p/2 for all points v, then G is hamiltonian. (Lhis is stated below as Corollary 7.3(b).) For the above argument implies that every pair of points of G are adjacent, so G is complete. But this is a contradiction since K, is hamiltonian for all p > 3. Therefore there isa point » in G with deg » < p/2. Let mbe the maximum degree among all such points and choose v, so that deg v, = m. By hypoth- esis the number of points of degree not exceeding m is at most m < p/2. Thus there must be more than m points having degree greater than m and hence at least p/2. Lheretore there is some point, say v,, of degree at least p/2 not adjacent to v,. Since v, and 2, are not adjacent, there is a spanning path 0, 02 +++» As above, we write v,,,°**, v,, as the points of G adjacent to », and note that », cannot be adjacent to any of the m points n, , for 1 3 and 1 1, and form the graph G = K,,,+;, then G is not hamiltonian but violates the theorem only68 TRAVERSABILITY Syeee eye Fig. 7.4. The Tutte graph. by having (p — 1)/2 + 1 points of degree (p — 1)/2. The graph Gy = Ky,3 of Fig. 7.3 illustrates this construction for p = 5. By specializing Pésa’s Theorem, we obtain simpler but less powerful sufficient conditions due to Ore [03] and Dirac [D6] respectively. Corollary 7.a) If p > 3 and for every pair u and v of nonadjacent points, deg u + deg v > p, then G is hamiltonian. Corollary 7.6) If for all points v of G, deg v = p/2, where p > 3, then G is hamiltonian. Actually, the cubic hamiltonian graph G, of Fig. 7.3 has four spanning cycles. The smallest cubic hamiltonian graph, Ks, has three spanning cycles. These observations serve to illustrate a theorem of C. A. B. Smith which appears in a paper by Tutte [T6]. Theorem 7.4 Every cubic hamiltonian graph has at least three spanning cycles. Tait [T1] conjectured that every 3-connected planar graph* contains a spanning cycle. Tutte ['T6] settled this in the negative by showing that the 3-connected planar graph with 46 points of Fig. 7.4 is not hamiltonian. The smallest known nonhamiltonian triply connected planar graph, having 38 points, was constructed independently by J. Lederberg, J. Bosak, and D, Barnette; see Griinbaum [G10, p. 359]. The apparent lack of any relationship between eulerian and hamiltonian graphs is illustrated in Fig. 7.5 where each graph is a block with eight points. * See Chapter II for a discussion of planarity. Tait’s conjecture, if true, would have settled the Four Color Conjecture.EXERCISES, 69 Hamiltonian Nophamiltonian Fig. 7.5. Eulerian and/or hamiltonian graphs. However, in the next chapter we shall relate eulerian and hamiltonian graphs by way of the “line graph.” Incidentally, M. D, Plummer conjectures that the square of every 2-connected graph is hamiltonian, EXERCISES 7.1 Find an eulerian trail in the graph G of Fig, 7.1 and a partition of the lines of G into cycles. 72. Hevery block ofa connected graph Giseulerian, then G is eulerian, and conversely. 73 In Corollary 7.1(a), the partition cannot be done with fewer than 7 trails. State and prove the converse of Corollary 7.1(b). 74. A graph is arbitrarily traversable from a point vif the following procedure always results in an eulerian trail: Start at point vp by traversing any incident line; on arriving at a point u depart by traversing any incident line not yet used, and continue until no new lines remain. a) An eulerian graph is arbitrarily traversable from vy if and only if every eyele b) 1fG is arbitrarily traversable from vo, then vo has maximum degree, (Babler [B1) ¢) If G is arbitrarily traversable from vo, then cither vg is the only cutpoint or G has no cutpoints. (Harary [H17]) 75 Prove or disprove: If a graph G contains an induced theta subgraph, then G is not hamiltonian, 76 a) For any nontrivial connected graph G, every pair of points of G? are joined by a spanning path. Hence every line of G? is in a hamiltonian cycle. (Karaganis [K2]) b) If every pair of points of G are joined by a spanning path and p > 4, then G is 3-connected.70 TRAVERSABILITY 7.7 Give an example of a nonhamiltonian graph with 10 points such that for every pair of nonadjacent points u and v, degu + deg > p — 1. 7.8 How many spanning cycles are there in the complete bigraphs K3,; and K,,3? 79 A graph G is called randomly traceable [randomly hamiltonian] if a spanning path [hamiltonian cycle] always results upon starting at any point of G and then successively proceeding to any adjacent point not yet chosen until no new points are available. a) Agraph G with p = 3 points is randomly traceable if and only ifit is randomly hamiltonian, b) A graph G with p > 3 points is randomly traceable if and only ifit is one of the graphs C,, K,, or K,,, with p = 2n, (Chartrand and Kronk [CK1]) 7.10 Theorem 7.3 can be regarded as giving sufficient conditions for a graph to be 2-connected. This can be generalized to the n-connected case. Let G be nontrivial and let 1 3. and q > (p? ~ 3p + 6)/2, then Gis hamiltonian. {Ore [04]) 7.14 If for any two nonadjacent points wand v of G, deg u + deg v > p + 1, then there is a spanning path joining every pair of distinct points. (Ore [06]) 7.15 If G isa graph with p > 3 points such that the removal of any set of at most 7 points results in a hamiltonian graph, then G is (n + 2)-connected. (Chartrand, Kapoor, Kronk [CKK1]) 7.16 Consider the nonhamiltonian graphs G such that every subgraph G — v is hamiltonian. There is exactly one such graph with 10 points and none smaller. (Gaudin, Herz, and Rossi [GHR1]) 7.17 Do there exist nonhamiltonian graphs with arbitrarily high connectivity?CHAPTER 8 LINE GRAPHS A straight line is the shortest distance between two points. Eucup The concept of the line graph of a given graph is so natural that it has been independently discovered by many authors. Ofcourse, each gave it a different name* : Ore [O5] calls it the “interchange graph,” Sabidussi [S7] “derivative” and Beineke [B8] “derived graph,” Seshu and Reed [SR1] “edge-to-vertex dual,” Kasteleyn [K4] “covering graph,” and Menon [M10] “adjoint.” Various characterizations of line graphs are developed. We also introduce the total graph, first studied by Behzad [B4], which has surprisingly been discovered only once thus far, and hence has no other names. Relationships between line graphs and total graphs are studied, with particular emphasis on eulerian and hamiltonian graphs. Si THES RAP! Consider the set X of lines of a graph G as a family of 2-point subsets of V(G). The line graph of G, denoted L(G), is the intersection graph Q(X). Thus the points of L(G) are the lines of G, with two points of L(G) adjacent whenever the corresponding lines of G are. If x = uv isa line of G, then the degree of x in L(G) is clearly deg u + deg v — 2. Two examples of graphs and their line graphs are given in Fig. 8.1. Note that in this figure G, = L(G,), so that L(G,) = L(L{G,)). We write L(G) = L(G), E(G) = L(L(G)), and in general the iterated line graph is I'(G) = L(Z~"(G). As an immediate consequence of the definition of L(G), we note that every cutpoint of L(G) isa bridge of G which is not an endline, and conversely. When defining any class of graphs, it is desirable to know the number of points and lines in each; this is easy to determine for line graphs. * Hoffman [H46] uses “line graph” even though he chooses “edge.” Whitney | WII] was the first to discover these graphs but didn’t give them a name, nm2 LINE GRAPHS LG): 3 Gr: UG): Fig. 8.1, Graphs and their line graphs. Theorem 8.1 If G is a (p, g) graph whose points have degrees d;, then L(G) has q points and q, lines, where qu= ~9 +45 dP Proof. By the definition of line graph, L(G) has q points. The d, lines incident with a point v; contribute (4) to 4, so a= ¥(H)=sEaa~ 9 45a 454 40a The next result can be proved in many different ways, depending on ‘one’s whimsy. Theorem 82 A connected graph is isomorphic to its line graph if and only if it is a cycle. ‘Thus for a (not necessarily connected) graph, G = L(G) if and only if G is regular of degree 2. If G, and G, are isomorphic, then obviously L(G,) and L(G,) are. Whitney [W11] found that the converse almost always holds by displaying the only two different graphs with the same line graph. The proof given here is due to Jung [3]. Theorem 8.3 Let G and G’ be connected graphs with isomorphic line graphs. Then G and G’ are isomorphic unless one is K and the other is K,,. Proof. First note that among the connected graphs with up to four points, the only two different ones with isomorphic line graphs are K3 and K;,3.CHARACTERIZATIONS OF LINE GRAPHS 73 Note further that if ¢ is an isomorphism of G onto G’, then there is a derived isomorphism ¢, of L(G) onto L(G’). The theorem will be demonstrated when the following stronger result is proved. If G and G’ have more than four points, then any isomorphism ~, of L(G) onto L(G’) is derived from exactly one isomorphism of G to G'. We first show that ¢, is derived from at most one isomorphism. Assume there are two such, ¢ and y. We will prove that for any point v of G, d(v) = Y(v). There must exist two lines x = uv and y = uw or ow. If y = vw, then the points #(v) and ¥(v) are on both lines $,(x) and ¢,(9), so that since only cone point is on both these lines, ¢(v) = W(v). By the same argument, when y = un, gu) = Yu) so that since the line #,(x) contains the two points @(v) and pu) = yu), we again have (v) = y(v). Therefore ¢, is derived from at most one isomorphism of G to G’. We now show the existence of an isomorphism ¢ from which ¢, is derived. The first step is to show that the lines x, = uv,, x, = uv, and X3 = uv3 of a K,_; subgraph of G must go to the lines of a K, , subgraph of G' under ¢,. Let y be another line adjacent with at least one of the x;, which is adjacent with only one or all three. Such a line y must exist for any graph with p > 5 and the theorem is trivial for p < 5. If the three lines #,(x;) form a triangle instead of K,,; then $,(y) must be adjacent with precisely two of the three. Therefore, every K:,3 must go to a Ky, Let S(v) denote the set of lines at v. We now show that to each v in G, there is exactly one v’ in G’ such that S(v) goes to S(v') under $,. Ifdeg v > 2, let y, and y, be lines at v and let v’ be the common point of $,(y,) and ,()2). Then for each line x at v, ’ is incident with @ (x) and for each line x’ atv’, v is incident with gj '(x’). If deg v = 1, let x = uv be the line at v. Then deg u > 2 and hence S(u) goes to S(u’) and $,(x) = w’v’. Since for every line x’ at v’, the lines py '(x’) and x must have a common point, u is on #7 *(x') and u' is on x’. That is, x’ = @,(x) and deg v’ = 1. The mapping ¢ is therefore one-to-one from V to V’ since S(u) = S(v) only when u = v. Now given v’ in V’, there is an incident line x’. Denote 7 1(x’) by wv. Then either @(u) = v' or p(v) = v’ so @ 18 onto. Finally, we note that for each line x = uv in G, @,(x) = 6(u)@(v) and for each line x’ = u'r’ in G’, 67 (x) = 6 (wo '(v), so that @ is an iso- morphism from which ¢, is derived. This completes the proof. CHARACTERIZATIONS OF LINE GRAPHS A graph G is a line graph if it is isomorphic to the line graph L(H) of some graph H. For example, K, — x is a line graph; see Fig. 8.1. On the other hand, we now verify that the star K, 5 is not a line graph. Assume K,,, = L(H). Then H has four lines since K,,, has four points, and H must be connected. All the connected graphs with four lines are shown in Fig. 8.2.4 LINE GRAPHS Tt fe Fig. 8.2. The connected graphs with four lines, Since L(C,) = C, by Theorem 8.3 and L(K,,3 + x) = K, — x (see Fig. 8.1), it follows that H is one of the three trees. But the line graphs of these trees are the path P,, the graph K3~ K>, and K,, showing that K,,, is not a line graph. We will see that the star K,,, plays an important role in characterizing line graphs. The first characterization of line graphs, statement (2) of the next theorem and due to Krausz [K12], was rather close to the defini- tion. The situation was improved by van Rooij and Wilf [RW1] who were able to describe in (3) a structural criterion for a graph to be a line graph Finally, Beineke [B8] and N. Robertson (unpublished) displayed exactly those subgraphs which cannot occur in line graphs. Recall that an induced subgraph is one which is maximal on its point set. A triangle T of a graph G is called odd if there is a point of G adjacent to an odd number of its points, and is even otherwise. ‘Theorem 8.4 The following statements are equivalent: (1) Gisa line graph. (2) The lines of G can be partitioned into complete subgraphs in such a way that no point lies in more than two of the subgraphs. (3) G does not have K,,. as an induced subgraph, and if two odd triangles have a common line then the subgraph induced by their points is K,. (4) None of the nine graphs of Fig. 8,3 is an induced subgraph of G. Proof. (2)implies (2) Let G be the line graph of H. Without loss of generality we assume that H has no isolated points. Then the lines in the star at each point of H induce a complete subgraph of G, and every line of G lies in exactly ‘one such subgraph. Since each line of H belongs to the stars of exactly two points of H, no point of G is in more than two of these complete subgraphs. (2) implies (1) Given a decomposition of the lines of a graph G into complete subgraphs 5,,S2,-*~, S, satisfying (2), we indicate the construction of a graph H whose line graph is G. The points of H correspond to the set S ofCHARACTERIZATIONS OF LINE GRAPHS 75 Gy: Oy Ge: La Se ged Fig. 8.3. ‘The nine forbidden subgraphs for line graphs. subgraphs of the decomposition together with the set U of points of G belonging to only one of the subgraphs S,. Thus S U U is the set of points of H and two of these points are adjacent whenever they have a nonempty intersection ; that is, H is the intersection graph Q(S U U). (2) implies (4) it can be readily verified that none of the nine graphs of Fig. 8.5 can have its set of lines partitioned into complete subgraphs satisfying the given condition. Since every induced subgraph of a line graph must itself hea line graph, the result follows (4) implies (3) We show that if G does not satisfy (3), then it has one of the nine forbidden graphs as an induced subgraph. Assume that G has odd triangles abe and abd with c and d not adjacent. There are two cases, depending on whether or not there is a point v adjacent to an odd number of points of both odd triangles.76 LINE GRAPHS CASE 1. There is a point v adjacent to an odd number of points of triangle abcand of triangle abd. Now there are two possibilities : either v is adjacent to exactly one point of each of these triangles or it is adjacent to more than one point of one of them. In the latter situation, » must be adjacent to all four points of the two triangles, giving G3 as an induced subgraph of G. In the former, either v is adjacent only either to a or b, giving G,, or to both cand d, giving G. CASE 2, There is no point adjacent to an odd number of points of both triangles. In this case, let u and v be adjacent to an odd number of points in triangles abc and abd, respectively. There are three subcases to consider: Case 2.1. Each of u, v is adjacent to exactly one point of the corre- sponding triangle. Case 2.2. One of u, v is adjacent to all three points of “its” triangle, the other to only one. Case 2.3. Each of u, vis adjacent to all three points of the corresponding triangle. Before these alternatives are considered, we note two facts. If u or v is adjacent to a or b, then it is also adjacent to c or to d, since otherwise G, is an induced subgraph. Also, neither u nor v can be adjacent to both c and dsince then G, or Gs is induced. CASE 2.1. Ifuc, ud € G then, depending on whether or not line wv is in G, we have G, or G, as an induced subgraph. If ub, vd ¢ G then it follows from the preceding remarks that udeG while ve ¢G; if wv ¢G then points {a,d, u,v} induce G,, while if uv eG, then {a,b,c,d,u,v} induce Gs. If ub, va € G then necessarily ud, ve € G, so that if uv ¢ G, Gg is induced, while if uv ¢ Gthen G, appears. Finally ifub, vb ¢ G, then again ud, ve € G from which it follows that either Gy or G, is an induced subgraph of G, depending on whether or not uve G. CASE 2.2. Let ua, ud, uc eG. Clearly if ude G then G, is induced; thus ud ¢G. Now v can be adjacent to d or b. If vd eG, then depending on whether or not uv € G, we find G, or Gs induced. If vb eG then either G, or G, is induced, depending on whether or not v is adjacent to both ¢ and u. CASE 23. Ifud, ve, or uv € G, then G3 is induced. The only other possibility gives Gy. 3. implies 1. Suppose that G is a graph satisfying the conditions of the statement. We may clearly take G to be connected. Now, exactly one of the following statements must be true: 1. G contains two even triangles with a common line. 2, Whenever two triangles in G have a line in common, one of them is odd.SPECIAL LINE GRAPHS 77 Fig. 8.4, Three line graphs It can be shown that if G satisfies the first statement, then it is one of the graphs 1, = L(K,,3 + %, H, = L(H,) or H3 = {K,) displayed in Fig. 8.4. So suppose that G satisfies the second statement. We indicate the method of constructing a graph H such that G = L(H) Let F, be the family of all cliques of G which are not even triangles, where each such clique is considered as a set of points. Let F, be the family of points (taken as singletons) of G lying in some clique K in F, but not adjacent to any point of G — K. Finally, let F be the family of lines (each taken as a set of two points) of G contained in a unique and even uiangle. It is not difficult to verify that G is isomorphic to the line graph of the inter- section graph H = Q(F, u F, UF). This completes the proof. This iasi consiruciion is ifiusiraied in Fig. 8.3, in which ihe given graph G has families F, = {{1,2,3, 4}, (4,5, 6}}, F2 = {{1}, (2}, (3}}, and F3 = {{5, 7}, (6, T}} leading to the intersection graph H; thus G = L(H). iy 1 2 {5,7} {6.7} Fig. 8.5. A line graph and its graph. SPECIAL LINE GRAPHS In this section, characterizations are presented for line graphs of trees, complete graphs, and complete bigraphs The next result, due to G. T. Chartrand, specifies when a graph is the line graph of a tree.7B LINE GRAPHS _— 3 a a bs Fig. 8.6. The line-graph G of a tree T. ‘Theorem 8S A graph is the line graph of a tree if and only if it isa connected block graph in which each cutpoint is on exactly two blocks. Proof. Suppose G = L(T), T some tree. Then G is also B(T) since the lines and blocks of a tree coincide. Fach cutpoint x of G corresponds to a bridge uv of T, and is on exactly those two blocks of G which correspond to the stars at u and v. This proves the necessity of the condition. To see the sufficiency, let G be a block graph in which each cutpoint is on exactly two blocks. Since each block of a block graph is complete, there exists a graph H such that L(H) = G by Theorem 8.2. If G = K, we can take H = K,,3. IfGis any other block graph, then we show that H must be atree. Assume that H is not a tree so that it contains a cycle. If H is itself acycle, then by Theorem 8.3, L(H) = H, but the only cycle which isa block graph is K, a case not under consideration. Hence H must properly contain a cycle, thereby implying that H has a cycle Z and a line x adjacent to two lines of Z, but not adjacent to some line y of Z. The points x and y of L(H) lie on a cycle of L(H), and they are not adjacent. This contradicts the condition of Theorem 3.5 that L(H) is a block graph. Hence H is a tree, and the thcorem is proved. In Fig. 8.6, a block graph G is shown in which each cutpoint lies on just two blocks. The tree T of which G is the line graph is constructed by first forming the block graph B(G) and then adding new points for the non- cutpoints of G and the lines joining each block with its noncutpoints. The line graphs of complete graphs and complete bigraphs are almost always characterized by rather immediate observations involving adjacencies of lines in K, and K,,,. The case of complete graphs was independently settled by Chang [C7] and Hoffman [143], [H44].LINE GRAPHS AND TRAVERSABILITY 79 Theorem 8.6 Unless p = 8, a graph G is the line graph of K, if and only if 1. G has (§) points, 2. Gis regular of degree 2p — 2), 3. Every two nonadjacent points are mutually adjacent to exactly four points, 4. Every two adjacent points are mutually adjacent to exactly p — 2 points. It is evident that L(K,) has these four properties. It is not at all obvious that when p = 8, there are exactly three exceptional graphs satisfying the conditions. For complete bigraphs, the corresponding result was found by Moon [M13], and Hoffman [H46]. Theorem 8.7 Unless m = only if = 4, a graph G is the line graph of K,,,, if and 1. G has mn points, 2. G is regular of degree m +n — 2, 3. Every two nonadjacent points are mutually adjacent to exactly two points, 4, Among the adjacent pairs of points, exactly n(’) pairs are mutually adjacent to exactly m — 2 points, and the other m(3) pairs to n — 2 points. There is only one exceptional graph satisfying these conditions. It has 16 points, is not L(K,,,), and was found by Shrikhande [S12] when he proved Theorem 8.7 for the case m = n. LINE GRAPHS AND TRAVERSABILITY We now investigate the relationship of eulerian and hamiltonian graphs with line graphs. If x = wv isa line of G, and w is not a point of G, then x is subdivided when it is replaced by the lines uw and wv. Ifevery line of G is subdivided, the G S@): Fig. 8.7. A graph and its subdivision graph.80 LINE GRAPHS a” Fig. 8.8. A counterexample. > OO G LG) LAG) Fig. 8.9, Another counterexample. resulting graph is the subdivision graph S(G); see Fig. 8.7. If we denote by $,(G) the graph obtained from G by inserting n new points of degree 2 into every line of G, so that S(G) = S,(G), we can then define a new graph L,(G) = L{S,_,(G)). Note that, in general, L,(G) # L(G), the nth iterated line graph of G. Theorem 8.8 If G is eulerian, then L(G) is both eulerian and hamiltonian. If G is hamiltonian, then L(G) is hamiltonian, For example i in mes 88, L(G) is culerian and hamiltonian while G is not eulerian;; in Fig. 8.9, L(G) is hamiltonian while G is not A refinement of the second statement in Theorem 8.8 is provided by the following result of Harary and Nash-Williams [HN1] which follows readily from the preceding theorem and the fact that L(G) = L(S(G)). ‘Theorem 8.9 A sufficient condition for L(G) to be hamiltonian is that G be hamiltonian and a necessary condition is that L(G) be hamiltonian. The graphs of Figs. 8.10 and 8.9 show that the first of these conditions is not necessary and the second is not sufficient for L3(G) to be hamiltonian. Wenote also (see Fig. 8.11) that L(G) = L,(G) and L(G) may be hamiltonianLINE GRAPHS AND TRAVERSABILITY 81 A Fig. 8.10. Still another counterexample. ip OY LAG) L(G) Fig. 8.11. A sequence of graphs L(G). without G being culerian. However, the next graph L(G) in this scrics provides the link between these two properties. ‘Theorem 8.10 A graph G is eulerian if and only if L(G) is hamiltonian. For almost every connected graph G, however, nearly all of the graphs 1G) are hamiltonian, as shown by Chartrand [C9]. Theorem 8.11 If G is a nontrivial connected graph with p points which is not a path, then Z(G) is hamiltonian for all n > p — 3. An example is given in Fig. 8.12 in which a 6-point graph G, as well as L(G), 2(G), and the hamiltonian graph [3(G) are shown. YY Ps LO LG) Fig. 812. A sequence of iterated line graphs.82 LINE GRAPHS { = | Fig. 813. Formation of a total graph. TOTAL GRAPHS The points aud lines of a graph are called its elements. Two elements of a graph are neighbors if they are either incident or adjacent. The total graph T(G) has point set V(G) U X(G), and two points of T(G) are adjacent whenever they are neighbors in G. Figure 8.13 depicts the formation of the total graph T(K;). It is easy to see that T(G) always contains both G and L(G) as induced subgraphs. An alternative characterization of total graphs was given by Behzad [B4]. Theorem 8.12 The total graph T(G) is isomorphic to the square of the sub- division graph S(G). Corollary 8.12(a) If v is a point of G, then the degree of point v in T(G) is 2degv. If x = w is a line of G, then the degree of point x in T(G) is deg u + deg v. Corollary 8.12(b) If G is a (p, q) graph whose points have degrees d,, then the total graph T(G) has pr = p + q points and qr = 2q + $Z d lines. The Ramsey function r(m, n) was defined in Chapter 2, where it was noted that its general determination remains an unsolved problem. Behzad and Radjavi [BR1] defined and solved an analogue of the Ramsey problem, suggested by line graphs. The line Ramsey number r,(m, n) is the smallest positive integer p such that every connected graph with p points contains either n mutually disjoint lines or m mutually adjacent lines, that is, the star K,,,- Thus r;(m, n) is the smallest integer p such that for any graph G with p points, L(G) contains K,, or L(G) contains K,, ‘Theorem 8.13 For n >1, we always have r,(2, n) = 3. For all other m and n, r(m, n) = (m — 1Xn — 1) + 2. Note that it is not always true that r,(m, n) = r,(n,m). Furthermore, in contrast with Ramsey numbers, r,(m,n) is defined only for connected graphs.EXERCISES 83 EXERCISES 8.1 Under what conditions can the lines of a line graph be partitioned into complete subgraphs so that cach point lies in exactly two of these subgraphs? 8.2 Determine the number of triangles in L(G) in terms of the number n of triangles of Gand the partition of G. 83 Determine a criterion for a connected graph to have a regular line graph. 84 A graph G can be reconstructed from the collection of q spanning subgraphs G — x, ifand only if its line graph L(G) satisfies Ulam’s Conjecture (p. 12). (Hemminger [H41)) 85 IfG is mline-connected, then 1. L(G) is n-connected, 2. 1{G) is Qn — 2}-line-connected, and 3. L(G) is (2n — 2)-connected. (Chartrand and Stewart [CS1]) 8.6 a) Construct a connected graph G with p > 4 such that L(G) is not eulerian but LG) is. on Lod gra BG) is. 8&7 The smallest block whose line graph is not hamiltonian is the theta graph with 8 points in which the distance between the points of degree 3 is 3. (J. W. Moon) 88 L(G) is hamiltonian if and only if there is a closed trail in G which includes at least one point incident with each line of G. 89 The graph L,(G) is hamiltonian if and only if G has a closed spanning trail. (Harary and Nash-Williams [HN1]) 8.10. The following statements are equivalent (1) L(G) is eulerian. (2) The degrees of all the points of G are of the same parity. (3) T(G)is eulerian 8.11 T(K,) is isomorphic to L(K,+,). (Behzad, Chartrand, and Nordhaus [BCN1]) 8.12 Define a family F of subsets of elements of G such that T(G) — Q(F). 8.13 a) IfG is hamiltonian, so is T(G). If G is eulerian, then T(G) is both eulerian and hamiltonian. b) The total graph T(G) of every nontrivial connected graph Gcontainsa spanning eulerian subgraph, ©) Ifa nontrivial graph G contains a spanning eulerian subgraph, then T(G) is hamiltonian, 4) If G is nontrivial and connected, then T2(G) is hamiltonian. (Behzad and Chartrand [BC2]) 8.14 For any multigraph M, define the line graph L(M) by V(L(M)) = X(M) and x adj y in L(M) whenever x and y are distinct lines meeting at either one or two points. Then @ graph G is the line graph of some multigraph if and only if it has no induced subgraph of the form G,, Gg, or Gy of Fig. 8.3. (. P. Geller)CHAPTER 9 FACTORIZATION The whole is equal to the sum of its parts. Fuicun, Elements ‘The whole is greater than the sum of its parts. Max WeRTHEIMER, Productive Thinking A problem which occurs in varying contexts is to determine whether a given graph can be decomposed into line-disjoint spanning subgraphs possessing a prescribed property. Most frequently, this property is that of regularity of specified degree. In particular, a criterion for the existence in a graph of a spanning regular subgraph of degree 1 was found by Tutte. Some observa- tions are presented concerning the decomposition of complete graphs into spanning subgraphs regular of degree 1 and 2. The partitioning of the lines of a given graph into spanning forests is also studied and gives rise to an invariant known as “arboricity.” A formula for the arboricity of a graph in terms of its subgraphs was derived by Nash- Williams, and explicit constructions for the minimum number of spanning forests in complete graphs and bigraphs have been devised. 1-FACTORIZATION A factor of a graph G is a spanning subgraph of G which is not totally dis- connected. We say that G is the sum* of factors G; if it is their line-disjoint union, and such a union is called a factorization of G. An n-factor is regular of degree n. If G is the sum of n-factors, their union is called an n-factorization and G itself is n-factorable. Unless otherwise stated, the results presented in this chapter appear in or are readily inferred from theorems in Kénig [K10, pp. 155-195], where the topic is treated extensively. When G has a 1-factor, say Gy, it is clear that p is even and the lines of G, are independent. In particular, K2,,, cannot have a 1-factor, but K3, certainly can. * Some call this product; others direct sum.1-FACTORIZATION 85 “LA +/ / ) pea NU ON | | a Fig. 9.1. A L-factorization of Kg Theorem 9.1. The complete graph K , is 1-factorable. Proof. We need only display a partition of the set X of lines of K3, into (2n — 1) I-factors. For this purpose we denote the points of G by Dy, Vy,"*, Vg, and define, for i= 1,2,---,2n —1, the sets of lines X= {02q} U (0) i253) = 1217+, — 1}, where each of the subscripts i ~ jandi + jis expressed as one of the numbers 1, 2,---, (2n — 1) modulo (2n — 1). The collection {X;} is easily seen to give an appropriate partition of X, and the sum of the subgraphs G; induced by X; is a |-factorization of Ka, For example, consider the graph K shown in Fig. 9.1. The 1-factoriza- tion presented in the proof of the theorem produces the five !-factors G,. Although the complete bigraphs K,,, have no I-factor if m # n, the graphs K,,, are 1-factorable, as seen by the next statement. ‘Theorem 9.2. Every regular bigraph is 1-factorable. It is not an easy problem to determine whether a given graph is 1- factorable, or, indeed, to establish whether there exists any 1-factor. Beineke and Plummer [BP2] have shown, however, that many graphs cannot have exactly one 1-factor. Theorem 9.3 If a 2-connected graph has a 1-factor, then it has at least two different 1-factors. ‘The graph G in Fig. 9.2 is a block with exactly two 1-factors, and they ave one common line. The most significant result on factorization is due to Tutte [T7] and characterizes graphs possessing a 1-factor. In general, this test for a I-factor is quite inconvenient to apply. The proof given here is based on Gallai [G1].86 FACTORIZATION ? wl 4% \ | Fig. 9.2. Two I-factors of a block. A set of mutually nonadjacent lines is called independent. By an odd component of G we mean one with an odd number of points. Theorem 9.4 A graph G has a 1-factor if and only if p is even and there is no set S of points such that the number of odd components of G — S exceeds ISI. Proof. The easier half of this theorem is its necessity. Let S be any set of points of G and let H be a component of G — S. Inany !-factor of G, each point of H must be paired with either another point of H or a point of S. But if H has an odd number of points, then at least one point of H is matched with a point of S. Let ko be the number of odd components of G — S. If G has a 1-factor then |S] > ko, since in a 1-factor each point of S can be matched with at most one point of G — S and therefore can take care of at most one odd component. In order to prove the sufficiency, assume that G does not have a 1-factor, and let S be a maximum set of independent lines. Let T denote the set of lines not in S, and let uo be a point incident only with lines in T. A path is called alternating if the lines allernately lie in S and T. For each point v # ug, call va 0-point if there are no up-v alternating paths; if there is such a Path, call v an S-point if all these paths terminate in: a line of § at v, a T- in pry type of line. The following s statements a are immediate consequences, Every point adjacent to ug is a T- or an ST-point. No S- or O-point is adjacent to any T- or ST-point. No T-point is joined by a line of $ to any T- or 0-point. Therefore, each S-point is joined by aline ofS toa T-point. Furthermore, each T-point v is incident with a line of S since otherwise the lines in an alternating ug~v path could be switched between S and T to obtain a larger independent set.1-FACTORIZATION 87 Let H be the graph obtained by deleting the 7-points. One component of H contains ug, and any other points in it are ST-points. The other com- ponents either consist of an isolated S-point, only ST-points, or only 0-points. We now show that any component H, of H containing ST-points has an odd number of them. Obviously H, either contains up or has a point u, joined in G to a T-point by a line of S such that some alternating ug—u, path contains this line and no other points of H,. If H, contains uo, we take uy = Uo. The following argument will be used to show that within H, every point v other than u, is incident with some line of S. This is accomplished by showing that there is an alternating u,—» path in H, which terminates in a line of S. The first step in doing this is showing that if there is an alternating u,-v path P,, then there is one which terminates ina line of S. Let P; be an alternating up-v path ending in a line of T, and let uv’ be the last line of P,, if any, which does not lie in H,. Then v’ must be a T-point and wv’ a line in S. Now go along P; from u, until a point w’ of P is reached. Continuing along P, in one of the two directions must give an alternating path. If going to v' results in an alternating path, then the original ug-u, path Po followed by this new path and the line v'u’ would be a up-u, path terminating in a line of S and u’ could not be a T-point. Hence there must be a u,—v path terminating in a line of S. Now we show that there is necessarily a u,-v alternating path by as- suming there is not. Then there is a point w adjacent to v for which there is a u,-w alternating path. If line wv is in S, then the u,-w alternating path terminates in a line of T, while if wo is in T, the preceding argument shows there is a uy~w path terminating in a line of S. In either case, there is a u,v alternating path. ‘his shows that ihe component fi, has an odd number of poinis, and that if H, does not contain up, exactly one of its points is joined to a T-point by a line of S. Hence, with the exception of the component of H containing up and those consisting entirely of 0-points, each is paired with exactly one T-point by a line in S. Since each of these and the component containing Ug is odd, the theorem is proved. The graph of Fig. 9.3 has an even number of points but contains no 1- factor, for if the set S = {0,, v,} is removed from G, four isolated points (and therefore four odd components) remain. ey va Fig. 9.3. A graph with no 1-factor.88 FACTORIZATION Building up on his criterion for the existence of a !-factor in a given graph, Tutte [T10] was able to characterize those graphs having a spanning subgraph with prescribed degree sequence, and later, [T11], proved this result asa straightforward consequence of Theorem 9.4. Consider a labeling of G and a function f from V into the nonnegative integers. Let S and T be disjoint subsets of V, let H be a component of G — (SU T), and let 4(H, T) be the number of lines of G joining a point of H with one in T. Then we may write ko(S, 7) as the number of components H of G ~ (S UT) such that q(H, T) + Eyer f(u) is odd. Theorem 9.5 Let G be a given graph and let f be a function from V into the nonnegative integers. Then Ghasa spanning subgraph whosedegree sequence is prescribed by f if and only if there exist disjoint sets S and T of points such that tr) < ko(S, T) + Lye ~ dg-s(v)]- 2-FACTORIZATION Ifa graph is 2-factorable, then each factor must be a union of disjoint cycles. If a 2-factor is connected, it is a spanning cycle. We saw that a complete graph is I-factorable if and only if it has an even number of points. ‘Since are not 2-factorable. The odd complete graphs are 2-factorable, and in fact a stronger statement can be made. Ze Zz |. A 2-factorization of K;.2-FACTORIZATION 89 Theorem 9.6 The graph K2,., is the sum of n spanning cycles. Proof. In order to construct n line-disjoint spanning cycles in K 2,4, first label its points v4, ¥2,°-*, ¥2941- Then construct n paths P; on the points Py Ba, “7s Vay aS follows: Py = vj 0-4 Bj) 4 277" Me View Thus the jth point of P; is », where k = i + (—1)/*![j/2] and all subscripts are taken as the integers 1, 2,---, 2n (mod 2n). The spanning cycle Z, is then constructed by joining v2,,, to the endpoints of P;. This construction is illustrated in Fig. 9.4 for the graph Ky. The lines of the paths P, are solid and the two added lines are dashed. ‘There is a decomposition of K,, which embellishes the result of Theorem 91. ‘Theorem 9.7 The complete graph K2, is the sum of a I-factor and n — 1 spanning cycles. Of course, every regular graph of degree 1 is itself a 1-factor and every regular graph of degree 2 is a 2-factor. If every component of a regular graph G of degree 2 is an even cycle, then G is also 1-factorable since it can be expressed as the sum of two I-factors. Ifa cubic graph contains a 1-factor, it must also have a 2-factor, but there are many cubic graphs which do not have 1-factors. The graph of Fig. 9.5 has three bridges. Petersen [P3] proved that any cubic graph without a 1-factor must have a bridge. Fig. 9.5. A cubic graph with no I-factor. Fig. 9.6. The Petersen graph. Theorem 98 Every bridgeless cubic graph is the sum of a I-factor and a 2-factor. a bridgeless cubic graph which is not the sum of three I-factors. This well known graph, shown in Fig. 9.6, is called the Petersen graph. By Theorem 9.8, it is the sum of a 1-factor and a 2-factor. The pentagon and pentagram together constitute a 2-factor while the five lines joining the pentagon with the pentagram form a 1-factor.90 FACTORIZATION A criterion for the decomposability of a graph into 2-factors was also obtained by Petersen [P3]. Theorem 9.9 A connected graph is 2-factorable if and only if it is regular of even degree, ARBORICITY In the only type of factorization considered thus far, each factor has been an n-factor. Several other kinds of factorizations have been investigated and we discuss one now and others in Chapter 11. Any graph G can be expressed as a sum of spanning forests, simply by letting each factor contain only one of the q lines of G. A natural problem is to determine the minimum number of line-disjoint spanning forests into which G can be decomposed. This number is called the arboricity of G and is denoted by ¥(G). For example, Y(K,) = 2 and Y(K;) = 3; minimal decompositions of these graphs into spanning forests are shown in Fig. 9.7. LINZ war Fig. 9.7. Minimal decompositions into spanning forests. A formula discovered by Nash-Williams [N2] gives the arboricity of any graph. Theorem 9.10 Let G be a nontrivial (p,q) graph and let g, be the maximum number of lines in any subgraph of G having n points. Then = m Wn 116) = max, | }. The fact that Y(G) > max, {q,/(n — 1)} can be shown as follows. Since G has p points, the maximum number of lines in any spanning forest isp — 1. Hence, the minimum possible number of spanning forests requiredARBORICITY — 91. to fill G, which by definition is Y(G), is at least g/(p — 1). But the arboricity of Gis an integer, so Y(G) > {q(p — 1)}. The desired inequality now follows from the fact that for any subgraph H of G, Y(G) > Y(H). Among all subgraphs H with n < p points, max Y(H) will occur in those induced subgraphs containing the greatest number of lines. Thus if H is a subgraph of G, Y(H) can be greater than {q/(p — 1)}. The (10, 15) graph in Fig. 98 illustrates this observation. Takingn = Sandq, = 10(for H = Ks), we have a \_ i 4 wy =f bas>2ef a}. For K,, the maximum value of q, clearly occurs for n = p so that Y(K,) = {p/2}- Similarly, for the complete bigraph K,,,, {g,/(n — 1)} assumes its maximum value when n = p = r + s. 2 Y \ Fig. 9.8. A graph G with a dense subgraph H. Corollary 9.10(a) The arboricities of the complete graphs and bigraphs are rik = {zh and nn,) {2}. Although Nash-Williams’ formula gives the minimum number of spanning forests into which an arbitrary graph can be factored, his proof does not display a specific decomposition. Beineke [BS] accomplished this for complete graphs and bigraphs, the former of which we present here. For p = 2n, K, can actually be decomposed into n spanning paths. Labeling the Points vj, V2, “+ ~, U2_, We consider the same n paths Py = 0; 0-1 ren Vin Via Pin Mam as in proof of Theorem 9.6, For p = 2n + 1, the arboricity of K, isn + 1 by Corollary 9.10(a). A decomposition is obtained by taking the paths just described, adding an extra point labeled v,, , to each, and then constructing92 FACTORIZATION Se Lge WIS Fig. 9.9. A minimal decomposition of K, into spanning forests. a star by joining v2,,, to the other 2n points. The construction for p = 9 is shown in Fig. 9.9. It is easily seen to consist of the star at one of the points of K, together with spanning subforests corresponding to the four spanning paths of Kg indicated above. EXERCISES 9.1 The graph K,, has a unique 1-factorization. Find the number of 1-factorizations of K3,3 and of Ky. 9.2 Display a 1-factorization for Ky. 9.3. The number of 1-factors in Ky is (2n)!/(2"n!). 9.4 Keq-2 has a 3-factorization. 95 Forn > 1, K4,., is 4-factorable. 9.6 Use Tutte’s Theorem 9.4 to show that the graph of Fig. 9.5 has no I-factor. 9.7 Ifan n-connected graph G with p even is regular of degree n, then G hasa 1-factor. (Tutte [T7)) 98 Let G be a graph with a I-factor F, A line of G is in more than one actor if'and only if it lies on a cycle whose lines are alternately in F. (Beineke and Plummer [BP2]) 9.9 Express K as the sum of four spanning cycles. 9.10 Is the Petersen graph hamiltonian?EXERCISES — 93 *9.11 Corresponding to any two integers d > 3 and g > 3, there exists a graph G with the following properties: L wR wND 9.12 9.13 Gis regular of degree d. . G has girth g. . Gis hamiltonian. .. The cycles of length g are line-disjoint and constitute a 2-factor of G. . G is the sum of this 2-factor and (d — 2) 1-factors, (Sachs [S9]) Display a minimal decomposition of K,,, into spanning forests. Find the smallest connected (p, q) graph G such that max, {q,/(r — 1)} > {al(p — 1)}, where q, is the maximum number of lines in any induced subgraph of G with r points.CHAPTER 10 COVERINGS Through any point not on a given line, there passes a unique line having no points in common with the given line. Through any point not on a given line, there passes no line having no points in common with the given line. RIEMANN, Through any point not on a given line, there pass more than one line having no points in common with the given line. BOLyal It is natural to say that a line x = uv of G covers the points wand v. Similarly, we may consider each point as covering all lines incident with it. From this viewpoint, one defines two invariants of G: the minimum number of points (lines) which cover all the lines (points). Two related invariants are the maximum number of nonadjacent points and lines. These four numbers associated with any graph satisfy several relations and also suggest the study of spovial puinis and lines witich are critical for covering purposes. These concepts lead naturally to two special subgraphs of G called the line-core and point-core. Criteria for the existence of such subgraphs are established in terms of covering properties of the graph. COVERINGS AND INDEPENDENCE A point and a line are said to cover each other if they are incident. A set of points which covers all the lines of a graph G is called a point cover for G, while a set of lines which covers all the points is a line cover. The smallest number of points in any point cover for G is called its point covering number and is denoted by %o(G) or @g. Similarly, «(G) or «, is the smallest number of lines in any line cover of G and is called its line covering number. For example, a(K,) = p — 1 and a,(K,) = [(p + 1)/2]. A point cover (line cover) is called minimum if it contains a (respectively «,) elements. Observe that a 94COVERINGS AND INDEPENDENCE 95 Fig. 10.1. The graph Ky Ky. point cover may be minimal without being minimum; such a sct of points is given by the 6 noncutpoints in Fig. 10.1. The same holds for line covers ; the 6 lines incident with the cutpoint serve. A set of points in G is independent if no two of them are adjacent. The largest number of points in such a set is called the point independence number of G and is denoted by fo(G) or Bo. Analogously, an independent set of lines of G has no two of its lines adjacent and the maximum cardinality of such a sei is the dine independence number f(G} or py. For ihe compicic grapir, Bo(K,) = 1 and B,(K,) = [p/2]. Obviously B,(G) = p/2 if and only if G has a I-factor. The numbers just defined are Bo(G) = 2 and B,(G) = 3 for the graph G of Fig. 10.1. For this graph as well as for K,,a + Bo = % + B, = p. Gallai [G2] proved that this identity always holds. Theorem 10.1 For any nontrivial connected graph G, a + Bo =P = % + fr Proof. Let My beany maximum independent set of points, so that |Mol = Bo. Since no line joins two points of Mg, the remaining set of p ~ Bo points constitutes a point cover for G so that %» < p — Bo. On the other hand, if No is a minimum point cover for G, then no line can join any two of the remaining p — 1, points af G, sa the set V — Ny is independent Hence, Bo = p — %p, proving the first equation. To obtain the second equality, we begin with an independent set M, of B, lines. A line cover Y is then produced by taking the union of M, and a set of lines, one incident line for each point of G not covered by any line in My. Since |M,| +|Y¥| = p and |¥| > ay, it follows that a, + B, p, completing the proof of the theorem.96 ‘COVERINGS Hedetniemi [H39] noticed that the proof of the first equation in Theorem 10.1, % + Bo = applies in a more general setting. A property P of a graph G is hereditary if every subgraph of G also has this property. Examples of hereditary properties include a graph being totally disconnected, acyclic, and bipartite. Aset S of points of G is called a P-set if the induced subgraph has property P; it is called a P-set if every subgraph of G without property P contains a point of S. Let Bo(P) be the maximum cardinality of a P-set of G and let ao(P) be the minimum number of points of a P-set. Then the proof of the next statcment is obtained at once from that of Theorem 10.1 Corollary 10.1(a) If P is an hereditary property of G, then ao(P) + Bo(P) = p. A collection of independent lines of a graph G is sometimes called a matching of G since it establishes a pairing of the points incident to them. For this reason, a set of B, independent lines in Gis called a maximum matching of G. If Gis bipartite, then more can be said. The next theorem due to Konig [K9] is intimately related to his Lheorem 5.18 on systems of distinct representatives stated in matrix form, in fact it is the same result. ‘Theorem 10.2 If G is bipartite, then the number of lines in a maximum matching equals the point covering number, that is, By = a. The problem of finding a maximum matching, the so-called matching problem, is closely related to that of finding a minimum point cover. Let M c X(G) be a matching of G. In an alternating M-walk, exactly ‘one of any two consecutive lines is in M. An augmenting M-walk is an alternating M-walk whose endpoints are not incident with any line of M. Such a walk must be a path because M is a matching. IfG has no augmenting M-walk, then matching M is unaugmentable. Clearly every maximum matching is unaugmentable; the converse is due to Berge [B10] and the proof given below appears in Norman and Rabin [NR1]. Theorem 10.3 Every unaugmentable matching is maximum. Proof. Let M be unaugmentable and choose a maximum matching M' for which |M ~ M'|, the number of lines which are in M but not in M’, is minimum. If this number is zero then M = M’. Otherwise, construct a walk W of maximum length whose lines alternate in M — M' and M’. Since M’isunaugmentable, walk W cannot begin and end with lines of M — M'‘and has equally many lines in M — M’ and in M’. Now we form a maximum matching N from M’ by replacing those lines of W which are in M’ by the lines of Win M — M'. Then|M — N| <|M — M1, contradicting the choice of M’ and completing the proof.CRITICAL POINTS AND LINES 97 Norman and Rabin [NR1] developed an algorithm, based on the next theorem, for finding all minimum line covers in a given graph. Let Y be a line cover of G. An alternating Y-walk is a Y-reducing walk if its endlines arc in Y and its endpoints arc incident to lines of Y which are not cndlines of the walk, Obviously every minimum line cover has no reducing walk. ‘Theorem 10.4 If Y is a line cover of G such that there is no Y-reducing walk, then Y is a minimum line cover. ‘The cover invariants a and a, of G refer to the number of points needed to cover all the lines and vice versa. We may also regard each point as covering itself and two points as covering cach other if they are adjacent, and similarly for lines. Then other invariants suggest themselves. Let agg be the minimum number of points needed to cover V, and let dog be the minimum number* of independent points which cover V. Then both these numbers are defined for any graph. Let a1, and a, have similar meanings for the covering of lines by lines. The relationships among these invariants were determined by Gupta [G11]. Theorem 10.5 For any graph G, oS %q and, = Hs CRITICAL POINTS AND LINES Obviously, if H is a subgraph of G, then ao(H) < a9(G). In particular, this inequality holds when H = G — v or H = G — x for any point vor line x. Iag(G — v) < ao(G), then v is called a criticalt point; if 1(G — x) < a9(G), then x is a critical line of G. Clearly, if v and x are critical, it follows that ao(G — v) = a(G — x) = &% — 1. Critical points are easily characterized. Theorem 10.6 A point v is critical in a graph G if and only if some minimum point cover contains v. Proof. If M isa minimum point cover for G which contains v, then M — {v} covers G — v; hence, x(G — v) <|M — {v}| = |M| — 1 = a(G) — 1 so that v is critical in G. Let v be a critical point of G and consider a minimum point cover M’ for G — v. The set M’ u {v} is a point cover for G, and since it contains one more element than M’, it is minimum. If the removal of a line x = uv from G decreases the point covering number, then the removal of u or v must also result in a graph with smaller point covering number. Thus, if a line is critical both its endpoints are * Berge [B12] calls ao the “external stability number” and fj, the “internal stability number.” + In this chapter, “critical” refers to covering; in Chapter 12, the same word will involve coloring. ‘The meanings should be clear by context.98 COVERINGS Fig. 10.2. Line-critical graphs. critical. If a graph has critical points, it need not have critical lines; for example, every point of C, is critical but no line is. A graph in which every point is critical is called point-critical while one having all lines critical is line-critical. Thus a graph G is point-critical if and only if each point of G lies in some minimum point cover for G. From our previous remarks, every line-critical graph is point-critical. Among the line-critical graphs are the complete graphs, the cycles of odd length, and the graphs of Fig. 10.2. A constructive criterion for line-critical graphs is not known at present ; however, the first two corollaries to the following theorem of Beineke, Harary, and Plummer [BHP1] place some rather stringent conditions on such graphs. Theorem 10.7. Any two adjacent critical lines of a graph lie on an odd cycle. Corollary 10.7(a) Every line-critical graph isa block in which any two adjacent lines lie on an odd cycle. Theorem 10.7 was derived by generalizing the next result due to Dulmage and Mendelsohn [DM1] Coroliary 10.7(6) Any two critical lines of a bipartite graph are independent. LINE-CORE AND POINT-CORE The line-core* C,(G) of a graph G is the subgraph of G induced by the union ofall independent sets Y of lines (if any) such that |Y| = 29(G). This concept was introduced by Dulmage and Mendelsohn [DM1], who made it an integral part of their theory of decomposition for bipartite graphs. It is not always the case that a graph has a line-core, though by Theorem 10.2, every bipartite graph which is not totally disconnected has one. As an example of a graph with no line-core, consider an odd cycle C,. Here we find that ao(C,) = (p + 1)/2 but that B,(C,) = (p — 1)/2, so C, has no line-core. * Called “core” in [DMI] and [HP19}.LINE-CORE AND POINT-CORE 99 Fig. 10.3. A graph and its line-core. Harary and Plummer [HP19] developed a criterion for a graph to have a line-core. A minimum point cover M fora graph G with point set V is said to. be external if for each subset M’ of M, |M’| < |U(M’], where U(M’) is the set of all points of V — M which are adjacent to a point of M’. Theorem 10.8 The following are equivalent for any graph G: (1) Ghas a line-core. 2) Ghas an external minimum point cover. (3) Every minimum point cover for G is external. ‘As an example, consider the graph G of Fig. 10.3. This graph has two minimum point covers: M, = {v, Us, Ug} and Mz = {vz, vs, 07}. Let us concentrate on M,. If Mj = M,, then U(M}) = {0,, 03, 04, 07}. For Mi Dg, Vg}, U(M7) = {03, U4, 07}. We observe that |M4| < |U(M))| and |M’| <|U(M")|, a fact which is true for every subset of M,; hence, by definition, M, is external. Obviously, M, is also external. On the other hand, there are graphs which are equal to their line-core. This family of graphs is characterized in the next theorem, given in [HP19]. Following the terminology of Dulmage and Mendelsohn [DMI], we consider a bigraph G whose point set V is the disjoint union $ U T. We say that G is semi-irreducible if G has exactly one minimum point cover M and either M.A S or MO Tisempty. Next, G is irreducible if it has exactly two minimum point covers M, and M, andeitherM, 0 S = gandM, \T = ¢ or M, T = ¢ and M, 0S = 9. Finally, G is reducible if it is neither irreducible nor semi-irreducible. Ge 7 Gey us Fig. 10.4. A semi-irreducible and an irreducible graph.100 COVERINGS Theorem 10.9 A graph G and its line-core C,(G) are equal if and only if G is bipartite and not reducible. Consider the bigraphs G, and G, of Fig. 10.4. In G,, let S; = {v3, ve} and T, = {vj, U2, U4, Us, V7}. The bigraph G, has the unique minimum point cover M, = {v3, vg}, and since M, 1 T, = ¢, G, is semi-irreducible and hence equals its line-core. In'Gz, set Sy = {uy, Us, us} and T, = {2, ua, Ug}. There are two minimum point covers, namely, M, = {u,, uy, us} and Ny = {uz, Uy, Ug}. However, My 0 T, = $ and Nz 7 Sz = ¢; therefore, G, is irreducible and also equals its line-core. EXERCISES 10.1 Prove or disprove: Every point cover of a graph G contains a minimum point cover. 10.2 Prove or disprove: Every independent set of lines is contained in a maximum independent set of lines. 103 For any graph G. aa(G) > B.(G) and a,(G) > o(G). 10.4 Find a necessary and sufficient condition that @,(6) = B,(G). 10.5 IfG has a closed trail containing a point cover, then L(G) is hamiltonian. 106 For any graph G, #(G) > 4(6). 10.7 IfGisa bigraph then q < doo, with equality holding only for complete bigraphs. 108 If Gis acomplete n-partite graph, then a) % -5—K— b) Gis hamiltonian if and only if p < 2a. ©) If G is not hamiltonian, then its circumference c = 2a and G has a unique minimum point cover. (M.D. Plummer) d) B; = min {6, [p/2]}. (Chartrand, Geller, Hedetniemi [CGH2]) 10.9 a) Let f, be the maximum number of points in a set Sc V(G) such that ¢S) is disconnected, Then x = p — By. b) Defining B, analogously, 4 = q — Py. (Hedetniemi [1139] 10.10 Calculate: a) @4(K,) —B) Ho0(Kmn 11 Kn 10.11. The “chess-queen graph” has the 64 squares of a chess board as its points, two of which are adjacent whenever one can be reached from the other by a single move of a queen; the chess-knight, chess-bishop, and chess-rook graphs are defined similarly. What is the number ago for each of these four graphs? (Solutions are displayed in Berge [B12, pp. 41-42] 10.12. Some relationships among a, doo, and oo are as follows: a) tag S to. b) For some graphs, ag < ag. ©) For some graphs, #9 < %. d) For some graphs, a9 < Mo.EXERCISES 101 10.13. Prove or disprove: A line x iscritical ina graph G ifand only ifthere is a minimum line cover containing x. 10.14 Prove or disprove: Every 2-connected line-critical graph is hamiltonian. 10.15 The converse of Corollary 10.6(a) does not hold. Construct a block which is not line-critical in which any two adjacent lines lie on an odd cycle. 10.16 A tree T equals its line-core if and only if T is a block-cutpoint tree. (Harary and Plummer [HP19]) 10.17 For any graph G, the following are equivalent : 1. Ghasa line-core. 2 a9(6) = B(G), 3. 43(6) = Bo(G). (Harary and Plummer [HP19]) 10.18 If Gis a connected graph having a line-core C,(G), then a) C,(G) is a spanning subgraph of G, b) C(C(G) = CG), ©) the components of C,(6) are bipartite subgraphs of G which are not reducible. (Harary and Plummer [HP19]) 10.19 If G is a graph with line-core C,(G) and B is a bipartite subgraph of G properly containing C,(G), then B is reducible. (Harary and Plummer [HP 19]) 10.20. The point-care C,(G) is the subgraph of Ginduced by the union of all indenendent sets S of @,(G) points. A graph G has a point-core if and only if it has a line-core. (Harary and Plummer [HP 18]) 10.21 IfG = CoG), then G has a I-factor. (Harary and Plummer [HP 18]) 10.22. If G is regular of degree n, then there is a partition of V into at most 1 + [n/2] subsets such that each point is adjacent to at most one other point in the same subset. (Gerencsér [G6))CHAPTER 11 PLANARITY Return with me a while to the plains of Flatland, and I will shew you that which you have often reasoned and thought about . . Epwin A. Aport, Flatland Topological graph theory was first discovered in 1736 by Euler (V — E + F = 2)and then was dormant for 191 years. The subject was revived when Kuratowski found a criterion for a graph to be planar. Another pioneer in topological graph theory was Whitney, who developed some important propertics of the cmbedding of graphs in the plane. All the known criteria for planarity are presented. These include the theorems of Kuratowski and Wagner, which characterize planar graphs in terms of forbidden subgraphs, Whitney’s result in terms of the existence of a combinatorial dual, and MacLane’s description of the existence of a pre- scribed cycle basis. Several topological invariants of a graph are introduced. The genus of a graph has been determined for the complete graphs and bipartite graphs, the thickness for “most” of them, and the crossing number for only a few. PLANE AND PLANAR GRAPHS A graph is said to be embedded in a surface S when it is drawn on S so that no two edges intersect. As noted in Chapter 1, we shall use “points and lines” for abstract graphs, “vertices and edges” for geometric graphs (embedded in some surface). A graph is planar if it can be embedded in the plane; a bOI <_< Fig. 11.1. A planar graph and an ambessng 102PLANE AND PLANAR GRAPHS — 103 plane graph has already been embedded in the plane. For example, the cubic graph of Fig. 11.1(a) is planar since it is isomorphic to the plane graph in Fig, 11.1(b). We will refer to the regions defined by a plane graph as its faces, the unbounded region being called the exterior face. When the boundary of a face of a plane graph is a cycle, we will sometimes refer to the cycle as a face. The plane graph of Fig. 11.2 has three faces, f,, f, and the exterior face fs. Of these, only f; is bounded by a cycle. xh Fig. 11.2. A plane graph. ‘The subject of planar graphs was discovered by Euler in his investigation of polyhedra. With every polyhedron there is associated a graph consisting only of its vertices and edges, called its /-skeleton. For example, the graph Q, is the 1-skeleton of the cube and K>,»,» that of the octahedron. The Euler formula for polyhedra is one of the classical results of mathematics. ‘Theorem 11.1 (Euler Polyhedron Formula). For any spherical polyhedron with V vertices, E edges, and F faces, V-E+F=2. (11.1) For the 3-cube we have V = 8, E = 12, and F = 6 so that (11.1) holds; for a tetrahedron, V = F = 4 and E = 6. Before proving this equation, we will recast it in graph theoretic terms. A plane map is a connected plane graph together with all its faces. One can restate (11.1) for a plane map in terms of the numbers p of vertices, q of edges, and r of faces, P-qtr=2. (HL) It is easy to prove this theorem by induction. However, this equation has already been proved in Chapter 4 where it was established that the cycle rank m of a connected graph G is given by m=q-ptlh Since it is easily seen that if (11.1’) holds for the blocks of G separately, then (11.1) holds for G also, we assume from the outset that G is 2-connected. Thus every face of a plane embedding of G is a cycle. We have just noted that p = V and q = E for a plane map. It only remains to link m with F. We now show that the interior faces of a plane graph G constitute a cycle basis for G, so that they are m in number. This holds because the edges of every cycle Z of G can be regarded as the symmetric104 PLANARITY difference of the faces of G contained in Z. Since the exterior face is thus the sum (mod 2) of all the interior faces (regarded as edge sets), we see that m = F — 1, Hencem =q ~ p+ 1 becomesF -1=E-V+1. Euler's equation has many consequences. Corollary 11.1(a)_ If G is a (p, q) plane map in which every face is an n-cycle, then q = nlp — In — 2) (11.2) Proof. Since every face of G is an n-cycle, each line of G is on two faces and each face has n edges. Thus nr = 2g, which when substituted into (11.1’) gives the result. ‘A maximal planar graph is one to which no line can be added without losing planarity. Substituting n = 3 and 4 into (11.2) gives us the next result. Corollary 11.1(b) If G is a (p, g) maximal plane graph, then every face is a triangle andq = 3p — 6. IfGisa plane graph in which every face is a 4-cycle, then g = 2p — 4. Because the maximum number of edges in a plane graph occurs when each face is a triangle, we obtain a necessary condition for planarity of a graph in terms of the number of lines. Corollary 11.1) If Gis any planar(p, q) graph with p > 3, theng < 3p — 6. If G is 2-connected without triangles, then g < 2p — 4. Corollary 11.1(d) The graphs K, and K3,; are nonplanar. Proof. The (5, 10) graph Ks is nonplanar because q = 10 > 9 = 3p ~ 6; for Ks,3,q = Yand 2p — 4 = 8. As we will soon see, the graphs Ks and K;,3 play a prominent role in characterizing planarity. The above corollaries are extremely useful in investigating planar graphs, especially maximal planar graphs. Corollary 11.1(e) Every planar graph G with p > 4 has at least four points of degree not exceeding 5. Clearly, a graph is planar if and only if each of its components is planar. Whitney [W12] showed that in studying planarity, it is sufficient to consider 2-connected graphs. ‘Theorem 11.2 A graph is planar if and only if each of its blocks is planar. It is intuitively obvious that any planar graph can be embedded in the sphere, and converscly. This fact enables us to embed a planar graph in the plane in many different ways.PLANE AND PLANAR GRAPHS 105 Theorem 11.3 Every 2-connected plane graph can be embedded in the plane so that any specified face is the exterior. Proof. Let fe a nonexterior face of a plane block G. Embed G on a sphere and call some point interior to fthe “North Pole.” Consider a plane tangent to the sphere at the South Pole and project* G onto that plane from the North Pole. The result is a plane graph isomorphic to G in which fis the exterior face. ———— Corollary 11.3(a) Every planar graph can be embedded in the plane so that a prescribed line is an edge of the exterior region. Whitney also proved that every maximal planar graph is a block, and more. Theorem 11.4 Every maximal planar graph with p > 4 points is 3-connected. Nv ALY @ () Fig. 11,3. Plane wheels. There are five ways of embedding the 3-connected wheel W, in the plane: one looks like Fig. 11.3(a), and the other four look like Fig. 11.3(b). However, there is only one way of embedding W, on a sphere, an observation which holds for all 3-connected graphs (Whitney [W13)). Theorem 11.5 Fvery 3-connected planar graph is uniquely embeddable on the sphere. 11.4. Two plane embeddings of a 2-connected graph. To show the necessity of 3-connectedness, consider the isomorphic graphs G, and G, of connectivity 2 shown in Fig. 11.4. The graph G, is embedded on the sphere so that none of its regions are bounded by five edges while G, has two regions bounded by five edges. * This is usually called stereographic projection.106 PLANARITY A polyhedron is convex if the straight line segment joining any two of its points lies entirely within it. The next theorem is due to Steinitz and Rademacher (SR2]. Theorem 11.6 A graphis the 1-skeleton ofa convex 3-dimensional polyhedron if and only if it is planar and 3-connected. One of the most fascinating areas of stndy in the theary of planar graphs is the interplay between considering a graph as a combinatorial object and as a geometric figure. Very often the question arises of placing geometric constraints on a graph. For example, Wagner [W1], Fary [F1], and Stein [S15] independently showed that every planar graph can be embedded in the plane with straight edges. ‘Theorem 11.7. Every planar graph is isomorphic with a plane graph in which all edges are straight line segments. OUTERPLANAR GRAPHS A planar graph is outerplanar if it can be embedded in the plane so that all its vertices lie on the same face ; we usually choose this face to be the exterior. Figure 11.5 shows an outerplanar graph (a) and two outerplane embeddings (b) and (0). In (©) all vertices lie on the exterior face. p= @) e © Fig. 11.5. An outerplanar graph and two outerplane embeddings. In this section we develop theorems for outerplanar graphs parallel with those for planar graphs. The analogue of Theorem 11.2 is immediate. ‘Theorem 11.8 A graph G is outerplanar if and only if each of its blocks is outerplanar. An outerplanar graph G is maximal outerplanar if no line can be added without losing outerplanarity. Clearly, every maximal outerplane graph is a triangulation of a polygon, while every maximal plane graph is a tri- angulation of the sphere. The three maximal outerplane graphs with 6 vertices are shown in Fig. 11.6, Theorem 11.9 Tet G be a maximal outerplane graph with p > 3 vertices all lying on the exterior face. Then G has p — 2 interior faces.OUTERPLANAR GRAPHS 107 w © © Fig. 11.6. Three maximal outerplanar graphs. Ke Rist Fig. 11.7. The forbidden graphs for outerplanarity. Proof. Obviously the result holds for p = 3. Suppose it is true for p = n and let G have p = n + 1 vertices and m interior faces. Clearly G must have a vertex v of degree 2 on its exterior face. In forming G — v we reduce the number of interior faces by 1 so that m— 1 =n — 2. Thusm=n—1 = p — 2, the number of interior faces of G. This theorem has several consequences. Corollary 11.9(a) Every maximal outerplanar graph G with p points has a) 2p — 3 lines, ) at ieast three points of degree not exceeding 3, c) at least two points of degree 2, a) x(G) = 2. All plane embeddings of K, and K 2,, are of the forms shown in Fig. 11.7, in-which each hasa vertex inside the exterior cycle. Therefore, neither of these graphs is outerplanar. We now observe that these are the two basic non- outerplanar graphs, following [CH3] Two graphs are homeomorphic if both can be obtained from the same graph by a sequence of subdivisions of lines. For example, any two cycles are homeomorphic, and Fig. 11.8 shows a homeomorph of K4. ‘Theorem 11.10 A graph is outerplanar if and only if it has no subgraph homeomorphic to Ky or Kz, except Ky — x. It is often important to investigate the complement of a graph with a given property. For planar graphs, the following theorem due to Battle,108 PLANARITY Fig. 11.8. A homeomorph of K,. Harary, and Kodama [BHK1] and proved less clumsily by Tutte [T16], provides a sufficient condition for the complement of a planar graph to be planar. Theorem 11.11 Every planar graph with at least nine points has a nonplanar complement, and nine is the smallest such number. This result was proved by exhaustion; no elegant or even reasonable proof is known. The analogous observation for outerplanar graphs was made in [G5]. Theorem 11.12 Every outerplanar graph with at least seven points has a WYVOe@ Fig. 11.9. The four maximal outerplanar graphs with seven points. Proof. To prove the first part, it is suflicient to verify that the complement of every maximal outerplanar graph with seven points is not outerplanar. This holds because there are exactly four maximal outerplanar graphs with p = 7 (Fig. 11.9) and the complement of each is readily seen to be non- outerplanar. The minimality follows from the fact that the (maximal) outerplanar graph of Fig. 11.6() with six points has an outerplanar complement. KURATOWSKI’S THEOREM Until Kuratowski's paper appeared [K14], it was a tantalizing unsolved problem to characterize planar graphs. The following proof of his theorem is based on that by Dirac and Schuster [DS!].KURATOWSKI'S THEOREM 109 ‘Theorem 11.13 A graph is planar if and only if it has no subgraph homeo- morphic to Ks or K3,3- Proof. Since Ks and K;, are nonplanar by Corollary 11.1(d), it follows that if a graph contains a subgraph homeomorphic to either of these, it is also nonplanar. The proof of the converse is a bit more involved. Assume itis false. Then there is a nonplanar graph with no subgraph homeomorphic to either Ks or K;,3. Let G be any such graph having the minimum number oflines. Then G must be a block with 6(G) > 3. Let xo = uovo be an arbitrary line of G. The graph F = G — xq is necessarily planar. We will find it convenient to use two lemmas in the development of the proof. Lemma 11.13(a) There is a cycle in F containing ug and vo. Proof of Lemma. Assume that there is no cycle in F containing up and 9. Then uo and vo lie in different blocks of F by Theorem 3.3. Hence, there exists a cutpoint w of F lying on every uo-vo path. We form the graph Fo by adding to F the lines wug and wv if they are not already present in F. In the graph Fo, uo and vg still lie in different blocks, say B, and B,, which necessarily have the point w in common. Certainly, each of B, and B, has fewer lines than G, so either B, is planar or it contains a subgraph homeo- morphic to K, or K3,3. If, however, the insertion of wg produces a subgraph H of B, homeomorphic to K or K3,3, then the subgraph of G obtained by replacing wu, by a path from up to w which begins with xo is necessarily homeomorphic to H and so to Ks or K;3,3, but this is a contradiction. Hence, B, and similarly B, is planar. According to Corollary 11.3(a), both B, and B, can be drawn in the plane so that the lines wu, and wv, bound the exterior region. Hence it is possible to embed the graph F, in the plane with both WUy and wv on the exterior region. Inserting x» cannot then destroy the planarity of Fo. Since G is a subgraph of Fo + Xo, G is planar; this contra- diction shows that there is a cycle in F containing uy and vp. Let F be embedded in the plane in such a way that a cycle Z containing Uo and vp has a maximum number of regions interior to it. Orient the edges of Z in a cyclic fashion, and let Z[u, v] denote the oriented path from u to v along Z. Ifv does not immediately follow u on Z, we also write Z(u, v) to indicate the subpath of Z[u, v] obtained by removing u and v. By the exterior of cycle Z, we mean the subgraph of F induced by the vertices lying outside Z, and the components of this suhgraph are called the exterior components of Z. By an outer piece of Z, we mean a subgraph of F induced by all edges incident with at least one vertex in some exterior component or by an edge (if any) exterior to Z meeting two vertices of Z. Ina like manner, we define the interior of cycle Z, interior component, and inner piece,110 PLANARITY Fig. 11.10. Separating cycle Z illustrating lemma. An outer or inner piece is called u—» separating if it meets both Z(u, ») and Z(v, u). Clearly, an outer or inner piece cannot be w-v separating ifu and vare adjacent on Z. Since F is connected, each outer piece must meet Z, and because F has no cutvertices, each outer piece must have at least two vertices in common with Z. No outer piece can meet Z(uo, v9) or Z(vo, vo) in more than one vertex, for otherwise there would exist a cycle containing ug and v9 with more interior regions than Z. For the same reason, no outer piece can meet uy or vo. Hence, every outer piece meets Z in exactly two vertices and is Up—vo separating. Furthermore, since x, cannot be added to F in planar fashion, there is at least one up-vo separating inner piece. Lemma 11.13(b) There exists a uv separating outer piece meeting Z(uo, Uo), say at u,, and Z(vo, uo), say at v,, such that there is an inner piece which is both u,~v, separating and u,—v, separating. Proof of Lemma. Suppose, to the contrary, that the lemma does not hold. It will be helpful in understanding this proof to refer to Fig. 11.10. We order the uo-v, separating inner pieces for the purpose of relocating them in the plane. Consider any uo-vo separating inner piece I, which isKURATOWSKI’S THEOREM IL Fig. 11.11. The possibilities for nonplanar subgraphs. nearest to ug in the sense of encountering points of this inner piece on moving along Z from up. Continuing out from up, we can index the up—vo separating inner pieces I, 15, and so on. Let uz and u, be the first and last points of I, meeting Z(uo, vo) and vy and v be the first and last vertices of I, meeting Z(vo, uo). Every outer piece necessarily has both its common vertices with Z on either Z[v3, u,] or Z[us, v2], for otherwise there would exist an outer piece meeting Z(uo, vo) atu, and Z(vp, uo) at v, and an inner piece which is both ug-vo separating and u,v, separating, contrary to the supposition that the lemma is false. Therefore, a curve C joining v; and u can be drawn in the exterior region so that it meets no edge of F. (See Fig. 11.10.) Thus, I, can be transferred outside of C in a planar manner. Similarly, the remaining u,—vy separating inner pieces can be transferred outside of Z, in order, so that the resulting graph is plane. However, the edge x, can then be added without destroying the planarity of F, but this is a contradiction, completing the lemma. Proof of Theorem. Let H be the inner piece guaranteed by Lemma 11.13(b) which is both uo-v, separating and u,—v, separating. In addition, let Wo, Wo, W,, and w be vertices at which H meets Z(ug, U9) Zo» Uo Zl, 0), and Z(v,, u,) respectively. There are now four cases to consider, depending on the relative position on Z of these four vertices. CASE 1. One of the vertices w, and w is on Z(uo, Vg) and the other is on Zr, uo). We can then take, say, Wo = w; and wy = wi, in which case G112 PLANARITY contains a subgraph homeomorphic to K 3,3, as indicated in Fig. 11.11(a), in which the two sets of vertices are indicated by open and closed dots. CASE 2. Both vertices w, and w, are on either Z(uy, vg) Or Z(vy uy). Without loss of generality we assume the first situation. There are two possibilities : either v, # wo or v, = w>. Ifo, # ws, then G contains a subgraph homeo- morphic to K3,3, as shown in Fig. 11.11(b) or (c), depending on whether w’y lies on Z(u,, v;) or Z(v,, u,), respectively. If v, = wy (see Fig. 11.114), then H contains a vertex r from which there exist disjoint paths to w,, w', and v,, all of whose vertices (except w,, w;, and p,) belong to H. In this case also, G contains a subgraph homeomorphic to K,5- CASE 3. W; = 09 and Ww; # uo. Without loss of generality, let w, be on Zuo, 9). Once again G contains a subgraph homeomorphic to K3,3. If ¥ ig. = the other hand, w is on Z(0,, up), there is a K,,, as indicated in Fig, 11.11(0. This figure is easily modified to show G contains K;,, if wy = 01. CASE 4, w, = vy and w', = uy. Here we assume wy = u, and wy — v,, for otherwise we are in a situation covered by one of the first 3 cases. We distinguish between two subcases. Let Py be a shortest path in H from uo to vo, and let P, be such a path from u, to v,. The paths P, and P, must intersect. If Py and P, have more than one vertex in common, then G contains a subgraph homeomorphic to K3,3, as shown in Fig. 11.11(g); otherwise, G contains a subgraph homeomorphic to K, as in Fig. 11.11(h). Since these are all the possible cases, the theorem has been proved. In his paper “How to draw a graph,” Tutte [T17] gives an algorithm for drawing in the plane as much of a given graph as possible and shows that whenever this process stops short of the entire graph, it must contain a subgraph homeomorphic to Ks or K3,3- Thus his algorithm furnishes an independent proof of Theorem 11.13. An elementary contraction of a graph G is obtained by identifying two adjacent points u and », that is, by the removal of u and v and the addition Fig. 11.12, Nonplanarity of the Petersen graph,OTHER CHARACTERIZATIONS OF PLANAR GRAPHS — 113 of a new point w adjacent to those points to which u or v was adjacent. A graph G is contractible to a graph H if H can be obtained from G by a sequence of elementary contractions. For example, as indicated in Fig. 11.12(a) and (b), the Petersen graph is contractible to Ks by contracting each of the five lines u,v; joining the pentagon with the pentagram to a new point w,, A dual form of Kuratowski’s theorem (in the sense of duality in matroid theory) was found independently by Wagner [W2] and Harary and Tutte [HT3]. Theorem 11.14 A graph is planar if and only if it does not have a subgraph contractible to K, or K3,3. We have just seen that the Petersen graph is contractible to K,. Since every point has degree 3, it clearly does not have a subgraph homeomorphic to Ks; Fig. 11.12(c) shows one homeomorphic to K3,3. OTHER CHARACTERIZATIONS OF PLANAR GRAPHS Several other criteria for planarity have been discovered since the original work of Kuratowski. We have already noted the “dual form” in terms of contraction in Theorem 11.14. Tutte’s algorithm for drawing a graph in the plane may also he regarded as a characterization Whitney ‘Tw, W14] expressed planarity in terms of the existence of dual graphs. Given a plane graph G, its geometric dual G* is constructed as follows: place a vertex in each region of G (including the exterior region) and, if two regions have an edge x in common, join the corresponding vertices by an edge x* crossing only x. The result is always a plane pseudo- graph, as indicated in Fig. 11.13 where G has solid edges and its dual G* dashed edges. Clearly G* has loop ifand only if G has an endvertex, and G* has multiple edges if and only if two regions of G have at least two edges in common. Thus, a 2-connected plane graph always has a graph or multi- graph as its dual, while the dual of a 3-connected graph is always a graph. Other examples of geometric duals are given by the Platonic graphs: the tetrahedron is self-dual, whereas the cube and octahedron are duals, as are the dodecahedron and the icosahedron. Fig. 11.13. A plane graph and its geometric dual.114 PLANARITY Fig. 11.14, Different geometric duals of the same abstract graph. As defined, the geometric dual of a plane graph G is also plane, and it follows that the dual of the dual of G is the original graph G. However, an abstract graph with more than one embedding on the sphere can give rise to more than one dual graph. This is illustrated in Fig. 11.14 in which graphs G and H are abstractly isomorphic, but as embedded they have different duals G* and H*. However, since a triply connected graph has only one spherical embedding, as noted in Theorem 11.5, it must have a unique geometric dual. Whitney gave a combinatorial definition of dual, which is an abstract formulation of the geometric dual. To state this, we recall from Chapter 4 that for a graph G with k components, the cycle rank is given by m(G) = q — p + kand the cocycle rank by m*(G) = p — k. The relative complement G — H of a subgraph H of G is defined to be that subgraph obtained by deleting the lines of H. A graph G* is a combina- torial dual of graph G if there is a one-to-one correspondence between their sets of lines such that for any choice Y and ¥* of corresponding subsets of lines, mG — Y) = *(G) — mCY*)), (11.3) where {¥*) is the subgraph of G* with line set Y*. This definition is illus- trated by Fig. 11.15 where the correspondence is x,++) Here ¥ — {%2, Xs Xa Xe)» SO that m*(G — Y) = 1, m*G) = 4, and m(¥*) = 3, soOTHER CHARACTERIZATIONS OF PLANAR GRAPHS 115 Ye Fig. 11.15. Combinatorial duals. the defining equation is satisfied. It is of course very difficult to check whether two graphs are duals using (11.3) since it involves verifying this equation for every set y of lines in G. As with geometric duals, combinatorial duals of planar graphs are not necessarily unique. However, if two graphs are combinatorial duals of isomorphic graphs, there is a one-to-one correspondence between their sets of lines which preserves cycles as sets of lines (that is, their cycle matroids are isomorphic). The correspondence x; + y, of G* and H* in Fig. 11.14 illustrates this. Whitney proved that combinatorial duals are equivalent to geometric duals, giving another criterion for planarity. Theorem 11.15 A graph is planar if and only if it has a combinatorial dual. Another criterion for planarity due to MacLane [M1] is expressed in terms of cyclic structure. Theorem 11.16 A graph G is planar if and only if every block of G with at least three points has a cycle basis Z;, Zz,*+*,Z, and one additional cycle Zy such that every line occurs in exactly two of these m | 1 cycles. We only indicate the necessity, which is much easier. As mentioned in the proof of Theorem 11.1, all the interior faces of a 2-connected plane graph G constitute a cycle basis Z,, Zz, ***, Zp Where m is the cycle rank of G. Let Z, be the exterior cycle of G. Then obviously each edge of G lies on exactly two of the m + 1 cycles Z,. To prove the sufficiency, it is necessary to construct a plane embedding of a given graph G with the stipulated properties.116 PLANARITY All of these criteria for planarity are summarized in the following list of equivalent conditions for a graph G. (1) Gis planar. (2) G has no subgraph homeomorphic to K's or Ks, (3) G has no subgraph contractible to Ks or K3,. (4) G has a combinatorial dual. (5) Every nontrivial block of G has a cycle basis Z,, Zz,***, Zp and one additional cycle Zp such that every line x occurs in exactly two of these m+ Lvydles. GENUS, THICKNESS, COARSENESS, CROSSING NUMBER In this section four topological invariants of a graph G are considered. These are genus: the number of handles needed on a sphere in order to embed G, thickness: the number of planar graphs required to form G, coarseness: the maximum number of linc-disjoint nonplanar subgraphs in G, and crossing number: the number of crossings there must be when G is drawn in the plane. We will concentrate on three classes of graphs—com- plete graphs, complete bigraphs, and cubes—and indicate the values of these invariants for them as far as they are known. ARN Fig. 11.16. Embedding a graph on an orientable surface. As observed by KSnig, every graph is embeddible on some orientable surface. This can easily be seen by drawing an arbitrary graph G in the plane, possibly with edges that cross each other, and then attaching a handle to the plane at each crossing and allowing one edge to go over the handle and the other under it. For example, Fig. 11.16 shows an embedding of K in a plane to which one handle has been attached. Of course, this method often uses more handles than are actually required. In fact, Kénig also showed that any embedding of a graph on an orientable surface with a minimum number of handles has all its faces simply connected.GENUS, THICKNESS, COARSENESS, CROSSING NUMBER 117 a a PAY ING 1 I ! I 1 | ! 1 1 1 1 I t 1 | 1 L Fig. IL17. An embedding of Ky on the Fig. 11.18. A toroidal embedding of torus Kaw We have already noted that planar graphs can be embedded on a sphere.. A toroidal graph can be embedded on a torus. Both K and K, ,are toroidal ; in fact Figs. 11.17 and 11.18 show embeddings of K, and K,,, on the torus, represented as the familiar rectangle in which both pairs of opposite sides are identified. No characterization of toroidal graphs analogous to Kura- towski’s Theorem has been found. However, Vollmerhaus [V6] settled a conjecture of Erdés in the affirmative by proving that for the torus as well as any other orientable surface, there is a finite collection of forbidden subgraphs. ‘The genus )(G) of a graph G is the minimum number of handies which must be added to a sphere so that G can be embedded on the resulting surface. Ofcourse, ;(G) = 0 ifand only if G is planar, and homeomorphic graphs have the same genus. The first theorem of this chapter presented the Euler characteristic equation, V — E + F = 2, for spherical polyhedra. More generally, the genus of a polyhedron* is the number of handles needed on the sphere for a surface to contain the polyhedron, Theorem 11.1 has been generalized to polyhedra of arbitrary genus, in a result also due to Euler. A proof may be found in Courant and Robbins [CR1]. Theorem 11.17 For a polyhedron of genus 7 with V vertices, E edges and F faces, V-E+F=2-2. (11.4) This equation is particularly useful in proving the easy half of the results to follow on the genus and thickness of particular graphs. Its corollaries, which offer no difficulty, are often more convenient for this purpose. * For a combinatorial treatment of the theory of polyhedra, see Griinbaum[G10].8 PLANARITY Corollary 11.17(a) If G is a connected graph of genus y in which every face isa triangle, then q = Xp — 2 + 2); (11.5) when every face is a quadrilateral, q = Up —2 + 2y). (11.6) ‘As mentioned in [BH2], it is easily verified from these two equations that the genus of a graph has the following lower bounds. Corollary 11.17(b)_ If Gis a connected graph of genus », then. y= eq — Hp - 2); (11.7) if G has no triangles, then 7249-4 - 2. (11.8) ‘The determination of the genus of the complete graphs has been a long, interesting, difficult, successful struggle. In its dual form, it was known as the Heawood Conjecture and stood unproved from 1890 to 1967. We return to this aspect of the problem in the next chapter. There have been many contributors to this result and the coup de grace, settling the conjecture in full, was administered by Ringel and Youngs [RY1]. Theorem 11.18 For every positive integer p, the genus of the complete graph is a W(K,) {oe 3h (11.9) The proof of the easier half of equation (11.9) is due to Heawood [H38]. It amounts to substituting q(K,) into inequality (11.7) to obtain x Ir) _ 16 -@=0- 49 AK) = (b) 30 -2)= 2. Then since the genus of every graph is an integer, = 3Xp - 4 AK) > {ee}, ‘The proof that this expression is also an upper bound for 7(K,) can only be accomplished by displaying an embedding of K, into an orientable surface of the indicated genus. When Heawood originally stated the conjecture 990, d that (K.) =< $90, he proved that 7{K7} Fig. 11.17, which triangulates the torus. Heffter proved (11.9) in 1891 for p = 8 through 12. Not until 1952 did Ringel prove it for p = 13. At that stage, it was realized that because of its form, it was natural to try to settle the question for one residue class of p