1: \input amstex.tex
2: \documentstyle{amsppt}
3: \input graphicx.tex
4:
5: \magnification=1200
6: \hsize=150truemm
7: \vsize=224.4truemm
8: \hoffset=4.8truemm
9: \voffset=12truemm
10:
11: \TagsOnRight
12: \NoBlackBoxes
13: \NoRunningHeads
14:
15: \def\Square{\rlap{$\sqcup$}$\sqcap$}
16: \def\cqfd {\quad \hglue 7pt\par\vskip-\baselineskip\vskip-\parskip
17: {\rightline{\Square}}}
18: \redefine\qed{\quad$\Square$}
19:
20: \define\E{{\Cal E}}% \E E gothique
21: \redefine\O{{\Cal O}}% \O O gothique
22: \redefine\P{{\Cal P}}% \P P gothique
23: \define\T{{\Cal T}}% \T T gothique
24: \redefine\D{{\Cal D}}% \D D gothique
25: \define\G{{\Cal G}}% \G G gothique
26: \define\Q{{\bold Q}}% \Q Q mathematique
27: \define\R{{\bold R}}% \R R mathematique
28: \define\Z{{\bold Z}}% \Z Z mathematique
29: \define\mi{^{-1}}% \mi moins un
30: \let\thm\proclaim% \thm \proclaim
31: \let\fthm\endproclaim% \fthm \endproclaim
32: \let\inc\subset % \inc inclus dans (\subset)
33: \let\ds\displaystyle% \ds \displaystyle
34: \let\ev\emptyset% \ev ensemble vide (\emptyset)
35: \let\ov\overline
36: \define\Aut{\text{\rm Aut}}
37: \define\Out{\text{\rm Out}}
38:
39: \newcount\tagno
40: \newcount\secno
41: \newcount\subsecno
42: \newcount\stno
43: \global\subsecno=1
44: \global\tagno=0
45: \define\ntag{\global\advance\tagno by 1\tag{\the\tagno}}
46:
47: \newcount\figno
48: \newcount\fihno
49: \global\figno=0
50: \global\fihno=1
51: \define\fig{\global\advance\figno by 1 \global\advance\fihno by 1
52: {\the\figno}}
53: \define\sta{\the\secno.\the\stno
54: \global\advance\stno by 1}
55:
56: \define\sect{\global\advance\secno by
57: 1\global\subsecno=1\global\stno=1\
58: \the\secno. }
59:
60: \def\nom#1{\edef#1{\the\secno.\the\stno}}
61: \def\eqnom#1{\edef#1{(\the\tagno)}}
62:
63: \newcount\refno
64: \global\refno=0
65: \def\nextref#1{\global\advance\refno by 1\xdef#1{\the\refno}}
66: \def\bref {\ref\global\advance\refno by 1\key{\the\refno}}
67:
68: \nextref\BK
69: \nextref\BS
70: \nextref\Cld
71: \nextref\Clc
72: \nextref\Coll
73: \nextref\Collev
74: \nextref\CV
75: \nextref\Fe
76: \nextref\Fed
77: \nextref\Forr
78: \nextref\Foc
79: \nextref\For
80: \nextref\GHMR
81: \nextref\Gu
82: \nextref\GL
83: \nextref\GLL
84: \nextref\Krn
85: \nextref\Kr
86: \nextref\KV
87: \nextref\LGD
88: \nextref\Ler
89: \nextref\Lrang
90: \nextref\LL
91: \nextref\McM
92: \nextref\Mold
93: \nextref\MSW
94: \nextref\Pett
95: \nextref\Pet
96: \nextref\RV
97: \nextref\RR
98: \nextref\Wh
99:
100: \topmatter
101:
102: \abstract A generalized Baumslag-Solitar group (GBS group) is a finitely
103: generated group $G$ which acts on a tree with all edge and vertex stabilizers
104: infinite cyclic. We show that Out$(G)$ either contains non-abelian
105: free groups or is virtually nilpotent of class $\le2$. It has torsion only at
106: finitely many primes.
107:
108: One may decide algorithmically whether Out$(G)$ is virtually nilpotent or
109: not. If it is, one may decide whether it is virtually abelian, or
110: finitely generated. The isomorphism problem is solvable among GBS groups
111: with Out$(G)$ virtually nilpotent.
112:
113: If $G$ is unimodular (virtually $F_n\times{\bold Z}$), then Out$(G)$ is
114: commensurable with a semi-direct product ${\bold Z}^k\rtimes\text{\rm
115: Out}(H)$ with
116: $H$ virtually free.
117:
118: \endabstract
119:
120: \title On the automorphism group of generalized Baumslag-Solitar groups
121: \endtitle
122:
123: \author Gilbert Levitt
124: \endauthor
125:
126: \toc
127: \widestnumber\head{0}
128: \head 1. Introduction and statement of results\endhead
129: \head 2. Basic facts about GBS groups\endhead
130: \head 3. The automorphism group of a GBS tree\endhead
131: \head 4. Unimodular groups\endhead
132: \head 5. The deformation space\endhead
133: \head 6. Free subgroups in $\Out(G)$\endhead
134: \head 7. Groups with $\Out(G)\not\supset F_2$\endhead
135: \head 8. Further results\endhead
136: \head { } References\endhead
137: \endtoc
138:
139: \endtopmatter
140:
141: \document
142:
143: \head \sect Introduction and statement of results \endhead
144:
145: The groups $BS(m,n)=\langle a,t\mid ta^mt\mi=a^n\rangle$ were introduced
146: by Baumslag-Solitar [\BS] as very simple examples of
147: non-Hopfian groups (a group $G$ is non-Hopfian if there exists a non-injective
148: epimorphism from $G$ to itself). It is now known that
149: $BS(m,n)$ is Hopfian if and only if $m=\pm1$, or $n=\pm1$, or $m,n$ have the
150: same set of prime divisors [\Collev]. In particular,
151: $BS(2,4)$ is Hopfian while $BS(2,3)$ is not.
152:
153: Though it has exotic epimorphisms, $BS(2,3)$ has very few automorphisms:
154: its automorphism group is generated by inner automorphisms and
155: the obvious involution sending $a$ to $a\mi$ [\Coll, \GHMR]. On the other hand,
156: $BS(2,4)$ has an incredible number of automorphisms, as {\it its
157: automorphism group is not finitely generated\/} [\Collev].
158:
159: The reason behind this drastic difference is that, because 2 divides 4 but not
160: 3, the presentation of
161: $BS(2,4)$ is much more flexible than that of $BS(2,3)$. By this we mean, in
162: particular, that $BS(2,4)$ admits the infinite sequence of
163: presentations
164: $$BS(2,4)=\langle a, b,t\mid tb t\mi=b{}^2,
165: b^{2^p}=a^2\rangle\tag{$1_p$}$$
166: obtained from the standard one by introducing a new generator
167: $b=t^{-p}a^2t^p$. It is clear already from $(1_p)$ that
168: $G=BS(2,4)$ has many automorphisms, as fixing $b ,t$ and
169: conjugating $a$ by $b$ defines an element of order $2^p$ in
170: $\Out(G)$.
171:
172: The presentations $(1_p)$ express $BS(2,4)$ as a generalized
173: Baumslag-Solitar group, or {\it GBS group\/}, or graph of $\Z$'s, namely as
174: the fundamental group of a finite graph of groups $\Gamma $ with all edge and
175: vertex groups infinite cyclic. This is visualized as a {\it labelled graph\/}, with
176: the absolute value of the labels indicating the index of edge groups in vertex
177: groups (see Figure \fig).
178:
179: \midinsert
180: \centerline
181: {\includegraphics[scale=.5]
182: {fig1.eps}}
183: \captionwidth{220pt}
184: \botcaption
185: {Figure \the\figno}
186: {The labelled graphs associated to the standard presentation of
187: $BS(m,n)$, and to $(1_p)$. }
188: \endcaption
189: \endinsert
190:
191:
192: In this paper, we study automorphisms of GBS groups. See [\Foc, \For, \Kr,
193: \Lrang,
194: \Wh] for various algebraic and geometric properties of these groups. As
195: pointed out in [\Foc], they are especially interesting in connection with JSJ
196: theory.
197:
198: Before giving general results,
199: let us review certain classes of GBS groups for which more specific
200: statements may be obtained. They are defined either by ``local'' conditions on
201: the labelled graph, or by ``global'' algebraic conditions on the group. In the
202: rest of this introduction, we always assume that $G$ is not one of the
203: elementary GBS groups: $\Z$,
204: $\Z^2$, and the Klein bottle group.
205:
206: \subhead{Algebraically rigid groups}\endsubhead
207:
208: As evidenced by the example of $BS(2,4)$, the main difficulty with GBS groups
209: is that they may be represented by many different labelled
210: graphs $\Gamma $. Sometimes, though, $\Gamma $ is essentially unique. By
211: [\GHMR, \Pet ], this {\it algebraic rigidity\/} holds in particular when there is
212: no {\it divisibility relation\/} in
213: $\Gamma
214: $: if $p,q$ are labels near the same vertex, then $p$ does not divide $q$
215: (see Section 2 for a precise definition and a characterization of algebraic
216: rigidity).
217:
218: Given $\Gamma $, let $T$ be the associated Bass-Serre tree, which we call a
219: {\it GBS tree\/}. Let $\Out^T(G)\inc \Out(G)$ be
220: the subgroup leaving $T$ invariant. Most elements of $\Out^T(G)$
221: may be viewed as ``twists'' (see Section 3). Algebraic rigidity implies
222: $\Out^T(G)=\Out(G)$, but in general $\Out^T(G)$ is smaller.
223:
224: \nom\un
225: \thm{Theorem \sta} Let $G$ be a GBS group, represented by a labelled
226: graph $\Gamma $, and let
227: $T$ be the Bass-Serre tree. Define $k$ as the first Betti number $b$ of
228: $\Gamma
229: $ if $G$ has a nontrivial center, as $b-1$ if the center is trivial.
230: \roster
231: \item The
232: torsion-free rank of the abelianization of $G$ is $k+1$.
233: \item
234: The group $\Out^T(G)$ is virtually $\Z^k$.
235: \item Up to
236: commensurability within $\Out(G)$, the subgroup $\Out^T(G)$
237: does not depend on
238: $\Gamma $.
239: \endroster
240: \fthm
241:
242: Conversely, any subgroup of $\Out(G)$
243: commensurable with a subgroup of
244: $\Out^T(G)$ is contained in $\Out^{T'}(G)$ for some GBS tree $T'$ [\Clc].
245:
246: For $G=BS(m,n)$, one has $k=0$ if $m\neq n$,
247: and
248: $k=1$ if $m=n$. For $G=BS(2,4)$ with the presentation $(1_p)$, the group
249: $\Out^T(G)$ has order $2^{p+1}$.
250:
251: \thm{Corollary \sta} If $G$ is algebraically rigid, then $\Out(G)$ is virtually
252: $\Z^k$.
253: \fthm
254:
255: The converse is also true (see Theorem 8.5).
256:
257: \subhead Unimodular groups\endsubhead
258:
259: A GBS group $G$ is {\it unimodular\/} if $xy^px\mi=y^q$ with $y\neq1$ implies
260: $|p|=|q|$, or equivalently if $G$ is virtually $F_n\times\Z$ (with $F_n$ a free
261: group of rank $n$). The group $G$ then has a normal infinite cyclic subgroup
262: with virtually free quotient, and we show:
263:
264: \thm{Theorem \sta} If $G$ is unimodular, there is a split
265: exact sequence
266: $$\{1\}\to
267: \Z^{k}\to\Out_0(G) \to \Out_0(H)
268: \to\{1\},$$ where $k$ is as above, $H$ is virtually free, and
269: $\Out_0$ has finite index in $\Out$.
270: \fthm
271:
272: Since $\Out(H)$ is VF [\KV], we get:
273:
274: \thm{Corollary \sta} $\Out(G)$ is virtually torsion-free and VF (it has a finite
275: index subgroup admitting a finite classifying space).
276: \fthm
277:
278: \subhead Groups with no nontrivial integral modulus\endsubhead
279:
280:
281: Now consider groups $G$ which do not contain a solvable Baumslag-Solitar
282: group
283: $BS(1,n)$ with $n\ge2$ (there is an equivalent characterization in terms of
284: the modular homomorphism $\Delta :G\to\Q^*$, see Section 2).
285:
286: Given any GBS group $G$, the group $\Out(G)$ acts on the space $P\D$ of all
287: GBS trees (see Section 5), with stabilizers virtually $\Z^k$ by Theorem \un.
288: Clay [\Cld] proved that the space $P\D$ is contractible (see also [\GLL]), and
289: Forester [\For] proved that the quotient is a finite complex if $G$ does not
290: contain
291: $BS(1,n)$ with $n\ge2$. This gives:
292:
293: \thm{Theorem \sta} If $G$ does not contain
294: $BS(1,n)$ for $n\ge2$, then $\Out(G)$
295: is $F_\infty$ (in particular, it is finitely presented). If furthermore $\Out(G)$
296: is virtually torsion-free, then it is VF.
297: \fthm
298:
299: \subhead{Arbitrary groups}
300: \endsubhead
301:
302: Now let $G$ be any GBS group.
303:
304: \nom\six
305: \thm{Theorem \sta} Either $\Out(G)$ contains a nonabelian free group, or it is
306: virtually nilpotent of class $\le2$.
307: \fthm
308:
309: A group is nilpotent of class $\le2$ if and only if every commutator is central.
310: As an example, let $G=\langle a,s,t\mid sa=as, ta^2=a^2t\rangle$ (see Figure
311: \fig). Then
312: $\Out(G)$ is virtually the integral Heisenberg group $H_3$, with
313: $\pmatrix1&i&j\\0&1&k\\0&0&1\endpmatrix$ mapping $(a,s,t)$ to
314: $(a,sa^k,ts^ia^j)$.
315:
316: \midinsert
317: \centerline
318: {\includegraphics[scale=.5]
319: {fig2.eps}}
320: \botcaption
321: {Figure \the\figno}
322: {$\Out(G)$ is virtually the integral Heisenberg group $H_3$. }
323: \endcaption
324: \endinsert
325:
326: Which possibility of Theorem \six{} occurs may be explicitly decided from
327: the divisibility relations in
328: any labelled graph $\Gamma $ representing $G$. We have seen that $\Out(G)$
329: is virtually abelian if there is none. A key observation is that certain
330: divisibility relations force the existence of $F_2$ inside $\Out(G)$.
331:
332: \midinsert
333: \centerline
334: {\includegraphics[scale=.5]
335: {fig3.eps}}
336: \captionwidth{220pt}
337: \botcaption
338: {Figure \the\fihno}
339: {$\Out(G)$ contains $F_2$. }
340: \endcaption
341: \endinsert
342:
343:
344: As a basic example, consider $G=\langle
345: a,b,c\mid a^5=b^4, b^2=c^3\rangle$ (see Figure
346: \fig). It is the amalgam of $G_1=\langle
347: a\rangle$ with $G_2=\langle b,c\rangle$ over $C=\langle b^4\rangle$. The
348: divisibility relation $2|4$ at the middle vertex implies that $C$ is central in
349: $G_2$. For any
350: $g\in G_2$, we may therefore define an automorphism $\varphi _g$ of $G$ as
351: being the identity on
352: $G_1$ and conjugation by $g$ on $G_2$. It is easy to show that the subgroup of
353: $\Out(G)$ generated by the $\varphi _g$'s is isomorphic to $\langle
354: b,c\mid b^2=c^3=1\rangle$, hence contains $F_2$.
355:
356: To prove Theorem \six{}, we assume that $\Out(G)$ does not contain $F_2$ and
357: we describe which divisibility relations may occur (Section 6).
358: In Section 7, we show that, though the
359: GBS tree
360: $T$ may not be
361: $\Out(G)$-invariant, some (non GBS) tree $S$ obtained from $T$ by collapsing
362: certain edges is. We then prove that $\Out^S (G) $
363: is virtually nilpotent.
364:
365: \midinsert
366: \centerline
367: {\includegraphics[scale=.5]
368: {fig4.eps}}
369: \captionwidth{210pt}
370: \botcaption
371: {Figure \the\fihno}
372: {$\Out(G)$ is virtually an infinitely generated abelian group if $n$ is not a
373: power of
374: $2$. }
375: \endcaption
376: \endinsert
377:
378:
379: For instance, let $G=\langle a,b,t\mid tbt\mi=b^2,b^n=a^2\rangle$ with $n$
380: not a power of 2 (see Figure
381: \fig). In this case,
382: $S$ is obtained from $T$ by collapsing edges projecting onto the loop of
383: $\Gamma
384: $. The group
385: $\Out(G)$ is virtually abelian (but not finitely generated).
386:
387: A special case of Theorem \six{} is:
388:
389: \thm{Theorem \sta} If no label of $\Gamma $ equals 1, then $\Out(G)$ contains
390: $F_2$ or is a finitely generated virtually abelian group.
391: \fthm
392:
393:
394: As a corollary of our analysis, we show:
395:
396: \thm{Theorem \sta} The isomorphism problem is solvable for GBS groups $G$
397: such that $\Out(G)$ does not contain $F_2$.
398: \fthm
399:
400: The isomorphism problem for GBS groups is to decide whether two
401: labelled graphs define isomorphic groups. It is solvable for groups with no
402: nontrivial integral modulus [\For] and $2$-generated groups [\Lrang], but
403: open in general.
404:
405:
406: We also show:
407:
408: \nom\huit
409: \thm{Theorem \sta} The set of prime numbers $p$ such that $\Out(G)$
410: contains non-trivial $p$-torsion is finite.
411: \fthm
412:
413:
414: The paper is organized as follows. In Section 2, we review basic
415: properties of GBS groups, such as algebraic rigidity and the modular
416: homomorphism $\Delta $. We extend to GBS groups a
417: result of [\Fed] about twisted conjugacy classes.
418: In Section 3, we study $\Out^T(G)$, proving Theorems \un{} and \huit. Section
419: 4 is devoted to unimodular groups, Section 5 to the action of $\Out(G)$ on
420: $P\D$. Theorem \six{} is proved in Sections 6 and 7. In Section 8 we discuss
421: several special cases and the isomorphism problem.
422:
423: \smallskip
424:
425:
426: \vskip8pt
427: {\baselineskip=9pt
428: {\eightpoint Acknowledgments. {\it I am grateful to M\. Forester and M\. Clay
429: for helpful suggestions, and to P\. Papasoglu for pointing out that
430: $\Out(BS(2,4))$ is not finitely generated.}}\par}
431:
432: \head \sect Basic facts about GBS groups \endhead
433:
434: \subhead Labelled graphs \endsubhead
435:
436:
437: A GBS group $G$ is the fundamental group of a finite graph of groups
438: $\Gamma $ whose vertex and edge groups are all infinite cyclic. It is
439: torsion-free.
440:
441: We denote by $b$ the first Betti number of the graph $\Gamma $. Note
442: the distinction between $G= \pi _1(\Gamma )$ and the topological
443: fundamental group $\pi _1^{top}(\Gamma )\simeq F_b$.
444:
445: If we choose generators for edge and vertex groups, the inclusion
446: maps are multiplications by non-zero integers. An oriented edge $e$
447: thus has a label $\lambda _e\in\Z\setminus\{0\}$, describing the inclusion of
448: the edge group $G_e$ into the vertex group $G_{o(e)}$ at the origin of $e$.
449: As in [\For], we visualize the graph of groups as a {\it labelled graph\/}
450: $\Gamma $, with the label $\lambda _e $ pictured near the origin
451: $o(e)$.
452:
453: A pair
454: $\varepsilon =(e,\ov e)$ of opposite edges is a non-oriented edge. It
455: carries two labels, one near either endpoint $o(e),o(\ov e)$, and we
456: say that $\varepsilon $ (or $e$) is a {\it $(\lambda _e,\lambda _{\ov
457: e})$-edge\/}. An edge is a {\it loop\/} if its endpoints are
458: equal, a {\it segment\/} if they are distinct.
459:
460: The group $G$ associated to a labelled graph $\Gamma $ may be
461: presented as follows. Choose a maximal subtree $\Gamma _0\inc
462: \Gamma $. There is one generator $x_v$ for each vertex $v$, and one
463: generator $t_\varepsilon $ for each non-oriented edge
464: $\varepsilon$ not in $\Gamma _0$. Each non-oriented edge
465: $\varepsilon =(e,\ov e)$ of $\Gamma $
466: contributes one relation. If $ \varepsilon $ is contained in
467: $\Gamma _0$, the relation is $(x_{o(e)})^{{\lambda (e)}}= (x_{o(\ov
468: e)})^{{\lambda (\ov e)}}$. If $\varepsilon $ is not in $\Gamma _0$, the
469: relation is
470: $t_e(x_{o(e)})^{{\lambda (e)}}t_e\mi= (x_{o(\ov e)})^{{\lambda (\ov
471: e)}}$.
472:
473:
474: Replacing the chosen generator of a vertex group $G_v$ by its
475: inverse changes the sign of all labels near $v$. Replacing an edge
476: group generator changes the sign of both labels carried by the
477: edge. These changes are {\it admissible sign changes\/}.
478: Labelled graphs will always be considered up to admissible sign
479: changes.
480:
481: When we focus on a particular edge, we always use admissible sign changes
482: to make it a $(p,q)$-segment with $p,q>0$, or a $(p,q)$-loop
483: with $1\le p\le|q|$.
484:
485: A
486: $(1,q)$-loop is an {\it ascending loop\/}. It is a {\it strict ascending
487: loop\/} if $|q|>1$; note that $G$ then contains a solvable
488: Baumslag-Solitar group
489: $BS( 1, q)$. A
490: $(p,q)$-loop with $p|q$ is a {\it pseudo-ascending loop\/}.
491:
492:
493: \subhead GBS trees \endsubhead
494:
495:
496: Let $G$ be the fundamental group of a labelled graph $\Gamma $. The
497: associated Bass-Serre tree is a locally finite $G$-tree $T$ with all
498: edge and vertex stabilizers infinite cyclic. Such $G$-trees will be
499: called {\it GBS trees\/}. Two trees are considered to be the same if
500: there is a $G$-equivariant isomorphism between them.
501:
502:
503: We always assume that the action is {\it minimal\/}: there is no proper
504: invariant subtree. In terms of $\Gamma $, this is equivalent to saying
505: that the label near every terminal vertex is bigger than $1$. We also
506: assume that actions are without inversions.
507:
508: Given a GBS tree $T$, one obtains a labelled graph $\Gamma
509: =T/G$, with the labelling well-defined up to admissible sign
510: changes (see Remark 2.3 in [\For]). This graph of groups is {\it marked\/}:
511: there is an isomorphism from its fundamental group to $G$, well-defined up to
512: composition with an inner automorphism. The valence of a vertex $v\in T$
513: is the sum of the absolute values of the labels near its image in $\Gamma $.
514:
515: GBS trees $T$ and marked labelled graphs $\Gamma $ are thus
516: equivalent concepts. We will work with both. We usually use the same letter
517: $v$ (resp\.
518: $e$) for a vertex (resp\.
519: edge) of $T$ and its image in $\Gamma $. When we need to distinguish, we
520: write $\ov v$ for the image of $v$ in $\Gamma $.
521: We denote vertex stabilizers
522: (vertex groups) by $G_v$, edge stabilizers (edge groups) by $G_e$.
523:
524:
525:
526: \subhead Collapses and algebraic
527: rigidity\endsubhead
528:
529:
530: Collapsing an edge $e$ of $\Gamma $ (or equivalently a $G$-orbit of edges of
531: $T$) yields a new tree $S$, which usually is not a GBS tree. It is a GBS tree if
532: and only if $e$ is a segment and at least one of the labels $\lambda _e$,
533: $\lambda _{\ov e}$ equals $1$. Such an edge will
534: be called a {\it collapsible\/} edge.
535:
536: In the proof of Theorem \six, we will
537: collapse
538: $(2,2)$-edges and $(1,q)$-loops; these are not collapsible edges. We usually
539: denote by $\Theta $ the collapsed graph of groups, by $\pi :T\to S$ the
540: collapse map. The image of a vertex $v\in T$ is denoted by $\pi
541: (v)$, or sometimes just
542: $v$. The stabilizer of $\pi (v)$ in $S$ contains the cyclic group $G_v$, we call
543: it
544: $H_v$ (it will often be a solvable Baumslag-Solitar group). We use the same
545: letter for a non-collapsed edge of
546: $T$ and its image in $S$. It has the same stabilizer in both trees.
547:
548:
549:
550:
551: Collapsing a collapsible edge is called an {\it elementary collapse\/}. The
552: reverse move is an {\it elementary expansion\/}.
553: Labels near
554: $o( e)$ get multiplied by
555: $\lambda _{\ov e} $ when we collapse an edge $e$ with $\lambda _e=1$ (see
556: Figure
557: \fig).
558:
559: \midinsert
560: \centerline
561: {\includegraphics[scale=.5]
562: {fig5.eps}}
563: \captionwidth{220pt}
564: \botcaption
565: {Figure \the\figno}
566: {Elementary collapse. }
567: \endcaption
568: \endinsert
569:
570:
571: The graph $\Gamma $, or the tree $T$, is {\it reduced\/} (in the sense
572: of [\Forr]) if there is no collapsible edge.
573: In terms of trees, $T$ is reduced if and only if
574: any edge $e=vw$ satisfying $G_e=G_v$ has its endpoints in the same
575: $G$-orbit. Any tree may be reduced by applying a finite
576: sequence of elementary collapses (the reduction is not always unique).
577:
578: \midinsert
579: \centerline
580: {\includegraphics[scale=.5]
581: {fig6.eps}}
582: \botcaption
583: {Figure \the\fihno}
584: {A labelled graph representing an algebraically rigid group.}
585: \endcaption
586: \endinsert
587:
588:
589: A reduced GBS tree $T$ is {\it rigid} if it is the only reduced GBS
590: tree (up to equivariant isomorphism). Building on work from [\Forr,
591: \GHMR, \Gu,
592: \Pet], it is shown in [\Ler] that, if $G$ is not solvable, $T$ is rigid if and only if
593: $\Gamma
594: $ satisfies the following condition (see Figure
595: \fig): if $e,f$
596: are distinct oriented edges of $\Gamma $ with the same origin
597: $v$, and the label of $f$ near $v$ divides that of $e$, then either $e=\ov f$ is a
598: $(p,\pm p)$-loop with $p\ge2$, or $v$ has valence 3 and bounds a
599: $(1,\pm1)$-loop.
600:
601:
602:
603: In particular,
604: $T$ is rigid whenever there is no divisibility relation in $\Gamma $ (recall
605: that a {\it divisibility relation\/} is a relation $p|q$ between two labels at the
606: same vertex).
607:
608: When there is a rigid GBS tree, we say that $G$ is {\it algebraically rigid\/}. In
609: this case, there is only one reduced marked labelled graph
610: representing
611: $G$. See [\MSW, \Wh] for quasi-isometric rigidity of GBS groups.
612:
613: \subhead Non-elementary groups \endsubhead
614:
615: We say that $G$ is {\it elementary\/} if
616: $T$ may be chosen to be a point or a line. As vertices of $T$ then have
617: valence at most 2, there are only four possibilities for $\Gamma $: a
618: point, a
619: $(1,1)$-loop, a
620: $(1,-1)$-loop, a $(2,2)$-segment. The corresponding groups are
621: $\Z$, $\Z^2$, and the Klein bottle group $\langle x,t\mid
622: txt\mi=x\mi\rangle=\langle a,b\mid a^2=b^2\rangle$, with $\Out(G)$ equal to
623: $\Z/2\Z$, $GL(2,\Z)$, and $\Z/2\Z\times\Z/2\Z$ respectively.
624:
625: Though non-elementary, the solvable groups $BS(1,n)$ are special. For $|n|>1$,
626: the group $\Out(BS(1,n))$ is virtually $\Z^{r-1}$, where $r$ is the number of
627: prime divisors of $n$ [\Coll]. More generally, see [\Coll, \Collev, \GHMR] for a
628: presentation of $\Out(BS(m,n))$.
629:
630:
631: From now on, we consider only non-elementary groups. Here are a few
632: simple properties (compare [\Foc]).
633:
634: From the action of $G$ on $T$, it is easy to see that a non-elementary
635: GBS group either is a solvable Baumslag-Solitar group $BS(1,n)$ (if it
636: fixes an end of
637: $T$), or contains non-abelian free groups (if the action on $T$ is
638: irreducible). In particular, $G$ always has exponential growth. A finitely
639: generated subgroup of
640: $G$ is free (if it acts freely on $T$) or is a GBS group. A non-elementary
641: GBS group is one-ended, coherent, and has cohomological dimension 2 [\Foc,
642: \Kr].
643:
644: Any GBS group maps onto $\Z$
645: (the presentation of $G$ given earlier has more generators than
646: relators). GBS groups are therefore locally indicable, hence orderable (see
647: [\RR]). Using the fact that $\langle a,b\mid a^p=b^q
648: \rangle$ is bi-orderable only if $|p|$ or $|q|$ equals $1$, one shows that the
649: only bi-orderable GBS groups are
650: $F_n\times\Z$ and $BS(1,n)$ for $n\ge1$.
651:
652:
653: \subhead Elliptic elements \endsubhead
654:
655: Two subgroups $H,K$ of $G$ are {\it commensurable\/} if $H\cap K$ has
656: finite index in both $H$ and $K$. Two elements $g,h$ are commensurable
657: if $\langle g
658: \rangle$ and $\langle h
659: \rangle$ are commensurable, equivalently if there is a relation
660: $g^p=h^q$ with
661: $p,q$ non-zero integers. The {\it commensurator\/} of $g$ is the
662: subgroup
663: $Comm(g)$ consisting of all $x\in G$ such that $xgx\mi$ is
664: commensurable to $g$.
665:
666: Given any $G$-tree $T$, an element $g\in G$, or a subgroup $H$, is {\it
667: elliptic\/} if it fixes a point. If
668: $g$ is not elliptic, it is {\it hyperbolic\/}: there is an invariant axis, on
669: which $g$ acts as a translation by some positive integer $\ell(g) $.
670: Conjugate or commensurable elements have the same type (elliptic or
671: hyperbolic). A relation
672: $ga^pg\mi=a^q$ with
673: $|p|\ne|q|$ implies that $a$ is elliptic, because its translation length
674: satisfies
675: $|p|\ell(a)=|q|\ell(a)$.
676:
677:
678: \thm{Lemma \sta{} [\Forr]} Let $T$ be a GBS tree, with $G$
679: non-elementary. Any two non-trivial elliptic elements
680: $g,g'$ are commensurable. An element $g\in G$ is elliptic if and only if
681: its commensurator equals $G$.
682: \fthm
683:
684: \demo{Proof} If $g,g'$ fix vertices $v,v'$, one shows that they are
685: commensurable by induction on the distance between $v$ and $v'$. If
686: $g$ is hyperbolic, its axis is $Comm(g)$-invariant, so $Comm(g)\neq G$
687: because $G$ is not elementary.
688: \cqfd\enddemo
689:
690: \nom\ellind
691: \thm{Corollary \sta} The set of elliptic elements depends only on $G$,
692: not on the GBS tree $T$. It is invariant under automorphisms of $G$.
693: \cqfd
694: \fthm
695:
696:
697:
698: As any two GBS trees have the same elliptic subgroups, Forester's
699: deformation theorem [\Forr] yields:
700:
701: \nom\defor
702: \thm{ Corollary \sta {}} Let $G$ be a non-elementary GBS group.
703: Any two GBS
704: trees are related (among GBS trees) by a finite sequence of elementary
705: expansions and collapses. \cqfd
706: \fthm
707:
708: The quotient $G/G_{ell}$ of $G$ by the subgroup generated by all
709: elliptic elements may be identified with the (topological) fundamental
710: group
711: $\pi ^{top}_1(\Gamma )$ of the graph
712: $\Gamma
713: $ (this is a general property of graphs of groups). All labelled graphs
714: representing $G$ thus have the same {\it first Betti number\/}, denoted
715: by $b$.
716:
717: All homomorphisms from $G$ to a free group with
718: non-abelian image factor through the quotient map $\theta :G\to
719: G/G_{ell}\simeq F_b$ (because in a free group the
720: commensurator of any non-trivial element is cyclic). Since $G$ always
721: maps onto $\Z$, the maximum rank of a free
722: quotient of $G$ is $\max(b,1)$.
723:
724:
725: \subhead The modular homomorphism $\Delta $ \endsubhead
726:
727: Let $G$ be a non-elementary GBS group. The set $\E$ consisting of all
728: non-trivial elliptic elements is stable under conjugation, elements of
729: $\E$ have infinite order, and any two elements of $\E$ are
730: commensurable. These properties yield a homomorphism
731: $\Delta $ from
732: $G$ to the multiplicative group of non-zero rationals $\Q^*$, defined as
733: follows.
734:
735: Given $g\in G$, choose any $a\in\E$. There is a relation
736: $ga^pg\mi=a^q$, with
737: $p,q$ non-zero, and we define $\Delta (g)=\frac p q$. As pointed out in [\Krn],
738: it is easily checked that this is independent of the choices made ($a$ and the
739: relation), and defines a homomorphism. We
740: call $\Delta (g)$ the {\it modulus\/} of $g$. Note that $\Delta
741: \circ\alpha =\Delta $ if $\alpha $ is any automorphism of $G$, because
742: $\E$ is $\alpha $-invariant.
743:
744: Let $H$ be a finite index subgroup of $G$. Any GBS $G$-tree is also a GBS
745: $H$-tree, so $H$ is a GBS group. The modular homomorphism of $H$ is
746: the restriction of that of $G$.
747:
748: Every elliptic element has modulus $1$, so $\Delta $
749: factors through the free group $G/G_{ell}\simeq F_b$. In particular, $\Delta $
750: is trivial when $\Gamma $ is a tree. If
751: $\Gamma
752: $ is a labelled graph representing $G$, one has $G/G_{ell}\simeq\pi
753: ^{top}_1(\Gamma )$ and the modulus may be computed as follows
754: (see [\BK, \For]): if
755: $\gamma \in \pi ^{top}_1(\Gamma )$ is represented by an edge-loop
756: $(e_1,\dots,e_m)$, its modulus is simply
757: $\ds\prod_{j=1}^m \frac{\lambda _{e_j}}{\lambda _{\ov e_j}}$.
758:
759: We denote by $M$ the image of $\Delta $. It is a
760: subgroup of $(\Q^*, \times)$. If $G=BS(m,n)$, then $M$ is generated by $\frac
761: mn$. If $G$ is represented by the labelled graph of Figure 6, then $M$ is
762: generated by $-1$ and $-\frac32$.
763:
764:
765: \nom\modu
766: \thm{Lemma \sta} Let $r=\frac pq$ be a nonzero rational number,
767: written in lowest terms. Assume $r\neq\pm1$.
768: \roster
769: \item $r\in M$ if and only if the equation $xy^px\mi=y^q$ has a solution
770: with
771: $y\neq1$.
772: \item If $r\in\Z$, then $r\in M$ if and only if $G$ contains a subgroup
773: isomorphic to
774: $BS(1,r)$.
775: \endroster
776: \fthm
777:
778:
779:
780:
781: \demo{Proof} If $\frac pq \in M$, the equation
782: $xy^{np}x\mi=y^{nq}$ has a non-trivial solution for some $n\in\Z$, so
783: $xy^px\mi=y^q$ has a non-trivial solution. Conversely, if
784: $|p|\neq |q|$ and
785: $xy^px\mi=y^q$ has a non-trivial solution, then $y$ must be elliptic
786: and therefore
787: $\frac pq=\Delta (x)\in M$. We have proved $(1)$.
788:
789: If $xy^rx\mi=y $ with $y\neq1$ and $r$ an integer different from $-1,0,1$,
790: then
791: $H=\langle x,y\rangle$ is a solvable GBS group, and $\Delta (x)=r$, so
792: $H\simeq BS(1,r)$ (one may also show $H\simeq BS(1,r)$ by arguing that
793: the only torsion-free proper quotient of $BS(1,r)$ is $\Z$).
794: \cqfd\enddemo
795:
796: \example {Remarks} $\bullet$ The values $ \pm1$ are special. If $G$ is
797: represented by a labelled tree
798: $\Gamma $ containing a $(2,2)$-edge, then $xy x\mi=y\mi$ has a
799: nontrivial solution because $G$ contains a Klein bottle group, but $-1\notin
800: M$ because $\Gamma $ is a tree. Conversely,
801: $1$ always belong to $M$, but
802: $BS(1,n)$ does not contain $\Z^2=BS(1,1)$ for $|n|>1$.
803:
804: $\bullet$ It is probably not true that $G$
805: always contains
806: $BS(p,q)$ if
807: $\frac pq\neq\pm1$ is a modulus.
808:
809:
810: $\bullet$ Using $\Delta $, it is easy to show that the isomorphism type of
811: $BS(m,n)$ determines $m$ and $n$ (normalized by $1\le m\le|n|$) [\Mold]. In
812: most cases,
813: $m$ and
814: $n$ are determined by $m/n$ (given by $\Delta $) and $|m-n|$ (given by
815: abelianizing). To obtain $m$ from $BS(m,m)$, observe that the quotient of
816: $BS(m,m)$ by its center is $\Z*\Z/m\Z$.
817: \endexample
818:
819: We say that $G$ has {\it trivial modulus\/} if $M=\{1\}$ (we often write this as
820: $\Delta =1$). It is {\it unimodular\/} if
821: $M\inc\{1,-1\}$, equivalently if $xy^px\mi=y^q$
822: with $y\neq 1$ implies $p=\pm q$. As in [\For], we say that $G$ has {\it no
823: nontrivial integral modulus\/} if
824: $M\cap\Z\inc\{1,-1\}$. This is equivalent to saying that $G$ contains no
825: solvable Baumslag-Solitar group $BS(1,n)$ with $n\ge2$ (we may take $n>0$
826: because $BS(1,-n)$ contains $BS(1,n^2)$).
827:
828: \subhead Unimodular groups \endsubhead
829:
830: \nom\cent
831: \thm{Proposition \sta} Let $G$ be a non-elementary GBS group. The
832: center $Z(G)$ of
833: $G$ is infinite cyclic if $G$ has trivial modulus, trivial otherwise. It acts as the
834: identity on any GBS tree.
835: \fthm
836:
837: \demo{Proof} Let $T$ be any GBS tree (recall that $T$ is always assumed to
838: be minimal). If
839: $a$ is central (more generally, if $\langle a \rangle$ is normal), it is elliptic,
840: as otherwise its axis would be a
841: $G$-invariant line. The fixed point set of $a$ is a $G$-invariant subtree,
842: so equals $T$ by minimality. This shows that
843: $Z(G)$ is contained in the kernel of the action (elements acting on $T$
844: as the identity). In particular, it is trivial or cyclic.
845:
846: If $\Delta (g)\neq1$ and $a\in\E$, there is a relation $ga^pg\mi=a^q$
847: with $p\neq q$, so $a$ cannot be central. This shows that $Z(G)$ is trivial if $G$
848: does not have trivial modulus. If $\Delta $ is trivial, choose any finite
849: generating system
850: $s_i$ for
851: $G$, and $a\in\E$. For each $i$, there is a relation
852: $s_ia^{n_i}s_i\mi=a^{n_i}$ with $n_i\neq0$. It follows that some power
853: of $a $ is central.
854: \cqfd\enddemo
855:
856:
857:
858: \thm{Proposition \sta} Let $G$ be a non-elementary GBS group. The
859: following are equivalent:
860: \roster
861: \item
862: $G$ is unimodular.
863: \item
864: $G$ contains a normal infinite cyclic subgroup $Z$.
865: \item
866: $G$ has a finite index subgroup isomorphic to $F_n\times \Z$ for some
867: $n>1$.
868: \endroster The quotient of $G$ by any normal infinite cyclic subgroup
869: $Z$ is virtually free.
870: \fthm
871:
872: \demo{Proof} Suppose $G$ is unimodular. The kernel of $\Delta $ has
873: index 1 or 2, and has trivial modulus. Its center is
874: infinite cyclic and characteristic in
875: $G$, so (1) implies (2).
876:
877: Suppose $Z$ is infinite cyclic and normal. Let $T$ be any GBS
878: tree. As in the proof of Proposition \cent, one shows that $Z$ is
879: contained in the kernel of the action. The quotient
880: $G/Z$ acts on
881: $T$ with finite stabilizers, so is virtually free. This easily implies that
882: $G$ is virtually $F_n\times\Z$.
883:
884: If $G$ is virtually $F_n\times\Z$, its modulus is trivial on a finite index
885: subgroup, so $M$ is finite, hence contained in $ \{1,-1\}$.
886: \cqfd\enddemo
887:
888: \example{Remarks} $\bullet$ Let $T$ be a GBS tree. If $G$ is
889: unimodular, we have seen that $ Z(\ker(\Delta ))$ is contained in the
890: kernel of the action on $T$. Conversely, if the action has a nontrivial kernel
891: $K$, then $G$ is unimodular (because $K$ is normal and cyclic), and furthermore
892: $K=Z(\ker(\Delta ))$. To see this, simply note that, if
893: $a$ generates $K$, one has $gag\mi=a^{\pm1}$ for any $g\in G$,
894: so $a$ commutes with $\ker(\Delta )$.
895:
896:
897: $\bullet$ Two non-solvable GBS groups are quasi-isometric if and only
898: if they are both
899: unimodular or both non-unimodular [\Wh]. Any torsion-free group
900: quasi-isometric to
901: $F_n\times
902: \Z$ with $n>1$ is a unimodular GBS group [\MSW].
903:
904: $\bullet$ A GBS group is residually finite if and only if it is solvable or
905: unimodular [\Lrang].
906: \endexample
907:
908:
909:
910: \subhead Twisted conjugacy classes \endsubhead
911:
912: Let $\alpha :G\to G$ be an endomorphism. Two elements $g,g'\in G$ are
913: {\it $\alpha
914: $-conjugate\/} if there exists $h$ such that $g'=hg\alpha (h)\mi$. The
915: number of
916: $\alpha $-conjugacy classes is the {\it Reidemeister number\/} of
917: $\alpha $, denoted by
918: $R(\alpha )$. It is relevant for fixed point theory (see [\Fed]).
919:
920: \thm{Proposition \sta} Let $\alpha :G\to G$ be an endomorphism of a
921: non-elementary GBS group. If one of the following conditions holds,
922: then
923: $R(\alpha )$ is infinite:
924: \roster \item $\alpha $ is surjective.
925: \item $\alpha $ is injective and $G$ is not unimodular.
926: \item $G=BS(m,n)$ with $|m|\neq|n|$, and the image of $\alpha $ is not
927: cyclic.
928: \endroster
929: \fthm
930:
931: This generalizes results of [\Fed] about Baumslag-Solitar groups.
932:
933: \demo{Proof} First suppose that $G$ is unimodular and $\alpha $ is
934: surjective. The group $G$ is residually finite (because it is virtually
935: $F_n\times\Z$), hence Hopfian. We therefore assume that $\alpha $ is an
936: automorphism. The subgroup
937: $Z=Z(\ker(\Delta ))$ is characteristic, so $\alpha $ induces an
938: automorphism
939: $\beta $ on the virtually free group $G/Z$. As $G/Z$ is a
940: non-elementary (word) hyperbolic group,
941: $R(\beta )$ is infinite [\LL, \Fe]. This implies that $R(\alpha )$ is infinite.
942:
943:
944: From now on, we assume that $G$ is not unimodular. If $\alpha $ is an
945: automorphism, we know that $\Delta \circ\alpha =\Delta $, so $\alpha
946: $-conjugate elements of $G$ have the same modulus. As $M$ is infinite, we
947: get $R(\alpha )$ infinite. This argument works in the general case, but we
948: have to prove $\Delta \circ\alpha =\Delta $ for endomorphisms satisfying (1),
949: (2), or (3).
950:
951:
952: We first claim that
953: $\alpha $ does not factor through $\theta :G\to G/G_{ell}\simeq F_b$. This
954: is clear if (2) or (3) holds. If a surjective $\alpha $ factors as $\rho
955: \circ\theta $, then $\theta
956: \circ\rho $ is a non-injective epimorphism from $F_b$ to itself, a
957: contradiction because free groups are Hopfian.
958:
959: We can now show $\Delta \circ\alpha =\Delta $. Since $\alpha $ does
960: not factor through
961: $\theta $, there is an elliptic
962: $a$ with
963: $\alpha (a)\neq 1$. As $G$ is not unimodular, there is a relation
964: $g_0a^mg_0\mi=a^n$ with $|m|\neq |n|$. From $\alpha (g_0)\alpha
965: (a)^m\alpha (g_0)\mi=\alpha (a)^n$, we deduce that $\alpha (a)$ is
966: elliptic. Thus $\alpha (a)$ is a nontrivial elliptic element, and may be
967: used to compute $\Delta $. Given any $g$, we have a relation $g a^pg
968: \mi=a^q$. We then write $\alpha (g )\alpha (a)^p\alpha (g )\mi=\alpha
969: (a)^q$, showing that $g$ and
970: $\alpha (g)$ have the same modulus $\frac pq$.
971: \cqfd\enddemo
972:
973:
974:
975:
976: \head \sect The automorphism group of a GBS tree \endhead
977:
978: \subhead General facts\endsubhead
979:
980: Let $G$ be any finitely generated group. As above, we consider
981: $G$-trees up to equivariant isomorphism. There is a natural action of
982: $\Out(G)$ on the set of $G$-trees, given by precomposing an action of
983: $G$ on $T$ with an automorphism of
984: $G$ (composing with an inner automorphism does not change the tree).
985:
986: Given $T$, we denote by $\Out^T(G)\inc\Out(G)$ its stabilizer: $\Phi $ is
987: in
988: $\Out^T(G)$ if and only if $T$, with the action of $G$ twisted by $\Phi $,
989: is equivariantly isomorphic to $T$ with the original action. When $T$ is
990: irreducible, this is equivalent to saying that the length function $\ell$
991: satisfies $\ell\circ\Phi =\ell$.
992:
993: We recall results of [\LGD] about $\Out^T(G)$. We assume that $T$ is
994: minimal and is not a line (but there is no condition
995: on edge and vertex stabilizers in this subsection). The quotient graph of
996: groups is denoted by
997: $\Gamma
998: $, its vertex set by
999: $V$.
1000:
1001: There is a natural homomorphism from $\Out^T(G)$ to the
1002: symmetry group of $\Gamma $ (viewed as a graph
1003: with no additional structure). The kernel is a finite index subgroup
1004: $\Out_0^T(G)$, and there is a homomorphism $\rho
1005: :\Out_0^T(G)\to\prod_{v\in V}\Out(G_v)$. All automorphisms of $G_v$
1006: which occur in the image of
1007: $\rho
1008: $ preserve the set of conjugacy classes of incident edge groups.
1009:
1010: The kernel of
1011: $\rho $ is generated by the group of twists $\T(T)$ together with
1012: automorphisms called bitwists (bitwists belong to $\T(T)$ when vertex
1013: groups are abelian). The group $\T(T)$, which we also denote by $\T(\Gamma
1014: )$, will play an important role in the sequel. Before defining
1015: it, we mention that, when edge groups are cyclic, there is a further
1016: finite index subgroup
1017: $\Out_1^T(G)\inc\Out_0^T(G)$ with $\Out_1^T(G)\cap\ker \rho =\T(T)$. It will
1018: be used in Section 7.
1019:
1020: To define $\T(T)$, we first consider an oriented edge $e$ of $ \Gamma $,
1021: with origin
1022: $o(e)=v$. Let
1023: $G_e$,
1024: $G_v$ be the corresponding edge and vertex groups, with $G_e$
1025: identified to its image in $ G_v$. We denote by $Z_{G_v}(G_e)$ the
1026: {\it centralizer\/} of $G_e$ in $G_v$.
1027:
1028: Given $z\in Z_{G_v}(G_e)$,
1029: we define the
1030: {\it twist\/}
1031: $D(z)\in\Out(G)$ by $z$ around $e$ as follows (see [\LGD] for details). If $e$ is
1032: separating, it expresses $G$ as an amalgam
1033: $G=G_1*_{G_e}G_2$. Then $D(z)$ is defined as the identity on $G_1$,
1034: and conjugation by $z$ on
1035: $G_2$. If $e$
1036: does not separate, $G$ is an HNN-extension and $D(z)$ maps the
1037: stable letter $t$ to $zt$ (keeping the base group fixed).
1038:
1039: The {\it group of twists $\T(T)$\/}, or $\T(\Gamma )$, is the
1040: subgroup of
1041: $\Out(G )$ generated by all twists. As twists around distinct edges commute,
1042: $\T(\Gamma )$ is a quotient of
1043: $\prod Z_{G_{o(e)}}(G_e)$, the product being taken over all oriented
1044: edges of $\Gamma $.
1045: Proposition 3.1 of [\LGD] says that only two types of relations are
1046: needed to obtain a presentation of $\T(\Gamma )$.
1047:
1048: For each pair of
1049: opposite edges $(e,\ov e)$, there are {\it edge relations\/} associated to
1050: elements $z$ in the center $Z(G_e)$ (twisting by $z$ near
1051: the origin of $e$ defines the
1052: same outer automorphism as twisting by $z\mi$ near the origin of
1053: $\ov e$). For each vertex $v$, there are {\it vertex relations\/} associated to
1054: elements
1055: $z\in Z(G_v)$ (twisting by $z$ simultaneously around all edges with origin
1056: $v$ defines an inner automorphism).
1057:
1058: \nom\te
1059: \example {Remark \sta} Let $e\inc\Gamma $ be a segment such that both
1060: adjacent vertex groups are abelian. Using the vertex relations, one sees that
1061: $\T(\Gamma )$ is generated by the groups $Z_{G_{o(f)}}(G_f)$ with $f\neq
1062: e,\bar e$. Collapsing $e$ yields a new graph of groups whose group of twists
1063: contains $\T(\Gamma )$.
1064: \endexample
1065:
1066:
1067: Our main tool for finding free groups $F_2$ in $\Out(G)$ will be:
1068:
1069: \nom\tw
1070: \thm{Lemma \sta} Let $\Gamma $ be a minimal graph of groups,
1071: with fundamental group $G$. Let $e$ be an edge
1072: with origin
1073: $v$, and let $G_e$, $G_v$ be the corresponding groups. The subgroup
1074: $\T(\Gamma )\inc \Out(G)$ maps onto
1075: $Z_{G_v}(G_e)/\langle Z( G_v),Z(G_e)\rangle$.
1076: \fthm
1077:
1078: We denote by $\langle Z(
1079: G_v),Z(G_e)\rangle$ the (obviously normal) subgroup
1080: generated by the centers of $G_v$ and $G_e$.
1081:
1082: \demo{Proof}
1083: Divide $\T(\Gamma )$ by (the image of) all factors $Z_{G_{o(f)}}(G_f)$ for
1084: $f\neq e$ (including $f=\ov e$). The only relations which remain are those
1085: involving
1086: $Z_{G_v}(G_e)$, namely edge relations associated to $(e,\ov e)$ and
1087: vertex relations associated to $v$. The quotient is precisely
1088: $Z_{G_v}(G_e)/\langle Z( G_v),Z(G_e)\rangle$.
1089: \cqfd\enddemo
1090:
1091:
1092:
1093:
1094:
1095:
1096: \subhead The group of twists of a GBS tree\endsubhead
1097:
1098: Now let $G$ be a non-elementary GBS group. We consider the action of
1099: $\Out(G)$ on the set of GBS trees. The corresponding action on the
1100: set of marked graphs is by changing the marking. If $T$ is
1101: rigid, then
1102: $\Out^T(G)=\Out(G)$.
1103:
1104: The group of twists $\T(T)$ is a finitely generated abelian group.
1105: The presentation recalled above may be rephrased as follows (we
1106: use additive notation).
1107:
1108: Given an oriented edge $e$ of $ \Gamma $, there is one generator
1109: $D_e$. If $e$ is separating, $D_e$ is the identity on $G_1$, and
1110: conjugation by $x_v$ on
1111: $G_2$ (with $x_v$ the generator of the vertex group
1112: at $v=o(e)$, and $G=G_1*_{G_e}G_2$ as above). If $e$
1113: does not separate, choose a maximal tree $\Gamma _0$ not containing
1114: $e$. In the corresponding presentation of $G$, define $D_e$ as mapping
1115: $t_e$ to $x_{o(e)}t_e$ and keeping all other generators fixed.
1116:
1117: In terms of these generators $D_e$, the relations are the following.
1118: For each pair of opposite
1119: edges $(e,\ov e)$, there is an edge relation
1120: $\lambda _e D_e+\lambda _{\ov e} D_{\ov e}=0$, implied by the relation
1121: $(x_{o(e)})^{{\lambda (e)}}= (x_{o(\ov e)})^{{\lambda (\ov e)}}$ or
1122: $t_e(x_{o(e)})^{{\lambda (e)}}t_e\mi= (x_{o(\ov e)})^{{\lambda (\ov e)}}$. For
1123: each vertex $v$, there is a vertex relation $\sum_{e\in E_v}D_e=0$, with
1124: $E_v$ the set of edges with origin $v$.
1125:
1126: \example{Remark} The group $G$ is in a natural way the fundamental group of
1127: a 2-complex consisting of annuli (corresponding to edges of $\Gamma $) glued
1128: to circles (corresponding to vertices). One can consider the subgroup
1129: $DT(\Gamma )$ of
1130: $\T(\Gamma )$ generated by Dehn twists supported in the annuli. It is easy
1131: to see that it has finite index. One may also show
1132: $\T(\Gamma )=DT(\Gamma ')$, where $\Gamma '$ is a (non reduced)
1133: graph obtained from
1134: $\Gamma
1135: $ by elementary expansions.
1136: \endexample
1137:
1138:
1139: Recall that
1140: $b $ is the first Betti number of any labelled graph $\Gamma $
1141: representing
1142: $G$.
1143:
1144: \nom \kk
1145: \thm{Proposition \sta} Let $G$ be a non-elementary GBS group. Define $k$ as
1146: $b$ if $G$ has trivial modulus,
1147: $b-1$ if not.
1148: \roster \item The torsion-free rank of the abelianization $G_{ab}$ is
1149: $k+1$.
1150: \item Let $\Gamma $ be any labelled graph representing $G$. The
1151: torsion-free rank of the abelian group $\T(\Gamma )$ is
1152: $k$.
1153:
1154: \item If $\Gamma _0\inc\Gamma $ is a
1155: maximal subtree, the twists $D_e$ around the
1156: edges of $\Gamma \setminus\Gamma _0$ generate a finite index subgroup of
1157: $\T(\Gamma )$.
1158: \endroster
1159: \fthm
1160:
1161: \demo{Proof} Killing all elliptic elements produces an epimorphism $\theta
1162: :G\to F_b$ (see Section 2), so
1163: $rk(G_{ab})\ge b$. If $\Delta $ is nontrivial, any elliptic element $a$ satisfies a
1164: relation $ga^pg\mi=a^q$ with
1165: $p\neq q$, so is mapped trivially to
1166: torsion-free abelian groups. This shows $rk(G_{ab})= b$ in this case.
1167:
1168: If $\Delta $ is trivial, fix $\Gamma $ and $\Gamma _0$. It
1169: follows from the presentation of $G$ given earlier that $G_{ab}$ is the direct
1170: sum of
1171: $\Z^b$ with the abelian group
1172: $G'$ defined by the following presentation: there is one generator $x_v$
1173: for each vertex of $\Gamma $, and one relation $\lambda _e x_{o(e)}=\lambda
1174: _{\ov e} x_{o(\ov e)}$ for each pair of opposite edges
1175: $(e,\ov e)$. We show that $G'$ maps non-trivially to $\Z$. It is easy to map the
1176: generators
1177: $x_v$ to
1178: $\Z$ in such a way that relations associated to edges in $\Gamma _0$ are
1179: satisfied. Using the formula
1180: $ \Delta (\gamma )=\prod_{j=1}^m \frac{\lambda _{e_j}}{\lambda _{\ov
1181: e_j}}$ (see Section 2), one sees that the remaining relations are
1182: automatically satisfied. We get $rk(G_{ab})= b+1$.
1183:
1184: Assertion (2) follows immediately from Proposition \cent{} and the exact
1185: sequence
1186: $$0\to Z(G)\to \Z^{\nu +\zeta }\to\Z^{2\zeta }\to\T\to0$$
1187: given by Proposition 3.1 of [\Ler], where $\nu $ (resp\. $\zeta $) is the number
1188: of vertices (resp\. edges) of $\Gamma $ (we are grateful to M\. Clay for
1189: suggesting this short argument).
1190:
1191:
1192: Assertion (3) follows from the presentation of $\T(\Gamma )$ in terms of the
1193: generators $D_e$. If we add the relations $D_e=D_{\ov e}=0$ for
1194: $e\notin\Gamma _0$, the quotient is the group of twists associated to the
1195: labelled graph $\Gamma _0$, so is finite by Assertion (2).
1196: \cqfd\enddemo
1197:
1198: \nom\foo
1199: \example{Remark \sta}
1200: One may decide whether a given
1201: $D_e$ has finite or infinite order. If $e$ does not separate, the order of
1202: $D_e$ is finite if and only if $G$ has nontrivial modulus, but every loop
1203: not containing $e$ has trivial modulus. If $e$ separates, the order
1204: is infinite if and only if each component of $\Gamma \setminus\{e\}$
1205: contains a loop with nontrivial modulus.
1206: \endexample
1207:
1208: The groups
1209: $\T(T)$ associated to different GBS trees are abstractly commensurable by
1210: Proposition
1211: \kk. We show that they are commensurable as subgroups of $\Out(G)$.
1212:
1213:
1214: \nom\commens
1215: \thm{Proposition \sta} If $T,T'$ are two GBS trees, then $\T(T)$ and
1216: $\T(T')$ are commensurable subgroups of $\Out(G)$.
1217: \fthm
1218:
1219:
1220:
1221: \demo{Proof} By Corollary \defor, it suffices to show that $\T(T')$ is
1222: commensurable with $\T(T)$ if
1223: $T'$ is obtained from
1224: $T$ by an elementary collapse. Consider the corresponding graphs
1225: $\Gamma ,\Gamma '$. Let $e=vw\inc \Gamma $ be the collapsed edge.
1226: We assume $\lambda _e=1$, and we denote
1227: $\lambda_{\ov e}$ by $\lambda $, so $G_v$ has index $\lambda $ in
1228: $G_w$.
1229:
1230: Let $F$ be the set of oriented edges of $\Gamma $ other than $e,\ov e$.
1231: The group
1232: $\T(T)$ is the subgroup of $\Out(G)$ generated by the twists $D_f$,
1233: $f\in F$ (see Remark \te). Similarly, $\T(T')=\langle
1234: D'_f\mid f\in F\rangle$, as $F$ may be viewed as the set of
1235: oriented edges of
1236: $\Gamma '$. Moreover, we have $D_f=D'_f$ if the origin of $f$ is not $v$,
1237: and
1238: $D_f=\lambda D'_f$ if it is because the collapse replaces the vertex
1239: group $G_v$ of
1240: $\Gamma $ by the larger group $G_w$. This shows that $\T(T')$ contains
1241: $\T(T)$ as a subgroup of finite index.
1242: \cqfd\enddemo
1243:
1244: \nom\divi
1245: \example{Remark \sta} The index of $\T(T)$ in $\T(T')$
1246: divides a power of the label $\lambda $. This will be used in the proof of
1247: Theorem 3.12.
1248: \endexample
1249:
1250:
1251: \subhead Applications \endsubhead
1252:
1253: We apply the preceding results to the study of $\Out^T(G)$, using the
1254: following fact:
1255:
1256: \thm{Proposition \sta{} [\LGD]} $\T(T)$ has finite index in $\Out^T(G)$.
1257: \fthm
1258:
1259: This follows from Theorem 1.6 of [\LGD], as edge and vertex groups
1260: have finite outer automorphism groups. More precisely, let us show:
1261:
1262: \nom\unib
1263: \thm{Proposition \sta} Given $G$, the index of $\T(T)$ in $\Out^T(G)$ is
1264: uniformly bounded (independently of
1265: $T$).
1266: \fthm
1267:
1268: \demo{Proof}
1269: Consider the chain of
1270: subgroups
1271: $\T(T)\inc
1272: \ker\rho
1273: \inc\Out_0^T(G)\inc\Out^T(G)$ mentioned at the beginning of this
1274: section. We check
1275: that each group has uniformly bounded index in the next.
1276:
1277: The index of
1278: $\Out_0^T(G)$ in
1279: $\Out^T(G)$ is bounded by the order of the symmetry group of $\Gamma $.
1280: The number of edges of $\Gamma $ is not always uniformly bounded, but the
1281: first Betti number
1282: is fixed, and there is a uniform bound
1283: for the number $d$ of terminal vertices
1284: (because adding the relations $x_v=1$, for
1285: $v$ non-terminal, maps $G$ onto the free product of
1286: $d$ nontrivial finite cyclic groups). This is enough to bound the symmetry
1287: group.
1288:
1289: The map $\rho $ describes how automorphisms act on vertex groups.
1290: Since these groups are all
1291: commensurable, and isomorphic to $\Z$, the image of
1292: $\rho $ has order at most 2, so $\ker\rho $ has index at most 2 in
1293: $\Out_0^T(G)$. Finally,
1294: $\ker\rho $ is generated by $\T(T)$ together with bitwists. As vertex groups
1295: are abelian, bitwists belong to $\T(T)$, so
1296: $\T(T)=\ker\rho $.
1297: \cqfd\enddemo
1298:
1299:
1300: If $T $ is rigid, then $\Out^T(G)=\Out(G)$. We get:
1301:
1302: \thm{Theorem \sta} If $G$ is algebraically rigid, then $\Out(G)$
1303: contains
1304: $\Z^k$ as a subgroup of finite index.\cqfd
1305: \fthm
1306:
1307: In general, we have:
1308:
1309: \nom\comm
1310: \thm{Theorem \sta} Up to commensurability, the subgroup $\Out^T(G)$
1311: of $\Out(G)$ does not depend on
1312: $T$. It contains
1313: $\Z^k$ with finite index. \cqfd
1314: \fthm
1315:
1316: Another proof of the first assertion (and therefore of Proposition
1317: \commens, but not of Remark \divi) is given in Section 5. Also note the
1318: following related result:
1319:
1320: \nom\clay
1321: \thm{Theorem \sta{} [\Clc]} Any subgroup of $\Out(G)$
1322: commensurable with a subgroup of
1323: $\Out^T(G)$ is contained in $\Out^{T'}(G)$ for some GBS tree $T'$. \cqfd
1324: \fthm
1325:
1326: We now prove:
1327:
1328: \nom\tooo
1329: \thm{Theorem \sta} The set of prime numbers $p$ such that $\Out(G)$
1330: contains non-trivial $p$-torsion is finite.
1331: \fthm
1332:
1333:
1334:
1335: \demo{Proof} As any torsion element of $\Out(G)$ is contained in some
1336: $\Out^T(G)$ by [\Clc], and the index of $\T(T)$ in $\Out^T(G)$ is
1337: uniformly bounded by Proposition \unib, it suffices to control torsion in
1338: groups of twists.
1339:
1340: First note that the set of prime numbers dividing a label of $\Gamma $
1341: does not depend on $\Gamma $, as it does not change during an
1342: elementary collapse. Call it
1343: $\P$. If
1344: $T$ and
1345: $T'$ are related by an elementary collapse, Remark \divi{} shows that
1346: $\T(T)$ and
1347: $\T(T')$ have torsion at the same primes, except possibly those in $\P$.
1348: This implies that only finitely many primes may appear in the torsion
1349: of a group of twists: those in
1350: $\P$, and those in the torsion of $\T(T_0)$ for some fixed $T_0$.
1351: \cqfd\enddemo
1352:
1353: We have seen that $\Out(BS(2,4))$ contains arbitrarily large $2$-torsion. The
1354: proof of Theorem \tooo{} also shows:
1355:
1356: \thm{Corollary \sta} If $p$ is a prime number such that $\Out(G)$
1357: contains $p$-torsion of arbitrarily large order, then $p$ divides at
1358: least one label of each labelled graph representing $G$. \cqfd
1359: \fthm
1360:
1361: We do not know whether $p$ must divide some integral modulus.
1362:
1363: \head {\sect Unimodular groups}\endhead
1364:
1365:
1366:
1367: Let $\Gamma $ be a labelled graph representing a non-elementary
1368: unimodular group $G$, and $T$ the associated Bass-Serre tree. We denote by
1369: $G^+$ the kernel of
1370: $\Delta:G\to\{\pm1\}
1371: $ ({\it positive\/} elements). All elliptic elements are positive. Let $Z$ be the
1372: center of
1373: $G^+$. We know that it is cyclic, characteristic in $G$, and acts as the identity
1374: on
1375: $T$.
1376:
1377: We fix a nontrivial $\delta \in Z$. If $\Delta =1$ we may take $\delta $ to be a
1378: generator $\delta _0$, but the study of $\Out(G)$ when $\Delta \neq 1$ will
1379: require
1380: $\delta $ to be $(\delta _0)^4$. Note that any
1381: generator of an edge or vertex group is a root of $\delta
1382: $.
1383:
1384: Let $Z'$ be the cyclic group
1385: generated by $\delta $.
1386: There is an exact sequence
1387: $\{1\}\to Z'\to G\to H\to\{1\}$ with $H$
1388: virtually free.
1389: The group $H$ is the fundamental group of a graph of groups with the same
1390: underlying graph. Vertex and edge groups are finite cyclic groups,
1391: the order being the index of
1392: $\langle\delta \rangle$ in the original group. We denote by $\bar g$ the image
1393: of $g\in G$ in $H$.
1394:
1395:
1396:
1397: Since $Z'$ is characteristic in $G$, there are natural homomorphisms
1398: $ \Aut(G)\to \Aut(H)$ and $ \Out(G)\to \Out(H)$. The basic example is
1399: $\Out(F_n\times\Z)$, which contains the semi-direct product
1400: $\Z^n\rtimes\Out(F_n)$ with index $2$ (the factor $\Z^n$ should be thought of
1401: as $\text{Hom}(F_n,\Z)$). But the following examples illustrate a few of the
1402: subtleties involved when trying to lift automorphisms from
1403: $H$ to
1404: $G$.
1405:
1406: \example{Examples}
1407:
1408: $\bullet$ Let $G$ be $\langle a,b\mid a^3=b^3\rangle $ and $H$ be $\langle \bar
1409: a,\bar b\mid
1410: \bar a^3=\bar b^3=1\rangle $. The automorphism of
1411: $H$ mapping $\bar a$ to $\bar a\mi$ and $\bar b$ to $\bar b$ does not
1412: lift to $G$.
1413:
1414:
1415: $\bullet$ $G$ is $BS(3,3)=\langle a,t\mid ta^3t\mi=a^3\rangle $ and $H$ is
1416: $\langle
1417: \bar a,\bar t\mid
1418: \bar a^3=1\rangle $. The automorphism fixing $\bar a$ and sending $\bar t$
1419: to
1420: $\bar t\bar a$ has order 3, but all its lifts have infinite order.
1421:
1422: $\bullet$ $G$ is $BS(2,-2)=\langle a,t\mid ta^2t\mi=a^{-2}\rangle $ and $H$ is
1423: $\langle
1424: \bar a,\bar t\mid \bar a^2=1\rangle $. Conjugation by
1425: $\bar a$ in
1426: $H$ has lifts of order 2, such as $a\mapsto a, t\mapsto ata$, or
1427: $a\mapsto a\mi,t\mapsto a ta\mi$, but no lift of order 2 is inner.
1428:
1429: \endexample
1430:
1431: Let $H^+$ be the image of
1432: $G^+$ in
1433: $H$. If $\Delta \neq1$, it has index 2 (because $\delta $ is positive).
1434: There are only finitely many conjugacy classes of torsion elements in
1435: $H$ (they all come from vertex groups). All torsion elements of $H$
1436: belong to
1437: $H^+$, but a conjugacy class in $H$ may split into two classes in $H^+$.
1438:
1439: We shall now define a homomorphism $\tau :G \to
1440: \text{Isom}(\R)$ (it is similar to the homomorphism $G'\to\Z$ constructed in
1441: the proof of Proposition \kk). We fix a maximal tree
1442: $\Gamma _0\inc\Gamma $. Recall the presentation of
1443: $G$ with generators $x_v,t_\varepsilon $ and relations of the form
1444: $x_v^m= x_w^n$ or $t_\varepsilon x_v^mt_\varepsilon \mi= x_w^n$.
1445:
1446: To define $\tau $, send
1447: $\delta $ to $x\mapsto x+1$, send $x_v$ to $x\mapsto x+1/n_v$ if
1448: $\delta=x_v^{n_v}$, send
1449: $t_\varepsilon $ to
1450: $x\mapsto \Delta (t_\varepsilon )x$, and check that the relations are
1451: satisfied. This
1452: $\tau $ is not canonical (it depends on the choice of
1453: $\Gamma _0
1454: $); it is uniquely defined on elliptic elements once
1455: $\delta $ has been chosen.
1456:
1457: The image of $\tau $ in $\text{Isom}(\R)$ is infinite cyclic if
1458: $\Delta =1$, infinite dihedral if $\Delta \neq1$. Its kernel contains no
1459: nontrivial elliptic element. The coefficient of $x$ in $\tau (g)$ is $\Delta
1460: (g)$, and
1461: $\tau (gcg\mi)=\tau (c)^{\Delta (g)}$ if
1462: $c$ is positive (in particular if $c$ is elliptic).
1463:
1464: The map $\tau $ induces a map $\ov\tau $ from $H$ to a finite group $F$
1465: (the quotient of the image of $\tau $ by $x\mapsto x+1$). The group $F$
1466: is cyclic if
1467: $\Delta =1$, dihedral if $\Delta \neq 1$.
1468:
1469: \definition{Definition} We define the finite index
1470: subgroup $\Aut_0(H)\inc\Aut(H)$ as the set of automorphisms $\ov\alpha $
1471: such that:
1472: \roster
1473: \item
1474: $ \ov\alpha (H^+)=H^+$;
1475: \item
1476: $\ov\alpha $ acts trivially on the set of
1477: $H^+$-conjugacy classes of torsion elements;
1478: \item
1479: $\ov\tau
1480: \circ\ov\alpha =\ov\tau $.
1481: \endroster
1482: \enddefinition
1483:
1484: \nom\liftt
1485: \thm{Lemma \sta} Let $\ov\alpha \in\Aut_0(H)$. There exists a
1486: unique lift
1487: $\alpha \in\Aut(G)$ such that
1488: $\tau \circ\alpha =\tau $. It satisfies $\alpha (\delta )=\delta $.
1489: \fthm
1490:
1491:
1492:
1493:
1494: \demo{Proof}
1495: Uniqueness is easy: $\alpha (g)$ is determined up to a power of
1496: $\delta $, and that power is determined by applying $\tau $.
1497:
1498: We define $\alpha $ on the generators of $G$.
1499: In $H$, the element $\bar x_v$ has finite order and
1500: therefore is mapped by $\ov\alpha $ to
1501: $\bar g_v \bar x_v \bar g_v \mi$ for some $\bar g_v\in H^+$. We define
1502: $\alpha (x_v)=g_vx_vg_v\mi$, where
1503: $g_v\in G^+$ is any lift of $\bar g_v$. Note that $\tau (\alpha (x_v))=\tau
1504: (x_v)$ because $x_v$ and $g_v$ are positive. If $x_v^{m} =x_w^{n} $
1505: is a relation, then $\alpha (x_v)^{m} \alpha (x_w)^{-n }$ is $1$ because
1506: it is killed both in
1507: $H$ and by $\tau
1508: $. Note that $\alpha (\delta )=\alpha (x_v^{n_v})=g_vx_v^{n_v} g_v\mi=\delta
1509: $.
1510:
1511: Now consider a generator $t_\varepsilon $, and a lift
1512: $u_\varepsilon $ of $\ov\alpha (\bar t_\varepsilon )$.
1513: Since $\ov\tau
1514: \circ\ov\alpha =\ov\tau $, the elements
1515: $t_\varepsilon $ and $u_\varepsilon $ have the same image in $F$, so $\tau
1516: (t_\varepsilon u_\varepsilon \mi)$ is translation by an integer
1517: $n_\varepsilon $. We define $\alpha (t_\varepsilon )$ as
1518: $ \delta {}^{n_\varepsilon } u_\varepsilon $, so that $\tau (\alpha
1519: (t_\varepsilon ))=\tau (t_\varepsilon )$. Given a relation
1520: $t_\varepsilon x_v^{m}t_\varepsilon \mi=x_w^{n }$, the relation
1521: $\alpha (t_\varepsilon )\alpha (x_v)^{m}\alpha (t_\varepsilon )\mi=\alpha
1522: (x_w)^{n }$ holds modulo $\delta $. It also holds when we apply $\tau $, so it
1523: holds in $G$.
1524:
1525:
1526: We have constructed an endomorphism of $G$ fixing $\delta $ and
1527: inducing $\ov\alpha $, and this forces it to be an automorphism. \cqfd
1528:
1529:
1530:
1531: \enddemo
1532:
1533:
1534: Let $\Aut_0(G)\inc\Aut(G)$ be the finite index subgroup consisting of
1535: automorphisms fixing
1536: $\delta
1537: $ and mapping into
1538: $\Aut_0(H)$. We know that the map $\varphi :\Aut_0(G)\to\Aut_0(H)$
1539: is onto and has a section. We consider its kernel.
1540:
1541: \nom\kernn
1542: \thm{Lemma \sta} The kernel $N$ of
1543: $\varphi :\Aut_0(G)\to\Aut_0(H)$ is isomorphic to
1544: $\Z^{b }$. It is generated by twists by $\delta $ around the edges of
1545: $\Gamma
1546: \setminus \Gamma _0$.
1547: \fthm
1548:
1549: Recall that $b$ is the first Betti number of $\Gamma $.
1550:
1551:
1552: \example{Remark} It is a general fact that, whenever $Z\inc G$ is
1553: characteristic, the kernel of the map
1554: $ \Aut(G)\to \Aut(Z)\times\Aut(G/Z)$ is abelian [\RV, Proposition 2.5].
1555: To see this, take $\alpha _1,\alpha _2$ in the kernel. Write
1556: $\alpha _1(g)=z_1g$ and
1557: $\alpha _2(g)=gz_2$ (with $z_1,z_2\in Z$, depending on $g$), and
1558: deduce $\alpha _1\alpha _2(g)=\alpha _2\alpha _1(g)=z_1gz_2$ (one can
1559: also prove that $z_1$ must be in the center of $G$)
1560: \endexample
1561:
1562: \demo{Proof} Suppose $\alpha \in N$. We have $\alpha (\delta )=\delta
1563: $. If $x$ is a root of $\delta $, we have $\alpha
1564: (x )=x
1565: \delta {}^p$ and
1566: $\delta =x ^{q}$, so that $\delta =\alpha (x ^{q})=\delta \delta
1567: {}^{pq}$ and
1568: $p=0$. Therefore
1569: $\alpha
1570: $ fixes every elliptic element.
1571: Furthermore $\alpha (t_\varepsilon )=\delta {}^{n_\varepsilon
1572: }t_\varepsilon $ for some $n_\varepsilon \in\Z$, so $\alpha $ is a product of
1573: powers of twists by $\delta $. Conversely, each choice of integers
1574: $n_\varepsilon $ determines an automorphism fixing all elliptic elements and
1575: belonging to $N$.
1576: \cqfd
1577: \enddemo
1578:
1579:
1580: We have proved:
1581:
1582: \nom\semidaut
1583: \thm{Theorem \sta} If $G$ is non-elementary and unimodular, there is a split
1584: exact sequence
1585: $$\{1\}\to
1586: \Z^{b }\to\Aut_0(G)\overset\varphi \to\rightarrow\Aut_0(H)
1587: \to\{1\},$$ where $H$ is virtually free and
1588: $\Aut_0$ has finite index in $\Aut$. \cqfd
1589: \fthm
1590:
1591: We shall now show:
1592:
1593: \nom\esout
1594: \thm{Theorem \sta} If $G$ is non-elementary and unimodular, there is a split
1595: exact sequence
1596: $$\{1\}\to
1597: \Z^{k}\to\Out_0(G)\overset\psi \to\rightarrow\Out_0(H)
1598: \to\{1\},$$ where $H$ is virtually free and
1599: $\Out_0$ has finite index in $\Out$.
1600: \fthm
1601:
1602: See Proposition
1603: \kk{} for the definition and properties of $k$.
1604:
1605: Since $\Out(H)$ is VF [\KV], this implies:
1606:
1607: \thm{Corollary \sta} $\Out(G)$ and $\Aut(G)$ are virtually torsion-free and VF
1608: (they have finite index subgroups with finite classifying spaces).
1609: \cqfd
1610: \fthm
1611:
1612: \demo{Proof of Theorem \esout} We denote by $\Out_0$ the image of
1613: $\Aut_0$ in
1614: $\Out$ (note that $\Aut_0$ does not contain all inner automorphisms if
1615: $\Delta \neq1$), and by
1616: $\hat N$ the image of
1617: $N$ in $\Out(G)$.
1618: Let
1619: $\psi :\Out_0(G)\to\Out_0(H)$ be the natural map. Note that $\hat N$ is
1620: contained in $\ker\psi $, and has torsion-free rank $k$ by Lemma
1621: \kernn{} and Assertion (3) of Proposition \kk. We shall show $\ker\psi =\hat
1622: N\simeq\Z^k$. We write
1623: $i_g$ for conjugation by $g$.
1624:
1625: First assume $\Delta =1$. Then $k=b$, and $\hat N\simeq\Z^b$ because it has
1626: torsion-free rank $b$ and is a quotient of $N\simeq\Z^b$. Since the image of
1627: $\tau $ is abelian, every conjugation $i_g$ in $G$ satisfies $\tau \circ
1628: i_g=\tau
1629: $, and Lemma \liftt{} lifts $i_h\in\Aut_0(H)$ to $i_g$, where $i_g$ is any lift of
1630: $h$. Thus $\psi $ has a section.
1631:
1632: There remains to show $\ker\psi \inc\hat N$. If
1633: $\alpha \in\Aut_0(G)$ represents an element of $\ker\psi$, its image
1634: $\ov\alpha $ in $\Aut_0(H)$ is conjugation by some $h\in H$. Lift $i_h$ to
1635: $i_g\in\Aut_0(G)$, and consider $i_g\mi\alpha $. It belongs to $N$, and
1636: has the same image as $\alpha $ in
1637: $\Out(G)$. This implies $\ker\psi =\hat N$.
1638:
1639:
1640: Now suppose $\Delta \neq1$. In this case we have to choose $\delta =(\delta
1641: _0)^4$, where $\delta _0$ is a generator of $Z$ (the center of $G^+$). We first
1642: show that
1643: $\ker\psi =\hat N$. The argument is the same as before, but we have to prove
1644: that
1645: $i_h$ has a lift $i_g\in\Aut_0(G)$ (we will see that Lemma \liftt{} lifts inner
1646: automorphisms to inner automorphisms, but we cannot claim it at this point).
1647: Since
1648: $\ov\tau
1649: \circ\ov\alpha =\ov\tau $, the image $\ov\tau (h)$ is central in $F$. Our
1650: choice of $\delta $ ensures that the center of $F$ has order $2$ (it is
1651: generated by the image of $x\mapsto x+1/2$). In particular, $h$ is
1652: positive. If $g\in G$ is a lift of $h$, it commutes with $\delta $ and
1653: therefore $i_g$ belongs to
1654: $\Aut_0(G)$.
1655:
1656:
1657: We now prove $\hat N\simeq
1658: \Z^{b -1}$. Recall that
1659: $\hat N$
1660: has torsion-free rank ${b -1}$.
1661:
1662: Consider the twist $D$ by $\delta =(\delta_0) ^4$ around the edges
1663: $\varepsilon $ of
1664: $\Gamma
1665: \setminus \Gamma _0$ such that $\Delta (t_\varepsilon )=-1$ (it fixes the
1666: generators $x_v$, and maps $t_\varepsilon $ to $t_\varepsilon$ if $\Delta
1667: (t_\varepsilon )=1$, to
1668: $\delta t_\varepsilon $ if $\Delta (t_\varepsilon )=-1$).
1669: Note that $D$ is conjugation by
1670: $(\delta _0)^{2}$. Indeed, the
1671: $t_\varepsilon $'s with modulus $1$, and the
1672: $x_v$'s, are fixed by $D$ and commute with $\delta _0$, whereas
1673: $(\delta_0) ^4t_\varepsilon =(\delta _0)^{2} t_\varepsilon (\delta _0)^{-2}$ if
1674: $t_\varepsilon
1675: \delta t_\varepsilon \mi=\delta
1676: \mi$. Since $D$ belongs
1677: to a basis of
1678: $N$ (see Lemma \kernn), the image $\hat N$ of $N$ in
1679: $\Out(G)$ is isomorphic to
1680: $\Z^{b -1}$.
1681:
1682: Finally, we show that Lemma \liftt{} lifts inner automorphisms to
1683: inner automorphisms (and therefore $\psi $ has a section).
1684:
1685: Suppose $i_h$ belongs to $\Aut_0(H)$. Then $\ov\tau (h)$ is central in
1686: $F$, and therefore $\ov\tau (h^2)$ is trivial. If $g\in G$ is a lift of
1687: $h^2$, then
1688: $\tau (g)$ is an integral translation and we can redefine $g$ (multiplying
1689: it by a power of
1690: $\delta $) so that
1691: $\tau (g)$ is trivial. Then $\tau (gug\mi)=\tau (u)$ for every $u\in G$,
1692: showing that Lemma \liftt{} lifts conjugation by $ {h^2}$ to
1693: conjugation by $g$ in
1694: $G$. Now
1695: consider the lift $\alpha $ of $i_h$ given by Lemma \liftt{}. It satisfies
1696: $\alpha ^2=i_g$, and its image
1697: in
1698: $\Out(G)$ belongs to $\ker\psi $. Since
1699: $\ker\psi =\hat N$ is torsion-free, we conclude that $\alpha $ is inner.
1700: \cqfd
1701:
1702:
1703:
1704: \enddemo
1705:
1706: \head \sect The deformation space \endhead
1707:
1708:
1709: Let $G$ be a non-elementary GBS group. In this section, we work with
1710: {\it metric\/} GBS trees: $T$ is a metric tree, and $G$ acts by
1711: isometries. Metric trees are considered up to $G$-equivariant
1712: isometry.
1713:
1714: Let $\D$ be the space of metric GBS trees, and $P\D$ its
1715: projectivization (obtained by identifying two trees if they differ by
1716: rescaling the metric). We call $P\D$ the (canonical) {\it
1717: projectivized deformation space\/} of
1718: $G$.
1719:
1720: Choosing a $G$-invariant metric on a given simplicial tree amounts to
1721: assigning a positive length to each edge of
1722: $\Gamma =T/G$.
1723: This makes
1724: $P\D$ into a complex. An open simplex is the set of
1725: trees with a given underlying simplicial tree, a closed simplex is the
1726: set of GBS trees that may be obtained from trees in an open simplex
1727: by collapse moves (closed simplices have ``faces at
1728: infinity'', as the length of a non-collapsible edge is not allowed to be 0). Every
1729: closed simplex contains reduced trees.
1730:
1731: The group $\Out(G) $ acts on $P\D$. There is a bijection between the set
1732: of orbits of open simplices and the set of (unmarked) labelled graphs
1733: representing $G$ (up to admissible sign changes).
1734: Standard techniques
1735: show that
1736: $P\D$ has a natural $\Out(G)$-equivariant deformation retraction onto
1737: a simplicial complex (see [\CV, \McM]).
1738:
1739:
1740:
1741: GBS trees are locally finite. This implies that the complex $P\D$ is {\it
1742: locally finite\/}. Indeed, closed simplices containing $T$ consist of
1743: simplicial trees obtained from $T$ by expansion moves. Performing
1744: such moves on
1745: $T$ amounts to blowing up each vertex
1746: $v$ of
1747: $T$ into a subtree. Since $v$ has finite valence, there are only finitely
1748: many ways of expanding (not taking the metric into account).
1749: As remarked by M\. Clay, this local finiteness gives another
1750: proof of the first assertion of Theorem \comm.
1751:
1752:
1753:
1754:
1755: In general, there are several ways to define a topology on spaces of
1756: trees (equivariant Gromov-Hausdorff topology, axes topology, weak
1757: topology), but because of local finiteness they all coincide on $P\D$
1758: (see the discussion in [\GL, \GLL]). Clay [\Cld] proved that {\it $P\D$ is
1759: contractible\/} (see also [\GLL]). By Theorem \comm{}, stabilizers for
1760: the
1761: action of $\Out(G)$ on $P\D$ are virtually $\Z^k$.
1762:
1763: To sum up:
1764:
1765: \nom\comp
1766: \thm{Proposition \sta}
1767: $\Out(G)$ acts on the locally finite, contractible, complex $P\D$ with
1768: stabilizers virtually $\Z^k$. \cqfd
1769: \fthm
1770:
1771:
1772: If $G$ is algebraically rigid, the unique reduced GBS tree belongs to every
1773: closed simplex, $P\D$ is a finite complex, and the action of $\Out(G)$ on
1774: $P\D$ has a fixed point. If $G$ is not algebraically rigid, we
1775: will show that
1776: $\Out^T(G)$ always has infinite index in
1777: $\Out(G)$ (see Theorem 8.5). All $\Out(G)$-orbits are therefore infinite.
1778:
1779:
1780: Under suitable hypotheses, we now show that
1781: $P\D$ is ``small'' and we deduce information on $\Out(G)$.
1782:
1783:
1784: \subhead Groups with no nontrivial integral modulus \endsubhead
1785:
1786: Suppose that $G$ has no integral modulus other than $\pm1$
1787: (equivalently, $G$ does not contain a solvable Baumslag-Solitar group
1788: $BS(1,n)$ with $n>1$). In this case, there are only finitely many
1789: $\Out(G)$-orbits of simplices consisting of reduced trees [\For,
1790: Theorem 8.2], and therefore $\Out(G)$ acts on the complex $P\D$ with
1791: only finitely many orbits. This implies:
1792:
1793: \thm{Theorem \sta} Let $G$ be a non-elementary GBS group with no
1794: integral modulus other than $\pm1$.
1795: \roster
1796: \item $\Out(G)$ is $F_\infty$ (it has a $K(\pi ,1)$ with finitely many cells
1797: in every dimension).
1798: \item There is a bound for the cohomological dimension of
1799: torsion-free subgroups of $\Out(G)$.
1800: \item If $\Out(G)$ is virtually torsion-free, it has a finite index
1801: subgroup with a finite classifying space.
1802: \item $\Out(G)$ contains only finitely many conjugacy classes of finite
1803: subgroups.
1804: \endroster
1805: \fthm
1806:
1807: \demo{Proof} The first three assertions follow from Proposition
1808: \comp{} by standard techniques. Unfortunately, we do not
1809: know whether $\Out(G)$ must be virtually torsion-free when $G$ is not
1810: unimodular. Assertion (4) follows from Theorems \comm{} and \clay: any
1811: finite subgroup is contained in some $\Out^T(G)$, and it is well-known that
1812: there are finitely many conjugacy classes of finite subgroups in a group
1813: which is virtually
1814: $\Z^k$ (a proof appears in [\LGD]).
1815: \cqfd\enddemo
1816:
1817: \subhead Groups with no strict ascending loop \endsubhead
1818:
1819: It is shown in [\GLL] that, if no reduced labelled graph contains a strict
1820: ascending loop, there is an
1821: $\Out(G)$-equivariant deformation retraction from $P\D$ onto a
1822: finite-dimensional subcomplex. This implies:
1823:
1824: \thm{Theorem \sta}
1825: If no reduced labelled graph representing $G$ contains a strict
1826: ascending loop, there is a bound for the cohomological
1827: dimension of torsion-free subgroups of $\Out(G)$. \cqfd
1828: \fthm
1829:
1830: \head \sect Free subgroups in $\Out(G)$ \endhead
1831:
1832: Let $\Gamma $ be a labelled graph. Recall that we consider it up
1833: to admissible sign changes. In
1834: particular, when we focus on an edge, we will always assume that it is a
1835: $(p,q)$-segment with $p,q\ge1$ ($\ge2$ if $\Gamma $ is reduced), or a
1836: $(p,q)$-loop with $1\le p\le|q|$. Recall that a strict ascending loop is a
1837: $(1,q)$-loop with
1838: $|q|\ge2$. A pseudo-ascending loop is a
1839: $(p,q)$-loop with $p|q$.
1840:
1841: \midinsert
1842: \centerline
1843: {\includegraphics[scale=.5]
1844: {fig7.eps}}
1845: \captionwidth{220pt}
1846: \botcaption
1847: {Figure \the\fihno}
1848: {Slide move. }
1849: \endcaption
1850: \endinsert
1851:
1852:
1853: If $e,f$ are distinct oriented
1854: edges with the same origin $v$, and the label $\lambda _f$ of $f$ at
1855: $v$ divides the label $\lambda _e$ of $e$, one may {\it slide\/} $e$
1856: across
1857: $f$, replacing its label by $\frac{\lambda _e}{\lambda _f}\lambda _{\ov
1858: f}$ (see Figure
1859: \fig); both $e$ and
1860: $f$ may be loops, but they have to be distinct geometric edges ($f\neq
1861: \ov e$). See [\For] for details about slide moves.
1862: The important thing for
1863: us here is that performing a slide move on a labelled graph gives another
1864: labelled graph representing the same group $G$
1865: (only the GBS tree changes).
1866:
1867:
1868:
1869: An edge $f$ is a {\it slid\/} edge if some other edge may slide across
1870: $f$ or $\bar f$ (we usually think of slid edges as non-oriented edges). For
1871: example, any
1872: $(1,q)$-loop is a slid edge if $\Gamma $ contains more than one edge (i.e\. if
1873: $G$ is not solvable).
1874:
1875: \midinsert
1876: \centerline
1877: {\includegraphics[scale=.45]
1878: {fig8.eps}}
1879: \botcaption
1880: {Figure \the\fihno}
1881: {If $ \Gamma $ contains one of these graphs, $\Out(G)$ contains
1882: $F_2$.}
1883: \endcaption
1884: \endinsert
1885:
1886:
1887: The goal of this section is to prove the following result (see Figure
1888: \fig, where the numbers within parentheses refer to the assertions of
1889: the theorem):
1890:
1891: \nom\pal
1892: \thm{Theorem \sta} Let $\Gamma $ be a reduced labelled graph
1893: representing $G$. Suppose $\Out(G)$ does not contain $F_2$. Then:
1894: \roster
1895: \item A slid edge is either a $(2,2)$-segment or a
1896: $(1,q)$-loop.
1897: \item Slid edges are disjoint.
1898: \item A pseudo-ascending loop is a $(p,\pm
1899: p)$-loop or a
1900: $(1,q)$-loop.
1901: \item If $v$ is the basepoint of a $(1,q)$-loop with $|q|\ge2$, then no
1902: other label at
1903: $v$ divides a power of $q$.
1904: \item
1905: If $v$ is the basepoint of a $(1,q)$-loop with $|q|\ge2$, and $r,s$ are two
1906: labels at $v$ not carried by the loop, then
1907: $s$ does not divide any $rq^n$.
1908: \item Let $vw$ be a $(2,2)$-segment. Let $r$ be a label at $v$, and $s$ a
1909: label at $w$ (other than those carried by $vw$). If $r|s$ and $s$ is even, then
1910: $r=s$ and the labels are carried by the same non-oriented edge.
1911:
1912: \endroster
1913: \fthm
1914:
1915: We prove Theorem \pal{} in several steps. Our two main tools will be
1916: slide moves and Lemma \tw.
1917:
1918: \subhead Slid segments are $(2,2)$-segments \endsubhead
1919:
1920: Let $f$ be a slid
1921: $(p,q)$-segment with $p,q\ge2$. Consider the graph of groups $\Theta $
1922: obtained by collapsing $f$. It has a vertex group $H_v$
1923: isomorphic to $\langle a,b\mid a^p=b^q\rangle$. Let
1924: $e$ be an edge of
1925: $\Gamma $ that may slide across $f$, viewed as an edge of $\Theta $.
1926: Its group $G_e$ is generated by a power of $a^p$, so is central in $H_v$.
1927: By Lemma \tw, the group $\T(\Theta )\inc\Out(G)$ maps onto $J=
1928: Z_{H_v}(G_e)/\langle Z( H_v),Z(G_e)\rangle=\langle a,b\mid
1929: a^p=b^q=1\rangle$. If
1930: $(p,q)\neq(2,2)$, the group $J$ contains
1931: $F_2$, so $\Out(G)$ contains $F_2$.
1932:
1933: \nom\tee
1934: \example {Remark \sta} For future reference, note that $\T(\Theta )$ contains
1935: $\T(\Gamma )$ by Remark \te, and furthermore the image of $\T(\Gamma )$ in
1936: $J$ is finite. To see this, recall that $\T(\Gamma )$ is generated by the groups
1937: $Z_{G_{o(f')}}(G_{f'})$ with $f'\neq f,\bar f$ (Remark \te). All these groups have
1938: trivial image in
1939: $J$, except
1940: $Z_{G_{o(e)}}(G_e)$ whose image is finite.
1941: \endexample
1942:
1943: \subhead Slid loops are
1944: $(1,q)$-loops \endsubhead
1945:
1946: We assume that $e$ slides across a $(p,q)$-loop $f$ with $2\le
1947: p\le|q|$, and we show that $\Out(G)$ contains $F_2$. If
1948: $|q|\ge 3$, we create a slid $(p,q)$-segment by performing an expansion
1949: move (replacing the loop by a $(p,q)$-segment and a $(1,1)$-segment),
1950: and we apply the previous argument (which is valid even if the labelled graph
1951: is not reduced). If
1952: $f$ is a
1953: $(2,\pm2)$-loop, we collapse it and we apply Lemma \tw. We now have
1954: $H_v=\langle a,t\mid ta^2t\mi=a^{\pm2}\rangle$, and $H_e$ is
1955: generated by a power of $a^2$. The quotient $Z_{H_v}(G_e)/\langle Z(
1956: H_v),Z(G_e)\rangle$ is isomorphic to $\Z*\Z/2\Z$, so contains $F_2$.
1957:
1958: \subhead Pseudo-ascending loops are $(p,\pm
1959: p)$-loops or
1960: $(1,q)$-loops \endsubhead
1961:
1962:
1963: This amounts
1964: to showing that $\Gamma $ cannot contain an $(r, rs)$-loop with $r\ge2$
1965: and $|s|\ge 2$. If it does, write
1966: $ G=G_1*_{\langle a\rangle}G_2$, with $G_1=\langle a,t\mid
1967: ta^rt\mi=a^{rs}\rangle$. By [\Collev], there exist two automorphisms of
1968: $G_1$ fixing $a$ and generating a free subgroup of rank 2 in
1969: $\Out(G_1)$ (in the notation of [\Collev], set $\alpha ^r=\varphi
1970: _0=\tau =1$ to see that the subgroup of
1971: $\Out(G_1)$ generated by $\alpha $ and $\gamma _2$ maps
1972: onto $\Z/r\Z*\Z/s^2\Z$). Extend the automorphisms by the identity
1973: on $G_2$ and check that they generate $F_2\inc\Out(G)$.
1974:
1975: Here is another argument, valid when $r$ or $|s|$ is bigger than 2:
1976: perform an expansion and a slide to obtain a graph with a slid
1977: $(r,s)$-segment.
1978:
1979:
1980:
1981: \subhead Slid edges are disjoint \endsubhead
1982:
1983: We argue by way of contradiction.
1984: There are several cases to consider (they are pictured from right to left on
1985: Figure 8). First suppose that
1986: $v$ belongs to a slid
1987: $(2,2)$-segment $f$ and a $(1,q)$-loop $f'$. If $q$ is even, one may slide $f'$
1988: across $f$ and then collapse $f'$. This creates a $(2,2q)$-loop, a contradiction.
1989: If $q$ is odd, some other edge
1990: may slide across $f$. Sliding $f$ around $f'$
1991: makes $f$ a slid $(2,2q)$-segment, a contradiction if $|q|>1$. If $q=\pm1$,
1992: collapse both
1993: $f$ and
1994: $f'$ and apply Lemma
1995: \tw.
1996:
1997: Now suppose $v$ belongs to a $(1,q)$-loop $f$ and a $(1,r)$-loop $f'$.
1998: If $|q|\ge2$, sliding $f'$ around $f$ makes it a $(q, r)$-loop, and then sliding
1999: $\bar f'$ twice makes it a
2000: $(q,q^2r)$-loop.
2001: If
2002: $|q|=|r|=1$, we may write $G$ as an amalgam $G_1*_{\langle
2003: a\rangle}G_2$, with $G_2=\langle a,t,t'\mid
2004: tat\mi=a^{\pm1},t'at'{}\mi=a^{\pm1}\rangle$. It is easy to embed $F_2$
2005: into $\Out(G)$ by using automorphisms of $G_2$ fixing $a$.
2006:
2007: Finally, suppose that $(2,2)$-segments $f$ and $f'$ have a vertex $v$ in
2008: common. We may assume that their other endpoints are distinct, as
2009: otherwise sliding $f$ across $f'$ would create a
2010: slid $(2,\pm2)$-loop.
2011:
2012: The fundamental group of the subgraph of groups $f\cup f'$ is
2013: $J=\langle a,b,c\mid a^2=b^2=c^2\rangle $. Consider the following
2014: automorphisms
2015: $\alpha ,\beta $ of $J$: $\alpha $ fixes $a$ and $b$ and conjugates $c$ by
2016: $ba$, while
2017: $\beta
2018: $ fixes $b,c$ and conjugates $a$ by $bc$. They extend to automorphisms
2019: of $G$ (they are twists in the graph of groups obtained from $\Gamma
2020: $ by collapsing
2021: $f'$ and
2022: $f$ respectively). We claim that $\alpha ,\beta $ generate a free
2023: nonabelian subgroup of $\Out(J)$ (hence also of
2024: $\Out(G)$ because $J$ is its own normalizer).
2025:
2026: Indeed, consider $\ov J\simeq
2027: \Z/2\Z*\Z/2\Z*\Z/2\Z$ obtained by adding the relation $a^2=1$. Let
2028: $J^+\inc \ov J$ be the subgroup of index $2$ consisting of elements
2029: of even length. It is free with basis
2030: $ \{\bar a\bar b,\bar b\bar c\}$. With respect to this basis, $\alpha $
2031: acts on the abelianization of
2032: $J^+$ as the matrix $\pmatrix 1 & 2 \\ 0 & 1
2033: \endpmatrix$, $\beta $ acts as $\pmatrix 1 & 0 \\ 2 & 1
2034: \endpmatrix$, and inner automorphisms of $\ov J$ act as $\pm Id$. It
2035: follows that there is no nontrivial relation between $\alpha $ and
2036: $\beta $ in $\Out(J)$.
2037:
2038:
2039:
2040:
2041:
2042:
2043: \subhead Labels near a $(1,q)$-loop \endsubhead
2044:
2045: Let $v$ be the basepoint of a $(1,q)$-loop. We already know that no
2046: other label $r$ at $v$ equals 1 or divides $q$ (it would be carried by a
2047: slid edge). Suppose that
2048: $r$ divides some $q^n$. Let $\ell$ be a prime divisor of $r$. Expand $v$
2049: so as to create a
2050: $(1,\ell)$-segment, and collapse the $(1,\frac q \ell)$-edge (see Figure
2051: \fig).
2052: The new labelled graph is isomorphic to $\Gamma
2053: $, except that $r$ has been divided by $\ell$ and other indices at $v$
2054: have been multiplied by $\frac q \ell$. Repeat this operation until
2055: $r$ divides $q$, a case already ruled out.
2056:
2057: \midinsert
2058: \centerline
2059: {\includegraphics[scale=.45]
2060: {fig9.eps}}
2061: \captionwidth{220pt}
2062: \botcaption
2063: {Figure \the\figno}
2064: {Replacing $r$ by $r/\ell$. }
2065: \endcaption
2066: \endinsert
2067:
2068:
2069:
2070: Now let $r,s$ be labels at $v$ carried by edges $e,f$, with $s|rq^n$. If
2071: $f\neq\ov e$, we can make $e$ a slid edge by sliding it
2072: $n$ times across the loop. If $f=\ov e$, we can create an $(s,
2073: rq^{n+1})$-loop, contradicting (3).
2074:
2075:
2076: \subhead Labels near a $(2,2)$-segment \endsubhead
2077:
2078: Let $vw$ be a $(2,2)$-segment. We already know that all other labels at
2079: $v$ and
2080: $w$ are bigger than 2 in absolute value. Furthermore, if $r,s$ are labels
2081: at the same vertex ($v$ or $w$), and
2082: $r|s$, then $r=\pm s$ and they are carried by a loop. Assertion (5) of
2083: the theorem follows: since $s$ is even, one can perform a slide across
2084: $vw$ so that
2085: $s$ becomes a label at $v$.
2086:
2087: \head \sect { Groups with $\Out(G)\not\supset F_2$ }\endhead
2088:
2089: In this section, we complete the proof of Theorem \six{} by showing:
2090:
2091: \nom\main
2092: \thm{Theorem \sta} If a non-elementary GBS group $G$ is represented by a
2093: reduced labelled graph
2094: $\Gamma $ satisfying the conclusions of Theorem \pal, then
2095: $\Out(G)$ is virtually nilpotent of class at most 2.
2096: \fthm
2097:
2098: The theorem is true if $G$ is a solvable Baumslag-Solitar group [\Coll], so we
2099: rule out this case.
2100: As we wish to study the whole automorphism group of $G$, it is important in
2101: this proof to think of
2102: $\Gamma
2103: $ as a marked graph. As usual, we denote by $T$ its Bass-Serre tree.
2104:
2105:
2106: In general, there are many graphs representing $G$, and the first step in the
2107: proof of Theorem \main{} will be to show that collapsing the slid edges of
2108: $\Gamma $ yields a marked graph of groups $\Theta $ and a (non GBS)
2109: tree $S$ which are canonical (they do not depend on $\Gamma $). In the
2110: language of [\Forr], we first show that all possible trees $S$ belong to the
2111: same deformation space, and then using [\Ler] that there is only one reduced
2112: tree in that space. In particular
2113: $\Out(G)=\Out^S(G)$, and we shall conclude the proof of Theorem \main{} by
2114: showing that
2115: $\Out^S(G)$ is virtually nilpotent.
2116:
2117: We know that the slid edges of $\Gamma $ are disjoint, and
2118: are either
2119: $(2,2)$-segments or $(1,q)$-loops. Define $\Theta $ and $S$ by collapsing
2120: them.
2121:
2122: Consider edge and vertex groups of $\Theta $.
2123: Edge groups are cyclic. Non-cyclic vertex groups are
2124: Klein bottle groups
2125: $\langle a,b\mid a^2=b^2\rangle$ arising from collapsed $(2,2)$-segments,
2126: and solvable Baumslag-Solitar groups
2127: $BS(1,q)= \langle a,t\mid tat\mi=a^q\rangle$ arising from collapsed
2128: $(1,q)$-loops. Note two special cases:
2129: $\Z^2$ if $q=1$, a Klein bottle group if $q=-1$.
2130:
2131: It is useful to think of $ BS(1,q)$, for $|q|\ge2$, as the subgroup of the affine
2132: group of $\R$
2133: generated by $a:x\mapsto x+1$ and
2134: $t:x\mapsto qx$. It consists of all maps of the form $x\mapsto q^\alpha
2135: x+\beta $ with $\alpha \in\Z$ and $\beta \in\Z[\frac1{|q|}]$. One deduces,
2136: for instance, that powers
2137: $a^i, a^j$ are conjugate if and only if
2138: $\frac ij$ is a power of $q$. The element $a^r$ has an $s$-root if and only if $s$
2139: divides some $rq^n$; the root is then unique.
2140:
2141:
2142: \nom\inva
2143: \thm{Lemma \sta} The set of vertex stabilizers of $S$ does not depend on the
2144: marked graph
2145: $\Gamma $. In particular, it is $\Out(G)$-invariant.
2146: \fthm
2147:
2148:
2149: \demo{Proof} By Corollary \ellind, it suffices to show that the vertex
2150: stabilizers of
2151: $S$ are determined by the elliptic subgroups of $T$. This will be done
2152: using the tree $T$, but the description of vertex stabilizers will involve
2153: only the algebraic structure of the set of elliptic subgroups.
2154:
2155: Let $v$ be a vertex
2156: of
2157: $T$, and
2158: $\ov v$ its projection in
2159: $\Gamma $. We denote by $\pi :T\to S$ the
2160: collapse map. We want to understand the stabilizer
2161: $H_v$ of $\pi (v)$ in $S$. It is also the stabilizer of the subtree $\pi \mi(\pi
2162: (v))\inc T$.
2163:
2164: If no strict ascending loop is attached at $\ov v$, the
2165: stabilizer
2166: $G_v$ of
2167: $v$ is a maximal elliptic subgroup of $T$. Conversely every maximal
2168: elliptic subgroup arises in this way.
2169: If a $(1,q)$-loop with $|q|\ge2$ is
2170: attached at $\ov v$, there is no maximal elliptic
2171: subgroup containing $G_v$; if $T'$ is another GBS tree,
2172: then
2173: $G_v$ is not necessarily a vertex stabilizer of $T'$ (see [\For]).
2174:
2175: We first determine the group
2176: $H_v$ containing a maximal elliptic subgroup $G_v$.
2177:
2178: Consider the normalizer $N(G_v)$. If it is $\Z^2$ or a
2179: Klein bottle group, then $\ov v$ bounds a $(1,\pm1)$-loop and
2180: $H_v=N(G_v)$. Otherwise, $N(G_v)=G_v$. Let $a$ be a generator of
2181: $G_v$. Then
2182: $H_v$ is the centralizer $Z(a^2)$ if $\ov v$ bounds a slid
2183: $(2,2)$-edge,
2184: $G_v$ if not.
2185:
2186: To decide which (in terms of $G_v$ only), first observe that $\ov
2187: v$ bounds a
2188: $(2,2)$-edge (slid or not) if and only if $Z(a^2)$ is a Klein bottle group.
2189: Assuming it is, consider (as in [\GLL]) the set of groups of
2190: the form
2191: $Z(a^2)\cap K$, where
2192: $K$ is an elliptic subgroup of $T$ not contained in $Z(a^2)$. It is easy to see
2193: that the edge is slid if and only if some maximal element of
2194: this set (ordered by inclusion) is contained in $\langle a^2\rangle$.
2195:
2196:
2197:
2198: Now suppose that a $(1,q)$-loop with $|q|\ge2$ is
2199: attached at $\ov v$.
2200: Then $H_v= \langle a,t\mid tat\mi=a^q\rangle$,
2201: where $a$ is a generator of $G_v$.
2202:
2203: Condition (4) of Theorem \pal{} implies that
2204: {\it $\pi (v)$ is the only point of $S$ fixed by $a^\ell$ if $\ell$ divides a power
2205: of
2206: $q$\/}: stabilizers of edges adjacent to $\pi (v)$ in $S$ are conjugate in
2207: $H_v$ to
2208: $\langle a^r\rangle$, where $r$ does not divide any $q^n$; since $a^\ell$
2209: is not conjugate (in $H_v$) to a power of such an $a^r$, it cannot fix an edge.
2210:
2211: As in [\For], say that an
2212: elliptic subgroup $K$ of $T$ is {\it vertical\/} if any elliptic subgroup $K'$
2213: containing $K$ is contained in a conjugate of
2214: $K$. For the action of $H_v$ on $T$, {\it the subgroup $\langle a^\ell\rangle$
2215: is vertical if and only if
2216: $\ell$ divides a power of $q$\/}. We show that the same result holds
2217: for the action of $G$ on $T$.
2218:
2219: Suppose that $\ell$
2220: divides a power of $q$. If $K'\supset \langle a^\ell\rangle$ is elliptic (in
2221: $T$, hence also in $S$), then $\pi (v)$ is the only point of $S$ fixed by
2222: $K'$, so $K'$ fixes a point $w\in \pi \mi(\pi (v))\inc T$. The
2223: stabilizer of $w$ is conjugate to $\langle a \rangle$ in $H_v$, and
2224: $\langle a \rangle$ is contained in a conjugate of $\langle a^\ell \rangle$
2225: because $\ell $ is a power of $q$. Thus $K'$ is contained in a conjugate of
2226: $\langle a^\ell\rangle$, and $\langle
2227: a^\ell\rangle$ is vertical as required. Conversely, if
2228: $\langle a^\ell\rangle$ is vertical, then it contains a conjugate
2229: $g\langle a \rangle g\mi$ with $g\in G$. Since $\pi (v)$ is the only point of $S$
2230: fixed by
2231: $a $, we have $g\in H_v$ and we deduce that $\ell$ divides a power of $q$.
2232:
2233:
2234: We now conclude the proof, by characterizing the vertex stabilizer $H_v$ of
2235: $S$ containing a vertical subgroup
2236: $K\inc G$ which is not maximal elliptic. We know that
2237: $K$ is generated by $ a^\ell $, where $a$ generates a vertex stabilizer $G_v$,
2238: a
2239: $(1,q)$-loop with $|q|\ge2$ is attached to $\ov v$, and $\ell$ divides a
2240: power of
2241: $q$. In particular,
2242: $K$ fixes a unique
2243: point $\pi (v)\in S$.
2244:
2245: The stabilizer $H_v$ of $\pi (v)$ is isomorphic to $BS(1,q)$. The set of
2246: elements of
2247: $H_v$ which are elliptic in $T$ is an abelian subgroup (isomorphic to
2248: $\Z[\frac1{|q|}]$), so
2249: $gKg\mi$ commutes with
2250: $K$ if
2251: $g\in H_v$.
2252: Conversely, if $gKg\mi$ commutes with
2253: $K$, then $g\in H_v$ because $\pi (v)$ is the only fixed point of $K$ in
2254: $S$. We
2255: may now characterize $H_v$ (independently of
2256: $T$) as the set of
2257: $g\in G$ such that
2258: $gKg\mi$ commutes with
2259: $K$.
2260: \cqfd\enddemo
2261:
2262: \nom\indep
2263: \thm{Lemma \sta} The $G$-tree $S$ does not depend on $\Gamma $.
2264: In particular,
2265: $\Out^S(G)=\Out(G)$.
2266: \fthm
2267:
2268: \demo{Proof} We apply the main result of [\Ler].
2269: Since $S$ is reduced (no inclusion $G_e\hookrightarrow H_v$ is onto), it
2270: suffices to check that the following holds. Let $e$ and
2271: $f$ be oriented edges of
2272: $S$ with the same origin such that $G_f\inc G_e$; if $e,f$ do not
2273: belong to the same
2274: $G$-orbit, then $e,\ov f$ are in the same orbit and $G_e=G_f$.
2275:
2276: Let $v,w$ be the origins of $e$ and $f$ in $T$. Let $ e_0$, $ f_0$ be the
2277: projections
2278: in $\Gamma $, and $r,s$ the corresponding labels. We distinguish several
2279: cases.
2280:
2281: First assume $\ov v=\ov w$ in $\Gamma $.
2282: If no collapsing takes place at
2283: $\ov v$, or if $\ov v$ bounds a $(1,\pm 1)$-loop, then $G_f\inc G_e$ implies
2284: that $r$ divides $s$. This is possible only if $e_0$ and $f_0$ are opposite
2285: edges forming a
2286: $(p,\pm p)$-loop.
2287:
2288: Now suppose that $\ov v$ bounds a $(1,q)$-loop with $|q|\ge2$. Write
2289: the corresponding vertex stabilizer $H_v$ of $S $ as
2290: $ \langle a,t\mid tat\mi=a^q\rangle$. Then $\langle a^s\rangle$ is conjugate
2291: in $H_v$ to a subgroup of $\langle a^r
2292: \rangle$, so there exists $n$ such that $\frac s{nr}$ is a power of $q$.
2293: This contradicts Assertion (5) of Theorem \pal, so this case cannot
2294: occur.
2295:
2296: Finally, suppose that $\ov v\ov w$ is a slid $(2,2)$-edge.
2297: The stabilizer of $\pi (v)$ in $S$ is then $ H_v=\langle
2298: a,b
2299: \mid a^2=b^2\rangle$. The subgroups $G_e$ and $G_f$ of $H_v$ are generated
2300: by conjugates of powers of $a$ or $b$. Distinct
2301: powers of
2302: $a$ (resp\. $b$) are not conjugate in $H_v$, while $a^i$ is conjugate to $b^j$
2303: only when $i$ and $j$ are equal and even. In particular, $r$
2304: divides $s$. If $e_0$ and $f_0$ have the same
2305: origin ($\ov v$ or $\ov w$), we conclude as in the first case. If not, then $s$
2306: must be even and we use Assertion (6) of Theorem \pal.
2307: \cqfd\enddemo
2308:
2309: We may now study $\Out(G)=\Out^S(G)$ using the results of [\LGD]
2310: recalled in Section 3. In particular, $\rho $ has a restriction
2311: $\rho _1:\Out_1^S (G)\to\prod_{u\in W}\Out(H_u)$ with $\Out_1^S (G)$ of
2312: finite index in $\Out(G)$ and $\ker\rho _1= \T(S)$ (we denote by $W$ the
2313: vertex set of
2314: $\Theta $, and by $H_u$ the vertex group of $u\in W$).
2315:
2316: We first show that $\T(S)$ is virtually abelian. It is generated by centralizers
2317: of edge groups in vertex groups $H_u$. If $H_u$ is $\Z$ or $\Z^2$, the
2318: centralizer is of course $H_u$. If
2319: $H_u=BS(1,q)$ with
2320: $|q|\ge2$, the centralizer is an infinitely generated abelian
2321: group isomorphic to $\Z[\frac1{|q|}]$. If
2322: $H_u$ is a Klein bottle group, the centralizer is $\Z^2$ if $u$ comes
2323: from a $(1,-1)$-loop, $\Z$ if $u$ comes from a slid $(2,2)$-segment and
2324: the edge group is not central, the whole of $H_u$ if the edge group is
2325: central. Since a Klein bottle group is virtually abelian, so is $\T(S)$.
2326:
2327: \nom\fingen
2328: \example{Remark \sta} Relations in the presentation of $\T(S)$ come from
2329: centers of edge and vertex groups of
2330: $\Theta
2331: $. Since these centers are $\{1\}$, $\Z$, or $\Z^2$, the group
2332: $\T(S)$ is finitely generated if and only if $\Gamma $ contains no strict
2333: ascending loop.
2334: \endexample
2335:
2336: Now fix a vertex
2337: $u$ of $\Theta $, and define $P_u\inc \Out(H_u)$ by projecting the
2338: image of
2339: $\rho _1$.
2340: If
2341: $H_u$ is
2342: $\Z$ or a Klein bottle group, $P_u$ is finite because $\Out(H_u)$ is
2343: finite. We claim that $P_u$ is finite also when $H_u$ is $BS(1,q)$ with
2344: $|q|\ge2$.
2345:
2346: Write $
2347: H_u=\langle a,t\mid tat\mi=a^q\rangle$. The vertex $u$ is obtained by
2348: collapsing a $(1,q)$-loop $f_u$ of $\Gamma $. Denote its basepoint by
2349: $v$.
2350: Since $G$ is assumed not to be solvable, we may consider an edge group
2351: $\langle a^r\rangle$, where $r$ is a label near $v$ not carried by $f_u$. Its
2352: image by an automorphism
2353: $\alpha
2354: \in P_u$ is also an edge group, so $\alpha (\langle a^r\rangle)$ is conjugate to
2355: $\langle a^s\rangle$ for some label $s$ (possibly equal to $r$). But $a^r$ has
2356: an $s$-th root only if $s$ divides some $rq^n$, so
2357: $r=s$ by Assertion (5) of Theorem \pal. By uniqueness of roots, $\alpha $
2358: maps
2359: $a$ to a conjugate of
2360: $a^{\pm1}$. Only finitely many outer automorphisms of $H_u$ have this
2361: property [\Coll], so $P_u$ is indeed finite.
2362:
2363: The group $P_u$ is infinite only when $u$ comes from collapsing a $(1,1)$-loop
2364: ${f_u}$. In this case, $H_u=\langle a,t\mid tat\mi=a \rangle$. As above, $a$
2365: must be mapped to a conjugate of $a^{\pm1}$, so $P_u$ contains with index at
2366: most 2 the group generated by the automorphism $D_{f_u}$ fixing $a$ and
2367: mapping $t$ to $at$. We view $D_{f_u}$ as an automorphism of $G$ (extend
2368: it by the identity). It is a twist relative to $\Gamma $, but not to $\Theta $
2369: (Remark \te{} does not apply here, as $f_u$ is not a segment; in general, none
2370: of the groups
2371: $\T(S )$, $\T(T )$ contains the other).
2372:
2373: This analysis shows that {\it
2374: $\T(S)$ and the automorphisms
2375: $D_{f_u}$ associated to
2376: $(1,1)$-loops of $\Gamma $ generate a finite index subgroup of
2377: $\Out(G)$\/}.
2378: We replace $\T(S)$ by an abelian subgroup $\T_0(S)$ of finite index,
2379: and
2380: we complete the proof of Theorem \main{} by showing that the subgroup
2381: generated by
2382: $\T_0(S)$ and the automorphisms $D_{f_u}$ is virtually nilpotent of
2383: class
2384: $\le2$: every commutator is central.
2385:
2386: Non-commutativity only comes from the fact that $D_{f_u}$ may fail to
2387: commute with $D(z)$, when
2388: $z\in H_u$ and
2389: $D(z)$ is a twist of $\Theta $ around an edge
2390: $e$ with origin $u$. Write $H_u=\langle a,t\mid tat\mi=a \rangle$.
2391: The group
2392: $G_e$ is generated by a power of
2393: $a$.
2394:
2395:
2396: Recall that $D_{f_u}$ fixes $a$ and maps $t$ to $at$. In particular, $D(z)$
2397: commutes with $D_{f_u}$ if $z$ is a power of $a$ (both automorphisms belong
2398: to
2399: $\T(T)$). The interesting case is when $z=t$ (geometrically, $u$ carries a
2400: 2-torus $T^2$, $e$ carries an annulus attached to a meridian of $T^2$,
2401: $D_{f_u}$ is a Dehn twist in $T^2$ around a meridian, and
2402: $D(t)$ drags the annulus around $T^2$ along the longitude). But
2403: conjugating $D(t)$ by
2404: $D_{f_u}$ gives $D(ta)$, so the commutator of $D(t)$ and $D_{f_u}$ is $D(a)$, a
2405: central element. This easily implies that every commutator is
2406: central, completing the proof of Theorem
2407: \main.
2408:
2409: \head \sect Further results\endhead
2410:
2411: \subhead Nilpotent vs abelian\endsubhead
2412:
2413: \thm{Corollary \sta} If $G$ is represented by a reduced labelled graph with no
2414: $(1,1)$-loop, then $\Out(G)$ contains $F_2$ or is virtually abelian.
2415: \fthm
2416:
2417: This follows immediately from the proof of Theorem \main. More generally,
2418: if $\Out(G)$ does not contain $F_2$, it is
2419: virtually abelian if and only if every commutator
2420: $D(a)$ as in the last paragraph of the proof has finite order. This happens in
2421: particular when the basepoint of every
2422: $(1,1)$-loop has valence 3.
2423: See Remark \foo{} for a more complete discussion.
2424:
2425: \subhead Finite generation\endsubhead
2426:
2427: Here is a general fact:
2428:
2429: \nom\finig
2430: \thm{Proposition \sta} Let $\Gamma $ be a labelled graph representing
2431: a GBS group $G$. If $\Gamma $ contains a strict ascending loop, but $G$
2432: is not a solvable Baumslag-Solitar group, then $\Out(G)$ has an
2433: infinitely generated abelian subgroup.
2434: \fthm
2435:
2436: \demo{Proof} Collapse the loop and apply Lemma
2437: \tw{} to an edge $e$ with origin at the collapsed vertex $v$ (there is such an
2438: edge because $G$ is not solvable). We have
2439: $H_v=BS(1,q)$ with $|q|\ge2$, and $Z_{H_v}(G_e)$ is
2440: infinitely generated abelian (it is isomorphic to $ \Z[\frac1{|q|}]$). The
2441: center of $H_v$ is trivial, and the center of $G_e$ is cyclic. The
2442: subgroup of $\T(\Theta )$ generated by $Z_{H_v}(G_e)$ is isomorphic
2443: to $ \Z[\frac1{|q|}]$, or is an infinite abelian torsion group.
2444: \cqfd\enddemo
2445:
2446: From the proof of Theorem \main{} we get:
2447:
2448: \thm{Corollary \sta} Let $\Gamma $ be a reduced labelled graph
2449: representing a non-solvable GBS group $G$ with $\Out(G)$
2450: virtually nilpotent. The group $\Out(G)$ is finitely
2451: generated if and only if $\Gamma $ contains no strict ascending loop.
2452: \fthm
2453:
2454: \demo{Proof} We have seen that $\Out(G)$ is generated by the union of
2455: $\T(S)$ and a finite set, so by virtual nilpotence $\Out(G)$ is finitely
2456: generated if and only if
2457: $\T(S)$ is finitely generated. The corollary now follows from Remark \fingen.
2458: \cqfd
2459: \enddemo
2460:
2461: \thm{Corollary \sta} If no label of $\Gamma $ equals 1, then $\Out(G)$ contains
2462: $F_2$ or is finitely generated and virtually abelian. \cqfd
2463: \fthm
2464:
2465: In the virtually abelian case, the torsion-free rank may be
2466: computed from $k$ and the labels near the $(2,2)$-edges.
2467:
2468:
2469: \subhead Algebraic rigidity\endsubhead
2470:
2471: \thm{Theorem \sta} If the GBS group $G$ is not a solvable
2472: Baumslag-Solitar group, the following are equivalent:
2473: \roster
2474: \item $G$ is algebraically rigid (there is only one reduced GBS tree).
2475: \item The deformation space $P\D$ is a finite complex.
2476: \item $\Out(G)$ is virtually $\Z^k$ (with $k$ defined in Proposition \kk).
2477: \item Let $\Gamma $ be any reduced labelled graph representing $G$. If $e,f$
2478: are distinct oriented edges of $\Gamma $ with the same origin
2479: $v$, and the label of $f$ divides that of $e$, then either $e=\ov f$ is a
2480: $(p,\pm p)$-loop with $p\ge2$, or $v$ has valence 3 and bounds a
2481: $(1,\pm1)$-loop.
2482: \endroster
2483: \fthm
2484:
2485: \example{Remarks} $\bullet$ If $G$ is unimodular, $(3)\Leftrightarrow(4)$
2486: follows from Theorem \esout{}
2487: and
2488: [\Pett, Corollary 5.14].
2489:
2490: $\bullet$ Suppose $|n|\ge2$. Then $BS(1,n)$ is algebraically rigid if and only if
2491: $|n|$ is prime [\Ler], while
2492: $\Out(BS(1,n))$ is virtually $\Z^k$ (i.e\. finite) if and only if $|n|$ is a prime
2493: power [\Coll].
2494: \endexample
2495:
2496: \demo{Proof} The equivalence $(1)\Leftrightarrow(4)$ is in [\Ler], and
2497: $(1)\Rightarrow(2)\Rightarrow (3)$ follows from Sections 3 and 5. We prove
2498: $(3)\Rightarrow (4)$.
2499:
2500:
2501:
2502: Suppose
2503: that $\Out(G)$ is virtually $\Z^k$ (equivalently, $\T(\Gamma )$ has finite
2504: index in $\Out(G)$). Let
2505: $e,f$ be adjacent edges with
2506: $\lambda _f|\lambda _e$. By Theorem \pal{} and Proposition \finig, the edge
2507: $f$ must be a slid $(2,2)$-segment or a $(p,\pm p)$-loop.
2508:
2509: It cannot be a
2510: segment because of Remark \tee: after collapsing $f$, the group $\T(\Theta
2511: )\inc\Out(G)$ would map onto the infinite dihedral group $J=\langle a,b\mid
2512: a^2=b^2=1\rangle$ with the image of
2513: $\T(\Gamma )$ finite, a contradiction.
2514: To prove (4), there remains to show that the basepoint of any $(1,\pm
2515: 1)$-loop has valence 3.
2516:
2517: Let $f$ be a $(1,\varepsilon )$-loop, with $\varepsilon =\pm1$, let $v$ be its
2518: basepoint, let
2519: $e_1,\dots, e_n$ be the oriented edges with origin $v$ (other than $f,\bar f$).
2520: We must show $n=1$.
2521:
2522: First consider the subgroup
2523: $\T_0$ of
2524: $\T=\T(\Gamma )$ generated by the twists $D_{e_i}$ and the twists around
2525: edges with origin other than $v$. The group $\T$ is generated by $\T_0$ and
2526: the twists
2527: $D_f$, $D_{\bar f}$. The only relations involving $D_f$, $D_{\bar f}$ are
2528: $D_f+\varepsilon D_{\bar f}=0$ (edge relation) and $D_f+D_{\bar
2529: f}+\sum_iD_{e_i}=0$ (vertex relation). It follows that $\T_0$ has index at
2530: most 2 in $\T$ if $\varepsilon =-1$, that $\T$ is the direct sum of $\T_0$ and
2531: the infinite cyclic group generated by $D_f$ if $\varepsilon =1$.
2532:
2533: Now let $\Theta $ be the graph of groups obtained by collapsing $f$, and
2534: consider $\T'=\T(\Theta )$. We will see that it is generated by
2535: $\T_0$ together with extra twists $D'_i$ around the edges
2536: $e_i$ (note that edge group centralizers are bigger in $\Theta $ than in
2537: $\Gamma $). To describe $D'_i$ precisely, we distinguish two cases.
2538:
2539:
2540: If $\varepsilon =-1$, the vertex group $H_v$ of $v$ in $\Theta $ is a Klein
2541: bottle group $\langle a,t\mid tat\mi=a\mi\rangle$. Its center is $\langle
2542: t^2\rangle$. The edge groups
2543: $G_{e_i}$ are generated by powers of $a$. Their centralizer in $H_v$ is the
2544: free abelian group generated by $a$ and $t^2$. In this case, $D'_i$ is the twist
2545: by
2546: $t^2$ around $e_i$. The only relation involving $D'_i$ is the vertex relation
2547: $\sum_iD'_i=0$, so $\T'$ is the direct sum of $\T_0$ and $\Z^{n-1}$. As $\T_0$
2548: has index at most 2 in $\T$, we must have $n=1$ since $\T$ has
2549: finite index in
2550: $\Out(G)$.
2551:
2552: If $\varepsilon =1$, we still have $\T'=\T_0\oplus\Z^{n-1}$ (the vertex group
2553: $H_v$ is $\Z^2=\langle a, t\rangle$, and
2554: $D'_i$ is the twist by $t$). There is a natural homomorphism from
2555: $\langle\T,\T'\rangle$ to
2556: $\Out(H_v)$, given by the action on
2557: $H_v$.
2558: The kernel contains $\T'$, but its intersection with $\T$ is $\T_0$ (as $D_f$
2559: acts on $H_v$ by fixing $a$ and mapping $t $ to $at$). Since
2560: $\T$ has finite index in $\Out(G)$, hence in $\langle\T,\T'\rangle$, we deduce
2561: that
2562: $\T_0=\T\cap \T'$ has finite index in
2563: $\T'$, so $n=1$.
2564: \cqfd\enddemo
2565:
2566:
2567:
2568: Combining with Proposition \kk, we obtain:
2569:
2570:
2571: \nom\outfini
2572: \thm{Corollary \sta} Let $\Gamma $ be a reduced labelled graph
2573: representing a group $G$. The group $\Out(G)$ is finite
2574: if and only if one of the following holds:
2575: \roster
2576: \item
2577: $\Gamma
2578: $ is a tree with no divisibility relation;
2579: \item
2580: $\Gamma $ is a graph with first Betti number 1, there is no divisibility
2581: relation, and $G$ has nontrivial modulus;
2582: \item
2583: $\Gamma
2584: $ is obtained from a tree with no divisibility relation by attaching one $(k,-
2585: k)$-loop. If
2586: $k\ge2$, no other index at the attaching point is a multiple or a divisor of
2587: $k$. If
2588: $k=1$, the loop is attached at a terminal vertex.
2589: \cqfd
2590: \endroster
2591: \fthm
2592:
2593:
2594: \subhead On the isomorphism problem
2595: \endsubhead
2596:
2597: Given a labelled graph $\Gamma $, it is easy to decide algorithmically
2598: whether the associated GBS group $G$
2599: is elementary, solvable, unimodular. By Theorems \pal{} and \main, we may
2600: decide whether $\Out(G)$ contains $F_2$ or is virtually nilpotent.
2601:
2602:
2603: The isomorphism problem for GBS groups is the problem of deciding whether
2604: two (reduced) labelled graphs represent isomorphic groups or not. It is
2605: solved for rigid groups (obviously), for groups with no nontrivial
2606: integral modulus [\For], and for $2$-generated groups [\Lrang].
2607:
2608:
2609: \thm{Theorem \sta}
2610: The isomorphism problem is solvable for GBS groups such
2611: that $\Out(G)$ does not contain a non-abelian free group.
2612: \fthm
2613:
2614: \demo{Proof} Let $\Gamma $ be a reduced labelled graph representing
2615: $G$. We assume that $\Out(G)$ does not contain $F_2$, so $\Gamma $ satisfies
2616: all six conditions of Theorem \pal.
2617:
2618:
2619:
2620: We describe three ways of producing new labelled graphs representing $G$
2621: (besides admissible sign changes). The first one changes the graph, the other
2622: two only change the labels.
2623:
2624: \roster
2625: \item
2626: Sliding an edge across a $(2,2)$-segment: it changes the attaching
2627: point of an edge (carrying an even label).
2628: \item If $v$ is the basepoint of a $(1,q)$-loop, one may multiply or divide
2629: by $q$ some other label near $v$, by sliding the corresponding edge around
2630: the loop.
2631:
2632: \item If $v$ is the basepoint of a $(1,q)$-loop, one may multiply or divide
2633: all other labels near $v$ by any number $p$ dividing $q$, by performing an
2634: expansion at $v$ followed by a collapse (see Figure
2635: \fig; this is called an induction move in
2636: [\Ler]).
2637: \endroster
2638:
2639: \midinsert
2640: \centerline
2641: {\includegraphics[scale=.45]
2642: {fig10.eps}}
2643: \captionwidth{220pt}
2644: \botcaption
2645: {Figure \the\figno}
2646: {Induction move. }
2647: \endcaption
2648: \endinsert
2649:
2650:
2651: Consider the set $\G$ consisting of all labelled graphs which may be obtained
2652: from
2653: $\Gamma $ by combining these moves. They are reduced by condition (4) of
2654: Theorem \pal, and it is easy to decide whether a given labelled graph
2655: $\Gamma '$ belongs to $\G$. We now
2656: complete the proof by showing that $\G$ contains all reduced
2657: graphs representing $G$.
2658:
2659: As above, consider the graph of groups $\Theta $ and the
2660: Bass-Serre tree $S $ obtained by collapsing the slid edges of $\Gamma $.
2661: We have seen that $S$ does not depend on the graph
2662: $\Gamma
2663: $ used to construct it (Lemma \indep). It thus suffices to show that the
2664: various ways of blowing up
2665: $S$ into a GBS-tree differ by the moves mentioned above.
2666:
2667: First consider a vertex group $H_v$ obtained by collapsing a $(2,2)$-edge. It
2668: is a Klein bottle group $ \langle a,b\mid a^2=b^2\rangle$. The generator
2669: of any adjacent edge group is conjugate to some $a^i$ or $b^j$. Distinct
2670: powers of
2671: $a$ (resp\. $b$) are not conjugate in $H_v$, while $a^i$ is conjugate to $b^j$
2672: only when $i$ and $j$ are equal and even. It follows that all ways of blowing up
2673: $v$ differ by slides across the $(2,2)$-edge.
2674:
2675: Now consider a vertex group $H_v= \langle a,t\mid tat\mi=a^q\rangle$
2676: obtained by collapsing a $(1,q)$-loop of $\Gamma $. Let $T'$ be another
2677: GBS-tree, associated to a labelled graph $\Gamma '$. As in the proof of
2678: Lemma \inva, we cannot say that $a$ generates a vertex group of $T'$, only
2679: that some $a^i$, with $i$ dividing a power of $q$, does.
2680:
2681: Suppose for a moment
2682: $i=1$. The generator
2683: of any edge group adjacent to $v$ is conjugate to a
2684: power of
2685: $a$. As
2686: $a^m$, $a^n$ are conjugate in $H_v$ only if $\frac mn$ is a power of $q$, the
2687: labelled graphs $\Gamma ,\Gamma '$ differ by moves of type (2) near the
2688: $(1,q)$-loop.
2689:
2690: If $i\neq1$, it is a product of divisors of $q$, and $\Gamma
2691: ,\Gamma '$ differ by moves of type (2) and (3).
2692: \cqfd\enddemo
2693:
2694: \example{Remark} The same technique may be used when
2695: $G$ is represented by a graph $\Gamma $ satisfying the six conditions of
2696: Theorem \pal, but with arbitrary slid segments allowed in condition (1) (not
2697: only
2698: $(2,2)$-segments). Condition (6) must then be rephrased as follows: Let $vw$
2699: be a $(p,q)$-edge. Let $r$ be a label at $v$, and $s$ a label at $w$. If $q|s$ and
2700: $qr|sp$, then $qr=sp$ and the labels are
2701: carried by the same non-oriented edge.
2702: \endexample
2703:
2704:
2705:
2706: \bigskip
2707: \Refs
2708: \widestnumber\no{99}
2709: \refno=0
2710:
2711:
2712: \bref \by
2713: H. Bass, R. Kulkarni\paper
2714: Uniform tree lattices \jour
2715: J. Am. Math. Soc. \vol3\pages843--902 \yr1990
2716: \endref
2717:
2718: \bref\by G. Baumslag, D. Solitar\paper Some two-generator, one-relator
2719: non-Hopfian groups\jour Bull. Amer. Math. Soc. \vol68\yr1962\pages
2720: 199--201\endref
2721:
2722: \bref\by M. Clay \paper Contractibility of deformation spaces
2723: \jour Alg. Geom. Topol. (to appear) \endref
2724:
2725: \bref\by M. Clay \paper A generalization of Culler's theorem \jour
2726: arXiv:math.GR/0502248\endref
2727:
2728: \bref \by D.J. Collins \paper The automorphism towers of some
2729: one-relator groups\jour Arch. Math.\yr1983\vol40\pages385--400
2730: \endref
2731:
2732: \bref \by D.J. Collins, F. Levin\paper Automorphisms and hopficity of
2733: certain Baumslag-Solitar groups\jour Arch.
2734: Math.\yr1983\vol40\pages385--400
2735: \endref
2736:
2737: \bref \by M. Culler, K. Vogtmann \paper Moduli of graphs and
2738: automorphisms of free groups\jour Invent. Math. \vol 84
2739: \yr1986\pages91--119 \endref
2740:
2741: \bref\by A. Fel'shtyn \paper
2742: The Reidemeister number of any automorphism of a Gromov
2743: hyperbolic group is infinite
2744: \jour J. Math. Sci.\vol119\pages117-123\yr2004
2745: \endref
2746:
2747: \bref\by A. Fel'shtyn, D.L. Gon\c calves \paper Twisted conjugacy
2748: classes of automorphisms of Baumslag-Solitar groups
2749: \jour arXiv:math.GR/0405590\endref
2750:
2751: \bref \by M. Forester\paper Deformation and rigidity of
2752: simplicial group actions on trees\jour Geom.
2753: \& Topol.\vol 6\yr2002\pages 219--267\endref
2754:
2755: \bref \by M. Forester\paper On uniqueness of JSJ
2756: decompositions of finitely generated groups\jour Comm. Math.
2757: Helv. \vol78\yr2003\pages 740--751 \endref
2758:
2759: \bref \by M. Forester\paper Splittings of generalized
2760: Baumslag-Solitar groups \jour arXiv:math.GR/0502060
2761: \endref
2762:
2763: \bref \by N.D. Gilbert, J. Howie, V. Metaftsis, E. Raptis\paper Tree
2764: actions of automorphism groups\jour J. Group
2765: Theory\yr2000\vol3\pages213--223
2766: \endref
2767:
2768: \bref \by V. Guirardel\paper A very short proof of Forester's
2769: rigidity result\jour Geom. \& Topol.\vol 7\yr2003\pages
2770: 321--328\endref
2771:
2772: \bref \by V. Guirardel, G. Levitt \paper The outer space of a free
2773: product \jour arXiv:math.GR/0501288\endref
2774:
2775: \bref \by V. Guirardel, G. Levitt \jour in preparation\endref
2776:
2777: \bref\by P.H. Kropholler\paper A note on centrality in $3$-manifold
2778: groups\jour Math. Proc. Camb. Phil. Soc.\yr1990\vol107\pages 261--266
2779: \endref
2780:
2781: \bref\by P.H. Kropholler\paper Baumslag-Solitar groups and some other
2782: groups of cohomological dimension two\jour Comm. Math. Helv.
2783: \yr1990\vol65\pages 547--558
2784: \endref
2785:
2786: \bref \by
2787: S. Krsti\'c, K. Vogtmann\paper Equivariant outer space and
2788: automorphisms of free-by-finite groups\jour Comm. Math. Helv.
2789: \vol 68 \yr1993\pages 216--262
2790: \endref
2791:
2792: \bref\by G. Levitt \paper Automorphisms of hyperbolic
2793: groups and graphs of groups \jour Geom.
2794: Dedic. \vol 114\yr2005
2795: \pages 49--70
2796: \endref
2797:
2798: \bref\by G. Levitt \paper Characterizing rigid simplicial actions on
2799: trees
2800: \inbook Geometric methods in group
2801: theory, Contemp. Math. \vol372\yr2005\pages27--33
2802: \endref
2803:
2804: \bref\by G. Levitt\jour in preparation
2805: \endref
2806:
2807: \bref\by G. Levitt, M. Lustig\paper Most automorphisms of a
2808: hyperbolic group have
2809: very simple dynamics \jour Ann. Sc. ENS\vol33\yr2000
2810: \pages507--517\endref
2811:
2812: \bref \by D. McCullough, A. Miller \paper Symmetric automorphisms
2813: of free products \jour Mem. Amer. Math. Soc.
2814: \vol122\yr1996\endref
2815:
2816: \bref
2817: \by D.I. Moldavanskij\paper
2818: Isomorphism of the Baumslag-Solitar groups
2819: \jour Ukr. Mat. Zh. \vol 43\pages1684-1686 \yr1991
2820: \transl\nofrills \jour Ukr. Math. J. \vol 43\pages 1569-1571 \yr1991
2821: \endref
2822:
2823: \bref \by L. Mosher, M. Sageev, K. Whyte\paper Quasi-actions on trees,
2824: I. Bounded valence\jour Ann. Math\yr2003\vol158\pages115--164
2825: \endref
2826:
2827: \bref \by M.R. Pettet\paper Virtually free groups with finitely many
2828: outer automorphisms\jour Trans. AMS
2829: \vol349
2830: \yr1997\pages4565--4587
2831: \endref
2832:
2833: \bref \by M.R. Pettet\paper The automorphism group of a graph
2834: product of groups\jour Comm. Alg.
2835: \vol27(10)
2836: \yr1999\pages4691--4708
2837: \endref
2838:
2839: \bref \by E. Raptis, D. Varsos\paper On the automorphism group of the
2840: fundamental group of a graph of polycyclic groups\jour Algebra
2841: Colloquium\yr1997\vol4\pages241--248
2842: \endref
2843:
2844: \bref \by A. Rhemtulla, D. Rolfsen \paper
2845: Local indicability in ordered groups: braids and elementary amenable groups
2846: \jour Proc. Am. Math. Soc. \vol130\pages 2569-2577 \yr2002
2847: \endref
2848:
2849: \bref\by K. Whyte\paper The large scale geometry of the higher
2850: Baumslag-Solitar groups\jour Geom. Funct. Anal. \yr
2851: 2001\vol11\pages 1327--1343\endref
2852:
2853: \endRefs
2854:
2855: \address LMNO, umr cnrs 6139, BP 5186, Universit\'e de Caen,
2856: 14032 Caen Cedex, France.\endaddress\email
2857: levitt\@math.unicaen.fr
2858: \endemail
2859:
2860: \enddocument