1: % submitted 1/25/06 Jorge Ramirex ramirez@math.jussieu.fr, special
2: % issue of EJC dedicated to OM05.
3:
4: % THIS DOCUMENT IS WRITTEN IN LATEX 2e
5: %
6: % TO FIND THE TITLE: search for the command \title using your word
7: % processor
8: %
9: \documentclass[12pt]{article}
10: \usepackage{amssymb,latexsym,start2e,graphs2,pstcol,pst-plot}
11:
12: \setlength{\topmargin}{-.1in}
13: \setlength{\textheight}{8.2in}
14: \setlength{\textwidth}{7in}
15: \setlength{\evensidemargin}{-.3in}
16: \setlength{\oddsidemargin}{-.3in}
17:
18:
19: \newtheorem{thm}{Theorem}[section]
20: \newtheorem{prop}[thm]{Proposition}
21: \newtheorem{cor}[thm]{Corollary}
22: \newtheorem{lem}[thm]{Lemma}
23: \newtheorem{conj}[thm]{Conjecture}
24: \newtheorem{exa}[thm]{Example}
25: \newtheorem{question}[thm]{Question}
26:
27: \newcommand{\BC}{\mathop{\rm BC}\nolimits}
28: \newcommand{\Mon}{\mathop{\rm Mon}\nolimits}
29:
30: \begin{document}
31: \pagestyle{plain}
32:
33: \title{Monomial Bases for Broken Circuit Complexes
34: }
35: \author{
36: Jason I. Brown\\[-5pt]
37: \small Department of Mathematics and Statistics\\[-5pt]
38: \small Dalhousie University\\[-5pt]
39: \small Halifax, NS B3H 3J5\\[-5pt]
40: \small CANADA\\[-5pt]
41: \small \texttt{brown@mscs.dal.ca}\\[10pt]
42: Bruce E. Sagan\\[-5pt]
43: \small Department of Mathematics\\[-5pt]
44: \small Michigan State University\\[-5pt]
45: \small East Lansing, MI 48824-1027\\[-5pt]
46: \small USA\\[-5pt]
47: \small \texttt{sagan@math.msu.edu}\\[10pt]
48: \it This paper is dedicated to Michel Las Vergnas\\
49: \it on the occasion of his 65th birthday.
50: }
51:
52: \date{\today\\[10pt]
53: \begin{flushleft}
54: \small Key Words: acyclic orientation, cocircuit, $f$-vector, $h$-vector,
55: homogeneous system of parameters, monomial basis, no broken
56: circuit, order ideal, Stanley-Reisner ring, theta graph, upper bound\\[5pt]
57: \small AMS subject classification (2000):
58: Primary 05E99;
59: Secondary 05C99, 13A02.
60: \end{flushleft}
61: }
62: \maketitle
63:
64: \begin{abstract}
65: Let $F$ be a field and let $G$ be a finite graph with a total ordering
66: on its edge set. Richard Stanley noted that the Stanley-Reisner
67: ring $F(G)$ of the broken circuit complex of $G$ is
68: Cohen-Macaulay. Jason Brown gave an explicit description of a
69: homogeneous system of parameters for $F(G)$ in terms of fundamental
70: cocircuits in $G$. So $F(G)$ modulo this hsop is a finite
71: dimensional vector space. We conjecture an explicit monomial basis
72: for this vector
73: space in terms of the circuits of $G$ and prove that the conjecture
74: is true for two infinite families of graphs. We also explore an
75: application of these ideas to bounding the number of acyclic
76: orientations of $G$ from above.
77: \end{abstract}
78:
79: \section{Simplicial complexes and chromatic polynomials}
80:
81: Let $E$ be a finite set and let $\De$ be an {\it abstract simplicial
82: complex on $E$}, i.e., a nonempty family of subsets of $E$ such that $S\in\De$
83: and $T\sbe S$ implies $T\in\De$. The elements $S$ of $\De$ are called
84: {\it faces\/}. We will assume henceforth that $\De$ is
85: {\it pure of rank $r$\/} which means that all maximal faces $S$ have
86: $|S|=r$ where the absolute value sign denotes cardinality.
87: Let $f_i=f_i(\De)$ be the number of $S\in\De$ with $|S|=i$.
88: Then $\De$ has {\it $f$-vector\/}
89: $$
90: \bff=\bff(\De)=(f_0,f_1,\ldots,f_r)
91: $$
92: as well as {\it $f$-polynomial\/}
93: $$
94: f(x)=f_\De(x)=f_0+f_1x+\cdots+f_r x^r
95: $$
96: where $x$ is a variable.
97: In the future we will continue the
98: practice of appending $\De$ in parentheses or as a subscript when we
99: wish to specify the complex, even if we do not do so in the
100: corresponding definition.
101:
102: Another important invariant of $\De$ is its $h$-vector. Define a
103: polynomial
104: $$
105: h(x):=(1-x)^r f\left(\frac{x}{1-x}\right)
106: =f_0(1-x)^r+f_1x(1-x)^r+\cdots+f_r x^r.
107: $$
108: Let $h_i$ be the coefficient of $x^i$ in $h(x)$ so that
109: $h(x)=\sum_i h_i x^i$. Then the {\it $h$-vector of $\De$\/} is
110: $$
111: \bh=(h_0,h_1,\ldots,h_r).
112: $$
113: It will sometimes be convenient to extend the range of
114: definition of the $f_i$ and $h_i$ by letting $f_i=h_i=0$ if $i<0$ or
115: $i>r$.
116:
117: Now suppose that $G$ is a finite graph with vertices $V=V(G)$ and edges
118: $E=E(G)$. We permit loops and multiple edges and will use the notation
119: $p=|V|$ and $q=|E|$. We will also write $v\in G$ for $v\in V(G)$ and
120: $e\in G$ for $e\in E(G)$ if it is clear from context whether we are
121: talking about the vertices or edges of $G$.
122: A {\it coloring\/} of $G$ is a function
123: $c:V\ra\{1,2,\ldots,\la\}$ and $c$ is {\it proper\/} if
124: $c(u)\neq v(v)$ for all edges $uv\in E$. Consider $G$'s {\it chromatic
125: polynomial}, $P(G)=P(G;\la)$, which is the number of such proper
126: colorings. Note that if $G$ has a loop then $P(G;\la)=0$.
127: It is well known that if $G$ is a loopless then $P(G;\la)$ is a monic
128: polynomial
129: of degree $p$ in $\la$ whose coefficients alternate in sign.
130: Writing
131: \beq
132: \label{Pf}
133: P(G;\la)=f_0\la^p-f_1\la^{p-1}+\cdots+(-1)^p f_p
134: \eeq
135: one can give the following interpretation to the coefficients $f_i$.
136:
137: Let $\cC=\cC(G)$ denote the set of {\it cycles\/} of $G$ which will also be called
138: the set of {\it circuits\/}. Suppose $G$ is {\it ordered\/} in that the edge set
139: $E$ has been given a linear ordering $e_1<e_2<\ldots<e_q$.
140: Then each $C\in\cC$ gives
141: rise to a {\it broken circuit\/}
142: $$
143: \Cb=C-\min C
144: $$
145: where $\min C$ is the
146: smallest edge of $C$ in the linear ordering.
147: The {\it broken complex of $G$}, $\De(G)$, is the
148: family of all subsets of $E$ which do not contain a broken
149: circuit.
150: It is easy to see that $\De(G)$ is a pure abstract simplicial complex.
151: Wilf~\cite{wil:wpc} was the first to consider this family of
152: sets as a complex.
153: In fact, $\De(G)$ is intimately connected with the chromatic
154: polynomial as can be seen in the following result which dates back to
155: Whitney~\cite{whi:lem}, although he did not state it in this form.
156: \bth[\cite{whi:lem}]
157: Let $P(G;\la)$ have coefficients $f_i$ as defined by~\ree{Pf}. Then
158: $$
159: f_i=f_i(\De(G)),\ 0\le i\le p.\qqed
160: $$
161: \eth
162:
163: One can think of the expansion~\ree{Pf} as being generated by a
164: sequence of deletions and contractions expressing $P(G;\la)$ as a
165: linear combination of chromatic polynomials of graphs with no edges. One
166: could use chromatic polynomials of trees instead, or equivalently expand
167: $P(G;\la)$ in terms of the basis $\{1\}\cup\{\la(\la-1)^i\ :\ i\ge0\}$
168: for the ring of polynomials in $\la$. So define coefficients $h_i$ by
169: \beq
170: \label{Ph}
171: P(G;\la)=h_0\la(\la-1)^{p-1}-h_1\la(\la-1)^{p-2}+\cdots+(-1)^p h_p.
172: \eeq
173: The next result follows easily from the previous theorem and the
174: definitions.
175: \bco
176: Define coefficients $h_i$ by~\ree{Ph}. Then
177: $$
178: h_i=h_i(\De(G)),\ 0\le i\le p.\qqed
179: $$
180: \eco
181:
182: Our goal is to give an explicit combinatorial description of the $h_i$
183: directly in terms of the broken circuits of the graph. To do this, we
184: will need some
185: machinery from the theory of Cohen-Macaulay rings.
186:
187:
188:
189: \section{Cohen-Macaulay rings and monomial ideals}
190:
191: Consider the polynomial ring $F[\bx]=F[x_1,x_2,\ldots,x_q]$ where
192: $F$ is a field and $\bx=\{x_1,x_2,\ldots,x_q\}$ is a set of variables.
193: If $E=\{e_1,e_2,\ldots,e_q\}$ then any $S\sbe E$ has
194: corresponding monomial
195: $$
196: \bx^S=\prod_{e_i\in S} x_i.
197: $$
198: Now given any simplicial complex $\De$ on $E$ we form its {\it
199: Stanley-Reisner ring}, $F(\De)$, by modding out by the non-faces of
200: $\De$, i.e.,
201: $$
202: F(\De)=F[\bx]/\spn{\bx^S\ :\ S\not\in\De}
203: $$
204: where $\spn{\cdot}$ denotes the ideal generated by the polynomials in the
205: brackets. Note that since we are generating an ideal, it suffices
206: to consider the $\bx^S$ where $S$ is a minimal non-face of $\De$.
207:
208: If $G$ is an ordered graph, then define
209: $$
210: F(G):=F(\De(G))=F[\bx]/\spn{\bx^\Cb\ :\ C\in\cC(G)}
211: $$
212: where we identify a (broken) circuit with its edge set.
213: This ring has a
214: {\it homogeneous system of parameters (hsop) of degree one\/},
215: i.e., a set of polynomials $\th_1,\ldots,\th_r\in F[\bx]$ which are homogeneous
216: of degree one and satisfy
217: \ben
218: \item $\th_1,\ldots,\th_r$ are algebraically independent, and
219: \item $F(G)/\spn{\th_1,\ldots,\th_r}$ is a finite dimensional vector space
220: over $F$.
221: \een
222: Brown~\cite{bro:cpo}
223: gave an explicit construction of an hsop as follows. To simplify
224: things, we will assume for now that $F=\bbZ_2$, the integers modulo
225: two. In the last section, we will describe how to modify these ideas
226: so that they will work over an arbitrary field.
227:
228: First note that if $G$ has blocks (maximal subgraphs having no
229: cutpoints) $G_1,G_2,\ldots,G_b$, then directly from the definitions we
230: have the ring isomorphism
231: \beq
232: \label{iso}
233: F(G)\iso \otimes_i F(G_i).
234: \eeq
235: So there is no
236: loss of generality in assuming that $G$ is a block and, in particular,
237: that $G$ is connected.
238: Let $T$ be a spanning tree of $G$. For each
239: edge $e\in T$, let $T'_e$ and $T''_e$ be the components of $T-e$. So
240: $e$ defines a {\it fundamental cocircuit\/}
241: $$
242: D_e=D_e(G)=\{uv\in E(G)\ :\ u\in T'_e,\ v\in T''_e\}
243: $$
244: as well as a homogeneous degree one polynomial
245: \beq
246: \label{th}
247: \th_e=\sum_{e_i\in D_e} x_i.
248: \eeq
249:
250: \thicklines
251: \setlength{\unitlength}{2pt}
252: \bfi
253: \bpi(80,70)(-10,-10)
254: \put(-10,30){\makebox(0,0){$G=$}}
255: \Gda
256: \Gad \Gdd \Ggd
257: \Gdg
258: \put(12,12){\makebox(0,0){$3$}}
259: \Gdaad
260: \put(48,12){\makebox(0,0){$7$}}
261: \Gdagd
262: \put(15,25){\makebox(0,0){$5$}}
263: \Gaddd
264: \put(12,48){\makebox(0,0){$1$}}
265: \Gaddg
266: \put(45,25){\makebox(0,0){$6$}}
267: \Gddgd
268: \put(33,43){\makebox(0,0){$2$}}
269: \Gdddg
270: \put(48,48){\makebox(0,0){$4$}}
271: \Ggddg
272: \epi
273: \hspace{50pt}
274: \bpi(80,70)(-10,-10)
275: \put(-10,30){\makebox(0,0){$T=$}}
276: \Gda
277: \Gad \Gdd \Ggd
278: \Gdg
279: \put(48,12){\makebox(0,0){$7$}}
280: \Gdagd
281: \put(15,25){\makebox(0,0){$5$}}
282: \Gaddd
283: \put(45,25){\makebox(0,0){$6$}}
284: \Gddgd
285: \put(48,48){\makebox(0,0){$4$}}
286: \Ggddg
287: \epi
288: \capt{A graph $G$ and spanning tree $T$}\label{GT}
289: \efi
290:
291:
292: Since this construction will be crucial, we illustrate it with an
293: example. Consider the graph $G$ and its spanning tree $T$ given in
294: Figure~\ref{GT}. For simplicity we have labeled the edges
295: $1,2,\ldots,7$ rather than $e_1,e_2,\ldots,e_7$. Then we have
296: \bea
297: \th_4&=&x_4+x_1+x_2,\\
298: \th_5&=&x_5+x_1+x_3,\\
299: \th_6&=&x_6+x_1+x_2+x_3,\\
300: \th_7&=&x_7+x_3.\\
301: \eea
302: For any graph $G$ we have the following result.
303: \bth[\cite{bro:cpo}]
304: If $G$ is a connected graph and $T$ a spanning tree then
305: the set of polynomials defined by~\ree{th} for $e\in T$ is an hsop for
306: $\bbZ_2(G)$.\qqed
307: \eth
308:
309: Continuing with the general development, let $\Mon(\bx)=\Mon(q)$
310: denote the set
311: of monomials in $F[\bx]=F[x_1,x_2,\ldots,x_q]$. When it will do no
312: harm, we will not distinguish between these monomials considered as
313: elements of $F[\bx]$ or considered as elements of some quotient of the
314: polynomial ring. A subset $L\subseteq\Mon(q)$ is a {\it lower order
315: ideal\/} (or {\it down set\/})
316: if whenever $m\in L$ and $n\in\Mon(q)$ divides $m$, then $n\in L$.
317: Similarly, $U \subseteq\Mon(q)$ is an {\it upper order
318: ideal\/} (or {\it filter\/}) if whenever $m\in L$ and $n\in\Mon(q)$ is
319: divisible by $m$, then $n\in L$. Note that $U$ is an upper order ideal
320: if and only if $\Mon(q)-U$ is a lower order ideal.
321: If $S\subseteq\Mon(q)$ then the {\it lower and upper order ideals
322: generated by $S$\/} are
323: \bea
324: L(S)&=&\{n\in\Mon(q)\ :\ \mbox{$n$ divides $m$ for some $m\in S$}\},\\
325: U(S)&=&\{n\in\Mon(q)\ :\ \mbox{$n$ is divisible by $m$ for some $m\in S$}\}.
326: \eea
327: Macaulay~\cite{mac:pet} showed that after modding out by an hsop, one can always
328: find a basis of monomials which forms a lower order ideal. And
329: Stanley~\cite{sta:ubc} connected such a basis with the $h$-vector.
330: \bth[\cite{mac:pet,sta:ubc}]
331: \label{mac}
332: Suppose that $I$ is an ideal of $F[\bx]$ and that
333: $\th_1,\ldots,\th_r$ be an hsop for $F[\bx]/I$. Then
334: the ring
335: $$
336: R=\frac{F[\bx]}{I+\spn{\th_1,\ldots,\th_r}}
337: $$
338: has a basis $L$ which is a lower order ideal of monomials.
339:
340: Suppose further that $F[\bx]/I$ is Cohen-Macaulay and $F[\bx]/I\iso
341: F(\De)$ for some simplicial complex
342: $\De$ with $h$-vector $\bh=(h_0,\ldots,h_r)$. Then
343: $$
344: h_i = \mbox{number of monomials of total degree $i$ in $L$}.\qqed
345: $$
346: \eth
347:
348: Now consider a graph $G$ with a spanning tree $T$ and define $I(G)$ to
349: be the ideal of $F[\bx]$ generated by the monomials $\bx^\Cb$ for
350: $C\in\cC(G)$. We wish to give an
351: explicit basis for the ring
352: $$
353: R(G)=\frac{F[\bx]}{I(G)+\spn{\th_e\ :\ e\in T}}
354: $$
355: which is a lower order ideal of monomials. First, however, we
356: wish to show that we have a basis inside $\Mon(\by)$ for a subset
357: $\by$ of $\bx$.
358:
359: An ordering $e_1<e_2<\ldots<e_q$ of
360: $E(G)$ will be called {\it standard\/} if the last $p-1$ edges in the
361: order form a tree. From now on we will assume that all our orderings
362: are standard and take our spanning tree $T=T(G)$ to be the one determined
363: the last edges in the order. It will also be convenient to denote the
364: number of edges not in $T$ by $k=q-p+1$. We will show that we that
365: our basis can be taken in $\Mon(\by)$ where $\by=\{x_1,x_2,\ldots,x_k\}$.
366:
367: We now return to working over $\bbZ_2$.
368: Suppose $k<j\le q$ and write $D_j$ for
369: $D_{e_j}$ and $\th_j$ for $\th_{e_j}$. Then since $\th_j=0$ in $R(G)$
370: we have
371: \beq
372: \label{xj}
373: x_j=\sum_{e_i\in (D_j-e_j)} x_i.
374: \eeq
375: where $x_i\in\by$ for all $x_i$ appearing in the sum. For each $C\in\cC$
376: let $p_\Cb=p_\Cb(\by)$ be the polynomial obtained from $x^\Cb$ by
377: substituting in
378: the sum in equation~\ree{xj} for $x_j$ for each $j>k$.
379: Consider the ideal
380: $$
381: J=J(G)=\spn{p_\Cb\ :\ C\in\cC}
382: $$
383: We immediately have the following result.
384: \bpr
385: If $G$ is a connected graph and $F=\bbZ_2$ then
386: $$
387: R(G)\iso \frac{\bbZ_2[\by]}{J(G)}.\qqed
388: $$
389: \epr
390:
391: Returning to our running example, we convert the list of circuits in
392: $G$ into polynomials using the equations for
393: $\th_4,\ldots,\th_7$.
394: $$
395: \barr{lll}
396: C_1=\{1,4,5,6\},\qquad &\bx^{\ol{C_1}}=x_4x_5x_6,\qquad
397: &p_{\ol{C_1}}=(x_1+x_2)(x_1+x_3)(x_1+x_2+x_3),\\
398: C_2=\{2,4,6\} &\bx^{\ol{C_2}}=x_4x_6,
399: &p_{\ol{C_2}}=(x_1+x_2)(x_1+x_2+x_3),\\
400: C_3=\{3,5,6,7\} &\bx^{\ol{C_3}}=x_5x_6x_7,
401: &p_{\ol{C_3}}= x_3 (x_1+x_3)(x_1+x_2+x_3),\\
402: C_4=\{1,2,5\} &\bx^{\ol{C_4}}=x_2x_5,
403: &p_{\ol{C_4}}= x_2 (x_1+x_3),\\
404: C_5=\{1,3,4,7\} &\bx^{\ol{C_5}}=x_3x_4x_7,
405: &p_{\ol{C_5}}= x_3^2 (x_1+x_2),\\
406: C_6=\{2,3,4,5,7\} &\bx^{\ol{C_6}}=x_3x_4x_5x_7,
407: &p_{\ol{C_6}}= x_3^2(x_1+x_2)(x_1+x_3)\\
408: C_7=\{1,2,3,6,7\} &\bx^{\ol{C_7}}=x_2x_3x_6x_7,
409: &p_{\ol{C_7}}= x_2 x_3^2(x_1+x_2+x_3).\\
410: \earr
411: $$
412:
413:
414: We will now pick a specific monomial $m_\Cb$ from
415: each $p_\Cb$ and these will be used to define the lower order ideal of
416: monomials being sought.
417: For $1\le i\le k$, the graph $T+e_i$ has a
418: unique circuit $C_i$ and these circuits will be called {\it fundamental}.
419: We label the nonfundamental circuits in some order as $C_i$ for $i>k$.
420: Also define
421: $$
422: d_i=\case{i}{if $i\le k$,}{\min \{j\ :\ e_j\in D_i\}}{if $i>k$.}
423: $$
424: Now let
425: $$
426: m_{\Cb_i}=\case{x_i^{|\Cb_i|}}{if $i\le k$,}
427: {\rule{0pt}{30pt}\dil\prod_{e_j\in\Cb_i} x_{d_j}}{if $i>k$.}
428: $$
429: It is easy to see from the definitions that $m_\Cb$ is indeed a term
430: in the polynomial $p_\Cb$.
431: Finally, define upper and lower order ideals
432: $$
433: U(G)=U(m_\Cb\ :\ C\in\cC(G))\qmq{and} L(G)=\Mon(k)-U(G).
434: $$
435: Note that all these quantities depend on the ordering imposed on the
436: edges and not just on the graph itself, even though our notation does
437: not reflect that. It is $L(G)$ which will be our candidate as a
438: monomial basis for $R(G)$
439:
440: Continuing with our example, $C_1$, $C_2$, and $C_3$ are fundamental
441: with 4, 3, and 4 edges (respectively) and so
442: $$
443: m_{\Cb_1}=x_1^3,\quad m_{\Cb_2}=x_2^2,\quad m_{\Cb_3}=x_3^3.
444: $$
445: The monomials $m_{\Cb}$ for the other four circuits are obtained by taking the
446: variable of smallest subscript in each factor of the corresponding
447: $p_{\Cb}$, so
448: $$
449: m_{\Cb_4}=x_1 x_2,\quad m_{\Cb_5}=x_1 x_3^2,\quad
450: m_{\Cb_6}=x_1^2 x_3^2,\quad m_{\Cb_6}=x_1 x_2 x_3^2.
451: $$
452: Thus $R(G)$ should have as basis
453: \bea
454: L(G)
455: &=&
456: \Mon(3)-U(x_1^3,\ x_2^2,\ x_3^3,\ x_1 x_2,\ x_1 x_3^2,\ x_1^2x_3^3,\ x_1 x_2 x_3^2)\\
457: &=&
458: \{1,\ x_1,\ x_2,\ x_3,\ x_1^2,\ x_3^2,\ x_1x_3,\ x_2x_3,\ x_1^2 x_3,\ x_2 x_3^2 \}.
459: \eea
460: and this can be verified directly.
461:
462:
463: A graph for which there is an ordering of $E$ such that $L(G)$ is a
464: basis for $R(G)$ will be said to have a {\it
465: no broken circuit basis\/} or {\it NBC basis\/}.
466: To outline the rest of the paper,
467: in the next section we will prove a
468: general theorem about when a graph has an NBC basis.
469: In Section~\ref{tf} we will apply these results
470: to show that two infinite
471: families of graphs do indeed have NBC bases.
472: Section~\ref{ub} will be devoted to
473: giving an upper bound for the number of acyclic orientations for a graph with an NBC basis.
474: We also compare this bound to others in the literature.
475: We end with some comments and open problems. This will include a
476: conjecture that every graph $G$ has ordering which produces an NBC
477: basis for $R(G)$, as well as a proposed line of attack on this idea.
478:
479:
480: \section{Graphs with NBC bases}
481: \label{gwb}
482:
483: One way to show that a graph has an NBC basis would be to use
484: induction. Since
485: the chromatic polynomial is involved, this would entail deletion and
486: contraction. If $e\in E(G)$ then let $G\sm e$ and $G/e$ denote $G$
487: with $e$ deleted and with $e$ contracted, respectively.
488: Since we are permitting loops and multiedges, both $G\sm e$ and $G/e$
489: will have exactly one less edge than $G$.
490: An elementary fact about the chromatic polynomial is that
491: $$
492: P(G;\la)=P(G\sm e;\la)-P(G/e;\la).
493: $$
494: Using this equation and~\ree{Ph} we easily obtain
495: the following proposition.
496: \bpr
497: Let $G$ be a graph and $e\in E(G)$. Then for all $i\ge0$ we have
498: $$
499: h_i(G)=h_i(G\sm e)+h_{i-1}(G/e).\qqed
500: $$
501: \epr
502:
503:
504: If we choose $e\in T$ then $T/e$ is a spanning tree of $G/e$ but
505: $T\sm e$ is no longer a tree. If, on the other hand, we choose
506: $e\not\in T$ then $T$ is still a spanning tree of $G\sm e$ but $T$ is
507: no longer a tree in $G/e$. However, we can get around these
508: difficulties if $G$ has a vertex $w$ with $\deg w =2$ where $\deg w$,
509: the {\it degree of $w$}, is the number of edges containing $w$.
510:
511: As noted before, it does no harm to restrict our attention to graphs
512: $G$ which are blocks so that $G\sm e$ and $G/e$ are connected for all
513: $e\in E$. We will say that a standard ordering $e_1<e_2<\ldots<e_q$ on $G$
514: imposes the {\it induced ordering\/} $e_1<e_2<e_3<\ldots<e_{q-1}$ on
515: $G\sm e_q$ and on $G/e_q$. Now suppose that $G$ has a
516: vertex $w$ with $\deg w=2$ and that $e_k,e_q$ are the two edges
517: containing $w$. Then if the ordering on $G$ is standard, so too
518: will be the induced orderings on $G\sm e_q$ and $G/e_q$.
519: Our primary tool for showing that certain graphs have NBC
520: bases will be the following theorem. Note that an example which
521: illustrates the proof of this result follows the demonstration, so the
522: reader may wish to read both in parallel.
523:
524: \bth
525: \label{deg2}
526: Let $G$ be a block with a standard ordering $e_1<e_2<\ldots<e_q$.
527: Suppose $G$ has a vertex $w$ of degree two such that the edges
528: containing $w$ are $e_k$ and $e_q$. If $R(G\sm e_q)$ and $R(G/e_q)$
529: have NBC bases in their induced standard orderings, then so does
530: $R(G)$.
531: \eth
532: \pf\
533: Let $\uplus$ denote disjoint union and if $S\sbe\Mon(k)$ and
534: $m\in\Mon(k)$ then let $mS=\{mn\ :\ n\in S\}$.
535: We first show that
536: \beq
537: \label{L(G)}
538: L(G)=L(G\sm e_q)\uplus x_k L(G/e_q)
539: \eeq
540: so that by our assumptions about
541: $R(G\sm e_q)$ and $R(G/e_q)$ and the previous proposition (summed
542: over all $i$) we have
543: $$
544: |L(G)|=|L(G\sm e_q)|+|L(G/e_q)|=\dim R(G\sm e_q)+\dim R(G/e_q)=\dim R(G)
545: $$
546: where dimension is being taken over the field $\bbZ_2$.
547:
548: Consider $G\sm e_q$. Note that $e_k$ is in the tree for $G\sm e_q$
549: and so the basis for $R(G\sm e_q)$ will be in $\Mon(k-1)$.
550: Also, from our assumptions on $w$, $e_k$ is the only
551: edge of $G\sm e_q$ containing $w$. So $C$ is a circuit of $G\sm e_q$ if
552: and only if $C$ is a circuit of $G$ not containing $e_k$. It follows
553: that $x_k$ is never a factor of $\bx^\Cb$ for such $C$. It also
554: follows that for $e_j\in T(G\sm e_q)$, $e_j\neq e_k$, we have
555: $D_j(G\sm e_q)=D_j(G)-e_k$. And both of these sets have the same
556: minimum since $e_k$ is the edge of largest index outside the tree for
557: $G$. Thus the generators for $J(G\sm e_q)$ are
558: obtained from those for $J(G)$ by setting $x_k=0$ wherever it
559: appears. So the monomials in $U(G\sm e_q)$ are precisely those in
560: $U(G)$ which do not have $x_k$ as a factor. Hence $L(G\sm e_q)$
561: consists of the monomials in $L(G)$ which do not have $x_k$ as a
562: factor.
563:
564: Now consider $G/e_q$. The circuits of $G$ are in bijection with the
565: circuits of $G/e_q$: If $C\in\cC(G)$ contains $e_q$ then it
566: corresponds to the circuit $C/e_q$ of $G/e_q$, while if $C$ does not
567: contain $e_q$ then it is
568: also a circuit of $G/e_q$ itself. We will call the former
569: circuits (in both $G$ and $G/e_q$) {\it type I\/}, and the latter {\it
570: type II}. Note that because of the assumptions on $w$, the type I and
571: type II circuits can also be characterized as those which do and do not
572: contain $e_k$, respectively. Since $e_q$ is the only edge of $T(G)$
573: containing $w$, we have $D_j(G/e_q)=D_j(G)$ for each
574: $e_j\in T(G/e_q)$. Thus, using $\pt_\Cb$ to denote the generators of
575: $J(G/e_q)$,
576: $$
577: p_\Cb=\case{x_k \pt_{\ol{C/e_q}}}{if $C$ is of type I,}
578: {\pt_\Cb}{if $C$ is of type II,}
579: $$
580: where the polynomials for the type II circuits have no factor of $x_k$.
581: Since $e_k$ has the largest index outside $T(G)$, the same relation
582: holds between the corresponding generators of $U(G)$ and of
583: $U(G/e_q)$, i.e., $m_\Cb= x_k {\tilde{m}}_{\ol{C/e_q}}$ or
584: ${\tilde{m}}_\Cb$ depending on whether $\Cb$ is type I or type II
585: (respectively), where the tilde indicates the the quantity is being
586: calculated in $C/e_q$.
587:
588: Now one sees that
589: $x_k L(G/e_q)$ consists precisely of the
590: monomials in $L(G)$ which have a factor of $x_k$:
591: Suppose that we have a monomial of $L(G)$ divisible by $x_k$. Then it can be
592: written as $x_k m$ for some $m\in\Mon(k)$. Since $x_k m$ is not
593: divisible by any type I generator of $U(G)$, and all such generators
594: have the form $x{\tilde{m}}_1$ for some type I generator ${\tilde{m}}_1$ of $U(G/e_q)$,
595: we see that $m$ is not divisible by ${\tilde{m}}_1$ for all type I generators
596: of $U(G/e_q)$. Also, $x_k m$ is not divisible by any type II
597: generator ${\tilde{m}}_2$ of $U(G)$, and all such generators do not have $x_k$
598: as a factor, so $m$ is not divisible by any type II generator ${\tilde{m}}_2$ of
599: $U(G/e_q)$. So $m\in L(G/e_q)$ and $x_k m \in x_k L(G/e_q)$. The
600: proof of the converse inclusion is similar.
601:
602: Since $L(G)$ is
603: clearly the disjoint union of its monomials with a factor of $x_k$ and
604: its monomials without a factor of $x_k$, we are done with the
605: demonstration of~\ree{L(G)}.
606: So we have proven that $L(G)$ contains $\dim R(G)$
607: monomials, and thus it will
608: suffice to show that these monomials span $R(G)$. For that, it
609: suffices to show that $L(G)$ spans $U(G)$. So take $m\in U(G)$.
610: Suppose first that $x_k$ is a factor of $m$ so that $m=x_kn$ for some
611: monomial $n$. Then from our work in the previous paragraph we see
612: that $n\in U(G/e_q)$. So by our assumption about $R(G/e_q)$, we can
613: write
614: \beq
615: \label{n}
616: n=\sum_{l\in L(G/e_q)} a_l l+ p
617: \eeq
618: where the $a_l$ are constants and $p\in J(G/e_q)$.
619: But $x_k l\in L(G)$ for $l\in L(G/e_q)$, and $x_k p\in J(G)$ for
620: $p\in J(G/e_q)$ since this is true for each of the generators of
621: $J(G/e_q)$. So multiplying~\ree{n} by $x_k$ expresses
622: $m=x_kn$ as a linear combination of elements of $L(G)$ modulo $J(G)$
623: as desired.
624:
625:
626: Now suppose that $x_k$ is not a factor of $m$. Then by our previous
627: results concerning $G\sm e_q$ we see that $m\in U(G\sm e_q)$. So
628: by our assumption about $R(G\sm e_q)$, we can write
629: \beq
630: \label{m}
631: m=\sum_{l\in L(G\setminus e_q)} a_l l + p
632: \eeq
633: where the $a_l$ are constants and $p\in J(G\sm e_q)$.
634: Now, as shown above, $l\in L(G\sm e_q)$ implies $l\in L(G)$.
635: Furthermore, there must be a $p'\in J(G)$ such that
636: $p'(x_1,\ldots,x_{k-1},0)=p$. It follows that $p'=p+x_k p''$ for some
637: $p''\in F[\by]$. But, from the previous paragraph, we have that
638: $x_kp''$ is spanned by $L(G)$ modulo $J(G)$. So substituting
639: $p=p'-x_kp''$ into~\ree{m} we have expressed $m$ as a linear
640: combination of elements of $L(G)$ modulo $J(G)$. Hence every monomial
641: is in the span of $L(G)$ and we are done.\Qqed
642:
643: \thicklines
644: \setlength{\unitlength}{2pt}
645: \bfi
646: \bpi(100,70)(-30,0)
647: \put(-20,30){\makebox(0,0){$G\sm e_7=$}}
648: \Gda
649: \Gad \Gdd \Ggd
650: \Gdg
651: \put(12,12){\makebox(0,0){$3$}}
652: \Gdaad
653: \put(15,25){\makebox(0,0){$5$}}
654: \Gaddd
655: \put(12,48){\makebox(0,0){$1$}}
656: \Gaddg
657: \put(45,25){\makebox(0,0){$6$}}
658: \Gddgd
659: \put(33,43){\makebox(0,0){$2$}}
660: \Gdddg
661: \put(48,48){\makebox(0,0){$4$}}
662: \Ggddg
663: \epi
664: \hspace{50pt}
665: \bpi(100,70)(-30,0)
666: \put(-20,30){\makebox(0,0){$T(G\sm e_7)=$}}
667: \Gda
668: \Gad \Gdd \Ggd
669: \Gdg
670: \put(12,12){\makebox(0,0){$3$}}
671: \Gdaad
672: \put(15,25){\makebox(0,0){$5$}}
673: \Gaddd
674: \put(45,25){\makebox(0,0){$6$}}
675: \Gddgd
676: \put(48,48){\makebox(0,0){$4$}}
677: \Ggddg
678: \epi
679: \capt{The graph $G\sm e_7$ and spanning tree $T(G\sm e_7)$}\label{Gd}
680: \efi
681:
682:
683: Returning to our example graph (which satisfies the conditions of the
684: previous theorem), $G\sm e_7$ and the tree for the induced order are
685: shown in Figure~\ref{Gd}. The relevant sets are
686: \bea
687: \{\th_e\}
688: &=&
689: \{x_3,\ x_4+x_1+x_2,\ x_5+x_1,\ x_6+x_1+x_2\},
690: \\
691: \{C\}
692: &=&
693: \{\{1,4,5,6\},\ \{2,4,6\},\ \{1,2,5\}\},
694: \\
695: \{\bx^\Cb\}
696: &=&
697: \{x_4 x_5 x_6,\ x_4 x_6,\ x_2 x_5\},
698: \\
699: \{p_\Cb\}
700: &=&
701: \{x_1(x_1+x_2)^2,\ (x_1+x_2)^2,\ x_1 x_2\},
702: \\
703: U(G\sm e_7)
704: &=&
705: U(x_1^3,\ x_2^2,\ x_1 x_2),
706: \\
707: L(G\sm e_7)
708: &=&
709: \Mon(2)-U(G\sm e_7)
710: \\
711: &=&
712: \{1,\ x_1,\ x_2,\ x_1^2\}.
713: \eea
714:
715: Making the same computations in $G/e_7$ yields
716: \bea
717: \{\th_e\}
718: &\hs{-10pt} = \hs{-10pt}&
719: \{x_4+x_1+x_2,\ x_5+x_1+x_3,\ x_6+x_1+x_2+x_3\},
720: \\
721: \{C\}
722: &\hs{-10pt} = \hs{-10pt}&
723: \{\{1,4,5,6\},\ \{2,4,6\},\ \{3,5,6\},\ \{1,2,5\},\ \{1,3,4\},\
724: \{2,3,4,5\},\ \{1,2,3,6\}\},
725: \\
726: \{\bx^\Cb\}
727: &\hs{-10pt} = \hs{-10pt}&
728: \{x_4 x_5 x_6,\ x_4 x_6,\ x_5 x_6,\ x_2 x_5,\ x_3 x_4,\ x_3x_4x_5,\ x_2x_3x_6\},
729: \\
730: \{p_\Cb\}
731: &\hs{-10pt} = \hs{-10pt}&
732: \{(x_1+x_2)(x_1+x_3)(x_1+x_2+x_3),\
733: (x_1+x_2)(x_1+x_2+x_3), (x_1+x_3)(x_1+x_2+x_3),\
734: \\
735: &&
736: \hs{5pt} x_2(x_1+x_3),\ x_3(x_1+x_2),\ x_3(x_1+x_2)(x_1+x_3),\ x_2 x_3(x_1+x_2+x_3)\},
737: \\
738: U(G/e_7)
739: &\hs{-10pt} = \hs{-10pt}&
740: U(x_1^3,\ x_2^2,\ x_3^2,\ x_1 x_2,\ x_1 x_3,\ x_1^2 x_3,\ x_1 x_2 x_3),
741: \\
742: L(G/e_7)
743: &\hs{-10pt} = \hs{-10pt}&
744: \Mon(3)-U(G/e_7)
745: \\
746: &\hs{-10pt} = \hs{-10pt}&
747: \{1,\ x_1,\ x_2,\ x_3,\ x_1^2,\ x_2 x_3 \}.
748: \eea
749: Note that we have $L(G)=L(G\sm e_7)\uplus x_3 L(G/e_7)$.
750:
751: \thicklines
752: \setlength{\unitlength}{2pt}
753: \bfi
754: \bpi(100,70)(-30,0)
755: \put(-20,30){\makebox(0,0){$G/e_7=$}}
756: \Gad \Gdd \Ggd
757: \Gdg
758: \put(30,5){\makebox(0,0){$3$}}
759: \put(30,30){\oval(60,40)[b]}
760: \put(15,25){\makebox(0,0){$5$}}
761: \Gaddd
762: \put(12,48){\makebox(0,0){$1$}}
763: \Gaddg
764: \put(45,25){\makebox(0,0){$6$}}
765: \Gddgd
766: \put(33,43){\makebox(0,0){$2$}}
767: \Gdddg
768: \put(48,48){\makebox(0,0){$4$}}
769: \Ggddg
770: \epi
771: \hspace{50pt}
772: \bpi(100,70)(-30,0)
773: \put(-20,30){\makebox(0,0){$T(G/e_7)=$}}
774: \Gad \Gdd \Ggd
775: \Gdg
776: \put(15,25){\makebox(0,0){$5$}}
777: \Gaddd
778: \put(45,25){\makebox(0,0){$6$}}
779: \Gddgd
780: \put(48,48){\makebox(0,0){$4$}}
781: \Ggddg
782: \epi
783: \capt{The graph $G/e_7$ and spanning tree $T(G/e_7)$}\label{Gc}
784: \efi
785:
786:
787: \section{Two families}
788: \label{tf}
789:
790: We will now consider two families of graphs and prove that they have
791: NBC bases. They are called (generalized) theta and phi
792: graphs.
793:
794: A {\it (generalized) theta graph\/} consists of two vertices $u,v$
795: together with $t$ internally-disjoint
796: $u\hor v$ paths $P',P'',\ldots,P^{(t)}$. Note that we are not insisting
797: that $t=3$ as is usually done for theta graphs.
798: To show that such a
799: graph has an NBC basis, we need to label its edges so that
800: $e_1<e_2<\ldots<e_q$ is a standard order.
801: First label all the edges in paths of length one with
802: $e_k,e_{k-1},\ldots,e_{l+1}$ for some $l\le k$. Now take any remaining
803: path of length at least two and label its edges, starting from the one
804: containing $u$, as $e_l,e_q,e_{q-1},e_{q-2},\ldots,e_{r+1}$ for some
805: $r$. Now take another path of length at least two (if any) and label
806: its edges $e_{l-1},e_r,e_{r-1},e_{r-2},\ldots,e_{s+1}$. Continue in
807: this way until all
808: the edges have been labeled. Note that this labeling does produce
809: a standard ordering and will be called a {\it theta labeling}.
810:
811:
812: \bth If $G$ is a (generalized) theta graph with a theta labeling then
813: $G$ has an NBC basis.
814: \eth
815: \pf\
816: We will induct on the number of edges of $G$. If $G$ is a single path
817: or if all paths are of length one, then the result is easy to verify.
818: So we may assume that $G$ is a block and has at least one $u\hor v$
819: path $P'$ of length two or greater
820:
821:
822: Let $w$ be the vertex on $P'$ adjacent to $u$. Then we have set
823: things up so that $w$ satisfies the hypotheses of Theorem~\ref{deg2}
824: except that $w$ is adjacent to $e_q$ and $e_l$ for some $l\le k$, not
825: necessarily $e_k$ itself. But the reason we chose $e_k$ in the
826: proof of the theorem was because $k$ was the largest index outside
827: $T(G)$. This guaranteed that for each circuit $C$, the monomials
828: $m_\Cb$ picked from the $p_\Cb$ in $G$, $G\sm e_q$, and $G/e_q$ would be
829: related in the correct way. And the reason for this was that given
830: any edge $e$ of $G$ which was both in a circuit and in $T(G)$, the
831: cocircuit $D_e$ would contain an edge of index smaller than $k$ and so
832: $x_k$ would not be picked from that factor. But because of the way we
833: have chosen to label the $u\hor v$ paths of length one, the preceeding
834: statements also hold if one replaces $e_k$ by $e_l$ everywhere. So
835: this change in index does no harm and will permit us to use induction,
836: as a theta labeling of $G$ will induce theta labelings of $G\sm e_q$
837: and $G/e_q$.
838:
839: Now consider $G\setm e_q$.
840: This is not a theta graph in general.
841: But the induced labeling on $G\sm e_q$ is a theta
842: labelling if we ignore the other edges on $P'$. This does not cause
843: any problems since each of these edges is now a block and so does not
844: contribute anything to $F(G)$ by~\ree{iso} and the fact that
845: $R(e)\iso F$ for any edge $e$. Hence, by induction, $R(G\sm e_q)$ has an
846: NBC basis.
847:
848:
849: Now look at $G/e_q$. This is still a theta graph and, since $P'$ has
850: length at least two, its induced
851: labeling is a theta labeling. So, by induction, $R(G/e_q)$ has an NBC
852: basis. Hence all the hypotheses of Theorem~\ref{deg2} are
853: satisfied and $G$ has an NBC basis, completing our proof.
854: \Qqed
855:
856: As a special case of the previous result, we obtain the following.
857: \bco
858: The complete bipartite graph $K_{2,t}$ with a theta labeling has an
859: NBC basis.\qqed
860: \eco
861:
862: Rather than thinking of theta graphs as unions of paths, one could
863: consider them as a set of cycles joined in parallel. We will now
864: define a family of graphs which can be thought of as joining cycles in
865: series. Suppose we are given $t$ cycles $C',C'',\ldots,C^{(t)}$ all
866: of length at least two, and in
867: each $C^{(i)}$ we are given a pair of distinguished edges
868: $e^{(i)},f^{(i)}$. Then
869: the associated {\it phi graph\/} is obtained by identifying $f^{(i)}$ with
870: $e^{(i+1)}$ for $1\le i< t$.
871: For example, if we let $P_p$ denote the
872: path on $p$ vertices then the cross product $P_2\times P_t$ is a phi
873: graph where all the cycles have length four.
874: (It is because of the shape of $P_2\times P_3$ that we call these phi graphs.)
875:
876: Again, we will need a specific labeling for our phi graphs. Label
877: edge $e^{(i)}$ with $e_{k-i+1}$, $1\le i\le t$. Now label the remaining
878: edges of $C^{(1)}$ as follows. We have
879: $C^{(1)}-e^{(1)}-f^{(1)}=P\uplus Q$ where $P, Q$ are paths. Label the
880: edges along $P$ (if any) starting with the one adjacent to $e^{(1)}$
881: with $e_q,e_{q-1},\ldots,e_{r+1}$. Now do the same along $Q$ using
882: the labels $e_r,e_{r-1},\ldots,e_s$. Continue in like manner to label
883: the rest of the cycles. (When one gets to the last one, there will be
884: only one path to label.) Call this a {\it phi labeling\/} of the
885: graph.
886:
887: Before proving that a phi graph has an NBC basis, we will
888: need a lemma to take care of the special case when the first cycle has
889: length two, so that attaching it to the second cycle creates an edge of
890: multiplicity two.
891: Let $G$ be a connected graph with standard ordering
892: $e_1<e_2<e_3<\ldots<e_q$ where $e_k$ and $e_{k-1}$ have the same
893: endpoints. Let $G\sm e_k$ have the induced ordering
894: $e_1<\ldots<\eh_k<\ldots<e_q$ where the hat indicates that $e_k$ has
895: been removed. Note that the induced ordering is standard.
896: Then the corresponding rings are related in the manner in which one
897: would expect given that the chromatic polynomials do not change.
898: \ble
899: Suppose that $G$ has a standard ordering such that $e_k$ and $e_{k-1}$
900: have the same endpoints. If $G\sm e_k$ is given the induced ordering
901: above then $R(G)\iso R(G\sm e_k)$.
902: \ele
903: \pf
904: Directly from the definitions one sees that one obtains the generators
905: for $J(G)$ from those for
906: $J(G\sm e_k)$ by substituting $x_{k-1}+x_k$ everywhere one has an $x_{k-1}$. The
907: additional cycle made by $e_{k-1},e_k$ also sets $x_k=0$ in the quotient $R(G)$.
908: Hence the isomorphism.
909: \qqed
910:
911: \bth
912: If $G$ is a phi graph with a phi labeling then
913: $G$ has an NBC basis.
914: \eth
915: \pf
916: Again, we induct on the number of edges in $G$. The case of a single cycle
917: is easy to do (and appears in~\cite{bro:cpo}). So suppose we have at
918: least two cycles. If $C'$ has length two, then its phi labeling is
919: exactly the type considered in the previous lemma. So
920: $R(G)\iso R(G\sm e_k)$ where the latter graph has a phi labeling and
921: fewer edges. So we are done in this case.
922:
923: If $C'$ has length at least three, then a deletion-contraction
924: argument similar to the one used for theta graphs will provide a
925: proof. We leave the details of the demonstration to the reader.
926: \qqed
927:
928: \bco
929: The graph $P_2\times P_t$ with a phi labeling has an NBC basis.\Qqed
930: \eco
931:
932: \section{Upper Bounds}
933: \label{ub}
934:
935: If graph $G=(V,E)$ has an NBC basis, then we can use this fact to give a
936: simple upper bound on its $h$-vector. (Lower bounds for $h$-vectors
937: of various types of complexes have been given
938: by Swartz~\cite{swa:lbh}.) This, in turn, bounds the values
939: of the chromatic polynomial $P(G;\la)$ at negative integers since then
940: all terms in the expansion~\ree{Ph} have the same sign. In
941: particular, this gives an upper bound on $\al(G)$, the number of
942: acyclic orientations of $G$, because of a famous theorem of
943: Stanley~\cite{sta:aog} which states that
944: $$
945: \al(G)=(-1)^p P(G;-1)
946: $$
947: where, as usual, $p=|V|$. To see why one could only expect to bound
948: these quantities, rather than obtaining their exact values, we need to
949: say a few words about the theory of \#P problems which was introduced by
950: Valiant~\cite{val:ccp,val:cer}.
951:
952: If $A$ and $B$ are two problems then we say that $A$ is {\it
953: polynomially reducible\/} to $B$ if it is possible, given a subroutine
954: to solve $B$, to solve $A$ in polynomial time, where we count calls to
955: the subroutine for $B$ as a single step.
956: The {\it class \#P\/} consists of those enumeration problems where the
957: structures being counted can be recognized in polynomial time. In
958: other words, there is an algorithm which is polynomial in the size
959: of the input problem that can verify whether a given structure should
960: be included in the count. So the class \#P is to enumeration
961: problems as the class NP is to decision problems. An enumeration
962: problem is {\it \#P-complete\/} if any problem in \#P is polynomially
963: reducible to it. So the \#P-complete problems are the hardest in
964: \#P.
965:
966: Linial~\cite{lin:hep} first showed that computing $\al(G)$ is
967: \#P-complete. Jaeger, Vertigan, and Welsh~\cite{jvw:ccj} derived more
968: general results about computing the Tutte polynomial of a matroid
969: which imply that computing $P(G;\la)$ is \#P-complete for all but
970: nine special values of $\la$.
971:
972: The case $\la=-1$ has attracted special interest because $\log\al(G)$
973: is a lower bound on the computational complexity of certain decision
974: and sorting problems, see for example the paper of Goddard, Kenyon,
975: King, and Schulman~\cite{gkks:ora}. Obviously the number of acyclic
976: orientations of $G$ is bounded above by the total number of
977: orientations, giving
978: $$
979: \al(G)\le 2^q
980: $$
981: where $q=|E|$. Fredman (whose work is reported in a paper of Graham,
982: Yao, and Yao~\cite[Section 7]{gyy:ibw}), and independently Manber and
983: Tompa~\cite{mt:enh} gave the first nonobvious upper bound for $\al(G)$ as
984: $$
985: \al(G)\le \prod_{v\in V} (\deg v+1),
986: $$
987: where, as usual, $\deg v$ is the degree of vertex $v$. This bound was improved by
988: Kahale and Schulman~\cite{ks:bcp} as follows.
989:
990: Given a graph $G$, consider its {\it cone\/}, $G^*$, obtained by
991: adding a new vertex adjacent to every vertex of $G$. Then Kahale and
992: Schulman show that $\al(G)$ is at most the number of spanning trees of
993: $G^*$. Using the Matrix-Tree Theorem, this bound can be expressed as
994: a determinant. Since the determinant itself could be costly to
995: compute, they give an upper bound for its value.
996: \bth[\cite{ks:bcp}]
997: \label{KS}
998: We have the upper bound
999: \beq
1000: \label{be}
1001: \al(G)\le\prod_{v\in V} (\deg v+1)
1002: \prod_{uw\in E} \exp\frac{-1}{2(\deg u+1)(\deg w+1)}
1003: \stackrel{\rm def}{=}\be(G).\qqed
1004: \eeq
1005: \eth
1006:
1007: Now suppose that $G$ has an NBC basis
1008: $\Mon(k)-U(m_\Cb\ :\ C\in\cC(G))$. If we remove the upper order
1009: ideal generated by just the fundamental circuits, then we will get a
1010: spanning set for the quotient which can be used to bound the
1011: $h$-vector from above. Furthermore, each of these monomials
1012: has the simple form
1013: $$
1014: m_{\Cb_i}=x_i^{|C_i|-1}.
1015: $$
1016: So by Theorem~\ref{mac} and equation~\ree{Ph}, we have proved the
1017: following result, where we use
1018: $L_d(S)$ to denote the set of monomials in the lower order ideal
1019: $L(S)$ which have total degree $d$.
1020: \bth
1021: If $G$ has an NBC basis with fundamental circuits $C_1,\ldots,C_k$ then,
1022: for $d\ge0$,
1023: \beq
1024: \label{ld}
1025: h_d(G)\le \left|L_d\left(x_1^{|C_1|-2}\cdots x_k^{|C_k|-2}\right)\right|
1026: \stackrel{\rm def}{=}l_d(G).
1027: \eeq
1028: Furthermore
1029: \beq
1030: \label{ga}
1031: \al(G)\le\sum_{d=0}^{p-1} l_d(G) 2^{p-d-1}
1032: \stackrel{\rm def}{=}\ga(G).\qqed
1033: \eeq
1034: \eth
1035: We note that it is an easy exercise to show that
1036: \beq
1037: \label{ch}
1038: l_d(G)\le|L_d(\Mon(k))|={d+k-1\choose k-1}.
1039: \eeq
1040: If one wishes, one can calculate the exact values of the $l_d(G)$
1041: using the Principle of Inclusion-Exclusion (see Stanley's
1042: text~\cite[Chapter 2]{sta:ec1}).
1043:
1044: We will now compare the bounds $\be(G)$ and $\ga(G)$ for certain theta and phi
1045: graphs. When possible, we will compare the $\ga$ bound with the
1046: actual number of acyclic orientations. Of course, from a practical
1047: viewpoint, it is unnecessary to use a bound when the exact value is
1048: known. But this will give some sense of how close $\ga$ is to the
1049: truth.
1050:
1051: We keep the conventions of the previous section.
1052: Define $\Th_{n,t}$ to be the theta graph consisting of $t$
1053: paths of length $n$ with their endpoints identified to form the
1054: special vertices $u$ and $v$. There is an interesting change in the
1055: behaviour of the $\ga$ bound depending on whether $n$ is held fixed and
1056: $t$ varies, or vice-versa.
1057: \bth
1058: As $n\ra\infty$ we have
1059: $$
1060: \ga(\Th_{n,3})\sim\al(\Th_{n,3}).
1061: $$
1062: As $t\ra\infty$ we have
1063: $$
1064: \be(\Th_{2,t})=o(\ga(\Th_{2,t})).
1065: $$
1066: \eth
1067: \pf\
1068: First consider $\Th_{n,3}$ where $p=3n-1$ and $q=3n$. Since this
1069: graph only has 3 circuits, it is easy to use Inclusion-Exclusion to
1070: calculate $\al(G)$, from which one sees that the count is asymptotic
1071: to the first term
1072: $$
1073: \al(G)\sim 2^q=2^{3n}.
1074: $$
1075:
1076: To compute $\ga$, first note that from~\ree{ld} and~\ree{ch} we have
1077: $$
1078: h_d(\Th_{n,3})\le l_d(x_1^{2n-2}\ x_2^{2n-2})\le d+1.
1079: $$
1080: Plugging this bound into~\ree{ga} gives
1081: $$
1082: \ga(\Th_{n,3})
1083: \le
1084: \sum_{d\ge0} (d+1) 2^{3n-2-d}
1085: =
1086: 2^{3n-2}\cdot\frac{1}{(1-1/2)^2}
1087: =
1088: 2^{3n}.
1089: $$
1090: So we must also have $\ga(\Th_{n,3})\sim 2^{3n}$ since $\ga$ is an
1091: upper bound.
1092:
1093: For $\Th_{2,t}$ note that $k$, the number of edges not in a spanning
1094: tree, satisfies $k=t-1$. We also have $p=t+2$ and $q=2t$.
1095: Using~\ree{ld}, we get
1096: $$
1097: l(d,t)\stackrel{\rm def}{=}l_d(\Th_{2,t})=
1098: \left|L_d\left(x_1^2 x_2^2\cdots x_{t-1}^2\right)\right|
1099: $$
1100: which is the coefficient of $y^d$ in the expansion of
1101: the generating function $(1+y+y^2)^{t-1}$. From this, it follows that
1102: the $l(d,t)$ satisfy the recursion
1103: \beq
1104: \label{rr}
1105: l(d,t+1)=l(d,t)+l(d-1,t)+l(d-2,t).
1106: \eeq
1107: Let $\ga_t=\ga(\Th_{2,t})$. So multiplying~\ree{rr} by $2^{t+2-d}$
1108: and summing over $0\le d\le t+2$, we can use~\ree{ga} to get the
1109: following equation, with the three expressions in brackets coming from
1110: the three terms of the recursion (respectively):
1111: \beq
1112: \label{gat}
1113: \ga_{t+1}=
1114: \left[2\ga_t + l(t+2,t) \right]+\left[\ga_t\right]
1115: +\left[\frac{1}{2}\ga_t - \frac{1}{2} l(t+1,t)\right]
1116: >\frac{7}{2}\ga_t - \frac{1}{4} \ga_t = \frac{13}{4}\ga_t
1117: \eeq
1118: where the inequality follows by noting $4 l(t-1,t)$ is a summand in
1119: $\ga_t$ and that, as provable from generating function, the sequence
1120: $(l(d,t))_{0\le d\le 2t-2}$ is symmetric and unimodal with maximum at $l(t-1,t)$.
1121:
1122: Finally, combining the estimates in~\ree{be} and~\ree{gat}, we see that
1123: for any $0<\ep<1/4$,
1124: $$
1125: \be(\Th_{2,t})=(t+1)^2 3^t \exp\frac{-2t}{6t+6}=o((13/4-\ep)^t)=
1126: o(\ga(\Th_{2,t}))
1127: $$
1128: as desired.
1129: \qqed
1130:
1131: Now for $n\ge4$, let $\Phi_{n,t}$ be a phi graph derived by pasting
1132: together $t$
1133: cycles of length $n$ in such a way that each cycle only intersects the
1134: cycle just preceding and the cycle just following it (if any). Note
1135: that $\Phi_{n,t}$ is actually a graph
1136: family since one can get a number of graphs with these
1137: specifications by pasting along different edges. But they all have a
1138: uniform description of their NBC bases and degree sequences, so the
1139: bounds under consideration will apply to any graph of the family.
1140:
1141: \bth
1142: As $n\ra\infty$ we have
1143: $$
1144: \ga(\Phi_{n,2})\sim\al(\Phi_{n,2}).
1145: $$
1146: As $t\ra\infty$ we have
1147: $$
1148: \ga(\Phi_{4,t})=o(\be(\Phi_{4,t})).
1149: $$
1150: \eth
1151: \pf\
1152: The proof for $\Phi_{n,2}$ is completely analogous to the proof given
1153: for $\Th_{n,3}$, so we leave it to the reader.
1154:
1155: Now considering $P_2\times P_{t+1}$ or any other member of $\Phi_{4,t}$,
1156: we see that $p=2t+2$, $q=3t+1$, and $k=t$.
1157: Using the bound~\ree{ch} and the Binomial Theorem
1158: in~\ree{ga} yields
1159: $$
1160: \ga(\Phi_{4,t})\le\sum_{d=0}^{2t+1} {d+t-1\choose t-1} 2^{2t+1-d}
1161: \le 2^{2t+1}\sum_{d=0}^{\infty} {d+t-1\choose t-1} 2^{-d}
1162: = 2^{2t+1}\frac{1}{(1-1/2)^t}=2\cdot 8^t.
1163: $$
1164: Now~\ree{be} gives
1165: $$
1166: \be(\Phi_{4,t})\sim a\cdot b^t,\quad b\approx 14.5682
1167: $$
1168: finishing the proof of the theorem.\qqed
1169:
1170: \section{Comments and Open Problems}
1171: \label{cop}
1172:
1173: \subsection{Arbitrary fields}
1174:
1175: We will now indicate how to generalize our construction to an
1176: arbitrary field. We first need to review what Brown's hsop looks like
1177: over a field $F$. Fix an orientation of $E(G)$. Also, for each $e_j\in
1178: T(G)$, orient all the edges of $D_j$ in one of the two possible
1179: directions. Now define signs
1180: $$
1181: \ep_{i,j}=
1182: \case{1}{if the orientation of $e_i$ in $G$ is the same as in $D_j$,}
1183: {-1}{if these orientations are opposite.}
1184: $$
1185: We have corresponding polynomials
1186: \beq
1187: \label{thF}
1188: \th_j=\sum_{e_i\in D_j} \ep_{i,j} x_i.
1189: \eeq
1190: \bth[\cite{bro:cpo}]
1191: If $G$ is a connected graph then
1192: the set of polynomials defined by~\ree{thF} for $e\in T(G)$ is an hsop for
1193: $F(G)$.\qqed
1194: \eth
1195:
1196: Solving for $x_j$ in the equation for $\th_j$ and plugging into the
1197: monomials $\bx^\Cb$, $C\in\cC(G)$, gives the generators $p_\Cb$ for an ideal
1198: $J(G)$ such that
1199: $$
1200: R(G)\iso \frac{F[x_1,\ldots,x_k]}{J(G)}.
1201: $$
1202: Note that the monomial $m_\Cb$ that was chosen from the expansion of
1203: $p_\Cb$ in the case $F=\bbZ_2$ will also appear with coefficient
1204: $\pm1$ for any field. So the proof of Theorem~\ref{deg2} will go
1205: through as before as long as the generators of $J(G)$, $J(G\sm e_q)$,
1206: and $J(G/e_q)$ can be related in the correct way.
1207:
1208: An orientation of $G$ induces orientations of $G\sm e_q$ and $G/e_q$
1209: merely by keeping each $e_i$, $i<q$, oriented the same way in all
1210: three graphs. Under the assumptions of Theorem~\ref{deg2} we showed
1211: that $D_j(G\sm e_q)=D_j(G)-e_k$ for $j>k$. So we can orient
1212: $D_j(G\sm e_q)$ the same way as $D(G)$ in this case. We also have
1213: $D_k(G\sm e_q) = \{e_k\}$, so it does not matter which way we orient
1214: $e_k$ in this cut set as $x_k$ is being set to zero in the quotient.
1215: Thus we get, as we did in the $\bbZ_2$ proof, that the generators for
1216: $J(G\sm e_q)$ are gotten from those for $J(G)$ by setting $x_k=0$.
1217: Similar considerations show that we can define orientations on the cut
1218: sets of $G/e_q$ so that the equalities we had before still hold. So
1219: Theorem~\ref{deg2} holds, and hence so do all the rest of the results of the
1220: previous sections, over any field.
1221:
1222:
1223:
1224: \subsection{Arbitrary graphs}
1225:
1226: We conjecture that any graph $G$, with its edge set suitably ordered,
1227: has an NBC basis.
1228: \bcon
1229: \label{C}
1230: Let $G$ be any graph. Then there is a standard ordering of
1231: $E(G)$ such that $L(G)$ is a basis for $R(G)$.
1232: \econ
1233:
1234: We will now outline a possible line of attack on Conjecture~\ref{C}.
1235: Even though we have not been able to push it through, it is possible
1236: that some of these ideas will be useful in finally proving or
1237: disproving this conjecture.
1238: Recall that it suffices to find a proof when $G$ is a block.
1239: But any block other than $K_2$ (the complete graph on 2 vertices)
1240: has a nice recursive structure in that it can be built
1241: from a cycle by adding a sequence of paths called {\it ears\/}. This
1242: result is due to Whitney~\cite{whi:cgc}. Proofs can also be found in
1243: the books of Diestel~\cite[Proposition 3.1.2]{die:gt} and
1244: West~\cite[Theorem 4.2.8]{wes:igt}.
1245: \bth[Ear Decomposition Theorem]
1246: Suppose $G\neq K_2$. Then
1247: $G$ is a block if and only if there is a sequence
1248: $$
1249: G_0, G_1,\ldots, G_l=G
1250: $$
1251: such that $G_0$ is a cycle and $G_{i+1}$ is obtained by
1252: taking a nontrivial path and identifying its two endpoints with two
1253: distinct vertices of $G_i$. \qqed
1254: \eth
1255:
1256: Note that the graph $G_1$ in the ear decomposition sequence is a theta
1257: graph. So one might try to prove Conjecture~\ref{C} by induction on
1258: $l$, the number of paths added. (Actually, one also needs to induct on the
1259: number of edges since one contracts an edge and not a whole path.) In
1260: fact, the induction step goes
1261: through in much the same way as our proof for theta graphs as long as
1262: the path added has length at least two. The difficulty comes if the
1263: path is a single edge. In that case, it is still easy to relate the
1264: circuits of $G\sm e_q$, where $e_q$ is the newly added edge, to those
1265: of $G$. But the situation is much more complicated in $G/e_q$, which
1266: may not even be a block. So a more delicate analysis is needed.
1267: Unfortunately, there are graphs (such as the complete graphs) where every ear
1268: decomposition requires the addition of a single edge at some stage.
1269:
1270:
1271: \subsection{Not quite arbitrary matroids}
1272:
1273: As a last point, the reader may have noticed that all of the graphical
1274: definitions we used to define NBC bases make sense for the broken
1275: cirucit complex of an arbitrary
1276: matroid. So a natural question is whether our construction goes through in that
1277: level of generality. Brown, Colbourn, and Wagner~\cite{bcw:cmr} have
1278: a way of producing an hsop for any representable matroid. (Actually,
1279: their construction is of an hsop for the independence complex of the
1280: matroid. But this will also give an hsop for the broken circuit
1281: complex since it is a subcomplex of the independence complex having the
1282: same rank.) So this
1283: would be the natural class of matroids in which to look for NBC bases.
1284:
1285: \subsection{Gr\"obner bases}
1286:
1287: We note that, in general, the monomials used to generate $U(G)$ are not the leading
1288: terms of a Gr\"obner basis for the ideal $J(G)$.
1289: As an example of this, one can take a theta graph
1290: consisting of three paths of length two in the theta labeling as
1291: described in Section~\ref{tf}. Then by choosing a suitable
1292: orientation for $G$ and its cocircuits, $J(G)$ will have generators
1293: \beq
1294: \label{gen}
1295: \{p_\Cb\}=\{x_1(x_1+x_2)^2,\ x_2(x_1+x_2)^2,\ x_1x_2^2\}
1296: \eeq
1297: from which we pick monomials
1298: \beq
1299: \label{mon}
1300: \{m_\Cb\}=\{x_1^3,\ x_2^3,\ x_1x_2^2\}
1301: \eeq
1302: for the NBC basis.
1303:
1304: Suppose, towards a contradiction, that there is a term ordering
1305: giving~\ree{mon} as the set of leading terms of a Gr\"obner basis. Then in
1306: that term ordering we either have $x_1<x_2$ or $x_1>x_2$. Suppose the
1307: former is true. Then $x_1^3$ is the smallest (monic) polynomial which
1308: is homogeneous of degree three. Also, the generators of $J(G)$ are
1309: homogeneous. So if $x_1^3$ were a leading term of a polynomial in
1310: $J(G)$ then, in fact, $x_1^3\in J(G)$. But it is easy to check that
1311: $x_1^3\not\in J(G)$ since it is not a linear combination of the
1312: polynomials in~\ree{gen}. (It suffices to consider linear
1313: combinations by homogeneity.) One gets a similar contradiction using
1314: $x_2^3$ if one assumes that $x_1>x_2$. So no such Gr\"obner basis can exist.
1315:
1316:
1317:
1318: \medskip
1319:
1320: {\it Acknowledgement.\/} We would like to thank David Forge for
1321: stimulating discussions.
1322:
1323: \bigskip
1324: \bibliographystyle{plain}
1325: %\begin{small}
1326: \bibliography{ref}
1327: %\end{small}
1328:
1329:
1330: \end{document}
1331:
1332: One might hope that every graph had a standard ordering such that the
1333: resulting ideal $L(G)$ would be a basis of $R(G)$. Unfortunately,
1334: that is not true. If one considers the graph in Figure~\ref{c-e} then
1335: it is easy, if tedious, to show that none of the possible standard
1336: orderings gives a basis, even over $\bbZ_2$. However, one can obtain a basis
1337: for $R(G)$ in this case by generalizing the construction for $L(G)$.
1338:
1339: \thicklines
1340: \setlength{\unitlength}{2pt}
1341: \bfi
1342: \bpi(80,70)(-10,-10)
1343: \Gba \Gfa
1344: \Gae \Gge
1345: \Gdg
1346: \put(30,-5){\makebox(0,0){$3$}}
1347: \Gbafa
1348: \put(2,20){\makebox(0,0){$2$}}
1349: \Gbaae
1350: \put(20,20){\makebox(0,0){$6$}}
1351: \Gbadg
1352: \put(58,20){\makebox(0,0){$4$}}
1353: \Gfadg
1354: \put(40,20){\makebox(0,0){$7$}}
1355: \Gfage
1356: \put(30,35){\makebox(0,0){$1$}}
1357: \Gaege
1358: \put(15,55){\makebox(0,0){$5$}}
1359: \Gaedg
1360: \put(45,55){\makebox(0,0){$8$}}
1361: \Ggedg
1362: \epi
1363: \capt{An ordering of a graph $G$}\label{c-e}
1364: \efi
1365:
1366:
1367: In generalizing the construction of the hsop from $\bbZ_2$ to any $F$
1368: we needed to consider more than one orientation on the edges of $G$.
1369: Similarly, to get a basis for $R(G)$ in our example graph we will
1370: need to consider two different orderings of the edge set. The first
1371: one, used to define the broken circuits, will be the usual
1372: $e_1<\ldots<e_q$. But now take any ordering of the first $k$ edges, say,
1373: $$
1374: e_{j_1},\ e_{j_2},\ \ldots,\ e_{j_k}
1375: $$
1376: and use it to define the $m_\Cb$ by picking from each factor the
1377: variable which comes first in the new ordering.
1378: Now define $U(G)$, and $L(G)$ as before except with respect
1379: to these new monomials. We will call a graph equipped with these two
1380: orderings a {\it doubly ordered
1381: graph\/}.
1382:
1383:
1384: Returning to the graph in Figure~\ref{c-e}, consider the given ordering on
1385: the edges together with the order
1386: \beq
1387: \label{ord}
1388: e_4,\ e_3,\ e_2,\ e_1.
1389: \eeq
1390: Computing the hsop over $\bbZ_2$ gives
1391: \bea
1392: \th_5
1393: &=&
1394: x_5+x_1+x_2,
1395: \\
1396: \th_6
1397: &=&
1398: x_6+x_2+x_3,
1399: \\
1400: \th_7
1401: &=&
1402: x_7+x_3+x_4,
1403: \\
1404: \th_8
1405: &=&
1406: x_8+x_1+x_4.
1407: \eea
1408: So, for example, if $C=\{2,3,5,7\}$ then the monomial chosen from
1409: $p_\Cb=x_3(x_1+x_2)(x_3+x_4)$ using~\ree{ord} will be
1410: $m_\Cb=x_2 x_3 x_4$. The rest of the $m_\Cb$ are computed similarly
1411: and one checks easily that the resultant $L(G)$ is a basis for $R(G)$.
1412:
1413: We conjecture that this can always be done.