1: \documentclass[amstex, amysymb, 12pt]{article}
2: \setlength{\oddsidemargin}{0.25in}
3: \setlength{\topmargin}{0in}
4: \setlength{\textheight}{8.5in}
5: \setlength{\textwidth}{6.125in}
6: \newtheorem{theorem}{Theorem}
7: \usepackage{amsmath, amsfonts, amssymb, amsthm}
8: %\usepackage[notref,notcite]{showkeys}
9: \usepackage{graphicx}
10:
11: \newtheorem{Pa}{Paper}[section]
12: \newtheorem{Tm}[Pa]{{\bf Theorem}}
13: \newtheorem{La}[Pa]{{\bf Lemma}}
14: \newtheorem{Cy}[Pa]{{\bf Corollary}}
15: \newtheorem{Rk}[Pa]{{\bf Remark}}
16: \newtheorem{Pn}[Pa]{{\bf Proposition}}
17: \newtheorem{Pb}[Pa]{{\bf Problem}}
18: \newtheorem{Dn}[Pa]{{\bf Definition}}
19: \newtheorem{Ex}[Pa]{{\bf Example}}
20:
21: \newcommand{\bn}{{\bf n}}
22: \newcommand{\bk}{{\bf k}}
23: \newcommand{\C}{{\mathbb C}}
24: \newcommand{\B}{{\mathbb B}}
25: \newcommand{\N}{{\mathbb N}}
26: \newcommand{\bM}{{\bf M}}
27: \newcommand{\cH}{{\mathcal H}}
28: \newcommand{\nline}{{\vspace{\baselineskip}}}
29:
30: \renewcommand{\theequation}{\thesection.\arabic{equation}}
31: % defining a command for cases \Cs
32: \newcounter{caser}[section]
33: \setcounter{caser}{0}
34: \newcommand{\Cs}
35: {\addtocounter{caser}{1}\vspace{5mm}\par\noindent \underline{{\bf Case} \arabic{caser}}:\space}
36: % end of \Cs def
37: \begin{document}
38:
39: \title{The Graphs for which the Maximum Multiplicity of an Eigenvalue
40: is Two}
41:
42: \author{
43: Charles R. Johnson\footnote{Corresponding author;
44: Department of Mathematics, College of William and Mary, Williamsburg, VA 23187-8795, USA;
45: email: crjohnso@math.wm.edu; phone (office): 1-757-221-2014; fax (identify recipient): 1-757-221-7400},
46: \and Raphael Loewy\footnote{Department of Mathematics, Technion -- Israel Institute of Technology, Haifa 32000, Israel}, \and Paul Anthony Smith\footnote{UCLA Mathematics Department, Los Angeles, CA 90095-1555, USA}}
47:
48: \maketitle
49:
50: \pagebreak
51:
52: \begin{abstract}
53: Characterized are all simple undirected graphs $G$ such that any
54: real symmetric matrix that has graph $G$ has no eigenvalues of
55: multiplicity more than 2. All such graphs are partial 2-trees
56: (and this follows from a result for rather general fields), but only certain
57: partial 2-trees guarantee maximum multiplicity 2. Among partial
58: linear 2-trees, they are only those whose vertices can be covered
59: by two "parallel" induced paths. The remaining graphs that
60: guarantee maximum multiplicity 2 are comprised by certain
61: identified families of "exceptional" partial 2-trees that are not
62: linear.
63: \end{abstract}
64:
65: \noindent{{\it AMS classification:}} 05C50; 15A57
66:
67: \nline
68:
69: \noindent{{\it Keywords:}} graph; partial 2-tree; linear partial 2-tree; exceptional partial 2-tree; eigenvalue; minimum rank of a graph
70:
71: \pagebreak
72:
73: \section{Introduction}
74: \setcounter{equation}{0}
75:
76: Throughout, $G$ denotes a simple, undirected graph on $n$ vertices
77: without loops. Associate with $G$ the set $S(G)$ of all $n$-by-$n$,
78: real symmetric matrices $A$ whose graph is $G$. No restriction
79: (other than reality) is placed upon the diagonal entries of $A$ by
80: $G$. For each $A \in S(G)$, let $M(A)$ be the largest multiplicity
81: for an eigenvalue of $A$ and let rank($A$) denote the rank of $A$.
82: Then, over $A \in S(G)$,
83: $$M(G) = \text{max }M(A)$$
84: and
85: $${\rm msr}(G) = \text{min rank }A,$$
86: the maximum multiplicity for
87: $G$ and the minimum symmetric rank in $S(G)$, respectively.
88: Because all eigenvalues of matrices in $S(G)$ are real and translation
89: by a real multiple of the identity does not change membership in
90: $S(G)$, of course
91: $$M(G) + {\rm msr}(G) = n,$$
92: and the two may be viewed interchangeably. This allows us to
93: implicitly assume, that, when working with $M(A)$ and when
94: convenient, 0 is an eigenvalue that attains $M(A)$. The same holds
95: for $M(G)$. When $\mathbb{R}$ is replaced by a field F, then
96: $M(G)$ is defined as the maximum corank of all symmetric matrices
97: with entries in F, and whose graph is G.
98:
99: In \cite{F}, it was observed that the only graph $G$ for which
100: $M(G) = 1$ is the path on $n$ vertices. In \cite{JLD}, the maximum
101: multiplicity $M(G)$ has been determined whenever $G$ is a tree.
102: Our purpose here is to describe all graphs $G$ for which $M(G) =
103: 2$, a much larger (than $M(G) = 1$) and more subtle to describe
104: class.
105:
106: \section{Partial 2-trees and preliminaries}
107: \setcounter{equation}{0}
108:
109:
110: Recall that a {\em k-tree} is a graph sequentially constructed
111: from $k+1$-cliques ($K_{k+1}$) via articulation along $k$-cliques.
112: Thus, a traditional tree is a 1-tree. We are particularly
113: interested here in 2-trees, in which the building blocks are
114: triangles ($K_3$'s) and the articulation is along edges. A {\em
115: partial $k$-tree} is a $k$-tree from which some edges (without
116: incident vertices) have been deleted. We call a 2-tree linear if
117: it has precisely two vertices of degree two; we also consider
118: $K_3$ to be a (degenerate) linear 2-tree. In this event, there is
119: a natural order to the triangles and a linear 2-tree is somewhat
120: analogous to a path (though it should be noted that a linear
121: 2-tree may have vertices of arbitrarily high degree).
122:
123: A graph $H$ is a {\em homeomorph} of a graph $G$ if $H$ may be
124: obtained from $G$ by a sequence of edge subdivisions. We use
125: $hK_4$ and $hK_{2,3}$ to denote graphs that are homeomorphs of
126: $K_4$ and $K_{2,3}$ (the complete bipartite graph on two and three
127: vertices) respectively. An $hK_{2,3}$ is just the result of
128: articulation of two cycles along a common induced path of at least
129: two edges.
130:
131: \noindent{\bf Examples.} Let
132: $$
133: A_1 =
134: \begin{bmatrix}
135: 1 & 1 & 1 & 1\\
136: 1 & 1 & 1 & 1\\
137: 1 & 1 & 1 & 1\\
138: 1 & 1 & 1 & 1
139: \end{bmatrix} \in S(K_4)
140: $$
141: and
142: $$
143: A_2 =
144: \begin{bmatrix}
145: -1 & 0 & 1 & 1 & 1\\
146: 0 & 1 & 1 & 1 & 1\\
147: 1 & 1 & 0 & 0 & 0\\
148: 1 & 1 & 0 & 0 & 0\\
149: 1 & 1 & 0 & 0 & 0
150: \end{bmatrix} \in S(K_{2,3}).
151: $$
152: Then ${\rm rank}(A_1) = 1$ and ${\rm rank}(A_2) = 2$, and it is easy
153: to see that $M(K_4) = M(K_{2,3}) = 3$. Note that the Schur
154: complement (see \cite{HJ1}, Ch.~0) in $A_2$ of the $(1,1)$ entry gives $A_1$ and, thus,
155: something whose graph is $K_4$.
156:
157: \begin{La}\label{L2-1}
158: Let $G^\prime$ be the graph resulting from an edge subdivision in
159: the graph $G$. Then $M(G^\prime) = M(G)$ or $M(G^\prime) = M(G) +
160: 1$ (i.e., $M(G^\prime) \geq M(G)$).
161:
162: \end{La}
163: \begin{proof} Denote by $e = (v_1,v_2)$ the edge in $G$ that is
164: subdivided to obtain
165: $G^\prime$. After subdividing $e$, we get a new vertex $v$
166: whose only neighbors are $v_1$ and $v_2$.
167: Let us number vertices $v_1$, $v_2$, and $v$ by the numbers
168: $n-1$, $n$, and $n+1$, respectively.
169: Here and in the sequel we shall assume that if some vertices
170: of a graph $G$ have been numbered, then any matrix in
171: $S(G)$ that we consider is consistent with the numbering
172: (we shall only use integers in the set $\{1, \ldots, n\}$).
173: Note that by permutation similarity we may always transform
174: an arbitrary matrix $B \in S(G)$ to one consistent with a
175: numbering.
176: Let $A \in S(G^\prime)$
177: satisfy $M(A) = M(G^\prime)$, i.e.,
178: ${\rm rank}\text{ } A = (n+1) - M(G^\prime)$.
179: We split the proof into two (mutually exclusive) cases:
180: \begin{eqnarray*}
181: &(a)& \text{ the $(n+1)^{st}$ diagonal entry of $A$ is nonzero.} \\
182: &(b)& \text{ the $(n+1)^{st}$ diagonal entry of $A$ is zero.}
183: \end{eqnarray*}
184: Let us first suppose that our $A$ as defined above satisfies
185: condition $(a)$. Only the last two off-diagonal entries of the
186: $(n+1)^{st}$ row and of the $(n+1)^{st}$ column are nonzero. We
187: may therefore add multiples of the $(n+1)^{st}$ column of $A$ to
188: columns $n-1$ and $n$ so that the entry in the last row of each
189: column is zero. By symmetry we may simultaneously perform the same
190: operation with the roles of rows and columns interchanged. Call
191: the matrix we so obtain $\tilde{A}$. As a result of our
192: operations, $\tilde{A}$ is a direct sum of a (real symmetric)
193: matrix $B$ with graph $G$ and a single nonzero number $x$, i.e.,
194: \begin{equation*}
195: \tilde{A} = \begin{bmatrix}B & 0 \\ 0 & x\end{bmatrix}.
196: \end{equation*}
197: Since $A$ was chosen to be of minimum
198: possible rank, it follows that $B$ has minimum possible rank also, and
199: so ${\rm rank}\text{ } B = n - M(G)$. Therefore
200: the chain of equalities
201: \begin{equation}
202: n - M(G) + 1 = {\rm rank }\text{ }B + 1 = {\rm rank}\text{ }A = (n+1) - M(G^\prime)
203: \label{2.1}
204: \end{equation}
205: holds, whence $M(G^\prime) = M(G)$.
206:
207: Now let us suppose that $A$ satisfies case $(b)$
208: and moreover that
209: there is no matrix $A^\prime \in S(G)$ with $M(A^\prime) = M(G)$ that
210: satisfies condition $(a)$.
211: Add $1$ to the $(n+1)^{st}$ diagonal entry of $A$
212: and call the new matrix $\tilde{A}$.
213: Due to our assumption, ${\rm rank}\text{ } \tilde{A} =
214: 1 + {\rm rank}\text{ } A$, and so
215: \begin{equation}
216: {\rm rank}\text{ } \tilde{A} = (n+1) - M(G^\prime) + 1.
217: \label{2.2}
218: \end{equation}
219: We now apply to the matrix $\tilde{A}$ the procedure
220: used to prove part $(a)$.
221: By (\ref{2.1}) and (\ref{2.2}), we obtain
222: $M(G^\prime) = M(G) + 1$.
223: \end{proof}
224:
225: \noindent{\bf Remark.} We note that both eventualities may occur.
226: If $G$ is a cycle, then $M(G^\prime) = M(G)$, and if
227: %$G = $ Insert picture of two cycles joined along an edge],
228: $G$ consists of two cycles that overlap in one (and only one) edge,
229:
230: %\includegraphics[bb=0 0 437 400, width=70mm]{mypics/TwoCycles3}
231: \includegraphics[width=70mm]{mypics/TwoCycles3}
232:
233: \noindent then
234: $M(G^\prime) = M(G) + 1$ if the overlapping edge is subdivided.
235:
236: \vspace{\baselineskip}
237:
238: Because of Lemma \ref{L2-1}, we see that any graph $G$ that is either an
239: $hK_4$ or an $hK_{2,3}$ satisfies $M(G) \geq 3$.
240:
241: The following combinatorial characterization of partial 2-trees
242: is known (see \cite{WC} or \cite{BLS})
243: and will be useful to us.
244:
245: \begin{La}
246: The graph $G$ is a partial 2-tree if and only if $G$ does not
247: contain an induced subgraph that is a supergraph of an $hK_4$.
248: \label{L2-2}
249: \end{La}
250:
251: We may now establish a key step in our characterization of graphs
252: for which $M(G) = 2$. Our proof of a statement for more general
253: fields is given in Appendix A.
254:
255: \begin{La}
256: If $G$ is a graph such that $M(G) = 2$, then $G$ is a partial
257: 2-tree.
258: \label{L2-3}
259: \end{La}
260:
261: In fact, a stronger result than Lemma \ref{L2-3} holds (cf.
262: \cite{vdH} ), namely: if G is not a partial 2-tree, then there
263: exists a positive semi-definite matrix in $S(G)$ with nullity
264: $\geq 3$. But this result has no natural analogue over general
265: fields.
266:
267: Of course, not all partial 2-trees have maximum multiplicity two.
268: For example, $K_{2,3}$ is a partial 2-tree (simply add an edge
269: between the two vertices in the first part to produce a ``book''
270: of triangles articulated at a single edge, a graph for which the
271: maximum multiplicity is also greater than two). The rest of our
272: work is to sort out which partial 2-trees do have maximum
273: multiplicity two. In the next section, we identify the major
274: portion of them, but certain ``exceptions'' will be identified
275: later.
276:
277: \section{Graphs of two parallel paths}
278: \setcounter{equation}{0}
279:
280: \begin{Dn}
281: A graph $G$ is a graph of two parallel paths if there exist two
282: independent induced paths of $G$ that cover all the vertices of
283: $G$ and such that any edges between the two paths can be drawn
284: so as to not cross. A simple path is not considered to be
285: such a graph (and two paths not connected is considered to be
286: such a graph).
287: \end{Dn}
288: We shall call two independent induced paths satisfying the conditions
289: in the above definition a {\em pair of parallel paths}.
290: We note that $K_3$ is a graph of two parallel paths, and,
291: in any given pair of two parallel paths of $K_3$, one of these paths
292: is degenerate (a vertex). We note also that each graph of two
293: parallel paths is a partial linear 2-tree. This we shall prove
294: after we elaborate on the meaning of the
295: the requirements in the definition.
296:
297: \begin{Rk}
298: The matrix structure of graphs of two parallel paths: \label{R3-2}
299: \end{Rk}
300:
301: Here we express more precisely the requirement in the above
302: definition that a pair of parallel path may be drawn so that edges
303: between the path do not cross. Suppose we have a graph $G$ on $n$
304: vertices such that there exists a pair of independent paths $P_1$
305: and $P_2$. Let $k_i$ denote the number of vertices of $P_i$.
306: Number the vertices of $P_1$ consecutively from $1$ to $k_1$,
307: starting from a pendant vertex of $P_1$. We shall number the
308: vertices of $P_2$ similarly, but in this case we shall require
309: that we start numbering from a pendant vertex of $P_2$ such that
310: we minimize the number of times the following situation occurs: a
311: vertex $j$ of $P_2$ is adjacent to a neighbor $s$ of $P_1$ and
312: some vertex $k
313: > j$ of $P_2$ is adjacent to some $t < s$ of $P_1$. The vertices
314: of $P_2$ may be numbered so that this situation never occurs if
315: and only if $G$ is a graph of two parallel paths. This may be
316: rephrased as following. $G$ is a graph of two parallel paths if
317: and only if there exists $A \in S(G)$ of the following form:
318: \begin{equation*}
319: A = \begin{bmatrix}T_1 & B \\ B^t & T_2\end{bmatrix},
320: \end{equation*}
321: where $T_1$ and $T_2$ are irreducible and tridiagonal and $B$
322: satisfies the following:
323: \begin{equation*}
324: \begin{array}{l}
325: \text{Whenever $b_{ij} \neq 0$ for some entry, then $b_{kl} = 0$}\\
326: \text{for $k > i$ and $l < j$, and for $k < i$ and $l > j$}.
327: \end{array}
328: \end{equation*}
329: and in addition $B$ is such that whenever $B \neq 0$ and
330: $b_{k_1,k_1+1} \neq 0$, this entry is not
331: the only nonzero entry of $B$ (this excludes
332: paths).
333:
334: \begin{Rk}
335: If $G$ is a graph of two parallel paths, and $P_1$ and $P_2$
336: constitute a pair of two parallel paths, then we shall assume that
337: $P_1$ and $P_2$ are numbered so that vertices of $P_1$ have lower
338: numbers than those of $P_2$, and so that $A \in S(G)$ is of the
339: form given in Remark \ref{R3-2}.
340: \end{Rk}
341:
342: \begin{La}
343: If a graph $G$ is a graph of two parallel paths, then $G$ is a partial
344: linear 2-tree.
345: \end{La}
346: \begin{proof}
347: Let $G$ be a graph of two parallel paths on $n$ vertices, and let
348: $p_1$ and $p_2$ be a pair of parallel paths, with $n_1$ and $n_2$
349: vertices, respectively. We shall construct a finite sequence of
350: graphs, the last of which we shall show must be a linear 2-tree.
351: Suppose $H$ is a graph of two parallel paths with a pair of paths
352: $q_1$ and $q_2$. Let $v$ be a vertex in $q_1$ numbered with $j$.
353: We shall define a function ${\rm upper}_H(j)$ on the vertices of
354: $q_1$ as follows. Let $S$ be the set of vertices of $q_1$ with
355: label $j+1$ or greater. Let $S^\prime$ be the subset of all
356: vertices in $S$ that are adjacent to some vertex in $q_2$. If
357: $S^\prime$ is empty, define ${\rm upper}_H(j)$ to be the last
358: vertex in $q_2$. If $S^\prime$ is not empty, take the vertex in
359: this set with the lowest number, call it $w$, and define ${\rm
360: upper}_H(j)$ to be the lowest numbered neighbor of $w$. Define
361: ${\rm lower}_H(j)$ similarly, but with the roles of ``greatest''
362: and ``lowest'' interchanged; S is now defined as the set of
363: verices of $q_1$ with label $j-1$ or lower, and $S^\prime$ is
364: defined in the obvious way. Let us now construct the following
365: sequence of graphs. Let $G_0 = G$ and, for $i = 1, \ldots, n_1$,
366: let $G_i$ be the supergraph of $G_{i-1}$ obtained by articulating
367: edges $(i,j)$, where $j$ runs from ${\rm lower}_{G_{i-1}}(i)$ to
368: ${\rm upper}_{G_{i-1}}(i)$. It may happen that some of these edges
369: already exist. Note that the graph $G_{n_1}$, by construction,
370: consists of only triangles, and so it a 2-tree. Moreover, either
371: $1$ or $n_1 + 1$ is of degree two, and either $n_1$ or $n$ is of
372: degree two, and all other vertices are of degree at least three.
373: To see this, note that not both $1$ and $n_1 + 1$ can be of degree
374: greater than two, for in that case $p_1$ and $p_2$ could not be
375: drawn so that edges do not overlap. If, in $G$ we do not already
376: have that each of these vertices is of degree at least two, then,
377: by considering the remaining cases, it is easy to see that our
378: construction gives the result claimed. We may argue similarly for
379: vertices $n_1$ and $n$. On the other hand, any other vertex in
380: $p_1$ must be of degree at least three, since ${\rm
381: lower}_{G_{i-1}}(j) \leq {\rm upper}_{G_{i-1}}(j)$ for each $i$
382: and $j$. It is also easy to see that each vertex of $p_2$ of
383: degree two relative to $p_2$ is numbered such that, for some $i$,
384: $j$, ${\rm lower}_{G_{i-1}}(j) \leq p_2 \leq {\rm
385: upper}_{G_{i-1}}(j)$. Therefore $G_{n_1}$ is a 2-tree with
386: precisely two vertices of degree two, whence a linear a 2-tree.
387: \end{proof}
388:
389: \begin{Dn}
390: We say that a graph is $C_2$ if it is connected and no vertex has
391: degree less than two (no pendant vertices).
392: \end{Dn}
393:
394: \begin{Dn}
395: The core of a connected graph $G$ is the maximal induced subgraph that is
396: $C_2$.
397: \end{Dn}
398:
399: \begin{La}
400: Suppose that $G$ is a graph of two parallel paths on $n$ vertices.
401: Then ${\rm msr}(G) = n - 2$, or, equivalently, $M(G) = 2$.
402: \label{L3-7}
403: \end{La}
404: \begin{proof}
405: Suppose that the two parallel paths are $P_1$ and $P_2$ with $k_1$
406: and $k_2$ vertices respectively, $k_1 + k_2 = n$. By definition,
407: $k_1,k_2 \geq 1$. We number the vertices of $P_1:1,2,\ldots,k_1$
408: consecutively along the path and the vertices of $P_2:k_1 + 1,
409: \ldots,n$ also consecutively along the path, but beginning with
410: $k_1 + 1$ at the same end of $P_2$ as $1$ is of $P_1$ (if
411: unambiguous, and at either end otherwise). Let $A \in S(G)$. We
412: show by induction on $n$ that $B=A(\{k_1,n\},\{1,k_1+1\})$, the
413: submatrix of $A$ obtained by deleting rows $k_1$ and n and columns
414: 1 and $k_1+1$, is permutation equivalent to a triangular matrix
415: with nonzero diagonal. In the event that $k_1 = 1$ or $k_2 = 1$,
416: this is immediate, as then the other is $n - 1$ and the indicated
417: submatrix is a necessarily triangular submatrix whose diagonal is
418: nonzero because of an irreducible tridiagonal principal submatrix.
419: Thus, we may assume that $k_1,k_2 \geq 2$ and that the cases $n =
420: 2,3$ have been verified. To start the induction, the case $n = 4,
421: k_1 = 2 = k_2$, is also easily checked. Because of the ``no
422: crossing'' requirement upon edges between $P_1$ and $P_2$, not
423: both vertices $1$ and $k_1 + 1$ can have degree more than $2$ and
424: not both vertices $k_1$ and $n$ can have degree more than $2$.
425: Thus, in the matrix $B$, either column $k_1$ or $n$ (original
426: numbering) has exactly one nonzero entry. We may assume without
427: loss of generality that it is column $n$ and that it appears in
428: the last (current numbering) position of that column of $B$.
429: Deletion of this row from $B$ leaves, either a path, or, by the
430: induction hypothesis, a matrix containing an $(n-3)-$triangle,
431: which is extended to an $(n-2)-$triangle by the nonzero in the
432: last column of $B$, to complete the induction.
433:
434: We have shown that msr$(G)\geq n-2$. In fact, the method of proof,
435: which is purely combinatorial, shows that the same conclusion
436: holds for any field F.
437:
438: To complete the proof, we note that G is not a path and apply
439: \cite{F} (or, for any infinite field, apply \cite{BD}) to conclude
440: that msr$(G)\leq n-2$.
441: \end{proof}
442:
443:
444: \noindent{\bf Example.}
445: Let $G$ be the unique 5-vertex linear 2-tree. Let $P_1$
446: and $P_2$ be two parallel paths and suppose that $P_1$
447: has two vertices. Furthermore, suppose that the first
448: and last vertices of $P_2$ are each of degree two.
449: In this example we shall use
450: $P(A^\prime)$ to denote the pattern of $A^\prime$.
451: Numbering according to our prescription
452: implies that matrices in $S(G)$ consistent with this numbering
453: have the pattern
454: \begin{equation*}
455: P(S(G)) :=
456: \begin{bmatrix}
457: \cdot & \times & \times & \times & 0 \\
458: \times & \cdot & 0 & \times & \times\\
459: \times & 0 & \cdot & \times & 0\\
460: \times & \times & \times & \cdot & \times\\
461: 0 & \times & 0 & \times & \cdot
462: \end{bmatrix}
463: \end{equation*}
464: where $\times$ denotes a nonzero entry and $\cdot$ a completely
465: free entry. Let $A \in S(G)$ satisfy $M(A) = M(G)$. In
466: particular, $A$ has the sign pattern given above, i.e. P(A) =
467: P(S(G)). The pattern of $P_1$ is given by the upper diagonal $2
468: \times 2$ block in $P(A)$, and $P_2$ by the lower diagonal $3
469: \times 3$ block. The pattern of $B$ is given by the uppermost
470: rightmost $2 \times 3$ block in $P(G)$. Striking out rows 2 and 5
471: and columns 1 and 3, we obtain a submatrix $\tilde{A}$ with
472: pattern
473: \begin{equation*}
474: P(\tilde{A}) =
475: \begin{bmatrix}
476: \times & \times & 0\\
477: 0 & \times & 0\\
478: \times & \cdot & \times
479: \end{bmatrix}
480: \end{equation*}
481: By taking the Schur complement with respect to the (3,3) nonzero
482: entry, we see that $\tilde{A}$ has full rank, and therefore
483: ${\rm msr}(A) \geq 3$, or equivalently, $M(A) < 3$. Since $A$ is not
484: a path, $M(A) > 1$, and therefore we conclude that $M(A) = 2$.
485:
486: \section{Graphs of minimum degree two and M = 2}
487: \setcounter{equation}{0}
488:
489: \begin{La}
490: If a $C_2$ graph $G$ contains a cut-vertex, then $M(G) > 2$.
491: \label{L4-1}
492: \end{La}
493: \begin{proof} Let us denote the cut-vertex by $v$ and consider
494: the induced subgraph $G- v$. We are left with connected components
495: $K_1, \ldots, K_n$, where $n$ is at least two.
496: Let us introduce the induced subgraphs
497: \begin{equation*}
498: \tilde{K}_1 = G-\{K_2, \ldots, K_n\}
499: \end{equation*}
500: and
501: \begin{equation*}
502: \tilde{K}_2 = G-K_1.
503: \end{equation*}
504: %It is at once verified that $\tilde{K}_1$ and $\tilde{K}_2$
505: %each contain at least one cycle, and that by superimposing these two graphs
506: %at the vertex $v$, we obtain the graph $G$. Let $n_i$ denote
507: %the number of vertices of the graph $\tilde{K}_i$, for $i = 1,2$.
508: %In the graph $\tilde{K}_1$, number the vertex $v$ by $n_1$, and
509: %in $\tilde{K}_2$, number $v$ by $1$.
510: %Choose $A_i \in S(\tilde{K}_i)$ so that
511: %$M(A_i) = M(\tilde{K}_i)$,
512: %for $i = 1,2$.
513: %We now embed $A_1$ and $A_2$ in $(n_1+n_2-1) \times (n_1+n_2-1)$
514: %matrices as given by the following block decompositions:
515: %\begin{equation}
516: %\tilde{A}_1 := \begin{bmatrix}A_1 & 0 \\ 0 & 0\end{bmatrix},
517: %\text{ }
518: %\tilde{A}_2 := \begin{bmatrix}0 & 0 \\ 0 & A_2\end{bmatrix}.
519: %\end{equation}
520: %Note that, even though these matrices have the same dimension,
521: %their partitionings are not conformal.
522: %However, they are nearly conformal in the following sense:
523: %in the Hadamard (entrywise) product of
524: %$\tilde{A}_1$ and $\tilde{A}_2$,
525: %the only entry that may be nonzero is
526: %the $n_1^{st}$ diagonal
527: %entry (corresponding to the vertex $v$).
528: %This and the fact that
529: %our graphs are $C_2$ implies that
530: %\begin{equation}
531: %M(\tilde{A}_1 + \tilde{A}_2) \geq
532: %M(A_1) + M(A_2) + 1 \geq
533: %2 + 2 - 1 = 3.
534: %\label{CutVertexInequality}
535: %\end{equation}
536: %The sum $\tilde{A}_1 + \tilde{A}_2$ has the graph that
537: %results from
538: %articulating $\tilde{K}_1$ and $\tilde{K}_2$ at the vertex
539: %$v$, i.e., the graph $G$.
