1: \documentclass[11pt]{article}

2: \setlength{\textheight}{8.75in} \setlength{\textwidth}{6.5in}

3: \setlength{\topmargin}{0.0in} \hoffset= - 0.75 in

4: 

5: \usepackage{amsfonts, amsmath, amssymb, amsthm}

6: \usepackage[xdvi]{graphics}

7: \usepackage{latexsym}

8: \newtheorem{thm}{Theorem}[section]

9: \newtheorem{defn}[thm]{Definition}

10: \newtheorem{claim}[thm]{Claim}

11: \newtheorem{conj}[thm]{Conjecture}

12: \newtheorem{obs}[thm]{Observation}

13: \newtheorem{lem}[thm]{Lemma}

14: \newtheorem{cor}[thm]{Corollary}

15: \newtheorem{prop}[thm]{Proposition}

16: \newtheorem{fact}[thm]{Fact}

17: \newtheorem{prob}[thm]{Problem}

18: \newtheorem{alg}[thm]{Algorithm}

19: 

20: \newtheorem{Example}[thm]{Example}

21: \newenvironment{exmp}{\begin{Example}\rm}

22:                 {\mbox{}~\hfill~\end{Example}}

23: 

24: \newcommand{\ZZ}{\mathbb{Z}}

25: \newcommand{\NN}{\mathbb{N}}

26: \newcommand{\RR}{\mathbb{R}}

27: \DeclareMathOperator{\id}{id}

28: 

29: \begin{document}

30: 

31: \title{Markov Bases of Binary Graph Models}

32: \author{Mike Develin \\ American Institute of Mathematics\\ 360 Portage Ave.

33: \\ Palo Alto, CA 94306-2244, USA \\ develin@post.harvard.edu

34: \and Seth Sullivant \\ Department of Mathematics \\ UC-Berkeley \\

35: Berkeley, CA 94720-3840, USA \\ seths@math.berkeley.edu}

36: 

37: \maketitle

38: 

39: \begin{abstract}

40: This paper is concerned with the topological invariant of a graph

41: given by the maximum degree of a Markov basis element for the

42: corresponding graph model for binary contingency tables.  We

43: describe a degree four Markov basis for the model when the

44: underlying graph is a cycle and generalize this result to the

45: complete bipartite graph $K_{2,n}$. We also give a combinatorial

46: classification of degree two and three Markov basis moves as well

47: as a Buchberger-free algorithm to compute moves of arbitrary given

48: degree.  Finally, we compute the algebraic degree of the model

49: when the underlying graph is a forest.

50: 

51: \smallskip

52: 

53: \noindent {\bf Keywords:}  Markov bases, contingency tables,

54: graphical models, hierarchical models, toric ideals.

55: \end{abstract}

56: 

57: 

58: 

59: 

60: \section{Introduction and Definitions}

61: 

62: The study of multidimensional tables and their marginals is of

63: central importance whenever one wishes to make inferences based on

64: statistical samples.  In general, one is presented with a

65: nonnegative integral table of data of size $d_1 \times \cdots

66: \times d_n$ and a simplicial complex $\Delta$ on $\{1, \ldots,

67: n\}$ which encodes the specific marginals we would like to

68: compute; this is called a hierarchical model.

69: 

70:  Certainly, the oldest example of such a model

71: is the case of computing the row and column sums of a matrix.  If

72: the matrix is a square $m \times m$ matrix and we require that all

73: the row and column sums have the same value $k$, one calls such a

74: matrix a semi-magic square with magic sum $k$. Here, our

75: simplicial complex consists of two isolated points.

76: 

77: In the more general statistical situations, each node of the

78: simplicial complex corresponds to a feature of a population sample

79: (e.g. eye color) and the levels of the table correspond to

80: different states of the feature (e.g. green, brown, hazel, etc.).

81: The faces of the simplicial complex are intended to model

82: interactions between the features.  One of the most fundamental

83: questions in statistical analysis is:  do the data appear to be

84: satisfied by a given model?  One way to test the hypothesis is to

85: compare the sample data to the maximum likelihood estimate using

86: the $\chi^2$ or $G^2$ statistic.  One problem with applying this

87: approach directly is that the data is always an integral table,

88: while the maximum likelihood estimate almost never is.  As a

89: result, there might be no integral table with the same marginals

90: which has a small $\chi^2$ statistic.  This problem is especially

91: dramatic when analyzing the large, sparse data sets which occur in

92: real world situations (e.g. census data). To remedy this

93: situation, one may attempt to decide whether or not a model fits

94: the data by comparing statistics of the data table with statistics

95: of random integral tables with the same marginals. If the

96: statistic of the table of data is exceptional one could hope to

97: conclude that the table was also exceptional. For example, if the

98: $\chi^2$ statistic of the table of data was exceptionally small,

99: one could conclude that the data did, in fact, fit the model.

100: 

101: We are now left with the problem of generating random integral

102: tables from the set of all nonnegative integral tables with fixed

103: marginals.  One solution is to perform a random walk over the set

104: of all nonnegative integral tables with given fixed marginal. Such

105: a random walk can be taken by first finding a suitable set of

106: ``moves'' (these are tables with integral entries which have all

107: their marginals equal to zero) and randomly adding moves to some

108: starting table.  It is at this point in the story that

109: computational commutative algebra enters the picture: finding such

110: a set of moves is equivalent to finding a generating set for the

111: associated toric ideal.  For a detailed introduction to the

112: connections between toric algebra and multidimensional contingency

113: tables, see \cite{DS}, \cite{HS}, and \cite{St2}.

114: 

115: In this paper, we restrict attention to tables where $d_i = 2$ for

116: all $i$ and for which the underlying simplicial complex is a

117: graph; that is, we compute only two- and one-way marginals of our

118: binary table. We refer to such models as binary graph models.

119: These are generally not the usual graphical models studied so

120: frequently in statistics, where the simplicial complex consists of

121: the cliques of the underlying graph $G$ \cite{L}.  Our notion of

122: graph model coincides with the more familiar graphical model if

123: and only if the graph has no three-cycle.  Now, we will give two

124: formal presentations of the objects of interest in this paper.

125: 

126: 

127: Let $G$ be a graph on the $n$-element vertex set $[n] := \{1,2,

128: \ldots,n\}$ with edge set $E(G)$. Denote by $Iso(G)$ the set of

129: isolated vertices of $G$.  For each edge $\{j,k\}$ of the graph

130: $G$ consider the linear transformations $\pi_{j,k}$

131: 

132: $$ \pi_{j,k} : \ZZ^{2^n} \longrightarrow \ZZ^{4}$$

133: $$  e_{i_1, \ldots i_n} \longmapsto  e_{i_j,i_k}$$

134: 

135: \noindent and for each isolated vertex $k$ of $G$ consider the

136: linear transformations $\pi_k$

137: 

138: $$ \pi_k : \ZZ^{2^n} \longrightarrow \ZZ^2 $$

139: $$ e_{i_1 \ldots i_n} \longmapsto  e_{i_k}.$$

140: 

141: \noindent  We think of the maps $\pi_{j,k}$ as computing the 2-way

142: marginal of a $2 \times \cdots \times 2$ table corresponding to

143: the edge $\{j,k\}$ and the maps $\pi_{k}$ as computing the 1-way

144: marginal of a $2 \times \cdots \times 2$ table corresponding to

145: the vertex $k$.  We define the map $\pi_G$ by taking all the

146: marginal computations induced by a given graph as

147: 

148: $$\pi_G:  \ZZ^{2^n} \longrightarrow \bigoplus_{\{j,k\} \in E(G)}

149: \ZZ^{4} \bigoplus_{k \in Iso(G)} \ZZ^{2}$$

150: 

151: 

152: $$ v \longmapsto \bigoplus_{\{j,k\} \in E(G)}

153: \pi_{j,k}(v)  \bigoplus_{k \in Iso(G)} \pi_k(v).

154: $$

155: 

156: We say that $\pi_G$ is the map which computes the marginals of a

157: $2 \times \cdots \times 2$ table according to the graph $G$.  The

158: matrix which represents this linear transformation will be denoted

159: $A_G$ and the polytope which is the convex hull of the columns of

160: $A_G$ is denoted $P_G$ where we consider the columns of $A_G$ as

161: vectors in $\RR^d$ for an appropriate $d$.  A move for $G$ is an

162: element of the integral kernel of $\pi_G$; that is, a move is an

163: integral table which does not change the $G$-marginals of a table

164: it is added to.  In general we are interested in sets of moves

165: with special properties.

166: 

167: \begin{defn}

168: A finite subset of moves $B \subset \ker_\ZZ( \pi_G)$ is called a

169: {\bf Markov basis} for the graph $G$ if for every pair of

170: nonnegative integral tables $v_1, v_2 \in \NN^{2^n}$ with the same

171: $G$-marginals $\pi_G(v_1) = \pi_G(v_2)$, there is a sequence of

172: moves $\{u_i\}_{i=1}^l \subset \pm B$ such that

173: 

174: $$

175: v_1 + \sum_{i=1}^l u_i= v_2

176: $$

177: and

178: $$ v_1 + \sum_{i =1}^j u_i  \in \NN^{2^n} \quad \mbox{ for all } 1

179: \leq j \leq l .$$

180: \end{defn}

181: 

182: 

183: There is also a much shorter and more algebraic way to arrive at

184: this definition.  Recall that $G$ is a graph on the $n$-element

185: vertex set $[n] := \{1,2, \ldots,n\}$ with edge set $E(G)$, and

186: isolated vertices $Iso(G)$. Consider the map of polynomial rings

187: 

188: $$\phi_G:  \mathbb{C}[p_{i_1,\ldots i_n} | i_j \in \{0,1\}] \longrightarrow

189: \mathbb{C}[t^{(j,k)}_{i_j, i_k}, t^{(l)}_{i_l} | \{j,k\} \in E(G)

190: \mbox{ and } l \in Iso(G)]

191: $$

192: 

193: $$ p_{i_1 \ldots i_n} \longmapsto \prod_{\{j,k\} \in E(G)}

194: t^{(j,k)}_{i_j,i_k} \prod_{l \in Iso(G)} t^{(l)}_{i_l}.$$

195: 

196: \noindent  The object of interest in this paper is the ideal

197: denoted $I_G = \ker(\phi_G)$, which we call the ideal of a binary

198: graph model.  It is a {\bf toric ideal}: a prime ideal generated

199: by monomial differences whose leading and trailing terms have

200: disjoint support.

201: 

202: Markov bases and toric ideals are connected by the following

203: fundamental theorem.

204: 

205: \begin{thm}\cite{DS} \label{thm:mbgs}

206: A finite subset of moves $B = \{u_i\}_{i=1}^l \subset

207: \ker_\ZZ(\pi_G)$ is a Markov basis for $G$ if and only if the set

208: of binomials $\{p^{u_i^+} - p^{u_i^-} \}$ is a generating set for

209: $I_G$.

210: \end{thm}

211: 

212: Here, we write $u_i = u_i^+ - u_i^i$ as the difference of two

213: positive vectors of disjoint support.  In light of Theorem

214: \ref{thm:mbgs}, we will use the expressions ``Markov basis for

215: $G$'' and ``generating set for $I_G$'' interchangeably throughout

216: this paper.  Similarly, we can interchange the words ``move'' and

217: ``binomial'' whenever we are discussing the Markov

218: bases/generating sets of $I_G$.  These definitions are best

219: illustrated by a simple example.

220: 

221: 

222: \begin{exmp}

223: Consider the graph $G$ on four nodes with two edges $\{1,2\}$ and

224: $\{2,3\}$ and one isolated vertex $4$.  The map $\phi_G$ is a map

225: from a polynomial ring in sixteen variables to a polynomial ring

226: in ten variables.  It is the map of rings

227: 

228: $$\phi_G : \mathbb{C} [p_{ijkl} | i,j,k,l \in \{0,1\}]

229: \longrightarrow \mathbb{C} [r_{ij}, s_{jk}, t_l]$$

230: $$ p_{ijkl} \longmapsto r_{ij} \cdot s_{jk} \cdot t_l. $$

231: 

232: \noindent  On the other hand, the marginal map $\pi_G$ is given by

233: a $10 \times 16$ matrix $A_G$.  It is the matrix

234: 

235: $$A_G = \left( \begin{array}{cccccccccccccccc}

236: 1 & 1 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\

237: 0 & 0 & 0 & 0 & 1 & 1 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\

238: 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 1 & 1 & 0 & 0 & 0 & 0 \\

239: 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 1 & 1 \\

240: 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 \\

241: 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 \\

242: 0 & 0 & 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 0 & 0 \\

243: 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 \\

244: 1 & 0 & 1 & 0 & 1 & 0 & 1 & 0 & 1 & 0 & 1 & 0 & 1 & 0 & 1 & 0 \\

245: 0 & 1 & 0 & 1 & 0 & 1 & 0 & 1 & 0 & 1 & 0 & 1 & 0 & 1 & 0 & 1

246: \end{array} \right). $$

247: 

248: The polytope $P_G$ in $\RR^{10}$ has dimension 6 and has 10

249: facets.  These facets are indexed naturally by the rows of $A_G$

250: and the facet defining inequalities are given by $y_i \geq 0$ with

251: one inequality for each row.  The ideal $I_G$ has a Markov basis

252: consisting of  quadratic moves.  These are

253: 

254: $$ p_{0j0l_1}p_{1j1l_2} - p_{0j1l_1}p_{1j0l_2} \mbox{  with  } j,

255: l_1, l_2 \in \{0,1\}$$

256: 

257: \noindent and

258: 

259: $$ p_{i_1 j_1 k_1 l_1} p_{i_2 j_2 k_2 l_2} - p_{i_1 j_1 k_1 l_2} p_{i_2 j_2 k_2

260: l_1} \mbox{  with  } i_1, j_1, k_1, l_1, i_2, j_2, k_2, l_2 \in

261: \{0,1\}.$$

262: 

263: \noindent  These generators are also a Gr\"obner basis with

264: respect to the reverse lexicographic term order with $p_{0\cdots

265: 0} \prec \cdots \prec p_{1 \cdots 1}$.

266: \end{exmp}

267: 

268: In general, when the underlying graph is a forest, the toric ideal

269: $I_G$ is relatively well understood.

270: 

271: \begin{thm}\label{thm:sqfr2}\cite{D, GMS, T}

272: The ideal $I_G$ is minimally generated by quadrics if and only if

273: $G$ is a forest.  In this case, the set of quadratic squarefree

274: binomials in $I_G$ forms a Gr\"obner basis with respect to the

275: reverse lexicographic term order with $p_{0 \cdots 0} \prec p_{0

276: \cdots 0 1} \prec \ldots \prec p_{1 \cdots 1}$.

277: \end{thm}

278: 

279: In this paper, we are primarily concerned with investigating

280: graphs which contain cycles. One fundamental question is to compute the

281: following invariant of a graph.

282: 

283: \begin{defn}

284: Let $G$ be a graph. The {\bf Markov width} $\mu(G)$ is the degree of the

285: largest minimal generator of the toric ideal $I_G$.

286: \end{defn}

287: 

288: Studying the Markov width of a graph is of fundamental importance

289: for statistical applications because it relates the complexity of

290: analyzing data to the complexity of the underlying graphical

291: structure.  Note that, Theorem~\ref{thm:sqfr2} states that the

292: graphs with $\mu(G)=2$ are precisely forests, which are certainly

293: topologically simple. Indeed, the Markov width of a graph $G$ is

294: topological in nature, by which we mean that $\mu(G)$ can only

295: decrease under the operations of vertex deletion and edge

296: contraction; we will show this in Section 4.  Since these

297: operations interact nicely with the toric ideals of the initial

298: and final graph, we will use the following definition throughout.

299: 

300: \begin{defn}

301: Let $G$ be a graph. By a {\bf minor} of $G$, we mean a graph $H$ which can

302: be obtained from $G$ via a sequence of edge contractions and vertex

303: deletions.

304: \end{defn}

305: 

306: This is different from the usual definition of a graph minor in

307: that we do not allow edge deletion, whose interaction with the

308: toric ideal is more complicated.

309: 

310: The rest of the paper is organized as follows.  In the next

311: section, we discuss computational results for graphs with few

312: vertices and suggest some conjectures based upon these data.  In

313: particular, we have computed the Markov width $\mu(G)$ for all

314: graphs on five vertices and many of the graphs on six vertices. In

315: the third section, we prove that the $n$-cycle and $K_{2,n}$ have

316: Markov width 4. Furthermore, we are able to explicitly describe

317: the moves needed in the Markov bases for these graph. The fourth

318: section is devoted to the inverse problem: that is, studying which

319: graphs may have Markov basis elements of a given degree.  We give

320: an algorithm which does not depend on computing $S$-pairs for

321: computing all the minimal generators of a given degree for the

322: ideals $I_G$.  As a consequence of our algorithm, we give a

323: combinatorial characterization of moves of degree two and three.

324: In the final section, we return to the study of forests and use

325: the reverse lexicographic Gr\"obner basis from Theorem

326: \ref{thm:sqfr2} to derive combinatorial formulae for the algebraic

327: degree of $I_G$ whenever $G$ is a forest.

328: 

329: 

330: 

331: \section{Graphs with Few Vertices}

332: 

333: In this section we discuss and display computational results about

334: the ideals of binary graph models.  In particular, we describe

335: generating sets for the ideals $I_G$ for all graphs $G$ with fewer

336: than five vertices and all the graphs on six vertices with at most

337: eight edges.  These computational results suggest many conjectures

338: and open problems which we describe at the end of the section.

339: All of our computations were carried out using the toric Gr\"obner

340: basis program \verb"4ti2" \cite{H} and the computational algebra

341: system \verb"Macaulay 2" \cite{M2}. We limit our description to

342: graphs which cannot be ``glued'' together from smaller graphs

343: based on the following definition and theorem.

344: 

345: \begin{defn}

346: Let $(V_1, S, V_2)$ be a partition of the vertex set of a graph

347: $G$ such that

348: \begin{enumerate}

349: \item there are no edges in $G$ between $V_1$ and $V_2$ and

350: \item $S$ is either the empty set or $S$ is a common vertex or

351: edge of the induced subgraphs $G_1$ and $G_2$ with vertex sets

352: $V_1 \cup S$ and $V_2 \cup S$, respectively.

353: \end{enumerate}

354: 

355: \noindent  Then $G$ is called {\bf reducible} with decomposition

356: $(V_1, S, V_2)$.

357: \end{defn}

358: 

359: \begin{thm}\label{reducible}\cite{DS2, HS}

360: Let $G$ be a graph which is reducible and $G_1$ and $G_2$ the

361: induced subgraphs arising from the vertex decomposition.  Then

362: there is a degree preserving operation with which one can build

363: generating sets and Gr\"obner bases for $I_G$ from generating sets

364: and Gr\"obner bases of $I_{G_1}$ and $I_{G_2}$. In particular,

365: $\mu(G) = \,{\rm max}(\mu(G_1),\mu(G_2))$.

366: \end{thm}

367: 

368: There are precisely one graph on three vertices, two graphs on

369: four vertices, six graphs on five vertices, and six graphs with

370: six vertices and at most eight edges which are not reducible.  We

371: will briefly describe these graphs and their Markov bases.

372: 

373: \subsection{Three and Four Vertices}

374: 

375: The only graph on three vertices which is not reducible is the

376: complete graph $K_3$. The Markov basis of $I_{K_3}$ consists of

377: the single degree four binomial

378: $$p_{000}p_{011}p_{101}p_{110} - p_{001}p_{010}p_{100}p_{111}.$$

379: 

380: The two irreducible graphs on four vertices are $C_4$ and $K_4$.

381: The graph $C_4$ is the four cycle with $E(C_4) = \{12,23,34,14\}$.

382: The Markov basis of $C_4$ consists of eight quadrics such as

383: $$ p_{0000}p_{0101} - p_{0001}p_{0100} $$

384: and eight quartics.  The complete graph $K_4$ has Markov basis

385: consisting of 20 moves of degree four and 40 sextic binomials such

386: as

387: $$ p_{0000}^2p_{0111}p_{1011}p_{1101}p_{1110} -

388: p_{0001}p_{0010}p_{0100}p_{1000}p_{1111}^2.$$

389: 

390: \subsection{Five Vertices}

391: 

392: There are six graphs on five vertices which cannot be decomposed

393: into subgraphs. These are the graphs we denote $C_5$, $K_{2,3}$,

394: $\widetilde{K_4}$, $SP$, $BP$, and $K_5$.  The graph $C_5$ is the

395: five-cycle $E(C_5) = \{12,23,34, 45,15\}$.  Its Markov basis

396: consists of 80 quadrics and 40 quartics.  The graph $K_{2,3}$ is

397: the complete bipartite graph $E(K_{2,3}) = \{13,14,15,23,

398: 24,25\}$. Its Markov basis consists of 44 quadrics and 420

399: quartics.  The graph $\widetilde{K_4}$ is the graph obtained from

400: $K_4$ by subdividing an edge, $E(\widetilde{K_4}) =

401: \{12,15,23,24,34,35,45\}$.  The Markov basis for

402: $I_{\widetilde{K_4}}$ consists of 32 quadrics, 473 quartics, and

403: 160 sextics.  The graph $SP$ is the edge graph of the square

404: pyramid, $E(SP) = \{12,13,15,23,24,34,35,45\}$.  The Markov basis

405: of $SP$ consists of 16 quadrics, 671 quartics, and 320 sextics.

406: The graph $BP$ is the edge graph of the bipyramid over a triangle,

407: $E(BP)  =  \{12,13,15,23,  \\ 24,25,34,35,45\}$.  Its Markov basis

408: consists of 8 quadrics, 436 quartics, and 2872 sextic binomials.

409: Finally, $K_5$ is the complete graph on five vertices.  The Markov

410: basis of $K_5$ consists of 260 degree four moves, 3952 sextic

411: binomials, 846 binomials of degree eight such as

412: $$

413: p_{00000}^3p_{01111}p_{10111}p_{11011}p_{11101}p_{11110} -

414: p_{00001}p_{00010}p_{00100}p_{01000}p_{10000}p_{11111}^3 $$

415: 

416: \noindent and 480 degree ten binomials like

417: 

418: $$

419: p_{00000}^2 p_{01111}^2 p_{10001}^2 p_{10010}^2p_{10100}p_{11000}-

420: p_{00010}p_{00101}p_{01001}p_{01110}p_{10000}^4p_{10011}p_{11111}.$$

421: 

422: \subsection{Six Vertices}

423: 

424: There are a total of 29 graphs on six vertices which are not

425: reducible.  We were able to compute Markov bases for the six

426: irreducible graphs on six vertices which have at most eight edges.

427: It remains a major computational challenge to determine Markov

428: bases of the other 23 irreducible graphs on six vertices.  The six

429: irreducible graphs on six vertices with less than nine edges will

430: be denoted $C_6$, $K_{2,4}$, $G_{129}$, $G_{151}$, $G_{153}$, and

431: $G_{154}$.

432: 

433: The graph $C_6$ is the six cycle with edge set $E(C_6) =

434: \{12,23,34,45,56,16\}$.  The graph $K_{2,4}$ is the complete

435: bipartite graph with edge set $E(K_{2,4}) =

436: \{13,14,15,16,23,24,25,26\}$.  The remaining graphs do not have

437: special names:  the labels we have chosen come from \cite{RW}.

438: These four graphs have edge sets $E(G_{129}) =

439: \{12,15,23,26,34,45,56\}$, $E(G_{151}) = \{12,14,23,26,34,36,45,46

440: \}$, $E(G_{153}) = \{12,15,16,23,24,45,46,56\}$, and $E(G_{154}) =

441: \{12,14,23,25,34,36,45,56\}$.  The data regarding the Markov bases

442: of these graphs as well as all the irreducible graphs on five and

443: fewer vertices is summarized in the following table. The columns

444: are labeled by the particular irreducible graph, the rows are

445: labelled by degree of minimal generators and the table entries are

446: the number of minimal generators of a given degree.

447: 

448: $$

449: \begin{array}{r||r|r|r|r|r|r|r|r|r|r|r|r|r|r|r|}

450:  & K_3 & C_4 & K_4 & C_5 & K_{2,3} & \widetilde{K_4} & SP & BP &

451:  K_5 & C_6 & K_{2,4} & G_{129} & G_{151} & G_{153} & G_{154} \\

452:  \hline 2 & 0 & 8 & 0 & 80 & 44 & 32 & 16 & 8 & 0 & 528 & 236 & 360& 280 & 320 & 256\\

453:  \hline 4 & 1 & 8 & 20 & 40 & 420 & 473 & 671 & 436 & 260 & 160 & 11696 & 2636 & 4949 & 4149 & 7784 \\

454:  \hline 6 & 0 & 0 & 40 & 0 & 0 & 160 & 320 & 2872 & 3952 & 0& 0 & 0 & 640 & 480 & 640 \\

455:  \hline 8 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 846 & 0& 0 & 0 & 0 & 0 & 0 \\

456:  \hline 10 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 480 & 0 & 0 & 0 & 0 & 0 & 0 \\

457:  \hline \hline \mbox{Total} & 1 & 16 & 60 & 120 & 464 & 665 & 1007 & 3316

458:  & 5538 & 688 & 11932 & 2996 & 5869 & 4949 &  8680\\

459:  \hline \mu(G) & 4 & 4 & 6 & 4 & 4 & 6 & 6 & 6 & 10 & 4 & 4 & 4 & 6 & 6  & 6 \\ \hline

460: 

461: \end{array}

462: $$

463: 

464: For these graphs, all generators are in even degree.  This is not

465: true in general, however, as we will demonstrate in Section 4.

466: Theorem~\ref{thm:sqfr2} characterizes graphs with $\mu(G)=2$ as

467: forests, but the next case is already quite interesting.

468: 

469: \begin{prob}

470: Characterize those graphs with Markov width $\mu(G) = 4$.

471: \end{prob}

472: 

473: In the next section we will show that cycles and the complete

474: bipartite graphs $K_{2,n}$ have Markov bases consisting of moves

475: of degree four or less, but from the data we see that this is not

476: yet a complete characterization.

477: 

478: A natural class of graphs which one would hope to

479: understand is planar graphs.  The data above suggest the following

480: optimistic conjecture.

481: 

482: \begin{conj}\label{planar}

483: There is a universal constant $C$ such that the Markov width

484: $\mu(G) \leq C$ whenever $G$ is a planar graph.  Even stronger, $C

485: = 6$.

486: \end{conj}

487: 

488: On the other hand, the data also suggest the following

489: conjecture.

490: 

491: \begin{conj}\label{treewidth}

492: The invariant $\mu(G)$ is a function only of the tree width of

493: $G$.

494: \end{conj}

495: 

496: The tree width is a topological invariant of a graph $G$ which is

497: equal to one less than the size of the largest clique in the

498: chordal graph containing $G$  which has the smallest maximal

499: clique.  For example, forests are precisely those graphs with tree

500: width zero or one, and indeed Theorem~\ref{thm:sqfr2} tells us

501: that these graphs all have Markov width two.

502: Conjecture~\ref{treewidth} also agrees with

503: Theorem~\ref{reducible}, since the tree width of a reducible graph

504: is the maximum of the tree widths of its components.

505: 

506: While the limited information we have suggests both

507: Conjecture~\ref{planar} and Conjecture~\ref{treewidth}, they

508: cannot both be true:  there are planar graphs with arbitrarily

509: large tree width.  For example, grid graphs can have arbitrarily

510: large clique size in their minimal chordal triangulations.  This

511: suggests another research problem.

512: 

513: \begin{prob}

514: Study the binary graph model $I_G$ for the family of $m \times n$

515: grid graphs.

516: \end{prob}

517: 

518: We do know that $\mu(G)$ can be arbitrarily large.  For example,

519: for the complete graph $K_{m}$ we can construct generators of

520: large degree.

521: 

522: \begin{prop}

523: The complete graph $K_m$ with $m \geq 3$ has Markov width

524: $\mu(K_m) \geq 2m -2$.

525: \end{prop}

526: \begin{proof}

527: It suffices to show that there is a minimal generator of $K_m$ of

528: degree $2m-2$.  For this consider the binomial

529: 

530: $$ p_{\mathbf{0}}^{m-2} \prod_{i=1}^m p_{\mathbf{1} - e_i} -

531: p_{\mathbf{1}}^{m-2} \prod_{i=1}^m p_{e_i}$$

532: 

533: \noindent where $\mathbf{0}$ is the string of all zeros,

534: $\mathbf{1}$ is the string of all ones, and $e_i$ is the $i$th

535: unit vector.  Then the monomials coming from the leading and

536: trailing terms are the only monomials which have the given image

537: under $\phi_{K_m}$.  Equivalently, the corresponding tables are

538: the only two tables which have these same fixed marginals under

539: $\pi_{K_m}$.  Since the leading and trailing terms have disjoint

540: support, this binomial must appear in every Markov basis for

541: $I_{K_m}$.

542: \end{proof}

543: 

544: Of course, this bound is already not tight for $m=5$,  where it

545: yields $\mu(K_5) \ge 8$ despite the fact that $K_5$ has Markov

546: width 10. In general we suspect that $\mu(K_m)$ grows

547: exponentially in $m$.

548: 

549: 

550: \section{Cycles and Bipartite Graphs}

551: 

552: In this section we confirm the observations from the second

553: section: the ideal of the cycle and the complete bipartite graph

554: $K_{2,n}$ are generated in degrees two and four.

555: 

556: 

557: \subsection{Cycles}

558: 

559: For ease of notation, we will represent a binomial such as

560: $p_{1011}p_{1110} - p_{1111} p_{1010}$ in {\bf tableau notation}

561: as

562: 

563: \[

564: \left[

565: \begin{array}{cccccc}

566: 1 & 0 & 1 & 1 \\

567: 1 & 1 & 1 & 0 \\

568: \end{array}

569: \right] - \left[

570: \begin{array}{cccccc}

571: 1 & 1 & 1 & 1 \\

572: 1 & 0 & 1 & 0 \\

573: \end{array}

574: \right].

575: \]

576: 

577: The tableau are obtained from a binomial by recording the indices

578: of each variable which appears in the monomial, repeating indices

579: when a variable appears to a power greater than one.  We say that

580: one binomial {\bf contains} another if it does so in the Graver

581: sense; that is, $p^u - p^v$ contains $p^a-p^b$ if $a \leq u $ and

582: $b \leq v$ componentwise.

583: 

584: We first prove the following theorem bounding the degree of

585: minimal generators for the $n$-cycle.

586: 

587: \begin{thm}\label{ncycle}

588: Let $C_n$ be the $n$-cycle graph. Then $\mu(C_n)=4$, and in particular $I_{C_n}$ is generated in degrees 2

589: and 4.

590: \end{thm}

591: 

592: \begin{proof}

593: We will start with an arbitrary binomial $f$ in the ideal, and

594: express it as a linear combination of elements either of lower

595: degree or of degree at most 4.

596: 

597: Given any binomial, take one variable from each monomial such that

598: the two variables chosen agree in first and last index. Our

599: strategy will be as follows: by adding multiples of ideal elements

600: of degree 4 and less, we will eventually obtain a binomial in

601: which both monomials have the same variable. Dividing out by this

602: variable (which clearly does not affect membership in $I_G$)

603: yields a binomial of lower degree which must still be in the

604: ideal, completing the proof.

605: 

606: We now start this process. We have a variable $p_{1?\cdots ?1}$ in

607: the first term of $f$, and a variable $p_{1?\cdots ?1}$ in the

608: second term. We wish to eliminate all disagreements between these

609: indices by ``moving'' the table entries corresponding to these

610: binary strings using binomials of degree 4 or less.

611: 

612: Consider any block of disagreements, in which, without loss of

613: generality, the indices of these variables look like

614: $(\cdots?10\cdots 01?\cdots)$ and $(\cdots?11\cdots11?\cdots)$. We

615: propose to add some multiple of an ideal element of degree $g$ at

616: most 4 so that the resulting binomial contains the two variables

617: in question, and is unchanged except that some of the

618: disagreements in the block have been removed. The two sets of

619: index strings in $g$ will agree on the portion of $C_n$ outside of

620: the block in question, counting the boundary elements.

621: Essentially, we are performing a local move by changing indices on

622: a subgraph of $C_n$. Let $\tilde{f}$ represent the image of $f$ in

623: this subgraph, i.e. under the map sending $p_{\cdots?I?\cdots}$ to

624: $p_I$, where $I$ is the index substring on the block we have.

625: 

626: 

627: Continuing in this manner, by induction on the number of

628: disagreements we eventually obtain a binomial for which the same

629: variable appears in both monomials, completing the proof. We now

630: construct the ideal element $g$ which we will add a multiple of.

631: We first construct the part $\tilde{g}$ which corresponds to the

632: block in question; in the tableaux that follow, we consider only

633: the indices corresponding to this block.

634: 

635: Because this element is in the ideal $I_{C_n}$, considering the

636: marginal in the first two directions, since an element in the

637: first term of $f$ has a $10$ marginal, so must an element in the

638: second term. So $\tilde{f}$ contains

639: 

640: \[

641: \left[

642: \begin{array}{cccccc}

643: 1 & 0 & 0 & \cdots & 0 & 1 \\

644: \end{array}

645: \right] - \left[

646: \begin{array}{cccccc}

647: 1 & 1 & 1 & \cdots & 1 & 1 \\

648: 1 & 0 & ? & \cdots & ? & ? \\

649: \end{array}

650: \right] .

651: \]

652: 

653: Now, if any of the unspecified elements is 1, we let $\tilde{g}$

654: be the binomial which switches the intervening substrings, i.e.

655: something of the form

656: 

657: 

658: \[

659: \left[

660: \begin{array}{ccccc}

661: 1 & 1 & 1 & 1 & 1 \\

662: 1 & 0 & 0 & 1 & A \\

663: \end{array}

664: \right] - \left[

665: \begin{array}{ccccc}

666: 1 & 0 & 0 & 1 & 1 \\

667: 1 & 1 & 1 & 1 & A \\

668: \end{array}

669: \right],

670: \]

671: where $A$ is the remainder of the index string of the

672: element that the second term of $\tilde{f}$ contains starting with

673: $10$. We fill in the rest of both terms of $g$ as the two index

674: strings corresponding to the two variables contained in the second

675: term of $\tilde{f}$ are filled in, so that the two variables of

676: the first term of $g$ are contained in the second term of $f$.

677: Adding $g$ to $f$ then has the effect of eliminating some

678: disagreements between the two variables as desired, while leaving

679: everything unchanged outside the block in question.

680: 

681: Otherwise, all of the elements marked $?$ must be 0. We can apply

682: the same argument to the terminal string and to the other

683: binomials to obtain $g$ in all cases except where $\tilde{f}$

684: contains

685: 

686: \[

687: \left[

688: \begin{array}{ccccc}

689: 1 & 0 & \cdots & 0 & 1 \\

690: 1 & 1 & \cdots & 1 & 0 \\

691: 0 & 1 & \cdots & 1 & 1 \\

692: \end{array}

693: \right] - \left[

694: \begin{array}{ccccc}

695: 1 & 1 & \cdots & 1 & 1 \\

696: 1 & 0 & \cdots & 0 & 0 \\

697: 0 & 0 & \cdots & 0 & 1 \\

698: \end{array}

699: \right].

700: \]

701: 

702: In this case, because we have a $00$ marginal in the first two

703: coordinates of the second term of $\tilde{f}$, we must have one in

704: the first term. If that element contains any 1, by adding a

705: multiple of a binomial of degree 2 involving it and the third

706: element in the first term, and then another multiple of a binomial

707: of degree 2 involving the third element and the first element in

708: the first term, we can construct a $g$ essentially as before which

709: reduces disagreements. The only case where we cannot apply this

710: argument to this fourth element is when $\tilde{f}$ contains

711: 

712: \[

713: \left[

714: \begin{array}{ccccc}

715: 1 & 0 & \cdots & 0 & 1 \\

716: 1 & 1 & \cdots & 1 & 0 \\

717: 0 & 1 & \cdots & 1 & 1 \\

718: 0 & 0 & \cdots & 0 & 0 \\

719: \end{array}

720: \right] - \left[

721: \begin{array}{ccccc}

722: 1 & 1 & \cdots & 1 & 1 \\

723: 1 & 0 & \cdots & 0 & 0 \\

724: 0 & 0 & \cdots & 0 & 1 \\

725: 0 & 1 & \cdots & 1 & 0 \\

726: \end{array}

727: \right].

728: \]

729: 

730: In this case, we let $\tilde{g}$ be this binomial of degree 4,

731: corresponding to switching the middle substrings of all 0's and

732: all 1's. Again, we extend this to $g$ by copying the indices from

733: $f$ outside this block to the relevant variables, and we can add

734: this element to $f$ to eliminate this patch of disagreements

735: between the variables in question.  This completes the proof by

736: induction. Note that we did not use any elements of degree 3 in

737: this process, so $I_{C_n}$ is in fact generated in degrees 2 and

738: 4.

739: \end{proof}

740: 

741: This theorem not only shows that the minimal generators are all of

742: degree 2 or 4, but it also gives a complete description of these

743: generators. The degree-2 generators come from separations of the

744: graph; we will prove a general statement characterizing degree-2

745: generators of graph ideals in Section 5.  As for the minimal

746: generators of degree four, we have the following categorization.

747: 

748: \begin{thm}

749: The minimal generators of degree 4 in the graph ideal $I_{C_n}$

750: are those elements of the form

751: \[

752: \left[

753: \begin{array}{cccc}

754: A_1 & 1 & B & 1 \\

755: A_2 & 1 & 1 -B & 0 \\

756: A_3 & 0 & 1 -B & 1 \\

757: A_4 & 0 & B & 0 \\

758: \end{array}

759: \right] - \left[

760: \begin{array}{cccc}

761: A_1 & 1 & 1-B & 1 \\

762: A_2 & 1 & B & 0 \\

763: A_3 & 0 & B & 1 \\

764: A_4 & 0 & 1-B & 0 \\

765: \end{array}

766: \right],

767: \]

768: where the columns correspond to $V_1, x_1, V_2$, and $x_2$, $V_1$

769: and $V_2$ are contiguous blocks of elements, and these elements in

770: this order comprise the $n$-cycle. Here, $1 -B$ represents the

771: opposite string of $B$.

772: \end{thm}

773: 

774: Note that these generators of degree 4 are very similar to the

775: generator of degree 4 in $K_3$, namely

776: \[

777: \left[

778: \begin{array}{ccc}

779: 1 & 1 & 1 \\

780: 1 & 0 & 0 \\

781: 0 & 0 & 1 \\

782: 0 & 1 & 0 \\

783: \end{array}

784: \right] - \left[

785: \begin{array}{ccc}

786: 1 & 0 & 1 \\

787: 1 & 1 & 0 \\

788: 0 & 1 & 1 \\

789: 0 & 0 & 0 \\

790: \end{array}

791: \right].

792: \]

793: 

794: We will prove a general similarity  theorem in this vein in

795: Section 4, when we classify generators of a given degree.

796: 

797: \subsection{Complete Bipartite Graphs}

798: 

799: Another nice class of models is the $K_{m,n}$ model, where

800: $K_{m,n}$ is the complete bipartite graph with partite sets of $m$

801: and $n$ vertices. We first prove the following theorem.

802: 

803: \begin{thm}\label{k2n}

804: The graph ideal for $G=K_{2,n}$ is generated in degrees 2 and 4

805: (for $n\ge 2$).

806: \end{thm}

807: 

808: \begin{proof}

809: 

810: As in the proof of Theorem~\ref{ncycle}, we use binomials of small

811: degree ($\le 4$) to transform a binomial in this ideal to one

812: whose two terms share a variable, completing the proof by

813: induction. Let the vertices of the two-element partite set be

814: $V=\{v_1,v_2\}$, and let the vertices of the $n$-element partite

815: set be $W=\{w_1,\ldots,w_n\}$.

816: 

817: For each monomial $M$, we define the submonomial $M_{ij}$, $i,j\in

818: \{0,1\}$, to be the product of the variables with $ij$ in the

819: index string corresponding to the two-element partite set; we will

820: write that index string first. For a monomial $M$, we define

821: $a_{ij}(M)$ to be the total degree of $M_{ij}$, and

822: $b_{ij,k,l}(M)$ to be the number of appearances of the digit $k$

823: in the $w_l$-position of the index strings of the variables in

824: $M_{ij}$. Here $k\in \{0,1\}$ and $l\in \{1,\ldots,n\}$.  In other

825: words, these function values enumerates the marginal in the

826: direction $(v_1,v_2,w_l)$ with the set values $(i,j,k)$.  We can

827: of course recover $a_{ij}(M) = b_{ij,0,l} + b_{ij,1,l}$ for any

828: $l$.

829: 

830: Then we have the following easy lemma.

831: 

832: \begin{lem}\label{deg2kmn}

833: If $M_1$ and $M_2$ are monomials such that $b_{ij,k,l}(M_1) =

834: b_{ij,k,l}(M_2)$ for all $i,j,k,l$, then their difference can be

835: expressed as a sum of multiples of quadratic elements of $I_G$.

836: \end{lem}

837: 

838: We do this simply by, for each $M_{ij}$, using quadratic

839: generators corresponding to the separation $(w_l, V, W\setminus

840: \{w_l\})$ to move around the $b_{ij,1,l}(M)$ 1's in the $l$th

841: column. Consequently, we need only to connect monomials with the

842: same marginals and different $F$-values. We introduce an

843: additional definition.

844: 

845: Given a monomial $M$, the function $c_{ij,l}(M)$ is defined to be

846: the subset of ${0,1}$ which appears in the $w_l$-position in the

847: variables of $M_{ij}$. Explicitly, this contains 0 when

848: $b_{ij,0,l}(M) > 0$, and 1 when $b_{ij,1,l}(M) > 0$.

849: 

850: We now unspool a series of moves designed to connect all of the

851: remaining $F$-values of monomials with the same marginals.

852: 

853: \begin{lem}\label{deg4sh1}

854: Suppose we have a monomial $M$ and a column $l$ such that $1\in

855: c_{01,l}(M), c_{10,l}(M)$ and $0\in c_{00,l}(M), c_{11,l}(M)$.

856: Then $M$ is equivalent by adding a multiple of a binomial of

857: degree 4 to a monomial with the following changes to the $a_I$'s

858: and $b_I$'s:

859: \[

860: \begin{array}{c}

861: +1: b_{01,0,l}, b_{10,0,l}, b_{11,1,l}, b_{00,1,l}, \\

862: -1: b_{01,1,l}, b_{10,1,l}, b_{11,0,l}, b_{00,0,l}. \\

863: \end{array}

864: \]

865: \end{lem}

866: 

867: \begin{proof}

868: This corresponds merely to adding a multiple of the degree 4

869: binomial

870: 

871: \[

872: \left[

873: \begin{array}{cccc}

874: 1 & 1 & 1 & I_1\\

875: 1 & 0 & 0 & I_2\\

876: 0 & 0 & 1 & I_3\\

877: 0 & 1 & 0 & I_4\\

878: \end{array}

879: \right] - \left[

880: \begin{array}{cccc}

881: 1 & 0 & 1 & I_1\\

882: 1 & 1 & 0 & I_2\\

883: 0 & 1 & 1 & I_3\\

884: 0 & 0 & 0 & I_4\\

885: \end{array}

886: \right],

887: \]

888: where the columns are $v_1, w_l$, and $v_2$, and all

889: other vertices in some order. This binomial comes from the minor

890: $K_3$ given by contracting all the $w_i$, $i\neq l$, into either

891: $v_1$ or $v_2$.

892: \end{proof}

893: 

894: \begin{lem}\label{deg2elt}

895: 

896: Suppose we have a monomial $M$ such that for each column $l$,

897: there exists some index $i_l$ such that $i_l\in c_{01,l}(M),

898: c_{10,l}(M)$. Then $M$ is equivalent by adding a multiple  of a

899: binomial of degree $2$ to a monomial with the following changes to

900: the $a_I$'s and $b_I$'s:

901: \[

902: \begin{array}{c}

903: +1: a_{11}, a_{00}, b_{11,i_l,l}, b_{00,i_l,l}, \\

904: -1: a_{10}, a_{01}, b_{01,i_l,l}, b_{10,i_l,l},

905: \end{array}

906: \]

907: for all $l\in \{1,\ldots,n\}$.

908: 

909: \end{lem}

910: 

911: \begin{proof}

912: If $I$ is the index string composed of the $i_l$, this corresponds

913: to adding a multiple of the degree 2 binomial

914: 

915: \[

916: \left[

917: \begin{array}{ccc}

918: 1 & 1 & I\\

919: 0 & 0 & I\\

920: \end{array}

921: \right] - \left[

922: \begin{array}{ccc}

923: 1 & 0 & I\\

924: 0 & 1 & I\\

925: \end{array}

926: \right],

927: \]

928: where the columns are indexed by $v_1$, $v_2$, and the $w_l$.

929: \end{proof}

930: 

931: Now, suppose we have two monomials with the same marginals, that

932: is the same image under $\phi_G$.  Add multiples of the binomials

933: from Lemma~\ref{deg4sh1} and Lemma~\ref{deg2elt} until one can no

934: longer apply these; since both increase $a_{11} + \sum b_{11,l}$,

935: one will not go around in circles. Our monomials $M$ and $N$ are

936: now in ``reduced'' form, in the sense that neither move can be

937: applied. We break the situation down into cases.

938: 

939: {\it Case 1. Suppose that $M_{11}$ and $N_{11}$ are both not equal

940: to $1$, so that both $M$ and $N$ have an entry which is $11$ in

941: the $(v_1,v_2)$ direction.}

942: 

943: If, for each $l$, there exists an index $i_l$ such that $i_l\in

944: c_{11,l}(M)$ and $i_l\in c_{11,l}(N)$, then, as desired, we can

945: simply extract the variable $p_{11,(i_l)}$ from both $M$ and $N$;

946: in other words, for both $M$ and $N$, we can find a monomial with

947: the same values of $a$ and $b$ containing this variable.

948: 

949: If this is not the case, then there exists an $l$ for which

950: without loss of generality $c_{11,l}(M) = \{1\}$ and $c_{11,l}(N)

951: = \{0\}$. Looking at the $(v_1,w_l)$ marginal, there exists at

952: least one marginal $11$ because of the first condition;

953: consequently, there must exist at least one of these marginals in

954: $N$. Since $c_{11,l}(N)=\{0\}$, the only other option is that

955: $1\in c_{10,l}(N)$. Similarly, considering the $(v_2,w_l)$

956: marginal, we must have $1\in c_{01,l}(N)$.

957: 

958: Now, we have $b_{11,0,l}(N) = k>0$. Since $b_{11,0,l}(M)=0$,

959: looking at the $10$-count in the $(v_1,w_l)$ direction, we must

960: have $b_{10,0,l}(M) = b_{10,0,l}(N)+k$. However, looking at the

961: $00$-count in the $(v_2,w_l)$ direction, it now follows that we

962: must have $b_{00,0,l}(N) = b_{00,0,l}(M)+k$, and in particular

963: $0\in c_{00,l}(N)$. This is a contradiction, since we can now

964: apply a move as in Lemma~\ref{deg4sh1} to $N$, contradicting the

965: assumption that $N$ is reduced.

966: 

967: {\it Case 2. Exactly one of $M_{11}$ and $N_{11}$ is equal to 1.}

968: 

969: Suppose without loss of generality that $M_{11}\neq 1$ and

970: $N_{11}=1$. Take any $i_l\in c_{11,l}(M)$; this $i_l$ must be in

971: $c_{10,l}(N)$ and $c_{01,l}(N)$. This means that we can apply a

972: move as in Lemma~\ref{deg2elt} to $N$, again contradicting the

973: hypothesis that $N$ is reduced.

974: 

975: {\it Case 3. Both $M_{11}$ and $N_{11}$ are empty.}

976: 

977: In this case, it follows immediately that

978: $b_{10,i,l}(M)=b_{10,i,l}(N)$ for all $i$ and $l$ by considering

979: the $(v_1,w_l)$ marginals equal to $(1,i)$. If $M_{10}\neq 1$,

980: this means that we can find an $i_l$ for all $l$ such that this

981: number is nonzero, and we can then pull the corresponding variable

982: out of both $M$ and $N$. Similarly, if $M_{01}\neq 1$, we can find

983: a shared variable there, and if both of these are 1 then applying

984: the same argument to $M_{00}=M$ and $N_{00}=N$ finishes the job.

985: 

986: Our litany of cases has come to an end, completing the proof of

987: Theorem~\ref{k2n}. Note again that we have not only  shown that

988: $\mu(K_{2,n})=4$, but also given an explicit generating set in

989: degrees 2 and 4 for $I_{K_{2,n}}$.

990: \end{proof}

991: 

992: For $K_{m,n}$ where $m,n>2$, the answer is less clear. The

993: statement and proof of Theorem~\ref{k2n} indicate that for $m$

994: fixed, as $n$ gets large, the maximum degree of an element in the

995: Markov basis of the graph ideal of $K_{m,n}$ stabilizes. The

996: degree, on the other hand, certainly goes up as $\text{min}(m,n)$

997: does; for instance, there is an element of degree $2m$ in the

998: Markov basis of $K_{m,m}$, and we have the following result.

999: 

1000: \begin{prop}

1001: Fix $m\ge 2$. Then for $n \geq {m \choose 2} 2^{m-2}$, there   is

1002: an element of degree $2^{m-1}$ in the graph ideal of $K_{m,n}$.

1003: \end{prop}

1004: 

1005: \begin{proof}

1006: Let the vertices of $K_{m,n}$ be $\{v_1,\ldots,v_m\}\cup \{w_I\}$,

1007: where $I=(i_1,\ldots,i_m)$ is an index string of length $m$,

1008: consisting of precisely two 1's, and some number of 0's and 2's.

1009: There are precisely ${m \choose 2} 2^{m-2}$ such strings.

1010: 

1011: Specify the marginals as follows: between $v_j$ and $w_I$, insist

1012: upon $i_j$ marginals of $11$, $2-i_j$ marginals of $01$,

1013: $2^{m-2}-i_j$ marginals of $10$, and $2^{m-2}-2+i_j$ marginals of

1014: $00$. What this means is that exactly two variables with

1015: coordinate $w_I$ equal to 1 occur, and that the sum of the

1016: $v$-coordinates (considered as vectors) is precisely $I$; it

1017: furthermore specifies that each of 0 and 1 occurs $2^{m-2}$ times

1018: in each $v_j$-coordinate.

1019: 

1020: For each $w_I$, there are only two ways to express the vector  $I$

1021: as the sum of two 0-1 vectors. Since we consider all index strings

1022: $I$, we obtain that for each diamond in the natural Boolean

1023: partial order of binary strings of length $m$, either the top and

1024: bottom elements are in the set of $v$-coordinates of table

1025: entries, or the middle two entries are. By an easy induction, it

1026: follows that the set of $v$-coordinates, which numbers only

1027: $2^{m-1}$, must consist of either all strings with an even number

1028: of 1's, or all strings with an odd number of 1's. From here, we

1029: can easily compute the $w$-coordinates of each of these entries.

1030: 

1031: These resulting tables are the only two which satisfy  these

1032: marginals, and thus their difference, an element of degree

1033: $2^{m-1}$, must be in the Markov basis of $I_{K_{m,n}}$ as

1034: desired.   For all $n \geq {m \choose 2} 2^{m-2}$ there is a move

1035: of degree $2^{m-1}$ by Corollary \ref{cor:minor}.

1036: \end{proof}

1037: 

1038: We suspect that the following conjecture, an  extension of the

1039: result for $K_{2,n}$, holds.

1040: 

1041: \begin{conj}

1042: The graph ideal for $G=K_{m,n}$ is generated in degree at most

1043: $2^{\min\{m,n\}}$.

1044: \end{conj}

1045: 

1046: \section{Combinatorial Classification of Minimal Generators of Low Degree}

1047: 

1048: In this section, we give algorithms for computing  all generators

1049: of a given degree in the graph ideal $I_G$. For degrees two and

1050: three, we give an explicit combinatorial characterization of these

1051: generators, giving a generating set which generates $I_G$ in

1052: degree less than or equal to 3; for arbitrary degree $d$, we

1053: categorize these generators as pullbacks of a distinguished

1054: generator in the graph ideal of a fundamental graph $X_d$. The key

1055: lemma is the following, relating generators in $I_G$ to generators

1056: in a minor of $G$.

1057: 

1058: \begin{lem}\label{gphom}

1059: Let $G$ be a graph, and let $f=\Pi p_{I_j} - \Pi p_{I_k}$ be a

1060: binomial contained in $I_G$. Then we have the following.

1061: 

1062: (a) If $v_i$ corresponds to a column where all the index strings

1063: $I_j$ have the same value, then $f$ is a minimal generator if and

1064: only if $\tilde{f}$ is a minimal generator of the graph ideal

1065: $G\backslash v_i$, where $\tilde{f}$ is the natural image of $f$

1066: with the column $v_i$ deleted from each index string.

1067: 

1068: (b) If $v_i$ and $v_j$ are adjacent and correspond to columns

1069: where for each index string $I_j$ or $I_k$, the value of that

1070: string in each column is identical, then $f$ is a minimal

1071: generator if and only if $\tilde{f}$ is a minimal generator, where

1072: $\tilde{f}$ is the natural image of $f$ with the two columns $v_i$

1073: and $v_j$ fused. Here, $\tilde{f}$ is an element of the graph

1074: ideal of $G$ with those two vertices fused.

1075: 

1076: (c) Suppose $v_i$ and $v_j$ are any two vertices and correspond to

1077: columns where for each index string $I_j$ or $I_k$, the value of

1078: that string in each column is identical. In this case, if $f$ is a

1079: minimal generator of the graph ideal of $G$, then $\tilde{f}$ is a

1080: minimal generator of the graph ideal of $G$ with those two

1081: vertices fused.

1082: \end{lem}

1083: 

1084: \begin{proof}

1085: In each case, a decomposition of $\tilde{f}$ into generators of

1086: lower degree can be lifted via the obvious method to a

1087: decomposition of $f$. In (a), this is simply inserting the shared

1088: value of $v_i$ into each index string to form a valid index string

1089: for $G$; in (b) and (c), this is simply duplicating the value of

1090: each index string in the obvious manner.

1091: 

1092: In cases (a) and (b), any decomposition of $f$ must necessarily

1093: satisfy the property that each binomial used has the property in

1094: question, by considering in (a) any marginal containing $v_i$ and

1095: in (b) the marginal corresponding to the edge $v_iv_j$. Therefore,

1096: a decomposition of $f$ naturally yields a decomposition of

1097: $\tilde{f}$, so if $\tilde{f}$ is a minimal generator $f$ must be

1098: also.

1099: \end{proof}

1100: 

1101: This lemma has a corollary legitimizing our notion of minor.

1102: 

1103: \begin{cor} \label{cor:minor}

1104: If $H$ is a minor of $G$, then $\mu(H)\le \mu(G)$.

1105: \end{cor}

1106: 

1107: \begin{proof}

1108: It suffices to show $\mu(H)\le \mu(G)$ if $H$ is obtained  from

1109: $G$ by a single vertex deletion or edge contraction. However, if

1110: it is obtained by a vertex deletion, then by part (a) of

1111: Lemma~\ref{gphom}, every minimal generator of a given degree in

1112: $I_H$ lifts to a minimal generator of the same degree in $I_G$.

1113: Similarly, part (b) of Lemma~\ref{gphom} guarantees that

1114: $\mu(H)\le \mu(G)$ if $H$ is obtained from $G$ via an edge

1115: contraction.

1116: \end{proof}

1117: 

1118: A natural extension of this is the following, which  agrees with

1119: the data in Section 2, but which we have been unable to prove.

1120: 

1121: \begin{conj}

1122: If $H$ is obtained from $G$ by deleting an edge, then $\mu(H)\le \mu(G)$.

1123: \end{conj}

1124: 

1125: This set of minimal generators comes with a group action. In

1126: particular, the group $(\ZZ/2)^n$ acts naturally on

1127: $\mathbb{C}[p_I]$, via the element $(c_1,\ldots,c_n)$ sending a

1128: variable $p_{i_1\cdots i_n}$ to $p_{j_1,\ldots,j_n}$, where $j_r =

1129: i_r + c_r$; the sum is evaluated in $\ZZ/2$. This action consists

1130: merely of flipping 0's and 1's in some positions.

1131: 

1132: Furthermore, the automorphism group

1133: $\text{Aut}(G)$ acts naturally on $\mathbb{C}[p_I]$ as well, by

1134: permuting the indices according to the permutation of the vertices

1135: of the graph, so we have a natural action of $(\ZZ/2)^n \oplus

1136: \text{Aut}(G)$ on $\mathbb{C}[p_I]$. This action maps $I_G$ onto

1137: itself; we make the following natural definition.

1138: 

1139: \begin{defn}

1140: Two generators are {\bf equivalent} if they lie in the same orbit

1141: of $\mathbb{C}[p_I]$ under the action of $(\ZZ/2)^n \rtimes

1142: \text{Aut}(G)$.

1143: \end{defn}

1144: 

1145: If two generators of graph ideals $I_{G_1}$ and $I_{G_2}$ reduce

1146: to equivalent generators in a basic graph $H$ by means of the

1147: above manipulations, we say that they are {\bf weakly similar}. If

1148: they furthermore reduce to equivalent generators using only

1149: manipulations of type (a) and (b), we say that they are {\bf

1150: strongly similar}. We are now prepared to define  the object

1151: pivotal in our categorization of generators of degree $d$.

1152: 

1153: \begin{defn}\label{fungraph}

1154: Fix a degree $d\ge 2$. The {\bf fundamental graph}  $X_d$ has

1155: vertex set $(S_i,T_i)$, where $S_i$ and $T_i$ are subsets of

1156: $\{1,\ldots,d\}$ with cardinalities $|S_i| = |T_i| \leq d/2$, and

1157: if $|S_i|\,=d/2$ then $1\in S_i$. Two vertices $(S_1,T_1)$ and

1158: $(S_2,T_2)$ are connected by an edge if $|S_1\cap

1159: S_2|\,=\,|T_1\cap T_2|$.

1160: \end{defn}

1161: 

1162: To this fundamental graph is associated a distinguished  element

1163: of $I_{X_d}$.

1164: 

1165: \begin{defn}

1166: Fix a degree $d\ge 2$. Then the {\bf distinguished generator}

1167: $f_d\in I_{X_d}$ is the binomial

1168: 

1169: \[

1170: \left[

1171: \begin{array}{c}

1172: I_1\\

1173: \cdots\\

1174: I_d\\

1175: \end{array}

1176: \right] - \left[

1177: \begin{array}{c}

1178: J_1\\

1179: \cdots\\

1180: J_d\\

1181: \end{array}

1182: \right],

1183: \]

1184: 

1185: where $I_{ji}=1$ if $j\in S_i$ and 0 otherwise,  and similarly

1186: $J_{ji}=1$ if $j\in T_i$ and 0 otherwise.

1187: \end{defn}

1188: 

1189: It is clear that the distinguished generator is  actually in

1190: $I_{X_d}$, since for all adjacent $v_i$ and $v_j$, the number of

1191: 11-marginals in $I$ is equal to $S_i\cap S_j$, while the number of

1192: 11-marginals in $J$ is equal to $T_i\cap T_j$. By definition of

1193: $X_d$, these are equal, and furthermore, the number of

1194: 11-marginals determines the numbers of all other marginals (along

1195: with the numbers of 1's in each column of $I$ and $J$, which are

1196: of course identical.)

1197: 

1198: \begin{exmp}

1199: The fundamental graph $X_2$ has two vertices $(1,1)$ and $(1,2)$

1200: which are not connected by an edge.  The distinguished generator

1201: of $I_{X_2}$ is the binomial

1202: 

1203: $$ \left[ \begin{array}{cc} 1 & 1 \\ 0 & 0 \end{array} \right] -

1204: \left[ \begin{array}{cc} 1 & 0 \\ 0 & 1  \end{array}\right]$$

1205: 

1206: \noindent which is just as $2 \times 2 $ determinant.  The

1207: fundamental graph $X_3$ has nine vertices  which are $(1,1),

1208: (1,2), \ldots , (3,3)$.  Two vertices $(i,j)$ and $(k,l)$ are

1209: connected if and only if $i \neq k$ and $j \neq l$.  Each vertex

1210: has degree four:  $X_3$ is the edge graph of the 4-polytope

1211: $\Delta_2 \times \Delta_2$, the product to two triangles pictured

1212: in Figure 1. Note the six triangular prisms which appears as

1213: minors of $X_3$. The fundamental generator of $I_{X_3}$ is the

1214: binomial

1215: 

1216: $$ \left[ \begin{array}{ccccccccc}

1217: 1 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 \\

1218: 0 & 0 & 0 & 1 & 1 & 1 & 0 & 0 & 0 \\

1219: 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 1 \end{array} \right] - \left[

1220: \begin{array}{ccccccccc}

1221: 1 & 0 & 0 & 1 & 0 & 0 & 1 & 0 & 0 \\

1222: 0 & 1 & 0 & 0 & 1 & 0 & 0 & 1 & 0 \\

1223: 0 & 0 & 1 & 0 & 0 & 1 & 0 & 0 & 1 \end{array} \right].$$

1224: \end{exmp}

1225: 

1226: \begin{figure}

1227: \begin{center}\includegraphics{x31.eps} \caption{The fundamental graph

1228: $X_3$}

1229: \end{center}

1230: \end{figure}

1231: 

1232: With these definitions, we can formulate the main  theorem of this

1233: section, categorizing all generators of degree $d$.

1234: 

1235: \begin{thm}\label{degreed}

1236: Let $G$ be any graph. Then the minimal generators of  degree $d$

1237: in $I_G$ can be enumerated by the following procedure.

1238: \begin{enumerate}

1239: \item Consider all graph homomorphisms $\phi$ from minors $H$ of $G$ to

1240: $X_d$.

1241: \item Determine if the natural image of the fundamental generator $\widetilde{f_d}$

1242: is a minimal generator on the image subgraph.

1243: \item If so, consider the pullback $\phi^\star \widetilde{f_d}$, a binomial of degree

1244: $d$ in $I_G$.  We can pull this back to $H$  using

1245: Lemma~\ref{gphom} part (b) and (c), and then to $G$ using

1246: Lemma~\ref{gphom} part (a) and (b).

1247: \end{enumerate}

1248: \end{thm}

1249: 

1250: A graph homomorphism $G\rightarrow H$ is simply a map $\phi$  from

1251: vertices of $G$ to vertices of $H$ such that if $u$ and $v$ are

1252: adjacent in $G$, $\phi(u)$ and $\phi(v)$ are adjacent in $H$. Note

1253: that since we have used Lemma \ref{gphom} part (c), the pullback

1254: will not always be a minimal generator but the set of moves

1255: calculated in this way contains all minimal generators $I_G$ of

1256: degree $d$.

1257: 

1258: \begin{proof}

1259: Suppose we have a generator $f$ of degree $d$ in $I_G$,  written

1260: in tableau notation as

1261: 

1262: 

1263: \[

1264: \left[

1265: \begin{array}{c}

1266: I_1\\

1267: \cdots\\

1268: I_d\\

1269: \end{array}

1270: \right] - \left[

1271: \begin{array}{c}

1272: J_1\\

1273: \cdots\\

1274: J_d\\

1275: \end{array}

1276: \right],

1277: \]

1278: where $I$ and $J$ are 0-1 matrices. Each column of $I$ has the

1279: same number of 1's as the corresponding column of $J$. If this

1280: number is either 0 or $d$, we can delete that vertex via

1281: Lemma~\ref{gphom} (a) to obtain an equivalent generator in a minor

1282: of $G$, proving the theorem by induction.

1283: 

1284: Therefore, by flipping 0 with 1 if necessary, we can assume that

1285: each column has at most $d/2$ 1's, and if it has $d/2$ 1's then

1286: $I_i$ has a 1 in that column. To each vertex $v_i$ associate the

1287: pair $(S_i,T_i)$, where $j\in S_i$ if $I_j$ has a 1 in the column

1288: corresponding to $v_i$, and $j\in T_i$ if $J_j$ has a 1 in that

1289: column. If two adjacent vertices have the same pair, we can

1290: contract the edge between them via Lemma~\ref{gphom} (b) to again

1291: obtain an equivalent generator in a minor of $G$.

1292: 

1293: If this is not the case, then we claim that the map given by

1294: sending $v_i$ to the vertex $(S_i,T_i)$ in $X_d$ is a graph

1295: homomorphism. Indeed, all that we need to check is that if $v_i$

1296: and $v_j$ are connected by an edge, $S_i\cap S_j = T_i\cap T_j$

1297: (and it is not the case that $(S_i,T_i)=(S_j,T_j)$, which is true

1298: since no two adjacent vertices have the same pair.) But this must

1299: be true, since each is just the number of 11-marginals along the

1300: edge $v_iv_j$ in the corresponding tables $I$ and $J$, which are

1301: equal since $f$ is in $I_G$.

1302: 

1303: Furthermore, this map fuses two vertices $v_i$ and $v_j$ only if

1304: the corresponding columns of $I$ and of $J$ are identical.

1305: Therefore, by Lemma~\ref{gphom} (c), the image of $f$ in $I_R$,

1306: where $R$ is the image subgraph of $X_d$, must be a minimal

1307: generator of $X_{I_R}$. However, this generator is precisely the

1308: natural image of the distinguished generator in the graph ideal

1309: $I_R$. If this is irreducible, then $f$ is produced by the

1310: procedure in Theorem~\ref{degreed}, which we have just done in

1311: reverse. If not, then $f$ cannot be a minimal generator by the

1312: contrapositive of Lemma~\ref{gphom}.

1313: \end{proof}

1314: 

1315: In this manner, we have reduced the computation of all generators

1316: of degree $d$ to the process of determining which images of the

1317: distinguished generator in subgraphs of $X_d$ are minimal

1318: generators, and of enumerating graph homomorphisms from $G$ to

1319: $X_d$. While the problem of computing graph homomorphisms is a

1320: difficult one, we can use symmetry techniques to greatly aid us in

1321: many cases. Theorem~\ref{degreed} also divulges which generators

1322: are weakly similar: those whose corresponding graph homomorphisms

1323: have images which are isomorphic subgraphs of $X_d$.

1324: 

1325: We now apply Theorem~\ref{degreed} to degrees 2 and 3. Consider

1326: first generators of degree 2. The fundamental graph $X_2$ has two

1327: vertices, $(1,1)$ and $(1,2)$, which are not connected by an edge.

1328: Given a minor of $G$, it has a homomorphism onto $X_2$ if and only

1329: if there are no edges between the vertices mapped to $(1,1)$ and

1330: the vertices mapped to $(1,2)$. In this case the homomorphisms

1331: correspond to partitions of the vertices into $V_1$ and $V_2$.

1332: Extracting the definition of minor yields the following Corollary.

1333: 

1334: \begin{cor}

1335: Let $G$ be a graph.  Then equivalence classes of generators of

1336: degree 2  of $I_G$ correspond to partitions $V_1\cup V_2\cup V_3$,

1337: where there are no edges between $V_1$ and $V_2$. The generator

1338: corresponding to this is precisely

1339: 

1340: \[

1341: \left[

1342: \begin{array}{cc}

1343: 1 0 1 \\

1344: 0 1 1\\

1345: \end{array}

1346: \right] - \left[

1347: \begin{array}{cc}

1348: 1 1 1\\

1349: 0 0 1\\

1350: \end{array}

1351: \right],

1352: \]

1353: where the three columns correspond to $V_1, V_2$, and $V_3$.

1354: \end{cor}

1355: 

1356: Readers familiar with the study of graphical models and their

1357: induced independence statements will recognize this as a theorem

1358: which says that the only independence statements induced by a

1359: graphical model are the global independence statements.  See, for

1360: example \cite{L}.

1361: 

1362: Next, we turn our attention to the case of generators of degree 3,

1363: using the methods of Theorem~\ref{degreed} to obtain a

1364: combinatorial classification of all such cubic minimal generators

1365: of $I_G$. The fundamental graph $X_3$ has nine vertices

1366: $\{(1,1),\ldots,(3,3)\}$, with $(i,j)$ connected to $(k,l)$ if

1367: $i\neq k$ and $j\neq l$.  By direct computation, the image of the

1368: fundamental generator $d_3$ in $I_R$, $R$ a subgraph of $X_3$, is

1369: a minimal generator if and only if $R$ contains one of the six

1370: triangular prism subgraphs of $X_3$. Therefore, we can classify

1371: cubic generators by finding all homomorphisms from minors of $G$

1372: to the fundamental graph $X_3$ whose image contains a triangular

1373: prism. In particular, if no such homomorphism exists, then $I_G$

1374: cannot have a generator of degree 3. This yields the following

1375: corollary.

1376: 

1377: \begin{cor}\label{no3kn}

1378: For all $n$, the graph ideal $I_{K_n}$ has no minimal generators of degree 3.

1379: \end{cor}

1380: \begin{proof}

1381: The only minors of $K_n$ are copies of $K_m$ for $m\le n$.

1382: Furthermore, all homomorphic images of complete graphs are again

1383: complete graphs (indeed, of the same degree.) The only complete

1384: graphs occurring in $X_3$ are $K_2$ and $K_3$, neither of which

1385: contains a triangular prism.

1386: \end{proof}

1387: 

1388: Corollary~\ref{no3kn} shows why edge deletion  does not behave

1389: well with respect to the graph ideal $I_G$ and its Markov basis;

1390: this operation obviously can introduce elements of new degrees,

1391: since there exist graphs with cubic minimal generators, and every

1392: graph can be obtained from a complete graph by edge deletion. The

1393: same technique can be used to show that there are no minimal

1394: generators of degree 5 in the graph ideal $I_{K_n}$, since the

1395: largest clique in $X_5$ has size 5, and the distinguished

1396: generator cannot be minimal in the graph ideal of these subgraphs

1397: since in fact there are no minimal generators of degree 5 in $K_i$

1398: for $i\le 5$.

1399: 

1400: In addition to the description of  Theorem~\ref{degreed}, we can

1401: obtain a more straightforward combinatorial characterization of

1402: minimal generators of degree 3 in graph ideals. We start with a

1403: pair of definitions.

1404: 

1405: \begin{defn}

1406: Let $G$ be any graph. Then the {\bf 3-coloring graph} $C_3(G)$ has

1407: vertices equal to the set of proper 3-colorings of $G$. Two

1408: 3-colorings are connected by an edge if one can be obtained from

1409: the other by the following (reversible) procedure: pick a color

1410: $i\in \{1,2,3\}$, pick a connected component of the induced

1411: subgraph consisting of all vertices with colors not equal to $i$,

1412: and switch the other two colors on this component.

1413: \end{defn}

1414: 

1415: \begin{defn}

1416: A graph $G$ is {\bf 3-rigid} if the coloring space $C_3(G)$ is disconnected.

1417: \end{defn}

1418: 

1419: The crucial example of a 3-rigid graph is the same triangular

1420: prism that arose in the analysis of degree 3 via

1421: Theorem~\ref{degreed}. This graph is 3-rigid, since there are only

1422: two proper 3-colorings up to permutation of the colors: one

1423: triangle $(v_1,v_2,v_3)$ is colored $(1,2,3)$, and the other

1424: triangle $(w_1,w_2,w_3)$ must be colored either $(2,3,1)$ or

1425: $(3,1,2)$. Here, the vertex labels are chosen so that $v_i$ is

1426: adjacent to $w_i$. It is easy to check that these colorings lie in

1427: different connected components of $C_3(G)$; the connected

1428: component of each consists of the colorings obtained from it by

1429: permutation of the colors.

1430: 

1431: We now present a complete  description of cubic generators of $G$

1432: based on these combinatorial objects.

1433: 

1434: \begin{prop}\label{3rigid}

1435: Let $G$ be any graph. Then  we can enumerate the (equivalence

1436: classes of) cubic minimal generators of the graph ideal $I_G$ as

1437: follows.

1438: 

1439: (a) Find all the 3-rigid minors of $G$.

1440: 

1441: (b) For each such 3-rigid minor $H$, consider all of the connected

1442: components of the 3-coloring graph of $H_3$. For each connected

1443: component, pick a representative 3-coloring $R_i$.

1444: 

1445: (c) For each $R_i$, $i>1$, take the element

1446: \[

1447: f= \left[

1448: \begin{array}{ccc}

1449: a_1\\

1450: a_2\\

1451: a_3\\

1452: \end{array}

1453: \right] - \left[

1454: \begin{array}{ccc}

1455: b_1\\

1456: b_2\\

1457: b_3\\

1458: \end{array}

1459: \right],

1460: \]

1461: where $a_{jk}=1$ if either the vertex $k$ of $G$ is deleted in

1462: obtaining $H$ or its image in $H$ is colored with color $j$ in

1463: $R_1$, and $b_{jk}=1$ if either $k$ is deleted or its image in $H$

1464: is colored with color $j$ in $R_i$.

1465: 

1466: Each of these elements $f$ is a minimal generator of $I_G$ in

1467: degree 3, and these elements together with the quadratic elements

1468: described above generate $I_G$ up to degree 3.

1469: \end{prop}

1470: 

1471: \begin{proof}

1472: Suppose we have a cubic minimal generator of $I_G$, given by

1473: 

1474: \[

1475: f= \left[

1476: \begin{array}{ccc}

1477: a_1\\

1478: a_2\\

1479: a_3\\

1480: \end{array}

1481: \right] - \left[

1482: \begin{array}{ccc}

1483: b_1\\

1484: b_2\\

1485: b_3\\

1486: \end{array}

1487: \right].

1488: \]

1489: 

1490: We construct two 3-colorings $C_A$ and $C_B$ of a minor of $G$ as

1491: follows. For the vertex $v_j$, consider the $j$-th column of $A$

1492: and $B$. If this consists of all 1's or all 0's, delete $v_j$. If

1493: not, exchange 0's and 1's if necessary so that it has exactly one

1494: 1. Then, in $C_A$ (resp. $C_B$), color $v_j$ with the position in

1495: which this 1 appears in $A$ (resp. B).

1496: 

1497: Next, if two adjacent vertices have the same color in $A$

1498: (equivalent to having the same color in $B$ by counting

1499: 11-marginals along this edge), then contract that edge. What

1500: remains is two proper 3-colorings of a minor $H$, and when these

1501: 3-colorings are converted to a binomial in $I_G$ via the method in

1502: part (c), we recover precisely the element $f$. By

1503: Lemma~\ref{gphom}, the image $\tilde{f}$ in $I_H$ is a minimal

1504: generator if and only if $f$ is a minimal generator of $I_G$.

1505: 

1506: We claim that the image $\tilde{f}$ is a minimal  generator of

1507: $I_H$, i.e. inexpressible as a sum of multiples of quadrics, if

1508: and only if $C_A$ and $C_B$ are in different connected components

1509: of $C_3(H)$. Indeed, consider a multiple of a binomial generator,

1510: say

1511: \[

1512: g= \left[

1513: \begin{array}{ccc}

1514: a_1\\

1515: a_2\\

1516: a_3\\

1517: \end{array}

1518: \right] - \left[

1519: \begin{array}{ccc}

1520: c_1\\

1521: c_2\\

1522: a_3\\

1523: \end{array}

1524: \right].

1525: \]

1526: 

1527: If we convert the two monomials of $g$ into colorings, the same

1528: set of vertices will have color 3 in these colorings. Therefore,

1529: the difference between these colorings consists of changing colors

1530: from 1 to 2 or vice versa. The only way this can be done while

1531: preserving the properness of the coloring is if the change

1532: consists of switching the colors 1 and 2 on some connected

1533: components of the induced subgraph of $H$ consisting of vertices

1534: not colored 3. Therefore, when we add a multiple of a binomial

1535: generator, we end

1536:  1

1537: up with an element corresponding to a 3-coloring in the same

1538: component of $C_3(H)$, and any two elements connected by an edge

1539: in $C_3(H)$ differ by a multiple of a binomial generator.

1540: 

1541: Consequently, the colorings $C_A$ and $C_B$ are in  the same

1542: component of $C_3(H)$ if and only if $\tilde{f}$ is not

1543: expressible as the sum of multiples of binomials in $I_H$, which

1544: is equivalent to $\tilde{f}$ being a minimal generator of $I_H$,

1545: as desired.

1546: \end{proof}

1547: 

1548: Classifying 3-rigid graphs is an interesting problem;  the graph

1549: $C_3(G)$ has been studied in connection with the problem of

1550: picking a random 3-coloring of a graph~\cite{Vigoda}. Indeed, the

1551: flip interchanging two colors on a connected component is

1552: precisely the move used in the Wang-Swendsen-Koteck\'{y} algorithm

1553: to pick a random $k$-coloring of a graph, and this scheme has ties

1554: to mathematical physics~\cite{WSK}. There are simple operations to

1555: produce 3-rigid graphs from other ones, but the triangular prism

1556: seems to be the only 3-rigid graph without a proper 3-rigid minor.

1557: 

1558: This method, paralleling Theorem~\ref{degreed}, can be  extended

1559: to higher degrees. However, it rapidly becomes unwieldy, as the

1560: vertices can now be colored with sets of $d/2$ colors, and the

1561: moves are more complicated, consisting of all moves keeping one of

1562: the colors fixed.  For instance, the fundamental graph $X_4$ has

1563: 34 vertices and understanding the homomorphisms to this graph

1564: seems difficult.

1565: 

1566: \section{Algebraic Degree of Forests}

1567: 

1568: A recent series of results gave a very thorough description of the

1569: family of ideals of decomposable models \cite{D, GMS, T}. A

1570: special case of these results is the following fundamental

1571: theorem.

1572: 

1573: \begin{thm}\label{thm:sqfr}

1574: The ideal $I_G$ is minimally generated by quadrics if and only if

1575: $G$ is a forest.  In this case, the set of quadratic squarefree

1576: binomials in $I_G$ forms a Gr\"obner basis with respect to the

1577: reverse lexicographic term order with $p_{0 \cdots 0} \prec p_{0

1578: \cdots 0 1} \prec \ldots \prec p_{1 \cdots 1}$.

1579: \end{thm}

1580: 

1581: Sturmfels \cite{St2} posed the natural follow-up problem of

1582: calculating the degree of the toric ideal $I_G$.  The degree of

1583: the toric ideal is interesting in statistics because it provides a

1584: natural upper bound for the maximum likelihood degree of the toric

1585: ideal \cite{St2}.  In this section we give combinatorial formulae

1586: for the degree of the graph model ideal $I_G$ when $G$ is a

1587: forest.  As the maximum likelihood degree of a forest is always 1,

1588: we see that the degree can be arbitrarily far from the maximum

1589: likelihood degree.  To perform these degree computations, we first

1590: recall a result about the degree of a general toric ideal. This

1591: result can be found in \cite{St}.

1592: 

1593: \begin{thm}

1594: Let $A$ be a $d \times n$ matrix whose toric ideal $I_A$ is

1595: homogeneous in the standard $\ZZ$-grading.  Then the degree of the

1596: ideal $I_A$ ($=$ degree of the projective toric variety defined by

1597: $I_A$) is equal to the normalized volume of the lattice polytope

1598: $Q = \mathrm{conv}(A)$.

1599: \end{thm}

1600: 

1601: Henceforth, the normalized volume of a lattice polytope $Q$ will

1602: be denoted $Vol(Q)$.  This theorem reduces the problem of

1603: calculating degree to computing the normalized volume of polytope.

1604: We now record some some basic facts about the polytope $P_G$ when

1605: the underlying graph is a forest.

1606: 

1607: \begin{lem}

1608: The polytope $P_G$ has dimension equal to the sum of the number of

1609: vertices and the number of edges of the graph (this is true for

1610: any graph).  There are precisely $4 \cdot | E(G) |  + 2 \cdot |

1611: Iso(G) | $ facets of $P_G$ when $G$ is a forest.  If the variables

1612: in marginal space are labelled $y^{(j,k)}_{i_j, i_k}$ for the

1613: variables coming from an edge and $y^{(l)}_{i_l}$ for the

1614: variables coming from an isolated vertex, then the facets are

1615: given by the inequalities

1616: 

1617: $$ y^{(j,k)}_{i_j, i_k} \geq 0 \mbox{  and   } y^{(l)}_{i_l} \geq 0.$$

1618: 

1619: \end{lem}

1620: 

1621: \begin{proof}

1622: The dimension formula appears in \cite{HS}.  The polyhedral

1623: results are a direct consequence of the closed form expressions

1624: for maximum likelihood estimates in decomposable models in

1625: \cite{L}.

1626: \end{proof}

1627: 

1628: The following lemma implies that to compute the degree of $I_G$

1629: when $G$ is a forest, one need only describe a formula which is

1630: valid for trees.  Furthermore, this lemma is important for

1631: carrying out the recursive computations implied by Theorem

1632: \ref{thm:vol}.

1633: 

1634: \begin{lem}\label{lem:prod}

1635: Let $G$ be a graph with a partition of the vertices $\{V_1,V_2\}$

1636: such that there is no edge in $G$ between the $V_1$ and $V_2$. Let

1637: $G_1$ and $G_2$ be the corresponding induced subgraphs.  Let $d_1$

1638: and $d_2$ be the corresponding dimensions of the polytopes

1639: $P_{G_1}$ and $P_{G_2}$; that is, $d_i = |V_i| + |E(G_i)|$ .  Then

1640: we have

1641: 

1642: $$\label{eqn:prod} Vol(P_G) = { d_1 + d_2 \choose d_1} \cdot Vol(P_{G_1}) \cdot

1643: Vol(P_{G_2}). $$

1644: 

1645: \end{lem}

1646: 

1647: \begin{proof}

1648: With these restrictions on the graph $G$, we have $P_G = P_{G_1}

1649: \times P_{G_2}$.  Equation \ref{eqn:prod} is then the usual

1650: formula for the normalized volume of the direct product in terms

1651: of the normalized volumes of the components of the product.

1652: \end{proof}

1653: 

1654: We now come to the main theorem of this section.

1655: 

1656: \begin{thm}\label{thm:vol}

1657: Let $G$ be a tree with $n$-vertices.  Then the degree of the toric

1658: ideal $I_G$ can be calculated by the formula

1659: 

1660: $$deg(I_G) = \frac{1}{2} \sum_{e \in E(G)} deg(I_{G \setminus e})

1661: $$

1662: 

1663: \noindent where the notation $G \setminus e$ denotes the graph $G$

1664: with the edge $e$ removed.

1665: \end{thm}

1666: 

1667: \begin{proof}

1668: As previously indicated, we prove the theorem by calculating the

1669: volume of the corresponding polytope $P_G$.  Theorem

1670: \ref{thm:sqfr} implies that the pulling triangulation of $P_G$

1671: induced by the reverse lexicographic term order above is

1672: unimodular. This in turn, implies that the normalized volume of

1673: $P_G$ is equal to the sum of the normalized volumes of the facets

1674: of $P_G$ which do not contain the vertex indexed by the variable

1675: $p_{1\cdots 1}$ (see \cite{St} for all the definitions and

1676: relevant theory). This is where the polyhedral description of

1677: $P_G$ when $G$ is a forest becomes essential.  We see from the

1678: polyhedral characterization that there are exactly $n-1$ facets of

1679: $P_G$ which do not contain this ``last'' vertex, and that they are

1680: indexed by the edges of $G$.  We will show that the normalized

1681: volume of the facet $F_e$ of $P_G$ which is indexed by the edge

1682: $e$ has volume precisely $\frac{1}{2} Vol(P_{G \setminus e})$

1683: which will complete the proof.

1684: 

1685: There are two cases to consider:  either the edge in question has

1686: one node a leaf or not (the case of the graph on two vertices with

1687: a lone edge is clear, by a direct calculation). We will handle the

1688: two cases separately.

1689: 

1690: \emph{Case 1}:  We may suppose without loss of generality that our

1691: edge $e$ is $\{1,2\}$, the vertex 1 is a leaf and the vertex $2$

1692: has the edge $\{2,3\}$ incident to it.  Then the matrix whose

1693: columns correspond to the vertices of $P_G$ has the top eight rows

1694: which look like

1695: 

1696: $$ \left( \begin{array}{cccccccc}

1697: 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 \\

1698: 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 \\

1699: 0 & 0 & 0 & 0 & 1 & 1 & 0 & 0 \\

1700: 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 \\

1701: 1 & 0 & 0 & 0 & 1 & 0 & 0 & 0 \\

1702: 0 & 1 & 0 & 0 & 0 & 1 & 0 & 0 \\

1703: 0 & 0 & 1 & 0 & 0 & 0 & 1 & 0 \\

1704: 0 & 0 & 0 & 1 & 0 & 0 & 0 & 1 \end{array} \right) $$

1705: 

1706: \noindent with this $8 \times 8$ block repeated $2^{n-3}$ times

1707: across the first eight rows.  We claim that the facet which

1708: corresponds to the inequality $y^{(1,2)}_{1,1} \geq 0$

1709: (corresponding to the fourth row of the above matrix) has volume

1710: equal to $\frac{1}{2} Vol(P_{G \setminus \{1,2\}})$.  Note the the

1711: vertices of $P_G$ which lie on this facet are precisely the $3

1712: \cdot 2^{n-2}$ vertices of $P_G$ which have a zero in the fourth

1713: row.  First we show that this facet is naturally isomorphic to a

1714: sub-configuration of $P_{G \setminus \{1,2\}}$. Consider the

1715: matrix $A_{G \setminus \{1,2\}}$ whose columns give the vertices

1716: of $P_{G \setminus \{1,2\}}$.  This matrix has two fewer rows than

1717: the matrix $A_G$ above and is almost the same: its first six rows

1718: look like

1719: 

1720: $$ \left( \begin{array}{cccccccc}

1721: 1 & 1 & 1 & 1 & 0 & 0 & 0 & 0 \\

1722: 0 & 0 & 0 & 0 & 1 & 1 & 1 & 1 \\

1723: 1 & 0 & 0 & 0 & 1 & 0 & 0 & 0 \\

1724: 0 & 1 & 0 & 0 & 0 & 1 & 0 & 0 \\

1725: 0 & 0 & 1 & 0 & 0 & 0 & 1 & 0 \\

1726: 0 & 0 & 0 & 1 & 0 & 0 & 0 & 1

1727:  \end{array} \right).$$

1728: 

1729: \noindent To show the natural isomorphism mentioned above, it

1730: suffices to show that there is a unimodular linear transformation

1731: from the first six columns of the first matrix above to the first

1732: six columns of the second matrix above.  Such a linear

1733: transformation is obtained by replacing the first row by the sum

1734: of the first and second rows, and then deleting the second and

1735: fourth rows.  We can delete the second and fourth rows because

1736: they are linear combinations of other rows and hence do not change

1737: the polyhedral description.  Such a transformation is clearly

1738: unimodular.

1739: 

1740: Now that we have shown that our configuration of $3 \cdot 2^{n-2}$

1741: points sits naturally inside $P_{G \setminus \{1,2\}}$, we wish to

1742: compute its volume.  For this, we show that there is a hyperplane

1743: which divides $P_{G \setminus \{1,2\}}$ into two congruent pieces,

1744: one of which is the convex hull of our new configuration of $3

1745: \cdot 2^{n-2}$ points.  This hyperplane is given by the equation

1746: $$ y^{(1)}_0 - y^{(1)}_1 + y^{(2,3)}_{0,0} + y^{(2,3)}_{0,1} - y^{(2,3)}_{1,0}

1747: -y^{(2,3)}_{1,1} = 0.$$

1748: 

1749: \noindent  Note that exactly $2^{n-1}$ vertices of $P_{G \setminus

1750: \{1,2\}}$ lie on this hyperplane (these are the ones corresponding

1751: to the middle four columns of the submatrix  of $A_{G \setminus

1752: \{1,2\}}$ displayed above) and the remaining $2^{n-1}$ vertices

1753: are split equally on each side of the hyperplane.  In particular,

1754: all of the vertices from our configuration of $3 \cdot 2^{n-2}$

1755: points lie on the nonnegative side of this hyperplane and the

1756: remaining $2^{n-2}$ points are on the negative side.  Furthermore,

1757: there is a natural reflexive symmetry across this hyperplane.  To

1758: complete the proof, we must show more: not only is the point

1759: configuration naturally ``cut in half'' by this hyperplane, but so

1760: is the polytope $P_{G \setminus \{1,2\}}$. This follows from a

1761: direct computation with the eight points listed above.  We

1762: performed the computation using the program \verb"PORTA" \cite{C}.

1763: 

1764: \emph{Case 2}: The argument is very similar to case 1;   we will

1765: sketch the relevant details.  We may assume that our edge is the

1766: edge $\{2,3\}$.  Since neither 2 nor 3 is a leaf we may assume $G$

1767: also contains the edges $\{1,2\}$ and $\{3,4\}$.  With these

1768: conditions, the first 12 rows of our matrix looks like

1769: 

1770: $$ \left( \begin{array}{cccccccccccccccc}

1771: 1 & 1 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\

1772: 0 & 0 & 0 & 0 & 1 & 1 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\

1773: 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 1 & 1 & 0 & 0 & 0 & 0 \\

1774: 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 1 & 1 \\

1775: 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 \\

1776: 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 \\

1777: 0 & 0 & 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 0 & 0 \\

1778: 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 \\

1779: 1 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 1 & 0 & 0 & 0 \\

1780: 0 & 1 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 1 & 0 & 0 \\

1781: 0 & 0 & 1 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 1 & 0 \\

1782: 0 & 0 & 0 & 1 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 1 \\

1783: \end{array} \right) $$

1784: 

1785: \noindent  with this block repeated $2^{n-4}$ times.  We wish to

1786: show that the facet defined by the inequality $y^{(2,3)}_{1,1}

1787: \geq 0 $ (corresponding to the eighth row in the above matrix) has

1788: volume equal to $\frac{1}{2} Vol(P_{G \setminus \{2,3\}})$.  The

1789: vertices which lie on this facet are precisely the $3 \cdot

1790: 2^{n-2}$ vertices with a zero in the eighth row in this matrix

1791: representation.  First we show that this facet naturally appears

1792: as a sub-configuration of $P_{G \setminus \{2,3\}}$.  This can be

1793: achieved by applying a unimodular transformation to the

1794: configuration:  the key point is that once we restrict attention

1795: to the facet, the middle 4 rows of the above configuration can be

1796: written as linear combinations of the other rows and hence are

1797: redundant in terms of the polyhedral description.

1798: 

1799: Now we show that there is a hyperplane which divides the polytope

1800: $P_{G \setminus (2,3)}$ in half.  This is just the hyperplane

1801: given by

1802: $$y^{(1,2)}_{0,0} - y^{(1,2)}_{0,1} + y^{(1,2)}_{1,0} - y^{(1,2)}_{1,1} +

1803: y^{(3,4)}_{0,0} + y^{(3,4)}_{0,1} - y^{(3,4)}_{1,0} -

1804: y^{(3,4)}_{1,1} = 0 .$$

1805: 

1806: \noindent  Note that our configuration of $3 \cdot 2^{n-2}$ points

1807: are precisely the points on the nonnegative side of this

1808: hyperplane.  Furthermore, there is a natural reflexive symmetry

1809: across the hyperplane.  A direct calculation shows that this

1810: hyperplane not only separates the point configuration, but also

1811: divides the polytope into two symmetric pieces with the the

1812: desired integral points as vertices.  Thus we deduce the desired

1813: equation of volumes.

1814: \end{proof}

1815: 

1816: For some special families of trees we use this recurrence relation

1817: to deduce simple formulae for the degree.  These appear in the

1818: following corollaries.

1819: 

1820: 

1821: \begin{cor}

1822: Let $K_{1,n}$ denote a star graph with $n$ leaves.  Then

1823: $deg(I_{K_{1,n}}) = (n!)^2$.

1824: \end{cor}

1825: 

1826: \begin{proof}

1827: Removing any edge of the graph $K_{1,n}$ produces the graph

1828: consisting of the disjoint union of a $K_{1, n-1}$ and an isolated

1829: point.  Hence from theorem \ref{thm:vol} and lemma \ref{lem:prod}

1830: we deduce that

1831: 

1832: $$deg(I_{K_{1,n}}) = n \cdot \frac{1}{2} {2n \choose 1} \cdot

1833: deg(I_{K_{1,n-1}}) = n^2 \cdot deg(I_{K_{1,n-1}}).$$

1834: 

1835: \noindent  Since $deg(I_{K_{1,1}}) =1$ we have the desired result.

1836: \end{proof}

1837: 

1838: 

1839: \begin{cor}

1840: Let $T_n$ denote the graph of the $n$-chain and $d_n =

1841: deg(I_{T_n})$.  Then $d_n$ satisfies the recurrence

1842: 

1843: $$ \label{eqn:rec} d_{n+1} = \frac{1}{2} \sum_{i=1}^n  {2n \choose 2i -1} d_i

1844: d_{n+1-i}$$

1845: 

1846: \noindent with $d_1 = 1$.  Furthermore, we have the equality of

1847: generating functions

1848: 

1849: $$\sum_{n=1}^{\infty} \frac{d_n}{(2n-1)!} x^{2n-1} = \sqrt{2}

1850: \tan(\frac{x}{\sqrt{2}}).$$

1851: \end{cor}

1852: 

1853: \begin{proof}

1854: The recurrence relation \ref{eqn:rec} follows immediately from the

1855: formula in Theorem \ref{thm:vol} and by applying Lemma

1856: \ref{lem:prod} to the graph consisting of two disjoint chains of

1857: length $i$ and $n+1-i$.

1858: 

1859: To deduce the equality of generating functions, let $y =

1860: \frac{1}{2} \sum_{n=1}^{\infty} \frac{d_n}{(2n-1)!} x^{2n-1}$. The

1861: recurrence relation implies that $2y' -2 = y^2$.  Solving the

1862: differential equation yields the desired formula.

1863: \end{proof}

1864: 

1865: The recurrence relation and generating function in the case of the

1866: $n$-chain also appears in a paper by Poupard \cite{Po}, but we do

1867: not know how to show that the objects we are counting (simplices

1868: in a regular unimodular triangulation) are in bijection with the

1869: objects she was counting (types of binary trees).

1870: 

1871: 

1872: 

1873: \section{Acknowledgements}

1874: 

1875: Mike Develin was supported in part by an NSF Graduate Research

1876: Fellowship and by the American Institute of Mathematics.  Seth

1877: Sullivant was supported by an NSF Graduate Research Fellowship.

1878: 

1879: 

1880: \begin{thebibliography}{99}

1881: 

1882: \bibitem{C} T.~Christof.

1883: \verb"PORTA": a polyhedron representation transformation

1884: algorithm, version 1.3.1 available at

1885: \verb"ftp://elib.zib.de/pub/mathprog/polyth/index.html".

1886: 

1887: \bibitem{DS} P.~Diaconis and B.~Sturmfels.

1888: Algebraic algorithms for sampling from conditional distributions.

1889: \emph{Annals of Statistics}, 26 (1998), 363--397.

1890: 

1891: \bibitem{D} A.~Dobra.

1892: Markov bases for decomposable graphical models.  \emph{Bernoulli},

1893: 9 (2003), no. 6, 1--16.

1894: 

1895: \bibitem{DS2} A.~Dobra and S.~Sullivant.

1896: A divide-and-conquer algorithm for generating Markov bases of

1897: multi-way tables.  To appear in \emph{Computational Statistics}.

1898: 

1899: \bibitem{Dyer} M.~Dyer, L.A.~Goldberg, C.~Greenhill, M.~Jerrum, M.~Mitzenmacher. An extension of path

1900: coupling and its application to the Glauber dynamics for graph colorings. {\em SIAM J. Comput.},

1901: 30 (2001), 1962--1975.

1902: 

1903: \bibitem{GMS} D.~Geiger, C.~Meek, and B.~Sturmfels.

1904: On the toric algebra of graphical models. Manuscript, 2002.

1905: 

1906: \bibitem{M2}

1907: D. Grayson and M. Stillman. \verb"Macaulay 2": a software system

1908: for research in algebraic geometry.  Available at

1909: \verb"http://www.math.uiuc.edu/Macaulay2/".

1910: 

1911: \bibitem{H} R.~Hemmecke.

1912: \verb"4ti2": computation of Hilbert bases, Graver bases, toric

1913: Gr\"{o}bner bases and more.  Available at

1914: \verb"http://www.4ti2.de/".

1915: 

1916: \bibitem{HS}

1917: S. Ho\c{s}ten and S. Sullivant. Gr\"{o}bner bases and polyhedral

1918: geometry of cyclic and reducible models,  \emph{Journal of

1919: Combinatorial Theory: Series A}, 100 (2002), no. 2, 277--301.

1920: 

1921: \bibitem{L} S.~Lauritzen. {\em Graphical Models}. Oxford

1922: University Press, New York, 1996.

1923: 

1924: \bibitem{Po}

1925: C. Poupard. Deux proprietes des arbres binaires ordonnes stricts,

1926: \emph{European J. Combin.}, 10 (1989), 369--374.

1927: 

1928: \bibitem{RW}

1929: R. Read and R. Wilson. {\em An Atlas of Graphs}.  Oxford

1930: University Press, New York, 1999.

1931: 

1932: \bibitem{St}

1933: B.~Sturmfels. \emph{Gr{\"o}bner Bases and Convex Polytopes},

1934: American

1935:   Mathematical Society, Providence, 1995.

1936: 

1937: \bibitem{St2}

1938: B.~Sturmfels.

1939: \newblock {\em Solving Systems of Polynomial Equations}, volume~97 of {\em CBMS

1940:   Regional Conference Series in Mathematics}.

1941: \newblock American Mathematical Society, Providence, 2002.

1942: 

1943: % \bibitem{Su} S.~Sullivant.

1944: % {\em Algebraic Geometry and Combinatorics of Hierarchical Models}.

1945: %  Master's thesis, San Francisco State University, 2002.

1946: 

1947: \bibitem{T} A.~Takken.

1948: {\em Monte Carlo Goodness-of-Fit Tests for Discrete Data}. PhD

1949: thesis,  Stanford University,  1999.

1950: 

1951: \bibitem{Vigoda} E. Vigoda. Improved bounds for sampling colorings,

1952: \emph{J. Mathematical Physics}, 41 (2000) no. 3, 1555--1569.

1953: 

1954: \bibitem{WSK} J.S. Wang, R.H. Swendsen, and R. Koteck\'{y}.

1955: Antiferromagnetic Potts models. \emph{Phys. Rev. Lett}, 63 (1989), no. 2, 109--112.

1956: 

1957: \end{thebibliography}

1958: \end{document}

1959: