Progress in Mathematics, Vol 144, © 1990 Birkhauser Verlag Baselwitrettnd Carnot-Carathéodory spaces seen from within Mixuass. Growov" Contents §0. Basic definitions, examples and problems ++ 85 1. Polarizations, horizontal curves and Carnot-Carathéodory metrics 85 2. Basie contact example 86 A. Connectivity theorem for the contact polarization H 86 A’. Contact C-C metrie on (V; H) 87 B. Internal versus external in C-C geometry 88 3. Heisenberg group view on the contact exam 89 A. Lie group theoretic proof of connectivity 90 A’. Connectivity theorem for general Lie groups 90 90 mal versus asymptotic 92 D. Self-similar spaces appearing as tangent cones of equiregular ones 93 DY. Pansu convergence theorem im D”. Mitchell cone theorem for equiregular spaces 95 4, Chow connectivity theorem 95 5. The shape of C-C balls Mitchell-Gershkovich-Nagel-Stein-Wainger theorent 97 A. Ball-box-theorem 98 B, Holder exponent evaluation 0 C. Holder mapping problem 0 D. Holder surfaces in contact 3-manifolds 99 E. Homotopy count of Hélder maps 100 6. The volume of C-C balls and the Hausdorff dimension 101 A. On the intrinsic size of C-C balls 102 B. Hausdorff dimension of submanifolds 103 C. Hausdorff dimension problem for subspaces 105 hank Richard Montgomery for reading the manuseript ‘of errors. “Acknowledgment. and locating n multi0 MIKHAEL GROMOV. 7. Isoperimetrie filling problem A. Isoperimetric inequality for k = B. Filling in eurves in V ©, Filling for dim $> 1 8. Carot-CarathGodory metrics as limits of Riemannian ones A. Ricimannian homogeneous spaces and their Limits B. Contact example in the spherical clothing, C. The asymptotic mapping problem D. The intermediate volume problem E. Families of metrics associated to dynamical systems F. The intermediate entropy problem G. Tntrinsie approximation of C-C spaces by Riemannian 9, Conformal C-C geometzy and hyperbolic geometry 10. Dimension and growth in the asymptotic geometzy $1. Horizontal curves and small C-C balls. 1. Proof of the Chow connectivity theorem ‘A. Quantitative version of the Chow theorem for deg.X, <2 and a preliminary Hélder bound on the C-C metric AV. Upper box bound on O-C dist for deg X, <2 B, Lower box bound on C-C dist for deg X, <2 €. Corollary: Ball-box theorem for deg 2. A new proof of the Chow theorem and the Hilder bound on the C-C metric for arbitrary degree A. Upper box bound ou the C-C distance for arbitrary degree ACh Lemma B. Making “smooth” instead of “piecewise smooth” in the Chow connectivity theorem for polarizations HC T(V) 3. Lower box bound on the C-C distance for deg > 2 A. Doubling and covering properties for balls; equisingularity and the Hansdorff dimension 4, Canonical coordinates, almost Lie groups, ullpotent tangent cones and a sharp version of the ball-box theorem for equiregular polarizations A. Convergence Proposition AY. Uniformity of the convergence AY. On C’-convergence 105 105 106 107 108. 109 109 109 110 110 0 10 1 uz 2 13 4 15 16 u7 us 120 120 120 WI 128, 130 1 134C-€ SPACES SEEN FROM WITHIN st Approximation of equiregular Carnot-Carathéodory, spaces by self-similar nilpotent groups 135 C. Pinching and related problems for C-C metries 140 1D. Riemannian and piecewise Riemannian approximation of C-C motries 144 D’. Approximation by nerves 146 D". Anisotropic blow-up of (V, #1) Ms E, Smoothing Lipschitz functions on C-C manifolds 149 EY, On the homotopy count of Lipschitz maps 150 5. Anisotropic metries beyond Carnot-Carathéodory 151 §2. Hypersurfaces in C-C spaces. - ae 1. A lower bound on the Hausdorff dimension of a hypersurface WeVv 152 A. A systolic bound on mesy_1 V’ of (+) 16a 2. Lower systolic bounds for families of metrics on compact manifolds V A. A lower systolic bound via closed horizontal (n — 1)-forms 3, Isoperimetric inequality 159 A. Flow tube estimate for small C-C balls 159 A’. Corollary: Flow tube estimate for hypersurfaces 160 B. Measure moving lemma 161 B', Corollary: Local isoperimetric inequality 161 C. Vitali covering lemma 162 D. Isoperimetric inequality in compact regions Vp CV. 163 DY, Excluding Va from the game 164 1D!" Isoperimetric inequality in non-equiregular spaces 166 E. Green forms, pencils of curves and an integral geometric proof of the isoperimetric inequality 107 4, Singular integrals and Sobolev inequalities 170 A. Dilation and homotopy a B, High levels of functions f with df | H|1.. <00 on (codim 1)-stable C-C manifolds of (formal) Hansdorif dimension N_ 172 5, Homotopy bounds by [Df | H|zyes taut maps and the bubbling phenomenon 174 A. Weak stability of homotopy classes lis B. On the topology of the space F, of maps f with |DF |My Se 179 C. Surface maps of small area and related questions 179s2 MIKHAEL GROMOV D. On the homotopy role of the L4-norn of the differential Df on H for q n—1/2 198 2, Polarizations with degenerate curvature w 199 43. Differential forms and straight Alexander-Spanier cocycles 200 A. Rumin complex 201 B. Construction of straight Alexander-Spanier cocyeles with controlled growth at the diagonal 202 4. Width and filling radius 204 A. A bound on wid by mes 204 B. Contact triangulation 205 BY. Integral-geometric intersection incquality 205 C. Asymptotic Riemannian version of 3.4.4 206 5. Lipschitz maps of Riemannian manifolds into contact C-C ‘ones 207 A, First Lipschitz approximation theorem 209 AV. Lipschitz approximation of families of maps W—> V 2u B. Extension of piecewise horizontal maps 212 C. Smoothing, Lipschitz maps 215 CC’. Sccond Lipschitz approximation theorem 215 D. Construction and extension of non-piccewise stavoth Lipschitz. maps 216 Controlled integration of differential forms and bounds on the rational homotopy invariants of maps 200 6.C-€ SPACES SEEN PROM WITHIN A. (LyA*)-energies in the contact case A’. Controlled integration of the Rumin complex B. Controlled integration and filling in Riemannian manifolds V BY, Thick families of filling-in codimension > 1 in Riemannian manifolds V_ C. Filling-in curves in Riemannian manifolds and ||A®D) CC’, Nantow eurves proposition D. Thick filling of horizontal curves in contact manifolds E, On the global contact geometry §4. Pfaffian geometry in the internal light 1. A brief metricly guided Pfaftinn tour A. The H-filtration and the type numbers n A’. On local connectedness of smooth submanifolds B, Submanifolds V" in V of a given type and Thom horizontal homology B', On the type of » morphisms C. Pfaffian systems in jet spaces D. Horizontal chains and cycles E. Horizontal forms and cohomology EY, Intrinsic metric evaluation of horizontality of forms 2, Analytie techniques for local construction of integral (H-borizontal) submanifolds A. Q-regularity and infinitesimal invertibili AN’. Local h-principle A”. Dimension count for regular isotropie subspaces B. Calculus of variations for regular horizontal submanifolds C. Partially horizontal submanifolds in V he global h-prineiple for smooth. horizontal submanifolds A. On the /-prineiple for morphisms of a given type 4. Folded integral submanifolds ‘A. The proof of Thom’s theorem (A) in the jet bundle B. Extension of folded and subfolded maps C. Deformations of regular horizontal folded maps: 5. Lower bounds ou the Hausdorff dimension of subsets in tal triangulations and a hound on the width of subsets in nilpotent Li 7. Lipschitz ma st 223 225, 26 227 229at MIKHAEL GROMOV. A. Construction and extension of Hélder maps 270 8. Dehn isoperimetry in nilpotent Lie groups 270 A. Hdlder maps D —+ V and isoperimetric inequalities of degree 2 B. Regular (i, -surfaces in V and isoperimetric inequalities of degree i+ j ©. On filling in dimension > 3 9. Metric properties of submanifolds partitions and maps A. Parabolic metrie spaces 10. Anosov endomorphism A. Entropy in codimension 1 B, Entropy in high codimension 11. Horizontal forms on polarized manifolds A. Horizontality via the anisotropic blow-up B. Horizontal cobomology on nilpotent Lie groups and algebras BY. Horizontal eyeles and coeyeles on ni-manifolds V/P BY, Harmonic forms and anisotropic blow-up of V/T C. Lower bound on the Minkowski dimension of eyeles in nncous spaces (V, 1) 298 compact homes Small harmonic forms and a lower bound on dimyiin tuon-compact homogeneous spaces C, On the limit of Pye for e +0 §5. Anisotropic connections 1. Curvature Qe A. On the equation 2¢ = 0 Norms and metrie associated to orthogonal connections, 3, Lardistanee in the space of connections A. On Ly-non-flatness of V B. Non-flatuess: measured by horizontal monodromy and curvature 3i2 C. Radial gauge fixing 313 4. Geometric and topological effects of the bound glu, 0. (Here one may take C= V4), ‘This equivalence makes? and dist, rather non-interesting from our present point of view. 0.1, Polarizations, horizontal curves and Carnot-Carathéodory metrics. A polarization of a manifold V is, by definition, « subbundic of the tangent bundle, say HC L(V). One may think of Has a dis tinguished set of directions (tangent veetors) in V- which are called in sequel horizontal. (This terminology is motivated by the picture where V is smoothly fibered over some manifold B and H is normal to the fibers.) A piecewise smooth eurve in V is called horizontal with respect to A if the tangent vectors to this curve are liorizontal, ‘The metric defined with (the set 1 of) the horizontal curves in V is called the Carnot-Carathéodory metric associated to H and denoted86 MIKHAEL GROMOY 50 distyy. Notice that the definition of distyy also involves an auxiliary Rie- mannian metrie g as dlistu(o,22) = inf(y-tengths of H-horizontal curves between vy and ¥), but the effect of g on dist; is non-essential from our point of view. Namely, two metrics defined with different g, and gy and some are bi-Lipschitz eqnivalent (on each compact subset in V7). On the other hand, the role of AZ is crucial as is seen in the following 0.2. Basic contact example. Let V = R* and H be the standard contact subbundle, which is the kernel of the (contact) I-forma 9 = d=-+edy on RS. This means that the tangent space (plane) Hy < Ty, (R3) = B3 is given at each vp = (2o,4o,70) € R® by the equation = + xoy = 0. Notice, that AT is generated by the following two independent veetor fields, = ff ond O = 2 — 1%. These fields do not commute. In fact, their coummutator equals ~ j£ and so the three fields 2, 82, and the Lie bracket [01.02] span the tangent bundle T(R*) at each point » € R*, 0.2.A, Connectivity theorem for the contact polarization H. Theorem. Bvery two points in BY can he joined by a smooth H- horizontal curve. Proof. Tak (e(t),y(), #€ [0.1], in the (x, y)-plane join- ing two given points (e1,¥1) and (ra, 2) and such that the formal area “bounded” by c, defined by the integral fardy = [3 s(A)y’(f)dt, equals. a given number a. (One easily finds sue c, say among curves of con stant curvature). Then we take the horizontal lift of ¢ = (x(t), y(t) to RY by letting 2(t) = 21 — fyx(t)y'(t)de for a given value 2 of 2. The lifted curve e = (2(t), y(t), 2(t)) is indeed horizontal as d2(t) = 2 (Qt = —a(t)y!(t)dt = —a(O)y(t) and it joins the given points (r),9.21) and (242,20 = 21 +) . Historical Remarks. ‘This result (which seems obvious by the mod- em standards) appears (in a more general form) in the 1909-paper by Carathéodory on formalization of the classieal thermodynamics where horizontal curves roughly correspond to adiabatic processes. In fact, th above proof may be performed in the language of Carnot (cycles) and for02 © SPACES SEEN FROM WITHTS ar this reason the metrics disty wore christened! “Carnot-Carathéodory” in [G-L-P]. T suspect that some form of the connectivity theorem was known to Lagrange in the framework of the nonholonomic mechanics. (Compare [Ve-Fa] and [Ver-Geria]: an instance of a nonholonomic system is given by a billiard ball rolling on the plane, such that the velocity at the low- est point of the ball where it touches the plane must be zero. Here V equals the configuration space, that is R?x SO(3), and the nonholonomic constraint (on the velocity) is represented by subbundie Hc T(V) of rank 3. Now, a child can roll the ball from one given position to another thus proving the connectivity property for H.) Various forms of the con- ivity theorem have been persistently appearing, in the literature but Thave not tried to keep track of them. By tradition, general commonly uused connectivity theorem (see (1.4.) is attributed to Chow (soe (Cho), though his paper was neither the first (see, e.g. [Rasb]) nor the last on that matter. Finally we notice that the phenomenon of H-connectivity is also seen in the theory of optimal control and in roboties where it is named “controllability” (see [Brock)) 0.2.4’, Contact C-C metric on (V,H). Now, let us look more closely at the C-C metric (C-C = Camot-Carathéodory) associated to contact. subbundle HC T(R*), We recall that the H-horizontal curves c in R¥ are the lifts of curves in the (,y)-plane, such that the 2-coordinate of c ‘equals the formal area of the (2, y}-projection ¢ of c. If two points v and ‘vy in R* ie on the samme vertical line (or 2-Hine), i. have equal (x, y)- coordinates, then the (z,y)-projections ¢ of curves ¢ joining these points are closed in the (z,y)-plane and so the (formal) area af theso curves ¢ is bounded by area¢ < const(length c)* < const(lengih ¢)?, where const = (4r)~'. It follows that the C-C distance between v and vg is bounded fiom below by the Euclidean distance as follows, dist yy > const} (distr)! w Since the Euclidean distance diste,, between our points equals 2) — aareac. One also has the upper bound dist < const” dist,ss MIKHAEL GROMOV 40 where @ is a certain positive function depending on the Euclidean norms of the points vy and vz. (One may take a = (1 + [lv] + llo2ll)*). In fact, one can join v and v2 by a curve ¢ in R* which projects to a cirele ¢ in R?, such that dist = areae = (47)"(lengthie)® while distyy (te; ty, 122) for the (:r,4y,2}-basis, where = is central and [x,y] these are, clearly, automorphisins and they exponentiate to automorphisms of G, called Ar: G+ G. These Ar preserve HT (which is generated by «x and y) and they scale the Riemannian metric on H (but not on all of T(G)) by 4 Therefore, the length of all horizontal (i. tangent A’) curves is also scaled hy t and then dlstyr is sealed by as well . General self-similar Lie groups. Let G be a simply connected nilpo- tent Lie group where the Lie algebra L admits a grading, L = @%, Li, such that (Z,, Lj] C L,4). Then the operator a : LL, t € R%, defined by a f= té for £ € L, are automorphisms of L. These a, integrate to a L-parameter group of automorphisms A; : G—+ G, which are similarities for the Carnot metric defined with the polarization corresponding to L1. (Here as earlier we should assume that L, Lie generates L im order to have dist < 00.) ‘Two-step example. Let G be a two-step nilpotent group. Then L can be graded with Lz = [L,L) and some subspace Ly CE complementary to La. This Ly obviously Lie-generates Zand thus gives us a self-similar C-C metric on G, Remarks. If V is a homogeneous Riemannian manifold which ad- mits & non-trivial self-similarity, then V, clearly, is isometric to R" for some 1. Analogous non-Riemannian examples are provided by Banach spaces of finite or infinite dimension but the self-similar nilpotent Carnot Carathéodory manifolds do not spring in one’s mind so readily. One knows now-a-days that there are no additional essential examples among finite dimensional manifotds (see Be-Ve]) but there are interesting #nfinite di- mensional and/or disconnected homogencous self-similar metric spa Infinite dimensional and totally disconnected examples (a) Let L = @#.,L; be an infinite dimensional graded nilpotent Banach- Lie algebra which is (algebraically) Lie-generated by Z;. Then the corresponding nilpatent Banact-Lie group cntvies a natural homoge- neous self-similar C-C metric. The simplest example is that of the infinite dimensional Heisenberg group for L = L; @ Lo, where Ly is the Hilbert space and L2 = center L = R.92 MIKHABL GROMOV fo (b) Every field with a norm, e.g. the field of padie numbers, is homoge- neous and self-similar, Furthermore, vector spaces and suitable nilpo- tent groups over such fields share this property. (c) Let @ be a topological group generated by a family of I-parameter subgroups. We give a metric to each of these subgroups by choosing parameters and then we define the corresponding C-C metric in as the maximal (or “supremal”) left invatiant metric for which the inclusions of all our I-parameter subgroups into G are I-Lipschit2, i.e. (non-strictly) distance docreasing maps. (This is similar to the definition of the word metric in G with a given generating subset). If G admits a (dilating) automorphism A: G — G, whieh preserves the distinguished subset. of subgroups and dilates the corresponding, parameters with a fixed constant A> 1, then this A also dilates the C-C motric by A, which means A is a nontrivial self-similarity. An especially interesting class of such self-similar groups is associated to infinitely graded Lie algebras L = PL; which are Lie-generated by Ls. (This was pointed out to me by I. Babenko.) 0.3.C. Uses of self-similarity: infinitesimal versus asymptotic. If a metric space V admits a non-trivial (Le. non-isometrie) self-similarity fixing a point vm € V then all of geometry of V can be seen in an ar- bitrary stall neighbourhood of up. In particular, if V is a C-C manifold then the asymptotic geometry of V (at infinity) can be read in terms of ‘esimal data of the implied polarization at vp and vice versa. in Let us apply this to the Heisenberg group G with the dilations Ay (defined by (r,y,2) > (tx, ty, 22) on the Lie algebra L = L(G)) and determine an approximate shape of the Carnot-Carathéodory balls in G. Denote by B'(p) cL the box defined by the inequalities =e el 0) and some constant C > 0.03 (-C SPACES SEEN FROM WITHIN 99 ‘The relation (+) for p — 0 justifies our earlier elaims (sce 0.2.4’) on the shape of the small C-C balls associated to the contact structure. On the other hand, the same (+) for p —> oc provides an asymptotic information on the ordinary Riemannian left invariant metric in G. Namely the balls in this metric, call them B*(p) ¢ G, (obviously) have the same asymptoties for p+ 20 as the balls B(p) (since the two metric involved are both left invariant and determined by length of enrves). In fact Bip) < B*(p) < B(C*p), for all p> 1 and some constant C* (where the first constantless inclusion is due to the fact that the Riemannian metrie we speak about: is the one which underlies the C-C metric and so the corresponding Riemannian distance < C-C distance). It follows that the Riemannian balls £ (p) are asymptotically approx- imated by the (exponentiated) boxes B'(p). In fact, this remains valid for all simply connected nilpotent Lie groups with graded Lie algebras by the above self-similarity argument. (An arbitrary simply connected nilpotent Lie group G is axvmptotic to a group Gro which does admit: a self-similarity by a theorem of Pansu cited in Remark (b) below. Thus the large balls in G are box-shaped as well as those in Gao. Also see [Bass] and [Kari] on this matter.) 0.3.D. Self-similar spaces appearing as tangent cones of equireg- ular ones. Let V = (V,dist) be a metric space with a reference point vv. We set £V = (V,tdist) for t €]0,oo{ and we want to go to some limits of (V for £ + 0 and for ¢ + 00. An appropriate notion of a limit for our present purpose is the one associated to the Hausdorff topology on the “set” of isometry classes of pointed metric spaces (see [Grocec] and [G-L-P]). If such a limit exists for t + co it can be thought of as the tangent, space (cone) of V at uy while the limit of ¢V for t + 0 looks like the asymptotic cone of V or the tangent space of V at infinity. Notice, that the very definition of the limit makes these tangent cones (spaces) self-similar whenever they exist, On the other hand, if V already aciits t-self-similarities fixing vy for all ¢ > 0 then (tV, vp) is isometric to (V, m) for all ¢ > 0 and so the tangent cones to V at ug and at infinity exist and are isometric to V.94 MIKHAEL GROMOV Pt Examples (i) ICV is a Riemannian manifold then the Hausdorif limit exist and it equals the ordinary tangent space T,4(V), (ii) Let G be the Heisenberg group with a left invariant Riemannian met- rie g and set gp = f-'Aj(q) for the above Ay. Tt is obvious, that the distance function disty associated to g converges (in the usual sense) to the Camot-Carathéodory metric dist. on G for t+ oc. On the ther hand the space (G, disty) is isometsic tot-1G = (G,t- dist). It follows that. the Hausdorff limit of the spaces 1G for { — 0 also ex and is isometric to (G,distx.). Hence, the asymptotic (tangent) com of the Riemannian manifold (G,dist) equals the Carnot-Carathéodory space (G, iste). nm 1V does Thus the (infinitesimal data of the) C-C geometry ean he recap- tured from the asymptotic Riemannian geometry of G. Remarks on C-C limits of discrete groups. ‘The above example ap- pears in [Grogpe] in the surrounding of diserete groups G of polynomial grouth. The polynomial growth ensures the existence of a Hausdorff sub- limit of tG, t + 0, an asymptotic cone, which is a nilpotent Lie group with a C-C metric where the degree of the growth of G translates to the Hausdorff dimension of the limiting C-C metric. In fact, the asymp- totic geometry of discrete nilpotent groups provided the major source of inspiration for the initial study of Canot-Carathéodory spaces, 0.3.D’, Pansu convergence theorem, The existence of an actual Hausdorff limit (not just a sublimit) lim #G was proven by P. Pansu (sce [Pancns]) for an arbitrary nilpotent Lie group with a Teft invariant Riemannian metric. (In fact, Pansu allows in [Pancpy| a more general class of spaces G including discrete virtually nilpotent groups with word metties.) Pansu shows that such a limit is isometric to a nilpotent Lie group Gac (which is, in general, not isomorphic to G) with a self-similar G-C metric. Thus the asymptotic geometry of every nilpotent group G reduces to the local O-C geometry of Gao. In particular the asymptotics of the p-balls in G for p + 90 is encoded in the (infinitesimal) behavior of prballs in Gap as p— 0.on C-C SPACES SEEN FROM WITIIN 95 0.3.D”. Mitchell cone theorem for equiregular spaces. Consider -C mettic in V defined with a polarization HC T(V) which satisfies the following (xenericity) assumption. Choose some tangent vector Fields in H which span H and denote by H,(v) CT.(V), v € V, the subspace in T,(V) spanned by all commutators of the chosen fields of order < i. Clearly, this H, does not depend on the choice of the spanning fields. Equiregularity Assumption. The dimension dim Hj(v) is constant in v € V foreach i. proves in [Mity,2| under this assumption that the Carnot Carathéodory metric in V (for an arbitrarily given underlying Riemannian metic) admits the tangent cone at each point uy € V (ie. the Hausdorff Jimit Jim ¢V exists) and this cone is isometric to some self-similar nilpo- tent Lie group (see 14.) ‘This shows that the local Carot-Carathéodory geometry essentially reduces to (asymptotic) geometry of a nilpotent group with a dilation, (One can imagine by looking at Pansu and Mitchell theorems that there an “inversion 1 —+ t!" which interchanges local C-C spaces with nilpotent Lie groups, stich that the “fixed point set” of this “inversion” consists of the self-similar groups.) 0.4. Chow connectivity theorem. (see 1.1), Theorem. Let X1,...,Xm be C™-smooth vector fields on a connected manifold V,, such that successive commutators of these fields span each tangent space Ty(V), v © V. Then every two points in V can be joined by a piecewise smooth curve in V where each piece is a segment of an integral curve of one of the fields X;. Remarks and corollaries (a) Lie group motivation. This theorem is quite obvious on the for- mal level. Tn fact, let: L be the Lie algebra generated by the fields Xy, i= 1...,m, and GC DIV be the subgroup of diffeomor- ms generated by the one-parameter subgroups corresponding to Xai sm. The theorem claims that G is transitive on V-pro- vided L spans T(V). This is immediate if L is finite dimensional as I, can be identified with the Lie algebra of G (which makes G finite d mensional as well) and then the “L spans T(V)"-condition amounts to surjectivity of the differential of the orbit map G + V for g + (vo)6 MIKHARL GROMOV fo for every given point vy © V. In fact, this argument applies in the infinite dimensional case as well (see 1.1). (b) Polarizations. If the dimension of the span Hy © T,(V) of the fields X, atv is independent of v, the span of these fields is a subbundle in T(V), ic, a polarization of V in our sense where the orbits of X; are horizontal, Thns Chow theorem implies the connectivity property for H-horizontal enrves mentioned earlier (©) Generic fields. The assumption of the theorem is satisfied by generic C™-ficlds whenever m > 2 as successive commutators of two generic C*-fields on V span T(V) as is easy to show, On the other hand, the conclusion of the theorem may be satisfied by some non-smooth cou- tinuous fields where the commutators are not. even defined. In fact, the comnectivity property is valid for a single (!) generic continuous field as everybody knows (compare ITS in [Hart]). Probably (this seems obvious) a generic pair of C*-fields have the connectivity property for every &, (4) Polarizations defined by 1-forms. As we mentioned in (b) the Chow theorem applies to polarizations HC T(V) viewed as spans of systerns of vector fields. But one also can define Has the common zero set of a system of L-form on V, This suggests a dual approach to (the proof of) the connectivity property of H which does not directly use orbits of vector fields tangent to H but rather appeals to leaves of 1- dimensional foliations obtained by intersecting A with submanifolds W CV with coclimW = rank H — 1. For example, if H is a contact subbundle on 3-dimensional manifold Vand Wy C V is a curve transversal to H, one takes 2-dimensional cylinders W. CV, ¢ > 0, around Wy and HV T(W-) give us (spiral) curves in W. tangent to HT whieh elosely follow Wj for small e, see Fig, 1 Pigure 1os (-C SPACES SEEN FROM WITHIN ” 0.5, The shape of C-C balls: Mitchell-Gershkovich-Nagel-Stein- ‘Wainger theorem. Let HC 7(V) be a smooth polarization (ie. sub- bundle) spanned by some vector fields Xy.....Xyne (Every Hf of rank JT ny can be spanned by m < m+n fields for n = dim V and locally one needs only m = ny fields. In fact, our considerations are local and 30 m= ny suffices.) We denote successive commutators of our fields by X, tor suitable indices i > m and we assign to each X, the mumber deg X which is the degree of the corresponding commutator. Thus, for example, dog X= 1 i Sm, and deg Xy = 2 Xp = [Xp Ny] 1S wy ve Sm. (It may happen that X; = X; for i # j and, moreover, deg(X:) # dea(X)) for X, = Xj, ic. the degree is assigned to the name (i.e, the subindex) of a field rather than the field itself.) Now, let Xj, 7 = 1j..4yM, be the successive commutators of X1....,Xyp of degree < d and let us assume that these Xy,7=1,...,M, span 7(V). We want to characterize the C-C metric in V correspond- ing to H in terms of the integral curves (J-parameter subgroups) of the fields X,, i= 1....,,and of the linear combinations of these fields with constant coefficients, First, for an arbitrary field X ou V, we recall the notation exp, X © V which is the result of applying to v the I-parameter group (flow) corresponding to X at the time ¢ = 1. In other words, we obtain vy = exp,(X) by taking the integral curve of X issuing from w and following it with speed X for time t = 1. (If V is non-compact, the ficld may be globally non-integrable but this is irrelevant. at the moment.) Next, we define the exponential map exp, : RM —+ V by (ty This is, clearly, a smooth p (as we assume X, are smooth) and the differential of exp, is surjective at the origin 0.€ R™ stu) > expyltr Xi +--+ ta Xa). Notice that if V is non-compact, the map exp, : IRM —+ V need not: be M defined on all of BY! (as some field X = t,X; may fail to be globally integrable) but it is always defined in a small neighbourhood of the origin of R™ and this is all we need for our purpose. Consider the following box in R™, Box(p) = {| < p™***. a MPCR™,os MIKHAEL GROMOV 50 0.5.A. Ball-box-theorem. (*) ‘Theorem. The small C-C balls B,(p) in V around v € V are uniformly equivalent to the exponential images of the boxes. This means, there are strictly positive continous functions C = C(w) and py = py(0), stuch that exp, Box(C~!p) © By(p) C exp, Box(Cp) for all v € V and p < po(v). Remarks and corollaries (a) Hélder equivalence of O-C and Euclidean. The first inclusion By(p) > exp, Box(C~'p) gives us a quantitative version of Chow's theoret as it shows that every point vy in the exponential image of Box(C~!p) can be joined with by an H-horizontal curve of length < p. Since the exponential map has surjective differential at the origin, the image of this box contains the Euctidean (or Riemannian) ball in V around v or radius ~ p* (where, recall, d = max deg X;). It follows that the identity map (V, Riemannian metric) “4 (V, C-C metric) is C*- Holder with the exponent «= d~). (Notice that the map (V,C-C metric) sy (V, Riemannian metric} is, obviously, Lipschitz.) (b) Let us generalize the above and determine the Holder class of a smooth map of a Riemannian or, more generally, Carnot- Carathéodory manifold W into V. ‘To simplify the matter we assume our polarization H is equiregular in the sense that the dimension of the subspace in T,(V) generated by the commutators of degree j < d oes not depend on » (compare Mitchell cone theorem cited in 0:3.D). In this ease the bundle T(V) is filtered by smooth subbundles HHL C Hy C- C HCC Hy =TV) such that # is spanned by the j-th degree commutators of the fields in H. Then we take some polarization H’ C T(W) whose successive 2 Soo [Mitya, (Gerrsel, [Gere], [N-S-W] and Lal = 1.3.os (CC SPACES SEEN FROM WITHIN 0 commutators span T(HW) and tum to the question of the Hilder ex- ponent for asmooth map f : W —+ V with respect to the C-C metrics associated to H’ and H. Here is the answer 0.5.B. Hilder exponent evaluation. The Hélder exponent of f cquals J-' for the minimal j, such that the differential D of f sends H' into H, (This includes the case of a Riemannian W for H’ = T(W).) Notice that this evaluation of the Hilder exponent contains two state- ments: the fact that f is C* for a = 7! and that it is not OF for 3 > j-*, ‘where the latter is immediate with the inclusion C-C ball ¢ exp Box in V while the former uses the Euclidean property of Hj-horizontal curves in V. It is not at all clear what is the precise geometrical (infinitesimal) sig- nificance of f being C* with respect to C-C metries without assuming f is smooth. Yet one can think of such maps f as of generalized solu- tious of the P.D.B.-system expressing the inclusion D(H!) © H, for 4 = (ontar') + 1. In faet, one of the general questions concerning O-C manifolds is the following, 0.5.C, Hilder mapping problem. Given two C-C spaces V and W and a real number 0 V — W is the degree (sce [Groxep] aud [G-L-P]). Now let V be C-C associated to a generic subbundle HC T(V) of codimension, one (e.g. a contact structure). Then, by an easy argument, Nm(X) ~ CX"** which agrees with the fact that the Hausdorff dimension of such a V is n-+1 (see below as well as 1.4.2’ and 2.5). Notice, that this is sharper than the bound Nm(X) $,A2" provided by the Riemannian estimate Nin(A) $ " via the Holder C!-equivalence between the contact and Riemannian structures, 0.6. The volume of C-C balls and the Hausdorff dimension. We assume here for simplicity’s sake that the polarization Hc T(V) defining our C-C structure is equiregular and take a frame of vec- tor fields X),...,Xy, n = dimV, which agrees with the comumuta- tors filtration Ho = Hy Cv © Hy C+ C Ha = T(V) defined above. (This means the first yr fields for m, = rank H) belong to Hj, the following ma fields for m2 = rank H2/H, belong to Hp, ete.) We denote by deg X;(= deg i) the minimal j such that X, belongs to Hy, Thus deg X, = 1 for i = 1,...,m = rank Hy, dogX; = 2 for i= m+1,...,rank Hy/Aj, etc. The corresponding box Bax(p) CR" given by the inequalities |f,| < p¥#** (obviously) has the Buelidean vol- tume equal 2"p? for D = S7/, deg X, = Xj, jramk(H,/H;—1), where we assume Hy = 0 C H, and by the ball-box theorem the Riemannian volumes of the C-C balls B,(p) in V are roughly the same, Namely, there102 MIKHAEL GROMOY 80 exist continnons strictly positive functions C\(v), Co(v) and po(v). on V, ‘such that ‘ *s(ur) < (Vol Bu(p))/p? < Calv) H forall p < py(v). Consequently, the Hansdorff dimension of V with respect to the C-C metric associated to H equals D. (See |Mity,2) and 1.3.A), equiregular V for some Notice that the inequality (+) partially extends to not structures (see [N-S-W]) which allows an eval non-equiregular Ff, e.g. for the real analytic ones (see 1.3.4) 0.6.A. On the intrinsic size of C-C balls. If p is small, the C-C ball B,(p) approximated hy the box with the sides p”, i = 1,...,n, (extrinsically) appears much smaller than the Euetidean ball of radius p (or the prenbe which is roughly the same as the p-ball from our view- point) if p is small und at least one of tbe degrees d; = deg Xj exceeds one (ie. if rank Hy Bp for a fixed (independent of has infinite avera nb>0. Recall that the average imultiplicity, also called the total volume or the variation of a map f with the range in a Euclidean ball B CR is defined by the integral f card f-M(a)dex, If the domain V of f is an n-dimensional Riemannian manifold and f is Lipschitz, this integral equals the total volume of the induced (singular) Riemannian metric on V and, hence, is finite for compact V. But if V is non-Riemannian of the Hausdorff dimension V > n, one can only claim the finiteness of the integral f tesiven (Fa) dry A0.6 (CC SPACES SEEN FROM WITHIN 103 ‘as this, for Lipschitz maps f, does not exceed const mesiy V by the coarea inequality (where const dim V’ and Dy(V") is maximaltoa MIKHAEL GROMOV go nanifold submanifolds V’, For instance if V" is a generic w/, then the number m{(v) is given for almost all v € V! by ingg obvious formula depending on codim V4 =n —n', feamee if codimV’ < rank Hy_1. for genes of dimen the Fallon my +n! =n if rank H)—, rank Hf; mi(v) It follows in the ease d = 2 that dimV—tankH if codimV! < rank H, dim V’ if codimV’ > rank H. ma Dy(V!) — din V" = { Evaluation of dimpauV’. The Hausdorff dimension of an arbitrar sinooth submanifold V" in a manifold V with a C-C metric associated to H equals the number Du(V") This is an easy corollary of the ball-box theorem for V (see 4.1.4) Warning. ‘The restricted C-C metric disty on V’ by no means equals the C-C metric on V associated to the polarization H’ = T(V’) 1H on V. In fact, this restricted metrie is of more general nature than Carnot- Carathéodory, (But it fits in many cases the definition in §1 of [N-S- W),) Yet there are important examples (e.g; hypersurfaces V" in contact ifolds of dimension > 5) where the restricted C-C metrics distr |V! rlent to the C-C metrie distyy: on V/ (see 2.4.B) is Lipschity equ Let us look at the inequality ditmyigy VW" < N! for a given NV? > 0 as equation imposed on V’ CV. If V! is a smooth: submanifold (and 0 dimyaaV! = Dy(V) this ean be indeed represented by a system of partial differential equations on V" expressing some degree of tangency of V! to H,. For example, the relation dittiigy V’-< dim V" says, in the case whore V" is smooth, that. V’ is everywhere tangent to 7. Thus non- smooth subspaces V/V with dimyyy V7 3 (see 2.1). Notice that this inequality immediately implies non-existence of C?-embeddings R? + V for a > 2 (as was claimed in 0.5.D earlier) since the Hansdorif dimen- sion (obviously) agrees with C*-Hélder maps f : W — V by the rule dition S(W) < a7! ditmiian We Exercise, Show that the Euelidean ball in the contact dspace V can bbe covered by =~! C-C balls of radius © while every C-C e-ball noeds a =~? Buclidean balls to cover it, Use this to compare the C-C and Euclidean Hausdorff dimensions of subsets in V. State and prove corre- sponding results for general C-C manifolds (V, H). 0.7. Isoperimetric filling problem. Let § he « k-dimensional cycle in V which is homologous to zero. We want to evaluate the minimal possible Hausdorff dimension of (the supports of) (K+ 1)-chains D in V filing in ie. having OD = $. More specifically we want to bound this minimal J = ditmys, D in terms of i = ditmiigy $ and then, moreover, we look for fa bound on the (minimal possible) j-dimensional Hausdort? measure of fillings D of S in terms of the i-dimensional Hausdorff measure of S. 0.7.A. Isoperimetric inequality for k = n — 1, If S is a closed hypersurface in V then there is little choice for D : this is domain in V bounded by 8. Thus we exercise no control over the Hausdorff dimension of D but we still may try to bound its Hausdorff measure of an appropriate dimension in terms of the Hausdorff measure of $ one dimension less. A first result of this kind is due to Pansu (see [Pattrnes] and [Pamiyic}) who proved the following isoperimetric inequality for the 3-dimensional Heisenberg group V with au equivariant C-C metric.106 MIKHAEL GROMOV so Pansu isoperimetric inequality. The 3-dimensional Hausdorff mea- sure of the boundary § = OD of a domain DC V restricts the 4- dimensional measure of D by mess D < const(mess $)4 () Notice that mes, in this ease equals the ordinary (3-dimensioual) Haar measure on the group V. Now we recall that every left invariant Riemannian metric on the group V has the same asymptotic geometry as the C-C metric. Thus (+) is essentially equivalent to the following inequality for the ordinary volume and area with respect to a left-invariant Riemannian metrie g in V, Vol D < consty(AreadD)3. (+) (This is stronger for large domains D than the Buclidean inequality ‘Vol D < const (ArcadD))). ‘The inequality (++) was extended by Varopoulos (see [Var-Sa-Co]) to an arbitrary simply connected nilpotent Lie group V, Voly D < const(Voly-1 3D) 81, where n = dim V and N is the asymptotic Hausdorff dimension or (he ‘exponent of the volume growth of the balls B(R) CV, that is N= jim, (log Vol (2))/log Re ‘The isoperimetric inequality for general C-C spaces V of the Hausdortf dimension NV reads mesy D < consty (mesy—1 0D)" w and we shall prove this in 2.3 under suitable assumptions on V. (‘The inequality (+) in the form of a Soholev inequality for smooth functions on V is due to Varopoulos, see [Var].) 0.7.B. Filling in curves in V. Here k = 1 and we look for a “minimal” surface DC V which fills in a given closed curve $C V. We assume for the moment that is reetifiable (i.e. dizmitan = 1) and, in fact, we do not loose much by assuming that $ is smooth and horizontal (with respect, to the polarization underlying the C-C-geometry). But even in this caseoF PACES SEEN FROM WITHIN wT the evaluation of the (minimal) Hausdorff dimension (and measure) of surfaces D filling in S is a rather snbtle matter as is seen in the following Contact example. Let V be the 3-dimensional Heisenberg group with a left invariant and self-similar C-C metric. We know already (see 0.6.) that surfaces in V have Hausdorff dimension > 3 and moreover, one eat show (using (+) for instance) that every filling D of a simple closed curve S has mess D > => 0 for ¢ = «($). On the other hand the Riemannian area of a smooth surface hounds mess which implies the (filling) inequality mess D < C{mess $)° for & suitable filling D of 8. The situation radically changes if we take the Heisenberg group V of di- ‘mension n > 3. (Recall that the Heisenberg group of dimension n = 2m+1 is characterized by its Lie algebra which admits a basis x;,ys,2, @ = 1,..-m, where = is central, [2.24] = [yyy] = 0 and the only non-zero ators between a, and y, are [2,, 44] --sm.) This V ad: ants an e derlying (contact) polarization H has codimension one (being spanned by x; and y)). Now, for n> 3, the filling of closed C-C-rectifiable curves X can be made more efficiently than for n = 3, Namely, 5 can be filled by surfnees D of Hausdorff dimension 2, which means, in effect, that these are horizontal, i.e. are everywhere tangent to H. Moreover, the area of these horizontal surfaces can be made as small as const(length $)°, Thus every S admits a filling D for which sy D < const(mes; 8)? ) (see 3.5 and 4.8 where (+) generalizes to some non-contact C-C manifolds). 0.7.C. Filling for dim $ > 1. Here the situation is rather unsatisfiac- tory, as we lack a non-trivial upper bound on the (best) filling in most cases, For example, we have the following, Open question. Let V be the 2m + 1-dimensional Heisenberg group with the C-C metric. Does there exist, for every A-dimensional cycle 5 for k m are expected to admit fillings D satisfying, mesi42 D < const (mese 15) Hy) ek dim V—1 as in the (only known) Remark (a) Since the Heisenberg group V admits selfsimilarities (dilations) Ay : V — V which scale the r-dimensional Hausdorff measure by ¢” for each r, every one of the above inequalities can be reduced to the case where the relevant measure of $ equals one and the problem boils down to finding any bound on the measure of a suitable filling D of Ss. (b) One knows that the constant in (++) is definitely different from evo In fact, for every closed! smooth k-dimensional submanifold X CV with & 2 mi, there exists a constant © = e(8) > 0, such that every filling D of 3 has mesne2D >. “Ww ‘This follows by the argument in 3.1.A. 0.8, Carnot-Carathéodory metrics as limits of Riemannian ones. Consider a smooth manifold V where the tangent. bundle is decomposed into the sum of two complementary subbundles, T(V) = H ® H+, and let Ae = T(V) = T(V) be defined by (yt) + (hth) for all t> 0. Then we pick up some Riemannian metric g on V andl look at the family 4 = Aj(ga) as ¢ varies between 1 and 00.os ©-0 spac EN FROM WITHIN 109 0.8.4. Riemannian homogeneous spaces and their limits. Let V be a Lie group and 1,1 and g be left invariant. Then the metrics gp ave also left invariant and one may naively think that their geometry is fairly simple. But even in the first non-trivial (ie. non-commutative) case of V = SU(2) aud rank H = 2, one has insufficient, understanding of the asymptotic behavior of the geometry of (V4) for 1 —> 20. For example fone still does not know if the path (ray) (Va), # € [1.20] is (at least roughly) minimizing in the space M of Riemannian manifolds with a suitable metric on M. (I have more respect for this problem now than ten years ago when I first faced it, see my note [Gro,gus] following a tmeeting, in Torino in 1983 on homogeneous spaces.) If the complementary bundle A has rank > 0, then the family of the Riemannian metries (quadratic forms) ge diverges at each point © € V. Yet the associated distance functions disty = dist, may converge for t 50. Im fact, if the subbundle Z (polarization) Lie generates the tangent bundle (i.e. snecessive commutators of H-borizontal, fields span T(V)) then dist, (obvionsly) converges to the Carnot-Carathéodory met- rie dist = dist, on V (compare 1.4.D). 0.8.B. Contact example in the spherical clothing. Let V = S* and H © T(V) be the (2-dimeusional horizontal) subbundle normal to the fibers of the Hopf fibration $® — S?. As we take the metrics ge for ¢ getting larger and larger, the Hopf fibers are becoming longer and Ionger but the diameter of S? with respect to the metric dist, remains bounded for t + o0, as every two points in Scan be joined by an I~ horizontal curve whose gi-length is independent. of t. (This drastically contrasts with what happens to the trivial fibration V = S2xS! — $3, where diam(V,dist,) > 50 for t + 00.) Lot us formulate two basic problems concerning asymptotic geometry of Vi = (Vig) (lor t+ 20). 0.8.C. ‘The asymptotic mapping problem. Find the asymptotics for f + 56 of the Lipschitz constant Lip, of the homotopy class of the identity map (Vm) —* (Var), where the Lipschitz: constant of a class ® of maps ¢ is defined as the infimm over g € # of the Lipschitz constants L(e) =sup De(v)]. More specially one wants to know if the Lipschitz constant of the identity map can be significantly decreased (for large ¢) by homotopying this map. One also asks this question for other dilation characteristies such as the Ly-norms of Df on A‘T(V) (compare 0.5.E.).110 MIKHAEL GROMOV §0 0.8.D. The intermediate volume problem. Take a k-dimensional homology class € Hy(V) and evaluate the asymptotics of Vol(a) with, respect to gy as t > oc, where the volume of a homology class in Riemannian manifold is defined as the infimum of the volumes of the cycles inthis class. Notice that these two problems are vaguely parallel to the Hilder map- ping problem and the Hausdorff dimension problem correspondingly (sce 0.5.C and 0,6.B) and we sce later clearer relations between this kind of probles 0.8.E. Families of metries associated to dynamical systems. In- stead of moving g, by automorphisms A, : T(V) T(V) one may use a family of diffeomorphistns ay : V+ V and define dist, = sup_ 05 (dist) {or a fixed (Riemannian) distance function on V. Many dynamical charac teristics of, e.g. the fopological entropy’, can be expressed as asymptot invariants of disty. On the other hand, for some dynamical systems (fitst of all for Anosov systems) one eau teuormalize dist; such that the limit for f + 06 exists and is of Carnot-Carathéodory type. Here is a typical question where these ideas are useful (see 4.10). 0.8.F. The intermediate entropy problem. Recall that for every self-homeomorphism a : V — V and each compact subset Kone can define the topological entropy ent(a: IC) (by using suitable e-covers of IC for the metries dist, = sup_(a!)* dist). ‘Then ane defines entg(a) as the infimum of ent(a; ) over all compact subsets K of topological dimension 1k. The question is how to evaluate this ents for specific (e.g, Anosov) transformations @ 0.8.G. Intrinsic approximation of C-C spaces by Riemannian ones. The approximation of a C-C metric by the Riemannian metric i of 0.8 makes an essential use of the polarization HT ud to see in purely metric derlying the C-C structure, and this is (intrinsic) C-C terms. An alternative intris (where © corresponds Lo t-?) appeals to the nerve of a suitable e-covering of V where each simples of the nerve is given the metrie of the standarc je approximation Ve to V09 cc SPACE: SEEN FROM WITHIN un Enclidean é-simplex. (These V- are piecewise Riemannian rather than Riemannian but this hardly matters.) We shall see later on that these V- rather closely approximate V for ¢ — 0/in a suitable category (albeit V. are not tiecessatily homeomorphic to V) and many geometric invariants of V can be extracted from those of V- for © +0 (see 1.4.D and |Gersisen)) 0.9. Conformal C-C geometry and hyperbolic geometry. The most profound geometric applications of the Carnot-Carathéodory struc- tures are centered around the rigidity problems for non-compact. sym- netric spaces of rank one and are due to Mostow and Pansu (see [Most], [Pangina])- Recall that every compact metric space V serves as the ideal boundary of certain hyperbolic spaces CV. Namely, we take CV = Vx{0,20| with the maximal (or better to say supremal) metric satisfying the following two conditions. (3) for each v © V the embedding [0,20[+ CV for t — (0,1) is (non- strictly) distance decreasing, (ii) for each £ € [0,o0[ the embedding V+ CV for 0 (vt) is 2 Lipschitz. Then one can show that the quasi-isometry type of CV determines {suitably defined) quasi-conformal types of V (compare [Mar], [Groac, Example (a) Let V = $* with the C-C metrie associated to the above (contact) po- larization (normal to the Hopf fibers). ‘Then the corresponding, space CV is quasi-isometric to the complex hyperbolic plane and quasic on of CV induce (passibly uon-sinooth) contact maps on the x CV = V which are qnasi-conformal for the C-C. that the quasi-conformality is automatic for smooth ideal boundan metric (no contact maps). (h) Let V = S? and HC T(V) be horizontal for the Hopf fbr S*. Then the corresponding CV is quasi-isometrie to the quaternionic hyperbolic plane, Quasi-isometries of CV induce C-C quasi-conformal transformations of V. But these are quite special (unlike the contact inaps in the complex: hyperbolic caso) by the following12 MIKHAEL GROMOV yo Pansu rigidity theorem. The above transformations of V are, in fact, conformal aud the corresponding quasi-isometries of CV are equivalent (asymptotic) to actual isometries of the quaternionic hyperbolic plane,” Remark. Usually one starts with the complex (or quaternionic) hyper- bolic space (e.g. plane) X and reconstructs the C-C structure at the sphere at infiuity Om. as the limit. of normalized Riemannian metrics wduced on the concentric spheres Vi CX’ around a fixed point tp € X. Namely, oue takes gon V; equal to (induced metrie)/Diamny, where Dian, denotes the diameter of the induced Riemannian metric on V, (which is about exp? for t 20), and then the distance funetions dist, = dist, (om q.X identified with Vj) converge to the Carnot-Carathéodory metric on the correspouding sphere with the horizontal subbundle (polarization) as- sociated to the Hopf fibration, (If one does all that to the real hyperbolic space of dimension n one gets just the usual round Riemannian metric on S"~1 and one does not know what happens for non-symmetrie spaces of negative curvature, compare [Groai]). 0.10. Dimension and growth in the asymptotic geometry. We have seen in 0.8. how certain problems for C-C manifolds appear in the limit for families of growing Riemannian metries on a compact manifold. “Then such families can be realized on growing concentric spheres in a sin gle non-compact complete manifold X whose ideal boundary comes along with a C-C geometry which is determined by the asymptotic geometry of LX. One can generalize further and take a quite general (hot necessarily iyperbolie) complete manifold X and transplant our basic C-C problems to X. For example, one may look for lower bounds on the (volume) growth, of subsets ¥ CX in terms of a suitable asymptotic dimension of Y in the spitit of (Groay) 1. Horizontal curves and small C-C balls We analyze the behavior of short horizontal curves issuing from a given point v & V by a systematic (and somewhat: boring) use of the Taylor remainder formula and thus prove several versions of the ball-box theorem (soe [Bell] for a more elegant treatment of these problems). We apply this * See the original paper [Pangie] for a more precise and general statPACES SEEN PROM WITHIN ns to the evaluation of the Hausdorff dimension of V and submanifolds V" in V and also to # count of homotopy classes of Lipschitz maps V — W. Our first step is 1.1. Proof of the Chow connectivity thearem, (compare (Herm) Recall that we are given vector fields Xy,-..,Xjq on a connected manifold V which Lie-generate T(V) and we want to join a given pair of points 7 bay a piecewise smooth curve where cach piece is a segment of am al curve of some field Xi,t = 1,....mb Since the problem is local, we may assume the fields X; integrate to one- parameter groups of diffeomorphisms of V and we must join given pointss by piecewise orbit curves. In other words we must prove that: the group G of diffeomorphisins of V gencrated by these subgroups is transitive on V. We start with the following, ‘Trivial Lemma. If G contains one-parameter subgroups, say Yi(t). Ya(t)...+.Yolt). where the corresponding vector fields V1, ¥a,-...Yy span T(V) (without taking commutators), then G is transitive on V. Proof. This follows from the implicit function theorem. Namely, for each v €V we consider the composed! action map E),:R” — V defined by (tays-+ sty) > Yalta) 0+ 0 Yoltoe) ‘The differential of Ey at the origin 0 €R” sends RB” onto the span of the ficlds ¥; in 7,(V) and, hence, is surjective in our ease. ‘Thus the orbit Go) is open in V for each v € V by the implicit function theorem and, as V is connected, G(v) = V . Now, let X(t) be # one-parameter group (flow) on VY be # vector field and let us look atthe transport of ¥ by X(t), denoted X.(QY. We observe that for small t+ 0 (QV =¥ 4eX,Y] + oft) racket) and conclude that, since the p> (by the very definition of the Li commutators of X, span T(V), there exist veetor fields Yj. j m,on V which span 7(V) and such that ()¥ =X for i= 1....,m,mM MIKHAEL GROMOV i (ii) each field ¥j for J > m equals (Xi(t))) Yj, Le. the transport of some Vy for j! mg is of the form Xj +2[X;, Xj] +0 Auerofore the vectors ¥),...,¥%, amv linearly independent at v for s nul so the differential D sR" — T,(V) of the composed or B= Ett. sta) + Vidor o Vale) does not degenerate, Moreover, it is clear that the norm of the inverse operator is bounded by |") const!=®, provided © > 0 is sufficiently small. This follows from the implicit function theorem since the second derivatives of the map £ are uniformly bounded in & (as th is true for the Fields ¥; of the form (X,(«)), Xp) . Reformulation in terms of the C-C metric. ‘The length of piecewise integral curves defines a metric on V which is of a slightly more general type than Carnot-Carathéodory since the fields X, are not supposed to be independent. But if these fields are independent this metric essentially majorizes the Carnot-Carathéodory one as our eurves are (special) H- horizontal for H = Span{ X;}. In fact, the two metrics are equivalent as wwe shall see Inter and we call our present metric C-C anyway, Now we can reformulate the above proposition in terms of the following Hélder hound on this C-C metric by a Riemannian one, CC dist $ (Riemalist)}. (+) 1.1.4’. Upper box bound on C-C dist for deg Xi < 2. Let us refine (+) by taking into account the difference in the behavior of the CC distance in different directions. Namely, we want to show that in the direction of H the C-C metric is equivalent to the Riemannian one, Namely, let c(t) be a smooth curve in V issuing from » © V parametrized by the length parameter t. I(t) is tangent to Hat (=0, then CC dist(w=c(o).c(t)) & t (++) Proof. The tangent veetor of e(t) at 1 combination of X; at v, say e(4) = ¥ composed orbit map (0) can be written as a line a,X,. Then the yalue of the E(theeses ln) HP Xi(f oe OX yu (HCO) at the point (lay....,a,) approximates e(t) for small # as Ella, --- stan) = (0) + OF), Jor remainder theorem (applied to e(t) and to B with respect uclidean structure in V near e). It follows, with (s). that C dist(c(t), E(tar,...,tam)) St16 MIKHAEL GROMOV il and the proof of (4) is implied by the triangle inequality as CC dist(v, E(tar,...,tam)) 0 independent of «). Now we want to show that the e-C-C ball is contained in such a box. Namely, we want to show that along a curve co(¢) transversal to H the C-C distance is 2 VRiemannian distance. Here are two slightly different proofs. First proof, Assume for the moment that the dimension of the span {X;} is constant at v and let o be a smooth T-form defined near v, such that a(X)=0, and : a(e(t)) =1, for our curve e(l) issuing from » in the direction e'() transversal to HL Let ¢ be a piecewise suiooth curve tangent to H joining 0 with we ea(t = £) and let b denote the closed curve formed by ¢9[0,2] and cy see Fig. 2 below. Figure 2 This 6 bounds a disk D of Riewannian area about @ for lengthy b= length ¢ + ¢, and by the Stokes formula AveaD2 ff d= Lovat®*ry ©-€ SPACES 81 FROM WITHIN ur since «vanishes on ¢ (which is tangent to £1). Tt follows that (length e+ 2)! 2 © and consequently lengthe 2 VE. . ‘Now, even without assuming that the span of X, has constant rank, we have a form o which annihilates all X; at v and has a (c)(0)) = 1. Then the integral ff, da = for acquires an extra term, namely Ja, which is of the order at most (length e}? sinee a(X,) is of order < in the Riemannian e-ball around v. This gives ns the rela (lengthe: +2)? 2 ¢ + (lengthe)?, which implies length e > y just the same. Second Proof. Take smooth hypersurface Vp CV passing through v and tangent to Hy © T,(V). Then every curve e issuing from v and tangent to H is also tangent to Vj at v and by the Taylor remainder theorem the Riemannian distance from e(t) to Va is bounded by = ¢2. Hence, the Riemaunian distance between v = e(0) and e(t) is bounded by # on each curve cy transversal to Vo. . Remark. Notice that both proofs need the fields X, to be C!-smooth ‘and fail for continuous ficlds where, in fact, the quadratic Holder bound may be invalid. 1.1.C, Corollary: Ball-box theorem for deg = 2. The sinall e-C-C Dall in V around! 1 is equivalent to the following ¢xe*-box. Take a smooth mo-dimensional submanifold Vi C V through wv with Z,(Vi) = Hy, take the Riemannian c-ball in Vi and then the Riemannian ¢?-neighbourhood of this ball in V. This is our exe*-box. It is equivalent to the &C-C ball in the sense: that Box( 8%) C e-C-C ball C Box(AexA*=?), where & and A are positive constants independent of = (which can be chosen eontinuonsly depending on 1 jn a suitable sense).bs MIKHAEL GROMOY aa 1.2. A new proof of the Chow theorem and the Hélder bound on the C-C metric for arbitrary degree. (compare [Mit,,2] and [N- S-W]). Our proof of the bound CO dist § (Riemdist)! w for the case where 7(V) is spanned by the commutators of the fields X; of degree d < 2, does not easily extend for d > 3 but the following more siraightforward approach proves out to be more forceful as it yields (+) for all d. The key ingredient is the following (well known) approximate expression for the commutator of one-parameter groups Xy(t) and Xo(t) in Diff V in terms of the I-parameter group corresponding to the Lie brackets of the fields Xy and X2, [Xi(t), Male)? XK) 2 Nalt) oXL Oo. XZ() = [Ni Xal(P) + 047), (+) where the additive notation refers to some Euclidean structure ina relevant neighbourhood (and where one should note that X;1(t) = Xi(-t), 2). Proof of (#). We nood the following three elementary formulas () (FY) = ¥ (tr) (2) (X47Y)()=X(Ho¥ (rt) +0(tr)=¥ (rt)oX(t)+oltr), for tyr — 0. (3) X()Xa(r)X7 '() = ( [XL Xa|)t + oft), Notice, that (1) is obvious, (2) follows from the Taylor remainder theorem for the composition X(f) © ¥(rt), and (3) is implied by (1), (2) and the definition of the Lie bracket (compare 1.1). Now we obtain (+) in the form X(t) © Xa(t) o XP) = [Xa Xal(l*) o Xalt) + 0(t*) by applying first (3) and then (2) and (1) to the left-hand side. Next we observe that (+) by induction implies [Xu(8), X24), Xa(OP"]° = [Xi Koa] (4) + off) [Xai [ee Mal [Xu Xu}. [tet + oft!)12 ©-€ SPACES 81 FROM WITHIN no Proof of (+). We concentrate on the simplest case where dim V = 3 and T(V) is generated hy Xj, Xz and ¥ = [X;, Xz). We denote by ¥°(2) the one-parameter fannily (not a subgroup) of diffeomorphistns defined by _f [xude). ratte)" for #0 [xeculy.xucesa}” for eso and we observe that the composed map E® : (ty, ta, ts) > Xi(t) 0 Xa(t) o ¥(A)(w) sends the box B(e) CR defined hy |t:| c2. To sce that we compare E® with the composed map EB: (tistayty) 4 X(t) © Xalt) o YUN), for which the image of the c-box is 6-large by the implicit function theo- rem, In fact, the E-image of the cube defined by jt)| < 0. (ii) Every two points in V within distance & < 6) can be joined by a unique geodesic segment of length 5. (iii) The map B is a homeomorphisin of B onto its image E(B) in V. Then the image E°(B) contains every point v € E(B) for which the b-ball in V around w is contained in E(B) ‘The (standard) proof of this is left to the reader. Finally we observe that. with the provisions we have made the above proof of (+) extends to the general case (of any number of fields and arbitrary d= 1,2,3,...) by just adjusting the notations.120 MIKHAEL GROMOV 1 1.2.A, Upper box bound on the C-C distance for arbitrary de- gree. ‘The above argument does not provide decent box-shaped domains inside the C-C balls but the following simple modification of this argu- ment does just that, To see this we concentrate again on the ease of three fields 1, Xz and ¥ = [X1.Xa] spanning T(V) and we observe the following 1.2.4’. C!-Lemma. The above family ¥°(t) is C!-smooth. Proof. What we know about ¥°(¢) is the relation ¥°(t2) = [X1, Xal(t) + o(¢?) (s)! where ¥°((2) and [X1,X2](#2) are smooth (as we assume our fields X; are ™-smooth). Hence («)' implies that Y(t) = [Xi Xal() + pit) where y(t) is smooth, and so. ¥°(th= [Xr XA +B (vay is C'-smooth, . We observe that the differential of Y°(t) at t = 0 equals that of Y(t) and so £(é) and £%(t) also have equal differentials at 1 = 0 (where the Cl-smootiness of B° is ensured by that of Y°(t)), and so the E°-image of the box Ble) = {Ita 0. fy+e-+ 4 and small (i) [Yel] < const for soine const > 0 independent of «. Now we define f(t) as the integral curve of the field Y; issuing from ep, f(O) = w and f'(t) = Y; at w= f(t), and observe that f > y for £ — 0. It easily follows (e.g. with Elementary topology lemma) that the map f H+ f(1) contains v in its image for a sufficiently sinall <. . Acknowledgment. ‘The smoothing problem in Chow's theorem was brought to my attention by Lucas Hsn. 1.3. Lower box bound on the C-C distance for deg > 2. We wan to show that on each smooth curve c(t) in V issuing from v in the direction c4(t) € T,(V) transversal to the span of the commutators of iven smooth fields X,, 7 = 1,....m, of degrees < 5, the C-C distance satisfies for small ¢ > 0, CC dist(c(o) = v,e(t)) 2 eH We shall do it by adopting the second proof from L.1.B for which we need the following12 MIKHAEL GROMOV a Definition of tangency. A smooth function f on V is called constant of order » at v € V with respect to given fields Xi, if for every differential operator of order r obtained by composing r < s of our fields, denoted Xp = XeXigee Keg for 1 Gtyeeoste)s the function X7(/) vanishes at v, where veetor ficlds are thought of ax differential oporators of first order acting on functions. Next, a subman- ifold Vo CV passing through » is called s-order tangent to X,, if every function on V vanishing on Vp is constant of order s at v. Examples (a) If s = 1 then this tangeney means that the tangent space T,(Va) C T,(V) contains the span of the fields X, at v (b) AF the system of fields X, is integrable in the sense that it gives a tangent frame to a foliation of V then each leaf is tangent to X; of infinite order and the same is true for every submanifold containing, a leat. ‘The following obvious lemma relates the above definition with our prob- Jem. Lemma, (a) Let f be constant of order # with respect to X; at v and e(t) be a smooth curve issuing from » and tangent to X; in the sense that (1) = DM a(t)Xi, for some smooth finctions a,(t). (If the span H of X; has dimension independent of v, then this tangency atnounts to the inclusion o(t} € H for all t.) Then f(o(t)) e+ (b) Let Vj be tangent to X, at v with onder § then the Riemannian distance from o(t) 0 Vy is bounded by Riem.dist (e(#),Va) Se" In other words, the C-C e-neighbourhood of V near vy is contained in the Riemannian ¢**"-neighbourhood. In particular, the C-C dis- tance on cach curve c(t) transversal to Vy atv iy 2 (Riemannian distance) =! * Notice, that the only assumption on the fields X, is C~-smootlinessLa CC SPACES SEEN FROM WITHIN 125 In order to use this Iemma and exhibit actual boxes in small O-C balls, we uced sufficiently many submanifolds tangent to X, with prescribed order. These are provided by the following Infinitesimal lemma. Let Ty © T,(V) be a linear subspace containing the commutators of X; of orders < s at v. Then there exists a submanifold Vi CV passing through v, having T(Vo) = Ty and tangent to X; at 0 with order 3. One may prove this by a straightforward linear algebra in the space of d-th onder jets of submanifolds (and functions) at v € V. But one can also make a short-cut with the following Exponential lemma. Let ¥i,.+-s¥injs---+¥inaree Yinyseoes¥ine be linearly independent vector fields on V,, where for each r = 1,....5 the 42;---Yin, are taken among commutators of X; of degree r, such that the fields Yy,..., Yin, have the same span at 1 as the commutators of X, of degree < r. (Obviously, such ¥; exist.) Then the image Va of the exponential map exp, : R* —+ V corresponding to Y; is tangent to X, at v with order s. Proof. To make it simple we start with the case » = 2 and assume for the moment that the fields X, are linearly independent and so 21 = m. and ¥; =X, for j= 1,...,m. Let. f be a smooth function vanishing on Vj (or rather on a germ of Yj av) and observe that () [X1.X1F) = 0. degree commutators of X, (2) For every field X = SO! a:X;, the operator X? satisfies X?/(v) = 0. In fact XPf(v) =O for all p = 1.2.3)... a8 fF vanishes on the orbit X(A)(v) (which is contained in. the exponential image of corresponding to X,). 4, ‘The proof of this is clear by now. We conclude by explicitly deseribing a box-shaped domain in V con- taining, the C-C ball around » provided by the above lemmas, Step 1. For every r= 1,2)... we take a submanifold V, C V passing through v such that ‘7,(V-) equals the span of the commutators of X; of degrees 1,...,r. A specific such V, is provided by the exponential map corresponding to Yi,..., Yin, from the exponential lemma. Step 2. We take the Riemannian e'*+!-neighbourhood of V, for exch r and intersect these over r= 0.1,-.- (with the convention Vy = {v}). This intersection is, indeed, box-shaped if we take Vp CV) C-s+C Va C= (which is possible with the exponential lemma) and it contains the <-C-C all around v € V for e” =e.13 C-C SPACES SEEN FROM WITHIN 135 Matching upper and lower box bounds. Our (families of) boxes in side (see 1.2.4) and outside (see above) the C-C balls are slightly different but this does not bring any confusion into the geometrie picture as these families are equivalent in att obvious sense as a simple argument shows. (See [N-S-W] for further information.) 1.3.A. Doubling and covering properties for balls; equisingular- ity and the Hausdorff dimension. The ball-box theorem (as stated in 0.5.A) immediately implies the following universal bound on the Rie mannian volume of concentric C-C balls in a compact manifold V, Vol B,(2p) < C'Vol By(p) CS) for all » € V and real p and some constant C = C(V). (This is one of the major applications of the ball-box theorem indicated in (N-S-W].) Consequently we obtain, as an obvious corollary, the following (purely internal) metric property of V. Beery ball Be(2p) can be covered by at mast & balls of radii p for some & depending only an V (but. not on 0 oF p). Now, suppose H is equiregular. Then, clearly, the ball-box theorem shows that small balls have volumes = p” (see (+) in 0.6) and thus diya, V = D with 0 < mesp < 00, where D is computed in terms of the commutator filtration 0 CH = fy © Hy C «++ C Hy = T(V) by D=S%, irank(H,/H,-1). Furthermore, let V" CV be a submanifold in V such that the intersections H/ = H, ()T(V") have constant ranks for all i=1,...,d. Then in the (proof of the) ball-hox theorem one may use frames adapted to V': if some vector from the frame is tangent to V! at a given point ¢’, then the corresponding vector field is tangent to V’ near v'. ‘Thus we extend the ball-box theorem to the restricted metric dist yy |V", and see that a small p-ball in this metric is approximated by a box in V/ having k! (among his 572_, ki) sides of length « p! for i =1,...,d, where Kk, denotes rank H!/H?_,. Now yre sce as earlier that the p-balls in v’ have volumes ~ p”’ for D! = 54 ik} and so satisfy the same properties as those in V. In particular we see that dimita.(V', dist) = Di Non-equiregular fields. First, let us try to understand how generic those frames of fields Xi,...,Xvm are which Lie-generate T(V). As the Lie generation condition involves jets of arbitrary large orders, one might k that for m > 2 the (jets of) non-generating frames have infinite codimension. In fact this is true modulo the following trivial