540: %This combined with
541: %(\ref{CutVertexInequality}) implies that
542: %$M(G) > 2$.
543: Neither $\tilde{K}_1$ nor $\tilde{K}_2$ is a path. By \cite{F} ,
544: \cite{BD}, the lemma holds for every infinite field $F$.
545: \end{proof}
546:
547: The following lemma is a special case of Theorem 2.3 in \cite
548: {BFH}, and is brought here for the sake of completeness. It holds
549: for any field F.
550: \begin{La}
551: Let $G$ be a graph containing a pendant vertex $v$ with unique
552: neighbor $u$. Then
553: $M(G) = {\rm max}\left\{M\left(G-v\right), M\left(G -
554: \{u,v\}\right)\right\}$.
555: \label{L4-2}
556: \end{La}
557: \begin{proof} Number the vertex $v$ by $n$ and $u$ by $n-1$.
558: Let $A \in S(G)$ satisfy $M(A) = M(G)$. We consider
559: separately two
560: (mutually exclusive) cases:
561: \begin{eqnarray*}
562: &(a)& \text{ the $n^{th}$ diagonal entry of $A$ is nonzero.} \\
563: &(b)& \text{ the $n^{th}$ diagonal entry of $A$ is zero.}
564: \end{eqnarray*}
565: Let us suppose that we may find a matrix $A$ as defined above
566: such that condition $(a)$ holds. Taking into account that
567: the only nonzero entries in the last row are entries $n-1$ and $n$,
568: we add multiples of the last row of $A$ to the row $n-1$
569: so that its $n^{th}$ entry becomes zero.
570: We simultaneously perform the same procedure with the roles
571: of rows and columns reversed. Let us call the resulting matrix
572: $\tilde{A}$. As a result
573: of our operations, $\tilde{A}$ is a direct sum of a
574: (real symmetric) matrix $B$ with graph $G-v$ and a single
575: nonzero number $x$:
576: \begin{equation}
577: \tilde{A} = \begin{bmatrix}B & 0\\0 & x\end{bmatrix}.
578: \label{4.1}
579: \end{equation}
580: Note that $M(A) = M(G)$ forces $M(B) = M(G-v)$: clearly $M(B) \leq
581: M(G - v)$; if $M(B) < M(G - v)$, then ${\rm msr}(G) > {\rm msr}(G
582: - v) + 1$, a contradiction. Because $x \neq 0$ in our direct sum
583: decomposition, it holds that $M(G) = M(A) = M(B) = M(G-v)$.
584:
585: Now suppose that $A$ satisfies condition $(b)$ and that there
586: is no matrix $A^\prime \in S(G)$ with $M(A^\prime) = M(G)$
587: that satisfies condition $(a)$.
588: Since $(b)$ is satisfied, in the last row of $A$ only the $(n-1)^{st}$
589: entry is nonzero. We therefore may add multiples of the last
590: row to each row of $A$, canceling all nonzero entries
591: of the $(n-1)^{st}$ column of $A$ without affecting any other
592: entries. By symmetry, we may perform the same operation
593: with the roles of rows and columns reversed. Let us
594: call the resulting matrix $\hat{A}$. Note that $\hat{A}$
595: is a direct sum of a (real symmetric) matrix $B$ with graph
596: $G-\{u,v\}$ and a $2\times 2$ matrix $X$ given by
597: $X = \begin{bmatrix}0 & x \\ x & 0\end{bmatrix}$, where $x$
598: is nonzero:
599: \begin{equation}
600: \hat{A} = \begin{bmatrix}B & 0 \\ 0 & X\end{bmatrix}.
601: \label{4.2}
602: \end{equation}
603: Since $M(A) = M(G)$, it follows that $M(B) = M(G-\{u,v\})$.
604: By our direct sum decomposition and the fact that $X$ has
605: full rank, we get that $M(G) = M(A) = M(B) = M(G-\{u,v\})$.
606: Thus we have established that
607: $M(G) \in \left\{M(G-v), M(G-\{u,v\}\right\}.$ Now, suppose
608: that we start with a matrix $\tilde{A}$ as given in
609: (\ref{4.1}) where $B \in S(G-v)$, $M(B) = M(G-v)$,
610: and $x$ is nonzero. By reversing all of our row and column
611: operations, we may obtain a matrix $A \in S(G)$ with $M(A) = M(G-v)$.
612: Similarly, if we start with a matrix $\hat{A}$ as given
613: in (\ref{4.2}) where $B \in S(G-\{u,v\})$,
614: $M(B) = M(G-\{u,v\})$, and $X = \begin{bmatrix}0 & x\\x & 0
615: \end{bmatrix}$, where $x \neq 0$, then by reversing our row
616: and column operations there performed, we may obtain a matrix
617: $A \in S(G)$ with $M(A) = M(G-\{u,v\})$.
618: \end{proof}
619:
620: \begin{La}
621: Let $G$ be a graph containing an induced subgraph that is a
622: supergraph of an $hK_{2,3}$. Then $M(G) \geq 3$.
623: \label{L4-3}
624: \end{La}
625: \begin{proof}
626: Let the $hK_{2,3}$ consist of three internally independent paths,
627: of at least two edges each between vertices $u$ and $v$. Call them
628: $P_1$, $P_2$, and $P_3$. If $G$ contains an $hK_4$, then the
629: conclusion follows from Lemma \ref{L2-3}. Thus, we may assume
630: that there is no path in $G-\{u,v\}$ from an interior vertex of
631: $P_i$ to an interior vertex of $P_j$, $j \neq i$, $1 \leq i,j \leq
632: 3$. It follows that if $u$ and $v$ are deleted from $G$, then at
633: least three components result, including a component
634: ``corresponding'' to each $P_i$, $i=1,2,3$. We conclude that $A
635: \in S(G)$ appears, with proper numbering of vertices, as
636: \begin{equation*}
637: A =
638: \begin{bmatrix}
639: \cdot & ? & d_1 & d_2 & d_3 & ? \\
640: ? & \cdot & f_1 & f_2 & f_3 & ? \\
641: d_1^t & f_1^t & A_1 & 0 & 0 & 0\\
642: d_2^t & f_2^t & 0 & A_2 & 0 & 0\\
643: d_3^t & f_3^t & 0 & 0 & A_3 & 0\\
644: ? & ? & 0 & 0 & 0 & B
645: \end{bmatrix}
646: \end{equation*}
647: Here $A_i$ corresponds to component $G_i$ and includes the interior
648: vertices of $P_i$, $i=1,2,3$, $B$ may be empty, and the first two
649: rows and columns correspond to $u$ and $v$. Each $d_i$ has its
650: first entry nonzero, and each $f_i$ has its last
651: entry nonzero. The symbol ``?'' is either free nonzero or zero.
652: Now, identify the three
653: neighbors of $u$ along the path $P_1$, $P_2$, and $P_3$ together with
654: $u$ and $v$ to give the set $S$ contained among the vertices of $G$.
655: Let $A \in S(G)$ be chosen as follows: $A[S^c]$ is an $M$-matrix
656: (see \cite{HJ2}, Ch.~2),
657: the three edges from $u$ to its neighbors in $S$ are free, the
658: diagonal entries corresponding to $S$ are free and all other edges
659: in $G$ are chosen to be positive. The Schur complement
660: $A / A[S^c]$ then appears as
661: \begin{equation*}
662: \left[
663: \begin{array}{c|c}
664: \begin{matrix}\tilde{a}_{11} & ? \\ ? & \tilde{a}_{22}\end{matrix} &
665: \begin{matrix}\tilde{a}_{13} & \tilde{a}_{14} & \tilde{a}_{15} \\ a & b & c\end{matrix}\\
666: \hline
667: \begin{matrix}\tilde{a}_{13} & a \\ \tilde{a}_{14} & b \\ \tilde{a}_{15} & c\end{matrix} &
668: \begin{matrix}\tilde{a}_{33} & 0 & 0 \\ 0 & \tilde{a}_{44} & 0 \\
669: 0 & 0 & \tilde{a}_{55} \end{matrix}
670: \end{array}
671: \right]
672: \end{equation*}
673: in which ? is either free nonzero, fixed nonzero, or zero, $\tilde{a}_{13}$, $\tilde{a}_{14}$,
674: $\tilde{a}_{15}$ are free nonzero, $a$, $b$, $c$ are nonzero and
675: $\tilde{a}_{11}, \ldots, \tilde{a}_{55}$ are free.
676:
677: To see that $a$, $b$, and $c$ may be chosen to be nonzero, consider
678: the following two situations. Denote by $w$ the neighbor of $u$
679: along $P_1$. If $w$ is not a neighbor of $v$, then the sign of
680: the second summand in the Schur complement $A / A[S^c]$ shows that
681: $a < 0$. If $w$ is a neighbor of $v$, then $a$ is given
682: as the difference of two positive numbers. In this case we may
683: treat the entry corresponding to the edge $(w,v)$ as a free variable,
684: and therefore guarantee that $a \neq 0$. The same argument holds for
685: $b$ and $c$.
686:
687: If we choose
688: $\tilde{a}_{33} = \tilde{a}_{44} = \tilde{a}_{55} = 0$, $\begin{bmatrix}\tilde{a}_{13}
689: & \tilde{a}_{14} & \tilde{a}_{15}\end{bmatrix}$
690: proportional to $\begin{bmatrix}a & b & c\end{bmatrix}$ and either
691: $\begin{bmatrix}\tilde{a}_{11} & ? \\ ? & \tilde{a}_{22}\end{bmatrix}$ to be
692: 0 if $? = 0$ or to be so that
693: $\begin{bmatrix}\tilde{a}_{11} & ? & \tilde{a}_{13} \\
694: ? & \tilde{a}_{22} & a\end{bmatrix}$ is rank one, otherwise, the proof
695: is complete.
696: Note that, because of the Schur complement step, $\tilde{a}_{13}$,
697: $\tilde{a}_{14}$, $\tilde{a}_{15}$ and $?$ could each have a single forbidden
698: value other than zero. This does not debilitate the argument
699: because of the flexibility in proportionality and in the
700: upper left 2-by-2 block.
701: \end{proof}
702: \begin{Rk}
703: Lemma \ref{L4-3} holds for any infinite field. Indeed, a
704: straightforward computation shows that the entries in the
705: $3,4;3,5$ and $4,5$ positions of $A/A[S^c]$, as well as the
706: entries in the transpose positions are 0. The results about the
707: other entries of $A/A[S^c]$ follow the discussion in Appendix A,
708: and in particular Lemma \ref{LA-2}
709: \end{Rk}
710: \begin{Dn}
711: A graph $G$ is a $SEAC$ (singly edge-articulated cycle) graph if
712: it is sequentially built from cycles via articulation of each
713: successive cycle along an edge of the previous graph. No such edge
714: may be used more than once for articulation.
715: \end{Dn}
716:
717: \begin{La}
718: If $G$ is a $C_2$ graph such that $M(G) = 2$, then $G$ is a $SEAC$
719: graph.
720: \label{L4-5}
721: \end{La}
722: %\begin{proof} Let $G$ be a $C_2$ graph. If $M(G) = 2$, then by Lemma
723: %\ref{L2-3}, $G$ must be a partial 2-tree. Suppose that $G$
724: %is a $C_2$ partial 2-tree that is not a $SEAC$ graph. It is
725: %trivial that $G$ then satisfies at least one of the following:
726: %\begin{enumerate}
727: %\item An edge of $G$ was used for articulation more than once.
728: %\item A cycle of $G$ was articulated along an induced path
729: % of at least two edges.
730: %\item $G$ cannot be constructed by sequentially articulating
731: %cycles along an induced path of the previous graph.
732: %\end{enumerate}
733: %Closer inspection, though, shows that condition $(3)$ can never
734: %arise when $G$ is a $C_2$ partial 2-tree. The argument goes
735: %as follows. Since $G$ is a partial 2-tree, we may find a
736: %2-tree $H$ from which we may obtain $G$ by successively
737: %removing edges from $H$. Since $H$ is a 2-tree, we may
738: %number (with distinct positive integers)
739: %the cliques in such a way that $H$ may be
740: %sequentially built
741: %by articulating cliques along single edges of previous
742: %cliques, going in order from the lowest to highest number. Now,
743: %keeping this ordering in mind, successively remove edges
744: %from $H$ to obtain $G$. Cliques may merge to become (larger) cycles.
745: %Whenever this occurs, assign the cycle the lesser of the
746: %two integers corresponding the merging cliques.
747: %Continuing in this way, we see that $G$ is a graph consisting
748: %of cycles (it is a $C_2$ graph)
749: %and, moreover, we obtain a numbering of the cycles.
750: %Because of our choice of numbering for $H$, we are left
751: %with a numbering for $G$ that shows that $G$ may
752: %be sequentially built by articulating cycles along
753: %paths of previous cycles, articulating in order from lowest
754: %to highest number.
755: %Therefore, a $C_2$ partial 2-tree
756: %satisfies either condition $(1)$ or condition $(2)$
757: %(or both) given above.
758: %
759: %Suppose that $G$ satisfies condition $(2)$
760: %above. Then it
761: %contains at least two cycles, $Z_1$ and $Z_2$, that share a
762: %maximal induced path
763: %$P$ in common, where $P$ consists of at least two edges.
764: %Let $v_1$ and $v_2$
765: %denote the two end-vertices of the induced path $P$. Then
766: %vertices $v_1$ and $v_2$, with induced paths $P$, $Z_1 - P$,
767: %and $Z_2 - P$ satisfy the conditions of Lemma \ref{L4-3},
768: %which therefore implies that $M(G) > 2$.
769: %
770: %Now suppose that $G$ is $C_2$ and satisfies condition
771: %$(1)$ above. Then it
772: %contains an edge $e = (v_1,v_2)$ such that $e$
773: %is common to cycles $Z_1$, $Z_2$, $Z_3$. Let us assume that $e$
774: %is the path of maximal length common to all three cycles
775: %(otherwise we satisfy the conditions of case $(a)$).
776: %Now
777: %$v_1$ and $v_2$ along with induced paths $Z_1 - e$, $Z_2 - e$,
778: %and $Z_3 - e$ satisfy the conditions of Lemma \ref{L4-3}
779: %and so $M(G) > 2$, a contradiction.
780: %
781: %We have shown that any graph $G$ that is a $C_2$ graph satisfying
782: %$M(G) = 2$ is a $C_2$ partial 2-tree, and that any $C_2$ partial
783: %2-tree $H$ that is not a $SEAC$ graph has $M(H) > 2$. The result
784: %now follows.
785: %\end{proof}
786: \begin{proof}
787: By Lemmas \ref{L2-3} and \ref{L4-1} we can assume that $G$ is a
788: partial 2-tree which does not contain any cut vertex. The lemma is
789: proved by induction on $n=|G|$. It holds for $n=3$, so we consider
790: the general induction step.\\
791:
792: G is a partial 2-tree, so it has a supergraph $\tilde{G}$ with
793: $|\tilde{G}|=|G|=n$, and such that $G$ is obtained from
794: $\tilde{G}$ by removing some edges. $\tilde{G}$ consists of $n-2$
795: triangles. Denote by $T_{n-2}$ the last one that was articulated
796: in the construction of $\tilde{G}$. Denote its vertices by
797: $u,v,w$, where we may assume that the degree of w in $\tilde{G}$
798: is 2.\\ \\
799: \centerline{\includegraphics[width=70mm]{mypics/5}}\\
800:
801: Since $G$ contains no cut vertices, $uw$ and $vw$ are edges of
802: $G$.%
803: \Cs Suppose that $uv$ is an edge of $G$. Then, since $G$ contains
804: no cut vertices, the degrees of $u$ and $v$ in $G$ are at least 3.
805:
806: Let $G'=G-\{w\} .$ Clearly, $G'$ is a $C_2$ graph with $|G'|=n-1$.
807: By \cite{F} we cannot have $\text{msr}(G')=(n-1)-1 .$ We cannot
808: have $\text{msr}(G')\leq (n-1)-3$, because this implies
809: $\text{msr}(G)\leq n-3$. We use here $\text{msr}(K_3)=1$. Hence
810: $\text{msr}(G')=n-3=(n-1)-2$, so we can apply the induction
811: hypothesis and conclude that $G'$ is a SEAC. We are done if $uv$
812: is incident to exactly one cycle in $G'$, and this is indeed the
813: case by Lemma \ref{L4-3}.%
814: \Cs Suppose that $uv$ is not an edge of $G$. Let $G'$ be obtained
815: from $G$ by compressing $w$, so $|G'|=n-1$, and $uv$ is an edge of
816: $G'$. $G'$ is $C_2$, and as in Case 1, msr$(G')=(n-1)-2$, so the
817: induction hypothesis implies that $G'$ is a SEAC. If the edge $uv$
818: (in $G'$) is incident to only one cycle we are done. Otherwise,
819: going back to $G$ we get a contradiction by Lemma \ref{L4-3}.
820: \end{proof}
821:
822:
823: \begin{Dn}
824: In a $SEAC$ graph, the cycles used to build it are well defined, as
825: are the edges of articulation. Each edge of articulation uniquely
826: defines two of the cycles. We say that two of the cycles are
827: neighbors if they share an edge of articulation. A $SEAC$ graph is
828: called linear $(LSEAC)$ if each of its constituent cycles has at
829: most two neighbors. An $LSEAC$ graph that consists of more than a
830: cycle has just two cycles with only one neighbor each (the two ends
831: of the linear path of cycles).
832: \end{Dn}
833:
834: \begin{La}
835: If $G$ is a $C_2$ graph and $M(G) = 2$, then $G$ is an $LSEAC$ graph.
836: \label{L4-7}
837: \end{La}
838: \begin{proof} Let $G$ be a $C_2$ graph with $M(G) = 2$.
839: By Lemma \ref{L4-5}, $G$ is a $SEAC$ graph.
840: Thus it remains to be shown that $G$ is in
841: fact an $LSEAC$ graph. If $G$ is not an $LSEAC$ graph, then
842: there exists a cycle $Z$ with at least $k \geq 3$ neighbors, say
843: $Z_1, \ldots, Z_k$ ($k \geq 3$).
844: These neighbors uniquely determine
845: the connected components of the graph $G - Z$. Let us superimpose
846: the connected component corresponding to $Z_i$
847: with the intersection of $Z_i$ and $Z$, and denote it by
848: $\tilde{Z}_i$. It is clear $\tilde{Z}_i$ is a $SEAC$ graph.
849: We shall assume that each is in fact an $LSEAC$ graph. Once
850: the lemma is proved for this case, the general case follows
851: by induction. Let $s$ denote the number of vertices of $Z$
852: and $s_i$ the number of vertices of $\tilde{Z}_i$. The cycle
853: $Z$ contains $s-k$ edges not shared with some $\tilde{Z}_i$.
854: It therefore follows that
855: \begin{equation}
856: {\rm msr}(G) \leq {\rm msr}(\tilde{Z}_1) + \ldots + {\rm msr}(\tilde{Z}_k) + (s-k).
857: \label{4.3}
858: \end{equation}
859: Since each $\tilde{Z}_i$ is an $LSEAC$ graph,
860: \begin{eqnarray}
861: {\rm msr}(\tilde{Z}_1) + \ldots + {\rm msr}(\tilde{Z}_k) + (s-k) &=&
862: (s_1 - 2) + \ldots + (s_k - 2) + (s-k) \nonumber \\
863: &=& (n-s) + (s-k) = n - k,
864: \label{4.4}
865: \end{eqnarray}
866: By (\ref{4.3}) and (\ref{4.4}), ${\rm msr}(G) \leq n - k$, i.e.,
867: $M(G) \geq k$. In particular, $M(G) > 2$.
868: The general case follows by induction.
869: \end{proof}
870:
871: \begin{Tm}
872: If $G$ is a $C_2$ graph, then the following three statements are
873: equivalent:
874: \begin{enumerate}
875: \item M(G) = 2
876: \item G \text{ is a graph of two parallel paths, and }
877: \item G \text{ is an LSEAC graph.}
878: \end{enumerate}
879: \label{T4-8}
880: \end{Tm}
881: \begin{proof} Note that $2. \implies 1.$ follows from Lemma
882: \ref{L3-7} and that $1. \implies 3.$ follows from Lemma
883: \ref{L4-7}. It remains to be shown that $3. \implies 2.$.
884: This is trivial if $G$ is a single cycle, and so let us
885: suppose that $G$ contains at least two cycles. Let
886: $Z_s$ and $Z_t$ denote the two cycles with only one neighbor
887: each. Let $s_1, s_2$ be two adjacent vertices belonging
888: to $Z_s$ such that the edge that they determine is not an
889: edge of articulation. Define $t_1, t_2$ belonging to $Z_t$
890: similarly.
891: %Let $P_1$ be the path of minimal length that
892: %connects vertex $s_1$ to one of $t_1$ or $t_2$ and that
893: %satisfies the following:
894: %\begin{enumerate}
895: %\item \text{Vertex $s_2$ does not belong to $P_1$.}
896: %\item \text{Only one of $\{t_1, t_2\}$ belongs to $P_1$.}
897: %\item \text{No edge of $G$ that is used for articulation belongs to $P_1$.}
898: %\end{enumerate}
899: %Let $P_2$ be the path covering the remaining vertices.
900: %That such paths exist and are uniquely defined after choosing
901: %$s_1,s_2,t_1,t_2$
902: %may easily be verified by inspecting
903: %the definition of an $LSEAC$.
904: %The paths $P_1$ and $P_2$ are independent induced paths of $G$ that
905: %cover all the vertices of $G$ and are such that they can be drawn
906: %so that no edges between the two paths overlap. This means
907: %precisely that $G$ is a graph of two parallel paths.
908: Now remove from $G$ all edges of articulation. The resulting
909: graph is a cycle that clearly defines the two
910: parallel paths in $G$.
911: \end{proof}
912:
913: \section{The exceptional graphs and the general result}
914: \setcounter{equation}{0}
915:
916: In this section we shall prove the following general result:
917:
918: \begin{Tm}
919: The graph $G$ satisfies $M(G) = 2$ if and only if $G$ is a graph
920: of two parallel paths or $G$
921: is one of the types listed in Table B1.
922: \label{T5-1}
923: \end{Tm}
924:
925: The bold lines in Table B1 indicate edges that may be subdivided
926: arbitrarily many times, whereas the dotted lines indicate paths
927: (possibly degenerate) of arbitrary length. Thus each
928: ''exceptional graph'' that appears in Table B1 stands not for just
929: one graph, but rather for a certain countable collection of
930: graphs.
931:
932: \begin{Rk}
933: We note that among partial linear 2-trees, the graphs for which
934: $M(G) = 2$ are precisely the graphs of two parallel paths. The
935: additional graphs for which $M(A) = 2$, the exceptional families
936: of Table B1, are, of course, partial (not linear) 2-trees with
937: very special structure.
938: \end{Rk}
939:
940: In the previous section, we completely characterized the $C_2$
941: graphs with $M = 2$. Consider now a general connected partial
942: 2-tree $G$. If $G$ is a tree, then $M(G)$ is determined by the
943: path covering number of $G$. In particular when $G$ is a tree,
944: $M(G) = 2$ if and only if $G$ is a graph of two parallel paths.
945: Let us assume then here and throughout this section that $G$
946: denotes a connected partial 2-tree that is a not a tree. By
947: sequentially stripping away all degree-one vertices, we arrive at
948: the maximal induced subgraph of $G$ that is $C_2$, which, by
949: definition, is the core of $G$. Let us denote this induced
950: subgraph by ${\rm Core}(G)$. By inductively applying Lemma
951: \ref{L4-2}, we see that
952: \begin{equation*}
953: M(G) \geq M({\rm Core}(G)),
954: \end{equation*}
955: which, combined with Theorem \ref{T4-8}, proves the following lemma:
956: \begin{La}
957: Let $G$ be a partial 2-tree. If $M(G) = 2$, then the core of $G$
958: is an $LSEAC$ graph and a graph of two parallel paths.
959: \label{L5-2}
960: \end{La}
961: In light of this Lemma,
962: asking for what connected graphs we have
963: $M = 2$ is
964: equivalent to asking under what conditions does
965: sequentially adding degree one vertices to $LSEAC$ graphs
966: preserve maximum multiplicity. Lemma \ref{L4-2} will allow us
967: to prove several ``forbidden subgraph'' lemmas in which it will
968: be shown that certain induced subgraphs of a graph $G$ preclude
969: $M(G) = 2$. Indeed, if we take an $LSEAC$ graph and sequentially
970: add degree one vertices until we have created a forbidden subgraph,
971: then, by Lemma \ref{L4-2} adding subsequent degree one vertices cannot
972: decrease maximum multiplicity.
973:
974: \begin{La}
975: Let $H$ be a $C_2$ graph and let $u$ be an arbitrary vertex of $H$.
976: Let $T$ be a tree that is a not a path and let $v$ be a degree-one vertex of $T$.
977: Let $G$ be the graph that is the result of identifying vertices
978: $u$ and $v$ of $H$ and $T$, respectively. Then $M(G) > 2$.
979: \label{L5-3}
980: \end{La}
981: \begin{proof} Since $T$ is not a path, it has a vertex of degree at
982: least three, say, vertex $w$. We may then find a minimal induced
983: path connecting $w$ to a pendant vertex of $G$ that also belongs
984: to the induced subgraph $T$ in $G$. Let us sequentially
985: remove the vertices of this path, including vertex $w$.
986: The resulting graph has at least two connected
987: components, one of which is
988: a graph that has a nontrivial core, the other being a tree
989: (possibly degenerate).
990: A graph with nontrivial core has $M > 1$, and so the resulting
991: induced subgraph of two connected components has $M > 2$.
992: Applying Lemma
993: \ref{L4-2} inductively shows that this implies that $M(G) > 2$.
994: %Raphy's Proof:
995: %Note that, due to [F-BD],
996: % ${\rm msr}(T) \leq |T| - 2$ and ${\rm msr}(H) \leq |H| - 2$.
997: %Hence ${\rm msr}(G) \leq |T| - 2 + |H| - 2 = |H| + |T| - 4$.
998: %On the other hand, $|G| = |H| + |T| - 1$.
999: \end{proof}
1000:
1001: The immediate consequence of this lemma is that if a given graph $G$
1002: cannot be constructed from an $LSEAC$ graph $H$ by sequentially
1003: articulating paths onto vertices of $H$ only, then $M(G) > 2$.
1004: Suppose now we are given a graph $G$ that can be constructed in such a
1005: manner. Inductively applying Lemma
1006: \ref{L4-2} then shows that in fact we need only consider
1007: the induced subgraph of $G$ obtained by sequentially removing
1008: pendant vertices whose neighbors have degree two, i.e.,
1009: we may assume without loss of generality that all paths
1010: with pendant vertices are of length one.
1011: We shall use this fact implicitly throughout.
1012:
1013: \begin{Dn}
1014: A simple partial 2-tree is a partial 2-tree $G$
1015: whose core is a nontrivial $LSEAC$ graph and which may be constructed
1016: from its core through sequential articulation of vertices to
1017: vertices belonging to its core.
1018: \end{Dn}
1019: We consider $LSEAC$ graphs to be simple partial 2-trees.
1020: The remarks immediately preceding the definition show that if
1021: a graph $G$ is not a simple partial 2-tree, then
1022: $M(G) > 2$.
1023: Whenever $v$ is a vertex of $G$ with a neighbor $u$ that is a
1024: pendant vertex in $G$, we shall call $u$ a pendant neighbor of
1025: $v$.
1026:
1027: \begin{La}
1028: Let $G$ be a simple partial 2-tree and $u$ a vertex of $G$
1029: with at least three pendant neighbors. Then $M(G) > 2$.
1030: \label{L5-5}
1031: \end{La}
1032: \begin{proof}
1033: Let $v_1,v_2,v_3$ denote three (distinct) pendant neighbors of $u$.
1034: Then the induced subgraph $G - \{u,v_3\}$ has three connected
1035: components: isolated vertices $v_1$ and $v_2$, and the induced subgraph
1036: $G - \{u,v_1,v_2\}$. It follows that
1037: \begin{equation*}
1038: M(G - \{u,v_3\}) = M(v_1) + M(v_2) + M(G - \{u,v_1,v_2\}) \geq 1 + 1 + 1 = 3
1039: \end{equation*}
1040: By Lemma \ref{L4-2}, $M(G) \geq M(G - \{u,v_3\})$, since $v_3$ is
1041: pendant, and therefore $M(G) \geq 3$.
1042: \end{proof}
1043:
1044: Next we introduce the definition
1045: of a terminal cycle. We shall see that these cycles play a rather
1046: more important role than the other cycles of a graph.
1047:
1048: \begin{Dn}
1049: A terminal cycle of an $LSEAC$ graph is a cycle that has at most one
1050: neighbor.
1051: \end{Dn}
1052: If $G$ is a graph whose core is an $LSEAC$ graph, then a cycle $Z$ of $G$
1053: is said to be a terminal cycle if and only if $Z$ is a terminal cycle
1054: in the core of $G$. Note that a single cycle is an $LSEAC$ graph and
1055: is, by our definition, considered to be a terminal cycle.
1056:
1057: \begin{La}
1058: Suppose $G$ is a simple partial 2-tree with a cycle $Z$ that is
1059: not terminal. Suppose further that there exists a vertex $u$ in
1060: $Z$ that is of degree at least two in the core of $G$, does not
1061: belong to a terminal cycle, and has at least one pendant neighbor
1062: $v$ in $G$. Then $M(G)
1063: > 2$. \label{L5-7}
1064: \end{La}
1065: \begin{proof}
1066: First suppose that $u$ is of degree two in the core of $G$.
1067: Note that this requirement is
1068: equivalent to requiring that $u$ belong to cycle $Z$ and no other
1069: cycles.
1070: Observe that there exists at least one vertex of $Z$ that is
1071: a cut vertex in the (nontrivial) $C_2$ graph ${\rm Core}(G - \{u,v\})$. For example, consider
1072: a vertex $w$ belonging to $Z$ and one of $Z$'s neighbors such that
1073: a minimal path between $w$ and $u$ includes an edge that serves as
1074: an edge of articulation in the core of $G$.
1075: It follows from Lemma
1076: \ref{L4-1} that $M\left({\rm Core}(G - \{u,v\})\right) > 2$. Applying Lemma \ref{L4-2}
1077: repeatedly,
1078: we get the desired conclusion $M(G) > 2$.
1079:
1080: Now suppose that $u$ is of degree at least three in the core
1081: of $G$. Necessarily, $u$ belongs to at least two cycles, say
1082: $Z_1$ and $Z_2$. Since $Z_1$ and $Z_2$ have a unique edge
1083: of articulation, say $(u,w)$, it follows that $w$ is a cut
1084: vertex in the graph ${\rm Core}(G - \{u,v\})$.
1085: By Lemma
1086: \ref{L4-1}, $M\left({\rm Core}(G - \{u,v\})\right) > 2$, whence $M(G) > 2$,
1087: which completes the proof.
1088: \end{proof}
1089:
1090: \begin{Dn}
1091: A distinguished vertex of an $LSEAC$ graph $G$ is a vertex that belongs
1092: to every terminal cycle of $G$.
1093: \end{Dn}
1094: If $G$ is a graph whose core is an $LSEAC$ graph, then a vertex $v$
1095: of $G$ is said to be distinguished if and only if $v$ is distinguished
1096: in the core of $G$. A vertex $u$ of $G$ that is not a distinguished
1097: vertex shall be called nondistinguished. Note that if $G$ is a single
1098: cycle, then every vertex of $G$ is a distinguished vertex. If $G$
1099: is an $LSEAC$ graph that is not a cycle, then $G$ has either $0$,
1100: $1$, or $2$ distinguished vertices. In fact, if $G$ is any graph
1101: whose core is an $LSEAC$ graph with $n$ distinguished vertices,
1102: then $G$ has at least $|n - 3| + 3$ core vertices for $n = 0,1,2$
1103: and, for $n = 3,4,\ldots,$,
1104: exactly $|n - 3| + 3 = n$ core vertices.
1105: In other words, if we suppose that a graph has at least $n$
1106: distinguished vertices, then we have already established at least
1107: a lower bound on the number of vertices that the core of $G$
1108: may have. On the other hand, an assumption on the number of terminal
1109: cycles $G$
1110: restricts the possibilities for the number of
1111: distinguished vertices.
1112: \begin{Rk}
1113: A vertex of a simple partial 2-tree $G$ is a distinguished vertex
1114: if and only if $G$ has at least one cycle and the vertex belongs
1115: to each cycle of $G$.
1116: \end{Rk}
1117: \begin{proof} Since the core of $G$ is a nontrivial $LSEAC$ graph,
1118: and because $LSEAC$ graphs may be sequentially constructed
1119: by cycle articulation along edges, the statement immediately
1120: follows.
1121: \end{proof}
1122:
1123: \begin{La}
1124: If $G$ is an $LSEAC$ graph with at least one distinguished vertex
1125: $v$, then the induced subgraph $G - v$ is a path.
1126: \label{L5-10}
1127: \end{La}
1128: \begin{proof} This is a simple consequence of the fact that a
1129: distinguished vertex belongs to each cycle. For,
1130: if $G$ is a cycle, then this statement is trivial.
1131: If $G$ has two distinguished vertices, then $G$ must simply
1132: be a pair of cycles articulated along a single edge. In this case
1133: too it is clear that $G - v$ is a path. Finally, if $G$ has
1134: one distinguished vertex, then it follows that every edge of
1135: articulation intersects the distinguished vertex, and so only
1136: a path remains in the induced subgraph.
1137: \end{proof}
1138:
1139: \begin{La}
1140: Let $G$ be a simple partial 2-tree with at least one
1141: distinguished vertex with precisely
1142: two pendant neighbors. Then $M(G) = 2$ if and only
1143: if $G$ is a graph of two parallel paths.
1144: \label{L5-11}
1145: \end{La}
1146: \begin{proof}
1147: Let $u$ be a distinguished vertex of $G$ with precisely two
1148: pendant neighbors, $v_1, v_2$.
1149: The induced subgraph $G - \{u,v_1\}$ has two connected components:
1150: the isolated vertex $v_2$ and the graph $G - \{u,v_1,v_2\}$.
1151: By Lemma \ref{L4-2}, $M(G) > 2$, unless we have
1152: $M(G - \{u,v_1,v_2\}) = 1$, i.e., unless $G - \{u,v_1,v_2\}$ is
1153: a path. This occurs if and only if the pendant vertices of $G$
1154: added to the core of $G$ extend the path
1155: ${\rm Core}(G) - u$ to a path $P$ in $G$.
1156: In this situation,
1157: the path $P$ and the path connecting $v_1,u,v_2$ constitute
1158: a pair of two parallel paths, showing on account
1159: of Lemma \ref{L3-7} that $M(G) = 2$.
1160: \end{proof}
1161:
1162: \begin{La}
1163: Let $G$ be a simple partial 2-tree with a terminal cycle
1164: $Z$ that has two nondistinguished vertices $u_1,u_2$ satisfying the following
1165: conditions: if either vertex is of degree two relative to the core
1166: of $G$, then they are not neighbors; and, $u_1$ and $u_2$ each
1167: have at least one pendant neighbor, say $v_1$ and $v_2$,
1168: respectively. Then $M(G) > 2$.
1169: \label{L5-12}
1170: \end{La}
1171: \nline
1172:
1173: \noindent{\bf Remark.} Note that the conditions imposed on $G$ imply
1174: that the core of $G$ cannot be a cycle,
1175: since a cycle has no nondistinguished vertices.
1176: \nline
1177:
1178: \begin{proof}
1179: %The induced subgraph $G - \{u_1,v_1,u_2,v_2\} $
1180: %consists of two connected components.
1181: %If the two parallel
1182: %paths of the core of $G$
1183: %were drawn parallel with no crossing edges, a line
1184: %were drawn between $u_1$ and $u_2$, and the vertices $u_1,u_2,
1185: %v_1,v_2$ were removed, then one connected component would lie in one half plane
1186: %and the other connected component in the other half plane.
1187: %
1188: Indeed, by our choice of $u_1$ and $u_2$, there exists a path in
1189: $Z$ connecting $u_1$ and $u_2$ such that, with the possible
1190: exception of $u_1$ and $u_2$, all vertices in the path belong to
1191: cycle $Z$ only. This path is one of our connected components. Let
1192: us call it $P_1$. The remainder of the graph (excluding
1193: $u_1,u_2,v_1,v_2$) becomes the other connected component, which we
1194: shall call $P_2$. Since $P_2$ contains at least one cycle (e.g.,
1195: the other terminal cycle), it satisfies $M(P_2) > 1$. Hence
1196: \begin{equation*}
1197: M(G - \{u_1,v_1,u_2,v_2\}) = M(P_1) + M(P_2) > 2.
1198: \end{equation*}
1199: By Lemma \ref{L4-2}, it follows that $M(G) > 2$.
1200: \end{proof}
1201:
1202: \begin{La}
1203: Let $G$ be a simple partial 2-tree with a terminal
1204: cycle $Z$ that has a nondistinguished vertex $u$. Suppose
1205: that $u$ has at least two pendant neighbors, say $v_1$ and
1206: $v_2$. Then $M(G) > 2$.
1207: \label{L5-13}
1208: \end{La}
1209: \begin{proof} The induced subgraph $G - \{u,v_2\}$ consists of
1210: two connected components: the isolated vertex $v_1$ and the
1211: induced subgraph $G - \{u,v_1,v_2\}$. Since
1212: $G - \{u,v_1,v_2\}$ has at least one cycle, we see,
1213: by reasoning now entirely analogous to that used in
1214: Lemma \ref{L5-12}, that $M(G) > 2$.
1215: \end{proof}
1216:
1217: \begin{La}
1218: Let $G$ be graph whose core is a nontrivial $LSEAC$ graph.
1219: Moreover, suppose that if any distinguished vertex of $G$ has
1220: exactly one pendant neighbor, then there exists another
1221: distinguished vertex of $G$ with at least two pendant neighbors.
1222: Then $M(G) = 2$ if and only if $G$ is a graph of two parallel
1223: paths. \label{L5-14}
1224: \end{La}
1225: \begin{proof} Without loss of
1226: generality, we may assume that $G$ is a simple partial 2-tree
1227: (for if $G$ cannot be constructed from some simple partial 2-tree by
1228: a sequence of subdivisions of edges adjacent to pendant vertices,
1229: then $M(G) > 2$, and so by Lemma \ref{L3-7}, $G$ cannot
1230: be a graph of two parallel paths).
1231: By Lemma \ref{L5-5}, if some distinguished vertex of $G$ has
1232: more than two pendant neighbors, then $M(G) > 2$. By Lemma
1233: \ref{L3-7}, $G$ is not a graph of two parallel paths.
1234: If some distinguished vertex of $G$ has
1235: precisely two pendant neighbors, then the claim follows by Lemma
1236: \ref{L5-11}. If no distinguished vertex of $G$ has any
1237: pendant neighbors, then, for $M = 2$, it is necessary that the following be
1238: satisfied: pendant vertices must be adjacent to vertices of
1239: terminal cycles, by Lemma \ref{L5-7}; no vertex of a terminal cycle
1240: may have more than one pendant neighbor, by Lemma \ref{L5-13}; and
1241: no terminal cycle may have more than two nondistinguished vertices
1242: each with at least one pendant neighbor, and moreover,
1243: if a terminal cycle has two such nondistinguished vertices, then those
1244: two vertices
1245: in the terminal cycle are neighbors (Lemma \ref{L5-12}). When
1246: these conditions are satisfied, however, it follows that there
1247: exist disjoint induced paths $P_1$ and $P_2$ in the core of $G$,
1248: covering all vertices in the core, such that the pendant
1249: vertices of $G$ extend paths $P_1$ and $P_2$. Hence it is necessary
1250: that $G$ be a graph of two parallel paths. The sufficiency of
1251: being a graph of two parallel paths has already been shown.
1252: \end{proof}
1253:
1254: The following result is a useful criterion for ruling out graphs
1255: for which $M > 2$.
1256: \begin{La}
1257: Let $G$ be a (general) connected graph. If $G$ has more
1258: than five pendant vertices, then $M(G) > 2$.
1259: \label{L5-15}
1260: \end{La}
1261: \begin{proof}
1262: Without loss of generality, we may assume that $G$ is a simple
1263: partial 2-tree.
1264: If $G$ is a tree with more than five pendant vertices,
1265: then $G$ has path covering number at least three, and so $M(G) \geq 3$.
1266: It therefore
1267: suffices to consider the case where the core of $G$ is a
1268: simple partial 2-tree with at least six pendant vertices.
1269: If no distinguished vertex of $G$ has precisely one pendant neighbor,
1270: then by Lemma \ref{L5-14}, $M(G) > 2$ since clearly $G$ cannot
1271: be a graph of two parallel paths. We may therefore restrict ourselves
1272: to the case where some distinguished vertex of $G$, say $u$, has
1273: precisely one pendant neighbor, say $v$. Since $u$ belongs
1274: to each cycle of $G$, $G - \{u,v\}$ is a tree.
1275: By our assumptions, $G - \{u,v\}$ has at least five
1276: pendant vertices, meaning now that it is a tree with
1277: path covering number at least
1278: three, implying that $M(G - \{u,v\}) > 2$. Invoking Lemma
1279: \ref{L4-2} one more time, we see that $M(G) > 2$.
1280: \end{proof}
1281:
1282: It is a direct consequence of the lemmas we have proven thus far
1283: that in order
1284: for a graph $G$ to be ``exceptional'', i.e., have $M(G) = 2$
1285: and yet not be a graph of two parallel paths, $G$ must also satisfy
1286: the following:
1287: \begin{enumerate}
1288: \item \text{$G$ is a simple partial 2-tree.}
1289: \item \text{$G$ has at least one distinguished vertex.}
1290: \item \text{At least one distinguished vertex has exactly one
1291: pendant neighbor.}
1292: \item \text{No vertex has more than one pendant neighbor.}
1293: \item \text{$G$ has no more than five pendant vertices.}
1294: \end{enumerate}
1295: That each graph belonging to one of the classes of graphs listed in
1296: the table is exceptional may be verified readily: inductively
1297: use the pendant vertex lemma on the path extending from a
1298: distinguished vertex, considering both the case
1299: where the distinguished vertex remains but without the path
1300: attached to it, and the case where the path and the distinguished vertex
1301: are removed. In each case, the induced subgraph has $M = 2$, and
1302: therefore, by Lemma \ref{L4-2}, so does the original graph.
1303: Clearly, no graph in the table is a graph of two parallel paths.
1304: This is obvious when the number of pendant vertices is five. When
1305: there are fewer pendant vertices, a small, finite number of subcases
1306: may be considered showing that no graph may be covered by two
1307: disjoint induced paths satisfying the edge crossing condition.
1308:
1309: In the subsequent analysis we shall show that these are the
1310: only exceptional graphs. In proving the results that follow we
1311: shall frequently make use of the fact that the five conditions
1312: listed above are necessary for a graph to be exceptional. In other
1313: words, it suffices to consider only the cases where
1314: the above conditions hold.
1315:
1316: \begin{La}
1317: If a graph $G$ is exceptional, then $G$ has more than two pendant
1318: vertices.
1319: \end{La}
1320: \begin{proof}
1321: Let $G$ be a simple partial 2-tree with a distinguished vertex
1322: $u_1$ that has precisely one pendant neighbor, say $v_1$.
1323: Let
1324: $u_2$ be a vertex, different from $u_1$, that belongs to the core of
1325: $G$, and let $v_2$ be a pendant neighbor of $u_2$. Suppose that
1326: $v_1$ and $v_2$ are the only degree-one vertices of $G$. The vertex
1327: $u_2$ belongs to a terminal cycle, for otherwise $M(G) > 2$. In this
1328: case, $u_2$ has a neighbor $w$ (possibly equal to $u_1$)
1329: in the same terminal cycle such that there exists a minimal path
1330: $P_1$ connecting $w$ and $v_1$ such that $G - P_1$ has only
1331: one connected component. In this case, though, $G - P_1$
1332: is a path, which we denote by $P_2$. By the construction,
1333: $P_1$ and $P_2$ constitute a pair of parallel paths, showing
1334: that $M(G) = 2$ and $G$ is a graph of two parallel paths.
1335: We have
1336: thus shown that a simple partial 2-tree $G$ with two pendant
1337: vertices has $M(G) = 2$ if and only if $G$ is a graph of two parallel
1338: paths. This completes the proof.
1339: \end{proof}
1340:
1341: \begin{La}
1342: If a graph $G$ is exceptional, then $G$ has no more than three
1343: cycles.
1344: \end{La}
1345: \begin{proof}
1346: Let $G$ be a simple partial 2-tree whose core has at
1347: least four cycles and such that no vertex in the core of $G$
1348: has more than one pendant neighbor.
1349: Since $G$ has at least four cycles, the cycles determine a
1350: unique distinguished vertex,
1351: which we shall call $u$. Suppose $u$ has a unique pendant
1352: neighbor $v$, and that the graph $G - v$ has either $2$, $3$,
1353: or $4$ pendant vertices and is a graph of two parallel paths.
1354: Note that if $G - v$ is not a graph of two parallel paths,
1355: then, then by Lemma \ref{L5-14}, $M(G) > 2$.
1356: If $G$ is a graph of
1357: two parallel paths, then $G$ is not exceptional, and so we suppose
1358: that $G$ is not a graph of two parallel paths.
1359: In particular, this
1360: implies that in each terminal cycle there is at least one vertex
1361: different from $u$ that has a pendant neighbor. It also implies
1362: that, if in
1363: terminal cycle $Z$ the neighbor of $u$ that belongs only in cycle
1364: $Z$ supports a pendant vertex, then there are at least
1365: three vertices in $Z$ (including $u$) that have pendant neighbors.
1366: Now consider
1367: the induced subgraph $G - \{u,v\}$. Since $u$ is distinguished,
1368: the induced subgraph is a tree. Since $G$ has at least four cycles,
1369: no path shorter than two edges
1370: in length exists between vertices of opposite terminal cycles.
1371: On the other hand, at least
1372: two vertices of each terminal cycle are pendant in the graph
1373: $G - \{u,v\}$ by the conditions we have already established
1374: on pendant neighbors. Together, these statements
1375: imply that $G - \{u,v\}$ is a tree of path covering number three.
1376: By Lemma \ref{L4-2}, $M(G) > 2$ and so, in particular,
1377: $G$ is not exceptional.
1378: \end{proof}
1379:
1380: \begin{La}
1381: A graph $G$ whose core consists of three cycles is exceptional
1382: if and only if $G$ belongs to the collection of graphs
1383: given in Table B1 (see Appendix B).
1384: \end{La}
1385: \begin{proof}
1386: Let $G$ be a simple partial 2-tree whose core consists of
1387: three cycles and let us assume that $G$ has at least three pendant
1388: neighbors. We assume that only vertices of terminal cycles may have
1389: pendant neighbors.
1390: The graph $G$ has a unique distinguished vertex
1391: $u$. In order for $G$ to be exceptional, $u$ must have precisely
1392: one pendant neighbor, say $v$. In order for $M(G) = 2$,
1393: $G - \{u,v\}$ must be a tree of path covering number two (if it
1394: were one, then $G$ would be a graph of two parallel paths). Let
1395: $Z$ be the unique cycle of $G$ that is not a terminal cycle. If
1396: $Z$ contains more than three vertices, then $M(G - \{u,v\}) > 2$
1397: unless $G$ is a graph of two parallel paths. Henceforth assume
1398: that $Z$ has only three vertices. In this case, though, by
1399: considering the graph $G - \{u,v\}$, we see that, in order for
1400: $G$ to be exceptional, each vertex of $Z$ must have precisely
1401: one pendant neighbor.
1402:
1403: Suppose we have such a graph $G$, but with the additional
1404: requirement that only the vertices of $Z$ have pendant neighbors.
1405: Then it is easy to check that such a graph $G$ is exceptional and
1406: is in Table B1. Suppose we wish to add a pendant vertex to $G$ so
1407: that the resulting graph is still exceptional. Necessarily, we
1408: must add a new vertex to a terminal cycle, say $Z_t$. Considering
1409: $G - \{u,v\}$ shows that it is also necessary that $Z_t$ have only
1410: three vertices. Considering $G - v$ shows that this is
1411: sufficient. Such graphs are also in our table. To add a fifth
1412: pendant vertex, necessarily each terminal cycle has only three
1413: vertices, and in that case, there is a unique way pendant vertices
1414: may be arranged. This case is also exceptional and is covered in
1415: Table B1. Since this analysis exhausts all possibilities, the
1416: proof is complete.
1417: \end{proof}
1418:
1419: \begin{La}
1420: A graph $G$ whose core consists of two cycles is exceptional
1421: if and only if $G$ belongs to the collection of graphs given
1422: in Table B1.
1423: \end{La}
1424: \begin{proof}
1425: Let $G$ be a simple partial 2-tree whose core consists of
1426: two cycles and let us assume that $G$ has at least three pendant
1427: neighbors, and that no vertex of $G$ has multiple pendant neighbors.
1428: We first claim that $G$
1429: cannot be exceptional unless each distinguished vertex,
1430: $u_1,u_2$, has precisely
1431: one pendant neighbor.
1432: For suppose that $u_1$ has a unique pendant neighbor $v_1$, but
1433: that $u_2$ does not have a pendant neighbor. If $G - \{u_1,v_1\}$
1434: is a path, then $G$ is a graph of two parallel paths, and so
1435: suppose that $G - \{u_1,v_1\}$ is a tree of two parallel paths.
1436: Then at least one cycle has at most one nondistinguished vertex
1437: with a pendant neighbor. However, in this event $G$ must be a graph
1438: of two parallel paths
1439:
1440: We therefore may assume that $u_1$ has a pendant neighbor $v_1$ and
1441: $u_2$ has a pendant neighbor $v_2$.
1442: If $M(G) = 2$, then we must have $M(G - \{v_1,v_2\}) = 2$, which
1443: happens only when this graph is a graph of two parallel paths.
1444: In this case we see that if a cycle of $G - \{v_1,v_2\}$ has
1445: two vertices each with a pendant neighbor, then these two
1446: vertices must be adjacent. If no cycle of this graph has
1447: two vertices, then, by considering $G - \{u_1,v_1\}$, we
1448: see that nondistinguished vertices with pendant neighbors
1449: must be distributed so that one is adjacent to $u_1$ and
1450: the other to $u_2$ (or perhaps both). In this case, however,
1451: $G$ is a graph of two parallel paths.
1452:
1453: Thus we assume that there is a cycle $Z$ of $G$ that has two
1454: nondistinguished adjacent vertices, each with a pendant
1455: neighbor. By considering the trees $G - \{u_1,v_1\}$
1456: and $G - \{u_2,v_2\}$, we see that $Z$ must have only
1457: four vertices, for otherwiswe $M(G) > 2$. Let $G$ be a graph
1458: satisfying all of these requirements. Suppose in addition
1459: that $G$ has only four pendant vertices. Then $G$ is an
1460: exceptional graph and is in Table B1. Suppose we wish to
1461: add a fifth pendant vertex to $G$. This vertex must
1462: necessarily be added to a vertex in $G - Z$. Considering
1463: $G - \{u_1,v_1\}$ and $G - \{u_2,v_2\}$ shows that this
1464: may only be done if the cycle in $G$ that is not $Z$
1465: is a cycle on three vertices. The resulting graph is
1466: unique, exceptional, and in our table. Since we have
1467: exhausted all possibilities, the proof is complete.
1468: \end{proof}
1469:
1470: \begin{La}
1471: A graph $G$ whose core is a single cycle is exceptional
1472: if and only if $G$ belongs to the
1473: collection of graphs given in Table B1.
1474: \end{La}
1475: \begin{proof}
1476: Let $G$ be a simple partial 2-tree with between three and five
1477: pendant vertices and such that no vertex has more than one pendant
1478: neighbor. If the core of $G$ has only three or four vertices,
1479: then clearly $G$ is a graph of two parallel paths. If the core of
1480: $G$ has five vertices, then $G$ is a graph of two parallel paths
1481: unless $G$ has five pendant vertices. If $G$ has five vertices in
1482: its core and five pendant vertices arranged as specified above,
1483: then it is exceptional and is in our table. Now suppose that the
1484: core of $G$ has more than five vertices. If $G$ has at most two
1485: pendant vertices, then $G$ is a graph of two parallel paths. Note
1486: that, since $G$ has more than five vertices, there exists a set
1487: (in general, many) of three vertices of the core of $G$, say
1488: $u_1$, $u_2$, $u_3$, such that the induced subgraph $G -
1489: \{u_1,u_2,u_3\}$ has three connected components. If $u_1$, $u_2$,
1490: and $u_3$ all have a pendant neighbor, then, by Lemma \ref{L4-2}
1491: applied three times, $M(G) > 2$. If three such vertices, each
1492: with a pendant neighbor, cannot be found, then $G$ has at most
1493: four pendant vertices and is a graph of two parallel paths. This
1494: exhausts all possibilities and so completes the proof.
1495: \end{proof}
1496:
1497: \noindent{\bf Remark.} Let $G$ be as in the previous lemma, but with
1498: the additional assumption that the core of $G$ has at least
1499: four vertices. Suppose $G$ is a graph
1500: of two parallel paths. If $G$ has three pendant vertices, then
1501: there is at least one pair of pendant vertices that are such that
1502: their neighbors are adjacent. If $G$ has four pendant
1503: vertices, then we may partition the pendant
1504: vertices into two such pairs.
1505: \nline
1506:
1507: \begin{La}
1508: Let $G$ be a simple partial 2-tree with a distinguished
1509: vertex $u$ that has precisely one pendant neighbor, say $v$.
1510: Then $M(G) = 2$ if and only if $G - v$ is a graph of two parallel
1511: paths and $G - \{u,v\}$ is either a graph of two parallel paths
1512: or a path.
1513: \end{La}
1514: \begin{proof}
1515: It is clear that if $G - v$ is a graph of two parallel paths and
1516: $G - \{u,v\}$ is either a graph of two parallel paths or a path,
1517: then $M(G) = 2$. If $M(G) = 2$ and $G$ is a graph of
1518: two parallel paths, then the implication is trivial. If $G$
1519: is exceptional, then $G$ is included in our table. It may be
1520: easily checked that the claim holds for each of these graphs.
1521: \end{proof}
1522: \begin{Rk}
1523: Our main result, Theorem \ref{T5-1}, holds in fact over any
1524: infinite field. This has been explained in detail in the proof of
1525: Lemma \ref{L2-3}, and to a lesser degree in the proof of Lemma
1526: \ref{L4-3}. Many of our lemmas generalize in a straightforward way
1527: from $\mathbb{R}$ to any infinite field (and sometimes to any
1528: field F).
1529: \end {Rk}
1530: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%THE APPENDIX A
1531: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1532: \newtheorem{Ob}[Pa]{{\bf Observation}}
1533: \newenvironment{Rem}{\par\noindent{\bf Remark:}\hspace{5mm}}{.\vspace{2mm}\\}
1534: %\newtheorem*{Rem}{{\bf Remark}}
1535: \newcommand {\wT}[1]{\tilde{a}_{#1}}
1536: \newcommand {\hsp}{\hspace{5mm}}
1537: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1538: \newpage
1539: \appendix
1540: %%%%% appendix title trick
1541: \renewcommand {\thesection} {Appendix A}
1542: \section {}
1543: \renewcommand {\thesection} {A}
1544: %%%%% end of appendix title
1545: \indent
1546:
1547: The purpose of the appendix is to prove Lemma \ref{L2-3}. In fact,
1548: we prove a more general result, by replacing $\mathbb{R}$ by any
1549: infinite field $F$. The proof given has an algebraic-combinatorial
1550: flavor.
1551:
1552: We start with a few preliminaries. Let $S(F,G)$ denote the set of
1553: all $n\times n$ symmetric matrices with entries in $F$, and whose
1554: graph is $G$. We use $E$ to denote the set of edges of the graph $G$. Let
1555: \begin{equation*}
1556: \text{msr}(F,G)=\text{min rank}\,A ,
1557: \end{equation*}
1558: where $A$ ranges over all matrices in $S(F,G)$. Let
1559: \begin{equation*}
1560: \text{cork} (F,G)=n-\text{msr}(F,G) .
1561: \end{equation*}
1562: \begin{Rem} We clearly have cork$(\mathbb{R},G)=M(G)$ \end{Rem}
1563:
1564: Let $H$ be any subgraph of $G$ which is an $hK_4$. We call the
1565: original vertices of $K_4$, used to obtain $H$ from $K_4$ by a
1566: sequence of edge subdivisions, {\em initial vertices}. All other
1567: vertices of $H$ are called {\em intermediate vertices}.\nline
1568:
1569: The following well known result is used in the Appendix:
1570: \begin{Ob}
1571: Let $F$ be an infinite field and let $f\in F[t_1,t_2,\cdots,t_n]$.
1572: Then there exist $a_1,a_2,\cdots,a_n\in F^*$ such that
1573: $f(a_1,a_2,\cdots,a_n)\in F^* .$
1574: \end{Ob}
1575:
1576: The following discussion plays an important role in the proof of
1577: the main result of the Appendix.
1578:
1579: Let $l$ be an integer such that $3\leq l\leq n-1$. Let $G'$ be the
1580: induced subgraph on $\{l+1,l+2,\cdots,n\}$. Let $B$ be an
1581: $n-l\times n-l$ symmetric matrix defined as follows: its rows and
1582: columns are labeled from $l+1$ to $n$ ; its main diagonal entries
1583: are all zero. For $i,j \in \{l+1,l+2,\cdots,n\}, i<j$, we let
1584: $b_{ji}=b_{ij}=0$ if $ij \notin E$ and we let $b_{ij}$ be an
1585: indeterminate if $ij \in E$ (and let $b_{ji}=b_{ij}$). So, for
1586: example, if $l+1,l+2\in E$ then the entry in the first row and
1587: second column of B is $b_{l+1,l+2}$. We let $\underline{b}$ denote
1588: the set of all indeterminates in $B_{22}$.
1589:
1590: \noindent Let
1591: \begin{equation*} A_{22}=I-\mu B_{22} ,\end{equation*}
1592: where $\mu$ is an indeterminate. Let $A_{12}=(a_{ij}) ,\hsp
1593: i=1,2,\cdots,l \hsp j=l+1,l+2,\cdots,n$ be such that the $ij$-th
1594: entry is 0 if $ij \notin E$ and an indeterminate $a_{ij}$ if $ij
1595: \in E$.
1596:
1597: Let $R$ be the polynomial ring consisting of all polynomials in
1598: the indeterminates that appear in $A_{12}$ and $A_{22}$, and with
1599: coefficients in $F$. Let $K$ be the quotient field of $R$. Let
1600: \begin{equation*} d=d(\mu,\underline{b})=\text{det} A_{22} ,\end{equation*}
1601: \begin{equation*} W=W(\mu,\underline{b})=\text{adj} A_{22} ,\end{equation*}
1602: and
1603: \begin{equation*} Z=Z(\mu,\underline{b})=A_{22}^{-1} .\end{equation*}
1604: Note that $d \in K^*$, because the constant term in its expansion
1605: is 1. Hence $A_{22}$ is an invertible element of $K^{n-l,n-l}$ ,
1606: so $Z$ exists and we have
1607: \begin{equation*} W=dZ\in K^{n-l,n-l} .\end{equation*}
1608:
1609: \begin{Rem}
1610: As for $B_{22}$, the rows and columns of $A_{22}, W$ and $Z$ are
1611: labeled $l+1$ to $n$
1612: \end{Rem}
1613: \noindent Define
1614: \begin{equation*} C=A_{12}A_{22}^{-1}A_{12}^t=A_{12}ZA_{12}^t \in K^{l,l} ,\end{equation*}
1615: and let q be a positive integer, to be determined later. We have
1616: \begin{equation*} (I-\mu B_{22})(I+\mu B_{22}+\mu ^2B_{22}^2+\cdots +\mu ^qB_{22}^q)=I-\mu^{q+1}B_{22}^{q+1} , \end{equation*}
1617: hence
1618: \begin{equation*} I+\mu B_{22}+\mu ^2B_{22}^2+\cdots +\mu ^qB_{22}^q=Z-\mu ^{q+1}ZB_{22}^{q+1} . \end{equation*}
1619: Since $Z=d^{-1}W$, we get
1620: \begin{equation} Z=d^{-1}[d(I+\mu B_{22}+\mu ^2B_{22}^2+\cdots +\mu ^qB_{22}^q)+\mu ^{q+1}WB_{22}^{q+1}] , \label{A-1} \end{equation}
1621: and
1622: \begin{equation} W=d(I+\mu B_{22}+\mu ^2B_{22}^2+\cdots +\mu ^qB_{22}^q)+\mu ^{q+1}WB_{22}^{q+1} . \label{A-2} \end{equation}
1623: \\
1624: \begin{La}\label{LA-2}
1625: Let $i,j\in \{1,2,...,l\}$ with $i\neq j$. Suppose that there
1626: exists a path in $G \hsp ii_1,i_1i_2,...,i_ri_{r+1},i_{r+1}j$ such
1627: that $i_1,i_2,...,i_r,i_{r+1}\notin \{1,2,...,l\}$. Then
1628: $c_{ij}\in K^*$.
1629: \end{La}
1630: \begin {proof}
1631: We show first that $z_{i_1,i_{r+1}}\neq 0$. It suffices to show
1632: $w_{i_1,i_{r+1}}\neq 0$. Indeed, in computing $w_{i_1,i_{r+1}}$
1633: the expression $\mu^r\displaystyle\prod_{k=1}^r b_{i_k,i_{k+1}}$
1634: appears as one of the summands of $(\mu^rB_{22}^r)_{i_1,i_{r+1}}$.
1635: This term cannot appear in any matrix that comes after $\mu^r
1636: B_{22}^r$ in (\ref{A-2}) (we assume $q$ large enough, say
1637: $q>n-l$). It cannot come from preceding matrices either. Indeed,
1638: if it would, some contribution $\neq 1$ must come from $d$. But in
1639: every summand $\neq 1$ in the expression of d the indeterminates
1640: from $\underline{b}$ that appear are $a$ disjoint union of several
1641: cycles, at least one not the identity, and this is impossible
1642: here. We notice now that one of the terms in $dc_{ij}$ is
1643: $a_{i,i_1}a_{i_{r+1},j}\mu^r\displaystyle\prod_{k=1}^r
1644: b_{i_k,i_{k+1}}$ , and it cannot be canceled. Hence $dc_{ij} \in
1645: K^*$, implying that $c_{ij} \in K^*$.
1646: \end {proof}
1647:
1648: The main result of the appendix is:
1649: \begin {La} Let $F$ be an infinite field and let $G$ be a graph on
1650: $n \geq 4$ vertices, which is not a partial 2-tree. Then
1651: \begin{equation*} \text{cork}(F,G)\geq3 ,\end{equation*}
1652: \end {La}
1653: \begin {proof}
1654: By Lemma \ref {L2-2} G contains (as a subgraph) an $hK_4$ , which
1655: is denoted by $H$, and its initial vertices are denoted by
1656: $1,2,3,n$. We distinguish four cases :
1657:
1658: \Cs There are three initial vertices, say $1,2,3$, such that the
1659: three paths in $H$ contain
1660: no intermediate vertices.\\
1661: \centerline{\includegraphics[width=70mm]{mypics/1}}
1662: \begin{Rem}
1663: Here, and in subsequent figures,\space\space{\Large$\bullet\;$}
1664: denotes a vertex that might or might not be present in $H$
1665: \end{Rem}
1666: \noindent We apply Lemma \ref{LA-2} with $l=3$. Since F is
1667: infinite we can clearly assign nonzero values to all
1668: indeterminates in $K$ so that all off-diagonal entries of
1669: $C=A_{12}A_{22}^{-1}A_{12}^t$ are nonzero. Then the matrix
1670: $\begin{bmatrix}
1671: C & A_{12}\\
1672: A_{12}^t & A_{22}\\
1673: \end{bmatrix}$
1674: is in $S(F,G)$ and has rank $n-3$.%
1675: \Cs There are three initial vertices, say $1,2,3$, such that
1676: exactly one of the three corresponding paths in $H$ contains an
1677: intermediate vertex.\\
1678: \centerline{\includegraphics[width=70mm]{mypics/2}}\\
1679: We assume that 3 and 4 are adjacent, but 2 and 4 \underline{don't
1680: have} to be adjacent.\\
1681: If $24\in E$ we may assume it is also an edge of $H$, for if it
1682: isn't, we replace the path from 2 to 4 in $H$ by the edge $24 $.
1683: Given $i,j \in \{1,2,3,4\}$ such that $i\neq j$, we say there is
1684: an {\em external path} from $i$ to $j$ in $G$ if there is a
1685: (simple) path from $i$ to $j$, which is not an edge, and such that
1686: no intermediate vertices of this path belong to $\{1,2,3,4\}$. In
1687: our case there are external paths from 1 to 2; from 1 to 3 ; from
1688: 2 to 3 ; from 2 to 4 if they are not adjacent.\\
1689: We want to use Lemma \ref{LA-2} with $l=4$. Let
1690: \begin{equation*}
1691: A_{11}=A[1,2,3,4]= \begin{bmatrix}
1692: ? & ? & ? & a_{14}\\
1693: ? & ? & 0 & a_{24}\\
1694: ? & 0 & ? & ?\\
1695: a_{14} & a_{24} & ? & ?\\
1696: \end{bmatrix},
1697: \end {equation*}
1698: where we use the following notation: '?' means an indeterminate
1699: (which is NOT one of those that appear in the definition of K);
1700: $a_{14}=0$ if $ 14 \notin E$ , and is chosen in $F^*$ if $14 \in
1701: E$ ; put $a_{24}=0$ if $24\notin E$ and a '?' if $24 \in E$.
1702: \begin{Rem}
1703: We have $a_{23}=0$ because we may assume $23 \notin E $ (or else
1704: apply case 1)
1705: \end{Rem}\noindent
1706: We assign nonzero values to all indeterminates in $K$ so that
1707: $A_{22}$ is nonsingular and every $c_{ij}\in K^*$ becomes an
1708: element of $F^*$. Now, the Schur complement with respect to
1709: $A_{22}$ is
1710: \begin{equation*}
1711: A_{11}-C=A_{11}-A_{12}A_{22}^{-1}A_{12}^t\;.
1712: \end{equation*}
1713: Also, if $a_{24}=?$, we choose it so that the element in the $2,4$
1714: position of $A_{11}-C$ is nonzero. So $A_{11}-C$ has the form:
1715:
1716:
1717: \begin{equation*}
1718: \begin{bmatrix}
1719: ? & ? & ? & \wT{{14}}\\
1720: ? & ? & \wT{{23}} & \wT{{24}}\\
1721: ? & \wT{{23}} & ? & ?\\
1722: \wT{{14}} & \wT{{24}} & ? & ?\\
1723: \end{bmatrix},
1724: \end{equation*}
1725: where we know $\wT{{23}} , \wT{{24}}\in F^*$. Now let
1726: $x^t=(\dfrac{\wT{{14}}}{\wT{{24}}},1,\wT{{23}},\wT{{24}})$. Then
1727: \begin{equation*}
1728: xx^t=
1729: \begin{bmatrix}
1730: \vspace{2mm} \frac{\wT{{14}}^2}{\wT{{24}}^2} &
1731: \frac{\wT{{14}}}{\wT{{24}}} & \frac{\wT{{14}}\wT{{23}}}{\wT{{24}}}
1732: & \wT{{14}}\\ \vspace{2mm} \frac{\wT{{14}}}{\wT{{24}}} & 1 &
1733: \wT{{23}} & \wT{{24}}\\ \vspace{2mm}
1734: \frac{\wT{{14}}\wT{{23}}}{\wT{{24}}} & \wT{{23}} & \wT{{23}}^2 &
1735: \wT{{23}}\wT{{24}}\\
1736: \wT{{14}} & \wT{{24}} & \wT{{23}}\wT{{24}} & \wT{{24}}^2\\
1737: \end{bmatrix},
1738: \end{equation*}
1739: and it has the desired form. So we can choose $A\in S(F,G)$ with
1740: $\text{rank} A =n-4+1=n-3$.%
1741: \Cs There are three initial vertices, say $1,2,3$, such that
1742: exactly two of the three corresponding paths in $H$ contain an
1743: intermediate vertex.\\
1744: \centerline{\includegraphics[width=70mm]{mypics/3}}\\
1745: Here we assume that $34\in E, 35 \in E$ but $24$ and $15$ don't
1746: have to be edges of G. As in Case 2, if $24\in E$ we may assume it
1747: is an edge of $H$. A similar statement for $15$. We use the notion
1748: external path as in the previous case, so, for example, there is
1749: an external path from 1
1750: to 2. If $24 \notin E$, there is an external path from 2 to 4.\\
1751: We want to use Lemma \ref{LA-2} with $l=5$. Let
1752: \begin{equation*}
1753: A_{11}=A[1,2,3,4,5]= \begin{bmatrix}
1754: ? & ? & 0 & a_{14} & a_{15}\\
1755: ? & ? & 0 & a_{24} & a_{25}\\
1756: 0 & 0 & ? & ? & ?\\
1757: a_{14} & a_{24} & ? & ? & a_{45}\\
1758: a_{15} & a_{25} & ? & a_{45} & ?\\
1759: \end{bmatrix}\, ,
1760: \end {equation*}
1761: where '?' is used as before; $a_{14}=0$ if $14 \notin E$ and
1762: $a_{14}\in F^*$ if $14 \in E$; put $a_{15}=0$ if $15\notin E$ and
1763: a '?' if $15 \in E$; put $a_{24}=0$ if $24\notin E$ and a '?' if
1764: $24 \in E$ ; $a_{25}=0$ if $25 \notin E$ and $a_{25}\in F^*$ if
1765: $25 \in E$; $a_{45}=0$ if $45\notin E$ and $a_{45}\in F^*$ if $45
1766: \in E$. Note that we may assume $13 \notin E$ and $23 \notin E$ ,
1767: or else we can apply a previous case. Assign nonzero values to all
1768: indeterminates in K so that $A_{22}$ is invertible and every
1769: $c_{ij}\in K^*$ becomes an element of $F^*$. Now, the Schur
1770: complement with respect to $A_{22}$ is
1771: \begin{equation*}
1772: A_{11}-C=A_{11}-A_{12}A_{22}^{-1}A_{12}^t\;.
1773: \end{equation*}
1774: Also, if $a_{15}=?$ (resp. $a_{24}=?$), we assign it a value so
1775: that $1,5$ entry (resp. $2,4$ entry) of the Schur complement is
1776: nonzero. Hence, $A_{11}-C$ has the form:
1777:
1778: \begin{equation*}
1779: \begin{bmatrix}
1780: ? & ? & \wT{{13}} & \wT{{14}} & \wT{{15}}\\
1781: ? & ? & \wT{{23}} & \wT{{24}} & \wT{{25}}\\
1782: \wT{{13}} & \wT{{23}} & ? & ? & ?\\
1783: \wT{{14}} & \wT{{24}} & ? & ? & \wT{{45}}\\
1784: \wT{{15}} & \wT{{25}} & ? & \wT{{45}} & ?\\
1785: \end{bmatrix}\, ,
1786: \end {equation*}
1787: where $\wT{{13}},\wT{{23}},\wT{{15}},\wT{{24}}\in F^*$ . One needs
1788: to assign values to the indeterminates so that
1789: $A_{11}-C$ has rank two. \\
1790: The $2 \times 2$ principal submatrix in the top left corner is
1791: arbitrary, so we can assume it is invertible and its inverse is
1792: also an arbitrary invertible matrix. Denote it by$\begin{bmatrix}
1793: u & v\\
1794: v & w\\
1795: \end{bmatrix}$. Let $B$ denote its Schur complement. We need
1796: $B=0$. Because of the form of $(A_{11}-C)[3,4,5]$ , we get
1797: effectively only one equation, namely:
1798: \begin{equation*}
1799: 0=b_{23}=\wT{{45}}-[\wT{{14}}\;\wT{{24}}]%
1800: \begin{bmatrix}
1801: u & v\\
1802: v & w\\
1803: \end{bmatrix}
1804: \begin{bmatrix}
1805: \wT{{15}}\\
1806: \wT{{25}}\\
1807: \end{bmatrix}
1808: =\wT{{45}}-\wT{{14}}\wT{{15}}u-(\wT{{14}}\wT{{25}}+\wT{{15}}\wT{{24}})v-\wT{{24}}\wT{{25}}w\;
1809: .
1810: \end{equation*}
1811:
1812: Suppose first that $\wT{{14}}\wT{{25}}+\wT{{15}}\wT{{24}}\neq 0$.
1813: If $\wT{{45}} \neq 0$ we have a solution with $u=w=0$ and $v \neq
1814: 0$ . So suppose $\wT{{45}}=0$. If $\wT{{14}}=\wT{{25}}=0$ we have
1815: a solution with $v=0$, and $u \neq0 , w \neq 0$. If exactly one of
1816: $\wT{{14}},\wT{{25}}$ is zero , we have a solution with $v \neq 0$
1817: and exactly one of $u,v$ is zero. So we may assume now
1818: $\wT{{14}}\wT{{25}}+\wT{{15}}\wT{{24}}=0$, implying
1819: $\wT{{14}}\wT{{25}}=-\wT{{15}}\wT{{24}}\neq 0\, .$
1820: %So we have a
1821: %solution with $w=0$, and appropriately chosen $u \in F^*,v
1822: %\in F^*$. So we choose $A \in S(F,G)$ with $\text{rank} A=n-5+2=n-3$%
1823: It is clear that we can find $u,v,w \in F$ such that
1824: $\begin{bmatrix}
1825: u & v\\
1826: v & w\\
1827: \end{bmatrix}$ is invertible. We conclude that there exists $A \in S(F,G)$ with $\text{rank} A=n-5+2=n-3 .$
1828: \Cs The remaining case to consider is of the form:\\
1829: \centerline{\includegraphics[width=70mm]{mypics/4}}\\
1830: where $34 \in E,35 \in E ,16 \in E$ , while $24,26$ and $15$ don't
1831: have to be edges. As in previous cases, if any of $24,26$ and $15$
1832: are in $E$ we may assume that they are in $H$. The discussion is
1833: similiar to the previous cases and leads to
1834: \begin{equation*}
1835: A_{11}-C=
1836: \begin{bmatrix}
1837: ? & \wT{{12}} & \wT{{13}} & \wT{{14}} & \wT{{15}} & ?\\
1838: \wT{{12}} & ? & \wT{{23}} & \wT{{24}} & \wT{{25}} & \wT{{26}}\\
1839: \wT{{13}} & \wT{{23}} & ? & ? & ? & \wT{{36}}\\
1840: \wT{{14}} & \wT{{24}} & ? & ? & \wT{{45}} & \wT{{46}}\\
1841: \wT{{15}} & \wT{{25}} & ? & \wT{{45}} & ? & \wT{{56}}\\
1842: ? & \wT{{26}} & \wT{{36}} & \wT{{46}} & \wT{{56}} & ?\\
1843: \end{bmatrix}\, ,
1844: \end{equation*}
1845: where
1846: $\wT{{12}},\wT{{13}},\wT{{15}},\wT{{23}},\wT{{24}},\wT{{26}}\in
1847: F^*$. One needs to assign values to the indeterminates so that
1848: $A_{11}-C$ has rank 3. We compute the Schur complement $B$ with
1849: respect to the $3 \times 3$ principal submatrix of $A_{11}-C$
1850: based on raws $1,2,6$. We write this principal submatrix as
1851: \begin{equation*}G=
1852: \begin{bmatrix}
1853: x & \wT{{12}} & u\\
1854: \wT{{12}} & y & \wT{{26}}\\
1855: u & \wT{{26}} & z\\
1856: \end{bmatrix}\, ,
1857: \end{equation*}
1858: where $x,y,u,z$ are indeterminates. So
1859: \begin{equation*}
1860: B=
1861: \begin{bmatrix}
1862: ? & ? & ?\\
1863: ? & ? & \wT{{45}}\\
1864: ? & \wT{{45}} & ?\\
1865: \end{bmatrix}-
1866: \begin{bmatrix}
1867: \wT{{13}} & \wT{{23}} & \wT{{36}}\\
1868: \wT{{14}} & \wT{{24}} & \wT{{46}}\\
1869: \wT{{15}} & \wT{{25}} & \wT{{56}}\\
1870: \end{bmatrix}G^{-1}
1871: \begin{bmatrix}
1872: \wT{{13}} & \wT{{14}} & \wT{{15}}\\
1873: \wT{{23}} & \wT{{24}} & \wT{{25}}\\
1874: \wT{{36}} & \wT{{46}} & \wT{{56}}\\
1875: \end{bmatrix}=0 .
1876: \end{equation*}
1877: Because of the special form of the first (matrix) summand, this
1878: leads to just one scalar equation, namely,
1879: \begin{equation*}
1880: \wT{{45}}-
1881: \begin{bmatrix}
1882: \wT{{14}} & \wT{{24}} & \wT{{46}}\\
1883: \end{bmatrix}
1884: G^{-1}
1885: \begin{bmatrix}
1886: \wT{{15}}\\ \wT{{25}} \\ \wT{{56}}\\
1887: \end{bmatrix}=0 .
1888: \end{equation*}
1889: This is equivalent to : $G$ is invertible and
1890: \begin{equation}
1891: \wT{{45}}\text{det}G-
1892: \begin{bmatrix}
1893: \wT{{14}} & \wT{{24}} & \wT{{46}}\\
1894: \end{bmatrix}
1895: \text{adj}G
1896: \begin{bmatrix}
1897: \wT{{15}}\\ \wT{{25}} \\ \wT{{56}}\\
1898: \end{bmatrix}=0 .
1899: \label{A-3}
1900: \end{equation}
1901: We have
1902: \begin{equation*}
1903: \text{adj}G=
1904: \begin{bmatrix}
1905: yz-\wT{{26}}^2 & \wT{{26}}u-\wT{{12}}z & \wT{{12}}\wT{{26}}-uy\\
1906: \wT{{26}}u-\wT{{12}}z & xz-u^2 &
1907: \wT{{12}}u-\wT{{26}}x\\
1908: \wT{{12}}\wT{{26}}-uy & \wT{{12}}u-\wT{{26}}x &
1909: xy-\wT{{12}}^2\\
1910: \end{bmatrix}\, .
1911: \end{equation*}
1912: We consider the left hand side of (\ref{A-3}) as a linear function
1913: of $x$, with coefficients in $F[y,z,u]$ . Denote by $\psi (y,z) $
1914: and $-\varphi(y,z,u)$, respectively, the coefficient of $x$ and
1915: the free coefficient. Then \begin{equation*} \psi (y,z)
1916: =\wT{{45}}(yz-\wT{{26}}^2)-\wT{{24}}\wT{{25}}z-\wT{{46}}\wT{{56}}y+\wT{{26}}(\wT{{24}}\wT{{56}}+\wT{{46}}\wT{{25}})
1917: \, .
1918: \end{equation*}
1919: Suppose first that $\psi (y,z)=0 $. This implies
1920: $\wT{{45}}=\wT{{25}}=\wT{{56}}=0$, and hence the constant
1921: coefficient in the right hand side of (\ref{A-3}) is
1922: \begin{equation*}
1923: -\wT{14}\wT{15}(yz-\wT{26}^2)-\wT{24}\wT{15}(\wT{26}u-\wT{12}z)-\wT{46}\wT{15}(\wT{12}\wT{26}-uy)\,
1924: .
1925: \end{equation*}
1926: The coefficient of u in this expression is
1927: $-\wT{15}\wT{24}\wT{26}+\wT{15}\wT{46}y$. We choose $y$ and $z$ in
1928: $F$ so that this coefficient of $u$ is nonzero and $yz-\wT{26}^2
1929: \neq 0$ . We can determine $u$ so that (\ref{A-3}) is satisfied,
1930: and then we determine $x$ so that $\text{det}G\neq 0$.
1931:
1932: We assume now that $\psi (y,z) \neq 0$. We let $x=\frac{\varphi
1933: (y,z,u)}{\psi (y,z)}$ , so any choice of $y,z,u$ in $F$ such that
1934: $\psi (y,z) \neq 0$ will yield a solution of (\ref{A-3}). We have
1935: to make a choice such that $\text{det}G \neq 0$. We have
1936: \begin{equation*}
1937: \text{det}G=\frac{(yz-\wT{26}^2)\varphi (y,z,u)}{\psi
1938: (y,z)}+2\wT{12}\wT{26}u-u^2y-\wT{12}^2z .
1939: \end{equation*}
1940: Let
1941: \begin{equation}
1942: \label{A-4} p(y,z,u)=(yz-\wT{26}^2)\varphi
1943: (y,z,u)+(2\wT{12}\wT{26}u-u^2y-\wT{12}^2z)\psi (y,z) \, .
1944: \end{equation}
1945: If $\wT{25} \neq 0$ the coefficient of $uz$ in $p(y,z,u)$ is
1946: $-2\wT{12}\wT{26}\wT{24}\wT{25}\neq 0$, so $p(y,z,u) \neq 0$.
1947: Hence we may assume $\wT{25}=0$. If $\wT{45} \neq 0$ then the
1948: coefficient of $u^2y^2z$ in $p(y,z,u)$ is $\wT{45}$, so
1949: $p(y,z,u)\neq 0$. Hence we may assume $\wT{45}= 0$. Since $\psi
1950: (y,z)\neq 0$ we must have $\wT{56}\neq 0$, implying that the
1951: coefficient of $u^2y$ in $p(y,z,u)$ is $-\wT{24}\wT{26}\wT{56}\neq
1952: 0$, so $p(y,z,u)\neq 0$. It follows that $x,y,z,u$ can be chosen
1953: so that $G$ is invertible and (\ref{A-3}) holds, so $B=0$
1954: \end {proof}
1955:
1956: %%%%%%%%%%
1957: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%THE APPENDIX B
1958: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1959: \newpage
1960: \appendix
1961: %%%%% appendix title trick
1962: \renewcommand {\thesection} {Appendix B}
1963: \section {}
1964: \renewcommand {\thesection} {A}
1965: %%%%% end of appendix title
1966: \indent
1967:
1968: %\includegraphics[width=165mm]{mypics/TableB1}
1969: \includegraphics[width=.9\textwidth]{mypics/TableB1}
1970:
1971:
1972:
1973:
1974:
1975:
1976:
1977:
1978: \begin{thebibliography}{WWW}
1979:
1980: \bibitem{BFH} F.~Barioli, S.~Fallat and L.~Hogben, Computaion of
1981: minimal rank and path cover number for certain graphs,{\it Linear
1982: Algebra Appl.}, {\bf 392} (2004), 289--303.
1983:
1984: \bibitem{BD} A.~Bento and A.~Leal Duarte,
1985: On Fiedler's Characterization of Tridiagonal Matrices over
1986: Arbitrary Fields, {\it Linear Algebra Appl.}, {\bf 401}
1987: (2005), 467--481.
1988:
1989: \bibitem{BLS} A.~Brandst\"adt, V.~B.~Le, and J.~P.~Spinrad,
1990: {\it Graph Classes: A Survey}, SIAM Monographs on Discrete
1991: Mathematics and Applications, Philadelphia, 1999.
1992:
1993: \bibitem{F} M.~Fiedler, A Characterization of Tridiagonal
1994: Matrices, {\it Linear Algebra Appl.}, {\bf 2} (1969),
1995: 191--197.
1996:
1997: \bibitem{vdH} H.~van der Holst, Graphs whose positive
1998: semi-definite matrices have nullity at most two, {\it Linear Algebra
1999: Appl.}, {bf 375} (2003), 1--11.
2000:
2001: \bibitem{HJ1} R.~Horn and C.~R.~Johnson, {\it Matrix Analysis},
2002: Cambridge University Press, New York, 1990.
2003:
2004: \bibitem{HJ2} R.~Horn and C.~R.~Johnson, {\it Topics in Matrix Analysis},
2005: Cambridge University Press, New York, 1991.
2006:
2007: \bibitem{JLD} C.~R.~Johnson and A.~Leal Duarte,
2008: The Maximum Multiplicity of an Eigenvalue in a Matrix
2009: Whose Graph is a Tree, {\it Linear and Multilinear Algebra},
2010: {\bf 46} (1999), 139--144.
2011:
2012: \bibitem{WC} J.~A.~Wald and C.~J.~Colbourn, {\it Steiner trees,
2013: partial 2--trees, and minimum IFI networks}, Networks, {\bf 13}
2014: (1983), 159--167.
2015:
2016: \end{thebibliography}
2017:
2018: \end{document}
2019: