1: \documentclass{article}
2: \usepackage{amssymb}
3: \usepackage{amsthm}
4: \usepackage[dvips]{graphicx}
5:
6: \widowpenalty=10000
7: \clubpenalty=10000
8:
9: \def\pn{\mathrel{{\mathop+}_n}}
10: \def\mn{\mathrel{{\mathop-}_n}}
11: \def\tn{\mathrel{{\mathop\cdot}_n}}
12: \def\tnd{\mathrel{{\mathop\cdot}_{n/d_i}}}
13: \newcommand{\floor}[1]{\left\lfloor#1\right\rfloor}
14: \newcommand{\ceil}[1]{\left\lceil#1\right\rceil}
15: \newtheorem{fact}{Fact}
16: \newtheorem{lemma}[fact]{Lemma}
17: \newtheorem{theorem}[fact]{Theorem}
18: \newtheorem{corollary}[fact]{Corollary}
19:
20: \begin{document}
21:
22: \title{Minimal Chordal Sense of Direction and Circulant Graphs}
23:
24: \author{Rodrigo~S.~C.~Le\~ao\thanks{Corresponding author
25: ({\tt rleao@cos.ufrj.br}).}\\
26: Valmir~C.~Barbosa\\
27: \\
28: Universidade Federal do Rio de Janeiro\\
29: Programa de Engenharia de Sistemas e Computa\c c\~ao, COPPE\\
30: Caixa Postal 68511\\
31: 21941-972 Rio de Janeiro - RJ, Brazil}
32:
33: \date{March 3, 2005}
34:
35: \maketitle
36:
37: \begin{abstract}
38: A sense of direction is an edge labeling on graphs that
39: follows a globally consistent scheme and is known to considerably
40: reduce the complexity of several distributed problems. In this paper, we study aparticular instance of sense of direction, called a chordal sense of direction
41: (CSD). In special, we identify the class of $k$-regular graphs that admit a CSD
42: with exactly $k$ labels (a minimal CSD). We prove that connected graphs in this class
43: are Hamiltonian and that the class is equivalent to that of circulant graphs,
44: presenting an efficient (polynomial-time) way of recognizing it when the graphs'degree $k$ is fixed.
45:
46: \bigskip
47: \noindent
48: {\bf Keywords:} Chordal sense of direction, Cayley graphs, Circulant graphs.
49: \end{abstract}
50:
51: \section{Introduction}
52:
53: In this paper we model a distributed system as an undirected graph $G$
54: on $n$ vertices having no multiple edges or self-loops. Every edge of $G$ is
55: assumed to have two labels, each corresponding to one of its end vertices. For
56: terminology or notation on graph theory not defined here we refer the reader to
57: \cite{bondy}.
58:
59: In \cite{santoro}, a property of this edge labeling was introduced which can
60: considerably reduce the complexity of many problems in distributed computing
61: \cite{flocchini7}. This property refers to the ability of a vertex to
62: distinguish among its incident edges according to a globally consistent scheme
63: and is formally described in \cite{flocchini1}. An edge labeling for which the
64: property holds is called a \emph{sense of direction} and is necessarily such
65: that all the edge labels corresponding to a same vertex are distinct (the edge
66: labeling is then what is called a \emph{local orientation}). We say that a sense
67: of direction is \emph{symmetric} if it is possible to derive the
68: label corresponding to one end vertex of an edge from the label corresponding to
69: the other. We say that it is \emph{minimal} if it requires exactly
70: $\Delta(G)$ distinct labels, where $\Delta(G)$ is the maximum degree in $G$.
71: For a survey on sense of direction, we refer the reader to \cite{flocchini8}.
72:
73: A particular instance of symmetric sense of direction, called a \emph{chordal
74: sense of direction} (CSD), can be constructed on any graph by fixing an arbitrary
75: cyclic ordering of the vertices and, for each edge $uv$, selecting the
76: difference (modulo $n$) from the rank of $u$ in the ordering to that of $v$ as
77: the label of $uv$ that corresponds to $u$ (likewise, the label that corresponds
78: to $v$ is the rank difference from $v$ to $u$). In Figure~\ref{fig_example}(a),
79: an example is given of a minimal chordal sense of direction (MCSD). It is
80: relatively easy to see that there exist graphs that do not admit an MCSD, as for
81: instance the one in Figure~\ref{fig_example}(b).
82:
83: \begin{figure}[t]
84: \centering
85: \begin{tabular}{c@{\hspace{1.5cm}}c}
86: \includegraphics[scale=0.75]{example.eps}&
87: \includegraphics[scale=0.75]{no_MCSD.eps}\\
88: {\small (a)}&{\small (b)}
89: \end{tabular}
90: \caption{A graph with an edge labeling that is an MCSD (a) and another graph
91: with an edge labeling that is a CSD but not an MCSD (b). Vertices are ordered
92: clockwise.}
93: \label{fig_example}
94: \end{figure}
95:
96: Given a finite group $A$ and a set of generators $S\subseteq A$, a \emph{Cayley graph}
97: is a graph $H$ whose vertices are the elements of the group ($V(H)=A$)
98: and whose edges correspond to the action of the generators
99: ($uv \in E(H) \iff \exists s \in S: v=s*u$, where $*$ is the operation defined
100: for $A$). We assume that the set of generators is closed under inversion, so $H$
101: is an undirected graph. An edge labeling of $H$ assigning two labels to each edge
102: in such a way that each of an edge's labels corresponds to one of its end vertices
103: is called a \emph{Cayley labeling} if, for edge $uv$, the
104: label that corresponds to vertex $u$ is $s$ such that $v=s*u$.
105:
106: In \cite{flocchini6}, it was shown that a regular graph's edge labeling
107: is a symmetric sense of direction if and only if the graph is a Cayley
108: graph and the labeling is a Cayley labeling. This result was later extended to
109: directed graphs in \cite{boldi3}, where the problem of recognizing labeled
110: Cayley graphs was also demonstrated to be solvable in parallel polynomial time.
111: This latter result uses the same $O(n^{14.256})$-time algorithm of
112: \cite{boldi1}, where the problem of deciding whether a given labeling is a sense
113: of direction of a given graph was solved.
114:
115: A \emph{circulant graph} (also known as a \emph{chordal ring}) is a Cayley graph
116: over $\mathbb{Z}_n$, the cyclic group of order $n$ under the addition operation.
117: The relevance of circulant graphs is due to their connectivity properties (small
118: diameter, high symmetry, etc.), which render them excellent topologies for
119: network interconnection, VLSI, and distributed systems \cite{bermond}.
120: The problem of recognizing circulant graphs is still challenging: results
121: are known only for very specific instances, like the cases of $n$ prime
122: \cite{muzychuk2}, geometric circulant graphs \cite{muzychuk6}, and recursive
123: circulant graphs \cite{fertin}.
124:
125: In this paper, we identify and analyze the regular graphs that admit an MCSD.
126: We describe their structure, show that they are all Hamiltonian if connected, and moreover
127: demonstrate an equivalence between certain distinct labelings.
128: We also show that the class of regular graphs that admit an
129: MCSD and the class of circulant graphs are equivalent to each other. A
130: straightforward consequence of our analysis is that the problem of recognizing
131: circulant graphs can be polynomially solved when the graphs' degree is fixed.
132:
133: Throughout the text, the operators $\pn$, $\mn$, and $\tn$ represent,
134: respectively, the modulo-$n$ operations of addition, subtraction, and
135: multiplication.
136:
137: \section{MCSD's of Regular Graphs}
138:
139: Let $G$ be a $k$-regular graph that admits an MCSD, $\lambda$ a labeling that is
140: an MCSD of $G$, and $\Gamma\subseteq\{1,\ldots,n-1\}$ the set of labels used by
141: $\lambda$. Since $\lambda$ is minimal, we may write $\Gamma=\{\gamma_1,\ldots,
142: \gamma_k\}$ and assume, further, that $\gamma_1<\cdots<\gamma_k$.
143: We denote by $\lambda_u(uv)$ the label of edge $uv$ that corresponds
144: to vertex $u$. We also write $\lambda(uv)=\{\lambda_u(uv),\lambda_v(uv)\}=
145: \{\gamma_i,\gamma_j\}$ with $1\leq i,j\leq k$. It is easy to see
146: that for any CSD there exists a symmetry function $\psi$ such that $\psi(\lambda_u(uv))=
147: \lambda_v(uv)$, given by $\psi(\gamma_i)=n-\gamma_i$ for $\gamma_i\in\Gamma$.
148: We start by highlighting an important property of $\lambda$.
149:
150: \begin{lemma}
151: \label{labeling}
152: If $k$ is even, then the edges of $G$ are labeled by $\lambda$ with the label
153: pairs $\{\gamma_1,n-\gamma_1\},\ldots,\{\gamma_{k/2},
154: n-\gamma_{k/2}\}$.
155: If $k$ is odd, then a further label pair is
156: $\{\gamma_{\ceil{k/2}},n-\gamma_{\ceil{k/2}}\}=\{n/2,n/2\}$.
157: \end{lemma}
158: \begin{proof}
159: The $k/2$ label pairs for the case of $k$ even follow directly from
160: the definition of $\psi$ and from the fact that $|\Gamma|=k$. If $k$ is odd,
161: then the label $\gamma_{\ceil{k/2}}=n/2$ (necessarily an integer, since $k$
162: odd implies $n$ even when $G$ is $k$-regular) remains unused by any of those
163: pairs, so a further label pair is $\{n/2,n/2\}$.
164: \end{proof}
165:
166: By Lemma~\ref{labeling}, we can always refer to $\lambda$ by simply giving the
167: $\ceil{k/2}$ labels that are no greater than $n/2$.
168: Having established this property of $\lambda$, we now set out to describe more
169: about the structure of graphs that admit an MCSD. In what follows, we say that a
170: graph $H$ \emph{decomposes} into the two subgraphs $A$ and $B$ when $V(A) \cup V(B)=
171: V(H)$, $E(A)\cup E(B)=E(H)$, and $E(A)\cap E(B)=\emptyset$. Also, recall that a
172: $2$-factor of $H$ is a collection of vertex-disjoint cycles from $H$ that
173: spans all of its vertices.
174:
175: \begin{theorem}
176: \label{decomp}
177: $G$ decomposes
178: into $\floor{k/2}$ $2$-factors and, if $k$ is odd, a perfect matching as well.
179: For $1\leq i\leq \floor{k/2}$, the edges of the $i$th $2$-factor are labeled
180: with $\gamma_i$, and the edges of the perfect matching with $n/2$.
181: \end{theorem}
182: \begin{proof}
183: The $k$ edges incident to each vertex are labeled with distinct members of
184: $\Gamma$ on their near ends. So each of the $\floor{k/2}$ label pairs asserted
185: initially in Lemma~\ref{labeling} can be used to identify a different
186: $2$-factor. Such $2$-factors encompass all of $G$, with the exception of the edges
187: whose label pair is $\{n/2,n/2\}$ in the odd-$k$ case (again, in Lemma~\ref{labeling}).
188: But these clearly constitute a perfect matching in $G$.
189: \end{proof}
190:
191: \section{The Necessity of a Hamiltonian Cycle}
192:
193: In this section we assume that $G$ is connected and begin by asserting
194: a relationship between two vertices that belong to a same
195: cycle of one of the $2$-factors established in Theorem~\ref{decomp}. Let us
196: denote by $r(u)$ the rank of vertex $u$ in the cyclic ordering that underlies
197: the CSD.
198:
199: \begin{lemma}
200: \label{samecycle}
201: For $1\leq i\leq \floor{k/2}$, two vertices $u$ and $v$ belong to a common cycle of
202: the $2$-factor whose edges are labeled with $\gamma_i$ if and only if $r(v)
203: =r(u)\pn t{\gamma_i}$ for some integer $t\geq 0$.
204: \end{lemma}
205: \begin{proof}
206: If $u$ and $v$ share a common $\gamma_i$-labeled cycle, then traversing the
207: cycle from $u$ to $v$ in the direction that exits a vertex along the
208: $\gamma_i$-labeled end of the edge adds (modulo $n$) $\gamma_i$ rank units to
209: $r(u)$ for each edge traversed. Then there exists a nonnegative integer $t$ such
210: that $r(v)=r(u)\pn t\gamma_i$.
211:
212: Conversely, if $r(v)=r(u)\pn t\gamma_i$ for some $t\geq 0$, then,
213: since every vertex has an incident
214: edge labeled with $\gamma_i$ on the near end, $v$ can be reached by a path that begins at $u$,
215: exclusively uses edges labeled with $\gamma_i$, and has $t$ edges. Clearly, such a path
216: is part of a cycle of a $2$-factor whose edges are labeled with $\gamma_i$.
217: \end{proof}
218:
219: Recall now that two integers $a$ and $b$ are \emph{relative primes}, denoted by
220: $a \perp b$, if $\gcd(a,b)=1$.
221:
222: \begin{fact}
223: \label{numtheory}
224: Let $a$ and $b$ be integers. Then $a \perp n$ if and only if the smallest $b>0$
225: that satisfies $b \tn a=0$ is $b=n$.
226: \end{fact}
227:
228: We are now in position to demonstrate that $G$ is Hamiltonian. We do this by
229: splitting the proof into cases that bear on the relative primality between
230: each of $\gamma_1,\ldots,\gamma_{\floor{k/2}}$ and $n$.
231:
232: \begin{theorem}
233: \label{prime_gamma}
234: If there exists $\gamma_i \in \{\gamma_1,\ldots,\gamma_{\floor{k/2}}\}$ such
235: that $\gamma_i \perp n$, then $G$ has a Hamiltonian cycle whose edges are
236: labeled with $\gamma_i$.
237: \end{theorem}
238: \begin{proof}
239: By Fact~\ref{numtheory}, the smallest integer $t>0$ that satisfies
240: $t\tn \gamma_i=0$ is $t=n$. In the same way, for any vertex $u$, the smallest
241: $t>0$ that satisfies $r(u)\pn t\gamma_i=r(u)$ is also $t=n$. By Lemma~\ref{samecycle},
242: vertex $u$ is on an $n$-vertex cycle whose edges are labeled with $\gamma_i$.
243: \end{proof}
244:
245: Before we proceed with the case in which no $\gamma_i \in \{\gamma_1,\ldots,
246: \gamma_{\floor{k/2}}\}$ is such that $\gamma_i\perp n$, we give a necessary
247: condition for such a scenario to happen.\footnote{The
248: reader should note that Theorem~\ref{prime_gamma} and Lemma~\ref{cycles}
249: could be coalesced into one single result stating that the number of cycles in the
250: $2$-factor whose edges are labeled with $\gamma_i$ for $1\leq i\leq\floor{k/2}$ is
251: $\gcd(\gamma_i,n)$. We choose to do otherwise for clarity's sake only.}
252: We also prove a general property on the relative primality
253: between the greatest common divisor of all the labels that are no greater than
254: $n/2$ (including, if applicable, $n/2$ itself) and $n$.
255:
256: \begin{lemma}
257: \label{cycles}
258: If no $\gamma_i \in \{\gamma_1,\ldots,\gamma_{\floor{k/2}}\}$ exists such that
259: $\gamma_i \perp n$, then for $1\leq i\leq\floor{k/2}$ the edges of $G$ labeled with
260: $\gamma_i$ form a
261: $2$-factor with $d_i$ cycles of the same length, where $d_i=\gcd(\gamma_i,n)>1$.
262: \end{lemma}
263: \begin{proof}
264: First rewrite $t\tn \gamma_i=0$ as $t\tnd \gamma_i /d_i=0$ and notice that
265: $\gamma_i /d_i \perp n/d_i$. By Fact~\ref{numtheory}, it follows that $t=n/d_i$ is the
266: smallest integer that satisfies $t\tnd \gamma_i /d_i=0$, and hence also
267: $t\tn \gamma_i=0$. By Lemma~\ref{samecycle}, each cycle of the $2$-factor whose
268: edges are labeled with $\gamma_i$ comprises $n/d_i$ vertices, so by Theorem~\ref{decomp}
269: the number of such cycles is $d_i$.
270: \end{proof}
271:
272: \begin{lemma}
273: \label{gcd}
274: Let $d=\gcd(\gamma_1,\ldots,\gamma_{\ceil{k/2}})$. Then $d\perp n$.
275: \end{lemma}
276: \begin{proof}
277: Let $t_i\geq 0$, for $1\leq i\leq \ceil{k/2}$, be an integer. Thus,
278: $\sum_i t_i\gamma_i$ is a multiple of $d$ and represents, for an arbitrary walk in
279: $G$, the rank difference along the CSD's cyclic vertex ordering from the walk's
280: initial vertex to its final vertex. To see why any walk is thus contemplated,
281: notice that a walk that uses $t_i$ edges labeled with $n-\gamma_i$ on their
282: near ends as vertices are exited along the walk can be
283: substituted for by another one that connects the same two vertices and uses
284: $n/\gcd(\gamma_i,n)-t_i$ edges labeled with $\gamma_i$ instead. Therefore, for
285: a walk between arbitrary vertices $u$ and $v$, the rank difference between these
286: vertices in the cyclic ordering is given by $t\tn d$ for some
287: $t\in\{0,\ldots,n-1\}$. And, since $G$ is connected, every possible value of
288: $t\tn d$ (i.e., $0,\ldots,n-1$) must result from a distinct value of $t$.
289: If such is the case, then
290: $\gcd(d,n)=1$, that is, $d \perp n$.\footnote{A formal proof of this implication
291: can be found in Section~4.8 of \cite{knuth}.}
292: \end{proof}
293:
294: It is important to recall that, when $k$ is odd, $\gamma_{\ceil{k/2}}=n/2$ by
295: Lemma~\ref{labeling}. In this case, by Lemma~\ref{gcd} we must have $d=1$ (since
296: $d$ divides $n/2$, therefore $n$ as well, which for $d>1$ contradicts the lemma).
297: The existence of a Hamiltonian cycle when none of $\gamma_1,\ldots,
298: \gamma_{\floor{k/2}}$ is relatively prime to $n$ can now be proven.
299:
300: \begin{theorem}
301: \label{not_prime_gamma}
302: If no $\gamma_i \in \{\gamma_1,\ldots,\gamma_{\floor{k/2}}\}$ exists such that
303: $\gamma_i \perp n$, then $G$ has a Hamiltonian cycle.
304: \end{theorem}
305: \begin{proof}
306: Note, first, that every one of $\gamma_1,\ldots,\gamma_{\floor{k/2}}$ is
307: necessarily greater than $1$.
308: Let $d'=\gcd(\gamma_1,\ldots,\gamma_{\floor{k/2}})$. The proof is divided into
309: two cases: $d'=1$ and $d'>1$.
310:
311: Let $d'=1$. Then there exist $\gamma_i$ and $\gamma_j$ in $\{\gamma_1,\ldots,
312: \gamma_{\floor{k/2}}\}$ such that $\gamma_i \perp \gamma_j$. Suppose that two
313: vertices $u$ and $v$ are connected by an edge labeled with $\gamma_i$ and belong
314: to the same cycle of the $2$-factor whose edges are labeled with $\gamma_j$. In
315: this case, we have $\gamma_i=t \tn \gamma_j$ for some integer $t\geq 0$, which
316: is a contradiction, since $\gamma_i \perp \gamma_j$. It follows that the end
317: vertices of an edge labeled with $\gamma_i$ belong to distinct cycles of the
318: $2$-factor whose edges are labeled with $\gamma_j$. We can also say that, if $u$
319: and $v$ are adjacent on a cycle $C_1$ of the $2$-factor whose edges are labeled
320: with $\gamma_i$, then there exist edges labeled with $\gamma_j$ that connect $u$
321: and $v$ to vertices $u'$ and $v'$, respectively, where $u'$ and $v'$ are
322: adjacent on a cycle $C_2$ of the same $2$-factor. This is summarized in the
323: equality
324: \begin{equation}
325: \label{equality}
326: \gamma_j\pn\gamma_i\pn(n-\gamma_j) = \gamma_i,
327: \end{equation}
328: which refers to the illustration in Figure~\ref{fig_path}.
329: Let then $C_1,\ldots,C_{d_i}$ be the cycles of the $2$-factor whose edges are labeled
330: with $\gamma_i$ (by Lemma~\ref{cycles}, $d_i=\gcd(\gamma_i,n)$). We can easily
331: identify a Hamiltonian cycle using $C_1,\ldots,C_{d_i}$
332: and interconnecting these cycles through the edges labeled with
333: $\gamma_j$, as in Figure~\ref{fig_hamilt1}.
334:
335: \begin{figure}[t]
336: \centering
337: \includegraphics[scale=0.75]{path.eps}
338: \caption{A path connecting vertices $u$ and $v$ and whose edges are labeled
339: with $\gamma_j$, $\gamma_i$, and $n-\gamma_j$ (in this order). Cycles $C_1$ and
340: $C_2$ belong to the $2$-factor whose edges are labeled with $\gamma_i$.}
341: \label{fig_path}
342: \end{figure}
343:
344: \begin{figure}[t]
345: \centering
346: \includegraphics[scale=0.75]{hamilt1.eps}
347: \caption{Hamiltonian cycle using the cycles whose edges are labeled with
348: $\gamma_i$ and edges labeled with $\gamma_j$.}
349: \label{fig_hamilt1}
350: \end{figure}
351:
352: Now let $d'>1$. If no edges are labeled with $n/2$, then every path in $G$, say
353: from $u$ to $v$,
354: is such that $r(v)\mn r(u)$ is a multiple (modulo $n$) of $d'$. And since $G$ is
355: connected, this has to hold for all vertex pairs in the graph, even those whose
356: rank differences are not a multiple of $d'$. This is clearly contradictory, so
357: there have to exist
358: edges labeled with $n/2$ (in which case $k$ must be odd, by Lemma~\ref{labeling})
359: and we must have
360: $d'\perp n/2$ (by Lemma~\ref{gcd}, according to which $\gcd(d',n/2)\perp n$).
361: This latter conclusion allows us to
362: substitute $n/2$ for $\gamma_j$ in (\ref{equality}), and then we see that the edges
363: labeled with $n/2$ connect two distinct cycles of the $2$-factor whose edges are
364: labeled with $\gamma_i$. If we take all such cycles and alternately interconnect
365: them by an edge labeled with $n/2$ and another labeled with an
366: appropriate $\gamma_j$,\footnote{It suffices for $\gamma_i$ and $\gamma_j$ not to
367: be multiples of each other. The case in which every one of $\gamma_1,\ldots,
368: \gamma_{\floor{k/2}}$ is a multiple of $\gamma_i$ has a trivial Hamiltonian cycle
369: that uses only edges labeled with $\gamma_i$ and $n/2$.} then a
370: Hamiltonian cycle is easily constructed, as in Figure~\ref{fig_hamilt2}.
371: \end{proof}
372:
373: \begin{figure}[t]
374: \centering
375: \includegraphics[scale=0.75]{hamilt2.eps}
376: \caption{Hamiltonian cycle using the cycles whose edges are labeled with
377: $\gamma_i$ and edges labeled with $n/2$ and $\gamma_j$.}
378: \label{fig_hamilt2}
379: \end{figure}
380:
381: The following is then straightforward.
382:
383: \begin{corollary}
384: \label{hamilt}
385: $G$ has a Hamiltonian cycle.
386: \end{corollary}
387: \begin{proof}
388: By Theorems~\ref{prime_gamma} and \ref{not_prime_gamma}.
389: \end{proof}
390:
391: Note, finally, that even though the presence of a
392: Hamiltonian cycle is a necessary condition for an
393: edge labeling to be an MCSD in connected regular graphs,
394: it is not a sufficient condition.
395: In fact, it is easy to find $k$-regular graphs that have a Hamiltonian cycle but
396: do not admit a CSD with $k$ labels, as the example in Figure~\ref{fig_example}(b).
397:
398: \section{Deciding Whether an MCSD Exists}
399:
400: Given an arbitrary set $\{\gamma_1,\ldots,\gamma_{\ceil{k/2}}\}$ of labels such
401: that $\gamma_1<\cdots<\gamma_{\ceil{k/2}}$, if none of its members is
402: greater than $n/2$ with $\gamma_{\ceil{k/2}}=n/2$ in the odd-$k$ case,
403: then one can easily (polynomially) generate a $k$-regular graph $H$ with an MCSD by simply
404: arranging the $n$ vertices in a cyclic ordering and, for each $\gamma_i \in
405: \{\gamma_1,\ldots,\gamma_{\ceil{k/2}}\}$, connecting pairs of vertices whose
406: ranks in the ordering differ by $\gamma_i$ and labeling the resulting edges
407: with $\{\gamma_i,n-\gamma_i\}$ appropriately.
408: In order for $H$ to be connected, by Lemma~\ref{gcd} we require in addition that
409: $\gcd(\gamma_1,\ldots,\gamma_{\ceil{k/2}})\perp n$.
410: Thus, a possible
411: direction towards the development of an algorithm to check whether a given
412: $k$-regular graph $G$ admits an MCSD is to generate $H$ in this way for
413: every pertinent set of labels,\footnote{It is easy to see that $H$ is unique for
414: a given set of labels.} and then to check whether $H$ is
415: isomorphic to $G$.
416:
417: When we fix the input graph's degree (i.e., $k$ is a constant), the
418: maximum number of candidate labelings to be checked if we ignore the restriction that
419: $\gcd(\gamma_1,\ldots,\gamma_{\ceil{k/2}})\perp n$ in the connected case is
420: \[ {\floor{n/2}\choose \floor{k/2}}=O(n^k), \]
421: a polynomial in $n$. In \cite{luks}, the isomorphism of graphs of bounded degree
422: was shown to be testable in polynomial time. Thus, we can decide whether $G$
423: admits an MCSD also polynomially.
424:
425: The complexity of the overall MCSD test can be clearly improved if we consider
426: the possible
427: isomorphism between graphs generated from distinct valid label sets. We say
428: that two distinct labelings $\lambda$ and $\lambda'$ are \emph{equivalent},
429: denoted by $\lambda \equiv \lambda'$, if they generate isomorphic graphs. For
430: example, it can be easily seen that the labelings $\lambda$ and $\lambda'$,
431: drawing respectively on the label sets $\{1,2,5\}$ and $\{3,4,5\}$,
432: generate isomorphic $5$-regular graphs on $10$ vertices (see
433: Figure~\ref{fig_equiv}), so $\lambda \equiv \lambda'$.
434: Let us then consider a transformation of $\lambda$ into $\lambda'$ that
435: preserves the MCSD property. In what follows, $\lambda$ draws on the
436: set $\{\gamma_1,\ldots,\gamma_{\ceil{k/2}}\}$ for labels, $\lambda'$ on
437: $\{\gamma_1',\ldots,\gamma_{\ceil{k/2}}'\}$.
438:
439: \begin{figure}[t]
440: \centering
441: \begin{tabular}{c@{\hspace{1.5cm}}c}
442: \includegraphics[scale=0.75]{equiv_a.eps}&
443: \includegraphics[scale=0.75]{equiv_b.eps}\\
444: {\small (a)}&{\small (b)}
445: \end{tabular}
446: \caption{Isomorphic graphs with equivalent MCSD labelings, based on
447: the label sets $\{1,2,5\}$ (a) and $\{3,4,5\}$ (b).}
448: \label{fig_equiv}
449: \end{figure}
450:
451: \begin{theorem}
452: \label{transform}
453: Let $\alpha<n/2$ be an integer such that $\alpha \perp n$. For $1\leq i\leq
454: \ceil{k/2}$, let
455: \[ \gamma_i'= \left\{\begin{array}{ll}
456: \alpha\tn \gamma_i & \mbox{\rm if $\alpha\tn \gamma_i\leq n/2$} \\
457: n-\alpha\tn \gamma_i & \mbox{\rm if $\alpha\tn \gamma_i>n/2$}. \\
458: \end{array} \right. \]
459: If $\lambda$ is an MCSD for $G$, then so is $\lambda'$.
460: \end{theorem}
461: \begin{proof}
462: It suffices that we argue that no member of $\{\gamma_1',\ldots,\gamma_{\ceil{k/2}}'\}$
463: is greater than $n/2$ with $\gamma_{\ceil{k/2}}=n/2$ for $k$ odd,
464: that every two members of this set are distinct, and also that
465: the vertices can be rearranged cyclically so that $\lambda'$ is indeed a CSD. The first of
466: these properties holds trivially and the second follows from well-known number-theoretic
467: properties.\footnote{We once again refer the reader to Section~4.8 of \cite{knuth}.}
468: As for the third property, clearly it suffices for the vertices to be arranged into a
469: cyclic ordering in which vertex $u$ has rank $r'(u)$ such that
470: $r'(u)=\alpha\tn r(u)$.
471: \end{proof}
472:
473: It is easy to see that any $\alpha>n/2$ would produce the same results as its
474: symmetric modulo $n$. We can also see that any $\lambda$ comprising a $\gamma_i$
475: such that $\gamma_i \perp n$ can yield a $\lambda'$ with $\gamma_i'=1$. For
476: such, it is sufficient to take $\alpha$ as the multiplicative inverse (modulo
477: $n$) of $\gamma_i$.
478:
479: It is also curious to note that the transformation in Theorem~\ref{transform}
480: ensures that $\gamma_i\perp n$ if and only if $\gamma_i'\perp n$. To see this,
481: consider for example the case of $\gamma_i'=\alpha\tn\gamma_i$. By using
482: Euclid's Theorem \cite{knuth} and the fact that $\alpha\perp n$ in succession,
483: we have $\gcd(\alpha\tn\gamma_i,n)=\gcd(\alpha\gamma_i,n)=\gcd(\gamma_i,n)$,
484: thence $\gcd(\gamma_i',n)=\gcd(\gamma_i,n)$.
485:
486: \section{MCSD's and Circulant Graphs}
487: \label{circ}
488:
489: There is a clear equivalence between circulant graphs and
490: regular graphs that admit an MCSD. We describe it formally in the following
491: theorem.
492:
493: \begin{theorem}
494: \label{equiv}
495: $G$ is circulant of generator set $S$ if and only if it is $|S|$-regular and
496: admits an MCSD.
497: \end{theorem}
498: \begin{proof}
499: Let $G$ be a circulant graph of generator set $S$. Then $uv\in E(G)$ if and only
500: if there exists $s\in S$ such that $v=u\pn s$.
501: Let $\lambda$ be a labeling for $G$ such that
502: $\lambda_u(uv)=s$. Since the vertices of $G$ are elements of $\mathbb{Z}_n$,
503: they already have a natural cyclic ordering in which $r(u)=u$ for all $u \in
504: V(G)$. So $\lambda_u(uv)=r(v) \mn r(u)$ and $\lambda$ is a CSD of $G$. Also, as
505: every vertex is connected to the vertex ranking $s$ higher (modulo $n$) than
506: itself for every $s\in S$, $G$ is $|S|$-regular and $\lambda$ uses $|S|$ labels
507: (thence the CSD is minimal).
508:
509: Conversely, if $G$ is a $k$-regular graph that admits an MCSD, then
510: $|\Gamma|=k$ and there exists
511: a cyclic ordering of the vertices such that each edge $uv$ is labeled with
512: $\lambda_u(uv)=r(v) \mn r(u)$, where $0\leq r(u) \leq n-1$. Letting
513: $V(G)=\mathbb{Z}_n$ so that $u=r(u)$ and $S=\Gamma$ yields $\lambda_u(uv)=v\mn u$
514: for all $uv\in E(G)$, thence $v=u\pn\lambda_u(uv)$. $G$ is therefore circulant
515: of generator set $S$.
516: \end{proof}
517:
518: One first example of how Theorem~\ref{equiv} sheds new light on the two concepts
519: involved comes from considering the result on circulant graphs in
520: \cite{boesch}, which implies in the connected case that $\gcd(s_0,\ldots,s_k,n)=1$, where
521: $s_i$, for $0\leq i \leq k$, is an element of the set of generators.
522: This is of course coherent with
523: Lemma~\ref{gcd} and Corollary~\ref{hamilt}, and a straightforward
524: implication of the general result of \cite{boesch}
525: is that a circulant graph is Hamiltonian if and only if it
526: is connected \cite{burkard}. Our approach introduces new ways of
527: constructing Hamiltonian cycles in this case.
528:
529: Another interesting insight is the following. An $n \times n$ matrix
530: is said to be \emph{circulant} if its $i$th line is the
531: cyclic shift of the first line by $i$ positions,
532: where $0\leq i\leq n-1$. Another characterization of
533: circulant graphs is that a graph is circulant if its
534: adjacency matrix is circulant. By
535: the equivalence established in Theorem~\ref{equiv}, it then becomes possible to
536: approach the problem of recognizing regular graphs that admit an MCSD along
537: a different route: since it is well-known that the
538: isomorphism between two graphs $G$ and $H$ can be viewed as a permutation of
539: lines and columns of the adjacency matrix of $G$ ($A(G)$) that generates that of
540: $H$, we can test whether a regular graph $G$ admits an MCSD by
541: finding a permutation of lines and columns of $A(G)$ such that the resulting
542: matrix is circulant.
543:
544: We note, in addition, that the transformation defined in Theorem~\ref{transform}
545: also has an analogue in the
546: literature on circulant graphs. Let $G$ and $H$ be circulant graphs such that their
547: sets of generators are $R$ and $S$, respectively.
548: We say that $R$ and $S$ are $proportional$, denoted
549: by $R \sim S$, if, for some $a\perp n$, $r=a\tn s$ bijectively for $r\in R$ and $s\in S$.
550: Clearly, if $R\sim S$ then
551: $G$ is isomorphic to $H$.
552: The converse statement was conjectured in \cite{adam} and is known as
553: \emph{\'Ad\'am's conjecture}. However, in \cite{elspas} the conjecture was
554: proven false.
555:
556: The problem of recognizing circulant graphs, finally, is probably the one most affected
557: by the equivalence of Theorem~\ref{equiv}. Even though the algorithm we suggest
558: to test whether a $k$-regular graph admits an MCSD is polynomial only for
559: fixed $k$, when applied to the context of circulant graphs it is the only known
560: result on arbitrary topologies (without any restrictions on the structure or the
561: number of vertices) for that class.
562:
563: \subsection*{Acknowledgments}
564:
565: The authors acknowledge partial support from CNPq, CAPES, and a FAPERJ BBP
566: grant.
567:
568: \bibliography{wg2005}
569: \bibliographystyle{plain}
570:
571: \end{document}
572:
573: