1: \documentclass[10pt]{amsart}

2: \input{epsf.tex}

3: 

4: \usepackage{latexsym,amsfonts,amssymb,epsfig,verbatim}

5: \usepackage{amsmath, latexsym, graphics, textcomp,psfrag}

6: \usepackage{float}

7: %\usepackage{showlabels}

8: \usepackage{amsthm}

9: 

10: \usepackage[all,2cell,dvips]{xy}

11: %\usepackage[dvips]{graphicx}

12: 

13: %\renewcommand{\topfraction}{0.75}

14: %\renewcommand{\textfraction}{0.1}

15: 

16: \setlength{\parskip}{2ex}

17: \setlength{\parindent}{0in}

18: 

19: \newcommand{\nc}{\newcommand}

20: \nc{\dmo}{\DeclareMathOperator}

21: \nc{\nt}{\newtheorem}

22: 

23: \nc{\ds}{\displaystyle}

24: \nc{\ens}{\ensuremath}

25: 

26: 

27: \theoremstyle{plain}

28: \nt{thm}{Theorem}[section]

29: \nt*{main}{Main Theorem}

30: \nt{lem}[thm]{Lemma}

31: \nt{cor}[thm]{Corollary}

32: \nt*{cor*}{Corollary}

33: \nt{prop}[thm]{Proposition}

34: \nt*{q}{Question}

35: \nt{fact}{Fact}

36: \nt*{claim}{Claim}

37: \nt{conjecture}[thm]{Conjecture}

38: 

39: \theoremstyle{remark}

40: \newtheorem{example}[thm]{Example}

41: \newtheorem{exercise}[thm]{Exercise}

42: \newtheorem{definition}[thm]{Definition}

43: \newtheorem{remark}[thm]{Remark}

44: \newtheorem{definition-remark}[thm]{Definition-Remark}

45: \newtheorem{examples}[thm]{Examples}

46: \newtheorem*{definition*}{Definition}

47: \newtheorem*{remark*}{Remark}

48: \newtheorem{question}[thm]{Question}

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

50: \newtheorem{obsdef}[thm]{Observation/Definition}

51: 

52: 

53: \def\pmf{{\cal PMF}}

54: \def\mf{{\cal MF}}

55: \def\r{{\mathbb R_+}}

56: \def\H{{\mathbb H}}

57: \def\G{{\Gamma}}

58: \def\S{{\Sigma}}

59: 

60: \def\C{{\mathcal C}}

61: 

62: \nc{\N}{\mathcal{N}}

63: 

64: \def\S{{\mathcal S}}

65: 

66: 

67: \dmo{\out}{Out}

68: \dmo{\aut}{Aut}

69: \dmo{\gl}{GL}

70: \dmo{\SL}{SL}

71: \dmo{\sy}{Sp}

72: \dmo{\mcg}{Mod}

73: 

74: \dmo{\stu}{\tilde{St}}

75: \dmo{\st}{St}

76: \dmo{\lku}{\tilde{Lk}}

77: \dmo{\lk}{Lk}

78: \dmo{\dlk}{Lk_<}

79: \dmo{\cdlk}{Lk_\ll}

80: 

81: \nc{\T}{\mathcal{T}}

82: \nc{\tm}{\mathcal{M}}

83: \nc{\X}{\mathcal{X}}

84: \nc{\Y}{\mathcal{Y}}

85: \nc{\K}{\mathcal{K}}

86: \nc{\IA}{\textrm{IA}}

87: 

88: \nc{\sg}{\Sigma_g}

89: \nc{\modg}{\mcg(\sg)}

90: \nc{\mods}{\mcg(S)}

91: \nc{\spz}{\sy_{2g}(\Z)}

92: 

93: \nc{\Z}{\mathbb{Z}}

94: \nc{\R}{\mathbb{R}}

95: \nc{\Q}{\mathbb{Q}}

96: \nc{\I}{\mathcal{I}}

97: \nc{\W}{W}

98: 

99: \nc{\outf}{\out(F_n)}

100: \nc{\glz}{\gl_n(\Z)}

101: \nc{\outfn}{\out(F_n)}

102: \nc{\glnz}{\gl_n(\Z)}

103: \nc{\glzw}{W \backslash \! \glz}

104: 

105: %\nc{\kpo}{K(\pi,1)}

106: \nc{\kpo}{K(\T_n,1)}

107: 

108: \nc{\rt}{\tilde{\rho}}

109: \nc{\strt}{\st(\rt)}

110: \nc{\strto}{\st(\rt_1)}

111: \nc{\strtt}{\st(\rt_2)}

112: \nc{\strtop}{\st(\rt_1')}

113: \nc{\strttp}{\st(\rt_2')}

114: \nc{\lkrt}{\lk(\rt)}

115: \nc{\str}{\st(\rho)}

116: \nc{\lkr}{\lk(\rho)}

117: \nc{\dlkr}{\dlk(\rho)}

118: \nc{\cdlkr}{\cdlk(\rho)}

119: 

120: \nc{\li}{L_i}

121: \nc{\lit}{\tilde\li}

122: 

123: \nc{\seg}{\mathcal{S}}

124: \nc{\lseg}{\underline{\mathcal{S}}}

125: 

126: \nc{\slice}{\Sigma_I}

127: \nc{\bzt}{\ens{{\mathbb Z}^2}}

128: 

129: \nc{\norm}[1]{\|#1\|}

130: \nc{\bignorm}[1]{\left\|#1\right\|}

131: 

132: \nc{\tbt}[4]{\ensuremath{\left(\begin{array}{rr} #1 & #2 \\ #3 & #4\end{array}\right)}}

133: 

134: \nc{\twbo}[2]{\ensuremath{\left(\begin{array}{c} #1 \\ #2 \\  \end{array}\right)}}

135: 

136: \nc{\tbo}[3]{\ensuremath{\left(\begin{array}{c} #1 \\ #2 \\ #3 \end{array}\right)}}

137: 

138: \nc{\ttt}[9]{\ensuremath{\left(\begin{array}{rrr} #1 & #2 & #3 \\ #4 & #5 & #6 \\ #7 & #8 & #9  \end{array}\right)}}

139: 

140: \nc{\p}[1]{\paragraph{{\bf #1}}}

141: 

142: \nc{\pics}[3]{\epsfysize=#3 cm \begin{figure}[htb]

143: \center{\leavevmode \epsfbox{#1.eps}} \caption{#2}  \label{#1pic}

144: \end{figure}}

145: 

146: \nc{\beqn}{\begin{equation}}

147: \nc{\eeqn}{\end{equation}}

148: 

149: \nc{\bl}{ \begin{list}{$\cdot$}{

150: \setlength{\leftmargin}{.5in}

151: \setlength{\rightmargin}{.5in}

152: \setlength{\parsep}{0.5ex plus .2ex minus 0ex}

153: \setlength{\itemsep}{0.2ex plus 0.2ex minus 0ex}

154: }

155: }

156: \nc{\blwide}{ \begin{list}{$\cdot$}{

157: \setlength{\leftmargin}{.25in}

158: \setlength{\rightmargin}{.25in}

159: \setlength{\parsep}{0.5ex plus .2ex minus 0ex}

160: \setlength{\itemsep}{0.2ex plus 0.2ex minus 0ex}

161: }

162: }

163: 

164: \nc{\el}{\end{list}}

165: 

166: \nc{\bpf}{\begin{proof}}

167: \nc{\epf}{\end{proof}}

168: 

169: \nc{\si}{\sigma}

170: \nc{\ep}{\epsilon}

171: \nc{\al}{\alpha}

172: \nc{\be}{\beta}

173: \nc{\Ga}{\Gamma}

174: \nc{\la}{\lambda}

175: 

176: \nc{\margin}[1]{\marginpar{\scriptsize #1}}

177: 

178: \nc{\set}[1]{\{#1\}}

179: 

180: \nc{\bc}{\begin{comment}}

181: \nc{\ec}{\end{comment}}

182: 

183: 

184: \begin{document}

185: 

186: \title[Dimension of the Torelli group for $\out(F_n)$]{Dimension of the Torelli group for \boldmath$\out(F_n)$}

187: 

188: \author{Mladen Bestvina}

189: \author{Kai-Uwe Bux}

190: \author{Dan Margalit}

191: 

192: \address{Mladen Bestvina: Department of Mathematics\\ University of Utah\\ 155 S 1440 East \\ Salt Lake City, UT 84112-0090}

193: \email{bestvina@math.utah.edu}

194: 

195: \address{Kai-Uwe Bux: Department of Mathematics\\ University of Virginia\\ Kerchof

196:   Hall 229\\ Charlottesville, VA 22903-4137}

197: \email{kb2ue@virginia.edu}

198: 

199: \address{Dan Margalit: Department of Mathematics\\ University of Utah\\ 155 S 1440 East \\ Salt Lake City, UT 84112-0090}

200: \email{margalit@math.utah.edu}

201: 

202: \thanks{The first and third authors gratefully acknowledge support by

203:   the National Science Foundation.}

204: 

205: 

206: \keywords{$Out(F_n)$, Torelli group, cohomological dimension}

207: 

208: \subjclass[2000]{Primary: 20F36; Secondary: 20F28}

209: 

210: \maketitle

211: 

212: \begin{center}\today\end{center}

213: 

214: \begin{abstract}

215: Let $\T_n$ be the kernel of the natural map $\outf \to \glz$.  We

216: use combinatorial Morse theory to prove that $\T_n$ has an Eilenberg--MacLane space which is $(2n-4)$-dimensional and that

217: $H_{2n-4}(\T_n,\Z)$ is not finitely generated ($n \geq 3$).  In particular, this

218: recovers the result of Krsti\'c--McCool that $\T_3$ is not finitely

219: presented.  We also give a new proof of the fact, due to Magnus, that

220: $\T_n$ is finitely generated.

221: \end{abstract}

222: 

223: %%%

224: %%%

225: %%%

226: 

227: \section{Introduction}

228: 

229: There is a natural homomorphism from $\out(F_n)$, the group of outer

230: automorphisms of the free group on $n$ generators, to $\gl_n(\Z)$,

231: given by abelianizing the free group $F_n$.  It is a theorem of

232: Nielsen that this map is surjective \cite{jn}.  We call its kernel the

233: \emph{Torelli subgroup} of $\out(F_n)$, and we denote it by $\T_n$:

234: 

235: \[ 1 \to \T_n \to \out(F_n) \to \gl_n(\Z) \to 1 \]

236: 

237: \begin{main} \label{m1}

238: For $n \geq 3$, we have:

239: \begin{enumerate}

240: \item $\T_n$ has a $(2n-4)$-dimensional Eilenberg--MacLane space.

241: \item $H_{2n-4}(\T_n,\Z)$ is infinitely generated.

242: \item \label{fg} $\T_n$ is finitely generated.

243: \end{enumerate}

244: \end{main}

245: 

246: Part~(\ref{fg}) of the main theorem is due to Magnus; we give our own

247: proof in Section~\ref{section:finite generation}.  We remark that

248: $\T_1$ is obviously trivial and $\T_2$ is trivial by a classical

249: result of Nielsen \cite{jn} (we give a new proof of the latter fact in

250: Section~\ref{section:finite generation}).

251: 

252: The group $\T_n$, like any torsion free subgroup of $\out(F_n)$, acts

253: freely on the spine for outer space (see Section~\ref{section:kpo}),

254: and therefore has an Eilenberg--MacLane space of dimension $2n-3$, the dimension of this

255: spine.  Our theorem improves this upper bound on the dimension and

256: shows that $2n-4$ is sharp.

257: 

258: When $n=3$, we obtain that $H_2(\T_3,\Z)$ is not finitely generated,

259: and this immediately implies the result of Krsti\'c--McCool that

260: $\T_3$ is not finitely presented \cite{km}.

261: 

262: \p{Historical background} The question of whether $H_k(\T_n,\Z)$ is finitely generated, for

263: various values of $k$ and $n$, is a long standing problem with few

264: solutions.  This question was explicitly asked by Vogtmann in her

265: survey article \cite{kv}.  We now give a brief history of related

266: results, all of which are recovered by our main theorem.

267: 

268: Nielsen proved in 1924 that $\T_3$ is finitely generated \cite{jn}.  Ten years later, Magnus proved

269: that $\T_n$ is finitely generated for every $n$ \cite{wm}.

270: 

271: Smillie--Vogtmann proved in 1987 that, if $2 < n < 100$ or $n > 2$ is

272: even, then $H_\star(\T_n,\Z)$ is not finitely

273: generated \cite{sv1} \cite{sv2}.  Their method is to consider the

274: rational Euler characteristics of the groups in the short exact

275: sequence defining $\T_n$ (see \cite{kv}).

276: 

277: The Krsti\'c--McCool result that $\T_3$

278: is not finitely presented was proven in 1997, via completely algebraic

279: methods \cite{km}.

280: It is a general fact that if the second homology of a group is not

281: finitely generated, then the group is not finitely presented.

282: 

283: \p{Large abelian subgroups} It follows from the second part of the

284: main theorem that the first part is sharp; i.e., $\T_n$ does not have

285: an Eilenberg--MacLane space of dimension less than $2n-4$.  A simpler proof that the

286: cohomological dimension of $\T_n$ is at least $2n-4$ is to simply

287: exhibit an embedding of $\Z^{2n-4}$ into $\T_n$.  There is a subgroup

288: $\Z^{2n-4} \cong G < \T_n$ consisting of elements with representative automorphisms

289: given by:

290: \[

291: \begin{array}{rcl}

292: x_1 & \mapsto & x_1 \\

293: x_2 & \mapsto & x_2 \\

294: x_3 & \mapsto & [x_1,x_2]^{p_3}x_3[x_1,x_2]^{q_3} \\

295: & \vdots & \\

296: x_n & \mapsto & [x_1,x_2]^{p_n}x_n[x_1,x_2]^{q_n}

297: \end{array}

298: \]

299: for varying $p_i$ and $q_i$ (the $x_i$ are generators for $F_n$).

300: 

301: In Section~\ref{section:infinite generation}, we prove that specific

302: conjugates of $G$ represent independent classes in

303: $H_{2n-4}(\T_n,\Z)$, thus proving the second part of the main theorem.

304: These conjugates are exactly the generators of $H_{2n-4}(\tm_n,\Z)$,

305: where $\tm_n$, called the ``toy model'', is a particularly simple

306: subcomplex of the Eilenberg--MacLane space $\Y_n$ defined in

307: Section~\ref{section:kpo}.  In Section~\ref{section:infinite

308: generation}, we prove that $H_{2n-4}(\tm_n,\Z)$ injects into

309: $H_{2n-4}(\Y_n,\Z)$.  Since the homology of $\tm_n$ is not finitely

310: generated in any dimension greater than 1, we are led to the following

311: question:

312: 

313: \begin{q}

314: \label{question:middle homology}

315: Does $H_\ast(\tm_n,\Z)$ inject into $H_\ast(\Y_n,\Z)$?

316: \end{q}

317: 

318: \p{Mapping class groups.} The term ``Torelli group'' comes from the

319: theory of mapping class groups.  Let $\sg$ be a closed surface of

320: genus $g \geq 1$.  The \emph{mapping class group} of $\sg$, denoted

321: $\modg$, is the group of isotopy classes of orientation preserving

322: homeomorphisms of $\sg$.  The \emph{Torelli group}, $\I_g$, is the

323: subgroup of $\modg$ acting trivially on the homology of $\sg$.  As

324: $\modg$ acts on $H_1(\sg,\Z)$ by symplectic automorphisms, $\I_g$ is

325: defined by:

326: \[ 1 \to \I_g \to \modg \to \spz \to 1 \]

327: It is a classical theorem of Dehn, Nielsen, and Baer that the natural

328: map $\modg \to \out(\pi_1(\sg))$ is an isomorphism.  In this sense

329: $\T_n$ is the direct analog of $\I_g$.

330: 

331: Our (lack of) knowledge of the finiteness properties of $\I_g$ mirrors

332: that for $\T_n$.  Using the fact that $\mcg(\Sigma_1) \cong \SL_2(\Z) =

333: \sy_2(\Z)$, it is

334: obvious that $\I_1$ is trivial.  Johnson showed in 1983 that $\I_g$ is

335: finitely generated for $g \geq 3$ \cite{dj1}.  In 1986,

336: McCullough--Miller showed that $\I_2$ is not finitely generated

337: \cite{mcm}, and Mess improved on this in 1992 by showing that $\I_2$

338: is a free group of infinite rank \cite{gm}.  At the same time, Mess

339: further showed that $H_3(\I_3,\Z)$ is not finitely generated.  In

340: Kirby's problem list, Mess asked about finiteness properties in higher

341: genus \cite{rk}.

342: 

343: 

344: \p{Automorphisms vs. outer automorphisms.}  Strictly speaking, Magnus

345: and Krsti\'c--McCool study the group $\K_n$, by which we mean the

346: kernel of $\aut(F_n) \to \gl_n(\Z)$, where $\aut(F_n)$ is the

347: automorphism group of the free group.  By considering the short

348: exact sequence

349: \[ 1 \to F_n \to \K_n \to \T_n \to 1 \]

350: we see that $\K_n$ is finitely generated if and only if $\T_n$ is

351: finitely generated.  Moreover, it follows from our main theorem and the spectral

352: sequence associated to this short exact sequence

353: that $H_{2n-3}(\K_n,\Z)$ is not finitely generated and if $k$ is the

354: smallest index so that $H_k(\T_n,\Z)$ is not finitely generated, then

355: $H_k(\K_n,\Z)$ is not finitely generated.

356: 

357: From a topological point of view, $\T_n$ is the more natural group to

358: study.

359: 

360: In the literature, $\T_n$ is sometimes denoted by

361: $\IA_n$ for ``identity on abelianization'' (see, e.g. \cite{kv}).

362: However, since Krsti\'c--McCool use $\IA_n$ to denote the kernel of $\aut(F_n)

363: \to \gl_n(\Z)$, we avoid this notation to eliminate the confusion.

364: The notation $\K_n$ comes from Magnus \cite{wm}.

365: 

366: \p{Acknowledgements.}

367: We would like to thank Bob Bell, Mikhail Gromov, Jon McCammond, and Kevin

368: Wortman for helpful conversations.  We are especially grateful to Karen Vogtmann for

369: explaining her unpublished work.

370: 

371: 

372: %%%

373: %%%

374: %%%

375: 

376: \section{An Eilenberg--MacLane space}

377: \label{section:kpo}

378: 

379: In Section~\ref{section:spine}, we recall

380: the definition of Culler--Vogtmann's spine for Outer space.  Then, in

381: Section~\ref{section:quotient}, we describe the quotient of this space by

382: $\T_n$.  This quotient is a $(2n-3)$-dimensional Eilenberg--MacLane space for $\T_n$.

383: 

384: A \emph{rose} is a graph with one vertex.  The \emph{standard rose} in

385: rank $n$, denoted $R_n$, is a particular rose which is fixed once and

386: for all.  We denote the standard generators of $F_n \cong \pi_1(R_n)$

387: by $x_1, \dots, x_n$.

388: 

389: %%%

390: 

391: \subsection{Spine for Outer space} \label{section:spine}

392: Culler--Vogtmann introduced the \emph{spine for Outer space}, which we

393: denote by $\X_n$, as a tool for studying $\out(F_n)$ \cite{cv}.  This

394: is a simplicial complex defined in terms of marked graphs.

395: 

396: A \emph{marked graph} is a pair $(\Ga,g)$, where $\Ga$ is a finite

397: metric graph (1-dimensional cell complex with a metric) with no

398: separating edges and no vertices of valence less than 3 and $g:R_n \to

399: \Ga$ is a homotopy equivalence ($g$ is called the \emph{marking}).  We

400: say that two marked graphs $(\Ga,g)$ and $(\Ga',g')$ are

401: \emph{equivalent} if $g' \circ g^{-1}$ is homotopic to an isometry,

402: where $g^{-1}$ is any homotopy inverse of $g$.  We will denote the

403: equivalence class $[(\Ga,g)]$ by $(\Ga,g)$.

404: 

405: The vertices of $\X_n$ are equivalence classes of marked graphs where

406: all edges have length 1.  A set of vertices

407: \[ \set{(\Ga_1,g_1), \dots, (\Ga_k,g_k)} \]

408: is said to span a simplex if $\Ga_{i+1}$ is obtained from $\Ga_i$ by

409: collapsing a forest in $\Ga_i$, and $g_{i+1}$ is the marking obtained

410: from $g_i$ via this operation.

411: 

412: We can think of arbitrary points of $\X_n$ as marked metric graphs:

413: for instance, as we move along an edge between two vertices in $\X_n$,

414: the length of some edge in the corresponding graphs (more generally,

415: the lengths of the edges in a forest) varies between 0 and 1.

416: 

417: There is a natural right action of $\out(F_n)$ on $\X_n$.  Namely,

418: given $\phi \in \out(F_n)$ and $(\Ga,g) \in \X_n$, the action is

419: given by:

420: \[ (\Ga,g) \cdot \phi = (\Ga,g \circ \phi) \]

421: (here we are using the fact that every element $\phi$ of $\out(F_n)$

422: can be realized by a homotopy equivalence $R_n \to R_n$, also denoted

423: $\phi$, uniquely up to homotopy).

424: 

425: Culler--Vogtmann proved the following result \cite{cv}:

426: 

427: \begin{thm}

428: \label{cv thm}

429: For $n \geq 2$, the space $\X_n$ is contractible.

430: \end{thm}

431: 

432: This theorem has the consequence that the virtual cohomological

433: dimension of $\out(F_n)$ is equal to $2n-3$, the dimension of

434: $\X_n$.

435: 

436: The \emph{star of a rose} in $\X_n$ is the union of the closed

437: simplices containing the vertex corresponding to a rose.  The key idea

438: for Theorem~\ref{cv thm} is to think of $\X_n$ as the union of stars of

439: vertices corresponding to marked roses.  We take an analogous approach

440: in this paper.

441: 

442: %%%

443: 

444: \subsection{The quotient}\label{section:quotient}

445: 

446: Baumslag--Taylor proved that $\T_n$ is torsion free \cite{bt}.  We

447: also know that the action of $\T_n$ on $\X_n$ is free: by the

448: definition of the action, point stabilizers correspond to graph

449: isometries, and isometries act nontrivially on homology.

450: Finally, the action is simplicial, and so it follows

451: that the quotient of $\X_n$ by $\T_n$ is an Eilenberg--MacLane space for $\T_n$:

452: \[ \Y_n = \X_n/\T_n \]

453: 

454: 

455: \p{Homology markings.} Since $\out(F_n)$ identifies every pair of

456: isometric graphs of $\X_n$, points of $\Y_n$ can be thought of as

457: equivalence classes of pairs $(\Ga,g)$, where $\Ga$ is a metric graph

458: (as before), and $g$ is a \emph{homology marking}; that is, $g$ is an

459: equivalence class of homotopy equivalences $R_n \to \Ga$, where two

460: homotopy equivalences are equivalent if (up to isometries of $\Ga$)

461: they induce the same map $H_1(R_n,\Z) \to H_1(\Ga,\Z)$.

462: 

463: Via the marking $g$, we can think of the (oriented) edges of $\Ga$ as

464: elements of $H^1(R_n) \cong \Z^n$ (if $e$ is an edge and $x$ is a

465: simplicial 1-chain, then $e(x)$ is the number of times $e$

466: appears in $x$).  As such, if we think of the generators $x_1,

467: \dots, x_n$ of $\pi_1(R_n)$ as elements of $H_1(R_n,\Z)$, then we can

468: \emph{label} each oriented edge $a$ of $\Ga$ by the corresponding row

469: vector:

470: \[ (a(g(x_1)), \dots, a(g(x_n))) \]

471: where $a(g(x_i))$ is the number of times

472: $g(x_i)$ runs over $a$ homologically, with sign.

473: 

474: In this way, a point of $\Y_n$ is given by a labelled graph, and two

475: such graphs represent the same point in $\Y_n$ if and only if there is

476: a label preserving graph isomorphism between them (i.e. if the map

477: induces the identity on cohomology).  See Figure~\ref{figure:not

478:   frontier} for an example of a labelled graph.  We remark that this

479: example exhibits the fact that $\Y_n$ is not a simplicial

480: complex---there are two edge collapses (and hence two edges in $\Y_n$)

481: taking this point to the rose with the identity marking.

482: 

483: \begin{figure}[htb]

484: \psfrag{x1}{$(1,0,0)$}

485: \psfrag{x2}{$(0,1,0)$}

486: \psfrag{x3}{$(0,0,1)$}

487: \centerline{\includegraphics[scale=.35]{label}}

488: \caption{An example of a labelled graph.}

489: \label{figure:not frontier}

490: \end{figure}

491: 

492: When convenient, we will confuse the points of $\Y_n$ with the

493: corresponding marked graphs.

494: 

495: We will make use of the following generalities about marked graphs in $\Y_n$:

496: 

497: \begin{prop}

498: \label{lemma1}

499: Let $(\Ga,g)$ be a marked graph.

500: \begin{enumerate}

501: \item \label{labels dont change} If an edge of $\Ga$ is collapsed, the labels of the

502:   remaining edges do not change.

503: \item \label{parallel}Any two edges of $\Ga$ with the same label (up

504:   to sign) are \emph{parallel}

505:   in the sense that the union of their interiors disconnects $\Ga$.

506: \item \label{switch condition} The sum of the labels of the (oriented) edges coming into a

507: vertex of $\Ga$ is equal to the sum of the labels of the edges leaving the

508: vertex.

509: \end{enumerate}

510: \end{prop}

511: 

512: We leave the proofs to the reader.

513: 

514: 

515: \p{Roses.}  Let $\Ga$ be a rose with edges $a_1, \dots, a_n$, and let

516: $g:R_n \to \Ga$ be a homology marking.  Up to isometries of $\Ga$, the marking $g$ gives an element of

517: $\gl_n(\Z)$, called the \emph{marking matrix}; the rows are exactly the labels of the edges.

518: 

519: Since all edges have length 1, the isometry group of $\Ga$ is generated

520: by swapping edges and by reversing the orientations of edges; the

521: former operation has the effect of switching rows of the matrix, and

522: the latter corresponds to changing signs of rows.  Thus, in this case,

523: $(\Ga,g)$ gives rise to an element of $\W\backslash\gl_n(\Z)$, where

524: $\W=\W_n$ is the signed permutation subgroup of $\gl_n(\Z)$, acting on

525: the left.  In fact, this gives a bijection between roses in $\Y_n$ and

526: elements of $\W\backslash\gl_n(\Z)$, as $\out(F_n)$ acts transitively

527: on the roses of $\X_n$.

528: 

529: \medskip

530: 

531: The right action

532: of $\out(F_n)$ on $\X_n$ descends to a right action of $\gl_n(\Z)$ on

533: $\Y_n$.  In particular, the action on roses is given by the

534: right action of $\gl_n(\Z)$ on $\W\backslash\gl_n(\Z)$.

535: 

536: %%%

537: %%%

538: %%%

539: 

540: \section{Stars of roses}

541: \label{section:star of a rose}

542: \label{section:frontier}

543: 

544: As with $\X_n$, we like to

545: think of the quotient $\Y_n$ as the union of stars of roses. By

546: definition, the \emph{star of a rose} in $\Y_n$ is the image of the

547: star of a rose in $\X_n$.  Thus, it consists of graphs which can be

548: collapsed to a particular rose.

549: We now discuss some of the basic properties of the star of a rose.

550: 

551: %%%

552: 

553: \subsection{Labels in the star of a rose} We will need several observations

554: about the behavior of labels in the star of a rose.  The proofs of the

555: various parts of the propositions are straightforward and are left to

556: the reader.  In each of the statements, let $\rho$ be a rose in $\Y_n$

557: represented by a marked graph $(\Ga,g)$.  Say that its edges $a_1,

558: \dots, a_n$ are labelled by $v_1, \dots, v_n \in \Z^n$.

559: 

560: \begin{prop}

561: \label{lemma2}

562: If $(\Ga',g')$ is a marked graph in $\str$, we have:

563: \begin{enumerate}

564: \item \label{vi} For each $i$, there is an edge of $\Ga'$ labelled $\pm v_i$.

565: \item \label{forest}The edges not labelled $\pm v_i$ form a forest.

566: \item \label{circles}If each edge of $\Ga'$ is labelled $\pm v_i$,

567:   then the union of edges labelled $\pm v_i$ (for any particular $i$)

568:   is a topological circle (see, e.g., Figure~\ref{figure:not frontier}).

569: \item \label{coefficients} The label of any edge of $\Ga'$ is of the form

570: \[ \sum_{i=1}^n k_i v_i \]

571: where $k_i \in \{-1,0,1\}$.

572: \end{enumerate}

573: \end{prop}

574: 

575: We have the following converse to the first two parts of the previous proposition:

576: 

577: \begin{prop}

578: \label{converse}

579: If $(\Ga',g')$ is a marked graph which has, for each $i$, at least one

580: edge of length 1 labelled $\pm v_i$, then $(\Ga',g')$ is in

581: $\str$.

582: \end{prop}

583: 

584: We also have a criterion for when a marked graph is in the frontier of

585: the star of a rose:

586: 

587: \begin{prop}

588: \label{lemma:frontier condition}

589: A marked graph $(\Ga',g')$ in $\str$ is in the frontier

590: of $\str$ if and only if it has at least one edge of length 1 whose

591: label is not $\pm v_i$ for any $i$.  In this case, the given label is

592: a label for some rose whose star contains $(\Ga',g')$.

593: \end{prop}

594: 

595: %%%

596: 

597: \subsection{Ideal edges}

598: \label{section:ideal edges}

599: 

600: Let $\rho=(\Ga,g)$ be a rose whose edges $a_1, \dots, a_n$ are labelled $v_1, \dots, v_n$,

601: as above.  An \emph{ideal edge} is any formal sum:

602: \[ \sum_{i=1}^n k_i a_i \]

603: where $k_i \in \{-1,0,1\}$, and at least two of the $k_i$ are nonzero.

604: An ideal edge is a ``direction'' in $\str$ in the following sense:

605: for any ideal edge, we can find a marked graph $(\Ga',g')$ in the

606: frontier of $\str$ where one of the edges of $(\Ga',g')$ has the label

607: $\sum k_i v_i$.  If a marked graph in $\str$ has an edge of length 1

608: with label $\sum k_i v_i$, we say that the marked graph

609: \emph{realizes} the ideal edge $\sum k_i a_i$.

610: 

611: \begin{lem}

612: \label{lem:1 edge}

613: Given any ideal edge for a particular rose, there is a 1-edge blowup

614: of $\rho$ in the frontier of $\str$ which realizes that ideal edge.

615: \end{lem}

616: 

617: The lemma is proven by example.  See Figure~\ref{figure:ideal edge}

618: for a picture of a 1-edge blowup realizing the ideal edge

619: $a_1-a_3+a_4$ in rank 5 (apply Proposition~\ref{lemma1}(\ref{switch

620: condition})).  Also, we see that there are many graphs satisfying the

621: conclusion of the lemma---the other graphs are obtained by moving the

622: loops labelled $v_2$ and $v_5$ arbitrarily around the graph.

623: 

624: \begin{figure}[H]

625: \psfrag{v1}{$v_1$}

626: \psfrag{v2}{$v_2$}

627: \psfrag{v3}{$v_3$}

628: \psfrag{v4}{$v_4$}

629: \psfrag{v5}{$v_5$}

630: \psfrag{v1v3v4}{$v_1-v_3+v_4$}

631: \centerline{\includegraphics[scale=.5]{1edge}}

632: \caption{A 1-edge blowup realizing the ideal edge $a_1-a_3+a_4$.}

633: \label{figure:ideal edge}

634: \end{figure}

635: 

636: Our notion of an ideal edge is simply the homological version of the

637: ideal edges of Culler--Vogtmann \cite{cv}.

638: 

639: We say that an ideal edge $\iota'$ is \emph{subordinate} to the ideal

640: edge $\iota = \sum k_i a_i$ if $\iota'$ is obtained by

641: changing some of the $k_i$ to zero.

642: A \emph{2-letter ideal edge} is an ideal edge of

643: the form $k_{i}a_{i} + k_{j}a_{j}$.  Two ideal edges are said

644: to be \emph{opposite} if one can be obtained from the other by

645: changing the sign of exactly one coefficient.  The following facts are

646: used in Section~\ref{section:finite generation}:

647: 

648: \begin{lem}

649: \label{lemma:subordinate}

650: Let $\rho=(\Ga,g)$ be a rose whose edges $a_1, \dots, a_n$ are

651: labelled $v_1, \dots, v_n$.  Suppose that $\iota$ and $\iota'$ are

652: ideal edges and that either:

653: \begin{enumerate}

654: \item $\iota'$ is subordinate to $\iota$, or

655: \item $\iota$ and $\iota'$ are 2-letter ideal edges which are not opposite.

656: \end{enumerate}

657: In either case, there is a marked graph $(\Ga',g')$ in $\str$ which

658: simultaneously realizes $\iota$ and $\iota'$.

659: \end{lem}

660: 

661: \bpf

662: 

663: In each case, we can explicitly describe the desired graph.  If

664: $\iota'$ is subordinate to the ideal edge $\iota = \sum k_i

665: a_i$, we start with a 1-edge blowup realizing $\iota$

666: (Lemma~\ref{lem:1 edge}), and then blow

667: up another edge to separate the edges which appear in $\iota'$ from

668: those which do not.  Figure~\ref{figure:simultaneous} (left hand side)

669: demonstrates this for $\iota=v_1+v_2+v_3+v_4$ and $\iota' = v_1+v_2$

670: in rank 4.

671: 

672: For the case of two 2-letter ideal edges which are not opposite,

673: without loss of generality it suffices to demonstrate marked graphs

674: which simultaneously realize $v_1+v_2$ with $-v_1-v_2$, $v_2+v_3$, or

675: $v_3+v_4$ (the arbitrary case is obtained by renaming/reorienting

676: edges and by attaching extra 1-cells to any vertex).  See

677: Figure~\ref{figure:simultaneous} (right hand side) for a

678: demonstration.  One can use Proposition~\ref{lemma1}(\ref{switch

679:   condition}) to verify the labels.

680: %

681: \epf

682: 

683: \begin{figure}[H]

684: \psfrag{v1}{$v_1$}

685: \psfrag{v2}{$v_2$}

686: \psfrag{v3}{$v_3$}

687: \psfrag{v4}{$v_4$}

688: \psfrag{v1v2}{$v_1+v_2$}

689: \psfrag{v2v3}{$v_2+v_3$}

690: \psfrag{v3v4}{$v_3+v_4$}

691: \psfrag{v1v2v3v4}{$v_1+v_2+v_3+v_4$}

692: \psfrag{-v1v2}{$-v_1-v_2$}

693: \centerline{\includegraphics[scale=.5]{compatible}}

694: \caption{Marked graphs simultaneously realizing subordinate ideal

695:   edges (left) and 2-letter ideal edges which are not opposite

696:   (right).}

697: \label{figure:simultaneous}

698: \end{figure}

699: 

700: The reader may verify that opposite ideal edges are never

701: simultaneously realized.

702: 

703: We remark that, in the framework established by Culler--Vogtmann, one

704: can think of this lemma in terms of compatibility of partitions, in

705: which case the proof is immediate; see

706: \cite{cv}.

707: 

708: %%%

709: 

710: \subsection{Homotopy type} In the remainder of this section, we

711: prove that the star of any rose retracts onto the subcomplex consisting

712: of ``cactus graphs'', and this subcomplex is homeomorphic to a union

713: of $(n-2)$-tori.

714: 

715: We define a \emph{rank $n$ cactus graph} inductively as follows.  A

716: rank 1 cactus graph is a graph with 1 vertex and 1 edge (i.e. a circle

717: with a distinguished point).  In general, a rank $n$ cactus graph is

718: obtained by gluing a rank 1 cactus graph to a rank $n-1$ cactus graph

719: along the vertex of the rank 1 cactus graph.  The set of vertices of

720: the new graph is the union of the sets of vertices of the original two

721: graphs.

722: We note that a rank $n$ cactus graph has exactly $n$ embedded

723: circles, and every edge belongs to exactly one embedded circle

724: (Figure~\ref{figure:not frontier} is an example).

725: 

726: Let $\C(\rho)$ denote the space of cactus graphs in $\str$.

727: Given any $\rho'$, there is a canonical homeomorphism $\C(\rho) \to

728: \C(\rho')$, once we choose orderings of the edges of $\rho$ and

729: $\rho'$.  Thus, we can unambiguously use $\C_n$ to denote the space of

730: cactus graphs in the star of a rose in rank $n$.

731: 

732: In the remainder, assume that $\rho=(\Ga,g)$ is a rose in $\Y_n$ with

733: edges $a_1, \dots, a_n$ labelled $v_1, \dots, v_n$.

734: 

735: \begin{lem}

736: \label{lemma:retraction to cactus graphs}

737: $\str$ strongly deformation retracts

738: onto $\C(\rho)$.

739: \end{lem}

740: 

741: \bpf

742: 

743: For every marked graph in $\str$, the set of edges

744: whose label is not $\pm v_i$ is a forest

745: (Proposition~\ref{lemma2}(\ref{forest})).  We perform a strong deformation

746: retraction of $\str$ by shrinking the edges of each such forest in

747: each marked graph in $\str$.

748: 

749: Consider any marked graph $(\Ga',g')$ in the image of the retraction.

750: By Propositions~\ref{lemma2}(\ref{vi}) and~\ref{lemma2}(\ref{circles}),

751: there is a circle of edges labelled $\pm v_i$ for each $i$.  We

752: consider the ``dual graph'' obtained by assigning a vertex to each

753: such circle (the \emph{circle vertices}) and each intersection point

754: (the \emph{point vertices}) and we connect a point vertex to a circle

755: vertex if the point is contained in the circle.  It follows from

756: Proposition~\ref{lemma1}(\ref{parallel}) that this graph is a tree,

757: and hence $(\Ga',g')$ is a cactus graph.

758: %

759: \epf

760: 

761: \begin{cor}

762: \label{star dim}

763: For $n \geq 2$, the star of any rose $\str$ in $\Y_n$ is homotopy equivalent to a complex of dimension $n-2$.

764: \end{cor}

765: 

766: \bpf

767: 

768: By the definition of cactus graphs, we can see that the dimension

769: increases with slope 1 with respect to dimension, starting at $n=2$.

770: Since $\C_2$ is a point, $\C_n$ is a complex of dimension

771: $n-2$.  An application of Lemma~\ref{lemma:retraction to cactus

772: graphs} completes the proof.

773: %

774: \epf

775: 

776: We can filter $\C(\rho)$ by subsets according to the

777: number of vertices in the cactus graphs:

778: \[ \set{\rho} = V_0 \subset V_1 \subset \cdots \subset V_{n-2} = \C(\rho) \]

779: Each $V_i$ consists of cactus graphs with $i-1$ vertices.

780: 

781: Our goal now is to give a generating set for $\pi_1(\C(\rho))$.

782: Since $V_2$ is simple to understand, the following proposition will

783: make it easy to do this.

784: 

785: \begin{prop}

786: \label{prop:cactus space}

787: There is a cell structure on $\C(\rho)$ so that the $i$-skeleton is

788: exactly $V_i$.

789: \end{prop}

790: 

791: \bpf

792: 

793: We proceed inductively.  The 0-skeleton is one point $V_0 = \set{\rho}$.

794: 

795: Let $i > 0$.  Any marked graph $(\Ga',g')$ in $V_i - V_{i-1}$ lies in

796: a unique $i$-cell $C$.  For each $i$, let $k_i$ be the number of

797: edges of $\Ga'$ labelled $\pm v_i$.  If we reparameterize so that the sum of the

798: lengths of the edges of $\Ga'$ labelled $\pm v_i$ is 1, then we get a

799: $(k_i-1)$-simplex for each $i$, and $C$ is the product of these simplices.

800: 

801: The boundary of $C$ is the set of points where some edge is

802: assigned length 0.  Clearly, $\partial C \subset V_{i-1}$. and so

803: the proposition follows.

804: %

805: \epf

806: 

807: For the remainder of this section, we use the cell structure given by

808: Proposition~\ref{prop:cactus space}, which is different from the cell

809: structure inherited from $\Y_n$.

810: 

811: Let $V_1^1$ be the subset of $V_1$ consisting of graphs with a vertex

812: of valence 4 and a vertex of valence $2n-2$ (i.e. only a single loop

813: is ``travelling'' around another loop).  We will see in

814: Section~\ref{section:finite generation} that the obvious generators

815: for $\pi_1(V_1^1)$ correspond to one of the two types of Magnus

816: generators for $\T_n$.

817: 

818: \begin{prop}

819: \label{cor:gens for cactus space}

820: The subcomplex $V_1^1$ contains a generating set for $\pi_1(\C(\rho))$.

821: \end{prop}

822: 

823: \bpf

824: 

825: First, each of $\C_1$ and $\C_2$ is a single point.  In

826: rank 3, $V_1^1=V_1$.  Thus, in all of these cases, the proposition is

827: vacuously true.  For the remainder, assume $n \geq 4$.

828: 

829: As per Proposition~\ref{prop:cactus space}, $V_1$ can be thought of as

830: the 1-skeleton of the cell complex $\C(\rho)$.  This subcomplex has

831: 1 vertex (the rose $\rho$) and an edge for each combinatorial type of labelled

832: graph with 2 vertices.  We now need to show that any such \emph{standard loop} $\alpha$ in

833: $V_1$ can be written in $\pi_1(\C(\rho))$ as a product of loops in

834: $V_1^1$.  Our strategy is to show that $V_2$ is a union of 2-tori and

835: that $\pi_1(V_1^1)$ surjects onto $\pi_1(V_2)$.

836: 

837: If $(\Ga',g')$ is a point of $V_2-V_1$, then there are two

838: possibilities: the three vertices of $\Ga'$ either lie on the same

839: circle or they do not; see Figure~\ref{figure:v3}.  If they do all lie

840: on some ``central circle'', then we obtain a 2-torus by fixing one

841: vertex and letting the other two vertices ``move around'' the central

842: circle (really we are changing lengths so as to give the appearance of

843: this motion).  In the other case, there are two central circles.  By

844: fixing the middle intersection point and letting the other two

845: intersection points move around the respective circles, we again see a

846: torus.

847: 

848: Consider a standard loop $\al$ of $V_1$.  At an interior point of $\alpha$,

849: there is a central circle with two vertices, and the two vertices have

850: valence, say, $p=p(\al)$ and $q=q(\al)$.  By definition of $V_1$, we

851: have that $p$ and $q$ are even and at least $4$; say $p \leq q$.  We thus

852: have a filtration of $V_1$: $\al$ is in $V_1^k$ if $(p-2)/2

853: \leq k$.  The number $(p-2)/2$ is the number of loops glued to that vertex,

854: other than the central circle.

855: 

856: Now, suppose that $\al$ is a standard loop of $V_1^k$ for some $k

857: \geq 2$.  At any interior point of $\al$, we perform a blowup so that

858: we end up with a graph in $V_2-V_1$ of the first type (left side of

859: Figure~\ref{figure:v3}).  Moreover, we choose the blowup so that (at

860: least) one of the vertices has valence 4.  The fundamental group of the

861: corresponding torus is generated by a standard loop from $V_1^{k-1}$ and a standard loop

862: from $V_1^1$, and so $\al$ can be written as a product of such loops.

863: By induction, $\al$ can be written as a product of loops from $V_1^1$.

864: %

865: \epf

866: \begin{figure}[ht]

867: \includegraphics[scale=.35]{v3.eps}

868: \caption{Two types of graphs in $V_2$.}\label{figure:v3}

869: \end{figure}

870: 

871: \p{Remark} For completeness, we mention that the entire space

872: $\C_n$ can be thought of as a union of $(n-2)$-tori, and the

873: intersection between any two of these tori is a lower dimensional

874: torus which is a product of diagonals of coordinate subtori.  It is

875: straightforward to prove this, given what we have already done.

876: However, we will not need this fact.

877: 

878: %%%

879: %%%

880: %%%

881: 

882: \section{Cohomological Dimension}

883: \label{section:morse function}

884: 

885: We now give the argument for the first part of the main theorem, that

886: $\Y_n$ is homotopically $(2n-4)$-dimensional.  The basic strategy is to

887: put an ordering on the stars of roses of $\Y_n$ (we think of the

888: ordering as a Morse function) and then to glue the stars of roses

889: together in the prescribed order.  This is in the same spirit as the

890: proof of Culler--Vogtmann that $\X_n$ is contractible.

891: 

892: %%%

893: 

894: \subsection{Morse function}  The ordering on roses will come from an

895: ordering on matrices.  We start with vectors.

896: By the \emph{norm} of an element $v=(a_1,\cdots,a_n)$ of $\Z^n$, we mean:

897: \[ |v|=(|a_1|,\cdots,|a_n|)\in \Z_+^n \]

898:  where the elements of $\Z_+^n$ are ordered lexicographically.

899: Consider the matrix:

900: \[ M = \left(\begin{array}{c} v_1 \\ v_2 \\ \vdots \\ v_n

901: \end{array}\right) \]

902: The \emph{norm} of $M$ is:

903: \[ |M| = (|v_n|,\dots,|v_1|) \in (\Z_+^n)^n \]

904: where $(\Z_+^n)^n$ has the lexicographic ordering on the $n$ factors.

905: We say that $M$ is a \emph{standard representative} for an element of

906: $\W\backslash\gl_n(\Z)$ if $|v_n| < \cdots < |v_1|$ (i.e. if it is a

907: representative with smallest norm).  Note that two rows of a matrix in

908: $\gl_n(\Z)$ cannot have the same norm, for otherwise these

909: two rows would be equal after reducing modulo 2, and the resulting

910: matrix would not be invertible.

911: 

912: We declare the norm of an element of $\W\backslash\gl_n(\Z)$ to be the

913: norm of a standard representative, and the norm of a

914: rose in $\Y_n$ to be the norm of the corresponding element of

915: $\W\backslash\gl_n(\Z)$.

916: 

917: In what follows, the following fact will be important:

918: 

919: \begin{lem}

920: \label{lemma:neighbors}

921: If the stars of two roses

922: intersect, then the roses have different norms.

923: \end{lem}

924: 

925: \bpf

926: 

927: If $M$ and $M'$ are marking matrices for neighboring roses, then

928: $M'=NM$, where each entry of $N$ is either $-1$, $0$, or $+1$ (apply Proposition~\ref{lemma2}(\ref{coefficients})).

929: Then, if $|M|=|M'|$, it follows that $N$ is the identity modulo 2, and

930: so $N \in \W$.

931: %

932: \epf

933: 

934: The norm on roses turns the set of roses into a well-ordered set.  We

935: use this fact without mention in the transfinite induction arguments for Theorem~\ref{thm:dim} and Proposition~\ref{proposition:normal generation}.

936: 

937: %%%

938: 

939: \subsection{The induction}

940: \label{subsection:the induction}

941: 

942: We define an \emph{initial segment} of $\Y_n$ to be a union of stars of

943: a set of roses that is closed under taking smaller roses (i.e. a

944: sublevel set of the ``Morse function'' given on stars of roses).  Note

945: that, in general, an initial segment consists of infinitely many

946: roses.  If we show that each initial segment is $(2n-4)$-dimensional,

947: it will follow by transfinite induction that $\Y_n$ has the same

948: property.

949: 

950: To this end, we define the \emph{descending link} of a rose in $\Y_n$

951: to be the intersection of its star with the union of all stars of

952: roses of strictly smaller norm (by Lemma~\ref{lemma:neighbors}, we need not worry about roses of equal norm).  The descending link of a rose $\rho$, denoted $\dlkr$, is a subset of the frontier of its star.  We will prove the following in

953: Section~\ref{section:descending links}:

954: 

955: \begin{prop}

956: \label{desc link dim}

957: For $n \geq 3$, descending links are homotopically $(2n-5)$-dimensional.

958: \end{prop}

959: 

960: Given this, we can prove the first part of the main theorem:

961: 

962: \begin{thm}

963: \label{thm:dim}

964: For $n \geq 3$, the complex $\Y_n$ is homotopy equivalent to a complex of dimension at most $2n-4$.

965: \end{thm}

966: 

967: \bpf

968: 

969: We proceed by transfinite induction on initial segments.  The base step is Corollary~\ref{star dim}.

970: 

971: Whenever we

972: glue the star of a rose $\str$ to an initial segment $\S$ in order to

973: make a new initial segment $\hat \S$, we can think of this as a

974: diagram of spaces:

975: \[ \S \leftarrow \dlkr \rightarrow \str \]

976: By the inductive

977: hypothesis, $\S$ is homotopy equivalent to a $(2n-4)$-dimensional

978: space $\S'$.  Denote by $\str'$ the $(n-2)$-complex homotopy equivalent to $\str$ given by Proposition~\ref{star dim}.  By

979: Proposition~\ref{desc link dim}, the descending link $\dlkr$ is homotopy equivalent to a $(2n-5)$-dimensional space

980: $\dlkr'$.  We choose maps $\dlkr' \to \S'$ and $\dlkr' \to \str'$ so

981: that the following diagram commutes up to homotopy:

982: 

983: \begin{figure}[H]

984: \begin{center}

985: \scalebox{1}{ \xymatrix{\str' \ar@{<->}[r] & \str \\

986: \dlkr'  \ar@{->}[d] \ar@{->}[u] \ar@{<->}[r] & \dlkr  \ar@{->}[d] \ar@{->}[u] \\

987: \S' \ar@{<->}[r] & \S}}

988: \end{center}

989: \end{figure}

990: It follows that the colimit of the diagram of spaces in the left column is homotopy equivalent to the colimit of the diagram of spaces

991: in the right column (see e.g. \cite[Proposition 4G.1]{ah}).  The

992: former, call it $\hat \S'$, is $(2n-4)$-dimensional (consider the double mapping cylinder), and

993: the latter is $\hat \S$.  By construction, the homotopy equivalence $\hat S' \to \hat S$ extends the homotopy equivalence $S' \to S$.

994: 

995: By transfinite induction, we thus build a homotopy model $Z_n$ for

996: $\Y_n$.  By the inductive construction given above, $Z_n$ has a

997: filtration by subcomplexes $\set{Z_n^\al}$, each equipped with a

998: homotopy equivalence $h_\al:Z_n^\al \to \Y_n^\al$ for some some initial

999: segment $\Y_n^\al$.  What is more, the induced map $h:Z_n \to

1000: \Y_n$, when restricted to $Z_n^\al$, is precisely $h_\al$.  It follows

1001: that $h$ is a homotopy equivalence (see, e.g., the discussion

1002: following \cite[Proposition 4G.1]{ah}).  Since $Z_n$ has dimension at

1003: most $2n-4$ (by construction), we are done.

1004: %

1005: \epf

1006: 

1007: \p{Remark.} If one wants to avoid transfinite induction, it is

1008: possible to alter the Morse function so that it is the same locally

1009: (i.e. Proposition~\ref{desc link dim} and its proof do not change)

1010: but the image of the Morse function is order isomorphic to the

1011: positive integers.

1012: 

1013: 

1014: %%%

1015: %%%

1016: %%%

1017: 

1018: \section{Finite generation}

1019: \label{section:finite generation}

1020: 

1021: In this section, we recall the definition of the Magnus generating set

1022: for $\T_n$, and explain how our point of view recovers the result that

1023: these elements do indeed generate $\T_n$

1024: (Theorem~\ref{proposition:finite generation} below).

1025: 

1026: Throughout the section (and the appendix), we denote an element $\phi$ of $\out(F_n)$

1027: by:

1028: \[ [\Phi(x_1), \dots, \Phi(x_n)] \]

1029: where $x_1,\dots,x_n$ are the generators of $F_n$, and $\Phi$ is a

1030: representative automorphism for $\phi$.

1031: 

1032: %%%

1033: 

1034: \subsection{Magnus generators}

1035: 

1036: Magnus proved that $\T_n$ is generated by:

1037: \[ \begin{array}{rcl}

1038: K_{ik} & = & [x_1, \dots, x_kx_ix_k^{-1}, \dots, x_n] \\

1039: K_{ikl} & = & [x_1, \dots, x_i[x_k,x_l], \dots, x_n]

1040: \end{array}\]

1041: for distinct $i$, $k$, and $l$.

1042: 

1043: We can see the $K_{ik}$ as loops in the star of a rose in $\Y_n$.

1044: Consider the picture in Figure~\ref{figure:not frontier}.  As

1045: mentioned in Section~\ref{section:quotient}, shrinking

1046: either of the parallel edges gives a path leading to the rose with the

1047: identity marking, and so this is a loop in the star of that rose in

1048: $\Y_n$.  By considering what is happening on the level of homotopy (as

1049: opposed to homology), we see that this loop is exactly $K_{23}$ (see

1050: \cite{cv}).  By attaching more loops at one of the vertices, and

1051: renaming the edges, we see that we can obtain any $K_{ik}$ in the star

1052: of the identity rose.  In the stars of other roses, the

1053: analogously defined loops are conjugates of the $K_{ik}$.  What is

1054: more, we have:

1055: 

1056: \begin{prop}

1057: \label{proposition:generators of star of a rose}

1058: The fundamental group of the star of the rose with the identity

1059: marking is generated by the $K_{ik}$.

1060: \end{prop}

1061: 

1062: The proposition follows immediately from the fact that the loops in

1063: the above discussion corresponding to the $K_{ik}$ are exactly the

1064: standard generators for $\pi_1(V_1)$ from Proposition~\ref{cor:gens

1065: for cactus space}.

1066: 

1067: %%%

1068: 

1069: \subsection{Proof of finite generation}

1070: 

1071: Our proof that the Magnus generators generate $\pi_1(Y_n) \cong \T_n$

1072: rests on the following two topological facts about descending links

1073: which we prove in Section~\ref{section:descending links}:

1074: 

1075: \begin{prop}

1076: \label{proposition:descending links are nonempty}

1077: Descending links are nonempty, except for that of the rose with the identity marking.

1078: \end{prop}

1079: 

1080: \begin{prop}

1081: \label{proposition:descending links are connected}

1082: Descending links are connected.

1083: \end{prop}

1084: 

1085: Combining Propositions~\ref{proposition:generators of star of a rose},

1086: \ref{proposition:descending links are nonempty},

1087: and~\ref{proposition:descending links are connected} with Van

1088: Kampen's theorem and the transitivity of the action of $\out(F_n)$ on

1089: stars of roses, we see that the fundamental group of any initial segment of $\Y_n$ is normally generated by the $K_{ik}$.  By transfinite induction, we have:

1090: 

1091: \begin{prop}

1092: \label{proposition:normal generation}

1093: $\T_n$ is normally generated by the $K_{ik}$.

1094: \end{prop}

1095: 

1096: The group generated by the $K_{ik}$

1097: is not normal in $\out(F_n)$, as any element

1098: of this subgroup is of the form:

1099: \[ [g_1x_1g_1^{-1},g_2x_2g_2^{-1},\dots,g_nx_ng_n^{-1}] \]

1100: Thus, to find a generating set for $\T_n$, we need to add more

1101: elements.

1102: 

1103: We have the following result of Magnus:

1104: 

1105: \begin{prop}

1106: \label{magnus}

1107: For any $n$, the group generated by

1108: \[ \set{K_{ik},K_{ikl}: i \neq k < l \neq i} \]

1109: is normal in $\out(F_n)$.

1110: \end{prop}

1111: 

1112: It is now easy to prove the following, which is the third part of our

1113: main theorem:

1114: 

1115: \begin{thm}

1116: \label{proposition:finite generation}

1117: $\T_n$ is finitely generated.  In particular, it is generated by $\set{K_{ik},K_{ikl}}$.

1118: \end{thm}

1119: 

1120: Proposition~\ref{magnus} is also one of the steps in

1121: Magnus's proof that the $K_{ik}$ and $K_{ikl}$ generate $\K_n$

1122: \cite{wm}.  For completeness, we give Magnus's proof of

1123: Proposition~\ref{magnus} in the appendix.

1124: 

1125: %%%

1126: 

1127: \subsection{Proof that \boldmath$\T_2$ is trivial} Since there are two

1128: ways to blow up a rank 2 rose, it follows that the star of a rose in

1129: $\Y_2$ is homeomorphic to an interval and that the frontier is

1130: homeomorphic to $S^0$.  If we glue the stars of roses together

1131: inductively according to our Morse function as in

1132: Section~\ref{section:morse function}, then at each stage we are gluing

1133: a contractible space (the star of the new rose) to a contractible

1134: space (the previous initial segment is contractible by induction)

1135: along a contractible space (Propositions~\ref{proposition:descending links

1136: are nonempty} and~\ref{proposition:descending links are connected} and

1137: the fact that the frontier is $S^0$).  It follows that each initial

1138: segment, and hence all of $\Y_2$, is contractible; hence, $\T_2=1$.

1139: 

1140: It is more illuminating to draw a diagram of $\X_2=\Y_2$. It is a

1141: tree, with edges representing stars of roses.  This tree is naturally

1142: dual to the classical Farey graph, with the matrix

1143: $\left(\begin{smallmatrix} a&b\\ c&d\end{smallmatrix}\right)$

1144: corresponding to the unordered pair $\{\frac ba,\frac dc\}$. See

1145: Figure \ref{id18}.

1146: 

1147: \begin{figure}[H]

1148: \centerline{\includegraphics{id-det.18}}

1149: \caption{A part of the Farey graph and the dual tree $\Y_2$.}

1150: \label{id18}

1151: \end{figure}

1152: 

1153: 

1154: %%%

1155: %%%

1156: %%%

1157: 

1158: \section{Descending links}

1159: \label{section:descending links}

1160: 

1161: Recall that the descending link $\dlkr$ of a rose $\rho$ is the

1162: intersection of its star with the union of stars of roses of strictly

1163: smaller norm.  The goal of this section is to prove

1164: Propositions~\ref{proposition:descending links are nonempty},

1165: \ref{proposition:descending links are connected}, and~\ref{desc link

1166: dim}, that descending links are nonempty, connected, and homotopically

1167: $(2n-5)$-dimensional.

1168: 

1169: As in Section~\ref{section:frontier}, let $\rho$ be a rose represented by a marked graph $(\Ga,g)$ whose edges $a_i$ are labelled $v_i$.  We assume the $a_i$ are ordered so that the marking matrix

1170: \[ M = \left(\begin{array}{c} v_1 \\ \vdots \\  v_n \end{array}\right) \]

1171: is a standard representative.

1172: 

1173: %%%

1174: 

1175: \subsection{Descending ideal edges}

1176: 

1177: An ideal edge for $\rho$ is called \emph{descending} if any of the corresponding

1178: 1-edge blowups (Lemma~\ref{lem:1 edge}) lies in $\dlkr$.  Every edge of a marked graph in $\str$ which is not labelled $\pm v_i$ corresponds to some ideal edge; if the corresponding ideal edge is descending, we may say that the edge is descending.

1179: 

1180: We now give a criterion for checking whether or not a particular ideal

1181: edge is descending.

1182: 

1183: \begin{lem}

1184: \label{descending criterion}

1185: Let $\iota = a_1 + a_{i_1} + \cdots + a_{i_m}$ be an ideal edge.  The

1186: following are equivalent:

1187: \begin{enumerate}

1188: \item $\iota$ is descending

1189: \item any of the corresponding 1-edge blowups lies in $\dlkr$

1190: \item all of the corresponding 1-edge blowups lie in $\dlkr$

1191: \item $|v_1 + v_{i_1} + \cdots + v_{i_m}| < |v_1|$

1192: \end{enumerate}

1193: 

1194: Similarly, $\bar \iota = -a_1 + a_{i_1} + \cdots + a_{i_m}$ is descending if and only if

1195: $|v_1 - (v_{i_1} + \cdots + v_{i_m})| < |v_1|$.

1196: \end{lem}

1197: 

1198: \bpf

1199: 

1200: A 1-edge blowup which realizes the ideal edge $\iota$ lies in $m+1$

1201: stars of roses (Proposition~\ref{converse}).  Namely, for each of $v_1, v_{i_1}, \dots v_{i_m}$,

1202: we get a new marking matrix by replacing that vector with:

1203: \[ v_1 + v_{i_1} + \cdots + v_{i_m} \]

1204: and leaving all other row vectors the same.  To see if $\iota$ is

1205: descending, we look at the smallest of these matrices.

1206: We claim that the smallest is:

1207: \[ N = \left(\begin{array}{c} v_1 + v_{i_1} + \cdots + v_{i_m} \\ v_2 \\ \vdots \\  v_n \end{array}\right) \]

1208: Indeed, suppose we had replaced some other row vector, say $v_i$, with

1209: $v_1 + v_{i_1} + \cdots + v_{i_m}$, obtaining a matrix $N'$.  Now,

1210: forgetting the order of the rows, $N$

1211: and $N'$ share $n-1$ rows, and $N$ has the row vector $v_i$ whereas

1212: $N'$ has the row vector $v_1$.  By the assumption that $M$ is a

1213: standard representative, we have $|v_i| < |v_1|$.  Now, if we put

1214: $|N'|$ in standard form, it is easy to find a representative for the

1215: $N$-coset with smaller norm than the standard representative for

1216: $N'$---simply replace the row of $N'$ consisting of $v_1$ with the

1217: vector $v_i$.  The norm of $N$ is less than or equal to the norm of

1218: this representative, so the claim is proven.

1219: 

1220: Now both directions are easy: if $|v_1 + v_{i_1} + \cdots + v_{i_m}| <

1221: |v_1|$ then $|N|$ is obviously strictly less than $|M|$ (the given

1222: representative has smaller norm) and so $\iota$ is

1223: descending; conversely, if $|v_1 + v_{i_1} + \cdots + v_{i_m}| \geq

1224: |v_1|$, then the given representative is in standard form and

1225: obviously has norm at least $|M|$.  (We remark that the last

1226: inequality must be strict by Lemma~\ref{lemma:neighbors}.)

1227: 

1228: The second statement follows by symmetry.

1229: %

1230: \epf

1231: 

1232: It is not hard to prove a stronger statement than the one given here.  However, the relatively simple result given suffices for our purposes, and the generalities are notationally unpleasant.

1233: 

1234: %%%

1235: 

1236: \begin{cor}

1237: \label{lemma:description of descending}

1238: A marked graph in $\str$ is in $\dlkr$ if and only if it realizes a descending ideal edge.

1239: \end{cor}

1240: 

1241: As a consequence of Lemma~\ref{descending criterion}, we see that

1242: there exist pairs of marked graph which can never be simultaneously

1243: descending.

1244: 

1245: \begin{lem}

1246: \label{forbidden pairs}

1247: If the ideal edge $\iota = a_1 + a_{i_1} + \cdots + a_{i_m}$ is descending

1248: then $\bar \iota = -a_1 +

1249: a_{i_1} + \cdots + a_{i_m}$ is not descending.

1250: \end{lem}

1251: 

1252: Generalizations of Lemma~\ref{descending criterion} lead to analogous generalizations of the current lemma.

1253: 

1254: \bpf

1255: 

1256: To simplify notation, let $w_0=v_1$, $w_1 = v_{i_1}$,

1257: $w_2=v_{i_2}$, etc.  We will denote particular entries in each of

1258: these row vectors by using double indices; i.e., $w_{jk}$ is the

1259: $k^{\mbox{\tiny th}}$ entry of the row vector $w_j$.

1260: 

1261: Let $k$ be the smallest number so that

1262: \[ |w_{0k} + w_{1k} + w_{2k} + \cdots + w_{mk}| \neq |w_{0k}| \]

1263: Note that there is such a $k$, for otherwise, the

1264: original matrix $M$ would not be invertible (reduce modulo 2).

1265: 

1266: Applying Lemma~\ref{descending criterion}, we see that $\iota$ is descending if and only if

1267: \begin{equation}\tag{1}

1268: |w_{0k} + w_{1k} + w_{2k} + \cdots + w_{mk}| < |w_{0k}|

1269: \end{equation}

1270: (we are using the minimality of $k$).  It follows that $w_{1k} +

1271: w_{2k} + \cdots + w_{mk}\neq 0$ and that the sign of this sum differs

1272: from that of $w_{0k}$.  Thus, we have:

1273: \begin{equation}\tag{2}

1274: |w_{0k} - (w_{1k} + w_{2k} + \cdots + w_{mk})| > |w_{0k}|

1275: \end{equation}

1276: and so $\bar \iota$ is not descending.  By symmetry, we are done.

1277: %

1278: \epf

1279: 

1280: 

1281: %%%

1282: 

1283: \subsection{Proof of Propositions~\ref{proposition:descending links are nonempty} and~\ref{proposition:descending links are connected}}

1284: 

1285: As usual, let $\rho$ be a rose represented by

1286: a marked graph $(\Ga,g)$ with edges $a_1, \dots, a_n$ labelled by

1287: $v_1, \dots, v_n$, and assume that the edges are ordered so that the

1288: marking matrix

1289: \[ M = \left(\begin{array}{c} v_1 \\ \vdots \\  v_n \end{array}\right) \]

1290: is a standard representative.

1291: 

1292: We first give the proof that descending links are nonempty:

1293: 

1294: \begin{proof}[Proof of Proposition~\ref{proposition:descending

1295:   links are nonempty}]

1296: 

1297: Let $k$ be the first column of $M$ which is not a coordinate vector

1298: (since $M$ is a standard representative, it follows that the entries

1299: in the first $k-1$ column vectors agree with the identity matrix up

1300: to sign).  If we denote the $j^{\mbox{\tiny th}}$ entry of $v_i$ by $v_{ij}$,

1301: then $v_{kk}$ is nonzero.  This follows from the fact

1302: that $M$ is a standard representative and the fact that $M$ is

1303: invertible.

1304: 

1305: Since the $k^{\mbox{\tiny th}}$ column is not a coordinate vector (and since $M$ is

1306: invertible), there is a $j$, different from $k$, so that $v_{jk}$ is

1307: nonzero.  If there is a $j > k$ such that $v_{jk} \neq 0$, then, since

1308: $M$ is a standard representative, $|v_{jk}| \leq |v_{kk}|$, and $a_k +

1309: \epsilon_j a_j$ is a descending ideal edge for some $\epsilon_j = \pm 1$.  If

1310: $v_{jk}=0$ for all $j > k$, it follows that $v_{kk} = \pm 1$ (since

1311: $M$ is invertible) and there is some $j < k$ so that $v_{jk} \neq 0$.

1312: But then, again, $a_j + \epsilon a_k$ is descending for some $\epsilon

1313: = \pm 1$.

1314: %

1315: \epf

1316: 

1317: Here is the proof that descending links are connected:

1318: 

1319: \begin{proof}[Proof of  Proposition~\ref{proposition:descending

1320:   links are connected}]

1321: 

1322: We first claim that if $\iota$ is any descending ideal edge, then

1323: there is a subordinate 2-letter ideal edge $\iota'$ which is also

1324: descending; see Section~\ref{section:ideal edges} for definitions.  It

1325: will then follow from Lemma~\ref{lemma:subordinate} and

1326: Corollary~\ref{lemma:description of descending} that there is a path in

1327: $\dlkr$ between the 1-edge blowup realizing $\iota$ to the 1-edge

1328: blowup realizing $\iota'$ (the graph simultaneously realizing $\iota$

1329: and $\iota'$ is the midpoint of the path).

1330: 

1331: To prove the claim, we need some notation.  First, recall the

1332: notations $\rho$, $a_i$, $v_i$, and $M$ from above.  Also, say

1333: (without loss of generality) that $\iota = a_{i_1} + a_{i_2} + \cdots +

1334: a_{i_m}$, and denote $v_{i_j}$ by $w_j$.  Starting with the matrix

1335: with the $w_i$ as rows, we obtain a matrix $M'$ by deleting all

1336: columns without a nonzero entry.

1337: The $ij^{\mbox{\tiny th}}$ entry

1338: of $M'$ is denoted $w_{ij}$.

1339: 

1340: We proceed in two cases.  If the first column of $M'$ is not a

1341: coordinate vector, then at least two of the $w_{i1}$ are nonzero, in

1342: particular, $w_{11} \neq 0$.  Without loss of generality, say $w_{11}

1343: > 0$.  Since $\iota$ is descending, there must be a $k$ so that

1344: $w_{k1} < 0$, and since $M$ is a standard representative, we have

1345: $|w_{k1}| \leq |w_{11}|$.  It follows that $a_{i_1} + a_{i_k}$ is

1346: descending, and this completes the proof of the first case.

1347: 

1348: If the first column of $M'$ is a coordinate vector (i.e. $w_{11} = \pm

1349: 1$ and $w_{k1} = 0$ for $k > 1$), then we look at the second column of

1350: $M'$.  Without loss of generality, assume $w_{22} > 0$.  At this point

1351: there are three subcases.  If $w_{k2}=0$ for all $k > 2$, then

1352: $a_{i_1}+a_{i_2}$ is descending, since $\iota$ is descending.  If there is a $k > 2$ so that

1353: $w_{k2} < 0$ then $a_{i_2}+a_{i_k}$ is descending (since $M$ is a

1354: standard representative).  If $w_{k2} \geq 0$ for all $k > 2$ and

1355: $w_{k2} \neq 0$ for at least one $k > 2$, then, since $\iota$ is descending, it follows that $w_{12}

1356: < 0$ and so $a_{i_1}+a_{i_k}$ is

1357: descending for any $k > 2$ with $w_k > 0$.

1358: 

1359: We now claim that given any two descending 2-letter ideal edges, there

1360: is a path between the corresponding points in $\dlkr$.  This follows, as above, from

1361: Lemma~\ref{lemma:subordinate} and Corollary~\ref{lemma:description of

1362:   descending}, in addition to the fact that opposite 2-letter ideal

1363: edges cannot both be descending (Lemma~\ref{forbidden pairs}).

1364: This completes the proof.

1365: %

1366: \epf

1367: 

1368: 

1369: %%%

1370: 

1371: \subsection{Completely descending link}

1372: 

1373: We now shift our attention to Proposition~\ref{desc link dim}.  Let

1374: $\rho$ be a rose represented by a marked graph $(\Ga,g)$, and say that

1375: $\Ga$ has edges $a_1, \dots, a_n$ labelled by $v_1, \dots, v_n$.

1376: 

1377: The main argument for the proof (Section~\ref{section:dimension of

1378: descending links} below) is purely combinatorial, referring only to

1379: isomorphism types of labelled graphs.  As things stand, however, we

1380: cannot describe $\dlkr$ in terms of combinatorial graphs

1381: without metrics.  Indeed, given a marked graph in $\dlkr$,

1382: if we shrink the descending edges to have length less than 1 (while

1383: staying in the frontier by enlarging a nondescending edge), then the

1384: resulting marked graph is not in $\dlkr$

1385: (Corollary~\ref{lemma:description of descending}).

1386: 

1387: To remedy this problem we perform a deformation retraction of $\dlkr$

1388: onto the \emph{completely descending link}, which we define to be the

1389: subset of $\dlkr$ consisting of marked graphs where each edge not

1390: labelled $\pm v_i$ is descending.  The deformation retraction is

1391: achieved by simply shrinking all edges which correspond to

1392: nondescending ideal edges.  Recall that these edges form a forest (Proposition~\ref{lemma2}(\ref{forest})),

1393: so there is no obstruction.

1394: We denote the completely descending link of $\rho$ by $\cdlkr$.

1395: 

1396: \begin{lem}

1397: \label{def ret}

1398: For any given rose $\rho$, the completely descending link $\cdlkr$ is

1399: a strong deformation retract of the descending link $\dlkr$.  In particular, the two

1400: are homotopy equivalent.

1401: \end{lem}

1402: 

1403: We see that $\cdlkr$ has the desired cell structure: a cell

1404: is given by a combinatorial type of labelled graph and the cells are

1405: parameterized by the lengths of the edges in the graph.  To be more

1406: precise, let $(\Ga'g')$ be a marked graph in $\cdlkr$, and for each

1407: $i$, let $k_i$ be the number of edges of $\Ga'$ labelled $\pm v_i$.

1408: For each $i$ we thus get a $(k_i-1)$-simplex by projecting

1409: \[ \set{(t_1, \dots, t_{k_i}) \in [0,1]^{k_i}: t_j=1 \mbox{ for some

1410:   } j} \]

1411: to the simplex $\Delta_i = \set{\sum t_i=1}$.  This projection is a

1412: homeomorphism.  For each edge not labelled $\pm v_i$, we allow its

1413: length to vary arbitrarily within $[0,1]$, as long as one such edge

1414: has length 1.  If $k_0$ is the

1415: number of such edges, then, as above, we get a $(k_0-1)$-simplex

1416: $\Delta_0$.  Thus, the cell corresponding to $(\Ga',g')$ has a cell

1417: structure given by the product:

1418: \[ \Delta_0 \times \cdots \times \Delta_n \]

1419: We now summarize some of the important features of this cell

1420: structure:

1421: \begin{prop}

1422: \label{prop:cdlk}

1423: Consider a cell $C$ of $\cdlkr$ as above.

1424: \begin{enumerate}

1425: \item Passing to faces of $C$ corresponds to collapsing forests in $\Ga'$.

1426: \item $C$ is top-dimensional if and only if all vertices of $\Ga'$

1427:   have valence 3.

1428: \item If $\Ga'$ has $v$ vertices, then $C$ has dimension $v-2$.

1429: \end{enumerate}

1430: \end{prop}

1431: 

1432: %%%

1433: 

1434: \subsection{Proof of Proposition~\ref{desc link dim}}

1435: \label{section:dimension of descending links}

1436: 

1437: In this section we show that the completely descending link for any

1438: rose is homotopy equivalent to a complex of dimension $2n-5$ (Proposition~\ref{cdl dim}).  Since

1439: the completely descending link is a deformation retract of the

1440: descending link (Lemma~\ref{def ret}), Proposition~\ref{desc link dim} follows as a corollary.

1441: 

1442: As usual, let $\rho=(\Ga,g)$ be a rose in $\Y_n$, with edges $a_1,

1443: \dots, a_n$ labelled by $v_1, \dots, v_n$.  If $(\Ga',g')$ is any

1444: marked graph in $\str$, we define the \emph{$v_i$-loop} as the image of $a_i$

1445: under a homotopy inverse of the collapsing map $\Ga' \to \Ga$.

1446: 

1447: 

1448: \begin{prop}

1449: \label{cdl dim}

1450: Let $n \geq 3$.  For any rose $\rho$ in $\Y_n$, there is a strong deformation retraction of $\cdlkr$ onto a

1451: complex of dimension $2n-5$.

1452: \end{prop}

1453: 

1454: \bpf

1455: 

1456: If any top-dimensional cell of $\cdlkr$ has a free face in

1457: $\cdlkr$, then there is a homotopy equivalence (deformation

1458: retraction) of $\cdlkr$ which collapses away this cell.  We perform

1459: this process inductively until we arrive at a subcomplex $L$ where no top-dimensional cell

1460: has a free face.

1461: 

1462: We now suppose that $L$ is $(2n-4)$-dimensional,

1463: i.e., it has at least one top-dimensional cell.  Among

1464: these, choose a cell $C$ where the total number of edges

1465: $\ell$ of a $v_1$-loop is minimal.  Call the loop $P$ and choose one

1466: of its edges labelled $\pm v_1$ and call it $e$; see the leftmost

1467: diagram in Figure~\ref{fig:generic}.  Say that $C$ is given

1468: by a marked graph $(\Ga',g')$.

1469: 

1470: Firstly, note that $\ell$ is not 1, since there are no graphs with

1471: separating edges in $\Y_n$.

1472: 

1473: \begin{figure}[ht]

1474: \psfrag{v1}{$v_1$}

1475: \includegraphics[scale=.5]{induction.eps}

1476: \caption{The top edge is $e$.  The horizontal path plus $e$ is $P$.}\label{fig:generic}

1477: \end{figure}

1478: 

1479: If we collapse any edge of $P-e$ (middle of Figure~\ref{fig:generic}),

1480: we move to a codimension 1 face of $C$.  There are two ways to move to

1481: a new top-dimensional cell, since there are two other blowups of the

1482: resulting valence 4 vertex.  One way reduces the length of $P$ (top

1483: right of Figure~\ref{fig:generic}), so by the minimality assumption for

1484: $C$, this is not a cell of $L$.  Since we are assuming $C$ does not

1485: have any free faces, the other top-dimensional cell (bottom right of

1486: Figure~\ref{fig:generic}), call it $C'$, must be in $L$.  The marked

1487: graphs in $C$ and $C'$ have the same labels outside of $P$; the

1488: difference is that the order of the edges leaving $P$ has changed

1489: (Proposition~\ref{lemma1}(\ref{labels dont change}) is applied twice).

1490: 

1491: Continuing in this way, we see that if we permute the edges leaving

1492: $P$ in any way, we arrive at cells which are necessarily part of $L$.

1493: In particular, the graph obtained by taking the edge which leaves $P$

1494: at one endpoint of $e$ and moving it to the other endpoint of $e$

1495: gives a descending cell $\bar{C}$.

1496: 

1497: We now argue that $C$ and $\bar{C}$ are opposite in the sense of

1498: Lemma~\ref{forbidden pairs}.  Consider either endpoint of $e$ in

1499: $\Ga'$.  This is a valence 3 vertex, as shown in Figure~\ref{fig:end

1500: of e}.  By Proposition~\ref{lemma1}(\ref{switch condition}) and

1501: Proposition~\ref{lemma2}(\ref{coefficients}), the labels must be as in

1502: the left hand side of the figure.  When we move the edge labelled

1503: $\sum k_iv_i$ to the other end of $e$ (as above), the labels must be

1504: as shown in the right hand side of the figure; the key point is that

1505: the labels and orientations do not change for $e$ and the edge being

1506: moved.  It is then possible for us to determine the label for the

1507: third edge leaving the vertex where these edges meet.  By

1508: Lemma~\ref{forbidden pairs} and Corollary~\ref{lemma:description of

1509: descending}, we have a contradiction.

1510: %

1511: \epf

1512: 

1513: 

1514: \begin{figure}[ht]

1515: \psfrag{1}{$v_1$}

1516: \psfrag{2}{$\displaystyle \sum_{i\geq 2} k_iv_i$}

1517: \psfrag{3}{$v_1+\displaystyle \sum_{i\geq 2} k_iv_i$}

1518: \psfrag{4}{$v_1-\displaystyle \sum_{i\geq 2} k_iv_i$}

1519: \includegraphics[scale=.75]{endofe.eps}

1520: \caption{Labels at endpoints of $e$.}

1521: \label{fig:end of e}

1522: \end{figure}

1523: 

1524: %%%

1525: %%%

1526: %%%

1527: 

1528: \section{The toy model and infinite generation of top homology}

1529: \label{section:infinite generation}

1530: \label{section:toy model}

1531: 

1532: In this section we prove the second part of the main theorem, that

1533: $H_{2n-4}(\T_n,\Z)$ is not finitely generated when $n \geq 3$.  In order to do this,

1534: we define a subcomplex $\tm_n$ of $\Y_n$, called the ``toy model'', we

1535: find an explicit infinite basis for $H_{2n-4}(\tm_n,\Z)$, and then we

1536: show that the inclusion $\tm_n \to \Y_n$ induces a monomorphism

1537: $H_{2n-4}(\tm_n,\Z) \to H_{2n-4}(\Y_n,\Z)$

1538: (Theorem~\ref{prop:independent tori}).

1539: 

1540: %%%

1541: 

1542: \subsection{Description of the toy model}

1543: \label{section:description of the toy model}

1544: 

1545: Let $\rho=(\Ga,g)$ be the rose in $\Y_n$ with the identity marking,

1546: let $x_i$ denote the edges of the standard rose $R_n$, and let $a_i$ denote the corresponding edges of $\Ga$.

1547: 

1548: Consider the set of points $\tm_n^0=\set{(\Gamma',g')}$ in $\str$ where $g'(x_1) \cup

1549: g'(x_2)$ is a rank 2 rose.  We define the \emph{toy model} to be the

1550: subset $\tm_n$ of $\Y_n$ given by: \[ \tm_n = \bigcup_{p_i,q_i\in \Z} \tm_n^0 \cdot

1551: \left[

1552: \begin{array}{cccccc}

1553: 1 & 0 & p_3 & p_4 & \cdots & p_n \\

1554: 0 & 1 & q_3 & q_4 & \cdots & q_n \\

1555: 0 & 0 & 1& 0 & \cdots & 0 \\

1556: 0 & 0 & 0& 1 & \cdots & 0 \\

1557: \vdots & & & & \ddots & \vdots \\

1558: 0 & 0 & 0& 0 & \cdots & 1 \\

1559: \end{array}

1560: \right ]

1561: \]

1562: 

1563: \p{Another point of view}  We now give a different description of

1564: $\tm_n$, which will make it easier to find its homotopy type.

1565: 

1566: For each marked graph $(\Ga',g')$ of $\tm_n$, the union $g'(x_1) \cup g'(x_2)$

1567: is a rank 2 rose in $\Ga'$, and $\Ga'$ has $n-2$ edges $a_3, \dots,

1568: a_n$ labelled $v_3, \dots, v_n$, where $v_i$ is a coordinate vector

1569: with $+1$ in the $i^{\mbox{\tiny th}}$ spot.

1570: By considering the starting and ending points of $a_3, \dots, a_n$ as

1571: points in $g'(x_1) \cup g'(x_2)$, a path in $\tm_n$ can be thought of

1572: as a path in the configuration space of $n-2$ pairs of points in the

1573: universal abelian cover of $g'(x_1) \cup g'(x_2)$, which is $U = (\R

1574: \times \Z) \cup (\Z \times \R)$.  To make this precise, for each

1575: metric graph $(\Ga',g')$ in $\tm_n$, we rescale the metric so that

1576: $g'(x_1)$ and $g'(x_2)$ both have length 1.  After doing this, the

1577: endpoints of the $a_i$ give a well-defined subset of the metric cover

1578: $U$.

1579: 

1580: If, in the configuration space, we move the two points

1581: corresponding to the endpoints of some $a_i$ by the same integral

1582: vector, then the corresponding point in $\tm_n$ does not change.

1583: 

1584: \begin{prop}

1585: \label{proposition:toy model as config space}

1586: The above construction defines a homeomorphism:

1587: \[ (U^2)^{n-2}/(\Z^2)^{n-2} \to \tm_n \]

1588: \end{prop}

1589: 

1590: At this point, the proof is straightforward and is left to the reader.

1591: 

1592: A typical graph in $\tm_7$ is shown in Figure~\ref{fig:torus

1593: graph}.  That graph is ``maximally blown up'' in the sense that it has

1594: the greatest number of valence 3 vertices possible in $\tm_7$.

1595: 

1596: \begin{figure}[ht]

1597: \includegraphics[scale=.5]{torus.eps}

1598: \caption{A maximally blown up graph in $\tm_7$.}

1599: \label{fig:torus graph}

1600: \end{figure}

1601: 

1602: %%%

1603: 

1604: \subsection{Homotopy type of the toy model}

1605: \label{section:rank 3 toy model}

1606: 

1607: We start by focusing our attention on the rank 3 toy model

1608: $\tm_3$.  In general, we have $\tm_n \cong (\tm_3)^{n-2}$, and so we will be able to deduce the finiteness properties of $\tm_n$ from those of $\tm_3$.

1609: 

1610: Via Proposition~\ref{proposition:toy model as config space}, we can

1611: think of $\tm_3$ as pairs of points in $U$.  However, it will simplify

1612: our analysis if we thicken $U$ to a space $V$, which we now

1613: define.  First, for any integers $p$ and $q$, denote by $D_{p,q}$ the

1614: open disk of radius $r$ around $(p,q) \in \R^2$, for some fixed $r$

1615: close to zero.  Then, define $V = \R^2 - \cup D_{p,q}$.

1616: 

1617: The straight line retraction of $V$ onto $U$ gives a homotopy

1618: equivalence from $V^2/\Z^2$ to $U^2/\Z^2 \cong \tm_3$.  Thinking of

1619: $V^2/\Z^2$ as pairs of points in $V$, we immediately see the following

1620: features:

1621: 

1622: \begin{enumerate}

1623: \item The diagonal of $V^2/\Z^2$ is a torus with one boundary

1624:   component.

1625: \item For each $(p,q)$, there is a

1626:   2-torus $Z_{p,q} = \partial D_{0,0} \times \partial D_{p,q}$.

1627: \end{enumerate}

1628: 

1629: We will now use Morse theory to argue that these features capture

1630: the homotopy type of $\tm_3$.  We consider the Morse function

1631: $d: V^2/\Z^2 \to \R$ which assigns to a point in $V^2/\Z^2$ the

1632: Euclidean distance between the pair of points in $V^2$.

1633: 

1634: \begin{figure}[ht]

1635: \psfrag{minset}{minset}

1636: \psfrag{index 1}{index 1}

1637: \psfrag{index 2}{index 2}

1638: \includegraphics[scale=.4]{critpoints.eps}

1639: \caption{Critical points for the toy model in rank 3.\label{critpointspic}}

1640: \end{figure}

1641: 

1642: We see that $d$ has the following features, depicted in Figure~\ref{critpointspic}:

1643: \blwide

1644: 

1645: \item The minset is the torus with one boundary component

1646: corresponding to the diagonal of $V^2/\Z^2$.

1647: 

1648: \item There is a horizontal 1-cell of index 1 critical points

1649: corresponding to pairs of points lying diametrically

1650: opposite from each other on $\partial D_{0,0}$.

1651: 

1652: \item For every $(p,q)\neq (0,0)$, there is an index 2 critical point,

1653:   corresponding to the two points of tangency of $\partial

1654:   D_{0,0}$ and $\partial D_{p,q}$ with the unique circle tangent to both.

1655: 

1656: \item At all other points, there is a well-defined gradient flow, and

1657:   so there are no other critical points.

1658: \el

1659: 

1660: We notice that each index two critical point is the maximum point of

1661: the corresponding $Z_{p,q}$.  Thus, at each critical point of index 2,

1662: a 2-cycle is added.  One can also see that, at the horizontal 1-cell,

1663: there is another torus being added.  Since there are no critical

1664: points of index greater than 2, these classes are nontrivial in $H_2(\tm_3,\Z)$.

1665: 

1666: In higher rank, we define the torus $Z_{p,q}$ to be the $(n-2)$-fold

1667: product $(\partial D_{0,0} \times \partial D_{p,q})^{n-2}$.

1668: As $\tm_n \cong \tm_3^{n-2}$, we have:

1669: 

1670: \begin{prop}

1671: \label{proposition:generators for top homology of toy model}

1672: The $Z_{p,q}$ freely generate $H_{2n-4}(\tm_n,\Z)$.

1673: \end{prop}

1674: 

1675: To formalize the above argument, one can use Morse theory for

1676: manifolds with corners (see \cite{db}).  There is one technicality:

1677: the critical values of $d$ are not isolated; this is easily overcome

1678: by replacing each $D_{p,q}$ with an ellipse (or proving a more

1679: general Morse theory).

1680: 

1681: We remark that the image of $\pi_1(Z_{p,q}) \cong \Z^{2n-4}$ in

1682: $\pi_1(\Y_n) \cong \T_n$ is a conjugate of the the subgroup $G$

1683: of $\T_n$ described in the introduction.  To see this, one simply

1684: needs to understand the effect of blowups and blowdowns on the

1685: \emph{homotopy} classes of marked graphs; see~\cite{cv}.

1686: 

1687: %%%

1688: 

1689: \subsection{Independence in homology}  We now set out to prove

1690: that the $Z_{p,q}$ represent independent classes in $\Y_n$

1691: (Theorem~\ref{prop:independent tori}).  In particular, this will

1692: prove the second part of the main theorem.

1693: 

1694: By understanding the homotopy equivalences

1695: \[ (V^2)^{n-2}/(\Z^2)^{n-2} \to (U^2)^{n-2}/(\Z^2)^{n-2}

1696: \to \tm_n \]

1697: we can give a concrete description of the $Z_{p,q}$ in terms of marked

1698: graphs.  First of all, we have:

1699: 

1700: \begin{lem}

1701: \label{seg has tori}

1702: Each $Z_{p,q}$ is contained in the union of stars of roses with marking matrices of the form:

1703: \[ \left [

1704: \begin{array}{cccccc}

1705: 1 & 0 & p_3 & p_4 & \cdots & p_n \\

1706: 0 & 1 & q_3 & q_4 & \cdots & q_n \\

1707: 0 & 0 & 1& 0 & \cdots & 0 \\

1708: 0 & 0 & 0& 1 & \cdots & 0 \\

1709: \vdots & & & & \ddots & \vdots \\

1710: 0 & 0 & 0& 0 & \cdots & 1 \\

1711: \end{array}

1712: \right ]

1713: \]

1714: where each $p_i \in [p-1,p+1]$ and $q_i \in [q-1,q+1]$.

1715: \end{lem}

1716: 

1717: The main fact we will need about the stars of roses listed in

1718: Lemma~\ref{seg has tori} is that any ideal edge of the form

1719: \[ \pm v_1 \pm v_2 + \sum_{i \geq 3} k_i v_i \]

1720: is \emph{ascending}.

1721: We will also need the following observation about the $Z_{p,q}$:

1722: 

1723: \begin{lem}

1724: \label{lemma:weakly realized}

1725: For any $Z_{p,q}$ and any rose $\rho$, we can give $Z_{p,q}$ the same cell structure as $\cdlkr$.  In particular, if a marked graph in some $Z_{p,q}$ has an edge of length less than 1 corresponding to an ideal

1726: edge $\iota$, then there is a point in that $Z_{p,q}$ which realizes

1727: $\iota$.

1728: \end{lem}

1729: 

1730: For the remainder, fix a $Z_{p,q}$, and let $\rho$ be a rose of

1731: greatest norm which intersects $Z_{p,q}$.  By Lemma~\ref{seg has

1732: tori}, if $p$ and $q$ are both nonzero, then $\rho$ is unique; if

1733: one of them is zero, then there are $2^{n-2}$ choices for $\rho$; and if

1734: $p=q=0$, then there are $2^{2n-4}$ choices.

1735: From Lemma~\ref{seg has tori}, we deduce the following key fact:

1736: 

1737: \begin{lem}

1738: \label{big torus}

1739: Any map from $\set{Z_{p,q}}$ to roses, sending $Z_{p,q}$ to any rose $\rho$

1740: of maximal norm with $Z_{p,q} \cap \str \neq \emptyset$, is injective.

1741: In particular, given any finite subset of $\set{Z_{p,q}}$, the rose of

1742: maximal norm intersecting this set has nonempty intersection with

1743: exactly one torus in this set.

1744: \end{lem}

1745: 

1746: Let $\dlkr$ denote the descending link for $\rho$ as

1747: defined in Section~\ref{subsection:the induction}.

1748: 

1749: \begin{lem}

1750: \label{torus sphere}

1751: The intersection $Z_{p,q} \cap \dlkr$ is homeomorphic to $S^{2n-5}$.

1752: \end{lem}

1753: 

1754: \bpf

1755: 

1756: We assume $p,q > 0$, with the other cases handled similarly.

1757: 

1758: Under the homotopy equivalence $(V^2)^{n-2}/(\Z^2)^{n-2} \to

1759: (U^2)^{n-2}/(\Z^2)^{n-2}$, we can identify $Z_{p,q}$ with the

1760: configuration space of $n-2$ pairs of points in $U \subset \R^2$

1761: where the first point $z_i$ in each pair lies on the coordinate square

1762: with vertices at $(0,0)$ and $(1,1)$ and the second point $z_i'$ in

1763: each pair lies on the square with vertices $(p,q)$ and $(p+1,q+1)$.

1764: 

1765: The rose $\rho$ is realized when each $z_i$ is at $(0,0)$ and each

1766: $z_i'$ is at the point $(p+1,q+1)$.  The points of $Z_{p,q} \cap

1767: \dlkr$ are exactly the set of points where each $z_i$ is within a

1768: distance of $1/2$ from the origin, each $z_i'$ is within $1/2$ of

1769: $(p+1,q+1)$, and at least one $z_i$ or $z_i'$ has distance exactly $1/2$.

1770: 

1771: In other words, each $z_i$ and $z_i'$ is allowed to move within a

1772: closed interval, and such a configuration is in $Z_{p,q} \cap \dlkr$

1773: if at least one of the points is on the boundary of its interval.

1774: Thus, $Z_{p,q} \cap \dlkr$ is homeomorphic to $\partial I^{2n-4} \cong

1775: S^{2n-5}$.

1776: %

1777: \epf

1778: 

1779: The following completes the proof of the main theorem:

1780: 

1781: \begin{thm}

1782: \label{prop:independent tori}

1783: Let $n\geq 3$.  The $Z_{p,q}$ form an infinite set of independent

1784: classes in $H_{2n-4}(\Y_n,\Z) \cong H_{2n-4}(\T_n,\Z)$.  In other

1785: words, $H_{2n-4}(\tm_n,\Z)$ injects into $H_{2n-4}(\Y_n,\Z)$.

1786: \end{thm}

1787: 

1788: \bpf

1789: 

1790: Given a finite subset $A$ of $\set{Z_{p,q}}$, let $Z_{p,q}$ be an element

1791: which intersects a rose $\rho$ of highest norm (marking matrix as in

1792: Lemma~\ref{seg has tori}, the $i^{\mbox{\tiny th}}$ row corresponds to

1793: $v_i$).  We know that there is a strong deformation retraction of

1794: $\dlkr$ onto a complex of dimension $2n-5$ (Lemma~\ref{def ret} plus

1795: Proposition~\ref{cdl dim}).  The goal is to show that the image of the

1796: sphere $Z_{p,q} \cap \dlkr$ is embedded in this complex.  Even better,

1797: we will show that the deformation retractions of Lemma~\ref{def ret}

1798: and Proposition~\ref{cdl dim} do not move the points of $Z_{p,q} \cap

1799: \dlkr$.

1800: 

1801: To see why this proves this proposition, we consider the long exact

1802: sequence associated to the pair $(\str,\dlkr)$:

1803: \[

1804: \begin{array}{l} \cdots \to H_{2n-4}(\str) \to H_{2n-4}(\str,\dlkr)

1805:  \\  \qquad

1806: \to H_{2n-5}(\dlkr) \to H_{2n-5}(\str) \to \cdots

1807: \end{array}

1808:  \]

1809: By excision, $Z_{p,q}$ corresponds to a class in

1810: $H_{2n-4}(\str,\dlkr)$.  The image in $H_{2n-5}(\dlkr)$ is the class

1811: $Z_{p,q} \cap \dlkr$, which is nontrivial once we show $Z_{p,q} \cap

1812: \dlkr$ is embedded in the

1813: $(2n-5)$-dimensional deformation retract of $\dlkr$ (Lemma~\ref{torus sphere}).  It follows

1814: that  $Z_{p,q}$ is nontrivial in $H_{2n-4}(\str,\dlkr)$ and hence, via

1815: excision, in $H_{2n-4}(\seg',\seg)$ where $\seg$ is the largest initial segment

1816: not containing $\rho$ and $\seg'=\seg\cup\str$.  By Lemma~\ref{big

1817:   torus}, each element of $A$ other than $Z_{p,q}$ is trivial in

1818: $H_{2n-4}(\seg',\seg)$, and so $Z_{p,q}$ is linearly independent from

1819: these, which is what we wanted to show.

1820: 

1821: \begin{comment}

1822: Here is why this proves the proposition: by excision, $Z_{p,q}$

1823: corresponds to a class in $H_{2n-4}(\str,\dlkr)$

1824: 

1825:  since $Z_{p,q} \cap \dlkr$ is homeomorphic to $S^{2n-5}$ (Lemma~\ref{torus sphere}) and the deformation retraction of $\dlkr$ is $(2n-5)$-dimensional, it will follow that $Z_{p,q} \cap \dlkr$ represents a nontrivial class in $H_{2n-5}(\dlkr)$.  Then, we consider the long exact sequence associated to the pair $(\str,\dlkr)$:

1826: \[

1827: \begin{array}{l} \cdots \to H_{2n-4}(\str) \to H_{2n-4}(\str,\dlkr)

1828:   \\ \qquad \to

1829:   H_{2n-5}(\dlkr) \to H_{2n-5}(\str) \to \cdots

1830: \end{array}

1831:  \]

1832: The first term shown is trivial (Proposition~\ref{star

1833: dim}), and so (by excision) $Z_{p,q}$ corresponds to a nontrivial class in

1834: $H_{2n-4}(\str,\dlkr)$.  But by excision, this is a nontrivial class

1835: in $H_{2n-4}(\seg',\seg)$ where $\seg$ is the largest initial segment

1836: not containing $\rho$ and $\seg'=\seg\cup\str$.  It follows that

1837: $Z_{p,q}$ is not in the space spanned by the tori contained in $\seg$,

1838: and this gives the proposition.

1839: 

1840: \end{comment}

1841: 

1842: Thus, we are reduced to showing that the two deformation retractions do not move the sphere $Z_{p,q} \cap \dlkr$.  We handle each in turn.

1843: 

1844: For the deformation retraction of $\dlkr$ onto $\cdlkr$, we need to

1845: show that $Z_{p,q} \cap \dlkr$ is already contained in $\cdlkr$.

1846: Suppose that there were a point of $Z_{p,q} \cap \dlkr$ which were not

1847: contained in $\cdlkr$.  By

1848: Lemma~\ref{lemma:weakly realized}, there is a point which realizes an

1849: \emph{ascending} ideal edge (Lemma~\ref{lemma:neighbors}), and this implies that $Z_{p,q}$ intersects some rose

1850: of higher norm, contradicting the choice of $\rho$.

1851: 

1852: We now focus on the deformation retraction of Proposition~\ref{cdl

1853: dim}.  A marked graph representing a maximal cell of $\cdlkr$ must

1854: have disjoint $v_1$- and $v_2$-loops.  Indeed, there are no valence 4 vertices, so the overlap would have to contain an ascending edge by

1855: Proposition~\ref{lemma1}(\ref{switch condition}), the statement after

1856: Lemma~\ref{seg has tori}, and Lemma~\ref{descending criterion}.  The

1857: codimension 1 cells which get collapsed are obtained from these

1858: maximal cells by collapsing an edge of the $v_1$-loop.  Thus, the

1859: corresponding graphs still have disjoint $v_1$- and $v_2$-loops.  On

1860: the other hand, in any graph of $Z_{p,q}$, the $v_1$-loop and the

1861: $v_2$-loop intersect in exactly 1 point.  Thus, no points of $Z_{p,q}$

1862: are moved during this retraction, so we are done.

1863: %

1864: \epf

1865: 

1866: 

1867: %%%

1868: %%%

1869: %%%

1870: 

1871: \appendix

1872: 

1873: \section*{Appendix: Proof of Proposition~\ref{magnus}}

1874: \label{section:magnus}

1875: 

1876: This appendix contains Magnus's proof of Proposition~\ref{magnus}.  In

1877: this section, we freely use the notation of

1878: Section~\ref{section:finite generation}.

1879: 

1880: Let $K$ be the subgroup of $\out(F_n)$ generated by the $K_{ik}$ and

1881: $K_{ikl}$ for distinct $i$, $k$, and $l$.  We now prove

1882: Proposition~\ref{magnus}, that $K$ is normal in $\out(F_n)$.

1883: 

1884: \bpf[Proof of Proposition~\ref{magnus}]

1885: 

1886: We choose the following generating set for $\out(F_n)$:

1887: \[ \begin{array}{rcl}

1888: \delta_{12} & = & [x_1 x_2, x_2, \dots, x_n] \\

1889: \Omega_1 & = & [x_1^{-1}, x_2, \dots, x_n] \\

1890: \Pi_{i-1} & = & [x_1, \dots, x_{i-2}, x_i, x_{i-1}, x_{i+1},

1891:     \dots, x_n]

1892: \end{array}\]

1893: 

1894: It suffices to show that the conjugates of the $K_{ik}$ and

1895: $K_{ikl}$ by the chosen generators of $\out(F_n)$ (and their

1896: inverses) are elements of $K$.

1897: 

1898: We have the following simplifications:

1899: 

1900: \begin{enumerate}

1901: 

1902: \item Operations on disjoint sets of elements commute.

1903: 

1904: \item Since $\Omega_1$ and $\Pi_{i-1}$ have order 2, we don't need to

1905: conjugate by their inverses.

1906: 

1907: \item We don't need to conjugate by $\delta_{12}^{-1}$ since \[ (\Pi_1

1908: \Omega_1 \Pi_1) \delta_{12} (\Pi_1 \Omega_1 \Pi_1)^{-1} =

1909: \delta_{12}^{-1} \]

1910: 

1911: \item Since $K_{ikl} = K_{ilk}^{-1}$, we may assume $k < l$.

1912: 

1913: \item  Any outer automorphism $\psi$ of the form

1914: \[ [x_1, \dots, g' x_i g, \dots, x_n] \]

1915: where $gg'$ is an element of the commutator subgroup of the subgroup $H$

1916: of $F_n$ generated by $\set{x_k:k \neq i}$ is an element of $K$.

1917: 

1918: \bigskip

1919: 

1920: To see that $\psi \in K$, first note that, by postcomposing with a product of $K_{i\star}^{\pm 1}$,

1921: we may assume that $g'=1$.  Now, we know that the commutator subgroup of $H$ is normally generated

1922: by the $[x_k,x_l]$, where $k$ and $l$ are both different from $i$.

1923: Therefore, it suffices to handle the case of

1924: \[ g = h [x_k,x_l] h^{-1} = [hx_kh^{-1},hx_lh^{-1}] \]

1925: where $h = x_{i_1} \cdots x_{i_p}$ is an arbitrary element of $H$.

1926: It is elementary to check that

1927: \[ \psi = P^{-1} K_{ikl} P \]

1928: where

1929: \[ P = \prod_{j \neq i} K_{ji_p} \cdots \prod_{j \neq i} K_{ji_2}

1930: \prod_{j \neq i} K_{ji_1} \]

1931: \end{enumerate}

1932: 

1933: \bigskip

1934: 

1935: Given these simplifications, it is straightforward to check (case by case) that the

1936: conjugates by $\Pi_{i-1}$, $\Omega_1$, and $\delta_{12}$ of each

1937: $K_{ik}$ and $K_{ikl}$ are elements of  $K$.  There is one exception; we give Magnus's

1938: computation for this difficult case here:

1939: \[ \delta_{12} K_{2l1} \delta_{12}^{-1} =

1940: K_{l2}K_{l1}^{-1}K_{1l}K_{l1}K_{2l1}K_{12l}K_{l2}^{-1}K_{2l}^{-1} \]

1941: %

1942: \epf

1943: 

1944: 

1945: \bibliographystyle{plain}

1946: \bibliography{tn}

1947: 

1948: \end{document}