Algorithm Search⌘K

Proofs: Kuratowski's theorem

These are the human proofs of Kuratowski's theorem, at full length.

Proof of Kuratowski's theorem full

This proof shows: statement.

We prove that a finite graph GG is planar if and only if it contains no subdivision of K5K_5 or K3,3K_{3,3}. Graphs are simple: loops and parallel edges do not change planarity. We write THTH for a subdivision of HH, and say that GG contains THTH if some subgraph of GG is a THTH. A model of HH in GG is a family of disjoint vertex sets VhV_h (h∈V(H)h \in V(H)), the branch sets, each inducing a connected subgraph, with an edge of GG between VhV_h and Vh′V_{h'} for each edge hh′hh' of HH; GG has an HH minor exactly when it has a model of HH. The proof follows the route of Diestel, Sections 4.4 and 3.2, with the minimal counterexample of Thomassen.

The easy direction

A subgraph of a planar graph is planar: delete points and arcs from a drawing. If GG contains TK5TK_5 or TK3,3TK_{3,3} and GG were planar, a drawing of the subdivision would give a drawing of K5K_5 or K3,3K_{3,3}, by reading each subdivided path as one arc. But K5K_5 has 10>3⋅5−610 > 3 \cdot 5 - 6 edges, and K3,3K_{3,3} is triangle-free with 9>2⋅6−49 > 2 \cdot 6 - 4 edges, so neither is planar by the edge bounds from Euler's formula. So a planar graph contains neither subdivision.

From minors to subdivisions

Lemma 1 (a center in a tree). Let TT be a finite tree and let kk labels be placed on vertices of TT (several labels may sit on one vertex). If no edge ee of TT leaves at least 2 labels on each side of T−eT - e, then there is a vertex xx such that each component of T−xT - x carries at most one label.

Proof. Each edge ee has a side with at most one label. Orient ee away from such a side (if both sides have at most one label, orient it either way). Start anywhere and follow the oriented edges. A walk along oriented edges in a tree cannot use an edge twice (it would have to go back against its orientation) and cannot close a cycle, so it never repeats a vertex and stops at a vertex xx with no outgoing edge. For each edge xyxy, the edge points to xx, so the side of yy, which is the component of T−xT - x that contains yy, has at most one label. □\square

Call such an xx a center. For a center xx and each label ℓ\ell on a vertex aℓa_\ell, the paths xTaℓx T a_\ell meet pairwise only in xx: two of them lie in different components of T−xT - x, or one is the trivial path xx.

Lemma 2. (a) If GG has a K3,3K_{3,3} minor, then GG contains TK3,3TK_{3,3}. (b) If GG has a K5K_5 minor, then GG contains TK5TK_5 or TK3,3TK_{3,3}.

Proof. Let (Vh)(V_h) be a model of H∈{K5,K3,3}H \in \{K_5, K_{3,3}\} in GG. For each edge hh′hh' of HH fix an edge ahh′ah′ha_{hh'} a_{h'h} of GG with ahh′∈Vha_{hh'} \in V_h and ah′h∈Vh′a_{h'h} \in V_{h'}. In VhV_h, put the label h′h' on ahh′a_{hh'} for each neighbour h′h' of hh, and let ThT_h be a spanning tree of G[Vh]G[V_h].

Suppose first that every ThT_h has a center xhx_h. For each edge hh′hh' of HH, let Phh′P_{hh'} be the path xhThahh′x_h T_h a_{hh'}, the edge ahh′ah′ha_{hh'} a_{h'h}, and ah′hTh′xh′a_{h'h} T_{h'} x_{h'}. Two such paths meet only in common ends: inside one tree ThT_h, the parts toward different labels meet only in xhx_h by the remark after Lemma 1, and different trees are disjoint. So the paths Phh′P_{hh'} with the vertices xhx_h form a THTH in GG.

For H=K3,3H = K_{3,3} each vertex has 3 labels, and no edge of ThT_h can leave 2 labels on each side, so Lemma 1 gives a center in every tree. This proves (a).

For H=K5H = K_5, if every ThT_h has a center we get TK5TK_5. Otherwise, by Lemma 1, some tree, say T1T_1 with labels 2,3,4,52, 3, 4, 5, has an edge u1u2u_1 u_2 such that the component U1U_1 of T1−u1u2T_1 - u_1 u_2 that contains u1u_1 carries two labels, say 2 and 3, and the component U2U_2 that contains u2u_2 carries 4 and 5. Then the sets V(U1),V4,V5V(U_1), V_4, V_5 and V(U2),V2,V3V(U_2), V_2, V_3 are the branch sets of a K3,3K_{3,3} model with these two sides: U1U_1 and U2U_2 are connected and joined by the edge u1u2u_1 u_2; U1U_1 contains a12a_{12} and a13a_{13}, so it is joined to V2V_2 and V3V_3; U2U_2 is joined to V4V_4 and V5V_5 in the same way; and each of V4,V5V_4, V_5 is joined to each of V2,V3V_2, V_3 because the VhV_h form a K5K_5 model. So GG has a K3,3K_{3,3} minor, and contains TK3,3TK_{3,3} by (a). □\square

Since a THTH in GG gives an HH minor (contract the subdivided paths), Lemma 2 shows: GG contains TK5TK_5 or TK3,3TK_{3,3} if and only if GG has a K5K_5 or K3,3K_{3,3} minor. Minors of minors are minors, so if GG contains neither subdivision, neither does any minor of GG.

A contractible edge

A graph is 3-connected if it has more than 3 vertices and no set of at most 2 vertices separates it. In a 3-connected graph GG, if SS is a set of 3 vertices that separates GG, then each vertex s∈Ss \in S has a neighbour in every component of G−SG - S: otherwise the other two vertices of SS would separate that component from the rest.

Lemma 3 (Thomassen). Every 3-connected graph GG other than K4K_4 has an edge xyxy such that G/xyG / xy is 3-connected.

Proof. The only 3-connected graph with 4 vertices is K4K_4, so GG has n≥5n \ge 5 vertices and each G/xyG / xy has at least 4. Suppose that no contraction is 3-connected. Let xyxy be an edge and vxyv_{xy} the vertex of G/xyG/xy that replaces xx and yy. Then G/xyG / xy has a separating set SS of at most 2 vertices. If vxy∉Sv_{xy} \notin S, then SS separates GG too, which is impossible. So SS contains vxyv_{xy}, and S={vxy,z}S = \{v_{xy}, z\} for a vertex zz, since {vxy}\{v_{xy}\} alone would give the separating set {x,y}\{x, y\} of GG. Then {x,y,z}\{x, y, z\} separates GG.

Among all edges xyxy and such vertices zz, choose them so that G−{x,y,z}G - \{x, y, z\} has a component CC with as few vertices as possible. The vertex zz has a neighbour vv in CC. By assumption G/zvG / zv is not 3-connected either, so as above there is a vertex ww such that {z,v,w}\{z, v, w\} separates GG, and each of z,v,wz, v, w has a neighbour in every component of G−{z,v,w}G - \{z, v, w\}. Since xx and yy are adjacent, they lie in at most one component of G−{z,v,w}G - \{z, v, w\} together with possibly ww; so there is a component DD that contains neither xx nor yy. The vertex vv has a neighbour d∈Dd \in D. Since d∉{x,y,z}d \notin \{x, y, z\} and dd is adjacent to v∈Cv \in C, dd lies in CC. The subgraph DD is connected and avoids x,y,zx, y, z, so D⊆CD \subseteq C. And v∈C∖Dv \in C \setminus D, so ∣D∣<∣C∣|D| < |C|, which contradicts the choice of CC, with the edge zvzv and the vertex ww. □\square

Two facts about drawings

We use two standard facts about plane graphs.

(F1) In a 2-connected plane graph, every face is bounded by a cycle of the graph (Diestel, Proposition 4.2.6). The region bounded by a cycle is homeomorphic to an open disk with the cycle as its boundary circle, by the Jordan-Schönflies theorem (Mohar and Thomassen, Section 2.1). So a new vertex placed in a face bounded by a cycle CC can be joined to any set of vertices of CC by arcs inside the face that meet only at the new vertex.

(F2) For each face ff of a plane graph, there is a drawing of the same graph with the same faces in which ff is the outer face (project the drawing onto the sphere and back from a point inside ff). So for each edge ee there is a drawing with ee on the boundary of the outer face. If G=G1∪G2G = G_1 \cup G_2 with V(G1)∩V(G2)⊆{x,y}V(G_1) \cap V(G_2) \subseteq \{x, y\}, and if G1+xyG_1 + xy and G2+xyG_2 + xy are planar (or G1G_1, G2G_2 are planar and share at most one vertex xx), then GG is planar: draw both parts with xyxy (or xx) on the outer face, put the drawing of the second part inside a face of the first next to xyxy (or at xx), and delete xyxy if it is not an edge of GG (Diestel, Section 4.4).

The main proof

Suppose the theorem is false, and let GG be a non-planar graph that contains no TK5TK_5 and no TK3,3TK_{3,3}, with ∣V(G)∣+∣E(G)∣|V(G)| + |E(G)| as small as possible. Then GG has no isolated vertex, and every graph with a smaller value of ∣V∣+∣E∣|V| + |E| and no TK5TK_5, TK3,3TK_{3,3} is planar. Graphs with at most 4 vertices are planar (they are subgraphs of K4K_4), so ∣V(G)∣≥5|V(G)| \ge 5.

GG is 2-connected. Otherwise G=G1∪G2G = G_1 \cup G_2 for proper subgraphs G1,G2G_1, G_2 that share at most one vertex, with no edge between G1−G2G_1 - G_2 and G2−G1G_2 - G_1, each with a vertex outside the other. Each GiG_i has a smaller value of ∣V∣+∣E∣|V| + |E| and no TK5TK_5 or TK3,3TK_{3,3}, so it is planar, and GG is planar by (F2). This is a contradiction.

GG is 3-connected. Suppose that {x,y}\{x, y\} separates GG. Write G=G1∪G2G = G_1 \cup G_2 with V(G1)∩V(G2)={x,y}V(G_1) \cap V(G_2) = \{x, y\}, each GiG_i with a vertex outside {x,y}\{x, y\}, and no edge between G1−{x,y}G_1 - \{x, y\} and G2−{x,y}G_2 - \{x, y\}; put the edge xyxy, if it exists, into G1G_1. Let Hi=Gi+xyH_i = G_i + xy. Each component of G−{x,y}G - \{x, y\} inside G3−iG_{3-i} has a neighbour at xx and one at yy, since GG is 2-connected; so G3−iG_{3-i} contains an xx-yy path P3−iP_{3-i} whose inner vertices lie outside GiG_i. The graph HiH_i has fewer vertices than GG and at most one edge that GG lacks, while G3−iG_{3-i} has at least two edges not in GiG_i; so ∣V(Hi)∣+∣E(Hi)∣<∣V(G)∣+∣E(G)∣|V(H_i)| + |E(H_i)| < |V(G)| + |E(G)|. If HiH_i contained a TK5TK_5 or TK3,3TK_{3,3}, then replacing the edge xyxy in it, if it is used, by P3−iP_{3-i} would give one in GG. So H1H_1 and H2H_2 are planar, and GG is planar by (F2), a contradiction. So GG is 3-connected, and G≠K4G \ne K_4 since ∣V(G)∣≥5|V(G)| \ge 5.

The contraction. By Lemma 3, GG has an edge xyxy with G′=G/xyG' = G / xy 3-connected. Let vv be the vertex of G′G' that replaces xx and yy. Since GG contains no TK5TK_5 or TK3,3TK_{3,3}, it has no K5K_5 or K3,3K_{3,3} minor (by Lemma 2), so neither has its minor G′G', so G′G' contains no TK5TK_5 or TK3,3TK_{3,3}. It is smaller than GG, so it is planar. Fix a drawing of G′G'.

The graph G′−vG' - v is 2-connected, so by (F1) the face ff of G′−vG' - v that contains the point of vv is bounded by a cycle CC. The edges at vv run inside ff, so all neighbours of vv lie on CC. Let X=NG(x)∖{y}X = N_G(x) \setminus \{y\} and Y=NG(y)∖{x}Y = N_G(y) \setminus \{x\}. Both are subsets of NG′(v)⊆V(C)N_{G'}(v) \subseteq V(C), and ∣X∣≥2|X| \ge 2, ∣Y∣≥2|Y| \ge 2 since GG is 3-connected. Let x1,…,xkx_1, \ldots, x_k be the vertices of XX in their cyclic order on CC, and let PiP_i be the path on CC from xix_i to xi+1x_{i+1} that contains no other vertex of XX (indices mod kk).

Case 1: Y⊆V(Pi)Y \subseteq V(P_i) for some ii. Erase the edges at vv that do not go to XX, and rename vv as xx; the arcs from xx to XX divide ff into regions, and the region RR between the arcs to xix_i and xi+1x_{i+1} is bounded by the cycle x xi Pi xi+1 xx\, x_i\, P_i\, x_{i+1}\, x. Put yy inside RR and join it to xx and to each vertex of YY by arcs in RR, which is possible by (F1). The result is a drawing of GG, a contradiction.

Case 2: ∣X∩Y∣≥3|X \cap Y| \ge 3. Let a,b,c∈X∩Ya, b, c \in X \cap Y. Split CC into three disjoint paths, each containing exactly one of a,b,ca, b, c; they are connected and pairwise joined by edges of CC. With {x}\{x\} and {y}\{y\}, these are five branch sets of a K5K_5 model in GG: xx and yy are adjacent, and both are adjacent to a,b,ca, b, c. So GG has a K5K_5 minor and contains TK5TK_5 or TK3,3TK_{3,3} by Lemma 2, a contradiction.

Case 3: otherwise. We find y1,y2∈Yy_1, y_2 \in Y and xi,xj∈Xx_i, x_j \in X, four different vertices that alternate around CC. If some y1∈Yy_1 \in Y is not in XX, it lies inside a unique path PiP_i, between xix_i and xi+1x_{i+1}. Since Case 1 fails, some y2∈Yy_2 \in Y lies outside V(Pi)V(P_i), and then y1,xi+1,y2,xiy_1, x_{i+1}, y_2, x_i alternate around CC (also when k=2k = 2). If instead Y⊆XY \subseteq X, then ∣Y∣≤2|Y| \le 2 because Case 2 fails, so Y={xa,xb}Y = \{x_a, x_b\} with ∣Y∣=2|Y| = 2; since Case 1 fails, xax_a and xbx_b are not consecutive in the cyclic order of XX, so each of the two arcs of CC between them contains a vertex of XX, say xcx_c and xdx_d, and xa,xc,xb,xdx_a, x_c, x_b, x_d alternate. In both cases the four arcs of CC between consecutive vertices of y1,xi,y2,xjy_1, x_i, y_2, x_j (in this cyclic order), the edges xxix x_i, xxjx x_j, yy1y y_1, yy2y y_2 and xyxy form a TK3,3TK_{3,3} with sides {x,y1,y2}\{x, y_1, y_2\} and {y,xi,xj}\{y, x_i, x_j\}: xx reaches yy, xix_i, xjx_j by edges, and each of y1y_1, y2y_2 reaches yy by an edge and xix_i, xjx_j along CC. This is a contradiction.

All cases are contradictory, so no counterexample exists. ■\blacksquare

Back to Kuratowski's theorem