Rk−1 (R(s1 .

. . . Repeat. . such that if Y ⊂ {v1 . k. . n}(r) as Z1 .4 (Innite Ramsey). . . vi2 . because there is a subsequence which Proof. . . Then for each n. . on which the colour of Z1 (2) (2) is constant. . . . . for To construct these sequences. . 24} is big. vi3 . and Y ⊂N is big if |Y | ≥ min Y . s + 1. r ≥ 1 there exists B(s. So {3. . Hence we get a colouring c : {s. }. . innite k colours. k. Now let monochromatic on M Then Y is big and c is monochromatic on m = min M and Y (r) . max{k : 2k | i + j} is odd. . . j} {i. . . is not. . . . c2 . ci2 . c2 . suppose we so far found v1 < v2 < · · · < vj and colours c1 . So if the theorem is false. . is constant and then we take vi1 . j} red if red if red if i + j is odd. . . v3 . . . 9. (r) and min Y = v then Y has colour c . . . . k. . c2 . . .5. Then M = {4l : l ≥ 1} works. cs+1 . picking a subsequence c1 . For each subset Z ⊂ (Nj − {vj })(r−1) . contradiction. . . M = {vi1 . (j−1) (j−1) of c1 . then c agrees with c1 on Y (r) and c1 = cn . s + 1. on which (j) (j) the colour of Z2 is constant. Proof. Trivially. n}(r) are coloured with k colours then there is a big subset Y ⊂ {s. of But by Ramsey's Innite Theorem there is an innite (r) . Then there is an M ⊂N (r) is monochromatic. s + 1. } such that c is let Y be the rst m elements of M . 209} Theorem 5. Then M the set of even numbers works. c2 . r) such that if n ≥ B(s. . vi3 . . . some n. . such that M r. . ci3 . vi2 . We cs . . . colour it the same colour as {vj+1 } ∪ Z . r) {s. {5. . if the assertion holds for some N. . the assertion fails for all Suppose not. The case r = 1 is just the pigeon-hole principle. . n}(r) → [k] {s. Enumerate (1) (1) can pick an innite subsequence c1 . . . . . ci1 . 36. there is an innite subset Nj+1 ⊂ Nj − {vj+1 } where every Z has the same colour cj+1 . c3 . n}(r) → [k] with no monochromatic big subset: for if c were constant on (l) (l) Y (r) . . . vj } ∪ Nj and min Y = vi then Y has colour ci . n} with Y (r) monochromatic. which has no big monochromatic Y . call it c(Z1 ). . . s + 1. . . c2 . it will hold for all n ≥ N. Example. Z2 . Let be coloured with Theorem 5. By Ramsey for r − 1.37 (i) Colour (ii) Colour (iii) Colour set {i. with no monochromatic big subset. We shall nd a sequence v1 < v2 < v3 < · · · of elements of N and a sequence of colours c1 . Let vj+1 = min Nj . n≥s dene a colouring cn : {s. on which the colour of Zj is c(Zj ). No explicit N(r) M is known. Starting with N0 = N it suces to show we can get from j to j + 1. . } i i By induction on Given such a sequence we are home. v2 . s + 1. . M ⊂ {s. . c2 . . . . j} {i. cj and an innite set Nj ⊂ {n ∈ N : n > vj } such that if Y is an r-subset of {v1 . 37. Then pick an innite subsequence c1 . s + 1. let l = max{j : Zj ∈ Y (r) }. . i + j has an odd number of distinct prime factors. . A set Denition. Given s. . c(Z2 ) say.

6) = 18 R(4. 2. 9) = 36 R(3) (4. . actually using compactness. 7) = 23 R(4. However. 5) having been R(3. 9). 1991. Consider the Turán graph Ts−1 ((s − 1)2 ). 5) = 25 R(3. 1992. 4). 4) = 9 R(3. This is the rst natural example of Gödel's Incompleteness Theorem. • (Upper bounds) We have R(s) ≤ 2s − 2 s−1 ∼ 22s−2 √ . n. 4. 4) = 18 The following bounds have been established. There is no Ks . The best known construction gives log2 s R(s) > e 4 log log s . though we proved it via an in- nite argument. R(3.5 is a purely nite statement. 3) = 6 R(3. s • (Lower bounds) Trivially. 3. For example. 2πs 1 and the only improvement made in the last 70 years is an extra factor of √ . . The bound R(s) > s3 is more tricky.38 Remark. . 5) = 14 R(3. Harrington 1977). . with found in 1982. for suciently large The following Ramsey numbers are known. so no blue Ks . n with k colours there exists an l-term arithmetic progression in one colour. Other structures support Ramsey-like theorems. R(3) (4. this theorem cannot be proved without recourse to such argument (Paris. respectively. Amazingly. so in particular no red Ks . 8) = 28 R3 (3. Ramsey Theory Theorem 5. 4) = 13 R(3. . R(s) > (s − 1)2 . this is still basically zero compared to the upper bound. and no s vertices with no edges between them. 4. van der Waerden's theorem states if we colour the integers 1. 3) = 17 R(3. R(4.

The expectation of X is E X = ω∈Ω X(ω) P(ω). e. E(X + Y ) = ω∈Ω (X(ω) + Y (ω)) P(ω) X(ω) P(ω) + ω∈Ω ω∈Ω = Y (ω) P(ω) = E X + E Y. Probability. Let Ω be a nite probability space. ω∈A = P(A). Each is monochromatic with probability is 2 number of monochromatic n s 2 1−( s 2 1 2 ( ). 1947).Chapter 6 Probabilistic Methods Theorem 6.g. if X(ω) is the number of monochromatic Ks . But this would be to lose the potential of the pobabilistic approach. But if n= 2(s−1)/2 . with uniform distribution. Proof. e. s 2 So the expected ). Then R(s) ≥ 2(s−1)/2 . There Kn Ks independently red and blue with probability n s Ks 's in Kn . But the expected number can be less than chromatic 1 only if there is a colouring with no mono- Ks . A random variable is a function X : Ω → R. The indicator function of an event A is IA : Ω → {0.g. are Colour the edges of Let s ≥ 3. 0 1 if if IA (ω) = Note that ω∈A . in the above (n) red and blue colourings of K . 1 2 . argument Ω is the set of all 2 2 n An event is a subset A ⊂ Ω. then s n 1−(s) 2 2 2 ≤ ns 2−(2) s! s < (n2−(s−1)/2 )s ≤ 1.1 (Erd®s. 1}. Variables E IA = ω∈Ω IA (ω) P(ω) = ω∈A P(ω) X that count can often be written as sums of indicators. It is important to note that expectation Recall the following from is linear. We could derandomise this argument by observing that all we did was to prove the average number of monochromatic Ks over all colourings of Kn is less than 1.

then for each is monochromatic.e. d(v) + 1 d+1 Remark. Then P(Iv = 1) ≥ P(v X= v∈G Iv where Iv is the indicator precedes all its neighbours) = 1 . and Then X be the number of vertices coloured 1.e. where a given graph with m m edges has probability p (1 This is equivalent to inserting edges independently with probability p. d(v) + 1 1 X−ε + 1 X+ε ≥ 2 X .40 in the above proof. of size at least ≥ |G| d+1 . Proof. Let G(n. Theorem 6. The graph β Iβ then E Iβ = β P(Iβ = 1). Then Probabilistic Methods α ∈ [n](s) . we get v∈G 1 |G| ≥ . which is a special case of Markov's inequality. t labelled graphs with vertex set n − p)( 2 )−m . d(v) + 1 Thus EX = v∈G So there is some ordering where P(Iv = 1) ≥ v∈G 1 . We keep using I{|X|>t} < |X| t . where d is the average degree. let Take a random ordering of the vertices. use the greedy colouring algorithm. The full theorem can be recovered by examining cases of equality. e(G) = |G|d/2. take expectations to obtain P(X > E X) < 1. α Iα . so let Iα indicate that Ks on the vertex set α X= EX = α E Iα = α P(Ks on α is monochromatic) = α If 21−(2) = X= s n 1−(s) 2 2 . 1 d(v)+1 v∈G i. d(v) + 1 X≥ v∈G Since 1 . This is equivalent to Turán's theorem when applied to the complement. an edgeless set. s EX = β X is a counting variable. Since P(|X| > t) < E|X| . Denition. i. . that v gets colour 1. p) be the space of all n 2( 2 ) [n]. G has an independent set.2.

using approximations which are explained separately later. χ(G) ≥ n .) Theorem 6. k ≥ 2. X = X3 + X4 + · · · + Xg−1 . Let B be the event Y ≥ 1. Let Y be the number of independent sets of size n 2k . in one graph at least. (i) 1 − p ≤ e−p true for all p. (See Example Sheet 3. the size of a maximum complete subgraph. Form G n by removing a vertex from each short cycle. P(A ∪ B) < 1 for suciently large n.e.3 graph (Erd®s. p) where p = n−1+1/g . g ∈ N at least g . 2 |G| n . Hence for such a G. so it has at most 2 short cycles and no independent set of size 2k . a perhaps counterintuitive fact. χ(G ) ≥ Since G has no independent set of size at least ≥ k. Then g−1 g−1 g−1 Let Xl be the number EX = ≤ if E Xl ≤ l=3 g−1 gn g l=3 n l pl = l=3 l ng < n 4 {X > n }. We showed the 2 at most 2 log2 n + 1 and 2 log2 n + 1. In fact. . Consider a random graph of cycles of length l and let G ∈ G(n. Then G has girth at least g and |G | ≥ . 2 t= then by Markov n is suciently large. 2k n/2k Since We have made the following approximations. with g ≥ 3. 1959) . Then EY = t n (1 − p)(2) t t ≤ nt (e−p )(2) ≤ nt e(−p 9k2 ) n2 = exp t log n − p 2 9k   1 1+ g   n = exp t log n −  9k 2  < if n2 1 2 P(B) < 1 2 for suciently large n is suciently large.41 In our earlier Ramsey example the red graph was an element of clique size. Let A be the event P(A) < 1 2 if n is suciently large. there exists a graph G where neither A nor B n n happens. is the independent set size is at most graph G(n. Then there exists a Proof. 1 ). i. then by Markov n. there are constructions of triangle-free graphs with large chromatic number. 2 log2 n + 1 This shows that the chromatic number can be far larger than the clique number. Let G with χ(G) ≥ k and girth k.

then t 2 = n2 8k2 + O(n) > n2 for suciently large 9k2 n. P(Xn = 0) → 0. Now E J = pn2 and E K = n pt . so t n E(J − K) = pn − t 2 2 pt 2 2 ≥ pn2 − 1 n2t pt . t = c E Xn . P(|X − E X| > t) = P((X − E X)2 > t2 ) < Var(X) E(X − E X)2 = 2 t t2 which is Chebychev's inequality. In particular. (iii) For xed L. t) = O(n2−1/t ). n Then 3 z(n. Y be classes of vertices and generate random n×n bipartite graphs by inserting edges independently with probability the number of Kt. 4 we obtain E(J − K) ≥ 3 2−2/(t+1) . 4 Taking p = n−2/(t+1) . Then by Markov's inequality. → ∞ as m → ∞. The expected number of Kt.t is less n 2 t2 −2/t . What if we want to say a variable is likely to be large? The variance of X is E(X − E X)2 . so the expected number of Kt. (log m)L mε = (log m)L m L (log m) ε ε → ∞.t . We can do better by making use of the linearity of expectation. Let J be the size of this graph and K Kt. Proof. en n > → ∞ as n → ∞. To get a lower bound we would consider random n × n bipartite graphs with p as large as possible. so p ≈ n 2 2−2/t . L n (L + 1)! m n = log m.t 's is t p . z(n.5. Let Xn : Ωn → R be a sequence of random variables. Apply Chebyshev's inequality with c = 1.4. 2 E(Xn )/(E Xn )2 → 1 as n → ∞.42 (ii) Let Probabilistic Methods t= n 2k . . For the second part. so there is a graph with 4n So far we used Markov's inequality to show that some random variable is unlikely to be large. Lemma 6. of pn ≈ n Recall that Theorem 6.t 's. t) > 4 n2−2/(t+1) . By throwing out an edge from each Kt. giving an expected size than 1. take Proof. so p. t) > J − K . J − K ≥ 3 n2−2/(t+1) . Let Let n ≥ t ≥ 2. constant c > 0 we have P(|Xn − E Xn | ≥ c E Xn ) → 0 as Suppose that Then for any n → ∞. or equivalently Var(Xn )/(E Xn )2 → 0. X.t we get a graph with no 2 2 z(n.

43 Note if X is counting. Proof.B IA IB = A. Iv indicates X be the number of isolated vertices. n Let Denition.6.v P(u isolated) [P(v isolated|u isolated) − P(v isolated)] = n(1 − p)n−1 1 − (1 − p)n−1 + n(n − 1)(1 − p)n−1 (1 − p)n−2 − (1 − p)n−1 ≤ E X + n2 (1 − p)n−1 p(1 − p)n−2 . X2 = A So 2 IA = A. Suppose X is a sum of indicators. Then v is isolated. v Iv where Var(X) = u. vertices almost ω(n) → ∞. If p = . Then EX = A so E IA = A P(A). A. t2 So for all t < EX P(X = 0) ≤ Var(X)/(E X)2 . Note there is no contribution from a pair Theorem 6. X= A IA .B So E IA∩B = A.B IA∩B . Var(X) = E(X − E X)2 = E(X 2 − 2X E X + (E X)2 ) = E X 2 − (E X)2 .B P(A) [P(B|A) − P(B)] . Var(X) = A. P(X = 0) ≤ P(|X − E X| ≥ E X) ≤ P(|X − E X| > t) Var(X) < . If p = log n−ω(n) then G has isolated n log n+ω(n) surely. So E X = n(1 − p)n−1 . E X2 = A. Let G ∈ G(n.B P(A) P(B|A). Let whether Here an event A happens almost surely Also if P(A) → 1 X = as n → ∞. B that is independent. p).B P(A ∩ B) = A. then G has no isolated vertices almost surely. We compute Var(X) as follows.B P(A) P(B). (E X)2 = A Now 2 P(A) = A.

where by Chebychev's inequality. a d µ(d) → 0 then G ∈ G(n. This is a typical Remark. we have Var X 1 ≤ 2 (E X) µ(d) d l=2 d l n − d (d)−( l ) p2 2 d−l . (E X)2 EX 1−p so X=0 almost surely by Chebyshev's inequality. say. p). p). So 1 p Var X ≤ + → 0. p) almost certainly contains a Kd . 1−p 2 1 − p ≥ e−p−p if p< 1 4 ). Ignoring the nal d l d d l n−d d−l is the number of β 's meeting in exactly subtracted term.β Now P(Iα = 1) [P(Iβ = 1|Iα = 1) − P(Iβ = 1)] . Kd . (d) = µ(d). p= log n+ω(n) then n EX = ≤ so n 1 1−p n(1 − p) −pn 1 → 0. then E Xd = Kd ⊂ G ∈ G(n. Let µ(d) = p( 2 ) .44 = EX + If Probabilistic Methods p 2 1−p (E X) . We Var X = α. Theorem 6. edge. Hence Var X = where n (d) p2 d d l=2 d l n−d d−l α p(2)−(2) − p(2) l vertices. and in Iα general and Iβ are independent if |α ∩ β| ≤ 1. n d As a nal example. p) almost certainly does not contain µ(d) → ∞ then G ∈ G(n. threshold phenomenon. If p= log n−ω(n) then n EX ≈ (Here we could also use that 1 ne−pn → ∞.7. Now p2 µ(d) ≈ Let Xd be the number of Kd 's in [np(d−1)/2 ]d d! d < (2 − ε) log1/p n n d then so if d > (2 + ε) log1/p n Let then µ(d) → 0. runs over Xd = α [n](d) and Iα indicates G[α] is complete. 1−p ne X=0 almost surely by Markov's inequality. Proof. because there is no common P(Iβ = 1|Iα = 1) depends only on the value of l = |α ∩ β|. The rst assertion follows from Markov's inequality and for the second we need only show Now let know Var(X)/(E X)2 → 0 α Iα . whereas if µ(d) → ∞. consider G ∈ G(n. (i) If (ii) If 0<p<1 be xed.

choose 1 al ≤ da2 + ( R + C )ad .45 = 1 n d d d ≤ = n d d l l=2 d n−l l l d−l p−(2) n d l=2 2 d d d l l p−(2) =: n l l=2 l=2 a d l n − d −( l ) p 2 d−l al using and = n−l n .8. and → 0.e. Remark. the size of the is (2 + o(1)) log1/p n almost surely. By looking at the ratios l+1 . C so if d − l < C d If . . of G ∈ G(n. ad } and indeed al ≤ da2 + Cad where C is some constant. Since da2 ≤ 2d5 1 n2 p →0 ad = 1 p 1 µ(d) l because d ≤ (2 + ε) log1/p n when µ(d) → ∞. Therefore al+1 al l. largest complete 0 < p < 1 be xed. al al+2 a then > R l+1 . With more care. subgraph. We may assume al+1 (d − l)2 = al (l + 1)(n − l) d ≈ 2 log1/p n. a For larger 1<R< 1 p and l = o(log n) then l+1 < 1. we see al decreases then increases al d−l l al ≤ max{a2 . p) Let Then the clique number. we can compute the clique number to within 1. Remark. i. we are done. l=2 Corollary 6.

.

choosing y T corresponding to λmax (H). . and taking x= 1 √ (1. if n is odd. 2} ∈ E(G). so power of t is n − 2k + 2j . and hence involving n − k − (k − j) entries from bottom right. (A ij of length d from i to j . .) Since A is real symmetric. 1. A term in the expansion is obtained by taking j t's from top left block. Take a 2 + · · · + ξ 2 = 1. take an orthonormal basis unit vector x. Pick y corresponding to λmin (H). The adjacency matrix of G is the aij = Since 1 0 if ij ∈ E(G) otherwise .Chapter 7 Eigenvalue Methods Denition. . λmax := maxi λi then λmin = min|x|=1 xT Ax. . Now xT Ax = λ ξ 2 + ξ1 1 1 n E(G) = ∅. . if ξ1 e1 + · · · + ξn en and |x|2 e1 . which are easy 2 to compute and give useful information about G. {I. Adiam(G) } is linearly independent. say {1. it is t times such a polynomial. and likewise. . If G is bipartite then A = 0 ∗ . Let y be a vector in R|W | and let x be the vector corresponding to y with zeros in the coordinates [n] − W . after an appropriate relabelling of vertices. . 0. so x = = 2 · · · + λn ξn where λ1 . . 0) then 2 λmin ≤ −1. n×n matrix Let G be a A = (aij ). . . characteristic polynomial . In particular. Denition. . 0. λmax = max|x|=1 xT Ax. en of eigenvectors. involving k−j entries taken from bottom left. In general. (A )ij = k aik akj is d ) is the number of walks the number of walks of length 2 from i to j . A2 . W ⊂ [n] and let H be the induced subgraph G[W ]. −1. In particular. Hence Let λmin (G) ≤ λmin (H) ≤ λmax (H) ≤ λmax (G). Consider the ∗0 det(A − tI). For example. The set If G is a connected graph then the diameter diam(G) is the maximum distance between any two vertices. . taking x= 1 √ (1. . (Note that changing the labelling of the vertices of G is eectively just a change of basis and does not change the eigenvalues. A. . if λmin := mini λi . then (x ) A(x ) ≤ λmax (G). it is diagonalisable and has A is real symmetric. . if λ is a root. . λn are the eigenvalues of A. so in particular G has at least diam(G) + 1 distinct eigenvalues. 0) then 2 λmax ≥ 1. . . then (x )T A(x ) ≥ λmin (G). . In particular. Thus if n is 2 even. the characteristic polynomial is a polynomial in t . so is −λ with the same multiplicity. . graph with vertex set [n]. n real eigenvalues.

g. if if direction to each edge. Then j∈Γ(m) aj = ∆ which means d(m) = ∆ and aj = 1 for all j ∈ Γ(m). n i=1 xi = 0}. Given a graph G.  1  bvl = −1   0 l = uv l = vu for some for some u u. and 1 indeed taking x = √ (1. . we repeat to get that y = (1. Finally.48 Theorem 7. (i) (ii) If Let Eigenvalue Methods G be a graph. . 1. . if Similarly. For (iii) and (iv) note that we can recover any information about eigenvalues by considering components of is |aj | ≤ d(m) ≤ ∆(G) G. . is connected. n showing the rst part of (i). an orientation G of The incidence matrix B of G is the G is obtained by giving a n × e(G) matrix B = (bvl ). . Since G is connected. . ij∈E(G) ai aj . has if and only if G is regular and bipartite. µ1 ≤ µ2 ≤ · · · ≤ µn . L be G is L = BB T . . So G is bipartite as we can partition G according to whether ai = ±1. . T LX = (B T x)T (B T x) = 2 we have µ1 ≥ 0. d(j) = ∆ and ai = 1 for all i ∈ Γ(j). 1. 1. L=D−A The (combinatorial) Laplacian D = diag(d(1). . Then since L is positive semi-denite. otherwise on the graph so Denition. then n then (Ay)i = j∈Γ(i) aj and y T Ay = y T Ay = 2e(G) = d ≥ δ(G) where d is the average degree. . . 0 B Suppose now λ = ∆. Let and only if G is regular. 1. Note that where L does not depend on the orientation. . In particular. . j ∈ Γ(m). x ij∈E(G) (xi − xj ) . so µ1 ≥ 0. . )T . µ2 = min{xT Lx/ x 2 : x = 0. |λ| ≤ j∈Γ(m) proving (ii) and the second half of (i). proving (iii). −1.2. for all λ = −∆ then j∈Γ(m) aj = −∆. . 1). an ) ∈ Rn . Then δ(G) ≤ λmax ≤ ∆(G). 1 . Thus G is ∆-regular. The converse is clear since A(1. |λ| ≤ ∆(G) for all λ. ai = ±1 for all i and ai aj = −1 for all ij ∈ E(G). e. Then λ = (Ay)m = j∈Γ(m) aj . . n Let the eigenvalues of Proposition 7.e. We may assume am = 1. . . 1)T = −∆(1. . 1) we see µ1 = 0. . in which case λmax has multiplicity 1. 1.1. d(n)). Applying this argument with one of these j 's instead of m shows d(j) = ∆ and ai = 1 for all i ∈ Γ(j). −1. 2 Take y = (a1 . if G has two disjoint components then the adjacency matrix A 0 . y= 1 √ (1. a2 . . . . . i. . so d(m) = ∆ and aj = −1 for all j ∈ Γ(m). 1) and G is ∆-regular. in which case λmin Proof. λmin ≥ −∆(G). then also G (iii) (iv) λmax = ∆(G) if λmin = −∆(G) multiplicity 1. . 1. . Indeed. Let m ∈ [n] be such that am ≥ |ai | for all i.

is the graph Does there exist a is called a d-regular graph of diameter and order maximum possible order and equivalent to asking for Moore graph. Remark. girth 5. en with Lei = µi ei . . = 0. d = 2. . d-regular and contains no C3 2 or C4 . n n 2 ξi µi ≥ µ2 2 ξi = µ2 x i=2 2 xT Lx = i=2 But = µ2 kn(n − k). w) = 1}| = |Γ(v)| ≤ d |{w ∈ V : d(v. Since L is symmetric. . respectively. Such a These lead to Example. xT Lx = ij∈E(G) where (xi − xj )2 = |F |n2 U and F is the set of edges between If V − U. How big can n = |G| be if ∆(G) ≤ d and diam(G) ≤ 2? Pick any v ∈ V (G). i. U and V − U . .49 Proof. . . . 1). C5 . Then |{w ∈ V : d(v. xn ) where xi = n−k if i ∈ U and xi = −k if i ∈ V − U . . . . . the graph expands. d2 + 1? Note this diameter 2. and We have an orthonormal basis 1 √ (1. w) = 2}| ≤ d(d − 1) and hence n ≤ 1 + d + d(d − 1) = d2 + 1. . Then there are at least µ2 |U ||V − 0 < k = |U | < n = |V (G)|. . . there is an orthonormal basis of n 1 eigenvectors e1 . Consider the cases d = 1. Now suppose n n i=1 xi n e1 . G is If equality holds.3. . . µ2 is large. . en with Lei = µi ei where e1 = √ (1. and the Petersen graph. . . Writing x = n ξi ei we i=1 n 2 µi ξi ≥ 2 µ2 ξi = µ2 x i=1 2 We can take have e1 = ξ1 = x · e1 = 0 xT Lx = i=1 Theorem 7. Then x = i=2 ξi µi ei n We may assume and Proof. K2 . U |/|G| Let edges between G be a graph and U ⊂ V (G). d = 3.e. • • vv vv vv v vv A •A AA AA AA AA A • r • vv rr rr rr rr rr • ! !! !! !! !! !! • • vrr vv rrr rr vv rr vv rr vv A •A rr vv v r !! AA v… …v…… AA • • ! A iii! AA ……rr !! A viii !! • r • v A rr AA !! rr vvAvv !! AA r !! vvv rrAAA !! AA rT !! v !! • •T AA ØØv TT !! AA ØØ TT !! ØØ ! • • . 1). . Let x = (x1 . it has girth g(G) = 5.

So J be the matrix with all entries A3 = dA + aA2 + bBA = dA + aA2 + b[(d − 1 − a)A] + (d − b)[A2 − dI − aA]. a Moore graph is a strongly regular graph with parameters Let (d. a. BA = 0I + (d − 1 − a)A + (d − b)B . b). Kd+1 . λ2 with multiplicities r. Theorem 7. Remark. b). G be a strongly regular graph with parameters (d. we have A2 = dI + aA + bB . G being strongly regular with parameters d ˜˜˜˜ ˜’’’’˜˜˜˜˜˜ ’ v • ˜˜˜˜’’’’’’’’ ’’ a Γ(v) d−1−a b d−b V (G)−Γ(v)−{v} Let matrix of 1 and set B = J − I − A. Therefore. b) and order n. Eigenvalue Methods Denition. d is an eigenvalue of multiplicA satises a cubic. Hence d. so has only two other eigenvalues λ1 . a. a. λ1 . every pair of adjacent vertices has exactly adjacent vertices has exactly a common neighbours and every pair of non- common neighbours. Thus A satises a cubic. 1 2 n−1± (n − 1)(b − a) − 2d (a − b)2 + 4(d − b) are natural numbers. s.50 Note that all of these are unique solutions.4. . Now consider the eigenvalues of ity A. Equivalently to G being strongly regular with parameters (d. b) is regular of degree d. A3 − (d − b + a)A2 − [d(b − a) + d − b]A + d(d − b)I = 0. a. since bB = A2 − dI − aA. Then if if if  d  2 (A )ij = a   b which is equivalent to i=j ij ∈ E(G) ¯ ij ∈ E(G) (d. 1). so B is the adjacency ¯ G. But t3 − (d − b + a)t2 − [d(b − a) + d − b]t + d(d − b) = 0. Then A strongly regular b graph with parameters (d. λ2 are roots of 1. 0. Since G is d-regular. Proof. Let We may assume G is connected for otherwise b = 0 and each component is a A be the adjacency matrix of G. From the diagram.

s to obtain the result. and the special case d = 2. 5. 3. or 4d − 3 = l2 2 1 r = 2 {d2 + d −2d }. So l | 15. There exists a Moore graph only if d ∈ {1. Thus . λ2 = 1 {a − b ± 2 Now (a − b)2 + 4(d − b)}. in which case d2 − 2d = 0.51 so λ1 . Solve for r+s = n−1 and d + rλ1 + sλ2 = tr A = 0. and possibly Theorem 7. 57}. 7. 3.5. so d ∈ {1. l l ∈ {1. where So d = 2. 15}. r. 3. Recall that n = d 2 + 1. so l5 + l4 + 6l3 − 2l2 + (9 − 32r)l − 15 = 0. Proof. 7} d = 57. l ∈ Z. 2. so 1 2 Hence either d2 − 2d d2 + √ 4d − 3 ∈ N.