1: \documentclass{amsart}

2: %\documentclass{article}

3: %\documentclass{article}

4: \usepackage{amsfonts}

5: \usepackage[all]{xy}

6: \usepackage{amssymb, amscd}

7: 

8: \usepackage{amssymb}

9: \usepackage{graphicx}

10: \usepackage{epstopdf}

11: 

12: 

13: \usepackage[dvips,bookmarks=false]{hyperref}

14: 

15: \newcommand{\dsty}{\displaystyle}

16: 

17: 

18: %\usepackage[dvips]{graphicx}

19: %*********** myfullpagestyle ********************

20: %\textwidth=460pt\textheight=656pt\oddsidemargin=0pt\evensidemargin=0pt%%%\topmargin=0pt\headheight=0pt\headsep=0pt

21: 

22: %******************** \newtheorem

23: \newtheorem{theorem}{Theorem}

24: \newtheorem{lemma}{Lemma}

25: \newtheorem{proposition}{Proposition}

26: \newtheorem{corollary}{Corollary}

27: \newtheorem{definition}{Definition}

28: 

29: %******************** \Bbb

30: \newcommand{\R}{{\Bbb R}}

31: \newcommand{\C}{{\Bbb C}}

32: \renewcommand{\H}{{\Bbb H}}

33: \renewcommand{\O}{{\Bbb O}}

34: \newcommand{\Z}{{\Bbb Z}}

35: \newcommand{\vo}{\overrightarrow{\omega}}

36: 

37: 

38: %******************** \mathcal

39: \newcommand{\D}{\mathcal D}

40: \newcommand{\dperp}{\mathcal D ^{2 \perp}}

41: 

42: %\newcommand(\dtp}{{\mathcal D}^{2 \perp}}

43: \newcommand{\I}{{\mathcal I}}

44: \newcommand{\F}{{\mathcal F}}

45: \newcommand{\tF}{\widetilde{\mathcal F}}

46: 

47: 

48: %******************** \mathfrak

49: \renewcommand{\t}{\mathfrak{t}}

50: \newcommand{\h}{\mathfrak{h}}

51: \renewcommand{\l}{\mathfrak{l}}

52: \newcommand{\s}{\mathfrak{s}}

53: \renewcommand{\k}{\mathfrak{K}}

54: \newcommand{\hp}{{\h}^\perp}

55: \newcommand{\g}{\mathfrak{g}}

56: \newcommand{\p}{\mathfrak{p}}

57: \newcommand{\gl}{\mathfrak{gl}}

58: \renewcommand{\sp}{\mathfrak{sp}}

59: \newcommand{\m}{\mathfrak{m}}

60: %\newcommand{\n}{\mathfrak{n}} oopes used twice

61: \newcommand{\so}{\mathfrak{so}}

62: \renewcommand{\sl}{\mathfrak{sl}}

63: 

64: %******************** \rm

65: \newcommand{\End}{{\rm End}}

66: \newcommand{\GL}{{\rm GL}}

67: \newcommand{\Hom}{{\rm Hom}}

68: \newcommand{\Iso}{{\rm Iso}}

69: \newcommand{\Aut}{{\rm Aut}}

70: \newcommand{\Sp}{{\rm Sp}}

71: \newcommand{\SO}{{\rm SO}}

72: \newcommand{\SL}{{\rm SL}}

73: \newcommand{\TS}[1]{H^{2,0}(\g_{#1})}

74: \newcommand{\TSS}{H^{2,0}(\g)}

75: \newcommand{\W}{{\rm\bf W}}

76: \newcommand{\V}{{\rm\bf V}}

77: \newcommand{\E}{{\rm\bf E}}

78: \newcommand{\SU}{{\rm SU}}

79: \newcommand{\x}{{\bf x}}

80: \newcommand{\y}{{\bf y}}

81: \renewcommand{\P}{{\bf P}}

82: \newcommand{\bp}{{\bf p}}

83: 

84: \renewcommand{\c}{ {\bf c}}

85: 

86: 

87: \newcommand{\by}{{\bf y}}

88: \newcommand{\Y}{{\bf Y}}

89: 

90: 

91: %\renewcommand{\S}{{\rm\bf S}}

92: 

93: %******************** Misc

94: \renewcommand{\L}{{\mathcal L}}

95: \newcommand{\marco}[1]{\framebox{$\displaystyle #1 $}}

96: \newcommand{\bs}{\bigskip}

97: \newcommand{\ms}{\medskip}

98: \newcommand{\mn}{\medskip\noindent}

99: \newcommand{\n}{\noindent}

100: \newcommand{\bn}{\bs\n}

101: \newcommand{\bnbf}{\bs\n\bf}

102: \newcommand{\tensor}{\otimes}

103: \newcommand{\w}{\wedge}

104: \newcommand{\X}{\times}

105: \renewcommand{\o}{{\omega}}

106: \newcommand{\ov}{\overrightarrow{\omega}}

107: \newcommand{\pf}{\n\textsc{Proof. }}

108: %\newcommand{\qed}{ $\Box$}

109: \newcommand{\e}{ {\bf e}}

110: \renewcommand{\a}{ {\bf a}}

111: \renewcommand{\b}{ {\bf b}}

112: \newcommand{\tQ}{\widetilde {Q}}

113: \newcommand{\tD}{\widetilde {D}}

114: \newcommand{\tC}{\widetilde {C}}

115: 

116: 

117: %\usepackage{pdfsync}

118: %package which lets you jump back and forth from pdf and tex files easily

119: %getting to the corresponding  place within the tex file

120: 

121: \title{$G_2$ and the ``rolling distribution''}

122: 

123: \author{ Gil Bor and Richard Montgomery}

124: 

125: %\input{pdfsync}

126: 

127: \begin{document}

128: \maketitle

129: 

130: \tableofcontents

131: 

132: \section*{0 Introduction}

133: 

134:  

135: \begin{figure}[h]

136: \centerline{\includegraphics[width=8cm, angle=90]{roll2.eps}} 

137: \caption{Rolling a ball on another ball.}\label{roll}

138: \end{figure}

139: 

140:  

141: 

142: 

143: Consider two balls of different sizes,

144: rolling on each other, without slipping or spinning. 

145: % It may help to think of the balls  as covered in Velcro.  

146: The

147: configuration space for this system is a 5-dimensional manifold

148:  $Q \cong \SO_3\times

149: S^2$ on which  the no-slip/no-spin condition defines a rank 2

150: distribution $D\subset TQ$, the ``rolling-distribution''.

151: 

152: Now $D$  is a non-integrable distribution (unless the balls are of equal size) which has an ``obvious'' 6-dimensional transitive symmetry group 

153: $\SO_3\times\SO_3$ arising from the isometry groups of each ball, but for balls whose radii are in the ratio 3:1, and only for this ratio, 

154: something strange happens: the {\it local } symmetry group 

155: of  the distribution  increases  from  $\SO_3\times\SO_3$  to $G_2$,  a

156: 14-dimensional Lie group. 

157: 

158: More  precisely, let $\g_2$ be the  real  ``split

159: form"   of the complex $14$-dimensional exceptional simple Lie algebra $\g_{2 \C}$.

160: There are precisely two connected Lie groups whose Lie algebras

161: are $\g_2$.  (See Appendix  \ref{twoG2s}.) We choose the one corresponding to the adjoint

162: representation, and call it $G_2$.  (The other one is $\tilde G_2$,

163: the universal cover of the one we chose.)

164: $G_2$ is  a subgroup of $\SO(3,4)$ and its maximal compact subgroup  $K\subset G_2$ is isomorphic to $\SU_2\times \SU_2/\{\pm(1,1)\}$ which   double-covers  $\SO_3\times\SO_3$. 

165: Let $\tQ=S^3\times S^2$ be 

166: the universal  cover of $Q$ equipped with the distribution $\tD$ induced by the double covering $\tQ  \to Q$.

167: Let $\Aut(\tQ,\tD)$ be the group of diffeomorphisms of  $\tQ$ leaving   $\tD$ invariant. 

168: Then we have 

169: 

170: \begin{theorem}  The connected component of the identity in $\Aut(\tQ,\tD)$ for radius ratio 3:1 or 1:3 is isomorphic to $G_2$. The $G_2$ action on $\tQ$ 

171: does not descend to $Q$, but its   restriction to the maximal compact $K \subset G_2$ does, covering the   $\SO_3 \times \SO_3$ action on $Q$. For any other radius ratio (other then 1:1) $\Aut(\tQ,\tD)$ is ismorphic to $K$.  

172: \end{theorem}

173: 

174: This theorem was communicated to us by  Robert Bryant, for whom it is but a  variation on a theme of E. Cartan's work on the method of equivalence, contained in his notoriously difficult  ``Five Variables Paper" \cite{C1}  from 1910 .   Bryant  wrote to us recently: 

175: \begin{quote}

176: ``Cartan himself  gave a geometric description of the flat

177: $G_2$-structure as the differential system that describes space

178: curves of constant torsion $2$ or $1/2$ in the standard unit 3-sphere.

179: (See the concluding remarks of Section 53 in Paragraph XI in the

180: Five Variables Paper.)  One can easily transform the rolling balls

181: problem (for arbitrary ratios of radii) into the problem of curves

182: in the 3-sphere of constant torsion and, in this guise, one can

183: recover the 3:1 or 1:3 ratio as Cartan's torsion 2 or 1/2 with

184: a minimum of fuss.  Thus, one could say that Cartan's calculation

185: essentially covers the rolling ball case.'' 

186: \end{quote}

187: 

188: \bn\centerline{$*\qquad*\qquad*$}

189: 

190: \bn 

191: 

192: Our main purpose  in this note is to try to explain  this beautiful and mysterious theorem 

193: in a direct manner which does not appeal  to   Cartan's method of

194: equivalence. We consider it an  expansion  of Section 4 in  

195: Bryant's lecture notes \cite{Bryant}.  Our contribution consists basically of a  description of two constructions of  $(\tQ, \tD)$ with 

196: a built-in  $G_2$-invariance.   Using these constructions we show here   that, for radius ratio 3:1 or 1:3,  $G_2$ is {\em contained} in $Aut(\tilde Q, \tilde D)$.  But

197: we do not know how to show, without the more sophisticated  Cartan's methods 

198: (or its variants such as those of Tanaka)

199:  that  $G_2$ is the   {\em full} identity component

200: of $Aut(\tilde Q, \tilde D)$, 

201: nor that for radius ratio different from 3:1, 1:3 or 1:1, $Aut(\tilde Q, \tilde D)$ is not larger than $K$. 

202: 

203: 

204: A secondary purpose  of this article is to correct an error appearing in the book

205: \cite{Montbook} by one of us. We had mistakenly said there  that   the symmetry group for the rolling distribution for a ball on a plane   (ratio $1:\infty$) was $G_2$.  

206: 

207: A tertiary purpose is to obtain a bit of a feel for the simplest

208:  exceptional Lie algebra $\g_2$ and  its Lie groups, and 

209:  to provide  a refresher course on   roots and weights.

210: 

211: 

212: \bn {\bf Structure of Paper.} 

213: In the next section (section 1) we describe the background and wider context  of the problem,

214: with references to the literature.   In section 2 we give a detailed   description 

215: of the distributions associated with the rolling of balls, noting their $\SO_3\times\SO_3$-symmetries. 

216: In section 3 we describe  the homogeneous distributions of a Lie group $G$

217: in terms of data   $(G, H, W)$, where $H\subset G$ is a closed subgroup and $W\subset\g/\h$ is an $H$-invariant subspace. 

218: We then  identify this data for the rolling distribution with respect  the group $G=\SO_3\times\SO_3$.  In section 4 we use the root diagram of $G_2$ to give our first construction of a $G_2$-invariant distribution data $(G_2, P, W)$. The identification of the resulting $G_2$-homogenous distribution on $G_2/P$ with $(\tilde Q, \tilde D)$

219:  amounts to the embedding of  

220:  $\so_3\times\so_3$ in $\g_2$ and is   the subject of section 5 (and Appendix B) which forms the heart

221:  of this article.   

222: In section 6 we give the second $G_2$-invariant construction of $(\tilde Q, \tilde D)$ an

223: explicit construction applying   projective geometry to the space of

224: purely imaginary  split octonions $V$, the lowest dimensional non-trivial representation space

225: for $G_2$.  Appendix C is historical.  Following suggestions by Bryant

226: we looked into Cartan's thesis and found that much of content of section 6,

227: and hence of the rolling distribution already appears there. 

228: 

229: 

230: 

231: 

232: 

233: \bn\centerline{$*\qquad*\qquad*$}

234: 

235: \bn

236: 

237: 

238: Despite all our efforts,  the   ``$3$'' of the ratio $1:3$ remains    

239:   mysterious.  In this article it  simply arises  out of the structure constants for 

240:    $G_2$ and   appears  in the construction

241: of the embedding of $\so_3\times\so_3$ into $\g_2$ (section 5 and Appendix B).

242:  Algebraically speaking, this  `3'   traces   back   to the 3 edges

243:  in $\g_2$'s  Dynkin diagram

244:  and the consequent relative positions of   the long and short

245:  roots in the root diagram (see figure 2 below) for  $\g_2$ which the Dynkin diagram is encoding.   

246: 

247:  \bn {\bf Open problem.} Find a geometric or dynamical 

248: interpretation for the ``$3$'' of the $3:1$ ratio.

249: 

250: \bn 

251: 

252: For work in this direction see Agrachev \cite{Agrachev} and also Kaplan and Levstein

253: \cite{Kaplan}. 

254: 

255: 

256: 

257:  

258:  \bn {\bf Acknowledgements.} 

259:   Robert 

260: Bryant has   been crucial, at various key steps along the way,

261: in steering us in the right direction.  Marti Weissmann

262: supplied us with    key information regarding $G_2$, and the 

263: crucial Vogan reference.  

264: 

265:  \section{History and Background}

266: 

267: {\bf On distributions.}  By a distribution we mean here a linear subbundle of 

268: the tangent bundle of a manifold. 

269: The distributions    first encountered    are  usually

270: the integrable and the contact distributions and  have infinite dimensional symmetry groups.   In dimension 

271: $5$  we first encounter distributions whose symmetry groups  

272:  are  finite-dimensional. Indeed, the generic  distribution of rank $2$ or $3$  in $5$ dimension  has  {\em no} local symmetries. 

273:  Cartan  \cite{C1} investigated  rank $2$ and

274:  $3$ distributions in $5$ dimensions in   detail.  

275:  The growth vector of 

276:   a generic rank $2$ distribution on a $5$-dimensional manifold,  at a generic

277:   point of that manifold,   is $(2, 3, 5)$.  

278:   This  ``$(2,3,5)$  at a point''

279:     means that   if $X, Y$ are any local vector fields spanning

280:   the distribution in a neighbhorhood of the point , then $[X,Y] = Z$

281:   is pointwise linearly independent of $X, Y$ (in a neighborhood of  the point) 

282:   and $X, Y, Z, [X, Z], [Y, Z]$ span the tangent bundle in a neighborhood of  the point.

283:   Cartan  worked out the  

284: complete   local  invariants -- analogues of the Riemann curvature

285: tensor -- for these $(2,3,5)$ distributions. 

286: For the distribution's symmetry group to act transitively   all of

287: Cartan's  invariants must be constant.  To get the maximal dimensional  symmetry group 

288: all Cartan's  invariants must vanish, in which case we call 

289: the distribution ``flat''.   Any such distribution

290: is locally   diffeomorphic to  that of the  ``Carnot group'' distribution

291: associated to the unique graded nilpotent Lie group $\mathfrak{n} = \mathfrak{n}_{2,3,5}$ of this same  growth, and  its  local  symmetry

292: algebra   is $\g_2$.

293: (By  the ``local symmetry algebra'' of a distribution we mean the algebra of vector fields

294: $X$ satisfying $[X, \Gamma(D)] \subset \Gamma(D)$ where $\Gamma(D)$ is the sheaf of local sections of vector fields tangent to the distribution.)

295:  

296: As mentioned in the above quote from Bryant,  Cartan \cite{C1} 

297: presented several geometric realizations of the flat case.   Bryant  and Hsu 

298: \cite{BryantHsu}  (see section 3.4) pointed out the

299:  rolling incarnation of $G_2$.  A $(2,3,5)$

300:  distribution will    arise whenever  one  rolls one Riemannian surface on  another 

301:  provided their Gaussian curvatures are not equal.    

302:  The Cartan invariants vanish if and only if  the ratio of their    curvatures are  $1:9$ .

303:   Hence the $1:3$ radii for spheres. We could also achieve the maximal

304:   local symmetry algebra $\g_2$ by 

305:  rolling two hyperbolic planes along  each other, provided their ``radii''

306:  are in the ratio $i: 3i$.    

307: More history, and more instances of the flat $G_2$ system

308: are  explained in Byrant \cite{Bryant}.   

309:  

310: 

311: Non-integrable rank $2$ distributions in dimension $n$ ($n > 3$)  admit special families

312: of integral curves known as ``singular'' or ``abnormal'')

313: (\cite{Montbook}).  These are curves which

314: admit no local variations through integral curves and having endpoints fixed.    In the case of $(2,3,5)$ distributions

315: there is precisely $1$ singular curve (up to reparameterization)

316: through every point  {\it  in every direction} tangent to $D$. In the case of

317: rolling one Riemannian surface along another, these singular curves correspond

318: to rolling along geodesics.   Using  the symplectic geometry associated to

319: variations of    singular curves

320:  Zelenko and Agrachev have been  able to 

321:  rederive  Cartan's (2,3,5) invariants. See   \cite{Agrachev} and references therein. 

322: 

323: Tanaka and his school have established a wonderful

324: generalization of the passage from the flat nilpotent model $\mathfrak{n}_{2,3,5}$

325: to $\g_2$.      

326: Associated to each point $p$ of a manifold endowed with a non-integrable

327: distribution there is a graded  nilpotent Lie algebra $\m = \m (p)$ called  the

328: `nilpotentization' of the distribution, or sometimes the ``symbol algebra''.  The dimension of $\m$ is that of the underlying manifold.   

329: Call the distribution ``of type $\m$'' if the different algebras $\m(p)$

330: are all isomorphic to the same $\m$, i.e. the isomorphism type does not change

331: from point to point.  

332: (Every   (2,3,5) distribution is of type $\mathfrak{n}_{2,3,5}$.)   Associated to each

333: graded nilpotent $\m$ there is graded Lie algebra $\g \supset \m$, possibly infinite dimensional, 

334: called the `prolongation' of $\m$ and built from  $\m$ in  a purely algebraic manner.

335: This  $\g$ represents, roughly speaking,  

336: the maximal possible symmetry  of a distribution of type $\m$:

337: every symmetry algebra for a type $\m$-distribution, after applying a 

338: grading process to it,  

339: must be a subalgebra of $\g$.  The prolongation of

340: the (2,3,5) algebra is $\g_2$, and this fact can be viewed 

341: as the algebraic restatement of Cartan's work on the flat model. 

342: This Tanaka prolongation method thus yields a proof that $Aut(\tilde Q, \tilde D)

343: \subset G_2$ in theorem 1, alternative to Cartan's proof. 

344:   Yamaguchi \cite{Yamaguchi} has classified all $\m$'s whose $\g$'s are simple. 

345:   To each of these pairs $(\m, \g)$ is associated an intricate differential geometry

346:   and most of these have not been explored in any detail.  

347:   

348: 

349: 

350:  

351: 

352: \bn{\bf On $G_2$.}  The Lie algebra $\g_2$  is

353: the smallest   of the  exceptional simple Lie algebras.

354: In 1894 Killing uncovered

355: the   existence of the root lattice for   $\g_2$'s,

356: but without establishing the existence of the

357: corresponding Lie algebra.   Cartan, in his thesis,   established the  existence  of $\g_2$

358: in one page of  his thesis  \cite{C2}.  He did so by constructing

359: the $7$-dimensional representation of $\g_2$, in a way which is   closely related to our second ``projective split octonion'' 

360: model for $\tilde Q$, the universal cover of the rolling space.

361: We have devoted   appendix C to this page of his thesis

362: and its connection with this  second model.  

363:   In 1914

364: Cartan \cite{C3} showed that $G_2$ can be realized as the automorphism

365: group of the octonions.  For our split $G_2$ he used `split

366: octonions'.    The compact form of $G_2$ 

367:   appears  in the Berger list of

368: potential holonomy groups of Riemannian metrics. Recently,  the compact  $G_2$ has been 

369: featured in   string theories, but perhaps that fad has passed already.  

370:  

371: 

372: 

373: \section{Distribution for rolling balls}

374: 

375: \subsection{The distribution}

376: 

377:  \bn

378: 

379: Take the first  ball to be stationary, of radius $R$, with its center at the

380: origin. Roll  a second ball  of radius $r$ on the first ball. The position of the

381: second ball is given by an isometry (rigid motion)

382: $\varphi_{(g,\x)}:\R^3\to\R^3$, mapping a point $\P$ to

383: $$\bp = \varphi_{(g,\x)}(\P)=g\P+(R+r)\x,$$ where $(g,\x)\in\SO_3\times

384: S^2$.  Here, $R\x$ is the point of contact of the two balls,

385: $(R+r)\x$ is the center of the second ball and $g\in\SO_3$ describes the

386: rotation of the second ball relative to its initial position.  See figure 1 in the introduction. Thus the configuration

387: space $Q$ for our rolling problem has been identified with the manifold

388: $\SO_3\times S^2$.  (For a visceral account of rolling a sphere on a plane,

389: accessible to upper division undergraduates,  we recommend

390: \cite{Hammersley}.)

391:  

392: 

393:  

394: %[PICTURE MISSING]

395: 

396: %\includegraphics[width=10cm, angle=90]{roll2.eps}

397: %Figure 1. 

398: %\end{picture}

399: %\end{center}

400: \bn

401: 

402: 

403: 

404: 

405: Let $(g_t,\x_t)\in Q$ be a  differentiable rolling motion.  Let $\o_t\in

406: \R^3\cong\so_3$ be the angular velocity of the rolling ball relative to

407: its center, measured with respect to 

408: inertial axes.  In other words,  if $\P$  is a material point fixed 

409: on the second ball, $\dot\P=0$, and if

410: we write  $\bp_t=g_t\P$,  then $\dot \bp=\dot gg^{-1}\bp=\omega\times

411: \bp$.  Then we have

412: 

413: 

414: \begin{proposition} Let $Q=\SO_3\times S^2$ be the configuration space of two rolling balls of 

415: radii $R$ and $r$. Let $\rho=R/r$ . 

416: Then a curve  $(g_t,\x_t)\in Q$ describes a rolling motion without slipping and spinning iff

417: \mn \begin{quote}

418: (1) $(\rho+1)\dot \x=\omega\times \x$ (no-slip condition),

419: 

420: (2) $\langle\omega,\x\rangle=0$ (no-spin condition,

421:  i.e. $\omega$ need to be tangent  to the stationary ball at $R\x$).

422: \end{quote}

423: \end{proposition}

424: 

425: 

426: \pf (1) The contact point between the two balls is $\bp=R\x$ on the

427: first ball, $\P=-g^{-1}r\x$ with respect to the second ball. For

428: non-slip, their velocities must match: $\dot\bp= g\dot\P.$ Now

429: $\dot\bp=R\dot\x$ and

430: $$\dot\P= [-{d\over dt}g^{-1}]r\x-g^{-1}r\dot\x=g^{-1}\dot gg^{-1}r\x-g^{-1}r\dot\x=g^{-1}r(\o\X\x-\dot\x),$$

431: hence the non-slip condition $\dot\bp= g\dot\P$ is equivalent to $R\dot\x=r(\o\X\x-\dot\x),

432: $

433: from which (1) follows.

434: 

435: \mn (2) Let $\bf P$ be a material point {\em fixed} on the second ball

436: ($\dot\P=0$). From the inertial point of view, which is to say,

437: from the point of view  of the first ball with origin at its center,

438: the position of this material point is  $\bp=g\P+(R+r)\x$, and

439: so its  velocity

440: $$\dot\bp=\dot g\P+(R+r)\dot\x=\dot g

441: g^{-1}[\bp-(R+r)\x]+(R+r)\dot\x=\o\X[\bp-(R+r)\x]+(R+r)\dot\x.$$ Using

442: the no-slip equation, $(R+r)\dot \x=r\omega\times \x$, we get

443: $$\dot\bp=\o\X[\bp-(R+r)\x]+r\o\X\x=\o\X(\bp-R\x).$$ 

444: The equation

445: $\dot\bp=\o\X(\bp-R\x)$ asserts  that the instantaneous motion of the second ball is

446:  a rotation whose

447: axis of rotation (a line) passes through the point of contact $R\x$,

448: in the direction of $\o$ and with angular velocity of magnitude

449: $\|\o\|$. The no-spin condition is 

450: that the second ball does not spin about

451: the   point of contact of the two balls, which is to say that $\o$ should have no

452: component orthogonal to the common tangent plane of the two balls,

453: i.e. $\langle\omega,\x\rangle=0$. \qed

454: 

455: The two conditions in the last Proposition define together a rank 2 distribution on $Q$. This is the rolling distribution.

456: 

457: \vskip .3cm

458:  

459: 

460: \subsection{The ``obvious'' symmetry}

461: 

462: The group $\SO_3\times\SO_3$ acts  on $Q$ by

463: $\varphi_{(g,\x)}\mapsto g'\circ\varphi_{(g,\x)}\circ g''^{-1},$ where

464: $g',g''\in\SO_3.$ In terms of $(g,\x)$ this 

465: action is

466: $$(g,\x)\mapsto (g' g g''^{-1},g'\x), \quad g',g''\in\SO_3.$$

467: This action is transitive and preserves the rolling distribution $D$ for any value of $\rho=R/r$.  The proofs of these assertions are easy and left as exercises.  

468: 

469: 

470: \section{Group theoretic description of the rolling distribution

471: }

472: 

473: \bn

474: 

475: In the previous section we wrote down a distribution $D$ on $Q=\SO_3\times S^2$, depending on the real parameter $\rho$.  We showed

476: that $Q$ admits   an $\SO_3\times\SO_3$-transitive action which preserves the distribution. Our aim in this paper is to show that for two specific values of the parameter, $\rho=3$ and $\rho=1/3$, the distribution admits a larger {\em local} group of symmetries, namely the group $G_2$.  We do so by defining a $G_2$-homogeneous space $\tQ=G_2/P$, together with a $G_2$-invariant rank 2 distribution $\tD$ on it. 

477: We then define a 2:1 covering map $\tQ\to Q$ which maps $\tD$ to $D$. Furthermore, the group $G_2$ contains a maximal compact  subgroup $K\subset G_2$ which is a double cover of $\SO_3\times\SO_3$, such that the map $\tQ\to Q$ is $K$-equivariant  (with respect to the covering homomorphism $K\to\SO_3\times\SO_3$). 

478: The  constructions 

479: are most easily done on the group level or on the Lie algebra level. We describe in what follows the general set up required for ``working on the group level'' and 

480: then calculate the group theoretic data corresponding to the rolling distribution. 

481: 

482: \bn 

483: 

484: Let $G$ be a Lie group. A ``$G$-homogeneous distribution'' is a pair  $(Q,D)$ where  $Q$  is a manifold on which $G$ acts  transitively  and  $D\subset TQ$ is  a $G$-invariant distribution. Fixing  a base point 

485:  $q_0\in Q$ with isotropy  $H\subset G$ we obtain a $G-$equivariant identification    $ Q\cong G/H$  and an $H$-equivariant identification $T_{q_0}Q \cong \g/\h$

486:  where $\h \subset \g$ denote   the Lie algebras corresponding to $H \subset G$.

487:  Then  $D_{q_0}\subset T_{q_0}Q$ corresponds to an $H$-invariant subspace $W\subset\g/\h$. 

488:  In this way  every  $G$-homogeneous distribution $(Q,D)$ corresponds to data 

489:  $(G,H,W)$, where  

490: $H\subset G$ is a closed subgroup with Lie algebra $\h\subset\g$ and 

491: $W\subset\g/\h$ is an $H$-invariant subspace.

492:  The adjoint action of $G$ defines an equivalence relation on the set of pairs  $(H,W)$ so that 

493: different choices of base points on $Q$ correspond to equivalent pairs  $(H,W)\sim(H',W')$. 

494:  Conversely,  given the data $(G,H,W)$, we can construct a $G$-homogeneous distribution $(Q,D)$ by letting $G$  act by left translations on the

495:  right $H$-coset space $Q:=G/H$, and define a $G$-invariant distribution

496:  $D\subset TQ$  using the $G$-action to  push   $D_{[e]}:=W\subset\g/\h\cong

497:  T_{[e]}(G/H)$ around to all other points of $Q$. 

498:  

499: 

500:  

501: On the level of  Lie algebras, the  data $(\g,\h,W)$ determines $(Q,D)$

502:   up to a cover.  If, as in our case of $\g= \so_3 \oplus \so_3$, the simply

503:   connected Lie group $G$ realizing $\g$ is compact,

504:   then there are only finitely many homogeneous distributions    $(G, H, W)$ which realize the   given  Lie algebraic data $(\g, \h, W)$. 

505: 

506:  

507:  

508: \bn 

509:   

510: We now  determine  the data $(G,H,W)$ corresponding to the  rolling  distribution

511: $(Q,D)$ 

512: of section 2.1.   Here $G=\SO_3\X\SO_3$, $Q=\SO_3\times S^2$, $\dim H=1$, $\dim W =2$. 

513: Identify   the Lie algebra of $\SO_3\times\SO_3$ with $\R^3\times\R^3$, the set of pairs of angular velocities $(\omega', \omega'')$, with 

514: Lie bracket given by the cross product:

515: $$[(\omega', \omega''),(\eta', \eta'')]=(\omega'\X\eta',

516: \omega''\X\eta'').$$  The first factor $\omega '$ corresponds to the

517: first (stationary) sphere, of radius $R$, while the second $\omega ''$ factor corresponds to second (rolling) sphere of   radius $r$. 

518: 

519: Fix a base point, say $(1,\e_3)\in\SO_3\times S^2=Q$.  The

520: isotropy at this base point is the circle  subgroup $H$ consisting of elements of the form $(h,h)$, where $h$ is a

521: rotation around the $\e_3$ axis, so $\h=\R(\e_3,

522: \e_3)\subset\R^3\X\R^3.$ Using the Killing metric on $\g=\so_3\X\so_3=\R^3\X\R^3$ we can

523: identify $\g/\h\cong\h^\perp$, so that the plane of the distribution at the

524: base point is given by some 2-plane in $\h^\perp.$ Let us

525: determine explicitly this 2-plane.

526: 

527:  

528: 

529: \begin{proposition} The rolling distribution  on $\SO_3\times S^2$ 

530: corresponding to rolling a ball of radius $r$ along one of radius $R$  is given by the 2-plane

531: in $\R^3\times\R^3$ (the Lie

532: algebra of $\SO_3\X\SO_3$) defined by the equations

533: $$\langle\omega',\e_3\rangle=\langle\omega'',\e_3\rangle=0, \quad

534: \rho \omega'+\omega''=0,$$

535: where $\rho = R/r$.

536: \end{proposition}

537: 

538: \pf Since $\h\subset\R^3\times\R^3$ is  generated by the vector $(\omega', \omega'')=(\e_3,\e_3)$ and the Killing metric corresponds to some multiple of the standard metric on $\R^3\times\R^3$, $\h^\perp\subset \R^3\times\R^3$  is given by the equation $\langle\omega'+\omega'',\e_3\rangle=0.$

539: 

540: 

541: 

542: 

543: From the formula for the $\SO_3\X\SO_3$-action in \S2.2 we get the

544: infinitesimal action at the base point $$\omega=\omega'-\omega'',

545: \quad\dot \x=\omega'\X \e_3.$$ Substituting these into the

546: rolling  conditions at the base point (see \S2.1),

547: $$\langle\omega,\e_3\rangle=0, \quad (R+r)\dot \x= r\omega\times \e_3,$$

548: we obtain

549: $$\langle\omega'-\omega'',\e_3\rangle=0, \quad  [R\omega'+ r\omega'']\times \e_3=0.$$

550: Adding the condition of orthogonality

551: to $\h$, $\langle\omega'+\omega'',\e_3\rangle=0,$

552: we obtain the above equations.$\Box$

553: 

554: \bn

555: 

556: \subsection{Shrinking the group.} 

557: \label{shrinking} The following observation will be key to proving

558: that part of   theorem 1 which we are going to prove, namely

559: that $G_2 \subset Aut(\tilde Q, \tilde D)$ for $\rho=3$ or $1/3$.

560: Suppose that   $(Q,D)$ is a $G$-homogeneous distribution with $G$-data $(H,W)$. 

561: Let $G_1\subset G$ be a subgroup for which the restriction of the $G$-action on $Q$ to $G_1$ is still transitive. 

562: Then $(Q,D)$ is also $G_1$-homogeneous distribution and its   $G_1$-data is $(H_1,W_1)$  where $H_1=H\cap G_1$ and $W_1\subset \g_1/\h_1$

563: corresponds to 

564: $W$ under the  linear isomorphism   $\g_1/\h_1\to\g/\h$

565: induced by the diffeomorphism $Q = G/H = G_1/ H_1$.

566: Since $(Q, D)$ has not been changed, it follows that

567: the $G$-data $(H, W)$ and the $G_1$-data $(H_1, W_1)$

568: yield diffeomorphic manifolds with distributions. 

569:  At the Lie algebra level, this discussion asserts that

570: $(\g_1, \h_1, W_1)$ and $(\g, \h, W)$ define manifolds-with-distributions

571: which are diffeomorphic up to a cover.  To prove that

572: $G_2 \subset Aut(\tilde Q, \tilde D)$ we  will be applying  this observation to 

573: the case $\g_1 = \so_3 \oplus \so_3 \subset \g_2$.   

574: 

575: 

576: 

577: 

578: \section{A $G_2$-homogeneous distribution }

579: 

580:  

581: 

582: \bn

583: 

584: We now describe the other main actor  in this paper,

585: a distribution with Lie algebraic data $(\g_2, \p, W)$. 

586: Please see  the root diagram of $\g_2$ in figure 2.  This diagram will be explained

587: immediately below.  The decorations on the diagram are used to  indicate

588: the Lie algebraic data and  will   be explained a bit later. 

589: 

590: \bn

591: 

592: \begin{center}

593: \unitlength=1cm

594: \begin{picture}(5,4)(-2,-2)

595: \put(0,0){\circle*{.4}}

596: 

597: 

598: \put(.4,.75){$\oplus$}

599: \put(-.5,-.866){\circle*{.25}}

600: \put(.6,1.2){$\sigma_1$}

601: \put(-1.2,-1.3){$-\sigma_1$}

602: 

603: \put(-.5,.866){\circle*{.25}}

604: \put(.4,-.95){$\oplus$}

605: \put(-.95,1.2){$\sigma_2$}

606: \put(.3,-1.3){$-\sigma_2$}

607: 

608: \put(-1,0){\circle*{.25}}

609: \put(1,0){\circle{.25}}

610: \put(1.3,-0.1){$\sigma_3$}

611: \put(-2,-0.1){$-\sigma_3$}

612: 

613: 

614: \put(1.4,-.866){\circle{.25}}

615: \put(-1.5,.866){\circle*{.25}}

616: \put(1.6,  -1.1){$-\lambda_1$}

617: \put(-2.2,  0.866){$\lambda_1$}

618: 

619: \put(1.4,.866){\circle{.25}}

620: \put(-1.5,-.866){\circle*{.25}}

621: \put(1.7,.9){$\lambda_2$}

622: \put(-2.4,-1.1){$-\lambda_2$}

623: 

624: \put(0,1.5){\circle*{.25}}

625: \put(0,-1.5){\circle*{.25}}

626: \put(-.2,1.8){$\lambda_3$}

627: \put(-0.5,-2.1){$-\lambda_3$}

628: 

629: \put(0,-1.4){\line(0,1){3}}

630: \put(-1.5,  -0.866){\line(5,3){2.77}}

631: \put(-1.5,  0.866){\line(5,-3){2.77}}

632: 

633: \put(  -1, 0){\line(1  , 0   ){1.85}}

634: \put(-.5,-.866){\line(3,5){.96}}

635: \put(-.5,.866){\line(3,-5){.96}}

636: \end{picture}

637: 

638: \bn

639: 

640: {\small Figure 2: The root diagram of $\g_2$.}

641: \end{center}

642: 

643: 

644: \bn{\bf A reminder of the meaning of the root diagram.} The plane in which the  diagram is drawn is  the dual of a Cartan subalgebra $\t\subset\g_2$. A Cartan

645: subalgebra of a semi-simple Lie algebra $\g$ is a maximal abelian

646: subalgebra $\t\subset\g$ of semi-simple elements, i.e. each

647: $ad(T)\in\End(\g)$, $T\in\t$, is diagonalizable.  In the case of

648: $\g=\g_2$, $\t$ is 2-dimensional, hence the subscript 2 in $G_2$, the

649: {\em rank} of the group.  The root diagram of $\g$ encodes the adjoint

650: action of $\t$ on $\g$, from which one can recover the whole structure

651: of $\g$.

652: 

653: The commutativity of the Cartan

654: subalgebra $\t$ implies that the diagonalizable endomorphisms

655: $ad(T)\in\End(\g)$, $T\in\t$, are {\em simultaneously} diagonalizable,

656: resulting in a $\t$-invariant decomposition

657: $$\g=\t\oplus\sum_\alpha\g_\alpha,$$

658: where each  $\g_\alpha\subset\g$ is a 1-dimensional subspace  of $\t$-common eigenvectors called a {\em root space}.

659: The corresponding eigenvalue  depends linearly on

660: the acting element of $\t$, so is given by a linear functional $\alpha\in\t^*$, called {\em root}.  Thus

661: $$[T,X]=\alpha(T)X, \quad T\in\t,\quad X\in\g_\alpha.$$

662: 

663:  When we draw the root diagram in $\t^*$

664: we use the Killing metric in $\g$ to determine the size of the roots

665: and especially  the angles between them. The Killing metric in $\g$ is the inner

666: product $\langle X,Y\rangle=-{\rm tr}(ad(X)ad(Y))$.  It is

667: non-degenerate (this is equivalent to semi-simplicity) and its

668: restriction to $\t$ is positive definite.

669: 

670: 

671: \bn{\bf Example of $\g=\sl_3(\R).$} The more familiar example of

672: $\sl_3 (\R)$ is  useful   to keep

673: in mind before   proceeding with $\g_2$.   The Lie algebra $\sl_3(\R)$ is the 

674: vector space of 

675: of 3 by 3 traceless real matrices with Lie bracket the usual matrix Lie bracket.

676: It is  Lie algebra of the Lie group $\SL_3(\R)$ of 3 by 3 real matrices with determinant $1$.   Like $\g_2$, the Lie

677: algebra $\sl_3(\R)$ is a non-compact split form of its

678: complexification ($\sl_3(\C)$) and has  rank 2. We take as a Cartan

679: subalgebra the subspace  $\t\subset\sl_3(\R)$ of traceless diagonal

680: matrices,

681: $$\t:=\{\left(\begin{array}{ccc} t_1&0&0\\0& t_2&0\\0&0& t_3\end{array}\right) |   t_1+ t_2+ t_3=0, t_i\in\R\}.$$

682: 

683: $\sl_3(\R)$ has    6 roots: 

684: $$\alpha_{ij}:= t_i- t_j\in\t^*, \quad i \neq j, \quad

685: i,j\in\{1,2,3\},$$ with corresponding root spaces

686: $$\g_{\alpha_{ij}}=\R E_{ij}, $$ where $E_{ij}$ is the matrix whose

687: $ij$ entry is 1 and all of whose other   entries are 0. The corresponding root space

688: decomposition $$\sl_3=\t\oplus\sum_{i\neq j}\g_{\alpha_{ij}},$$ is

689: just the decomposition of a matrix as a diagonal matrix plus its off

690: diagonal terms.  The metric induced on $\t$ by the Killing metric is

691: some multiple of the standard euclidean metric, so that $\langle T,

692: T'\rangle=c\sum_i t_i t'_i$ for some $c>0$.

693: 

694: 

695: 

696: \bn

697: 

698: \begin{center}

699: \unitlength=1cm

700: \begin{picture}(5,4)(-2,-2)

701: \put(0,0){\circle*{.4}}

702: 

703: \put(-1.5,.866){\circle*{.25}}

704: \put(1.3,-1){$\oplus$}

705: \put(-2.3,1){$\alpha_{12}$}

706: \put(1.2,-1.3){$-\alpha_{12}$}

707: 

708: 

709: 

710: \put(0,1.5){\circle*{.25}}

711: \put(0,-1.5){\circle{.25}}

712: \put(-.3,1.8){$\alpha_{13}$}

713: \put(-.5,-2){$-\alpha_{13}$}

714: 

715: 

716: 

717: \put(1.4,.866){\circle*{.25}}

718: \put(-1.7,-1){$\oplus$}

719: \put(1.6,1){$\alpha_{23}$}

720: \put(-2.2,-1.3){$-\alpha_{23}$}

721: 

722: \put(-1.4,  -0.8){\line(5,3){2.9}}

723: \put(0,-1.4){\line(0,1){3}}

724: \put(-1.5,  0.866){\line(5,-3){2.8}}

725: \end{picture}

726: 

727: \bn

728: 

729: {\small Figure 3: The root diagram of $\sl_3$}

730: 

731: \end{center}

732: 

733: 

734: \bn

735: 

736: 

737: 

738: 

739: 

740: 

741: \bn{\bf Reading the root diagram}.  One can read much of the structure

742: of $\g$ from its root diagram  in a formula-free manner. Here

743: is the key observation. Let $\alpha, \beta$ be two roots with

744: (non-zero) root vectors $E_\alpha\in\g_\alpha,$ $ E_\beta\in\g_\beta$.

745: That is, $$[T,E_\alpha]=\alpha(T)E_\alpha,\quad T\in\t, $$ and

746: similarly for $\beta$. It then follows immediately from the Jacobi

747: identity that

748:  $$[T,[E_\alpha, E_\beta]]=(\alpha+\beta)(T)[E_\alpha, E_\beta].$$ This means that

749:  \begin{itemize}

750: 

751:  \item[(1)] if  $\alpha+\beta\neq 0$ and is not a root then $[E_\alpha, E_\beta]=0$;

752: 

753:  \item[(2)] if  $\alpha+\beta\neq 0$ and is a root then $[E_\alpha, E_\beta]\in\g_{\alpha+\beta}$;

754: 

755:  \item[(3)] if $\alpha+\beta= 0$, i.e. $\beta=-\alpha$, then $[E_\alpha, E_\beta]\in\t$.

756:  \end{itemize}

757: 

758: This set of 3 conclusions permit us to see at a glance from the

759: diagram a fair amount of the structure of $\g$.  In the last two cases

760: one can further show that $[E_\alpha, E_\beta]$ is non-zero and

761: determine, with some calculations, the actual bracket, as will be

762: illustrated in Appendix B.

763: 

764: \bn{\bf Example: reading the root diagram of $\sl_3$.} Let us consider

765: the subspace $\p\subset\sl_3$ spanned by $\t$ and the root spaces

766: corresponding to the roots marked with dark dots in figure 3. 

767: 

768: 

769: 

770:  The diagram shows that $\p$ is a 5-dimensional subalgebra, i.e. it is

771: closed under the Lie bracket (there are 4 dark dots, but remember that

772: the thick dot at the origin stands for the 2-dimensional Cartan

773: subalgebra). Indeed, $\p$ is the subalgebra of upper triangular

774: matrices (including diagonal ones), with corresponding subgroup

775: $P\subset\SL_3$, the subgroup of upper triangular matrices with

776: determinant=1.  The quotient space $\SL_3(\R)/P$   can be

777: identified with the space $F$ of full  flags in $\R^3$.  A full

778: flag is a  pairs

779: $(l,\pi)$, where $l$ is a line and $\pi$ is a plane, and

780: $l\subset\pi\subset\R^3$.    The ``standard flag'' consisting

781: of the $x$ axis sitting inside the $xy$ plane has isotropy group $P$. The tangent space to $F$ at this  base

782: point  is naturally identified with $\sl_3/\p$,

783: represented in the root diagram by the remaining three light dots.

784: %(labelled  $-\alpha_{12},-\alpha_{13},-\alpha_{23}$).

785: Two of the light dots are marked +. The diagram shows that the root spaces corresponding to these roots

786: %(labelled $-\alpha_{12}$ and $-\alpha_{23}$)

787: span a $\p$-invariant 2-dimensional subspace of $\sl_3/\p$ which Lie

788: generates the root space associated with the third light

789: dot. %($-\alpha_{13}$).  

790: This means that we have on $F$ an

791: $SL_3(\R)$-invariant rank 2 contact distribution, i.e. a

792: non-integrable distribution that Lie generates the tangent bundle. 

793: 

794: This distribution can be geometrically interpreted

795: as  the  ``tautological"

796: contact distribution on $F$.  This distribution  is spanned by two vector fields,

797: corresponding to the two +s in figure 3.  One vector field generates the flow

798: in which the line   $l$ spins within   the plane $\pi$, while  the plane remains  fixed.  The other vector field generates the flow in which   the plane $\pi$ rotates about

799: the line $l$, while the line remains fixed.

800:  

801: 

802: 

803: 

804: \bn{\bf Reading the $\g_2$ diagram.}  

805: Now let us draw conclusions in a

806: similar fashion from the $\g_2$ diagram.  There are   twelve roots 

807: in the diagram (figure 2) and so 12 root spaces.  The rank of $\g_2$ is $2$

808: and so the dimension of $\g_2$ is $14 = 2 + 12$.  

809:   Consider the 9-dimensional

810: subspace $\p\subset\g_2$ spanned by $\t$ and the root spaces

811: associated with the roots marked by the dark dots in the diagram of figure 2.

812: Then the diagram shows that

813: \begin{itemize}

814: \item $\p$ is closed under the Lie bracket, i.e. is a subalgebra (a

815: so-called parabolic subalgebra, a subalgebra containing a Cartan

816: subalgebra).

817: \item Let $P\subset G_2$ be

818: the corresponding subgroup. It follows that $G_2$ has a 5-dim

819: homogeneous space $G_2/P$, whose tangent space $\g_2/\p$ at a point is

820: represented by the remaining 5 light dots.

821: 

822: \item Two of the light

823: dots are marked with +. The diagram shows that their root spaces generate a 2-dim $\p$-invariant

824: subspace  $W_1\subset\g_2/\p$, hence a $G_2$-invariant rank 2 distribution on $G_2/P$.

825: 

826: \item This

827: distribution is not integrable, in fact, it is a distribution of type $(2,3,5)$,

828: since the diagram shows that bracketing once gives the light dot marked with

829: $\sigma_3$ and bracketing again gives the remaining two light dots.

830: 

831: \end{itemize}

832: 

833: To summarize, we have  assembled the ingredients for the data

834: $(G_2, P, W_1)$.   To prove the theorem

835: is to provide the geometric interpretation

836: of this distribution as $(\tilde Q, \tilde D)$ of theorem 1. 

837: The first step in doing so is to embed $\so_3\oplus\so_3$ in $\g_2$. 

838: 

839: 

840: 

841: 

842: 

843: 

844: 

845: 

846: 

847: \section{The  maximal compact subgroup of $G_2$}

848: 

849: \subsection{Algebraic strategy of the proof.}

850: 

851: In the previous sections we   assembled the  Lie algebraic   data,

852: $(\so_3 \oplus \so_3, \h, D)$ and $(\g_2, \p, W)$

853: with corresponding group data $(K, H, D)$ and   $(G_2, P, W)$.  

854: The key to  theorem 1  is to embed

855: $\so_3\oplus\so_3$   in $ \g_2$.  This embedding is constructed

856: in   the next section.    Once established, we obtain

857: a diffeomorphism between corresponding distributions

858: by following the   observation made in  section \ref{shrinking}.

859: 

860: We recap that observation,  adding a bit of topology.  

861: Suppose that $\k \subset \g$

862: and $\p \subset \g$ are   Lie subalgebras of the Lie algebra $\g$.

863: Suppose that the natural map  $\k/\k \cap \p \to  \g/ \p$ is a linear 

864: isomorphism.  Let  $W \subset \g/\p$ be an $ad$- $\p$-invariant subspace 

865: and   $W_1 \subset \k/\k \cap \p$

866:  the corresponding subspace.

867: Then   we will say that the   Lie algebraic distributional data

868: $(\g, \p, W)$ and

869: $(\k, \k \cap \p,  W_1)$ are isomorphic.  If the corresponding 

870: connected Lie groups are $K \subset G$ and

871: {\it if } $K$ is compact, then we can conclude

872: that the data $(G; P, W)$ and $(K; K \cap P, W_1)$

873: define isomorphic manifolds with distributions.

874: For when $K$ is compact we have that $K/K \cap P$ is a

875: compact and open submanifold of $G/P$ and hence is diffeomorphic

876: to $G/P$.  And under this diffeomorphism

877: the distribution corresponding to $W$ is the same as the one represented

878: by $W_1$.  

879: 

880: The compactness assumption on $K$ is neccessary to conclude

881: that $G/P = K/(K \cap P)$. 

882: Think of the case $K = \C^* \subset G = SL(2,\C)$

883: where $G$ acts on the sphere $Q = \C \cup \{ \infty \}$  by Mobius transformations

884: and where $\C^*$ corresponds to the complex scalings $z \mapsto \lambda z$,

885: $\lambda \ne 0$. 

886: The fixed points of the $\C^*$-action are  $0, \infty$.

887: The  $\C^*$ orbit through any   point $z_0 \ne 0, \infty$

888: is open, being the whole sphere minus the two fixed points.

889:  Thus $G/P \ne K/ (K \cap P)$ where $P$ is the isotropy group of $z_0$. 

890:  But we still have $\k/(\k \cap \p) = \g/ \p$ since the orbit of $z_0$ is open. 

891:  

892:  Lie algebraic data   defines the corresonding Lie group data  only  up to a covering.

893:  We can insure that there are only finitely many such coverings

894:  by knowing that  $\k$, like $\so_3 \oplus \so_3$, is  the compact

895: real form of its corresponding complex Lie algebra.

896: For in this case there are only a finite number of 

897: connected  Lie groups $K$  with    Lie algebra $\k$,

898: all of these being compact and covered by the simply connected $K$.

899: To say this in another way, suppose we are given  Lie algebra data

900: $(\g, \p, W)$ and $(\k, \k \cap \p,  W_1)$ as above, and

901: suppose that $\k$ is a compact real form.  Let 

902: $(G, P, W)$ and $(K, H, W_1)$ denote {\it any}   Lie-group data realizing these

903:  respective  Lie algebraic data  {\it where we are no longer assuming that $K \subset G$}.

904:  Then the   two manifolds-with-distribution which they stand for

905: are isomorphic up to a finite cover. 

906: By this we mean, there is a third  manifold-with-distribution $(X, E)$

907: and covering maps $\pi_G: X \to G/P$,

908: $\pi_K:  X \to K/H$ such that $\pi_G ^* W = \pi_K ^* W_1 = E$.

909: Indeed,  we can take   $X$  to be $\tilde G/ \tilde P$

910: where $\tilde G$ is the unique simply connected Lie group with algebra

911: $\g$. 

912: 

913: To establish theorem 1  we will apply these considerations

914: to the case  $G = G_2$ and   $K \subset G_2$

915:   its maximal compact subgroup.  

916: We will show that $(\k, \k \cap \p)$

917: is isomorphic to $(\so_3\X\so_3,  \h = \R(\e_3,\e_3))$.

918: And we show that under this isomorphism

919: $W_1 \subset \k/ \k \cap \p$

920: corresponds to the rolling distribution 

921: when the   ratios of the rolling spheres are   $1:3$.     

922: 

923: 

924: \subsection{ Finding   Maximal compacts.} How can we ``see'' a maximal compact subgroup of $G_2$ tangled within   its root diagram? Let us look back again

925: at the example of $\SL_3(\R)$. Here the

926:  maximal compact subgroup is $\SO_3$, with Lie algebra $\so_3$, the set of 3 by 3 antisymmetric matrices.

927:  These are spanned by the vectors $E_{ij}-E_{ji}$, $i>j$. So we see that corresponding to each pair

928: of ``antipodal" roots  $\pm\alpha_{ij}$ we have one generator of $\k$,  lying   in the sum of the two corresponding root spaces.

929: 

930: 

931: More generally, for the ``split'' real form of any semi-simple Lie

932: algebra (such as our $\g_2$), the situation is similar: we get the Lie

933: algebra $\k$ of a maximal compact subgroup $K\subset G$ by taking the

934: sum of 1-dimensional subspaces, one subspace for each pair of

935: antipodal roots $\pm\alpha$.  In fact, there is a certain

936: particulary ``nice" choice of root vectors $E_\alpha\in\g_\alpha$

937: (sometimes called a ``Weyl basis''), so that the sought-for line is

938: given by $\R(E_\alpha-E_{-\alpha}),$ as in the $\sl_3$ case.

939: 

940: 

941: 

942: \bn

943: 

944: In the case of $\g_2$ we thus have that \begin{itemize}

945: \item  $\k$ is the sum of six 1-dimensional subspaces  $\s_{i},\l_{i}$, $i=1,2,3,$ where

946: $\s_{i}$ lies in the sum of the root spaces corresponding to $\pm\sigma_i$, and

947:  $\l_{i}$ lies in the sum of the root spaces corresponding to $\pm\lambda_i$.

948: 

949: \item The isotropy of the $K$-action, $H=K\cap P\subset K$,  is given  in the diagram  by

950: the vertical segment, $\h=\l_{3}$.

951: 

952: \item The distribution plane $W\subset \k/\h$ is generated by  $\s_{1},\s_{2}$ (mod $\h$).

953: 

954: \end{itemize}

955: 

956: 

957: We have thus assembled the required ingredients for a   ``distribution data" $(\k,\h,W)$.

958: 

959: 

960: 

961: 

962: \subsection{$\k\simeq\so_3\oplus\so_3$}

963: 

964: Our task here is to define an isomorphism $\k\simeq\so_3\oplus\so_3$

965: that maps $(\k,\h,W)$ to the data of \S4 with $\rho=3$ or $1/3$. This entails the

966: decomposition of $\k$ into the direct sum of two ideals, each 

967: isomorphic to $\so_3$. It would have been quite nice and simple if the

968: sought-for decomposition of $\k$ had been the decomposition into

969: ``long" ($\l_{i}$) and   ``short" ($\s_{i}$). But this is not the case. For   the

970: diagram shows that although the $\l_{i}$ generate an

971: $\so_3$ subalgebra of $\k$, this subalgebra is not an ideal, so is not one

972: of the summands in the decomposition. And the $\s_{i}$

973: do not generate even a subalgebra. So we have to work harder, i.e. write

974: down the precise commutation relations.

975: 

976: 

977: \begin{proposition} There is

978:  a basis $\{S_i,L_i | i=1,2,3\}$ of $\k$, with $S_i\in \s_{i}$ and $L_i\in \l_{i}$,

979:  such that

980: $$[L_i,L_j]=\epsilon_{ijk}L_k,\quad [L_i,S_j]=\epsilon_{ijk}S_k,\quad [S_i,S_j]=\epsilon_{ijk}({3\over 4}L_k-S_k),$$

981: where  $\epsilon_{ijk}$ is the ``totally antisymmetric tensor on 3 indices" (

982: $\epsilon_{ijk}=1$

983: if $ijk$ is a cyclic permutation of 123, -1 if anticyclic permutation, and

984: 0 otherwise).

985: \end{proposition}

986: 

987: The proof of this proposition is relegated to Appendix B. 

988: It  consists of  simple but tedious

989: calculations which we could not ``see'' in the diagram. We tried.  We

990: were reduced to picking up as nice as possible basis for $\g_2$ and

991: calculating the corresponding structure constants with the help of  Serre  \cite{Serre}. 

992: 

993: \bn Now set $$\e_i':={3L_i+2S_i\over 4}, \quad \e_i'':={L_i-2S_i\over 4}, \quad i=1,2,3.$$

994: These 6 vectors form a new basis for $\k$ and satisfy the standard $\so_3\oplus\so_3$ commutation relations

995: \begin{equation}

996: [\e_i',\e_j']=\epsilon_{ijk}\e_k', \quad

997: [\e''_i,\e''_j]=\epsilon_{ijk}\e''_k, \quad [\e_i',\e''_j]=0,

998: \label{std}

999: \end{equation}

1000:  thus

1001: establishing the desired Lie algebra isomorphism $\k\simeq\so_3\oplus\so_3$.

1002: 

1003: 

1004: 

1005: 

1006: \begin{corollary} The map $\k\to\so_3 \oplus\so_3$ defined by

1007: $\e_i'\mapsto (\e_i,0),$ $\e_i''\mapsto (0,\e_i),$ $i=1,2,3,$ is a Lie

1008: algebra isomorphism. It maps $\h=\R L_3$ to $\R(\e_3,\e_3)$ and the

1009: 2-plane in $\k$ generated by $S_1,S_2$ to the 2-plane in

1010: $\so_3\X\so_3$ defined in the Proposition of \S4 for $\rho=3$.

1011: Interchanging the summands in $\so_3 \oplus\so_3$, i.e. mapping 

1012: $\e_i'\mapsto (0,\e_i),$ $\e_i''\mapsto (\e_i,0),$ correponds to $\rho=1/3$. 

1013: \end{corollary}

1014: 

1015: The first assertion is eq (\ref{std}). The second assertion

1016: is easily verified using the last Proposition. We have thus

1017: defined a $G_2$-action on some finite cover of the rolling configuration,

1018: one which preserves the pulled-back distribution when the 

1019: radii of the two balls are in the ratio  $3:1$ or $1:3$. QED

1020: 

1021: 

1022: \bn{\bf How we came up with the formulae for $\e'_i,

1023: \e''_i$.} The first thing to observe is that since $L_3$ generates the

1024: isotropy $H=P\cap K$ we should have $L_3=\e_3'+\e_3''.$ Since

1025: everything is symmetric in 1,2,3 we conclude that $L_i=\e_i'+\e_i'',$

1026: $i=1,2,3.$ Next since $S_3$ commutes with $L_3$ we should have

1027: $S_3=a\e_3'+b\e_3''$ for some constants $a,b$, and again by symmetry

1028: $S_i=a\e_i'+b\e_i''$, $i=1,2,3.$ Now by using the sought-after commutations

1029: relations for the $\e'_i, \e''_i$ and the known commutations for  $L_i, S_i$ we get that $a,b$

1030: are roots of the equation $x^2+x-3/4=0,$ i.e. $a=1/2, b=-3/2.$ Hence,

1031: $$L_i=\e_i'+\e_i'',\quad S_i=(\e_i'-3\e_i'')/2,\quad i=1,2,3.$$

1032: Inverting these equations we obtain the above equations for $\e_i', \e_i''$.

1033: 

1034:  

1035: \section{Split Octonions and the projective quadric realization of $\tilde Q$}

1036: 

1037: To show that $G_2 \subset Aut(\tilde Q, \tilde D)$

1038: (theorem 1),   it remains to identify $G_2/P$ with

1039: the $\tilde Q = S^3 \times S^2$ of the theorem   

1040: and to show that the covering map $S^3 \times S^2 \to Q$

1041: corresponds to the identification $G_2 /P = K/H$

1042: composed  with the projection $K/H \to (\pm 1, \pm 1) \backslash K / H$.

1043: In order to do these things  will use the fact, discovered by Cartan \cite{C3}

1044: in 1914, that $G_2$

1045: is the   group of automorphisms

1046: of the ``split octonions'' $\tilde \O$.  We will

1047: follow the treatment in the book  \cite{Harvey}, in  the section   ``The

1048: Cayley-Dickson process" (p.104).  There  further   consequences and motivation  

1049: can be found.    

1050: 

1051: The split octonions $\tilde \O$ are a real  eight-dimensional

1052: algebra with unit and  which is neither associative nor commutative. 

1053: We identify $\tilde \O$ with     $\H^2$, the 2 dimensional quaternionic vector space.

1054: Its   multiplication law is  

1055: \begin{equation}

1056: (a,b)(c,d)=(ac+\bar d b,da+b\bar c), \hskip .6cm a, b, c , d \in \H

1057: \label{product}

1058: \end{equation}

1059:  The unit  $1 \in  \tilde \O$ is   $(1,0) \in \H^2$

1060: 

1061:  The  automorphism group of  a real algebra is  $A$

1062:   is defined to be 

1063:   the space of nonzero real invertible linear maps  $g: A \to A$

1064:  satisfying  $g(xy) = g(x) g(y)$ for all $x, y \in A$.

1065:   $G_2$ is the automorphism group of $\tilde \O$.  

1066:     

1067:   The unit  $1$ 

1068:   is automatically invariant under any automorphism of $\tilde \O$,

1069:   so that $\R = \R 1 \subset \tilde \O$ is a $G_2$-invariant subspace.

1070:   This subspace has a $G_2$-invariant complement:  

1071:   $$ \tilde \O = \R 1 \oplus V = Re(\tilde \O) \oplus Im(\tilde \O) $$

1072:   In  quaternionic terms: 

1073:   \begin{equation}

1074:   V=Im\tilde\O=Im\H\oplus\H   \subset \H ^2 .

1075:   \label{V}

1076:   \end{equation}

1077:   To see the invariant nature of $V$, we

1078:   use   the split-octonion conjugation  $x \mapsto \bar x$

1079:   defined by  $x=(a,b)\in\tilde\O \mapsto \bar x=(\bar a, -b)$

1080:   for $x \in \tilde\O$.

1081:    Then $x=Re(x)+Im(x)$,

1082: $Re(x)=(x+\bar x)/2 \in \R 1$, and $Im(x)=(x-\bar x)/2$.

1083: Also $x \bar x = - \langle x, x \rangle 1$

1084: where  $\langle x,y\rangle=Re(x\bar y)$ is an inner product of signature $4,4$ on

1085: $\tilde\O$ which is  invariant under the action of $G_2$.     

1086: $V$ is the orthogonal complement of 

1087: $1 \in \tilde \O$  relative to this inner product.

1088: Alternatively, it  can be shown that $x \in V$ if and only if $x^2 = \langle x, x \rangle 1$

1089:  (see \cite{Harvey}, lemma 6.67), proving the

1090:    $G_2$-invariance of $V$.   

1091: $V$   forms  a 7-dimensional inner product 

1092: space of signature $(3,4)$ relative to the restriction of 

1093: $\langle \cdot , \cdot \rangle$. The $G_2$ action on $V$

1094: leaves this inner product invariant, so that  

1095: $G_2$ is realized as   a subgroup of $SO(3,4)$ through its representation

1096: on $V$. 

1097: 

1098: The maximal compact of $G_2$

1099: is $K = SO(4) = ( SU(2) \X SU(2) )/ \pm(1,1)$. 

1100: See Appendix B and \cite{Vogan}.  Upon restricting from  $G_2$ 

1101: to $K$, the representation  $V$ decomposes into irreducibles according

1102: to   (\ref{V}).  In other words, thinking of $SU(2)$ as unit quaternions, 

1103:   $(q_1, q_2) \in SU(2) \times SU(2) = \tilde K$ (the universal cover of $K$)

1104:    and $(a,b) \in Im(\H) \oplus \H = V$ we

1105: have $(q_1, q_2) \cdot (a,b) = (q_1 a \bar q_1, q_1 b \bar q_2)$.  

1106: 

1107: In quaternionic terms (\ref{V}) the quadratic form associated to 

1108: our $(3,4)$ inner product on $V$ is 

1109: $$\langle (v, q), (v, q) \rangle = -|v|^2 + |q|^2.$$

1110: Since $K$ acts transitively on the product of spheres

1111: $S^2 \times S^3 \subset Im (\H) \oplus \H = \tilde \O$

1112: we have that 

1113: $G_2$ acts transitively on the null cone

1114: $\{ x = (v, h): \langle x,x\rangle=0, x\neq 0 \}$.

1115: (To see that we can change the `length'

1116: of an $x$ in the null cone using $G_2$,

1117: use the fact that each such  null vector is a nonzero weight vector relative

1118: to some choice of maximal Cartan $T \subset G_2$.

1119: This maximal Cartan then acts on $x$ by scaling.  See the description

1120: following eq. (\ref{basis}) below.) 

1121: Thus $G_2$ acts  

1122:  transitively on the space  of null rays

1123:  $$ C=\{\R^+ x\subset V|\langle x,x\rangle=0, x\neq

1124: 0\}\subset P^+ (V) := \hbox{ rays in } V $$. 

1125: This $C$ is a nondegenerate 5-dimensional quadric

1126: sitting  in the 6-dimensional real ray space

1127: $P^+(V)$ (diffeomorphic to $S^6$).  We can describe  points 

1128: of the ray space  $P^+ (V)$ using homogeneous coordinates 

1129: $[x] = [v,h] = [\lambda v, \lambda h]$,

1130: $\lambda \in \R^+$ with $(v,h)\in Im\H\oplus\H =V$.

1131:  $C$ is

1132: defined  by the homogeneous equation $\|v\|^2=\|h\|^2.$ 

1133: Using the $\R^+$ action, we   normalize  $\|v \| = 1$,

1134: proving that  $C$ is diffeomorphic to

1135: the  product of spheres  $S^2 \times S^3 \cong S^3 \times S^2 =\tilde Q$

1136: which appears  in theorem 1.   

1137: 

1138:  

1139:   Given a point $\R^+ x = [x] \in C,$  set $$x^\perp=\{y\in V|\langle x,y\rangle=0\}, \quad

1140: x^0=\{y\in V| xy=0\}.$$ Then

1141: \begin{proposition}

1142:  $$\R x\subset x^0\subset (x^0)^\perp\subset x^\perp\subset V,$$

1143: and the dimensions are $1,3,4,6,7.$

1144: \end{proposition}

1145: 

1146: \pf Use  the   definitions of  the split octonion product (eq (\ref{product})) and inner product.\qed

1147: 

1148: When we projectivize,  $x^\perp$ maps to the tangent plane $T_{[x]} C$ to $C$ at $[x]$, and $x^0$

1149: maps to a 2-dimensional subspace $D_{[x]}\subset T_{[x]} C$. 

1150: Letting $[x]$ vary over $C$ we have

1151: defined a   rank 2 distribution $D\subset TC$. 

1152: This construction of $(C, D)$   depends only on

1153:  on the algebraic structure of   $\tilde \O$, so that  $G_2=Aut(\tilde\O)$

1154:  acts on $C$ preserving $D$. 

1155:  

1156:  

1157: \begin{proposition}

1158: \label{projective}

1159: The (ray) projective quadric $C$ is a homogeneous

1160: space for $G_2$.  $C$  is diffeomorphic to $\tilde Q =S^3 \times S^2$

1161: of theorem 1,  and is naturally

1162: endowed with a   $G_2$-invariant  distribution  $D$ of rank $2$.

1163: Viewed as a $G_2$-homogeneous space, the data for $(C,D)$ 

1164:  coincides  with the data  $(G_2, P, W)$ of  section 3.3.   Viewed

1165:  as a $K$-homogeneous space, its Lie algebraic data

1166:  coincides with that of the rolling distribution $(\k, \h, D)$ for 

1167:  the ratio $1:3$.   The distribution on $C$

1168: pushes down to the rolling distribution for ratios $3:1$

1169: under the two-to-one cover   $C = S^3 \to S^2 \to Q = SO_3 \times S^2$

1170: \end{proposition} 

1171: 

1172: This proposition immediately implies

1173: that part of the theorem we are going to prove:

1174: that $G_2 \subset Aut(\tilde Q, \tilde D)$. 

1175: 

1176: {\bf Steps of the proof.}

1177: In the paragraph   preceding the   

1178:  proposition we proved that $C$ is a homogeneous

1179:  space for  $G_2$, that $D$ is invariant under this $G_2$ action,

1180:  and that $C$ is diffeomorphic to $\tilde Q$.

1181:  Next, we will prove that the  coincidence of the $\g_2$-data for $(C,D)$ 

1182:  and the data   $(\p, W)$ of the previous section.

1183:  For this we will  use  the weights for the 

1184:  $G_2$-representation space $V = Im(\tilde \O)$. 

1185:  

1186:  

1187: {\bf   Weights for the $7$-dimensional representation.}

1188:  

1189:  Here is  the weight diagram  for this representation. 

1190:  

1191: \begin{center}

1192: \unitlength=1cm

1193: \begin{picture}(5,4)(-2,-2)

1194:  

1195: 

1196: \put(0,0){\circle{.3}}

1197: 

1198: \put(0,0){\circle{.18}}

1199: 

1200: 

1201: 

1202: \put(-.5,-.866){\circle{.25}}

1203: \put(-.8,-1){X}

1204: \put(.5,.866){\circle{.25}}

1205: \put(-.5,.866){\circle{.25}}

1206: \put(-.68,.866){X}

1207: \put(.5,-.866){\circle{.25}}

1208: 

1209: \put(-.5,-.866){\circle{.1}}

1210: \put(.5,-.866){\circle{.1}}

1211: 

1212: \put(.5,.866){\circle{.1}}

1213: \put(-.5,.866){\circle{.1}}

1214: 

1215: 

1216: 

1217: 

1218: \put(-1,0){\circle*{.25}}

1219: \put(1,0){\circle{.25}}

1220: \put(-1,0){\circle{.1}}

1221: \put(1,0){\circle{.1}}

1222: 

1223: %this marks the weights for  V : one circle inside the other 

1224: 

1225: %\put(1.3,-0.1){$\sigma_3$}

1226: %\put(-2,-0.1){$-\sigma_3$}

1227: \put(-1.4,-0.1){X}

1228: 

1229: 

1230: \put(1.4,-.866){\circle{.25}}

1231: \put(-1.5,.866){\circle{.25}}

1232: %\put(-1.5,.866){X}

1233: %\put(1.6,  -1.1){$-\lambda_1$}

1234: %\put(-2.2,  0.866){$\lambda_1$}

1235: \put(-1.9,  .8){ X  }

1236: 

1237: \put(1.4,.866){\circle{.25}}

1238: \put(-1.5,-.866){\circle{.25}}

1239: \put(-1.9,  -1){X}

1240: %\put(1.7,.9){$\lambda_2$}

1241: %\put(-2.4,-1.1){$-\lambda_2$}

1242: 

1243: \put(0,1.5){\circle{.25}}

1244: \put(-.2,1.6){X}

1245: \put(0,-1.5){\circle{.25}}

1246: \put(-0.3,-1.7){X}

1247: 

1248: \put(0,-1.4){\line(0,1){3}}

1249: \put(-1.5,  -0.866){\line(5,3){2.77}}

1250: \put(-1.5,  0.866){\line(5,-3){2.77}}

1251: 

1252: \put(  -1, 0){\line(1  , 0   ){1.85}}

1253: \put(-.5,-.866){\line(3,5){.96}}

1254: \put(-.5,.866){\line(3,-5){.96}}

1255: 

1256: \end{picture}

1257: 

1258: Figure 4: Weights and roots  associated with the representation $V$.

1259: 

1260: \end{center}

1261: 

1262: 

1263: The  weights of the representation $V$  form a subset of the roots of $\g_2$.

1264: In figure 4 we redrew the root diagram of $\g_2$, marking  

1265: those roots  which are  weights for   $V$   with bullseye's. 

1266: They   are the six short roots and one zero root.

1267: The corresponding weight spaces $V_w$ are all one-dimensional.

1268: The black dot is a selected weight vector and corresponds to 

1269: a `choice of base point' for $C$.  The meaning of the X's will be given below.  

1270: 

1271: 

1272: \bn{\bf A reminder of the meaning of the weight  diagram.} 

1273: 

1274: We   recall the general case of a  representation $V$ of a

1275: semi-simple Lie algebra $\g$

1276: with Cartan subalgebra $\t$.  

1277: A {\it weight} for the  representation $V$ of $\g$

1278: is an element $w \in \t^*$ such that there is a nonzero

1279: vector $v \in V$ with the property that

1280: $\zeta \cdot v = w(\zeta) v$ for all $\zeta \in \t$.    The space of $v$'s for a given weight $w$

1281: is called the {\it weight space} for $w$ and is denoted $V_w$.  

1282: If, for given $w \in \t^*$ there is no such nonzero $v$

1283: then we set $V_w = 0$.   For a finite-dimensional representation

1284: the set of weights is finite.       We have

1285: $$V = \bigoplus_{w\in\t^*} V_w.$$

1286: The roots of $\g$ are the non-zero weights of the adjoint representation. 

1287: 

1288: If, as in our situation, the roots for $\t$ are real, then its `torus' $T = exp(\t)$  is noncompact and  acts on the weight spaces by scaling, as follows. 

1289: If $\lambda = exp(\xi) \in T$,  with $\xi \in \t$,

1290: then $\lambda e_w = exp(w(\xi))e_w$ for $w \in V_w$.

1291:  

1292: From $\zeta \cdot \xi  \cdot v =  \xi \cdot \zeta   \cdot v + [\zeta, \xi ] \cdot v$

1293: it follows that if $v$ is in the weight space for $w$

1294: and $\xi \in \g_{\alpha}$ is in the root space for $\alpha$ then $\xi v$

1295: is in the weight space for $w + \alpha$ (which, as above, could be zero).

1296: In other words: 

1297: $\g_{\alpha} V_w \subset V_{w + \alpha}$. This inclusion is  half of the rule:

1298: \begin{equation}

1299: w \hbox{ a weight}, \alpha \hbox{ a root} \Rightarrow  \g_{\alpha} V_w =  V_{w + \alpha}

1300: \label{rule}

1301: \end{equation}

1302: which is true for $V$.

1303: It follows in particular that if $v \in V_w$

1304: and $\xi \in \g_{\alpha}$ and if $w + \alpha$ is not a weight for the representation,

1305: then $\xi (v) = 0$.  We will use this fact momentarily.  

1306: 

1307: 

1308: We now construct  the weight spaces and the action of

1309: the torus  for our $G_2$-representation

1310: $V = Im(\tilde \O)$. 

1311: Let $n$ be an imaginary quaternion.

1312: Then $(n, n)$ and $(n, -n)$ are both null vectors in $V$.

1313:  Take as  basis for $V$:

1314: \begin{equation}

1315: e_1 = \frac{1}{2}(i, i), e_2 = \frac{1}{2}(j, j), e_3 = \frac{1}{2}(k,k) ; f_1 = \frac{1}{2}(i, -i), f_2 = \frac{1}{2}(j, -j), f_3 = \frac{1}{2}(k, -k)

1316: \label{basis}

1317: \end{equation}

1318: and 

1319: $$U = (0,1).$$  Then we have

1320: the multiplication table:

1321: $$e_i ^2 = f_i ^2 = 0 $$

1322: $$e_i f_j = f_j e_i = 0,  \hskip .2cm \hbox{ if }  i \ne j$$

1323: $$e_i e_j = f_k;  i, j, k \hbox{ a  cyclic permutation of } 1,2,3 $$

1324: $$f_i f_j =   e_k  ;   i, j, k \hbox{ a  cyclic permutation of } 1,2,3 $$

1325:  $$e_i f_i =   -\frac{1}{2} + \frac{1}{2}U$$

1326:  $$f_i e_i =  -\frac{1}{2}  - \frac{1}{2}U$$

1327:  $$e_i U =  e_i$$

1328:  $$f_i U = -  f_i $$

1329:  To complete the multiplication table, use that the conjugate

1330:  of  $xy$ is  $\bar x \bar y$,

1331:  so that if $x, y \in V = Im(\tilde \O)$ we have 

1332:  $y x = \bar z$ where

1333:  $z =  xy$.  

1334:  Thus, for example since $\bar f_k = -f_k$

1335:  we see that  $e_j e_i =-f_k$,  for   $i,j,k$ a cyclic permutation of $1,2,3$. 

1336: Now let $\lambda_1, \lambda_2, \lambda_3$

1337: be nonzero reals with $\lambda_1 \lambda_2 \lambda_3 = 1$.

1338: Let $\alpha_i, \beta_i, \gamma_i$

1339: and  $\tilde \alpha_i, \tilde \beta_i,  \tilde \gamma_i$ be real exponents for $i = 1, 2, 3$

1340: satisfying $\alpha_i + \beta_i + \gamma_i = 0$

1341: Then the scaling  transformation

1342: $$e_i \mapsto \lambda_1 ^{\alpha_i} \lambda_2 ^{\beta_i} \lambda_3 ^{\gamma_i} e_i$$

1343: $$f_i  \mapsto \lambda_1 ^{\tilde \alpha_i} \lambda_2 ^{\tilde \beta_i} \lambda_3 ^{\tilde\gamma_i}  f_i$$

1344: together with $z \mapsto z$ 

1345: preserves the multiplication table, and hence defines an element of   $G_2$,

1346: provided 

1347: $$\tilde \alpha_i = - \alpha_i,  \tilde \beta_i = - \beta_i, \tilde \gamma_i = - \gamma_i$$

1348: and provided that 

1349: $(\alpha_i, \beta_i, \gamma_i)$ are multiples of the values from the following

1350: weight table 

1351: 

1352: $$\begin{array}{|c||c|c|cl}

1353: \hline

1354:  &\alpha_i&\beta_i&\gamma_i\\

1355: \hline\hline

1356: i = 1  &2 &-1& -1\\

1357: \hline

1358: i=2& -1 &2 &-1  \\

1359: \hline

1360: i=3  &-1 & -1& 2   \\

1361: \hline

1362: \end{array}

1363: $$

1364: These scaling transformations generate the Cartan $T$ of $G_2$,

1365: and the table  gives the corresponding weights.  Thus for example

1366: $e_1$ is a weight vector with corresponding weight, relative to

1367: the basis for $\t$,  being $(2,-1,-1)$.  Here we view $\t$

1368: as being the collection of vectors $(a, b, c)$ with $a + b + c = 0$.

1369:   Looking at the inner products of these vectors we see that they are arranged

1370: on the weight diagram according to:

1371: 

1372: \begin{center}

1373: \unitlength=1cm

1374: \begin{picture}(5,4)(-2,-2)

1375:  

1376: 

1377: \put(0,0){\circle{.3}}

1378: 

1379: \put(0,0){\circle{.18}}

1380: 

1381: 

1382: 

1383: \put(-.5,-.866){\circle{.25}}

1384:  \put(.5,.866){\circle{.25}}

1385: \put(-.5,.866){\circle{.25}}

1386:  

1387: \put(.5,-.866){\circle{.25}}

1388: \put(.8, 1){$e_2$}

1389: \put(.76, -1){$e_3$}

1390: 

1391: \put(-.5,-.866){\circle{.1}}

1392: \put(.5,-.866){\circle{.1}}

1393: \put(-.8, -1.18){$f_2$}

1394: \put(-.8, 1.1){$f_3$}

1395: 

1396: \put(.5,.866){\circle{.1}}

1397: \put(-.5,.866){\circle{.1}}

1398: 

1399: 

1400: 

1401: 

1402: \put(-1,0){\circle*{.25}}

1403: \put(1,0){\circle{.25}}

1404: \put(-1,0){\circle{.1}}

1405: \put(1,0){\circle{.1}}

1406: 

1407: \put(-1.45, 0){$e_1$}

1408: \put(1.2, 0){$f_1$}

1409: 

1410: %this marks the weights for  V : one circle inside the other 

1411: 

1412:   

1413:  \put(.2, .2){$z$} 

1414:  

1415: \put(  -1, 0){\line(1  , 0   ){1.85}}

1416: \put(-.5,-.866){\line(3,5){.96}}

1417: \put(-.5,.866){\line(3,-5){.96}}

1418: 

1419: \end{picture}

1420: 

1421: Figure 5:  The weight space basis. 

1422: 

1423: \end{center} 

1424: 

1425: We are now in a position to compute  the $\g_2$-data

1426: associated to $(C, D)$ from the proposition. 

1427: 

1428: {\bf Weight vectors are null vectors: }   Because the inner product is $G_2$-invariant,

1429: the $\g_2$ action on $V$ satisfies $\langle \xi x,  x \rangle = 0$

1430: for any $\xi \in \g_2$, $x \in V$. Take $x$ a weight vector with nonzero weight $w$,

1431: and   

1432: take  $\xi \in \t$ with $w(\xi) \ne 0$.  From 

1433: $\langle \xi x,  x \rangle =  w(\xi) \langle x, x \rangle$

1434: we have that   $x$ is a null-vector.     

1435: 

1436: 

1437: {\bf Computing the isotropy data.}

1438: Set $c_0 = [e_1]$, the ray through $e_1$.

1439: We must show that the isotropy group of $c_0$

1440: is $P$.  

1441: 

1442: We begin by showing that  the isotropy algebra

1443: $\g_{c_0}$  of $c_0 = [e_1] $ is  

1444: $\p$.   We have that   $\g_{c_0} = \{ \xi \in \g_2 :  \xi e_1 = \lambda e_1 \hbox{ for some

1445: real number } \lambda \}$.

1446:  The black dot in figure 4 indicates

1447: the weight space spanned by $e_1$,  

1448: with corresponding   weight 

1449: by $w_1$.  (This weight is the root

1450: marked $- \sigma_3$ in Figure 3.) 

1451: According to the addition rule, 

1452: (\ref{rule}) if $\alpha \in \t^*$ is a root

1453: and $w_1 + \alpha$  {\it is not a weight} for $V$,

1454: then $\xi_{\alpha} x_0 = 0$.    Those

1455: roots for $w_0 + \alpha$ {\it is not a weight} are marked by X's in figure 4.

1456: The sum of the corresponding $\g_{\alpha} \subset \g_2$ is a vector

1457: space of elements $\xi$ satisfying   $\xi (e_1) = 0$.

1458: Now  the isotropy algebra $\g_{c_0}$of the ray through $e_1$

1459: consists of all those $\xi$ such that $\xi e_1 = \lambda e_1$

1460: for some real scalar $\lambda$.   

1461: The elements  $ H \in \t$   act on $e_1$ by

1462: scalar multiplication by $\lambda = w_1 (H)$.

1463: Referring to the diagram then, we see that

1464: $\p \subset \g_{c_0}$.   But there is no

1465: subalgebra of $\g_2$ lying between $\p$ and all of $\g_2$.  

1466: It follows that the isotropy algebra for the ray {\it is} $\p$.

1467: 

1468:   It follows  from this Lie algebra computation 

1469:     that the isotropy subgroup $G_{c_0} $

1470: {\it contains} $P$ and has Lie algebra  equalling the Lie

1471: algebra $\p$ of $P$.  Now   $P$

1472: was defined to be the {\it connected} Lie subgroup of $G_2$

1473: whose Lie algebra is $\p$, thus to show

1474: $G_{c_0} = P$ it suffices to show that $G_{c_0}$ is connected.

1475: We demonstrate connectivity 

1476: by applying  the homotopy exact sequence

1477: to the fiber bundle $G_{c_0} \to G_2 \to C = G_2 / G_{[x]}$.

1478: This exact sequence

1479: is   $\ldots  \to \pi_1(C) \to \pi_0 ( G_{c_0}) = \pi_0 (G_2) \to \pi_0 (C)$.

1480: Since $C$ is simply connected and connected we

1481: get that  $\pi_0 ( G_{c_0} = \pi_0 (G_2)$

1482: and since $\pi_0 (G_2) = 0$ we have our connectivity:

1483: $\pi_0 (G_{c_0}) = 0$.  

1484: 

1485: We have   established the isotropy ($P$ part) of the data for $(C, D)$.

1486:  

1487: 

1488: 

1489: {\bf Computing the distribution data}.    

1490: The distribution plane  $D(c_0)$   at $c_0$ 

1491: corresponds to $e_1 ^0$ -- the 

1492: subspace $S \subset V$ consisting  of

1493: those  vectors $y \in V$ for which 

1494: $e_1 y = 0$.    From the multiplication table 

1495: following the description of  our basis (\ref{basis})  we see that 

1496: $S = span \{ e_1, f_2, f_3 \}$.

1497: From Figure 4, we see that weights

1498: corresponding to $f_2, f_3$, , say $w_2, w_3$,

1499: are given by $w_2 = w_1 + \sigma_1$,

1500: $w_3 = w_1 + (-\sigma_2)$.  Compare Figure 3.  

1501: Let $x_1, y_1 \in \g_2$ be the corresponding

1502: nonzero root vectors for $\sigma_1, - \sigma_2$. 

1503: (We follow the $x, y$ notation from Figure 5, Appendix 2.)

1504: It follows from   rule (\ref{rule})

1505: that  $f_2$ is a multiple of   $x_1 (c_0)$

1506: and $f_2$ is a multiple of $y_2 (c_0)$.

1507: In other words, $S = W(e_1)$ mod $\p (e_1)$

1508: where $W$ is the space spanned by the roots indicated

1509: by the pluses in Figure 2.  We have proved

1510: that the Lie algebraic   data for

1511: $(C, D)$ is  $(\g_2, \p, W)$.  

1512: 

1513: 

1514: 

1515: {\bf The covering map.} 

1516: On the Lie algebra level we have shown that the data for

1517: $(C, D)$ is  $(\g_2, \p, W)$. As computed in section 3.3, Cor. 1, upon

1518: restricting the action of $G_2$  to $K$ this Lie algebraic data

1519:  $(\g_2, \p, W)$ corresponds to  

1520: the data $(\so_3 \oplus \so_3, \h, D(3;1))$.

1521: Thus, up to a finite cover, $(C,D)$ is the rolling distribution.

1522: Now $C$ is simply connected, and  $2:1$ covers  $Q$.

1523: This covering map  $C = S^3 \times S^2 \to Q = SO_3 \times S^2$

1524: is realized by forming the quotient of

1525: $C$ by the $\Z_2$ subgroup

1526: generated by image of  $\sigma = (\pm 1, 1) \in K = SU_1 \X_{\pm(1,1)} SU_2$.

1527:  Being an element of the symmetry group   $\sigma$ preserves

1528: the distribution $D$ on $C$, and so $D$ does  push down to

1529: the rolling space $Q$. The $\k$ data of the pushed-down distribution

1530:  remains   $(\so_3 \oplus \so_3, \h, D(3;1))$. Thus the pushed-down distribution

1531: is isomorphic to the rolling distribution on $Q$.  

1532: 

1533: QED

1534: 

1535: 

1536: 

1537: \section{Summary. Lack of  action on the rolling space. The theorem is done.    }

1538: 

1539: We have proved that $G_2 \subset Aut( \tilde Q, \tilde D)$

1540: and that is all that we are going to prove of

1541: theorem 1,  with the exception of   the fact that the 

1542:   $G_2$-action  does {\it not} descend to  $Q$.  (Recall from the introduction

1543:   we are not going to prove that $G_2 = Aut( \tilde Q, \tilde D)$.)

1544:   To prove that the $G_2$ action does not descend to $Q$, we 

1545:   realize as above that   $Q = \Z_2 \backslash C$ 

1546: where  the  $\Z_2 \subset K \subset G_2$ is generated by $\sigma = (\pm 1, 1)$.

1547: Now we use the following fact about group actions.  Suppose

1548: that a group $G$ (here $G_2$) acts effectively on a set $C$

1549: and that $\Gamma \subset G$.   (``Effectively'' means that the only

1550: group element acting as the identity on $C$ is the identity.)  

1551: Then the action of an element $g \in G$ descends to the quotient space $\Gamma \backslash C$ if and only if $g \Gamma g^{-1} = \Gamma$.  

1552: In particular, if $\Gamma$ is not normal in $G$ then

1553: the action of all of $G$ does not descend to the quotient $\Gamma \backslash C$. 

1554: Returning to our situation, we see that  

1555: if the $G_2$ action were to descend then this

1556: $\Z_2$ generated by $\sigma$ 

1557: would have to be normal.   But a discrete normal subgroup

1558: of a connected Lie group is central, and $G_2$ has

1559: no center.  See Appendix A, or \cite{Vogan}.  So our $\Z_2$ is not normal,

1560: and the $G_2$ action does not descend.

1561:  

1562: 

1563: {\bf Remark.}  Had we used  lines instead of rays

1564: when constructing  $C = \tilde Q$, we would have

1565: arrived at a quadric  $Q_f$ in the standard real   projective space $P(V)$

1566: which is double covered by $C = \tilde Q$.

1567: (The subscript `f' is for `false'.) 

1568: $Q_f \subset P(V)$ is diffeomorphic to    $S^3 \times_{\Z_2} S^2 = {\pm I} \backslash C$ where

1569: the notation $\times_{\Z_2}$

1570: indicates that we divide out by the action of the involution $(v, h) \mapsto (-v, -h)$.

1571: (This involution does not lie in $G_2$.)

1572: $C = \tilde Q$ double-covers both  $Q_f$  and

1573: $Q$, and the distribution $\tilde D$ pushes down to both covered spaces. 

1574:  But $Q_f$ is   topologically distinct  from $Q$.   Both

1575:  $Q$ and $Q_f$ are $\SO_3$-bundles over $S^2$. $Q$ is  the trivial $SO(3)$-bundle. 

1576: $Q_f$ is the other one. (Since $\pi_1 (\SO_3) = \Z_2$ there are precisely

1577: two topologically distinct $\SO_3$ bundles over $S^2$.)

1578: Because $-I \in GL(V)$ commutes with the $G_2$ action on $V$

1579: the $G_2$-action on $\tilde Q$ does descend to $Q_f$. 

1580: We find it  curious that   the action of $G_2$

1581: on $\tilde Q$ does descend to this `false'  rolling configuration space $Q_f$,  but not to the real one $Q$. 

1582:  

1583: 

1584: 

1585: %\section{Appendices.}

1586:  \appendix{}

1587:  \section{Covers. Two   $G_2$'s. }

1588: \label{twoG2s}

1589:  

1590: To understand our results, it helps to understand that  

1591:  up to isomorphism,  there are precisely two connected $G_2$'s:

1592:  the adjoint one which is the one we have been using,

1593:  and the simply connected one, which is the universal cover

1594:  of the adjoint one.  

1595:  For a general semi-simple Lie algebra $\g$  we can always form

1596: the simply connected Lie group $\tilde G$ having $\g$ as its   Lie algebra.

1597: If $Z$ is the center of  $\tilde G$, then $Ad(\tilde G) = \tilde G/Z$

1598: where $Ad(\tilde G)$ is the image of $G$ under the adjoint map

1599: from $\tilde G$ to $Hom(\g)$.  If $Z \ne I$ then $G \ne Ad(G)$.

1600: There are as   many distinct connected Lie groups with

1601: algebra $\g$ as there are distinct subgroups of $Z$, these being

1602: the connected topological groups covered by $\tilde G$ and covering   $Ad(G)$. 

1603: So,  when  $Z = \Z_2$ there are precisely two such Lie groups,   $\tilde G$, the simply connected one, and $G = Ad(\tilde G)$, the adjoint one.  

1604: 

1605: We find on p. 3 of Vogan \cite{Vogan} that the center of the simply connected

1606: $G_2$ is  indeed $\Z_2$, and hence we have precisely two $G_2$'s.   

1607: It will be useful to explain a few details of this computation of $Z(G_2)$.

1608: The universal cover of any $G$ contracts onto its maximal compact.

1609: Thus, if the maximal compact of $Ad(G)$ has finite fundamental group,

1610: then the universal cover $\tilde G \to Ad(G) = \tilde G /Z$

1611: is a finite cover, and so  the center $Z$ must be  finite.

1612: (At the other extreme,  the maximal compact of 

1613: $SL(2, \R)$ is a circle group, corresonding to the fact that its 

1614: universal cover has infinite center  $\Z$.)   

1615: 

1616: We saw above  that the  Lie algebra of 

1617: the maximal compact of  any   $G_2$ realizing $\g_2$

1618:  is  $\k = \so_3 \times \so_3$.  The  connected Lie groups $K$ 

1619:   having $\k$ as Lie algebra have fundamental groups 

1620:  consisting of

1621:   either $1$, $2$ or $4$ elements.  

1622:   It follows that the center $Z(G_2)$ of any $G_2$   is finite, and 

1623: hence compact.  Being compact and central, this center lies

1624: in every maximal compact:  $Z(G_2) \subset K \subset G_2$. 

1625: If we take the simply connected  $G_2$, call

1626: it $\tilde G_2$,  then its maximal compact is 

1627: $\tilde K = \SU_2 \times \SU_2$. The center of  $\tilde K$   is  the group of the four elements   $(\pm 1, \pm 1)$.  

1628: The center of $\tilde K$ need not be the center of $G_2$

1629: but it must 

1630: contain it:   $Z(\tilde G_2) \subset Z(\tilde K)$.

1631: To see what the actual center of $\tilde G_2$  is, it suffices to see how  

1632: $Z(\tilde K) \subset \tilde K$ acts on 

1633: the Lie algebra $\g_2$  under the adjoint action.

1634: This can be done using roots. 

1635: The center of $\tilde G_2$ is that part of $Z(\tilde K)$ which

1636: acts trivially on $\g_2$. A computation using roots and

1637: the restriction of the adjoint representation to $\tilde K$ shows that

1638: this part is $(1,1)$ and $-(1,1)$.   

1639:  

1640: 

1641: %\appendix{}

1642: \section{The isomorphism of $\k$ and $\so_3 \oplus \so_3$ from Proposition 3. }

1643: \label{maxcpt}

1644: 

1645: We complete the  proposition 3 from section 5, in which the

1646: explicit identification of  $\so_3 \oplus \so_3$

1647: as the Lie algebra   $\k$ of the maximal compact in $\g_2$.   We  follow  Serre  \cite{Serre}, 

1648: page VI-11: $\g_2$ is Lie-generated by the elements

1649: $x,y,h,X,Y,H$, subject to the following relations, which one can read

1650: off the root diagram.

1651: 

1652: $$\begin{array}{llll}

1653: [x,y]=h,& [h,x]=2x,& [h,y]=-2y,& \\

1654: \left[X,Y\right]=H,& [H,X]=2X,&[H,Y]=-2Y;&\\

1655: \left[h,X\right]=-3X, &[h,Y]=3Y;&[H,x]=-x,&[H,y]=y;\\

1656: \left[x,Y\right]=[X,y]=[h,H]=0;&&&\\

1657: \left[ad(x)\right]^4 X=0; &\left[ad(X)\right]^2x=0;&&\\

1658: \left[ad(y)\right]^4Y=0;& \left[ad(Y)\right]^2x=0.&&\\

1659: \end{array}

1660: $$

1661: 

1662: 

1663: 

1664: Taking Lie brackets of the vectors $x,y,h,X,Y,H$ we generate a complete set  $\{x_i,X_i, y_i,Y_i|i=1,2,3\}$ of root vectors for

1665: $\g_2$, which, together with the basis $h,H$ for the Cartan subalgebra form a basis for $\g_2$ as follows:

1666: %

1667: $$\begin{array}{llllll}

1668: x_3=x,&X_1=X, &x_2=[x,X_1],&x_1=[x,x_2],&X_2=[x,x_1],&X_3=[X_1,X_2];\\

1669: y_3=y,& Y_1=Y,&y_2=-[y,Y_1],&y_1=-[y,y_2],&Y_2=-[y,y_1],&Y_3=-[Y_1,Y_2].

1670: \end{array}

1671: $$

1672: We label each root in the diagram

1673: with the corresponding root vector.

1674: 

1675: \bn

1676: 

1677: \begin{center}

1678: 

1679: \unitlength=1cm

1680: \begin{picture}(5,4)(-2,-2)

1681: \put(0,0){\circle*{.4}}

1682: %

1683: \put(.5,.866){\circle*{.25}}

1684: \put(.65,1.05){$x_1$}

1685: \put(-.5,-.866){\circle*{.25}}

1686: \put(-.75,-1.25){$y_1$}

1687: %

1688: \put(-.5,.866){\circle*{.25}}

1689: \put(-.95,1.05){$x_2$}

1690: \put(.5,-.866){\circle*{.25}}

1691: \put(.65,-1.25){$y_2$}

1692: %

1693: \put(1,0){\circle*{.25}}

1694: \put(1.2,-0.1){$x_3=x$}

1695: \put(-1,0){\circle*{.25}}

1696: \put(-2.4,-0.1){$y=y_3$}

1697: %

1698: \put(0,1.7){\circle*{.25}}

1699: \put(-.2,2){$X_3$}

1700: \put(0,-1.732){\circle*{.25}}

1701: \put(-.2,-2.3){$Y_3$}

1702: %

1703: \put(-1.5,.866){\circle*{.25}}

1704: \put(-3.1, 1.05){$X=X_1$}

1705: \put(1.4,-.866){\circle*{.25}}

1706: \put(1.6,  -1.25){$Y_1=Y$}

1707: %

1708: \put(1.4,.866){\circle*{.25}}

1709: \put(1.7, 1.05){$X_2$}

1710: \put(-1.5,-.866){\circle*{.25}}

1711: \put(-2.2,-1.25){$Y_2$}

1712: 

1713: 

1714: 

1715: \put(0,-1.6){\line(0,1){3.3}}

1716: \put(-1.4,  -0.8){\line(5,3){2.77}}

1717: \put(-1.5,  0.866){\line(5,-3){2.77}}

1718: \put(  -.9, 0){\line(1  , 0   ){1.85}}

1719: \put(-.46,-.75){\line(3,5){.96}}

1720: \put(-.5,.866){\line(3,-5){.96}}

1721: 

1722: \end{picture}

1723: 

1724: \vskip .8cm

1725: Figure 5.  A basis for the Lie algebra.

1726: \end{center}

1727: 

1728: 

1729: \bn

1730: 

1731: We end up with a ``nice'' basis wrt which the structure constants are

1732: particulary pleasant; they are integers and have symmetry properties

1733: which facilitate greatly the work involved in their determination; you

1734: can also apply some elementary $\sl_2$ representation theory that

1735: further facilitate the calculation; it helps to work with the root

1736: diagram nearby.

1737: 

1738: 

1739: \bn{\bf Symmetry properties of the structure constants.}  Suppose $\alpha,\beta$ are two roots such

1740: that $\alpha+\beta$ is also a root. Let $E_\alpha, E_\beta$

1741: be the corresponding root vectors, as chosen above. Then

1742: $[E_\alpha,E_\beta]=c_{\alpha,\beta}E_{\alpha+\beta}$, for some

1743: non-zero constant $c_{\alpha,\beta}\in\Z$. The nice feature of

1744: our base is that the structure constants satisfy

1745: $$c_{-\alpha,-\beta}=-c_{\alpha,\beta}.$$ This cuts in half the amount

1746: of work involved, since you need only consider say $\alpha>0$ (the positive roots are the six dots in the last root diagram marked 

1747: with $x$'s and $X$'s).

1748: Combining this with the obvious $c_{\alpha,\beta}=-c_{\beta,\alpha}$

1749: (antisymmetry of Lie bracket) you obtain

1750: $$c_{\alpha,-\beta}=c_{\beta,-\alpha}.$$

1751: This cuts in half again the amount of work.

1752: 

1753: \begin{proposition} The structure  constants of $\g_2$, with respect to the basis of root vectors $\{x_i,X_i,y_i, Y_i | i=1,2,3\}$

1754: and the Cartan algebra elements $\{h,H\}$ are given as follows. The

1755: basis elements are grouped in three sets: positive (three $x$'s and

1756: three $X$'s), negative (three $y$'s and three $Y$'s), and Cartan

1757: subalgebra elements ($h$ and $H$).

1758: \begin{itemize}

1759: \item $[Positive, positive]$: other then the ones given above, and

1760: those which are zero for obvious reasons from the root diagram (sum of

1761: roots which is not a root): $$[x_1, x_2]= X_3.$$

1762: 

1763: \item $[ Positive, negative ]$:

1764: 

1765: $$\begin{array}{|c||c|c|c|c|c|c|}

1766: \hline

1767: c_{\alpha,\beta}&y_1&y_2&y_3&Y_1&Y_2&Y_3\\

1768: \hline\hline

1769: x_1&1&4&-4&0&12&-12\\

1770: \hline

1771: x_2& 4 &1 &-3 &1 &0 &3 \\

1772: \hline

1773: x_3  &-4 & -3& 1& 0&-3&0  \\

1774: \hline

1775: X_1 &0 &1 &0 &1 &0&-1 \\

1776: \hline

1777: X_2 & 12& 0& -3& 0& 1&36\\

1778: \hline

1779: X_ 3& -12& 3& 0&-1 &36&1\\

1780: \hline

1781: \end{array}

1782: $$

1783: 

1784:  \mn The 1's on the diagonal stand for the relations $[x_i,y_i]=h_i,$

1785:  $[X_i,Y_i]=H_i,$ where, in terms of our basis $\{h,H\}$ for the

1786:  Cartan subalgebra, $$h_1=8h+12H,\quad h_2=h+3H, \quad h_3=h,$$

1787: $$ H_1=H,\quad H_2=36(h+H), \quad H_3=36(h+2H).$$

1788: \item $[Cartan, anything]$: this is coded directly by  the root diagram:

1789: 

1790: - $ad(x)$  has eigenvalues and eigenvectors

1791: $$\begin{array}{l||c|c|c|c|c|c|c}

1792: \hbox{eigenvalue}&3&2&1&0&-1&-2&-3\\

1793: \hline

1794: \hbox{eigenvectors}&X_2,Y_1&x_3& x_1, y_2&X_3,Y_3, h,H&x_2,y_1&y_3&X_1,Y_2

1795: \end{array}$$

1796: 

1797: - $ad(X)$ has

1798: eigenvalues and eigenvectors

1799: $$\begin{array}{l||c|c|c|c|c}

1800: \hbox{eigenvalue}&2&1&0&-1&-2\\

1801: \hline

1802: \hbox{eigenvectors}&X_1&X_3,x_2,y_3,Y_2&x_1,y_1,h,H&X_2,x_3,y_2,Y_3&Y_1

1803: \end{array}$$

1804: 

1805: 

1806: \end{itemize}

1807: \end{proposition}

1808: 

1809: 

1810: 

1811: 

1812: \pf This is elementary, using only the Jacobi identity, but takes time. We will give as a typical example

1813: the calculation of $[x_1, x_2]$:

1814: 

1815: $$\begin{array}{rcll}

1816: [x_1,x_2]&=&[x_1,[x,X]]&\quad\hbox{(by definition of $x_2$)}\\

1817: &=&[x,[x_1,X]]+[X, [x,x_1]]&\quad\hbox{(Jacobi identity)}\\

1818: &=&[X, [x,x_1]]&\quad\hbox{(since $[x_1,X]=0$)}\\

1819: &=&[X, X_2]=X_3&\quad\hbox{(by definitions of $X_2, X_3$).}

1820: \end{array}

1821: $$

1822: The rest of the relations are derived in a similar fashion. \qed

1823: 

1824: \bn

1825: 

1826: 

1827: Now we are ready to define the generators of the Lie algebra of a maximal compact subgroup $K\subset G_2$. Let

1828: $$L_1=X_1-Y_1, \quad L_2={X_2-Y_2\over 6}, \quad  L_3={X_3-Y_3\over 6},$$

1829: $$S_1= {x_1-y_1\over 4}, \quad S_2= {x_2-y_2\over 2}, \quad S_3= {x_3-y_3\over 2}.$$

1830: 

1831: 

1832: Using the commutation relations of the last Proposition  one checks easily  that

1833: $$[L_i,L_j]=\epsilon_{ijk}L_k,\quad [L_i,S_j]=\epsilon_{ijk}S_k,\quad [S_i,S_j]=\epsilon_{ijk}({3\over 4}L_k-S_k).$$

1834: 

1835: \bn  Note: the strange-looking   coefficients 2,4,6 in the definition of the $L_i, S_i$ are chosen precisely so that we get

1836: these pleasing commutation relations.

1837:  

1838: 

1839: \section{The rolling distribution in Cartan's thesis}

1840: 

1841: \subsection{Cartan's constructions and claims.} 

1842: \label{claims}

1843: 

1844: In E. Cartan's thesis \cite{C2}, p.146,  we find the following constructions:

1845: consider $V=\R^7=\R^3\X \R^3\X \R$ with coordinates $(\x,\y,z)$, where $\x,\y\in\R^3$, $z\in\R$,  

1846: and the following 15 linear vector fields (hence linear operators) on  $V $:

1847: 

1848: 

1849: 

1850: \begin{itemize}

1851: \item $X_{ii}=-x_i \partial_{x_i} + y_i\partial_{y_i}+

1852: {1\over 3}\sum_{j=1}^3 (x_j\partial_{x_j}-y_j\partial_{y_j})$, $i=1,2,3$.

1853: 

1854: \item $X_{i0}=2z\partial_{x_i}- y_i \partial_z - x_j \partial_{y_k} + x_k\partial_{y_j},$ $(ijk)\in A_3=\{(123),(231),(312)\}.$

1855: 

1856: \item $X_{0i}=-2z\partial_{y_i}+ x_i \partial_z +y_j \partial_{x_k} - y_k\partial_{x_j},$ $(ijk)\in A_3.$

1857: 

1858: \item $X_{ij}=- x_j \partial_{x_i} + y_i\partial_{y_j},$ $i\neq j,$ $i,j=1,2,3.$

1859: 

1860: \end{itemize}

1861: 

1862: Cartan makes the following claims without proof:  

1863: 

1864: \begin{enumerate}

1865: 

1866: \item The linear span of these 15 operators is a 14 dimensional Lie subalgebra $\g\subset\End(V)$ isomorphic to $\g_2$. 

1867: 

1868: \item $\g$ preserves the  quadratic form on $V$ given by 

1869: $$J=z^2+\x\cdot\y.$$

1870: 

1871: \item The linear group $G\subset\GL(V)$ generated by $\g$ acts transitively on the projectivized null cone of $J$. 

1872:  

1873: 

1874: \item  $G$ preserves the system of 6 Pfaffian  equations  on $V,$

1875: given by the 6 components of 

1876: 

1877: $$

1878: \left\{

1879: \begin{array}{rcl}

1880: \alpha&:=&zd\x-\x dz+\y\X d\y=0,\\

1881:  \beta &:=&zd\y-\y dz+\x\X d\x=0,

1882:  \end{array}

1883:  \right.

1884:  $$

1885: which have as a consequence 

1886: $$

1887: \left\{

1888: \begin{array}{rcl}

1889: \gamma_1&:=&zdz+\x\cdot d\y=0,\\

1890: \gamma_2&:=&zdz+\y\cdot d\x=0.

1891:  \end{array}

1892:  \right.

1893:  $$

1894:  

1895:  $$$$

1896: \item $G$ preserves  a 5 parameter family of 

1897: 3 dimensional linear subspaces of $V,$ contained in the null cone of $J$, 

1898: 

1899: $$

1900: \left\{

1901: \begin{array}{l}

1902: \x-z\a +\b\X \y=0,\\

1903: \y-z\b +\a\X \x=0,

1904:  \end{array}

1905:  \right., 

1906:  $$

1907: where 

1908: $$\a\cdot\b+1=0.$$ 

1909: 

1910: 

1911: \end{enumerate}

1912: 

1913: Our goal in this appendix is to sketch proofs of these claims,

1914: provide a minor correction in one place,  

1915: relate Cartan's  construction to the  octonions,

1916: and show how they contain, in essence, the 

1917: construction of the rolling distribution $\tilde Q$

1918: via projective geometry, as in the proposition \ref{projective}

1919: from section 5.  

1920: 

1921: 

1922: \subsection{Relation with Octonions}.

1923: \label{connectOctonions}

1924: 

1925: Recall the basis $e_i ,   f_i , U $ of section \ref{basis} for $V$ (imaginary split octonions) 

1926: with its consequent multiplication table. 

1927: Make the change of basis $e_i \mapsto -e_i$,

1928: keeping $f_i, U$ as they were,  thus changing the

1929: signs of some entries of the multiplication table.

1930: Use this new   basis   $E_i = -e_i ,   f_i , U$

1931:   to identify   $V$ with $\R^3 \times \R^3 \times \R$

1932: by setting   $(\x, \y, z)  = \Sigma x_i E_i + \Sigma y_i f_i + z U \in V$. 

1933:  Referring to the multiplication

1934: table we compute 

1935: \begin{eqnarray*}

1936: (\x, \y, z) (\x', \y', z') &= &(-\y \times \y' -z \x' + z' \x, \x \times \x' + z \y' - z' \y, \frac{1}{2}( \x \cdot \y' - \x' \cdot \y)) 

1937: \\

1938: &&+ 1 \{ z z' + \frac{1}{2} ( \x \cdot \y' - \x' \cdot \y) \}.

1939: \end{eqnarray*}

1940: The last term is in the real part of the split octonions, and not in $V$. 

1941: It follows from this formula that $(\x, \y, z)^2 =J$, of Cartan's claim

1942: 2 in the preceding paragraph.

1943: Multiplying out   $(\x, \y, z) (d \x, d \y, dz)$ we  find that

1944: $$(\x, \y, z) (d \x, d \y, dz)

1945:  = (\alpha, \beta, \frac{1}{2} (\gamma_1 - \gamma_2)) + 

1946:  1 \{  \frac{1}{2} (\gamma_1 +  \gamma_2) \},$$

1947:  where  $\alpha, \beta, \gamma_1, \gamma_2$

1948:  are as in   Cartan's claim 4 of the previous paragraph. 

1949:  It follows that the any element of $G_2 = Aut(\tilde \O)$

1950:  preserves $J$ and preserves   the Pfaffian system of Cartan's claim 4.  

1951: The distribution  $D$ defined by this system  is, upon restriction

1952: to the   null cone $\{J =0 \} \setminus \{ 0 \}$,  precisely the    distribution $D$  which  we defined  in the final section of

1953: our paper:  $D(\x, \y, z) : = 

1954: \{ (\a, \b , c) :  (\x, \y, z) (\a, \b, c) = 0 \}$.  

1955: It follows that Cartan's construction, 

1956: pushed down to the space of rays using the

1957: $\R^+$-action,   yields

1958: precisely our $\tilde Q$.  

1959: 

1960: \subsection{Commentary  and proofs of Cartan's claims.}

1961: 

1962: 

1963: \subsubsection{Definition of $\g_2$} 

1964: 

1965: {\bf The Cartan subalgebra.} The first 3 operators of claim 1  are linearly dependent since

1966: $\sum_i X_{ii}=0$. This is the only linear relation (proof below) and explains why $\g$ is 14 dimensional and not 15 dimensional.  The flows

1967: of $3 X_{ii}$ generate the scalings

1968: $x_i \mapsto \lambda_1 ^{\alpha_i} \lambda_2 ^{\beta_i} \lambda_3 ^{\gamma_i}x_i$

1969: $y_i \mapsto \lambda_1 ^{-\alpha_i} \lambda_2 ^{-\beta_i} \lambda_3 ^{-\gamma_i}y_i$

1970: $z \mapsto z$ as described in section 5.  Hence these operators  should span the

1971: Cartan $\t$ of $\g = \g_2$. 

1972: 

1973: 

1974: 

1975: 

1976: \begin{proposition} $\g$ is a  14 dimensional   Lie subalgebra of $\End(V)$,  isomorphic to $\g_2$, with Cartan subalgebra as just described. 

1977: \end{proposition}

1978: 

1979: \pf It is convenient to put $\g$ in block matrix form. For each ${\bf u}\in\R^3$ let  $R_{\bf u}\in\End(\R^3)$ be given by $ {\bf v}\mapsto{\bf u}\X {\bf v}$; i.e. 

1980: $$R_{\bf u}=\left(\begin{array}{ccc}

1981: 0&-u_3&u_2\\

1982: u_3&0&-u_1\\

1983: -u_2&u_1&0

1984: \end{array}\right).$$

1985: 

1986: \mn Define the linear  map $\rho:\sl_3(\R)\X\R^3\X\R^3\to\End(V)$ by

1987: 

1988: $$

1989: \rho(A,\b,\c)=\left(\begin{array}{ccc}

1990: A&R_{\c} &2\b\\

1991: -R_\b&-A^t&-2\c\\

1992: \c^t&-\b^t&0

1993: \end{array}\right).

1994: $$

1995: %

1996: Now $\rho$ is clearly   injective, hence its image is a 14 dimensional linear subspace of $\End(V)$. 

1997: Denote the components of $A,\b,\c$ by $a_{ij},b_i,c_i$ (resp.), then  it is easy to check that 

1998: $$\rho(A,\b,\c)=-\sum_{i,j}a_{ij}X_{ij}+\sum_i b_iX_{i0}+\sum_i c_iX_{0i}.$$

1999: %

2000: This shows that $\g$ is the image of $\rho$ and hence a 14 dimensional subspace of $\End(V)$. 

2001: 

2002: \bn To show that $\g$ is a lie algebra one calculates that $$[\rho(A,\b,\c),\rho(A',\b',\c')]=\rho(A'',\b'',\c''),$$

2003: where

2004: \begin{eqnarray*}

2005: A''&=&[A,A']+3(\b\c'^t-\b'\c^t)- [\b\cdot \c'-\b'\cdot\c]I,\\

2006: \b''&=&A\b'-A'\b-2\c\X \c',\\

2007:  \c''&=&-A^t\c'+A'^t\c+2\b\X \b'.

2008: \end{eqnarray*}\\

2009: 

2010: 

2011: These formulae show that $\{\rho(A,0,0)|A\in\sl_3(\R)\}$ forms  a lie subalgebra of $\g$ isomorphic to $\sl_3(\R)$.  This

2012: subalgebra corresponds to the sum of the long root

2013: spaces in the root diagram, and  the Cartan subalgebra

2014: (the sum of the $X_ii$) as identified earliter.

2015: The formulae also show that 

2016: the images of the   $\rho(0,\b,0)$  and $\rho(0,0,\c)$ are stable  under the adjoint action of the Cartan, hence they must correspond to the remaining short roots.   

2017: 

2018: A tedious computation now yields the root diagram and the structure

2019: constants of $\g_2$.   

2020: \bn 

2021: 

2022: \subsubsection{Invariance of $J$}

2023: Let $G_2\subset\GL_7(\R)$ be the subgroup generated by $\g$. 

2024: 

2025: 

2026: \begin{proposition}  $J$ is $G_2$-invariant.  

2027: \end{proposition}

2028: 

2029: \pf This is equivalent to showing that every $X\in\g$ is $J$-antisymmetric, i.e. that $X$ anti-commutes with 

2030: $$

2031: \left(\begin{array}{ccc}

2032: 0&I/2&0\\

2033: I/2&0&0\\

2034: 0&0&1

2035: \end{array}\right).

2036: $$

2037: 

2038: 

2039: One now checks easily that the set of $J$-antisymmetric matrices consists of the matrices of the form

2040: $$\left(\begin{array}{ccc}

2041: A&R_\c&2\tilde\b\\

2042: -R_\b&-A^t&-2\tilde\c\\

2043: \tilde\c^t&-\tilde\b^t&0

2044: \end{array}\right),

2045: $$

2046: where $A\in\End(\R^3)$ and $\b,\tilde\b,\c,\tilde\c\in\R^3$. Looking at the formula for 

2047: $\rho(A,\b,\c)$ we see that $\g$ is the subset of the $J$-antisymmetric matrices 

2048: satisfying $\rm tr A=0, \b=\tilde\b,\c=\tilde\c$ (a codimension  7 condition).\qed

2049: 

2050: 

2051: \subsubsection{Invariance of the Pfaffian system.}

2052:  

2053: \mn  

2054: 

2055: 

2056: \n {\bf Generalities.} A ``Pfaffian system''on  a manifold $M$ is given locally by the common kernels of a finite set of 1-forms, 

2057: $$\alpha_1=\ldots=\alpha_m=0.$$

2058:  Two sets of 1-forms $$\{\alpha_1,\ldots,\alpha_m\},\quad \{\beta_1,\ldots,\beta_n\},$$

2059: give equivalent systems if one can express each element of one set as a linear combination 

2060: (with coefficients in $C^{\infty}(M)$) of the elements of the other set. We write this as 

2061: $$\alpha_i\equiv 0 \quad  \mod \beta_1,\ldots,\beta_n, \quad i=1,\dots, m,$$

2062: and similarly for the $\beta$'s. 

2063: 

2064: 

2065: 

2066: 

2067: Consequently, if we want to prove that a system is preserved by some diffeomorphism $f:M\to M$

2068: we must show that $$f^*\alpha_i\equiv 0\quad  \mod \alpha_1,\ldots,\alpha_m, \quad i=1,\dots, m,$$

2069: and if we want to show that the flow of some vector field $X$ on $M$ preserves the system 

2070: we must show that 

2071: $$\L_X\alpha_i\equiv 0\quad  \mod \alpha_1,\ldots,\alpha_m, \quad i=1,\dots, m.$$

2072: 

2073: Given such a system we can consider the common kernels $D_x\subset T_xM$ of the 1-forms at each point $x\in M$. 

2074: This is well defined independently of the 1-forms chosen to represent the system. If $\dim D_x$ (the rank of the system) 

2075: is constant we obtain a distribution $D\subset TM$ (a subbundle of the tangent bundle). But the rank may vary. 

2076: For example, the system on $\R$ given by $xdx=0$ has rank 1 at $x=0$ and rank 0 for $x\neq 0$.  However, if $G$ acts on $M$ preserving a Pfaffian system, then the rank must clearly be constant along the $G$-orbits. 

2077: 

2078: 

2079: \bn {\bf Cartan's Pfaffian system.    Rank jumps.  A  correction.}  Due to jumping of rank,

2080: as discussed in the last remark, the  Pfaffian system

2081: which Cartan defined by 

2082:   the vanishing of  the 6 components of $\alpha,\beta$

2083:  cannot be $G_2$ invariant, even when restricted to $\tC$,  the $J$ null cone.

2084: For  at $(\e_1,0,0)$ the system   reduces 

2085: to $dx_2=dx_3=dz=0$ and so has rank $4$. On the other hand, at

2086:  the point $(\e_1,\e_2,0)$ the system is equivalent to  $dy_1=dx_2=dz-dy_3=dz+dx_3=0,$ and  so has rank  $3$. 

2087:  And both points lie in  $\tC\setminus\{0\}$,

2088:  which is a single $G_2$-orbit, contradicting $G_2$ invariance.  

2089: A related  problem with Cartan's claim  4 of subsection (\ref{claims}) is his claim that $\gamma_1=\gamma_2=0$ is a consequence of $\alpha=\beta$. But this is true only on the $z\neq 0$ part of $\tC$. 

2090: 

2091: Both errors are fixed by  imposing  the extra equation $\gamma:=\gamma_1-\gamma_2=0.$  Then, as in section (\ref{connectOctonions}),

2092:  we do obtain a $G_2$-invariant system on $V$.    Furthermore, 

2093:  as proved immediately below, the two equations $\gamma_1=\gamma_2=0$ are indeed a consequence of $\alpha=\beta=0, \gamma=0$ on $\tC$, and are a consequence $\alpha=\beta=0$ on the subset $z\neq 0$ of $\tC$. 

2094:  So   Cartan's claim is correct on the open dense set $z\neq 0$ of the null cone $\tC\subset V$. 

2095: (See also page 11 of Bryant's paper on Geometric Duality \cite{Bryant}, where he adds the equation $\gamma=0$ to 

2096: $\alpha=\beta=0$.)

2097: 

2098: 

2099: 

2100: \begin{proposition} The Pfaffian system on $V$ given by $\alpha=\beta=0, \gamma=0$ is  $G_2$-invariant. 

2101: On $\tC$ the system is equivalent to $\alpha=\beta=0, \gamma_1=\gamma_2=0$. On the subset $z\neq 0$ of $\tC$

2102: it is equivalent to $\alpha=\beta=0.$

2103: \end{proposition}

2104: 

2105: \pf We prove the   claims of the last two sentences  first.

2106: Note that  $\gamma_1+\gamma_2=dJ$. It follows that on  $\tC$, where $J=0$, 

2107: we have that $\gamma_1=\gamma_2=0$ 

2108: is a consequence of $\gamma:=\gamma_1-\gamma_2=0.$ Thus, restricted to $\tC$, the system $\alpha=\beta=0, \gamma=0$

2109: is equivalent to $\alpha=\beta=0,\gamma_1=\gamma_2=0$. 

2110:  Next, note that $\x\cdot\beta-\y\cdot\alpha=z\gamma.$ It follows that on $z\neq 0 $ the equation $\gamma=0$ is a consequence of  $\alpha=\beta=0$.\qed

2111: 

2112: It remains to establish invariance. 

2113: We need to show that $$\L_X\alpha_i\equiv\L_X\beta_j \equiv\L_X\gamma\equiv 0 \mod \alpha_i,\beta_j,\gamma,$$

2114: for all $X=\rho(A,\b,\c)\in\g.$ 

2115: Divide into  3 cases, corresponding to $(A, 0,0)$ , $(0, \a, 0)$ and $(0, 0, \b)$

2116: in our coordinatization of $\g$. 

2117: \begin{itemize}

2118: \item case 1: $X=\rho(A,0,0),$ $A\in\sl_3(\R)$.

2119: 

2120: \begin{lemma} If   $A\in\End(\R^3)$ and ${\bf u},{\bf v}\in\R^3$, then 

2121: $$A({\bf u}\X {\bf v})+A^t{\bf u}\X {\bf v}+ {\bf u}\X A^t{\bf v}={\rm tr }A({\bf u}\X {\bf v}).$$

2122: \end{lemma}

2123: \pf Sketch: divide in 2 cases.  If  $A^t=-A$ then ${\rm tr }A=0$ and the identity is a consequence of the fact the 

2124: $\SO_3$ preserves de cross product and that $\so_3$ are the antisymmetric matrices. If $A^t=A$ then can assume w.l.o.g. that $A$ is diagonal and do  an explicit easy calculation. \qed

2125: 

2126: 

2127: Now since $$X(\x,\y,z)=(A\x,-A^t\y,0), \quad \alpha=zd\x-\x dz+\y\X d\y,$$

2128: we get, using the lemma and  ${\rm tr }A=0$, that

2129: \begin{eqnarray*}

2130: \L_X\alpha&=&zAd\x-A\x dz-A^t\y\X d\y-\y\X A^td\y=\\

2131: &=&A(zd\x-\x dz+\y\X d\y)=A\alpha\equiv 0\quad \mod \alpha.

2132: \end{eqnarray*}

2133: Similarly, $\L_X\beta=-A^t\beta\equiv 0\quad (\mod \beta).$

2134: 

2135: Finally, $\L_X\gamma=(A\x)\cdot d\y- \x\cdot (A^t d\y)=0.$

2136: 

2137: \item case 2: $X=\rho(0,\b,0),$ $\b\in\R^3$.

2138: 

2139: Here $$X(\x,\y,z)=(2\b z, -\b\X  \x, -\b\cdot\y),   $$

2140: and one calculates that 

2141: $$\L_X\alpha=\b\gamma,\quad \L_X\beta=\b\X\beta,\quad\L_X\gamma=\b\cdot\beta.$$

2142: 

2143: 

2144: \item case 3: $X=\rho(0,0,\c), \c\in\R^3.$ The proof for this  case is very similar to the previous case.  Just interchange $\x$ and $\y$, and $\b$ and $\c$.

2145: \end{itemize}

2146: This completes the proof of invariance,

2147: and hence the proof of the proposition.

2148: 

2149: 

2150: 

2151: %{\bf Caveat. Rank jumps.  The $\gamma_i$.}

2152: %Cartan's Pfaffian system from claim 4, defined

2153: %by the vanishing of $\alpha$ and $\beta$   does not   agree with our

2154: %$D$  on the locus $z =0$ , due to the fact that its rank jumps along this locus.  This discrepancy is corrected by imposing   $\gamma_1 = \gamma_2 = 0$.

2155: %Contrary to Cartan's assertion, these last  equations are not a consequence of $\alpha =\beta = 0$ when$z = 0$.

2156: 

2157: 

2158: 

2159:  \begin{thebibliography}{9}

2160:  

2161:  \bibitem{Agrachev}

2162:  Andrei Agrachev, {\em  Rolling balls and Octonions},

2163:  math arXivs: 	math.OC/0611812.

2164:  

2165: \bibitem{Serre} J.P.~Serre,

2166: {\em Complex semi-simple Lie algebras},

2167: Springer,  reprinted and translated from

2168:  {\em Alg\`ebre de Lie semi-simple complexes}, W. A. Benjamin

2169:   (1966).

2170: 

2171:   \bibitem{BryantHsu}

2172:  {R. Bryant and L. Hsu},

2173:   {Rigidity of Integral Curves of Rank Two Distributions},

2174:   {Invent.\ Math.},

2175:     vol 114,

2176:     p.  {435--461},

2177:      1993.

2178: 

2179: 

2180:   \bibitem{Bryant}   Robert L. Bryant,

2181: Elie Cartan and Geometric Duality,

2182:       lecture notes from a lecture given at the Institut dÕHenri Poincare ,   on 19 June 1998;

2183:     available on Bryant's website.

2184: 

2185: 

2186:  \bibitem{C1}  E. Cartan,  {Les syst{\`e}mes de {P}faff {\`a} cinque variables et l{\`e}s

2187:                {\'e}quations aux d{\'e}riv{\'e}es partielles du second ordre},

2188:   {Ann.\ Sci.\ {\`E}cole Normale},

2189:     vol. 27,   no.  3,  p. {109--192}, 1910.  (Reprinted in Oeuvres completes, Partie III, vol 2.  137-288.)

2190: 

2191: 

2192: \bibitem{C2} E. Cartan, {Sur la structure des groupes de transformations finis et continus}, These, Paris, 1894.  (Reprinted in Oeuvres completes, Partie I, vol 1.)

2193: 

2194: 

2195: \bibitem{C3} E. Cartan.  Le groupse reels simples finis et continus, Ann. Ec. Norm. v. 31, pp 263-355, (1914).  (Reprinted in Oeuvres completes, Partie I, vol 1.)

2196: 

2197: 

2198: \bibitem{Chern} S.S Chern and C. Chevalley, {Elie Cartan and his mathematical work}

2199: (Reprinted in Oeuvres, partie III vol. 2. )

2200: 

2201: 

2202: \bibitem{Hammersley} J. M. Hammersley, {\em Oxford commemoration ball},

2203: in {\bf Probability, Statistics, and Analysis}, London Math. Society, Lecture Note

2204: Series, vol. 79, Cambridge U. Press, [1983].

2205: 

2206: 

2207: \bibitem{Harvey} Reese

2208: Harvey, {\bf Spinors and Calibration}, Acad. Press, 1990.

2209: 

2210: \bibitem{Kaplan} Aroldo Kaplan  and F. Levstein, {\em A split Fano plane}

2211: in preparation.  

2212: 

2213: 

2214:   \bibitem{Montbook}{R. Montgomery} {\bf A tour of sub-Riemannian Geometry},

2215:   AMS, 2001.

2216: 

2217: \bibitem{Serre}  J-P. Serre, (translated from the French by G.A. Jones)

2218:  {\bf Complex Semisimple Lie algebras}, Springer-Verlag. 

2219: 1987.

2220: 

2221: \bibitem{Tanaka} N.  Tanaka, {\em On the differential systems, graded Lie

2222: algebras and pseudo-groups},   J. Math. Kyoto Univ. , v. 10, 1-82, 1970.

2223: 

2224: \bibitem{Tanaka2}N. Tanaka, {\em On the equivalence problem associated

2225: with simple graded Lie algebras},  Hokkaido Math. Journal,

2226: v. 8, 23-84, 1979.

2227: 

2228: \bibitem{Vogan} D. Vogan, {\em The unitary dual of $G_2$}, Inventiones, vol. 116, 677-791,

2229: (esp. p. 679), 1994. 

2230: 

2231: \bibitem{Yamaguchi} K. Yamaguchi, {\em Differential Systems Associated with Simple graded Lie Algebras},  in {\bf Progress in Differential Geometry},  Advanced Studies in Math, v. 22, 413-494, 1993. 

2232: 

2233: 

2234: \end{thebibliography}

2235: 

2236: 

2237: \end{document}

2238: 

2239: 

2240: