Access the full text.
Sign up today, get DeepDyve free for 14 days.
Loading next page...
/lp/springer-journals/harmonic-maps-into-singular-spaces-andp-adic-superrigidity-for-UveCIPvChP
References (52)
-
K. Kwun, J. Tollefson (1973)
Uniqueness of orientation preserving ${\text{PL}}$ involutions of 3-spaceBulletin of the American Mathematical Society, 79
-
G. Mostow, G. Mostow, Gerhard Hochschild (1980)
On a remarkable class of polyhedra in complex hyperbolic spacePacific Journal of Mathematics, 86
-
H. Federer (1969)
Geometric Measure Theory -
Y. T. Siu (1980)
The complex analyticity of harmonic maps and the strong rigidity of compact Kähler manifoldsAnn. of Math., 112
-
M. Gromov, P. Pansu (1991)
Rigidity of lattices: An introduction -
H. Karcher (1977)
Riemannian center of mass and mollifier smoothingCommunications on Pure and Applied Mathematics, 30
-
Y. Chiang (1990)
Harmonic maps of V-manifoldsAnnals of Global Analysis and Geometry, 8
-
Shmuel Agmon (1966)
Unicité et convexité dans les problèmes différentiels -
P. Harpe, A. Valette (1989)
La propriété (T) de Kazhdan pour les groupes localement compacts -
H. Garland, M. Raghunathan (1970)
Fundamental Domains for Lattices in (R-)rank 1 Semisimple Lie GroupsAnnals of Mathematics, 92
-
H. Garland (1973)
P-adic curvature and the cohomology of discrete subgroupsAnn. of Math., 97
-
Richard Shoen, S. Yau (1976)
Harmonic maps and the topology of stable hypersurfaces and manifolds with non-negative ricci curvatureCommentarii Mathematici Helvetici, 51
-
P. Hartman (1967)
On Homotopic Harmonic MapsCanadian Journal of Mathematics, 19
-
K. Brown (1989)
Buildings -
(1966)
On a certain class of discrete Lobachevsky space groups with infinite fundamental domain of finite measure -
P. Deligne, G. D. Mostow (1986)
Monodromy of hypergeometric functions and non-lattice integral monodromyPubl. Math. IHES, 63
-
C. Morrey (1966)
Multiple Integrals in the Calculus of Variations -
DE L’I.H.É.S, M. Gromov, I. Piatetski-Shapiro (1987)
Non-arithmetic groups in lobachevsky spacesPublications Mathématiques de l'Institut des Hautes Études Scientifiques, 66
-
S. Alexander, I. Berg, R. Bishop (1987)
The Riemannian obstacle problemIllinois Journal of Mathematics, 31
-
F. Lin (1989)
Nonlinear theory of defects in nematic liquid crystals; Phase transition and flow phenomenaCommunications on Pure and Applied Mathematics, 42
-
M. Gromov (1986)
Partial Differential Relations -
(1977)
Discrete groups of motions of manifolds of nonpositive curvature -
R. Hamilton (1975)
Harmonic Maps of Manifolds with Boundary -
M. Gromov, I. Piatetski-Shapiro (1988)
Non-arithmetic groups in Lobachevsky spacesPubl. Math. IHES, 66
-
H. Garland, M. S. Raghunathan (1970)
Fundamental domains for lattices inR-rank 1 groupsAnn. of Math., 92
-
N. Garofalo, F. Lin (1986)
Monotonicity properties of variational integrals, ap weights and unique continuationIndiana University Mathematics Journal, 35
-
E. M. Landis (1963)
Some problems on the qualitative theory of second order elliptic equations (case of several variables)Uspekhi Mat. Nauk., 18
-
K. Corlette (1992)
Archimedean superrigidity and hyperbolic geometryAnnals of Mathematics, 135
-
(1963)
Some problems on the qualitative theory of second order elliptic equations (case of several variables), Uspekhi Mat. Nauk -
R. Schoen, Karen Uhlenbeck (1982)
A regularity theory for harmonic mapsJournal of Differential Geometry, 17
-
K. Miller (1964)
Three circle theorems in partial differential equations and applications to improperly posed problemsArchive for Rational Mechanics and Analysis, 16
-
N. Garofalo, F. H. Lin (1986)
Monotonicity properties of variational integrals, A p weights and unique continuationIndiana Math. J., 35
-
K. Stein (1956)
Analytische Zerlegungen komplexer RäumeMathematische Annalen, 132
-
(1975)
Lecture Notes 471 -
A. Selberg (1969)
Recent developments in the theory of discontinuous groups of motions of symmetric spaces -
Integrality of rigid local systems of rank two on a smooth projective variety -
(1969)
Geometric Measure Theory, Springer-Verlag -
J. Eells, J. Sampson (1964)
Harmonic Mappings of Riemannian ManifoldsAmerican Journal of Mathematics, 86
-
(1984)
Ergodic Theory and Semi-simple Groups -
(1963)
A three-sphere theorem -
M. Leite (1988)
Harmonic Mappings of Surfaces with Respect to Degenerate MetricsAmerican Journal of Mathematics, 110
-
F. Bruhat, J. Tits (1972)
Groupes réductifs sur un corps local. I. Données radicielles valuéesPubl. Math. IHES, 41
-
R. Schoen (1984)
Analytic Aspects of the Harmonic Map Problem -
F. Bruhat, J. Tits (1972)
Groupes réductifs sur un corps localPublications Mathématiques de l'Institut des Hautes Études Scientifiques, 60
-
I. Nikolaev (1979)
Solution of Plateau's problem in spaces of curvature no greater than KSiberian Mathematical Journal, 20
-
H. Fédérer (1970)
The singular sets of area minimizing rectifiable currents with codimension one and of area minimizing flat chains modulo two with arbitrary codimensionBulletin of the American Mathematical Society, 76
-
Y. Siu (1979)
Complex-analyticity of harmonic maps and strong rigidity of compact Kähler manifolds.Proceedings of the National Academy of Sciences of the United States of America, 76 5
-
H. Garland (1973)
p-Adic Curvature and the Cohomology of Discrete Subgroups of p-Adic GroupsAnnals of Mathematics, 97
-
E. Vinberg (1967)
DISCRETE GROUPS GENERATED BY REFLECTIONS IN LOBAČEVSKIĬ SPACESMathematics of The Ussr-sbornik, 1
-
(1989)
ZIEMER, Weakly Differentiable Functions, Springer-Verlag, Grad -
F. Almgren (1983)
$Q$ valued functions minimizing Dirichlet's integral and the regularity of area minimizing rectifiable currents up to codimension twoBulletin of the American Mathematical Society, 8
-
I. G. Nikolaev (1979)
Solution of Plateau problem in spaces with curvature ⩽KSib, Math. J., 20
- Publisher
- Springer Journals
- Copyright
- Copyright © 1992 by Publications Mathématiques de L’I.H.É.S.
- Subject
- Mathematics; Mathematics, general; Algebra; Analysis; Geometry; Number Theory
- ISSN
- 0073-8301
- eISSN
- 1618-1913
- DOI
- 10.1007/BF02699433
- Publisher site
- See Article on Publisher Site
Abstract
HARMONIC MAPS INTO SINGULAR SPACES AND p-ADIC SUPERRIGIDITY FOR LATTICES IN GROUPS OF RANK ONE By MIKHAIL GROMOV and RICHARD SCHOEN Introduction In Part I of this paper we develop a theory of harmonic mappings into non- positively curved metric spaces. The main application of the theory, which is presented in Part II, is to provide a new approach to the study of p-adic representations of lattices in noncompact semisimple Lie groups. The celebrated work of G. Margulis [Mar] establishes "superrigidity'" for lattices in groups of real rank at least two. The fact that superrigidity fails for lattices in the isometry groups of the real and complex hyperbolic spaces is known. In fact, Margulis deduced as a consequence of superrigidity the conclusion that lattices are necessarily arithmetic in groups of rank at least two. Arithmeticity of lattices was conjectured and proved in some cases by A. Selberg (see [Se] for discussion). Constructions of nonarithmetic lattices in the real hyperbolic case were given by Makarov [Mak], Vinberg [V], and Gromov-Piatetski-Shapiro [GPS]. For the complex hyperbolic case, nonarithmetic lattices have been constructed in low dimensions by G. D. Mostow [Mos] and Deligne-Mostow [DM]. In this paper we establish p-adic superrigidity and the consequent arithmeticity for lattices in the isometry groups of Quaternionic hyperbolic space and the Cayley plane (the groups Sp(n, 1), n~>2 and Fs176 Archimedian superrigidity for these cases has been esta- blished recently by K. Corlette [C] who used harmonic map theory together with a new Bochner formula and vanishing theorem to prove the result. We show here that representations of lattices in Sp (n, 1) and F 4 in almost simple p-adic algebraic groups have bounded image. This is accomplished by the construction of an equivariant harmonic map from the symmetric space into the Euclidean building of Bruhat- Tits [BT] associated to the p-adic group. We analyze the structure of such maps in detail, and show that their image is locally contained in an apartment at enough points so that differential geometric methods may be applied. In particular, we apply the Corlette vanishing theorem to show that the harmonic map is constant, and conclude that the representation has bounded image. We also prove that equivariant harmonic maps of finite energy from a K~ihler manifold into a class of Riemannian simplicial complexes (referred to as F-connected) are pluriharmonic. The class of F-connected complexes includes Euclidean buildings. * Research partially by NSF grant # DMS-03076. 166 HARMONIC MAPS INTO SINGULAR SPACES This result generalizes work of Y. T. Siu [Siu] which implies the same result for maps to manifolds with nonpositive curvature operator. We now briefly outline the contents of this paper. We consider maps into locally finite Riemannian simplicial complexes, by which we mean simplicial complexes with a smooth Riemannian metric on each face. In the first four sections of this paper we develop methods for constructing Lipschitz maps of least energy in homotopy classes or with the map specified on the boundary provided the receiving space (complex) has non-positive curvature in a suitable sense. This generalizes the theorems of J. Eells and J. H. Sampson [ES] and R. Hamilton [Ham] who proved these results for maps to manifolds of nonpositive curvature. We also prove and use convexity properties of the energy functional along geodesic homotopies to prove uniqueness theorems generalizing those of P. Hartman [Har]. A key property of harmonic maps which we exploit to prove these results is a statement to the effect that harmonic maps can achieve their value at a point only to a bounded order, and near the point they can be approximated by homogeneous maps from the tangent space of the domain manifold to the tangent cone of the image complex at the image point. These homogeneous maps have degree at least one and, at most points, they must have degree equal to one. The homogeneous maps of degree one are compositions of an isometric totally geodesic embedding of a Euclidean space into the tangent complex with a linear map of Euclidean spaces. In particular, these maps identify flat totally geodesic submanifolds of the tangent complex. In section 5 we define an intrinsic notion of differentiability for harmonic maps based on how well approximated they are near a point by maps which are homoge- neous of degree one in an intrinsic sense. We then prove a result which enables us to establish differentiability of a map based on the differentiability of maps into a totally geodesic subcomplex which approximately contains the local image of the map. This result is the main technical tool of the paper as it can be used to show that the local image of a harmonic map under appropriate conditions is actually in a subcomplex whose geometry is simpler than that of the ambient complex. We then apply this result to assert differentiability of harmonic maps into one-dimensional complexes. In section 6 we define a class of complexes which we refer to as F-connected. A k-dimensional complex is called F-connected if each of its simplices is isometric to a linear image of the standard simplex and any two adjacent simplices are contained in a k-flat, by which we mean a totally geodesic subcomplex isometric to a region in R k. We then show that harmonic maps into F-connected complexes are differentiable, and we give a detailed discussion of the size of the set of nonsmooth points, by which we mean points for which the local image of the map is not contained in a k-flat. In section 7 we carry through the Bochner method (in particular the Corlette vanishing theorem) for maps into F-connected complexes. We establish pluri-harmonic properties for maps of K~ihler manifolds, and show that finite energy equivariant M. GROMOV AND R. SCHOEN 167 maps are constant from the Quaternionic hyperbolic space or the Cayley plane. We also extend the existence theory to include the construction of finite energy equivariant maps into buildings associated to an almost simple p-adic algebraic group H. We show that either the harmonic map exists or the image of the representation lies in a parabolic subgroup of H. In particular, if the image of the representation is Zariski dense in H, then the harmonic map exists. The hypothesis on the domain manifold is very general here. One requires only that it be complete. In section 8 we establish our p-adic superrigidity results and discuss the arithmeticity of lattices. Finally in section 9 we discuss the structure of harmonic maps of K/ihler manifolds into trees and buildings. We describe an extension of our work to maps in Z-trees and use it to show that the fundamental group of a K/ihler manifold cannot be an amalgamated free product unless the manifold admits a surjective holomorphic map to a Riemann surface. Applications of harmonic maps into trees similar to those done in section 9 were also obtained by C. Simpson [Sim]. A technical device which plays an important role in determining the structure of harmonic maps into nonpositively curved complexes is the monotonicity in o of the ratio Ord(x, o, Q) defined in section 2. For harmonic functions on R" this is a classical fact which is the L 2 version of the Hadamard three spheres theorem. It says that the logarithm of the mean L 2 norm of a harmonic function on a sphere of radius r is a convex function of log r. We prove in section 2 a global geometric version of this result. Its proof relies on the usual monotonicity formula for harmonic maps (which plays an important role in the regularity theory of Schoen-Uhlenbeck [SU] for energy minimizing maps into manifolds) combined with the strong convexity of the distance function on a nonpositively curved complex. A ratio of this type has been used by a variety of authors on various elliptic PDE problems. A partial list includes Agmon [Ag], Almgren [A1], Garofolo-Lin [GL], Landis [Lal, La2], Lin [Lin], Miller [Mi]. The first author to realize the importance of this type of result for proving unique continuation properties of solutions of general classes of elliptic equations seems to have been S. Agmon [Ag] in 1965. (The earlier papers of Landis are also quite closely related.) The optimal unique continuation result was proved by this method only recently in [GL]. The work in this paper was initiated by a suggestion of the first author that it might be possible to develop a harmonic map theory into nonpositively curved metric spaces, and that, in interesting cases, the resulting maps might be regular enough so that the Bochner method could be applied. In particular, he had a conjecture on the singular structure of harmonic maps into trees. He also proposed a version of the heat equation method which might be used to produce such harmonic maps. The work in Part I of this paper comprises the second author's solution to this problem. The approach taken is a variational approach rather than a heat flow method. The conjectured behavior of harmonic maps to trees is shown to be substantially correct 168 HARMONIC MAPS INTO SINGULAR SPACES (with a slightly worse blow-up of derivatives near singular points than conjectured). Note that the theory developed in this paper is largely independent of discrete group theory, and should be viewed as a part of the geometric calculus of variations. The authors are grateful to Kevin Corlette for pointing out several errors in the first version of w 9 and to Ralf Spatzier for a useful conversation about buildings. The second author thanks Scot Adams and Alex Freire for several suggestions for improving the exposition. Many of these have been incorporated in the final version of the paper. History, motivation and examples We develop in this paper a theory of harmonic maps into certain singular spaces with non-positive curvature. The simplest example of such a space is the tripod (see Figure 1) that is the union of three copies of the segment [0, 1] identified at zero, such that the distance between every two points a and b lying in different copies of [0, 1] by definition equals a + b, V . FIG. I. Similarly, one may consider n-pods obtained by joining n intervals. Then we see further examples by looking at graphs (i.e. connected 1-dimensional simplicial complexes) which are endowed with metrics locally isometric to the above n-pods. Notice that the distance between every two points x and y equals the length of the shortest path in X between x and y that is an isometric embedding of the interval [0, 6 = Dist (x, y)] _~ R into X with 0 w-~ x and 8 ~-~ y. If the points x and y lie sufficiently close together, then the minimal path is unique. In fact, if X is a k-pod then this uniqueness holds for all pairs of points and this manifests the (not yet defined) non- positivity of curvature of X. The next important example is provided by Bruhat-Tits euclidean buildings asso- ciated to reductive p-adic Lie groups. For us, a building is a simplicial complex X which is accompanied by a simplicial action of a (p-adic Lie) group G. One knows that such an X carries a G-invariant metric (with curvature ~< 0) such that (a) Every simplex in X is isometric to an affine simplex in some Euclidean space. (b) Every two points in X can be joined by a unique shortest path (geodesic) lying in the union of some top-dimensional simplices. M. GROMOV AND R. SCHOEN 169 Moreover, the buildings have the following remarkable property: For every two points x and y in X there exists an isometric embedding of R a into X for k=dimX, such that the image of this R k, called a k-fiat in X, contains x and y. (See [BT] and [B] for information on buildings.) One can easily grasp the geometry of 1-dimensional buildings as these are just regular trees i.e., simply connected 1-dimensional simplicial complexes where all edges have the same length and where all vertices have the same number of adjacent edges. The geometry of higher dimensional buildings is somewhat more elaborate but one gains some insight by looking at the Cartesian products of regular trees. (These products are not quite buildings in the above sense as they are built of cubes rather than simplices but they are buildings in the sense of [BT] - which allows polysimplicial complexes-, they do have non-positive curvature and all pairs of points are connected by k-flats.) We shall not discuss at this stage the general notion of non-positive curvature, expressed by K (X)~< 0, for general metric spaces X, but rather indicate the following. Examples. - (1) If X is a smooth Riemannian manifold, then K(X)~<0 signifies that the sectional curvature of X is everywhere ~< 0. In particular, the symmetric spaces of non-compact type have K ~< 0. Recall that these symmetric spaces have the form X = G/H where G is a connected semisimple Lie group with finite center and without non-trivial compact factor groups and H is a maximal compact subgroup. Every such X admits a G-invariant Riemannian metric (since H is compact) and this metric has K ~< 0 by a theorem of E. Cartan. The basic example here is the space X = SL, (R)/SO (n) which may be thought of as the space of positive definite quadratic forms on the n-dimensional linear space. Notice that the buildings discussed earlier are substitutes, for p-adic groups, of the symmetric spaces. In the p-adic case one has several maximal compact subgroups in G corresponding to different G-orbits on the set of vertices of the building. Also observe that the dimension of a building corresponds to the rank (rather than dimen- sion) of a symmetric space X. Recall that rank (X) is the dimension of a maximal .flat in X, i.e., a totally geodesic submanifold isometric to R k. For example, rank (SL, (R)/SO (n))= n-1 and a maximal flat consists of the set of quadratic forms which are diagonal with respect to a fixed basis. (Super) rigidity. - Consider a symmetric space of non-compact type, called l~I, (it is secretly thought of as the universal covering of a compact manifold M) and let 170 HARMONIC MAPS INTO SINGULAR SPACES F be a discrete faithful group of isometries acting on M (if F acts freely on 1VI then, in fact, 1VI is the universal covering of M = I~,I/F). Then we take another space, say X, and let F act isometrically on X, where now the action need not be discrete or faithful. In the cases we are most concerned with, X is either a building or a symmetric space of non-compact type. In this case, X is topologically contractible and therefore there exists a continuous F-equivariant map u o : l~I ~ X which is unique up to F-equivariant homotopy. (If the actions of F on l~I and X are discrete and free then F-equivariant maps 1~I --. X correspond to continuous maps lVl/F --, X/F.) We call the above setup rigid if the map u o is F-equivariantly homotopic to a geodesic map u : l~I ~ X, which means that the graph F, m ~ x X is a totally geodesic subspace of the Cartesian product. The first instance of rigidity was discovered by Mostow in the case where 1VI and X are manifolds of constant negative curvature of equal dimension ~> 3 and where the action of F is discrete and cocompact on 1VI and on X. ("Co-compact" signifies "the quotient space is compact".) This was extended later by Mostow to other equidimensional symmetric spaces and then a similar result was proven by Prasad and Ragunathan for equidimensional buildings. Finally there came. Margulis' superrigidity theorem. - If M is an irreducible symmetric space of rank ~>2 and the action ofF has finite covolume (i.e., Vol 1VI/F<oo) then the above setup is rigid (i. e., u o is homotopic to a geodesic map whenever X is an arbitrary building or a symmetric space). The celebrated corollary of this rigidity is Margulis's arithmeticity theorem for F which says that F is obtained from the lattice SL~ (Z) c SL N (R) by certain elementary algebraic manipulations. (These are: taking the intersection of SL N (Z) with Lie subgroups in SL N (R), applying surjective homomorphisms between Lie groups with compact kernels, replacing discrete groups by subgroups of finite index or enlarging groups by finite index extensions.) Remarks. - (a) Margulis has also proved his theorem for certain reducible spaces (which are Cartesian metric products I~7I=IVI 1 x 1~I2) but we stick to the irreducible case for the purpose of the exposition. In this case every geodesic map u : 1VI ~ X is either constant or is an injective map onto a totally geodesic submanifold 1VI' c X. In fact, the map 1VI ~ 1V[' becomes an isometry if we change the metric in M by a multiplicative constant. Furthermore, if X is a building, then u is necessarily a constant map which sends all of 1VI to a fixed point of F acting on X. Thus the p-adic superrigidity amounts to the existence of a fixed point for every action of F on a building. (b) The superrigidity fails to be true for certain symmetric spaces 1VI of rank one. Such examples are easy to construct for 1VI a real hyperbolic space and there are M. GROMOV AND R. SCHOEN 171 (more complicated examples) for the complex hyperbolic space. What remains of rank one are the quaternionic hyperbolic spaces and the hyperbolic Cayley plane. These long have been suspected to be as rigid as rank >~2 spaces as Kostant has shown that they satisfy Kazhdan's T-property (see [HV]) and also a kind of metric rigidity proven by Pansu (see [GP]). Then, recently, the real superrigidity (i. e., for X a symmetric space) was proven by K. Corlette [C] using harmonic maps ff, l---, X and one of the goals of the present paper is to do the same in the p-adic case where X is a Bruhat- Tits building. (c) Margulis' approach, unlike those by Mostow and Corlette, does not directly involve symmetric spaces and geodesic maps but rather deals with Lie groups and continuous homomorphisms. Notice that a geodesic map between symmetric spaces immediately gives us a homomorphism between the relevant (isometry) groups as central symmetries of a geodesic subspace canonically extend to symmetries of the ambient space. But going from group homomorphisms to geodesic maps (which we do not need for this paper) requires a non-trivial Lie algebraic lemma by Mostow (see [GP]). Idea of harmonic maps. - The classical Dirichlet energy of a smooth function u, that is f llgradull2 or better, flldull2, can be defined for a smooth map between Riemannian manifolds, u" M -~ Y, by E (u) = fM e (u) dl.t where e (u) is the so called energy density of u whose value at m e M is given by e(u)(m) = Dm(u)[[ 2, where Dm (u) "T,. (M) ~ T. (m)(Y) is the differential of u whose norm is defined by [1D [1 = Trace D* D where D* denotes the adjoint operator. (If one uses orthonormal bases which diagonalize D the ][ D [[2 dim M becomes ~ )~2 for the diagonal entries )~i.) i=1 One should slightly modify the domain of integration for F-equivariant maps u: l~I--, X by first observing that the density function e(u) is F-equivariant and so 172 HARMONIC MAPS INTO SINGULAR SPACES descends to a function on M = 1VI/F, also denoted e (u). Then we set E e d.. A smooth map between Riemannian manifolds is harmonic if it satisfies the Euler- Lagrange equation for the energy functional. Thus every energy minimizing map M --. Y in a fixed homotopy class (if such a map exists at all) is harmonic. The similar conclusion applies to F-equivariant map 1~1--. X minimizing the above equivariant energy. The theory of harmonic maps into nonpositively curved manifolds starts with the following existence theorems proven by Eells and Sampson [ES] in 1964 which came before the first rigidity result by Mostow. If X is simply connected, K (X) ~< 0 and the actions of F on M and X are discrete and co-compact then there exists a smooth energy minimizing F-equivariant map u" M--*X. Moreover, one knows in the above situation that every harmonic map necessarily is energy minimizing, and such a map is unique up to a parallel translation in X. This means (apart from some irrelevant pathological examples) that there exists a F-invariant totally geodesic submanifold X'c X which isometrically splits as X'=X o x R' such that the image of every F-equivariant harmonic map 1VI ~ X is contained in X' and any two such maps can be obtained one from another by applying the (obvious) action of R' on X'. (This uniqueness result is due to Hartman [Har].) Notice that every geodesic map IVl ~ X is (obviously) harmonic and, moreover, energy minimizing (at least in the case where the total energy is finite). Furthermore, the above parallel translation moves geodesic maps again to geodesic maps and so the existence of a single geodesic F-invariant map lVl ~ X implies that every harmonic map is geodesic. This suggests the following approach to the (super) rigidity problem: First construct an energy minimizing F-equivariant map lVl---, X and then show that every harmonic map is geodesic. In fact, such a result appears in the original paper by Eells and Sampson, as they prove that every F-invariant harmonic map of IVl = R" into an arbitrary manifold X with K (X)~<0 is geodesic. Then they combine this with their existence theorem and come to the following conclusion. Let M and Y be closed (i. e. compact without boundaries) Riemannian manifolds where M is fiat (i. e., K (M) = 0) and K (Y) ~< 0. Then every continuous map M ~ Y is homotopic to a geodesic map. In particular, if the fundamental group of Y contains a subgroup isomorphic to Z 2 then Y contains an immersed totally geodesic flat torus. The proof of the implication harmonic =~ geodesic M. GROMOV AND R. SCHOEN 173 is similar to the following classical argument showing that every harmonic function u on a closed manifold M is constant. The basic formula here reads div (u grad u)= II grad u II 2 + u Au. If u is harmonic and Au=O, this formula shows that Itgradull 2 equals the divergence of a vector field and so integrates to zero M II grad u Ila = fM div (u grad u) = O. Thus grad u= 0 which means (as M is assumed connected) u is constant. The proof of Eells and Sampson uses a more elaborate expression which involves the Hessian of u (rather than the gradient) which measures the totality of the second derivations of u and which vanishes if and only if u is geodesic. Here is the Bochner formula of Eells and Sampson: If M is flat then every smooth harmonic map u:M ~ X satisfies at each point meM, II Hess (u)II 2 = Ae (u) + K* where e denotes the energy density (function) on M and K* is a certain (real valued) function on M obtained by pulling back the curvature tensor of Y to M by the differential of u and then by taking an appropriate trace of the resulting tensor on M. The explicit formula for K* is not important at the moment but we need the following crucial property of K*: if K (Y) ~<0 then also K* ~< 0 (this is true for all Ca-maps u, not only for harmonic ones). Now, since A = div grad, the integral of Ae over M vanishes and thus f][ Hess (u) [[ 2=fK*~<0 which for K (Y)~< 0 implies that Hess (u) = 0 as well as K*=0. The first relation, as we know, tells us that u is a geodesic map and the second relation says (once the explicit formula for K* is written down) that the curvature (2-form) of Y vanishes on the image of the differential of u. (In this particular case the second 174 HARMONIC MAPS INTO SINGULAR SPACES conclusion can be derived from the first one but for some Bochner-type situations one should keep track of both terms separately.) The idea of using harmonic maps for rigidity problems was widely discussed since the appearance of the first paper by Mostow. One also was encouraged by the local rigidity results established earlier by Calabi-Vesentini and A. Weil where the Bochner method had been successfully carried through in the infinitesimal (and hence linear) setting. However, it took more than 10 years before the first usable Bochner formula was found by Siu in 1978. Siu's formula applies to harmonic maps between Kiihler manifolds and shows, in the case where a certain curvature of the target space is nonpositive, that every harmonic map is either holomorphic or antiholomorphic. Siu's formula was modified by Sampson who considered harmonic maps u of a K/ihler manifold M into an arbitrary Riemannian manifold. Sampson expressed the complex Hessian H dJ du H 2 as a sum of a divergence term and a certain curvature expression (like the above K*) pulled back from the target manifold. Then, assuming his curvature is ~< 0, Sampson concludes the vanishing of the Hessian dJ du (here J stands for the complex structure operator in M and the two d's are appropriate differentials) which means (more or less by definition) that u is a pluriharmonic map, i.e., the restriction of u to every complex submanifold in M is harmonic. Furthermore, Sampson has shown that his curvature is ~<0 for symmetric spaces of non-compact type and thus proved the following p!uriharmonic (rather than geodesic) rigidity theorem. Every continuous map of a compact Kiihler manifold into a compact locally sym- metric space X with K (X)~< 0 (i.e. of non-compact type) is' homotopic to a pluriharmonic map. (This result together with the circle of surrounding ideas was explained by D. Toledo to the first author some time ago.) The simplest case where Sampson's theorem applies is that of the flat torus Y = T' where the Siu-Sampson formula reduces to the classical Hodge identity Au= 2 ~?* ~?u for some function u:M ~ C. (The corresponding pluriharmonicity theorem claiming that every continuous map M --, T" is homotopic to a pluriharmonic map can probably be dated back to Poincar6 or maybe to Riemann.) Maps to singular spaces. - Now we turn to the p-adic (super) rigidity problem where the receiving space X is a Bruhat-Tits building and we want to see what remains of the theory of harmonic maps when the target space is singular. First of all we should define a notion of the energy for maps into non-Riemannian metric spaces. This can be done in a variety of ways in a quite general situation. The most direct definition uses the squared ratio (stretch) between the distances in M and Y, i.e., S (m l, m2) = Dist,~ (u (ml), u (mz))/Dist M (ml, m2) which makes sense (for ml #m2) for maps between arbitrary metric spaces. Then one can integrate S over the ~-neighborhood of the diagonal in M x M, say N~ c M x M M. GROMOV AND R. SCHOEN (here one needs the Riemannian measure in M and one assumes dim M~> 1) and define the energy by (" supg-dimM| S(ma, mz)dmldm2. E(u) = lim ~--*0 ,,IN Notice that in the smooth case this agrees with the classical definition up to a normalizing constant. In fact, the above limit formula gives an intuitive explanation for the Dirichlet energy classically defined with infinitesimals. (One may slightly modify the above by using instead of the Riemannian measure restricted to N~ another approximation to the 5-measure on the diagonal, for example the normalized heat kernel on M for time tending to zero.) Another more practical (but indirect) definition for maps into buildings X (and similar spaces) can be made by locally isometrically embedding X into some R N thus reducing the definition to the classical case (see section 1). The simplest case to look at is where M=[0, 1] and where one can define the energy density of a map u at m by e (u) = 1 lim sup S (m', m") 711' ---~m and then set E (u) = e (u) dm. This energy E (u) on maps u : [0, 1] --. X is well behaved under rather general assump- tions on X and E (u) assumes its minimum at the geodesic curves in X parametrized by a multiple of the length parameter. (A geometric study of these curves in the situation where the singularity comes from an obstacle inside a smooth manifold Y was conducted in [ABB].) Notice that one can recapture some properties of the energy E (u) for dim M >~ 2 from dim M = 1 by restricting u to the unit geodesic segments in M, taking the energy of the restricted maps and then by integrating over the unit tangent bundle of M. Yet one does not expect a meaningful higher dimensional variational theory for E(u) unless the "curvature" of the receiving (singular!) space X is somehow bounded from above. The most convenient (for us) definition of K(Y)~<0 (suggested by A. D. Alexandrov many years ago) can be best seen in the universal covering X of Y. Here is a list of the properties of X given in the order of increasing strength which can be used as a definition for K (X)~< 0 (and thus for K (Y)~< 0, compare section 2.1). 176 HARMONIC MAPS INTO SINGULAR SPACES 1. Every two points in X can be joined by a unique distance minimizing geodesic segment. 2. The distance function Dist (x z, x2) on X x X is convex on the cartesian products of geodesic segments. 3. The distance function on X is "more convex" than the distance on X'=R 2. Namely, if we take geodesic segments of the same length in X and X' which are both identified with [0, 6] c R (for 6 the distance between the ends), then the equalities Dist (x o, 0) = Dist (x o, 0) X' X Dist (xo, 8) = Dist (Xo, 8) X' X imply that Dist (Xo, t) ~> Dist (Xo, t) X' X for all t e [0, 8]. The convexity of the distance implies that the energy E(u) is convex under geodesic deformations of maps which implies Hartman's uniqueness theorem (see section 4). Furthermore, the convexity of the distance allows one to define a good notion (in fact several non-equivalent notions) of the center of mass (see [K]) of a finite measure g on X. (The standard definition of center (12) refers to the point x o e X which minimizes tvDist2(x0, x)d12 x. Another possible definition uses the map c:X x X---, X assigning to (x 1, x2) the center of the geodesic segment between xl and x 2. This c pushes forward the measure g x g to a measure on X, say gl, and thus, by induction, one has 122 coming from 121 x 121, etc. Then one defines the center of g as the (one point!) support of the weak limit of 12i for i ~ o0.) This gives us a possibility to regularize a map u" l~I--, X using smoothing kernels K (m, m') on M as follows. Assign to each point m eM the measure Km=K(m, m')dm' for the Riemannian measure din' on M, and define the regularization u of a given map u by defining u (m) to be the center of the u-push-forward of K m to X, m e M. This operation works particularly nicely if M is a compact flat manifolds, say a flat torus T n, and K is of the form K (m, m')= L (m-m'). In order words, the g in this case comes from some measure )~ (corresponding to L) on the torus acting on itself by parallel translation. In fact, given any family of selfmappings of M (e.g., of isometries or more general selfdiffeomorphisms) and a measure on this family, we obtain by composing with u a family of maps M - Y, say u with a measure d'L. Now we can "average" this family over the parameter space with the measure )~ by applying a center of mass construction at each point m e M. M. GROMOV AND R. SCHOEN 177 In the case where the measure )~ is supported on the isometries of M, this smoothing decreases the energy density at each point of M (for any conceivable definition of the energy) provided the chosen center of mass operator is contracting in the appropriate sense. (Here it is appropriate to use the center of mass defined above with the map c: X x X ~ X where the needed contraction property directly follows from the convexity of the distance function on X x X.) It follows that every energy minimizing map u:M-~ X is invariant under such regularization and conse- quently, in the case M = T" (and more generally, for M flat), the minimizing map is geodesic. Thus the "flat => geodesic" theorem of Eells-Sampson extends to maps into singular spaces. Recall that the basic existence theorem of Eells and Sampson for harmonic maps is based on a certain construction of a heat flow (in the space of maps u'M ~ Y) which can be (at least formally) perforce used for singular spaces Y with K (Y)~< 0 using the above smoothing operators. (The smoothing operators, the way we describe them, must be performed for the corresponding maps between the covering mani- folds 1VI ~ X=Y on the global center of mass defined in X but not Y. Yet this causes no problem as all relevant constructions are invariant under F c Isom. group (X).) Namely, for every t>0 and i= 1,2 .... , we consider the ordinary heat kernel K(m, m, ~' t/i) on the covering 1VI on M (corresponding to the universal covering X of Y) and apply the smoothing i times to a given map u'M--,X. (Here we speak the language of F-equivariant maps 1VI ~ X corresponding to maps M = 1VI/F --, Y = X/F). The limit (or sublimit, whichever existence one is able to prove) of these iterated smoothings for i ---, oo defines the action of the heat flow on u at the time t. Since the (regularity) properties of the Eells-Sampson heat flow are "uniformly good" for K (Y) ~< - ~2 for ~ -, oo one may expect that every map M --, Y can be homotoped to a (essentially unique) sufficiently regular (at least Lipschitz) energy minimizing map. Evidence in favor of this conclusion is provided by those singular spaces with K~<0 which can be approximated by Riemannian manifolds with K~<0. For example, take a unit geodesic triangle in the hyperbolic plane with curvature _j g2 and let ~ ~ oo. Then we obtain in the limit the tripod described at the beginning of the introduction. In fact, an arbitrary finite graph Y admits a similar approximation. Namely, there exists a sequence of compact manifolds Y~e with convex boundaries and with constant curvatures -oug2 for ~'~ ~ ~, such that Y admits embeddings Y c Yg for all ~ with the following properties: (i) each edge of Y isometrically goes to a geodesic segment of Y; (ii) sup Dist(y',Y)~0 for ~-~ oo; y' ~ Yjg (iii) for each ~r there exists a homotopy retraction Y~ ~ Y (which is moreover Lipschitz and the implied Lipschitz constant is independent of ~). 178 HARMONIC MAPS INTO SINGULAR SPACES Now one can construct Lipschitz harmonic maps MeY by using (sub)limits of harmonic maps M --, Y~ for ~ ~ or, as the latter maps (in a fixed homotopy class) satisfy uniform Lipschitz bounds by the Eells-Sampson theorem. Another geometrically significant class of singular surfaces Y which admit a smooth approximation comes from ramified coverings of smooth manifolds Y0 with K ~< 0. If the ramification locus is totally geodesic (of codimension 2) in Yo, then the induced singular metric in Y has K ~< 0 and usually it can be approximated by smooth metrics with K~<-c for e ~ 0. (In fact, the same is true for more complicated ramification loci such as unions of totally geodesic submanifolds with 90~ On the other hand, a smooth approximation of higher dimensional buildings appears more difficult though not inconceivable as certain buildings (and building-like spaces) do appear in the limits of (parts of) symmetric spaces (of rank /> 2). On Bochner formulas in singular spaces. - In order to derive an interesting geometric conclusion from the general theory of harmonic maps u" M ~Y (or l~I ~ X = ~() one needs a Bochner formula showing that a map u, a priori only harmo- nic, is, under favorable conditions on the curvature of Y, more special, e.g. geodesic or pluriharmonic. If Y is a piecewise Euclidean polyhedron (e.g. the universal covering X of Y is a Euclidean building) then one may think that the required curvature condition is somehow encoded in the local combinatorial structure of X. For example, one knows a combinatorial formulation for K(X)~<0 in terms of the numbers of different simplices adjacent to every face and similar but stronger conditions might be responsible for Siu-Sampson type curvatures and their generalization. A comprehensive understanding of such conditions appears a rather difficult (and still unresolved) problem but the examples presented below indicate a way out of this difficulty for maps having certain regularity properties. Consider a map u of the plane R 2 near the origin to the unit tripod Y with the edges numbered 1, 2 and 3, see Figure 2. Here the sectors T, ~ and 3 represent the FIG. 2 u-pull-backs of the corresponding edges of Y and the tripod Z in R 2 formed by the boundaries of these sectors equals the pullback of the central point in Y. At first sight the map u seems necessarily singular at Z as Z goes to the singular locus (the center) of Y. Yet if we look at the map u restricted to two out of three sectors, say on 1 + 2, we see that this (]" + 2)-sector is mapped into the union of the edges 1 and 2 of Y. M. GROMOV AND R. SCHOEN This union, from the intrinsic point of view, has no singularity as it is just isometric to the real segment of double length and so our map u to 1 + 2 (as well as to 1 + 3 and 2 + 3) reduces to a real valued function. Thus the "true singularity" of u is concentrated at the origin of R 2, where the three sectors meet. Now let us see what we need of such a map u in order to prove some integral Bochner formula. We assume we do have some infinitesimal Bochner identity defined on M where the map u in question is smooth and the global identity is obtained by integrating this formula over M. A non-trivial effect is achieved by the presence of some divergence term in the identity which integrates to zero if M is a closed manifold and the vector field, call it 8, whose divergence we integrate is smooth. If Y has singularity and 8 is not everywhere defined, then, in general, I,, div 8 need not be zero. On the other hand if 8 decays near the singularity, then one may expect (and prove whenever the decay is sufficiently strong) that div 8 does integrate to zero. Here, let us recall that the field 8 in all relevant Bochner formulas appears as a bilinear expression in the first and the second derivatives of u, something like Z 0 i u ~?kt U. (In the Eells-Sampson formula the divergence term is It ~ u II 2 = div grad a iu a; u). If, for example, the map u in question has a singular set of codimension two or more (as in the above picture of maps from R 2 to Y), and if the first and the second derivatives are bounded, then the divergence of such 8 obviously integrates to zero as is seen by integrating 8 over the complement of 8-neighborhoods of the singular locus and then letting a --, 0. In fact what one needs for the vanishing ldiv 8 is the decay of ~- 11] @ u ]] 1] 9 2 u ]] for 8 ~ 0. (In fact, if the codimension of the singularity is two, one can relax the "bounded derivatives'condition to II~ul111~2ul]=0 (e -1) for the dis- tance e ---, 0 from the singularity of u). Let us exhibit actual examples of harmonic maps into our tripod Y. To do that we embed Y into a two-dimensional singular space, namely to the unit cone Y' over the circle of length 3 re, The natural (Z3-symmetric) embedding of Y to Y' is isometric and Y divides Y' into three sectors each isometric to the half-plane. The projections of these half-planes to the boundary lines define a projection of Y' to Y. Then we observe that Y' is conformally equivalent to the unit disk U 2 and we think of Y' as D 2 with a singular Kfihler metric (with K ~< 0). It is easy to see that every holomorphic map of an arbitrary K/ihler manifold M into Y' is harmonic (in fact pluriharmonic) and by composing with the projection Y' -, Y we obtain harmonic maps M ~ Y. The simplest of these corresponds to the identity map of M = D 2 (with the flat metric) to Y'= D 2 (with the singular metric). Notice that the derivative of such a map at zero 180 HARMONIC MAPS INTO SINGULAR SPACES decays with the rate ~1/2 (If we use the circle of length k~ we have the derivative 0 (d/2- The above considerations have led the first author to the following: Conjecture. - If Y is a building, every harmonic map to Y must have a singular set of codimension at least two and ]1 ~ u][ ]] 9 2 u]] should have a sufficient rate of decay in order to validate Bochner formulas. It is also natural to expect that harmonic maps into general spaces Y of nonposi- tive curvature exist and are Lipschitz. These conjectures were proved by the second author and the proof occupies sections 1-6 of the present paper. Before entering the analytic discussion of singular spaces on the technical level we add a few more motivating examples. First we observe that the harmonic maps into tripods (and graphs in general) coming from holomorphic maps do satisfy the above regularity properties. Furthermore by looking at the level curves of the simplest such map of the disk 9 2 into the tripod Y we recognize (see Fig. 3) the familiar pattern associated to a quadratic differential on a Riemann surface. FIG. 3 Then we show that every harmonic map u: D2~ Y gives rise to a holomorphic quadratic differential on D z which is equal away from the singular locus of u to the complexification of the form dy 2 for the length parameter y on the nonsingular part of the tripod Y. This is done by approximating Y by (regular) spaces Y~ D Y of constant negative curvature -~, for ~ ~ oo, and then by approximating u by harmonic maps u~Y~. Every such map gives rise to a holomorphic quadratic differential on D 2 coming from the (2, 0)-part of the pullback of the Riemannian metric of Y~ to D 2 and these differentials converge to the desired limit and are associated to the original harmonic map u: D2--+ Y. Thus we see that the harmonic maps of surfaces into graphs are non-singular apart from a discrete set (where the corresponding quadratic differential vanishes). Then one can express @u and ~2u M. GROMOV AND R. SCHOEN 181 (outside the singularity) in terms of the quadratic differential and check the regularity properties. If dim M >~ 3 it seems more difficult to obtain local regularity results by the above approximation. Yet the global Bochner type theorems do follow this way. For example, Every harmonic map of a compact Kiihler manifold into a graph is pluriharmonic. This immediately follows from the corresponding result by Jost-Yau and Carlson- Toledo for harmonic maps into spaces with constant negative curvature. We conclude with an example of a map into a surface with K ~< 0 with an isolated singularity where the singular locus of the map in the domain has codimension one. We take the unit cone over the circle S~ of length l>2 7t for Y and let M be the cylinder M=M o x [-1, 1] for some closed manifold M o. We map the boundary M = (M 0 x -1)t._)(M 0 x 1) to the boundary S 1= 0Y in such a way that the angular distance between the images of M 0 x - 1 and M 0 x 1 in S 1 is at least ~. For example we may send M ox{-1} to an arc of length ~< ll_rc and then send M oxl symmetrically to the opposite arc in the S 1. It is easy to see that the convex hull in Y of two such arcs equals the union of the cones over these arcs and so the harmonic map M ~ Y solving the Dirichlet problem will be contained in this union of cones and thus have a singular hypersurface in M (separating M o x - 1 from M o x 1) sent into the common vertex of the cones, where the metric of Y is singular. See Figure 4. FIG. 4 (Notice that the solvability and Lipschitz regularity of the solutions of the Dirichlet problem follows for this Y by an approximation of the singular metric on Y by regular metrics with K ~< 0.) Additional remarks. - (a) The conjectured regularity property of harmonic maps referring to Bochner formulas does not provide such formulas but rather allows us to reduce those to the nonsingular locus of the map. The relevant formula for our p-adic superrigidity is the one discovered by Corlette in the non-singular frame- work. (One could also use the Kodaira-Siu-Sampson formula applied to an auxiliary K/ihler foliation associated to the domain manifold. But the "foliated" approach becomes somewhat cumbersome, though quite interesting from a geometric viewpoint, 182 HARMONIC MAPS INTO SINGULAR SPACES when the curvature of the receiving space is not strictly negative. Yet there are certain cases where the global conclusion obtained by using foliations cannot be achieved with known Bochner formulas). Notice here that Corlette also arrived at the idea that the harmonic maps into buildings may provide the solution of the p-adic superrigidity problem. (b) Although there are some instance of harmonic maps into singular spaces which have been considered, the subject has not attracted a great deal of attention. (We should mention that Hodge theory and function theory on singular manifolds has been considered by a number of authors. A cohomology theory on Euclidean buildings was developed by H. Garland [G] to study p-adic group cohomology. The work of J. Cheeger has brought the subject of global analysis on singular spaces into prominence recently.) The uniform Lipschitz bound was exploited by the second author in the late seventies to construct Lipschitz harmonic maps into surfaces with cone metrics of nonpositive curvature. This work was refined and used to characterize the Teichmfiller map by M. L. Leite [L]. S. Alexander, I. D. Berg, and R. Bishop [ABB] have studied geodesics for obstacle problems from this point of view and Nikolaev [N] has constructed minimal surfaces in singular spaces with curvature bounded from above. The works of Almgren [A1] and Lin [Lin] both deal with harmonic maps into special singular spaces. The paper by Y. J. Chiang [Ch] discusses harmonic maps into V-manifolds. Part I: Harmonic maps into singular spaces In this first part of the paper we develop some basic existence, uniqueness, and regularity results for harmonic maps into a class of nonpositively curved singular spaces. 1. Preliminary results Let X be a locally compact Riemannian simplicial complex. By this we mean a space which is the geometric realization of a locally finite simplicial complex such that each geometric simplex is endowed with a Riemannian metric which is the restriction to the standard simplex of a smooth Riemannian metric defined in a neighborhood of that simplex. Moreover, assume that the maximal dimension of a simplex in X is k. Assume finally that X is properly isometrically embedded in a Euclidean space R N in the sense that the induced Riemannian metric on each simplex coincides with the given metric. Let M be a smooth Riemannian manifold of dimension n and Riemannian metric g. For a bounded domain f~_ M with smooth boundary, define the space of H 1 maps from f~ to X by M. GROMOV AND R. SCHOEN 183 Ht(fl, X)={ueHI(~, RN): u(x)eX a.e. xaf2} where H I (f~, R N) is the Hilbert space of R N vector valued L 2 functions on f~ with first distributional derivatives in L 2. The H I inner product is given by (u,v)i= Z ou 6. ~, ~: i 8x ~ 8x ~ g~ (x) dgo, In( ) where we use u.v to denote the Euclidean dot product of vectors in R r~. For ueH I (~, X), we denote the energy of u by E(u), so that Etu)= f. Iv ul da, 7l ~u ~u where I Vul 2= Z g'~ denotes the energy density. We now observe the ~, ~=i ~x ~ 0x ~ following result which allows us to construct energy minimizing maps in the space H 1 (f~, X) with arbitrarily specified boundary data. Recall that if u, v e H 1 (f~, RN), we say that u=v on ~?fl provided u-veH~(f~, R N) where He1 (f~, R N) is the Hi-norm closure of smooth compactly supported RN-valued maps on ft. Lemma 1. I. - Let q~Hl(f~, X). There exists uEHl(f~, X) such that u=q~ on ~s and E (u) <<. E (v) for all v e H 1 (~), X) with v = q~ on ~. Proof. - Let { ul} be a minimizing sequence of maps in H 1 (~,X) with ui=q~ on c~f~. Since bounded subsets of H 1 (fL R N) are weakly compact, there is a subsequence again denoted {ui} which converges weakly to a map ueHl(f~RN). Since X is a closed subset of R N and a subsequence of { u i } converges pointwise almost everywhere, it follows that ueH~(fLX). Since the set {veHl(f~, RN): v=q~ on ~} is a closed affine subspace of H 1 (fLRN), it is weakly closed. Thus u=q~ on c~f~, and we have established Lemma 1.1. We now discuss the distance function on X and give a treatment of harmonic maps of an interval into X. Assuming that X is connected (which we do without loss of generality), we see that any two points Po, P, eX can be joined by a path 7:[0, 1] ~ X which is Lipschitz as a map to R y. We can then define the Riemannian distance function d(Po, P1) by d(P0, P1)=inf{ L(7):7 a Lipschitz path from Po to P1 }. It is immediate that d(.,. ) is a metric, and that (X, at) is a complete metric space. We show that the infimum is attained, and describe the associated harmonic map. We consider paths 7:[0, 1] ~ X with 7 (0)= P0 and 7 (1)= P~. Fix such a path q~, and let 7 184 HARMONIC MAPS INTO SINGULAR SPACES be an energy minimizing path provided by Lemma 1.1. An easy application of the fundamental theorem of calculus shows that 7 is H61der continuous with exponent 1/2. We let dT/dt denote the LZ-vector valued function on [0, 1] which is the distributional derivative of 7. We will show that I dT/dt[ is a constant a.e. on [0, 1]. To see this, let ~(t) be a smooth real valued function with compact support in (0, 1), and consider the path of maps 7s (t)-- 7 (t + s ~ (t)). By the minimizing property of 7 = 7o we have E (7)~< E (Ts) for all sufficiently small s, and hence the function s ~ E (7~) has a mini- mum at s = 0. We examine this function more carefully: fo dvs dt=fo dTs where 9 = t + s ~ (t), s being a fixed small number. Now dvs -(1 + s~" (tl) d7 dt d~ by the chain rule, and hence ld72 (1 E(Vs)=fo Since the s dependence is explicitly exhibited here, we see that the function s ~-~ E (Ts) is a differentiable function of s. Thus its derivative vanishes at s = 0, and we have fo t d7 2~'(t)dt= 0 for every smooth function ~ with compact support in (0, 1). This implies that the L 1 function I dT/dt 12 is equal to a constant almost everywhere on [0, 1]. Thus we have established the following result for energy minimizing maps of an interval into X. This will be used later when we develop the theory for n ~> 2. Lemma 1.2. - A map 7 E H 1 ([0, 1], X) which minimizes energy among maps which coincide with 7 at t = O, 1 is Lipschitz and satisfies I dT/dt I = L (7) a.e. on [0, 1]. Moreover, 7 is a length minimizing curve among all Lipschitz curves from 7 (0) to 7 (1) in X. We showed above that I dT/dtl =c a.e. for some constant c. Integrating we find c=L(7). To see that 7 is length minimizing, let 71 be any Lipschitz curve with 71(0)=7(0), 71(1)=7(1). Assume that 71:[0, 1]~X is parametrized so that I d71/dtl = L (71). We then have L2 (7) = E (7) ~< E (71) = L2 (71), M. GROMOV AND R. SCHOEN 185 and hence L (7)~< L (~q). To see that any Lipschitz curve may be parametrized propor- tionate to arc length, we can consider all H 1 ([0, 1], X) which have image in the closed set ~'1 ([0, 1]) and which agree with Y1 at t = 0, 1. We can then minimize energy in this class and repeat our previous argument to show that the minimizer is parametrized with constant speed. Finally, we make some general remarks about H ~ maps which will be needed in the next section. Let u e H 1 (~, RN), and let x o e~. We will say that u is approximately differentiable at x o if there is a linear map l(x) of the form l(x)=A(x-xo)+B with A an N x n matrix and B e R N such that lim{o-2-"fB [u-l]2d~to+cy-" fB [Vu-Vl[2dBg}=O. o J, 0 ,~ (x0) o It is then a general result (see [Z, Theorem 3.4.2]) about H 1 maps that u is approxima- tely differentiable at almost every point x 0 ~ Y~. Observe also that l (Xo)= u (Xo) for a.e. Xo, and hence for a.e. Xo, B~X if u~H 1 (~,X). We will need the following result. Lemma 1.3. - Let ucH 1 (~, X) be a map whose image lies in a compact subset of X, and let x o ~ D be a point at which u is approximately differentiable with linear approximation l (x) = A (x - Xo) + B. If A # 0 and B ~ X, then we have lim d 2 (u (x), B) dE o ] V u 12 dg 0 = 1. o ~ 0 B~ (x0) o (x0) Proof. - We first observe that there is no loss of generality in assuming the metric g to be Euclidean near x 0 because we can introduce Riemannian normal coordinates centered at x o, and compare the integrals in the g metric with the corres- ponding Euclidean integrals in these coordinates. Note that the balls centered at x o are identical in the two metrics, and it is immediate that each of the two integrals appearing in the statement has ratio with the corresponding Euclidean integral which tends to 1. Thus it suffices to consider the Euclidean metric. By the triangle inequality we have iv 12 ) o (xo) ~ (xo) ,~ (x0) where the second inequality holds because u is approximately differentiable at x. Since A#0, we see that [ [Vl[2dg is a nonzero constant times ~". Thus it follows that .)B (xo) 186 HARMONIC MAPS INTO SINGULAR SPACES (1.1) lim IVl[2dl.t [Vul2dl.t=l. k 0 ~ (x0) a (xo) In order to complete the proof we need to compose the distance function of X with the Euclidean distance for points near B. We claim that for any e>0, there is r o >0 such that (1.2) I d(B' P) 1 ~<a IB-PI for P#B, P eX satisfying [B-P[~<r o. This follows from the hypothesis that each simplex containing B is smoothly embedded in R N so that there is a curve from B to P which is arbitrarily close to a straight line when P is close to B. Moreover, since the image of u lies in a compact subset of X, the function d 2 (u (x), B) is bounded and hence d 2 (u (x), B) dZ f x:[B-u(x)]>r 0}c~0B o(x0) ~<cVol{xE~Bo(xo)'lu(x)-B[>~r o }. For cr ~ r 0 we then have fe lu(x)-l(x)12dZ. B~ (xo) By the Sobolev trace inequality (see [Z]) we have f .o xolU(X)-t(x)l 2dz [Vu-Vl[2dl't+c~-l fB lu-ll2d . <<- c fB ~ (x0) o (xo) Thus it follows that lim (Y- 1 -n fO lu-l[Zd :0 o $ 0 Bcr (xo) (1.3) lim or- 1 -. f d 2 (u (x), B) dZ = 0. o ~ 0 ~{x~OB o(xO):lu(x)-B[/>r O} From (1.2) we have (1-OIP-B I~<d(P,B)<~(1 +a)IP-BI for PeX with IP-B [~<r0, and therefore we will write Vol{xe~B,,(Xo):lu(x)--Bl>~ro}<~Cro2 M. GROMOV AND R. SCHOEN 187 /' (1.4) | d 2 (u, B) dE J~ x~c~Ba(xo):[u(x)-B[~<r O} P -BI dE x e 0Ba (xo) : I u (x)-B [~<ro } where ~ means that the ratio of the quantities is arbitrarily close to one. We also have from the triangle inequality and (1.3) [/-B]2d~) 1/2 <" (fOB~ (.o) ~O((~n+ l/2). Since A r 0 we see that fo I l-B [2dZ is a positive constant times cy "+ x, and hence Bet (x0) it follows that f0 B~ (x0) lu-Bl dE f ,d 0B o (x0) Iz-Bl dE' and each term is of the order cy "+ 1. Combining (1.3) and (1.4) we see that fo d2 (u' B)dE~f lu-BlZdE' B~ (xo) d 0B~ (x0) and hence (1.5) fod2(u,B)dZ,~fo ]I-BI2dE. B~ (x0) B~ (x0) Direct calculation shows that for a linear function l(x)= A (x-Xo)+ B we have for all cy>0 [/-ul2dE ~ IVll2d. =1. B~ (x0) o (x0) Thus combining this with (1.1) and (1.5) we have established the conclusion of Lemma 1.3. (fOB~(xo)]bl--B]2d~)l/2--(fOB~(xo) 188 HARMONIC MAPS INTO SINGULAR SPACES 2. Behavior of harmonic maps into nonpositively curved spaces We showed in section one that minimizing maps of an interval into X are Lipschitz. In order to obtain Lipschitz bounds on minimizing maps M ~ X for dim M ~> 2 (or even continuity of these maps), it is necessary to require that (X, d) have nonpositive curvature as a metric space. We now discuss this notion. First observe that if y, Xo, x leR ~, and x(s) is the unit speed geodesic from x o to x 1 parametrized on [0,/], l= Ix0 - xl I, then the function D O (s) = [x (s)-y 12 is a quadratic polynomial in s of the form D O (s)= s2+ as + b where the constants a, b are uniquely determined by the boundary conditions D O (0)= I xo - y 12, D o (l)= Ixx - y 12. Indeed, 21x-v12 this is another way of saying that the function [x-y [2 satisfies " -2 8ij. We 0xi 0xj say that a simply connected space X has nonpositive curvature if for any three points Q, Po, P1 sx, the function D (s)= d2 (P (s), Q) satisfies D(s)<~Do(s ) where P(s), se[0,/], l=d (Po, P0, is a minimizing unit speed geodesic from Po to P1 and Do(s ) is the unique solution of D 0' (s)= 2 on [0,/] with D o (0)= D (0), D o (l)= D (l). Thus the condition states that points of the side of a geodesic triangle opposite to Q in the space X are at least as close to Q as they would be in a Euclidean triangle with the same side lengths. Note that the statement that the Lipschitz function D (s) lies below the monic quadratic polynomial with the same boundary data for every subinterval of [0,/] is equivalent to the distributional inequality D" (s)~> 2 on [0,/]. Precisely this means that for any nonnegative function ~(s) with compact support in (0, l) the following holds D (s) 4" (s) ds >~ 2 ~ (s) ds. /o ;o Thus X having nonpositive curvature is equivalent to the statement that the distance function is more convex than the Euclidean distance function. For a general space X we say that X has nonpositive curvature if its universal covering space has nonpositive curvature. The following properties can be derived from the definition of nonpositive curva- ture (see [B, VI. 3 B]). First, any two points in X can be joined by precisely one length minimizing path. Secondly, if Po, PI and Qo, Q1 are two pairs of points in x, and we parametrize the geodesic paths from Po to P1 and from Qo to Q1 by P(t), Q(t) for te[0, 1] where t is a constant speed parameter along each of the paths, then the function g (t)= d (P (t), Q (t)) is a convex function of t. Note that this second property implies that geodesics from a point spread more quickly than Euclidean geodesics since we may take Po = Qo and conclude that M. GROMOV AND R. SCHOEN d(P (t), Q (t))) ~,d(P (t'), Q (t')) for 0 ~< t'~< t~< 1. Finally, we observe that for any )~ e [0, 1] and Q e X we can define a map R~, o : X --. X by R~, Q (P) = P (Z), where P(t), te[0, 1], denotes the constant speed geodesic from Q to P parametrized on [0, 1]. By our previous discussion we see that R~, Q is a Lipschitz map; in fact d(R~, Q (Po), R~, o (P1)) ~< )~ d(Po, P1), so the Lipschitz constant is at most )~. Moreover, the family of maps Rz. o, ~ e [0, 1], defines a deformation retraction of X to the point Q, so that X is necessarily contractible. We now digress briefly to present a technical result which will be needed to justify a calculation below. First suppose F:R"I---.R m2 is a Lipschitz map, for a point Pc R "1 and a vector V we define the directional derivative D v F (P) by D v F (P) = lira F (P + h V) - F (P) h-*O h assuming the limit exists. Lemma 2.1. - Assume 7 : [a, b] --* R ml is absolutely continuous and F" R"I ~ Rm2 is Lipschitz. Then F ~ is absolutely continuous and at any point t o e(a,b) at which both 7 and F ~ 7 are differentiable it follows that D r, (to) F exists at 3' (to) and o r -- (F 3') (to) - Dr, .o) F (3' (to)). In particular this holds for almost all t o ~ [a, b]. Proof - That F ~ is absolutely continuous is clear. Since 3' is differentiable at t o we have 3' (to + h) = 3' (to) + 3" (to) h + o (h). Since fo 7 is differentiable at to we have the existence of the limit F(V(to+h))-F(3'(to)) lim h--*0 h Since F is Lipschitz, we have from above F (3' (t o + h)) = F (3' (to) + 3" (to) h + o (h)) = F (3' (to) + 3" (to) h) + o (h). 190 HARMONIC MAPS INTO SINGULAR SPACES It follows that Dr, (*o)F (7 (to)) exists and is equal to (F o 7)' (to) as required. This proves Lemma 2.1. Coming back to our situation, let us assume that u e H 1 (fl, X) is energy minimizing and that its image lies in a compact subset K of X. Observe that the map R:[0, 1] x K--. R N is Lipschitz with respect to distance measured along X. On the other hand, on compact subsets of X we have d(P~, P2)~< c[P~- P2 I, so R is Lipschitz in the Euclidean sense. By Kirzbraun's theorem (see [F1, 2.10.43]) R may be extended as a Lipschitz map from all of R N+I into R N. We now consider the family of maps u~ (x)= R1-< r o (u (x)) for x~ where z/> 0 and ~ is a nonnegative smooth function with compact support on ~. (We assume that "c is so small that 1- ~ (x)e [0, 1] for all xe~.) It is easy to see that u,~H ~ (fl, X) since u~ is a composition of a Lipschitz map with an H 1 map. We will justify the following equality of distributional derivatives (2.1) c3u~ 04 ORx-~; o Oxi (x) = Do./oxi (x) RI - < (~), 0 (u (x)) - ~ c~xi c~L " (u (x)) for i= 1,...,n. For example we consider the case i= 1, and observe that (2.1) is a local result near a given x0 efL Consider a neighborhood of x0 of the form I~ x C where Ii=(x~-r,x~+r) for some r>0, and (9 is an open subset of R "-1= { (0, x 2 ..... x") }. Denote by 2 the point of (9 whose final (n-1) coordinates are those of x. It then follows that for 5 ~ ~ almost every point 2 of (9 the map from 11 ~X given by t~--~u(t, 2) is in HI(I1, X). For such an 2, let 7:I1 ~R N+I be the map v (t) = (I - r; (t, u (t, It then follows that u~(t, 2)=Ro7(t). Since an H 1 map of an interval is equal a.e. to an absolutely continuous map, we may assume by redefining u on a set of measure zero in 11 x 0 that the map 7 is absolutely continuous for almost every 2eC. We then apply Lemma 2.1 to conclude that for almost all t~I 1 we have Our (t, 2) = D v, {,) R (7 (t)). 0xl Now 7, (t)=(_z c~ (t,2), Ou -~ 2)~, at(t'ax, 0x, .@= and hence (2.1) follows. Next we observe that the function P ~ d 2 (P, Q) is Lipschitz on X, and therefore its restriction to the compact set K has a global Lipschitz extension to R N. Thus by Lemma 2.1 the chain rule calculation M. GROMOV AND R. SCHOEN 191 -- d 2 (u (x), Q) = D0,/0x, d 2 (u (x), Q) ~xf is justified. We also have for a tangent vector V to X at P 1D D v R~., Q (P). Rx, Q (P) = ~ v d2 (R~, Q (P), Q) provided the indicated diectional derivatives exist. (Both sides are equal to d((R~, Q (P), Q) D v R~, Q (P). 7' (Rx, Q (P)) where 7 is the unit speed geodesic from Q to P.) Thus we square (2.1) and use this result on the cross term to get ~i(X) 2=[Dou/Ox(x) Rl_,r 2 -- "c 0~ ~d 2 (R 1 _< (:o, Q (U (X)), Q) ~?xi ~?xi axi/ & Q (u (x)) Using the contracting property of R~, Q we thus have E(u,)<<. f, (1 v u 12 dg - z fa V 4- V d 2 (R 1 _,~ ~:,~, o (u (x)), Q) d~t + 0 (z2). Since u is minimizing we therefore have 0~< -2~ fn 4[ V u 12 dp +z fn (At) d2 (u, (x), Q)dp +0 ('1~2). It follows that for every smooth nonnegative function ~ with compact support in f~ fn [(a~) d 2 (u (x), Q) - 2 ~ IV u 12] d~t >~ 0. We restate this as a formal result. Proposition 2.2. - If u~H 1 (~, X) is energy minimizing and has image lying in a compact subset of X, then the function d 2 (U (X), Q)for any Q ~ X satisfies the differential inequality Ad z (u (x), Q) - 21V u [ 2 >~ 0 in the weak sense. 192 HARMONIC MAPS INTO SINGULAR SPACES As a first application of this result, choose ~ (x) to approximate the characteristic function of a small geodesic ball B, (xo) centered at a point xo ~ f~. We then get for almost every cy (2.2) 2fB Ivu[2dg~f ~r(d2(u(x), Q))dZ. We now derive the usual monotonicity formula for harmonic maps which can be done for minima of our problem. Note that the nonpositive curvature condition on X is not needed in this derivation 9 Let ~ (x) be a smooth function with support in a small neighborhood of a point x o e f~. For [ x I small consider the diffeomorphism of f~ given in normal coordinates by F~ (x)= (1 + z~ (x))x in a neighborhood of 0 with F~ = id outside this neighborhood 9 Consider the maps u, = u o F t. These are clearly in H ~ (FL X), so the function x~--~E(u~) has a minimum at z=0. To analyze this condition we perform a change of variable as we did in the geodesic case in section 1. We assume that ~(x) has compact support in Bo(0) so that we may work in a single normal coordinate chart. We then set y = F, (x) and use the chain rule to compute [ V u~ [2 (x) = ~ g~J OYk OYt ( Ou O~y~ ) i,~,k~t OX ~ OX J Oyk The volume dement x/g dx then becomes x/g dx : det ( ~xV ~ x/g dy. \ ~yq / Thus we may write the energy of u~ in the form fB ]Vu~lZdg=fB ZaiJ(y, "c) Ou Ou dy, o co) u (o) ~Y~ c?Y j where a ~j (y, z) is a smooth function of y and z. Thus it follows that z ~ E (u~) is a smooth function of ~, and its derivatives may be computed by differentiation under the integral sign. In particular we have E' (0) = 0, and this gives us by direct calculation ~o) i ~ xi +2 Z xJ 0. 0u i, i, k Ox i ~?x~ ~x k + Remainder, M. GROMOV AND R. SCHOEN 193 where the remainder term arises from the fact that the metric is not exactly Euclidean. The remainder term is given precisely by fB I -~ ~ (~gij X k Olg OU ~/g_]_]Vul2 ~/~ X i k " (~X t o(o) ax axJ -- Observe in particular that this term is bounded by a constant x 0 .2 EBo (0)(u). Taking to be an approximation to the characteristic function of the ball B o (0) we get (2.3) 0 = (2- n + 0 (0"2)) ~ IV b/[2 dB dB o(0) +0. ]Vu12d ;-20- f~ f ~U 2 B o (0) 8B o(0) ~r We now introduce the notation E (o), 1 (0.) defined by E(o)= fB IVul2d~t, I(0.)= f~ d2(u(x), Q)dZ (x). o (0) B o (0) Since we are working in normal coordinates observe that if f(x) is a nonnegative function we have 0fdZ + (n - 1) 0. - 1 .]ofs. (o) f dZ (o) Or +0(0-) f e fd'Z B o (0) We now compute logarithmic derivatives I' (0.) _ n - 1 + (I (0.)) - 1 f (d 2 (u, Q)) dZ + 0 (0.). (0) Or I (0.) 0. JOBo (We should remark that I (0.) is an absolutely continuous function for 0- > 0.) From (2.3) we have E'(0.) _ n-2 +2(E(0.))_1 f e c~u 2dZ+0(0.)" E (0.) 0- Bo (0) 8r Therefore I' (0-) E' (0.) - 1 + (E (o) I (0.))- ~ IE (0.) f ~ ~(d2 (u, Q)) dZ I (0.) E (0-) 0. B~ (0) ~N~Idx.. 194 HARMONIC MAPS INTO SINGULAR SPACES fo #u 2 j -2I(cy) ~rr dZ +0(~) B~ which together with (2.2) implies the inequality (2.4) ~d_d log(I(cy) "]~<2(E(cy)i(cy))_l[(f ~ d(u,Q)~rd(U,Q)dZ) 2 acy \ cy E (~) J u, co) -(foB,~lo)d2(u,Q)d~) (fOBo(O) ~Ur 2d~)l -F0(ff)" Since I~rrd(U, Q)~< 0~ , it follows by the Schwarz inequality that __ eq ,,2 ~> 0 (2.5) d I ~E(~) 1 dcy I (cy) for a constant c 1 depending on the metric g. Of course (2.5) holds only under the assumption that I(~)>0 for cy>0. Notice however that from Proposition 2.2 it follows that the function d 2 (u (x), Q) is subharmonic, so that if I (~)=0 for some cy > 0, then the map u is equal almost everywhere to Q in a neighborhood of 0. For any x e f~, cy > 0, Q ~ X we define an order function Ord (x, ~, Q) by ~ Ivul2d a (x) Ord (x, ~, Q) = e q .2 f0 a 2 (u (x), Q) d's (x) B~ (x) The reason for this notation is that for a harmonic function lim Ord (x, cy, u (x)) = Order (u - u (x)), that is, the order with which u attains its value u (x) at x. Alternatively, it is the degree of the dominant homogeneous harmonic polynomial which approximates u-u(x) near x. In particular, for harmonic functions (or harmonic maps into smooth manifolds of nonpositive curvature) this limit is a positive integer. Generally, if x e f~ and cy > 0, then the function Q w-~ fo d2 (u,Q) dZ B~ (x) M. GROMOV AND R. SCHOEN 195 is a convex function on X with compact sublevel sets, and hence has a unique minimum point Qx,, e X. The function Q ~ Ord (x, cy, Q) thus has a unique maximum point at Qx, ~. We now define Ord (x) by Ord (x)= lim Ord (x, or, Qx, ~). o- ---~ 0 This limit exists because the function cy ~ Ord (x, cy, Qx, ,) is monotone increasing in cy. Moreover, for a fixed cy > 0, the function x ~ Ord (x, or, Qx, ,) is a continuous function, and hence it follows that the function x ~ Ord (x) is upper semicontinuous since it is the decreasing limit of a family of continuous functions. We now prove the following result. Theorem 2.3. - Suppose u e H 1 (fL X) is an energy minimizing map with image in a compact subset of a nonpositively curved complex X. Then u is (equal a.e. to) a locally Lipschitz map. Proof. - Since u e H 1 (f2, X), it is approximately differentiable in the sense of Lemma 1.3 at almost every point of f~. Consider a point x o in the closure of the set of points at which u has nonzero approximate derivative. By Lemma 1.3, x o is a limit of points xj at which Ord (xj) ~> 1. Therefore by the upper semicontinuity of the Ord (.) function it follows that Ord(x0)>~ 1. Let a=Ord(x0) and fix CYo>0 so that B,o (x0)e ~. Let ~1 e (0, %), and note that the monotonicity of the ratio implies Of B IVul2dI-t>/~e-ClC2 fo d2( u, Q1)dE (xo) Bo (x0) for all ct~ [%, Cyo) where QI = Qxo, ~1 Combining this with (2.2) yields e_Cx ,2 1 (cy) ~< ~ cy (d 2 (u (x), Q0) dE (x) B~ (x0) ~< - (c~ I' (cy) - (n - 1) I (cy)) + 0 (cy 2) I (~). This implies I'(cy) > n- 1 +2a 0 (cy), I (~) where ~ is any radius in [%, Go). Integrating from % to cy o and fixing % we obtain ; ~"- ~)I (~ 1) ~< c ~ ~ I (%). 196 HARMONIC MAPS INTO SINGULAR SPACES Since the function d 2 (u(x), Q1) is a subharmonic function (Proposition 2.2) on ~, the mean value inequality implies sup d 2 (u (x), Q1) ~< c o2 ~. x ~ Bcl/2 (x0) (Note that I (Oo) is bounded independent of Q1 because u (~) _ K.) In particular, by the triangle inequality, d(u(x), U(Xo))<,d(u(x), Q1)+d(u(Xo), Q1)~<2co] for x EBat/z (Xo). Thus if x EBoo/2 (Xo), we may choose cy I =2 [X-Xo[ and conclude d(u (x), u (Xo)) ~< c Ix- x o [" for x ~ Boo/2 (Xo). Since ~ >/1 this certainly implies that for any x 0 at which the approximate derivative exists and is nonzero we have ~x~ ~< c, i= 1 .... , n. It follows that u has bounded first derivatives locally in f~, and hence u is (equal a.e. to) a locally Lipschitz function. This completes the proof of Theorem 2.3. Remark. - The previous result leaves open the possibility that u might be constant on an open subset of ~. We will show in the next section that this does not occur. Thus it will follow that Ord (x) is defined for all x ~ ~. We will need the following local estimate in order to apply compactness arguments to gain more detailed information about harmonic maps. Theorem 2.4. - Let ueHl(Bl(0), X) be a least energy map (with image in a compact subset of X) for some metric g on B 1 (0)={ x~R":[x]~< 1 }. There is a constant c depending only on g (e.g. on the C 2 norm of the matrix valued functions (gij (x)), (gij (x)) such that sup IVul2 c IVulZd~. B1/2(0) dB 1 (0) Proof. - By Theorem 2.3 the map u is locally Lipschitz. We need to estimate its Lipschitz constant. We first observe that we can replace X by a dilated complex g X where we assume by translation of coordinates in R N that u (0)= 0. The complex g X still has nonpositive curvature, and we may choose g so that the map pu has energy equal to 1 on B 1 (0). Thus we may assume without loss of generality that I vu[ d =l" 1 (o) We also observe that it suffices to prove IV u 12 (0)~ C where we may assume that x = 0 is a point of approximate differentiability of u. This follows just by changing the M. GROMOV AND R. SCHOEN 197 center of balls. Now we may of course assume V u(0)50, and hence Ord (0)>/1. It follows that Ord (0, cy, Qo, ~) ~> 1 for cye (0, 1), and therefore fo d2(u' Qo,1/z)clZ~CfB IVulZdg~c B1/2 (0) 1/2 (0) for a constant c depending only on g. The fact that d 2 (u(x), Qo,1/2) is subharmonic then implies sup d E (u(x), Qo, 1/2)~<c x ~ B1/4 (0) for a new constant c. In particular it follows that the distance from Qo, ~, to 0 = u (0) is bounded for eye(0, 1/4). Taking Oo=1/4 we may then apply the argument of Theorem 2.3 to show that for xeB1/8 (0) we have from above d:(u(x), O)~<c[x[ 2 sup f~ dZ(u, Qo,~)d'Z~<c[x] 2. ~ (0, 1/4) B1/4 This gives the desired conclusion and completes the proof of Theorem 2.4. Remark. - For harmonic maps into smooth manifolds of nonpositive curvature the conclusion of Theorem 2.4 is a well known result of Eells and Sampson [ES]. The usual proof of this is based on the Bochner formula for the calculation of A IV u 12. This proof seems to rely heavily on the smoothness of X whereas the proof we have given is a "lower order" proof which works in a setting which allows X to be singular. 3. Approximation by homogeneous maps To begin this section we consider the case when the image complex X is a geometric cone in RN; that is, if Q e X, )~ e R +, then L Q e X. Under this assumption it is natural to consider energy minimizing maps u: R"~ X which are homogeneous of some degree 0~>~0. We make the standing assumption throughout the remainder of this paper that X has nonpositive curvature. Thus we assume that u ()~ x)= L~u (x) for x eR", )~>0. From the results of section two we know that the map u is locally Lipschitz on R" and hence if the map u is not identically zero we must have 0t ~> 1. There is a special class of homogeneous maps which we call regular homogeneous maps. To describe these we consider an embedding J:R m--,X which is isometric and totally geodesic. This means that d(J(x), J(y))= Ix-y[ and the image of a line under J is a geodesic in X. (Note that the set J (R m) need not be a plane in RN.) Now suppose v: R"--,R m is a homogeneous harmonic map. This simply means that v (x)= (v 1 (x) ..... Vm (X)) where each v i is a homogeneous harmonic polynomial of a 198 HARMONIC MAPS INTO SINGULAR SPACES given degree ~. If the map u = J o v is homogeneous, then we refer to such a map u as a regular homogeneous map. It is regular in the sense that it can be described in terms of a smooth (in fact polynomial) map. A priori it would seem that regular homogeneous maps would be quite rare. However, the next result implies that they occur in abundance. Proposition 3.1. - A homogeneous minimizing map u : R" ~ X is regular if it is of degree 1. Before we give the proof of this result we need to introduce a new concept. We will say that a minimizing map u:B 1 (0)--+ X is intrinsically homogeneous if there is ~> 1 such that for x ~ B 1 (0) we have d(u (x), u (0)) = Ix 1~ d(u (x/[ x [), u (0)), and for each xEOB 1 (0) the curve t~--~u(tx) is a geodesic in X. The following result gives us a simple criterion which guarantees that a map is intrinsically homogeneous. Lemma 3.2. - If u: B 1 (0) --+ X & a minimizing map from the unit ball in R" with Euclidean metric such that for each eye(0, 1) we have Ord(0, cy, Qo, o)=~ for some fixed ot >~ 1, then u is intrinsically homogeneous of degree ~. Proof - Since the domain metric is Euclidean, inequality (2.4) holds without the 0 (cy) term. Next observe that since u is Lipschitz we have lim Qo, o = u (0) so that o'~0 if we first fix cy 0 small we have for cy ~ [~ 1) Ord (0, o, Qo, oo) ~< Ord (0, o, Qo, o) by the maximizing property of Q0, o. Since the right hand side is equal to ~, and the left hand side is equal to ~ for cy = cy o and is monotone increasing, it follows that Ord(O, o, Q0,o0)=0~ for o~[~ 0, 1). Letting cy o tend to zero we see that Ord(0, cy, u(0))=~ for all ~(0, 1). We now apply (2.4) with Q = u (0) so that the left hand side vanishes. We have (k,0 (; Bo,0 ' )2 = d(u, u(O)) d(u, u(O)) r B o (0) M. GROMOV AND R. SCHOEN 199 It follows that (O/Or)d(u, u(0)) = [Ou/&c[ a.e., and that for c~e(0, 1), there is a constant h (cy) such that (O/Or) d(u, u (0)) = h (cy) d(u, u (0)). Integrating the first equality along the ray 7 : (cy, 1)~ R" given by 7(r)= r~ for some ~ e 0B 1 (0), we find L (u (7)) = d(u (r u (0))- d(u u (0)) d(u (r u In particular it follows that u(7) is a geodesic path in X. We now return to (2.2), and observe that since equality holds in (2.4), we must also have equality in (2.2). This then gives us B~ (0) On the other hand we have E (~)= 0~cy-1I (0"), SO we conclude that h (~)= ~o "-1. We may then integrate along a ray from x to x/[ x[ to obtain \ \ Ixl This completes the proof of Lemma 3.2. Proof of Proposition 3.1. - Suppose u:R"~ X is homogeneous of degree I. It follows immediately that Ord (0, cy, u (0)) is a constant independent of cy, and in fact that this constant is one. (To see this, observe that equality holds in (2.4) and (2.2) while (O/Or) d (u, u (0)) = r- 1 d(u, u (0)).) Since Qo, ~ approaches u (0) as ~ approaches 0, we have, for cy >/cy0, 1 ~< Ord (0, %, Qo, %) ~< Ord (0, ~, Qo, %). Letting % ~ 0 we then have Ord (0)= 1. The homogeneity of u then implies that Ord (0, cy, Q0, ~) is a constant independent of cy, and hence this constant is identically one. Therefore we have Qo,, = u (0) for all cy. Because x ~ Ord (x) is uppersemiconti- nuous, at least one for all x, equal to one for x= 0, and homogeneous of degree zero, we have Ord (x)= 1 for all x e R". On the other hand we have, for any k > 0, Ord (x, cy, u (x)) = Ord (k x, k cy, k u (x)), so we may take Z = cy-1 and conclude lim Ord (x, cy, u (x)) = 1. ~ ---~ oo It now follows that for all x~ R" and all cy > 0 we have Ord (x, cy, u (x))= 1. From Lemma 3.2 we conclude that u is intrinsically homogeneous of degree one about every point. It follows that the restriction of u to any line parametrizes a geodesic in X with 200 HARMONIC MAPS INTO SINGULAR SPACES constant speed. The fact that equality holds in (2.2) for Q=u(xo) on B~(xo) for arbitrary cy implies that for each x 0 the function d 2 (u, u (x0)) is a weak solution of A~ d 2 (u (x), u (Xo)) = 2 [ V u [z. Since for all x o ~ R" the function x ~ d 2 (u (x), u (Xo)) is homogeneous of degree 2 about x o, it follows that IV u]Z is homogeneous of degree zero about Xo for every Xo. Therefore I V u 12 is a constant, say e 0. Thus it follows that x l---+d 2 (u(x), u(0)) has constant Laplacian and quadratic growth. This function is therefore a quadratic polynomial. Since the function is everywhere positive and vanishes quadratically at x = 0 we must have d 2 (u (x), u (0)) = ~ gij xi xJ i,j=l for an n x n symmetric G = (gi). This matrix is positive semi-definite, and we can find an orthonormal basis e 1 .... , e, for R" such that tejGe~=)~i aij with )~>0. By reordering el,..., e,, we may assume that )~i>0 for i= 1 .... , m and X~=0 for i=m+ 1,..., n. By change of coordinates we assume that the x 1, ..., x" are coordinates associated to the basis e~, ...,e, so that d 2 (u (x), u (0)) = (xi) 2. i=1 Let v : R" ~ R m be the linear map given by v (x 1 ..... x") = (kl/2 x 2 ..... kid 2 xm), and let J : R" --+ X be given by j (yl, . .., y,,) = u (k; 1/2 Yl, 9 9 XZ, 1/2 ),,,, O, . .., 0). We then have u = J o v, and m m d 2 (J (y), J (0)) = ~ )~i ()~i- 1/2 yi)2 = ~ (yi)2, i=1 i=1 so that J is an isometric totally geodesic embedding. (Note that J is an embedding because J (x)= J (y) implies that the image of the segment xv is a geodesic with the same initial and final point, thus x=y.) This completes the proof of Proposition 3.1. Now we return to the general situation of a minimizing map u : ~ ~ X. Given a point x o e~, we will attempt to approximate the map u near x 0 by a homogeneous map. We choose coordinates so that Xo=0 and U(Xo)=0. We next observe that Ord (0)= lim Ord (0, cy, u (0)). To see this, note that Ord (0)>~ lim Ord (0, cy, u (0)) by ~--*0 a--+0 M. GROMOV AND R. SCHOEN the choice of Qo,,. On the other hand Qo,, approaches u(0) as cy tends to zero, so given s > 0 and % > 0 we have for cy sufficiently small Ord (0, ~, Qo, ,) ~< Ord (0, cy o, Qo, ~) ~< Ord (0, ct o, u (0)) + s. Since % is arbitrary, it follows that Ord (0)~< lim Ord (0, cy, u (0)) as required. Now o'-*0 let ~ = Ord (0) and fix a normal coordinate chart on B~o (0). For )~, g> 0, define the map u~,, (x)= g-lu (X x). This is then a minimizing map from B~-~o (0) with metric g~ (x) = g 0~ x) to the complex g- 1 X = { g- ~ P : P e X }. Notice that the complex g- ~ X again has non-positive curvature since distances are multiplied by a constant factor. We have, by a change of variable, (0) ;~ (0) Bcr (0) B;~ (0) In particular, Ord"~, ~ (0, cy, 0)= Ord" (0, Zcy, 0) for any o e (0, Z-~ Cyo). For any small )~>0, let g=(;O-"I0~)) ~/z, so that we then have f0 d 2 l x (uz, ,, 0) dZoz = 1. B I (o) Since Ord"~,~(0, 1, 0) tends to 0~=Ord"(0) as )~ ~ 0, we also have fB 2 1 (0) for ?~ small. Thus u~, ~ has uniformly bounded energy and then by Theorem 2.4, has uniformly bounded Lipschitz constant on compact subsets of B 1 (0). Thus for any sequence {Xi} tending to zero, the corresponding sequence of maps {u i} has a uniformly convergent subsequence, again denoted { ui }, which has Lipschitz limit which we denote u. : B 1 (0) ~ X o where X o denotes the tangent cone of X at 0. The next result shows that u. is a nonconstant homogeneous minimizing map of degree ~. We will refer to such a map u. as a homogeneous approximating map for u at the point 0. Proposition 3.3. - The map u. is a nonconstant homogeneous minimizing map of degree o~. 202 HARMONIC MAPS INTO SINGULAR SPACES Proof. - To show that u, is nonconstant, we will show that I"~(~)>~o o'a for o~(0, l) and constants go, ~,. We are using the obvious notation I"i (o) = fo d2~ 1 x (bli, O) d~o~ i. B,~ (0) Since u~ converges uniformly to u., it follows that u. is nonconstant. To derive this lower bound, let 0 ~ (0, 1) and r o ~ (0, 1]. We then have, by an obvious estimate, r o I"'(ro)-I'i(O)= i ~d I"(cY)dcY 30 do ~< JB~ o (0)- BO (0) 4-c fB,o d.2~ - ~ (u~, O) d~9,. (o) (o) Using the bound 2 ab ~ g a 2 4- ~3- 1 b 2, we then have I ui (ro) - I" (0) ~< a E "i (ro) + c ~- 1 f0 0 I,i (O) do. Since Ord"~ (0, r o, 0) is bounded above, we may fix e and obtain I"~ (ro) - I'i (0) ~< I "i (r0) + c I"' (cy) do. ;0 ~ This implies, since r o ~ (0, 1] is arbitrary, sup IUi(r)~<2I"i(0)+c(1-0) sup I'i(r). re(0,1] r~(0,1] Therefore, we may fix 0 close enough to 1 so that c(1-0)= 1/2 and we then obtain IUi(l)~<4I'i(0). This gives a lower bound on IUi(0) since I"i(1)= 1 by choice of Xi, ~ti- This already implies that u, is nonconstant. The lower bound I"~(o)>~a 0 o ~1 for all o E (0, 1) follows by iterating the previous argument. In order to show that u, is minimizing, we use the fact that there is a bi-Lipschitz 9 R N RN map, for any G0 > 0, Fi (laT* X) ~ Boo (0) ~ X0 5) Bo o (0) with F~ (0) = 0 and with Lipschitz constants of both Fi and FT* approaching 1, and Fi converging to the identity as i-* ~. Let v be a minimizing map from Bo(0 ) into X 0 with u, =v on ~Bo(0 ) for some o~(0, 1). We must show that E(u.)~E(v), so that u. is also a minimizer 9 To see this, consider the map F i ~ o v : B o (0) ~ ~- 1 X. Let o 1 ~ (o, 1), and observe that for i sufficiently large we have d~Tlx(Ui(X), F7 l~ smaller than M. GROMOV AND R. SCHOEN 203 (~1--(Y for all xc0B,(0). We may then extend the map FF 1 o v to the annular region B,t (0)- B, (0) so that it agrees with u i on the outer boundary. We do this by choosing for each ~eS "-1, the constant speed geodesic 7~:[cy, %]--,gf~X which satisfies 7~ (cy) = F F 1 (v (cy~)) and ?,~ (%) = u i (% ~). We then define vi: B,~ (0) --> gi- 1X by setting vi=F71ov on B,(0), and vi(r~)=7~(r ) for re[cy, %]. The nonpositive curvature condition then implies that the Lipschitz constant of v~ in B,, (0)-B,(0) is bounded by a constant depending only on the Lipschitz constants of u~ and u,. Since u~ is minimizing, we have for any a > 0 E"' (Cyl) ~< E~i (or1) ~< E (v) + ~:, for i large. By lower semicontinuity of the energy it then follows that E (u,) <~ E (v) on B~ (0) as required. This shows that u, is minimizing. Finally we show that for each eye(0, 1) we have lim E"i(cr)=EU*(~). Since i --* c~3 E(v)~<E(u,) on Bo(0) in the previous argument, we have for i large E"~ (~) ~< E u* (c~) + ~. This implies lim E"~ (or) ~< E"* (or) which, combined with lower semi- i ---~ oO continuity, establishes continuity of the energy. In particular, since we have also shown that I"~ (or) has a lower bound for ~ e (0, 1), we can now conclude that lim Ord "i (0, or, 0) = lim Ord u* (0, or, 0). i ---~ oO i-+oO In particular it follows that OrdU*(0, cr, 0)=~ for all eye(0, 1). Therefore u, : B x (0) ~ X o is intrinsically homogeneous of order oz. Since X 0 is a geometric cone in R N, the geodesics from 0 are simply Euclidean rays, and it follows that u, is homogeneous of order ~. This completes the proof of Proposition 3.3. Recall that Lipschitz functions are differentiable almost everywhere. We now generalize the notion of differentiability for minimizing maps into X to exploit the intrinsic geometry of X. We have seen that an intrinsically homogeneous map of degree 1 is essentially a linear map, so it is natural to consider these as derivatives of maps to X. We make the following definition. Definition. - We say that u has an intrinsic derivative at a point x o efl if there is a minimizing map l: Oxo ~ X, (x0)(f~xo = tangent space) which is intrinsically homoge- neous of degree 1 such that lim ]V] -1 ]u (eXpx o (V)) - u (Xo) - I(V) I = 0. V--,O V e ~xo Note that if x 0 is the center of a normal coordinate system x 1, .... x", then this condition reads 204 HARMONIC MAPS INTO SINGULAR SPACES lim Ix1-11 u(x)-u(O)-l(x)l=o. x--+O It is clear that if u has a derivative at x o, then u has an intrinsic derivative at x o, and l = Uxo, the usual linear approximation to u. While the best one can hope for is that u is differentiable at almost every x ef~ (true since u is Lipschitz), we will give general conditions on X in later sections of this paper which imply that u has an intrinsic derivative at every point x E ft. It is immediate that l is unique if it exists since two such maps l, l I would satisfy lim [x1-1 ]l(x)-I l(x) I:o. x~O If we fix a point ~ e S"- l and let Yg (r) = l(r ~) and '~1, { (r) = 11 (r ~), we then have lim r- 1 [Vl, ~ (r) - V~ (r)[ = 0, r ---~ 0 which implies that 71, ~ coincides with 7~ (r) for r small, since these are constant speed geodesics. It follows that l= l I in a neighborhood of x = 0. Note that if ~= Ord (Xo) > 1, then u has an intrinsic derivative at x o because we saw in the proof of Theorem 2.3 that for x near x o we have d(u (x), u (Xo)) ~< c ] x - x o ]~, which implies lim Ix-xol 'l,,o<)-u(.,<o)l=o, X --+ XO so that U'xo = 0. If Ord (Xo)= 1, we are not able to show that u has an intrinsic derivative at x o (although this will be shown under additional assumptions on X in later sections). In order to show this it would suffice to show that there is a unique homogeneous approximating map u, which would then be a nonzero constant times l. We close this section by establishing the result that a minimizing map which is constant on an open set is identically constant. This then implies that Ord(x) is defined for all x e f~ provided u : f~ ~ X is nonconstant. Proposition 3.4. - If ~ is connected, and u" ~ ~ X is a minimizing map which is constant on an open subset of f~, then u is identically constant in f~. Proof. - If u is not constant in ~1 but is constant on an open subset of fl we can find a ball B completely contained in fl such that u is constant in the interior of B, but for some boundary point x oe~B, u is not constant in any neighborhood of x o. Assuming without loss of generality that u-=0 inside B, we may then find a homogeneous approximating map u, at x 0 which satisfies u,-0 in a half space of R". In particular by Proposition 3.1 we know that the degree e~= Ord(x0) of u, M. GROMOV AND R. SCHOEN is strictly greater than one. Since u, is not constant, we may find a ball B 1 in R" such that u,=0 in B 1, and there is a point xleOB 1 with [xl]=l such that u, is not constant in a neighborhood of Xl. As above we must have Ord'*(x0> 1, and so if we let u, 1 be a homogeneous approximating map for u, at x 1, we must have u, independent of a direction in R". (If we take x 1 = (1, 0 ..... 0), then Oul,/Oxl = 0 a.e.). Thus u, 1 restricts to R "-x as a nonconstant minimizing homogeneous map which vanishes in a half space. Repeating this argument a finite number of times we produce a nonconstant minimizing homogeneous map of R to X o which vanishes on a half line. This contradicts the fact from section 1 that energy minimizing maps of R have constant speed parametrization. This contradiction then shows that u must have been constant in f~. This completes the proof of Proposition 3.4. 4. Existence in a homotopy class and uniqueness In this section we make the transition from the local problem of minimizers from a domain with given boundary data to the more global problem of existence of minimizers in a homotopy class. This can be done fairly directly with the help of convexity properties of the energy functional along geodesic homotopies. The approach we follow here was developed in [S]. In particular we will generalize the theorems of Eells-Sampson [ES] and P. Hartman [Har]. A non-simply connected complex X will be said to have nonpositive curvature if its universal covering space X has nonpositive curvature in the sense we have discussed. Given two Lipschitz maps u 0, ul:M ~ X which are homotopic, we can construct a unique geodesic homotopy ut:M ~ X by replacing each parameter curve of any given homotopy by the unique constant speed geodesic with the same endpoints in the same homotopy class. The convexity result which we need is the following. Proposition 4.1. - Each map ut: M ~ N is locally Lipschitz, and for any compact domain f~ ~_ M the function t ~ E n (ut) is a continuous convex function which is a weak solution of the differential inequality dZ En (ut) ~2~ ]Vd(uo, Ul)lZd~. dt 2 j f~ Proof. - For simplicity of notation we assume that M is compact with boundary and Y2 = M. First consider the one-dimensional case in which we have Lipschitz curves 7o, 71:(-6, 6)~X and a geodesic homotopy 7t for 0~<t~< 1. Assume that s=0 is a point at which both 7o, 71 are differentiable and 7~ is differentiable for almost every t~[0, 1]. We are going to do a calculation involving dTt/ds at s=0 only, so we may replace 70, 71 by constant speed geodesics with the same tangent vectors. Observe that at any t for which the original 7t is differentiable, the new curve is differentiable 206 HARMONIC MAPS INTO SINGULAR SPACES and has the same derivative. Let l(s) be the length of the curve t~'yt(s), and observe that l(s)=d(%(s), 5'a (s)) is a Lipschitz function of s. Assume that s~--+l(s) is differentiable at s = 0, and reparametrize the homotopy by setting ~ (s)= 7~/~ (~)(s) for E [0, l (s)]. Thus z ~ ~ (s) is now a unit speed geodesic. The fact that X has nonpositive curvature implies that for any h the function r ~-+ d e (~{~ (h), 7~ (0)) is convex. At any for which d/ds ~,~ (0) exists we have -- 2. lim h- 2 d 2 (~z (h), ~x (0)) = ~ (0) h~0 Since this derivative exists at z=0 and l(0), it follows that there is a sequence hi tending to zero such that the functions z F-+ h7 2 d 2 (~ (t0, ~, (0)) converge uniformly on [0, l (0)] to a convex function which agrees almost everywhere with the function "c~--~[ (dT/ds) (O) [2. In particular by redefining it on a set of measure zero we may assume that this function is convex. Now by the chain rule we have (0)--= ~ d%~/l (0) (0). This may be rewritten in terms of t: dyt (o)= dY~ (O) + tl (O)- l dl~s ) ~t Now for any "I;1, "17 2 ~(0, /(0)) with 2;1 <z2 we have d(~,~l (s), ~[x 2 (S))='r2 2 --2; 1. Differen- tiating with respect to s we then conclude that z=.rl Z Therefore we have, for almost every t e [0, 1], ~st (O)= V (t) + (a + tl (O)- l ~s(O)) ~t (O), for a constant a where V (t). (03,,/Ot) (0)= 0 for a.e. t. Since I v(t)12=l( Jds)(O)12+a2l(O) 2, it follows that Iv(t)[ 2 is a convex function of t. Therefore we have d'~t(O) = 12 +(al(O)+ [ d/(0)~2 7, =Iv(t) \ / and hence it follows that, in the weak sense, -'Cl(O)-2d~f(O)~'Yz/l(O)(O)§ M. GROMOV AND R. SCHOEN ~s_s(O) )~>2( d 2 ~s d(7o (s), 71 Now to prove the result in higher dimensions, observe that the map (x, t)~ u t (x) is Lipschitz, and hence for almost every line parallel to the t-axis it is differentiable at almost every point of the line. At such points of differentiability the previous results tell us d 2 ~t~(lVu, 12)>21Vd(uo, u,)l 2 in the weak sense. Thus if ~ (t) is a smooth nonnegative compactly supported function in (0, 1), we have, for almost every x E M, llVutl=(x)U(t)dt>2 IVd(uo(X), ul(x))12r ;o ;o Integrating and interchanging the order of integration we have bll)[2d~)f~(t) dt. This completes the proof of Proposition 4.1. We now derive two easy corollaries of this result. The first tells us that in case u o = ut on a nontrivial boundary and Uo, u I are nearly minimizing, u o is close to u~. Corollary 4.2. - Suppose u o, u~ :f~--* X are Lipschitz maps which agree on Of~ (Of~ ~ (25) and are homotopic through maps which are fixed on Of~. Let E o be given by E o = inf { E (v) : v : f~ --* X Lipschitz, homotopic to u o with fixed boundary }. Suppose ~o>~0 such that E(uo), E(u0~<Eo+ao. Then it follows that ~< d~t Ul) fnd (Uo, c ~o for a constant c depending only on fL Corollary 4.3. - Suppose X is simply connected. There is a unique minimizing map u : f~ ~ X with given Lipschitz boundary data. folE(ut)~"(t)dt~2(IMIVd(uo, 208 HARMONIC MAPS INTO SINGULAR SPACES Proof of Corollary 4.2. - Connect u o to ua with a geodesic homotopy ut, and let E (t) = E (ut). Define ~ >~ 0 by Iv d(uo, uO? Proposition 4.1 then implies that the function E (t) lies below the appropriate quadratic polynomial with leading term ~ t 2. This implies E(t)<<, o~t(t- 1)+E(1)t+(1-t)E(0). Since E(t)~>E o and E(0), E(1)~<Eo+e o, it follows that for every t this quadratic polynomial has value between E 0 and E o + eo- Setting t = 1/2 we find Eo~< 1_o~+ 12E(1)+ I_E(0)~ < 1 - - -~+Eo+~o 4 2 4 which implies ~<4~ o. Combining this with the Poincar6 inequality we then obtain the conclusion of Corollary 4.2. Proof of Corollary 4.3. - This is almost immediate modulo a minor technical detail; if u o, u 1 are both minimizing and equal to a given Lipschitz map q~ on c3f~, we know that u 0, ul are locally Lipschitz but not necessarily Lipschitz up to the boundary. First observe that Proposition 4.1 works for such maps since its proof involved local expression. Thus we have L ]Vd(uo, ul)12dta=0, and therefore integrating d(uo, ul)= const. Since Uo = u~ on 0~ and d(.,.) is a Lipschitz function it follows that Uo-ul, as required. This proves Corollary 4.3. We are now in a position to solve the homotopy problem for harmonic maps into nonpositively curved complexes. Theorem 4.4. - Let M be a compact Riemannian manifold without boundary, and let X be a compact nonpositively curved complex. Let (p" M ~ X be a Lipschitz map. There exists a Lipschitz map u: M ~ X which is freely homotopic to q~ and which minimizes energy in the sense that E (u) = inf { E (v)" v" M ~ X, Lipschitz, v homotopic to q~ }. Moreover, on simply connected regions f~ c M, the lift of u to the universal cover minimizes in the sense of the previous sections. Proof. - Let {u i} be a sequence of Lipschitz maps homotopic to q~ with E (ui) ~ Eo, where E o is given by M. GROMOV AND R. SCHOEN 209 E o = inf { E (v) : v Lipschitz, homotopic to q) }. Assume that X is embedded in ~N, and choose a subsequence, again denoted { u~}, which converges weakly in H t (M, X) to an H 1 map u. We claim that u is (equal a.e. to) a Lipschitz map homotopic to q~. First let x o e M and consider a small ball B. We may lift the map u~ to the universal cover X. Denote this lift by u~: B ~ i~. Let v i be a minimizing map from B ~ ~2 which is equal to u i on 0B. We then define a replaced map u i by ui (x) = { n (vi (x)) for x e B u~ for xeM-B. Since E (ui)~< E (ui), the sequence { ug } is again a minimizing sequence. In particular, given any a > 0, for i sufficiently large we have E o ~ E (ui) ~< E (ug) ~< E o + ~. Applying Corollary 4.2 in the region B we have fa d2 (Ui, di) dg <~ c ~. On the other hand by Theorem 2.4, the sequence u~ is uniformly Lipschitz on compact subsets interior to B. Thus a subsequence of { ~ } converges uniformly in a neighborhood of x 0 to a Lipschitz map u. Since ~ above was arbitrary we have u = u a.e., and hence u is a Lipschitz map. The sequence { t)'i } is then a minimizing sequence which converges uniformly near x 0 to u. If we consider the geodesic homotopy iz, t with iv o = q0, ~v 1 = ui, we see immediately that this homotopy converges in a neighborhood of x o to a geodesic homotopy vt with v o = % Vl = u. If we consider an overlapping ball, since replaced maps are uniformly close on the intersection, the corresponding geodesic homotopies agree. Therefore we have a global geodesic homotopy from q) to u. The fact that the lift of the restriction of u to f~ for a simply connected region f~_ M minimizes is a consequence of Corollary 4.3 which shows that the minimizer must agree with the lift of u. This completes the proof of Theorem 4.4. 5. Some smoothness results for harmonic maps For the main applications of this work it will be important to show that harmonic maps are better than Lipschitz in certain cases. First note that if u:f~--, X is mini- mizing, and for some x o e s u (x 0) is a regular point of X, then the usual regularity theory for harmonic maps (see e.g. [S]) implies that u is C ~ in a neighborhood of x 0. On the other hand, it happens in important cases that even if u (x o) is a singular point of X, the map u may be differentiable at x 0 in a strong sense. Of course differentiability should mean that u is well approximated by a linear map near x o. We have seen 210 HARMONIC MAPS INTO SINGULAR SPACES previously that for harmonic maps, a map which is homogeneous of degree 1 is essentially linear. Thus it is natural to use such maps to approximate a general harmonic map. In fact, homogeneous maps tend to exist only when the image complex is a cone, so we generalize the notion as follows. Let x ~ ..... x" be a normal coordinate system centered at x o, and let r=lx I, ~= x/lx I denote polar coordinates in B~o (Xo). We will say that a Lipschitz map l: B~o (x0) ~ X is essentielly homogeneous of degree 1 if there is a nonnegative function X : S"- ~ --, R and an assignment y~ to each ~ e S"- 1 of unit speed geodesic in X with y~ (0)= P (where P= l(0)) such that l(r~)=y~ (X(~)r) for x = r ~ e B,o (Xo). In short, a map is essentially homogeneous of degree 1 if the restriction of u to each ray is a constant speed geodesic. Of course maps of this type exist in great abundance because they are determined by their restriction to 0Bro (Xo) and by the assignment of the value at x 0. Thus given any Lipschitz map l: OB,o (Xo)~ X and any point P e X, there is a unique essentially homogeneous map of degree 1 which agrees with l on c~B~o (Xo) and sends x o to P. For a point x oe~ and a radius 0>0 such that B,(xo) is compactly contained in ~ we consider the error with which u can be approximated by degree 1 essentially homogeneous maps. Let l: B, (Xo) ~ X be such a map, and consider the quantity d o(u,l)= sup d (u (x), l (x)). x ~ B~ (x0) We then define R (Xo, cy) by R (Xo, cy) = infd, (u, l), where the infimum is taken over all essentially homogeneous maps, I:Bo (Xo)~ X, of degree 1. Since the constant map l (x) - u (Xo) is a competitor and the map u is Lipschitz we have R (x o, o) ~< ccy. Definition. - A minimizing map u:f~ ~ X is intrinsically differentiable on a compact subset K ~ f~ provided there exists r o, c > 0 and [3 e (0, 1] such that R(x, ~)~ccyl+~R(x, r 0) for all x e K, o E (0, ro]. The constants c, [3, ro may depend only on K, ~, X, and the total energy of the map u. Definition. - A subset S _ X is essentially regular if for any minimizing map u : ~ ~ X with u (~) _ S, the restriction of u to any compact subset of f~ is intrinsically differentiable. It is not difficult to see that a closed subset of the regular set of X is essentially regular; this will be explicitly discussed later in the section. More generally, an M. GROMOV AND R. SCHOEN 211 isometrically and totally geodesically embedded submanifold of any dimension in X has essentially regular image. The main result of this section will provide a criterion for a map to be intrinsically differentiable near a point. This criterion will say roughly that the map is well approximated by an essentially homogeneous map of degree 1 whose image is "effecti- vely" contained in a totally geodesic subcomplex X o which is essentially regular. In order to define what it means for a map to be effectively contained in X o we need to introduce some terminology. Given a Lipschitz map l: ?B~0 (Xo)~ X together with points x ~ B e (Xo) and P e X, there is a unique essentially homogeneous map of degree 1 denoted I~, p from B,o (Xo) into X which satisfies lx, p (X)= P and lx, p= l on aB~o (Xo). (Note we are assuming that r o is small so that B~o (Xo) is convex.) Definition. - Let X o be a totally geodesic subcomplex of X, and let l: B~o (Xo) ~ X be a map which is essentially homogeneous of degree 1. We say that l is effectively contained in X o near the point x o if for any ~ > 0 there exists 8 > 0 such that for all x e ~ sufficiently near Xo and all P e X o sufficiently near Po = l(xo) and all (~ ~ (0, to/2 ] we have Vol { y ~ B= (x) : BX, (Ix, p (y)) ~ X o } ~ ecy". We stress that B x(.) is used to denote the full ball in X, so that it follows that for any point y e B, (x) such that BX,, (l,,. p (y)) c_ Xo, a maximal simplex in X o which contains lx, p(y) is also maximal in X. Thus X o should be thought of as a top- dimensional (typically of dimension k = dim X) subcomplex of X. We are now ready to prove the main result of this section. Theorem 5.1. - Let u : ~ ~ X be a minimizing map. Let x o ~ ~ and r o > 0 be such that Bro (Xo) is compactly contained in ~. Let X o ___ X be a totally geodesic subcomplex, and let l: Bro (Xo) + X o be an essentially homogeneous degree 1 map with Po = l(xo) ~ Xo. Assume that a neighborhood of Po in Xo is essentially regular. There exists 5o>0 depending only on l, ~, X, Xo such that if l is effectively contained in Xo near x o and sup d(u(x), l(x))~<8o, x e Bro (xo) then u is intrinsically differentiable in a neighborhood of x o. In fact, there exists cy o > 0 such that u (B,o (Xo)) _ Xo. Proof - The idea of the proof is to compare u at small scales to a minimizing map having image in X o. To carry this out, we let H : X ~ X o denote the nearest point projection map. The map H is then a distance nonincreasing Lipschitz map. We need the following lemma. 212 HARMONIC MAPS INTO SINGULAR SPACES Lemma 5.2. - If veil i (fL X), then v=Hov is also in H 1 (~, X), and E (v) <~ E (v). Thus if 9:~-" Xo is Lipschitz, then the minimizer u o into X which agrees with q~ on 8~ has image in X o. We also need a second lemma. We postpone the proofs of both of these until we have completed the proof of the theorem. Lemma 5.3. - If ui, u z are minimizing maps .from a region f~i into X, then the function x F-+ d(u I (x), u (x)) is a Lipschitz weakly subharmonic function in ~1. Let xt be a point sufficiently close to x o such that u (x 1) lies in X o and such that u (x~) is close enough to u (Xo) so that we may find for any ~ > 0 a number 5 > 0 such that Vol { xE B~ (xO : BX~ (l~ (x)) ~ X o } <act" for eye(0, ro/2] where we have denoted by l I the map lxl,,~xl). By translation of coordinates in R kl we may assume u (xt)--0, and by multiplying the metric on f~ by a fixed constant factor we may for convenience take r o = 2, and hence both u and l 1 are defined on B 1 (Xl). Also note that we have (5.1) sup d(u(x), l 1(x))~<25 o x 9 (Xl) provided x 1 is sufficiently close to x o. We choose a normal coordinate system xl,..., x" centered at x~, and for any map v:B 1 (x 1)--+ X and any ere(0, 1] we let % denote the dilated map given by "v (x) = ~- i v (cy x). Thus we have %" Bo- 1 (0) -+ cy- 1 X, and if v is minimizing, then ~v is minimizing as a map from (B,-1 (0), (y-lg(cyx)) to or-1 X. Note also that the metrics or-1 g(cyx) become Euclidean as ~ tends to zero, and hence are uniformly controlled (in any C k topology) on B1 (0) while the complex cr-~X has nonpositive curvature for each cr and converges to the tangent cone to X at 0 as ~s tends to zero. Now assume that for some eye(0, 1] there exists an essentially homogeneous degree 1 map l 2 : B 1 (0) --+ cr- 1 Xo such that D is defined by (5.2) sup d(~u (x), 12 (x)) = D, x 9 B 1 (0) and for some ~1 >0 assume that (5.3) sup d(5~(x), ~l I (X))~<51. xeB 1 (0) We assume D, 61 are small positive numbers; in fact, for a given ei >0 assume that 51 is so small that M. GROMOV AND R. SCHOEN Vol { x 6 B, (0)" B~( 1 x ((~ll (X)) ~ (3-1 Xo } ~ E1 " In particular from (5.3) we have Vol {x~B 1 (0) : ct/(X) r O "-1X 0 } ~E1. We may therefore choose a number 01 e [3/4, 1] such that Vol { xE~B01 (0):"U(X)q~CY-1Xo } ~<4E1 . Now let "H:cy-IX--,cy-IX o be the nearest point projection map, and let "q) = q-I o "u. We then have (5.4) Vol { x E 0 Box (0) : "q~ (x) #" u (x) } ~< 4 ~l- Let v:B01(0)--*X o be the least energy map with v=%0 on 0B01(0). Since a neighborhood of Po in X o is essentially regular, it follows that for any 0~(0, 1/4) there is a homogeneous degree 1 map [2 :B0(0)~ ~-i X o such that for some [36(0, 1) sup d(v(x), [2(x))<cOl+~Rv(O, 1/2). x ~ B 0 (0) By definition of R v (0, 1/2) this implies (5.5) sup d(v(x), [2(x))<~cO 1+~ sup d(v(x), 12(X)). x~B 0(0) xeB1/2(0) We now show that "u is very close to v. To see this we note that (5.2) and (5.4) imply fe d("u, v)d~<481D. BO1 (0) Since the function x ~ d("u (x), v (x)) is subharmonic by Lemma 5.3, we have sup d(<'u, v) <-% c ~1 D. B1/2 (0) Combining this with (5.5) then gives us sup d("u, [2)~cE1D+cO 1+~ sup d(v, 12) B 0 (0) B1/2 (0) ~(C~.I-J-cOI+[~-]-Cgl 01 + [3) D. This clearly implies (5.6) sup d("u, 12) ~ (C~I + C 01 + [3) sup d("u, 12)" B 0 (0) B 1 (0) 214 HARMONIC MAPS INTO SINGULAR SPACES On the other hand we have, from (5.2) and (5.3) and the triangle inequality, that for xe0B1 (0) d(V, (x), (x)) + 61. Since ~ 1 (0)= ~ (0)= 0, we have from (5.2) d("/1 (0), 12 (0)) ~< D. The fact that X has nonpositive curvature then implies that for any xeB 1 (0) we have d(~ (x), l 2 (x))~D-~lx[ 61, so that, in particular, sup d(~ /2) ~< D + 61 0. B o (0) Combining this again with (5.2), we finally have (5.7) sup d(~ ~ ~< 061 + 2 D. B 0 (0) We now apply the previous argument for varying choices of o. First take cy = 1 and l 2 = l 1. Set 1l = ['2 and rewrite (5.6) as sup d(~ l/)~(Cgl0 l+c0t~) sup d(u, ll). (5.8) B 1 (0) B 1 (0) On the other hand, (5.7) becomes (5.9) sup d(~ ~ <~ 61 + 2 0 1 sup d(u, /1). B 1 (0) B 1 (0) Set 16 = 61 + 2 0-1 Do where we set D O = sup d(u, ll). Assuming we have defined B 1 (0) 1 l, 21,..., il, we set Di= sup d(~ il). B 1 (0) Assume by induction that for integers up to i(i~> 1) z6 has been defined, and that we have the inequality (5.10) sup d(~ ~ <~ i6. B 1 (0) M. GROMOV AND R. SCHOEN Now apply the previous argument with cr = 0 i, l 2 ----i l. This can be done provided i8 is sufficiently small depending on el. Set i+1/=/2 and observe that (5.6) then may be written (5.11) Di+l ~(CE 1 0 -1 + C 013) Di, while (5.7) yields (5.12) sup d(~ 0 i+1 ll)~<iS+2 0 -1D i. B 1 (0) Therefore we may take i+ 1S to be (5.13) i+lS=iS20 -1D i. We now fix 0 so small that r in (5.11), and we then fiX~x so small that ce 1 0 -1 = 1/4 in (5.11) so that we obtain Di+ 1 ~< 1/2D i provided i8 is sufficiently small. Assuming 16,..., kS are small enough we then obtain Dj~<2-JDo, j= 1,...,k+ 1. Putting this into (5.13) then gives j+ 1S~jS--[-2 0 -1 2-JDo, j = 1, . . . ,k+ I. In particular we have k+lS~<xS+20-1Do ~ 2-J~<lS+20-1Do . j=l Recalling the definition of 1 S we have k+IS~<SI +40-1Do, but by (5.1) we have D o ~< 2 So, and hence finally k+IS~S1-]- 8 0-1SO . Thus if 6o, S 1 are sufficiently small we may apply this argument for any k, and conclude from (5.12) sup d(~ ~ ~< S + 8 0- 1 SO B 1 (0) for all nonnegative integers i. Now if cy~ (0, 1], we choose the nonnegative integer i so that cre (0 i+ 1, 0i], and we conclude (5.14) sup d(u, /1)~(0 -1 ~1q-80-2S0) O ". Br (Xl) 216 HARMONIC MAPS INTO SINGULAR SPACES This then implies that most points in B, (Xl) are mapped by u into the interior of X o. Note that Xl was an arbitrary point near Xo such that u(x,)cXo. If for cy o>0, u(B~o (Xo))~ X o, then we can find a ball B_ b/-l(XNXo)(-'] B~o (Xo) such that at least one point x I in c~B maps to X o. By the choice of x 1 we have, for every c h >0, a substantial fraction (at least 1/2) of B~(xx) mapped into the closure of X-X o. This contradiction shows that for some cr o > 0 we have u (B~o (Xo)) ___ X o, we have completed the proof of Theorem 5.1. Proof of Lemma 5.2. - If we have some vector V at a point P e X such that the directional derivative D v H exists, where FI is considered as a vector valued function, then we have [DvH[<~[V [ since H is distance decreasing. Therefore we may use Lemma 2.1 to argue that for almost every point x e f~ we have ~H~ 0z, j= 1, . . .,n. Oxj[<<'[&~cj ' In particular we have E(I-Iov)~<E(v). The last statement of Lemma 5.2 follows immediately. Proof of Lemma 5.3. - First observe that if C _ X is a closed convex set, then the function x~-+d(u(x),C) is weakly subharmonic This was shown in Proposition 2.2 for the case that C is a point. (Note that IvulZ>_-IVd(u, P)L so Proposition 2.2 implies Ad(u, P)~>0.) The general case is an easy modification of this, and we omit the proof. Consider now the function g:X x X--+ N given by g(P1, Pz)=d(P~, P2). We take the product metric on X x X, and we claim that g is a multiple of the distance function to the convex set C__c_ XxX where C = {(P, P):PeX} is the diagonal in Xx X. To see this we observe that C is the fixed point set of the isometry F given by F (PI, P2)= (P2, P1); so that if (Q, Q) is any point of C, then for any path 7 from (Q, Q) to (P~, P2) the path 7 U F (~,) suitably oriented is a path from (P1, P2) to (P2, P1). In particular we have L(y)~> ~d((P1, P2), (P2, P1)) = X/2d(Pl'2 P2), and this is achieved when Q is the midpoint of the geodesic from P1 to P2" Therefore we have shown g(P1, P2)=x/2d((P~, P2), C). It then follows that the function x~--+d(u~(x), U2(X)) is weakly subharmonic, as required. This completes the proof of Lemma 5.3. M. GROMOV AND R. SCHOEN We now present two applications of the previous theorem. We first show that a minimizing map is strongly differentiable in a neighborhood of a rank k (= dim X) point provided a mild regularity condition is satisfied for X. First recall that if Ord (Xo)= 1 for some x o e ~, then there exists a degree 1 homogeneous approximating map u," Dx o ~ Xu (xo)- This map, by Proposition 3.1, is a linear map to a flat totally geodesic subcomplex of X, (~o)" The rank of u, is the dimension of this flat subspace. If a neighborhood of u (Xo) is isometric to a neighborhood of the origin in the tangent cone X, (~o), we then have a flat totally geodesic subcomplex of X containing the image of u, locally. This happens for example if the simplices in X are standard Euclidean simplices. In general we say that u, is a good homogeneous approximating map if there exists a smooth Riemannian metric go given in normal coordinates on the ball B,I(0) _~ Nk and an isometric totally geodesic embedding i: B,~(0)---,X with i (0) = u (Xo) such that the image i (B,1 (0)) is contained in a totally geodesic subcomplex X o whose tangent cone at u (x o) is the image of u,. We now state a theorem. Theorem 5.4. - If u : ~ ~ X is a minimizing map, and x o ~ f2 is a po&t at which u has a good homogeneous approximating map of rank k = dim X, then u is intr&sically differentiable in a neighborhood of x o. In fact, the map u in a ball B,o (Xo).for some cy o > 0 is given by u = i o v where v : B~0 (x o) ~ B1 (0) ~ ~k is harmonic with respect to the metric go described above. Proof. - Recall that u,=ioolo where io:~k--* Xu(xo ) is an isometric totally geodesic embedding of the Euclidean space ~k, and l o : flxo -~ ~k is a linear map of rank k. On a small ball B~l(Xo) we define /:B~l(Xo)-~X by l=iolo where B~(Xo) _~ D is identified with the ball of radius o I centered at 0 in Dxo via the exponential map. The map l is then essentially homogeneous of degree 1. We claim that l is effectively contained in X o. To see this, let X 1 be the subcomplex of X o consisting of those simplices which are faces of a simplex of X which is ,lot in X o. Since no k-simplex can lie in X1, it follows that X 1 is a subcomplex of codimension at least one in Xo. Thus l- 1 (X0 is a subset of B~I (Xo) consisting of a finite number of compact smooth submanifolds with piecewise smooth boundary, each having codi- mension at least one. (Note that n~>k.) It is then immediate that 1 is effectively contained in X o. We next observe that a smooth manifold is essentially regular. For suppose we have a harmonic map v:B~o (Xo)~ N k with bounded energy. We may then assume that the image lies in a normal coordinate ball with coordinates u ~ ..... u k centered at U(Xo). We also assume that xl,...,x n are normal coordinates centered at x o. By Taylor's theorem we have v (x) = l(x) + Q (x), where Co, given by Co =- sup [xl-21Q(x)l, B~o(O) 218 HARMONIC MAPS INTO SINGULAR SPACES is bounded in terms of the second derivatives of v. Since the second derivatives of a harmonic map are bounded in terms of the energy (and the manifolds), we have sup Iv-l[<....Co(y20-o z sup Iv-l[<~Cl ~2 sup [v-ll, B~ (0) B60(O) B~0(0) with cl depending only on the energy of u, the manifolds, and c~ 0. Since we have chosen normal coordinates in both domain and range, t is essentially homogeneous of degree 1, so we have R(xo, 0-)~<c~ 0 -2 sup Iv-l]. Bcso(0) Now if l 1 "Boo (x0) --+ N is any essentially homogeneous map of degree 1 with image near the image of v, then we have for cy 1 < ~o sup d(l, ll)<~c0-? 1 sup d(l, ll), B~o(O) B~I(O) because both l, l I are essentially homogeneous of degree 1 and our metrics are nearly Euclidean. Now we have sup d(l, ll)~ sup d(v, ll)+ sup d(v, l) B~I(0) Bc~0(0) B61(0) ~< sup d(v, ll)+ccy 2 sup d(v, l). B~0(O) B(~o(O) Therefore we have sup d(v, I)~ sup d(v, ll)+ sup d(l, ll) B~0(0) B~0(0) B~o(O) ~<(1 +c0-i*) sup d(v, /1)+CO'1 sup d(v, 1). Boo(O) B~o(O) Now, taking cy 1 so that c0-1 = 1/2, we have sup d(v, l)<~c sup d(v, 11). B~o(O) Boo(O) Since l 1 was an arbitrary essentially homogeneous degree 1 approximation to v, we have finally shown R (Xo, 0-) ~< ccy 2 R (x 0, cy 0) as required. M. GROMOV AND R. SCHOEN Since u, is a homogeneous approximating map for u at Xo, for any given 6 o (take the 6 o determined in Theorem 5.1), there exists a small radius 9o > 0 such that sup [go'U(poX)-U,( x)[<~6o. xeB 1 (0) This implies sup d(u (x), [(x)) <~ 6 o x 9 BO0 (0) where l=i o (go lo). By Theorem 5.1 we now conclude that u is intrinsically differen- tiable near x o, and the image of a small ball B,, o (Xo) under u lies in X o. This completes the proof of Theorem 5.4. The final result of this section will deal with the case dim X = 1. We now state it. Theorem 5.5. - If dim X = 1, then X is essentially regular. Moreover, if u : f~ ~ X is a minimizing map, there exists a constant ~> 0 depending only on n such that for all x o ~ ~ we have Ord (Xo) = 1 or Ord (Xo) >~ 1 + ~. Proof. - Since dim X = 1, any point x 0 s f~ where Ord (Xo) = 1 has a neighborhood in which u is intrinsically differentiable (in fact, defined by a smooth harmonic function to a geodesic of X). In particular we note that the set of points x where Ord (x) = 1 is an open subset, denoted fl,, of f~. In this case one can see explicitly that if Ord (Xo)> 1. then Ord (xo)~> 1 + ~. To see this, first observe that if u (xo) is not a vertex of X, then Ord (xo) >~ 2 since u is a smooth map near x o. Let ~t =Ord (xo), and assume that u (x o) is a vertex of X with at least three edges emanating from Po = u (Xo) (there cannot be one edge by the maximum principle, and if there are only two, then Ord (Xo) >/2 as above). Consider any homoge- neous approximating map u,: N"--, Xpo. If we choose an edge e emanating from Po and introduce an arc length parameter s along e which is zero at Po, then on the open region Oe = { x e ~": u, (x) e e- { Po }} the function h e = s (u, (x)) is a harmonic function. Of course O e is the cone over a domain D e ___ S"-', and he is homogeneous of degree a in Oe- It follows that the restriction of h e to D e is a first Dirichlet eigenfunction of the domain De, and in particular we have )~ (D e) = ~ (~ + n - 2). Since D e is non-empty for at least three distinct edges (otherwise Ord(xo)>~2 as above), there is an edge e for which D e is nonempty and Vol(De)~< 1/3 Vol(S"-1). Standard results about eigenvalues then imply that there exists a number 6,>0 such that )~1 (De)~>n-1 + 6,. It then follows that ~> 1 + ~ for a fixed constant ~ depending only on n. To show that X is essentially regular, we must show that for any minimizing map u : f~ -* X and any compact subset K _~ ~, there exists ro, c, [3 such that 220 HARMONIC MAPS INTO SINGULAR SPACES R(x o, c~)~<c~l+~R(xo, ro) for all x 0eK. Let r 0 be such that Bro(Xo) is compactly contained in ~. It dearly suffices to show that there exists 0e(0, 1) such that for all eye(0, %] and all xoEK we have (0cy)- 1 R (x 0, 0or) ~< (1/2) or- 1 R (Xo, or). This can be proved by contradiction. If this were not true there would be for any 0 e (0, 1) sequences { xi }, { cy, } both of which converge to limits xc K, S e [0, r o] such that (0%)- 1 R (xi, 0oi) > (1/2) of i R (xi, %). We then rescale the maps in the usual way by setting u~ equal to b/i (Y) = ~.l/- 1 ~/i (O'i X) for xeB 1 (0) so that sup d(u 1 (x), ui(O))= 1. We then have B 1 (0) 0- * R", (0, 0) > (t/2) R", (0, 1). We may assume by taking a subsequence that { u i } converges uniformly to a mini- mizing map u: B 1 (0)--+ X where the target complex is either a dilation of X or the tangent cone Xp, where P = lim ui (xi). Now Ord~-(0) = 1 or Ordr >~ 1 + ~, and in the first case Theorem 5.1 implies that for i sufficiently large the map u i is regular in a fixed neighborhood of 0. This contradicts the previous inequality if 0 is small enough. In case Ordr 1 +~, then we know from the proof of Theorem 2.3 that we have sup d (u, u (0)) ~< ccy ~ sup d (u, u (0)). B o (0) B 1 (0) This implies, by ah easy argument, that for 0 small depending on c we have 0 - 1 R';(0, 0) ~< 1/4 R~(0, 1), a contradiction. This completes the proof of Theorem 5.5. 6. Special structure of harmonic maps into building-like complexes In this section we consider complexes X which satisfy a very special hypothesis which we describe shortly. In a certain sense these are higher dimensional generaliza- tions of trees. For us the important property which a tree has is that two adjacent M. GROMOV AND R. SCHOEN edges lie in a geodesic (the union of the two closed edges). We assume that the simplices of X are Euclidean simplices; that is, images under a linear transformation of the standard Euclidean simplex. We make the following definition. Definition. - We say that a nonpositively curved complex X k is F-connected if any two adjacent simplices are contained in a totally geodesic subcomplex X o which is isometric to a subset of the Euclidean space Rk. The most important F-connected complexes are the locally finite Euclidean buildings of Bruhat and Tits (see [BT]). We first want to show that F-connected complexes are essentially regular, thus generalizing Theorem 5.5. We first need two elementary lemmas. Lemma 6.1. - If X1, X 2 are essentially regular complexes, then so is X 1 x X2" Proof. - This follows from the fact that a map u = (Ul, U2)" ~'~--~ X 1 X X2 has energy E(u)=E(uO+E(u2). Thus u is minimizing if and only if both tq, u2 are minimizing. Thus ul, u 2 are intrinsically differentiable on any compact subset of ~, and hence so is u. This proves Lemma 6.1. The next result enables us to find essentially regular totally geodesic subcomplexes which contain any given flat effectively. Since we are interested only in local construc- tions near a point Po~X, and since a neighborhood of Po in X is isometric to a neighborhood of the origin in the tangent cone Xpo, we may replace X by its tangent cone XP0. The fact that X is F-connected then implies that any two simplices (actually simplicial cones) in Xa0 are contained in a totally geodesic subcomplex isometric to the Euclidean space ~k. Let J'~"~ Xpo be an isometric totally geodesic embedding for some m with 1 ~< m ~< k. We may assume that J (0)= 0. If we choose a point x ~ 0, x~ ~" such that neither x nor -x lie in the J-1 ((m-1)-skeleton of Xao ), then any k-dimensional flat F which contains both x and -x must contain the full image J (~"). This is because J- ~ (F) contains cones over a neighborhood of both x and - x, and therefore contains the convex hull of these cones which is ~". It may happen that J (~") is contained in several distinct k-dimensional flats. We need the following result. Lemma 6.2. - Let X o be the union of all k-dimensional jlats in Xpo which contain J(R"). The subcomplex X o is totally geodesic and is isometric to R"x X]-" where X 1 is an F-connected complex of dimension k - m. If L : R" ~ R" is a linear map of rank m, then l: R"~ XPo, given by l=J ~ L, is effectively contained in X o. Proof - Let F1,...,F r denote the k-dimensional flats in XPo which contain J (R"). Let I/: R* ~ Xp0, i= 1,..., r, denote isometric embeddings to the F i normalized 222 HARMONIC MAPS INTO SINGULAR SPACES so that I i (0)= 0 and I i (R")= J (R"), where we denote by R" __c_ R k the plane spanned by the first m standard basis vectors; i.e. Rm= { (.~, 0)'x=(x' .... ,xm), xJ6R forj = 1,...,m}. Similarly we denote by R k-m ~_ R k the orthogonal complement R k-m= {(0, .~):.~=(Xm+',...,xk), xJER forj=m+ 1 .... ,k}. We may further normalize our isometrics so that I1, ..., I, and J are identical on R m since this can be achieved by right composition of each I i with an orthogonal transformation of R m. Each map I i induces on R g a cell decomposition by simplicial cones. We may describe X 0 as the disjoint union of r copies of R g where ceils Xi, ]~j in the i-th, j-th decompositions are identified if I i (Xi) = Ij (E j). By taking the intersection of cells with R", R k-" each cell decomposition of R g induces a cell decomposition of both R m, R g-m by simplicial cones. Since all of the embeddings agree on R m, the cell decompositions of R" all coincide, so we may speak of the induced cell decomposition of R". Now suppose we have two points (x, x--), (y_ y--) in R k such that I i (.}-, x)= I t (y, y). Since the point (x, 0)~ R" is the nearest point of R" to (x, x), and (y, 0) is nearest in R" to (y, y) and the maps Ii, Ij coincide on R", we must have x =y. It also follows that the image of the closed convex hulls of R" U { (-}-, ~{) } and R m k3 { (y, ~) } under I i and I t respectively must coincide in X. In particular we have I i (0, .~)= Ij (0, y). We form a complex X 1 of dimension k- m by taking the disjoint union of the r copies of R k-m with their induced cell decompositions and identifying points in the i-th and j-th copy if their respective images under I~ and I t coincide. The previous argument k - m then shows that X o is isometric to R"� X 1 . We now show that any two cells in X o are contained in some F t. This implies that X o is totally geodesic since the geodesic connecting two points P1, P2 ~ Xo lies in any Fj which contains P1 and P2, and hence lies in X o. It also obviously implies that X 1 is F-connected. Suppose we have a pair of cells in X o which we may assume to be k-dimensional without loss of generality. We may think of these cells as Z~, E~ in two of the cell decompositions of R k given by I~, Ij respectively. Fix a nonzero point (-~-o, 0)~R" such that both (x o, 0) and (-x o,0 ) are interior to m-cells in the cell decomposition of R m. Let C~, C t be the closed convex hulls of R" U Ei, R" k) gj respec- tively. Choose points yi, yj in the interior of C i, C t respectively such that both Yi and yj are interior to k-cells E i, ~ Ej 1 respectively. Assume also that Yi is close to (-{:o, 0) and y~ is close to (-x o, 0). It follows that the m-cell of R" containing (Xo, 0) is a face of Z 1 1 i, and the m-cell containing (-x o, 0) is a face of Z t . Let F be a k-dimensional flat totally geodesic subcomplex of X which contains Ii(Z~) and I~(Zx). Since F contains neighborhoods of a pair of antipodal points of J(R m) it follows that F contains J (R"), and hence F-F~o for some i o with 1 ~< i o ~<r. From the choices we M. GROMOV AND R. SCHOEN 223 have made, F~o must contain some interior point of both I~(Z~) and Is(Es. ), and therefore F~o contains both cells. This proves the desired statement. Finally, suppose l= J o L : R" ~ XPo where L : R" ~ R m is linear of rank m. We want to show that l is effectively contained in X o. To see this, let X 2 be the subcomplex of R" (with the above cell decomposition) consisting of those cells of R m whose image under J is contained in the closure of some cell of X which is not in X o. As in the proof of Theorem 5.4, if we can show that X 2 has codimension at least one in X o then the result follows. Let E be any m-cell of R", and let Z 1 be a cell of X which is not in X o such that J(Z)_ Z 1. Without loss of generality we may assume E~ is k- dimensional. Take an interior point (Xo, 0)e Z such that (-,~-o, 0) is interior to some cell 's of R". Let Z 2 be a cell of X which contains J (]~) in its closure. Then any k-flat of Xp o which contains both E~ and Z2 automatically contains all of J(Rm). This contradicts the fact that Z~ is not in X o. We have completed the proof of Lemma 6.2. Observe that if Poe X, then a neighborhood of Po is isometric to a neighborhood of the origin in the tangent cone XPo. In fact, if we define St (Po) to be the union of all simplices which contain Po in their closure, then St(Po) is a totally geodesic subcomplex of X which is canonically isometric to a neighborhood of the origin in Xpo. Note that if Po is a vertex, then St (Po) is the star of Po. Thus if X o is a totally geodesic subcomplex of Xpo, then we get a totally geodesic subcomplex, which we will also refer to as X o, of St(Po) which is isometric to a neighborhood of 0 in X 0. We now prove an important result concerning minimizing maps into F-connected complexes. Theorem 6.3. - Let X be an F-connected complex. The following three properties hold." (i) For any positive integer n and any compact subset K o of X, there exists ~> 0 depending only on K o and n such that, .for any minimizing map u:~"-~ X with u (~)) ___ K o, we have, for all x o ~ ~, either Ord (Xo) = 1 or Ord (Xo) ~> 1 + s. (ii) Let u:~-~X be a minimizing map, and let Xo~ with Ord(xo) = 1. There exists a totally geodesic subcomplex X o of X, ~xo) which is isometric to R"� X] " .for some integer m with 1 <~ m <~ min { n, k I and some F-connected complex X 1 of dimension k- m such that u(B~o(Xo))_~Xo for some CYo>0. Moreover if we write u = (ul, u2) : B, o (x) ~ R m � X 1, then u I is a harmonic map of rank m at each point of B,o (Xo) and Ord"2 (Xo) > 1. (iii) The complex X is essentially regular. Proof - To prove (i) we first observe that the tangent cone to X at a point P is the cone over the link of the open simplex which contains P. Therefore, there are only a finite number of cones which appear as tangent cones for points P in a compact subset K o of X. Thus to prove (i) we may restrict attention to a single cone XPo for 224 HARMONIC MAPS INTO SINGULAR SPACES some fixed PosKo. Suppose we have a sequence of homogeneous harmonic maps Cui}, u~'R"--.Xpo such that u~ is homogeneous of degree ~i with ~i>l and lim cz~ = 1. We may normalize ui so that i--+~ sup ]u i(x)]=l. x~B 1 (0) With this normalization a subsequence of {ui} again denoted {u i} converges to a map u: R" --. XPo which is homogeneous of degree 1. Thus by Proposition 3.1 there exists an integer m with 1 ~< m ~< min { n, k } such that u = J o L where L" R" ~ R" is a linear map of rank in and J:R" ~ Xp0 is an isometric embedding. By Lemma 6.2 the map u is effectively contained in a totally geodesic subcomplex X o of Xpo which is isometric to R" x X 1. By Theorem 5.1 we have the image of u i contained in X o for i sufficiently large. It then follows that the projection of u i to R" is a harmonic map which is close to a linear map of rank m. Therefore ui must be linear (since it is homogeneous harmonic of degree less than two), and we have a contradiction. This establishes property (i). Property (ii) is an easy consequence of Lemma 6.2 and Theorem 5.1. To establish (iii) we work by induction on k. For k= 1 the result was established in Theorem 5.5. Assume that k~>2, and that all F-connected complexes of dimension less than k are essentially regular. We first prove the weaker result that for any x o e FL there exists r o > 0 such that for c~ ~ (0, r0] R(xo, cr)~<c~l+~R(xo, ro) for constants C, r o, [3 depending on the point x o, the energy of u, and the spaces f2, X. If we can prove this, then the same compactness argument used in Theorem 5.5 will imply that u is intrinsically differentiable on any compact subset of ~2. There are two cases to consider. First suppose Ord(xo)>l; then from (i) we know that Ord (Xo) ~> 1 + ~. This implies, by the proof of Theorem 2.3, that sup d(u(x), tt(Xo))<~e~ 1+'~ sup d(u(x), U(Xo) ) x ~ B~ (xo) x ~ Bro (xo) for some constant c and r o > 0. This easily implies the desired decay on R (x o, cy). The other case we must consider is the case Ord (x o)= 1. In this case the result follows immediately from (ii), Lemma 6.1, and the inductive assumption. This completes the proof of Theorem 6.3. It will be important for our application to do smooth differential geometric calculations for our harmonic maps. To make sense of this we need to refine the notion of intrinsic differentiability. This we now do. 225 M. GROMOV AND R. SCHOEN Definition. - A point x o e f~ is a regular point of u if there exists cy o > 0 and a k-flat F __g_ X, ~:~o~ such that u (B,o (xo)) c_ F. We see in particular that the map u is actually a real analytic map to the Euclidean space R k in a neighborhood of a regular point. We then let ~(u) denote the open subset of f~ consisting of all regular points of u, and we let Oo (u) denote the singular set, 50 (u)= fl- ~, (u). We now prove the following result, Theorem 6.4. - Let u be a minimizing map from ~ to an F-connected complex X. The Hausdorff dimension of 5 e (u) is at most n-2. For any compact sub- domain ~ of~ there is a sequence of Lipschitz functions {q,i} with q,i==-O in a neighborhood of 5 p ~ ~1, 0 <~ qti <<- 1, and ~i (x) --+ 1 for all x ~ f~l - ~ such that lim fnlVVul JV~,]dg=0. i --+ oo Proof - We first observe that the singular set 5 P may be written as a union 5 e = 5eo O. 9 9 t._) 5Pko where k o = min { n, k- 1 } and 5Pj consists of those singular the rank of a point x o is the number m appearing in points having rank j, where and the rank is zero if Ord (Xo)> 1. In other words, the Theorem 6.3 if Ord (Xo) = 1, of the linear approximation to u near x o (which exists rank of u at x o is the rank by Theorem 6.3). We will show first that dim 5fo~n-2, and the result for 5 ~ will follow from an easy inductive argument based on Theorem 6.3. Let 5>0 (u) = { x o E f~" Ord (Xo) > 1 }, so that 5g o (u) ~ 5>0 (u). We deal with the larger set 5>o. The proof for 5>o is an application of the basic argument of H. Federer [F2]. For any subset E _ f~ and any real number s e [0, n] define an outer Hausdorff measure ~, ~(.) by ~s (E) = inf { ~ r~. all coverings { B,i (xi) }~= of E by open balls} i=1 The value of J~s (E) is not important, and bears little resemblance to the s-dimensional measure of a set; however, it is clear that the Hausdorff dimension of a set E is given by dim E = inf { s: ~s (E) = 0 }. We note the following result. Lemma 6.5. - If { ui } is" a sequence of minimizing maps into a compact subset of X from B1 (0) c_ R" equipped with metrics ig converging in the C z norm to a limit g, 226 HARMONIC MAPS INTO SINGULAR SPACES then a subsequence of { u i } converges uniformly on compact subsets of B x (0) to a minimizing map u : (B 1 (0), g) ~ X, and we have 3~s (~o (U) O Br (0))/> lim ff~ (5~o (u~) O B, (0)) i --+ oo for all re (0, 1). In particular, dim(5~o (u))>~ lim dim (~o (ui)). i~ oo Proof. - Since ~o(U)N B~(0) is compact we may consider finite coverings, So (u) N B, (0) c_ U B,~ (x j). For any ~ > 0 we have j=l ~0 (Ul) ('~ Be (0) ~ { X e 81 (0) : Dist (x, 5~ o (u) U) B, (0)) < e } for i sufficiently large. This follows immediately from the fact that if x i e fie o (ui) O Br (0), and x i ~ x, then u i ~ u lim Ord (xi)-.~ Ord (x) i~ao and hence by part (i) of Theorem 6.3, xe 5~o(U) ~Br(0). The conclusion of Lemma 6.5 now follows. We now show that dim 5~o (u) ~< n - 2. Suppose s e [0, n] with Jf~ (~o (u)) > 0. Then by [F1, 2.10.19] we may find a point Xoe~) such that lim r Yf~ (5~o (u) ~ B~ (Xo)) >~ 2-~. o~0 Let u,:R"~X,~xo) be a homogeneous approximating map for u at x o. Let 0~=Ord"(xo), so that u, is homogeneous of degree ~, and since xoeS o (u) we have ~> 1 + ~ by Theorem 6.3. We may apply Lemma 6.5 to suitable rescalings { ui } of u near x o, recalling that maps to a neighborhood of u (Xo) may be thought of as maps to the fixed space X, (~o), to conclude that ~ (5~ o (u,))> 0. Since 5~ o (u,) is a cone (u, is homogeneous), it follows that there is a point x 1 e S"- t ~ 5~o (u,) such that lim ~-~ ~ (57 o (u,) t'-') B, (Xl)) ~> 2 -~. t~ ---~ 0 Let ~/1 be a homogeneous approximating map for u, at x x. Note that Ord "1 (xl)~> 1 + ~, and hence Ord"l (txl))1 +~ for t E [0, 1]. Therefore the derivative of ul is zero along the ray t~txl, and hence ul(x0--0eX,(~o r Therefore u 1 is a homogeneous map to the same cone X, Ixo)' The homogeneity of u, together with the fact that OrdU* (x 1) > 1 translates to the statement that the map u~ is independent of a M. GROMOV AND R. SCHOEN direction. If we choose coordinates in which x I = (0 .... ,0, 1), then we have ~?ul/~?x"= O. Therefore the restriction of u~, denoted u~, to the R "-1 spanned by the first (n-1) standard basis vectors is a homogeneous map of degree ~ > 1 + ~. We then have ~o (u0 = 6% (Ul) x R, and thus Jutes-1 (5~o (ul)) > 0. If s> n- 2, we may repeat this argument inductively and produce finally an ~>0 and a minimizing map v'RZ---,,Xu(xo homogeneous of degree ~>1+~ such that ~s-("-2)(5~o(V))>0. In particular, there is a point of 5~ o (v) different from the origin. If we repeat the argument again we construct a geodesic. By the 1-dimensional analysis of Section 1 we know that for a geodesic we have Ord (Xo)= 1 for all x o. This contradiction shows that s ~< n- 2 for any s ~ [0, n] with ~fs (~o (u)) > 0. Therefore dim 5~o (u) ~< n - 2 as required. We now show by induction on k=dim X that we have dim 5 P (//)~<n-2. For k = 1 we have 5 ~ = Yo, and we have established this case. Assume that k ~> 2, and that for F-connected complexes X of dimension less than k we know that the singular sets of minimizing maps of any n-manifold into ~2 have dimension at most n-2. Let Xo be any point of 5~,, - 5Po for a minimizing map u" f~ ~ X k. We then have Ord (Xo) = 1, and by Theorem 6.3 there is a neighborhood B~o (x o) for some cr o > 0 such that the map u maps B,o (Xo) into a totally geodesic subcomplex Xo of X having the form R" x X 1. Thus we have u = (u 1, l/2) where u I 9 B,, o (Xo) ~ R m, U 2 " B,o (Xo) ~ X 1 are both minimizing. The set 5P,,(u)O B, o (Xo) is then the same as the set 5P o (u2)O B, o (Xo) since u~ has rank m at every point of B,o (Xo). By the inductive hypothesis we then have dim (Se,, (u) O B, o (Xo)) ~< n - 2. This shows that for each m, dim 5e,, ~< n - 2 and therefore dim 5 P ~< n-2 as required. To prove the final statement of Theorem 6.4, that is, to construct the functions ~i, let e>0 and d>n-2. Let ~"~2 be a fixed compact subdomain off~ with f~ f~2, and choose a finite covering {Br~(Xj)'j=l,...,l} of the compact set 5eo~fil satisfying ~ rj-,~e.d< Assume also that B 4rj (xj) ~ ~2 as will be true if ~ is small enough. j=l Let q~j be a Lipschitz function which is zero in B~j(xj) and identically one on f~-B2~2(xj) such that ]Vq~jl~<2rf ~. We assume also that x2e~oO~ 1. Let q~ be defined by q>:min {q~j" j: 1,...,l} and observe that q~ vanishes in a neighborhood of ~o~1 and q~-I on - U B2 ,j(xj). Now let 40 = q~2 and observe j=l 228 HARMONIC MAPS INTO SINGULAR SPACES (6.1) IVVul [goldla=2 q) lVVul {V~pld. Q)} = l B2 rj{Xj) ".~ 1/2 "<2(f 5 = 1B2rj(Xj) Ivvul21vuI-I(p2d~t) 5=lB2rj(Xj , by the Schwarz inequality. On the other hand, an elementary result for harmonic maps (see [ES]) implies that on the regular set we have (6.2) For j= 1,..., l let 9j be a Lipschitz function which is identically one on B 2 rj (Xj) and identically zero on f~ - B 3 rj (x j) with I V 9r [ ~< 2 rf ~. Define P by 9=max{pj: j= 1, ...,l} l l and observe that 9 -= l on (..) B 2 rj(Xj) while 9=0 outside L) B3rj(xfl. We therefore j=l j=l have (6.3) fu [ VV u [2 IV u [-~ q~a 02 dla" ~=1 B2rj(Xl) d~ We now state the following result whose proof we give later. Lemma 6.6. - The conclusion of Theorem 6.4 implies that inequality (6.2) holds distributionally on all of ~. Now we make the assumption either that k= 1, or that Theorem 6.4 holds for maps into F-connected complexes of dimension less than k. In the first case we have ~o = 5P, so (6.2) holds away from ~o- In the second case, for any point x o ~ ~- 5%, the map is given locally as u = (Ul, u2) where u I is a map to a Euclidean space and u 2 is a map to a lower dimensional F-connected complex. Since we have IVul2=iVUl [2-~-IVu2l 2, lVVul2=iVV/~li2~-igVu2[2 by our inductive assumption and Lemma 6.6 we have inequality (6.2) distributionally on f*-SPo . A result of [SY] implies that on the regular set we have (1 - ~,) I VV u [2 ~>[V[V u [[2 for some ~, > 0 depending only on n. Therefore we have the distributional inequality alvul> .lVvulelvui-'-civu[ [VVul21Vul-'q~2d"~f M. GROMOV AND R. SCHOEN 229 on f2-50 o. Using p2 {p2 as a test function we find gn fa[VV u[2 lV u[-l p2 ~2 cll, t <~ - 2 fn pq~ ( V IV u l, V (pqo)) dt.t +c I~ ] V u [ P2 q~2 d,. This immediately implies f IVVu[2lVul-Xp2,#2dp (, ~cJa[Vu[(q~2[Vp[2+p2[Vq~[2+02q)2)dg. Combining this with (6. l) and (6.3) we have f lvv. I V*o I d.~ c ~ I v u r (,p= I v 012 + o~ I v ~, I ~ + f)2 (p2) rill an Using the definition of q0 and P we have the result If [VVHI [V*o[d"~c 2 F;2IB [gu[d.. j = 1 3 rj (xj) On the other hand since xj e 5#o we have Ord (x) ~> 1 + e, and therefore by Theorem 2.4 sup IVul<~cr3. B2 rj (x j) Thus we have n 2+~< 8 Ivv.I IV.old.-<,' E rj j=l provided d<n-2+~ and ~ is small. This gives the result for k = 1 since 5 P = 50o in this case. For k> 1, we cover the compact set Brj(x) ~ ~1 with balls { Brp 09 P = 1,..., q } such that in each ball Brp(yv) the map can be written u=(ul, /J2) as in part (ii) of Theorem 6.3. By the inductive assumption there is a function t)p vanishing in a 230 HARMONIC MAPS INTO SINGULAR SPACES neighborhood of 5P O Brp (yfl and identically one outside a slightly larger neighborhood with f IVVu] IV*vld~tK~a2 -v. We finally set ~--min { t~o, ~1,-.., ~q } and conclude [VVu I IV*ldl-t~< IVVu[ IVqtv[dbt~<2a. ;o iofo This completes the proof of Theorem 6.4 except for the proof of Lemma 6.6 which we now give. Proof of Lemma 6.6. - Let P be any nonnegative function with compact support in ~. Suppose ~1 is a domain in ~ with compact closure which contains the support of 9. Let {e,} be a sequence tending to zero and let {~i} be the corresponding sequence of functions coming from Theorem 6.4. We then have -1- V'V u'2' V (q"P) ) fnlVV u'2q"Pd -c f.[V since #i P has support in ~/(u). This implies Ivvul2*'Pd -~ I v u 12" P d" -12 fn( V'V u12, V p ) *~d~t>~ fn -supofnlVuln IVVu] ]V,ld,. Since IV u I is bounded on compact subsets of ~ we conclude that the third term on the right tends to zero. The dominated convergence theorem then allows us to let i ~ ~ to conclude - l fn(V]Vu]2' V p )dP>>" fnlVV u12 pdp-c f ]Vu]2 as required. This completes the proof of Lemma 6.6. M. GROMOV AND R. SCHOEN 231 Part II : Applications to discrete groups In this second part of the paper we prove some rigidity theorems for discrete groups with the help of the theory developed in Part I. In particular, we prove our p-adic superrigidity results, and discuss fundamental groups of K/ihler manifolds. 7. Equivariant harmonic maps to buildings and the Bochner method We begin this section by extending the homotopy existence result of Section 4 to a more general setting for certain image complexes X. Suppose X is the Euclidean building associated to an almost simple p-adic algebraic group H (see [BT] for the construction of X). It will be important to know (see [B]) that X has a compactification X=X L)~X such that any h e H acts continuously on X (of course h is an isometry of X). Moreover, if { P, } is a sequence from X with lira P,= P* e 0X, and if { Qz } is another sequence from X with d(Pi, Qi)~<C independent of/, then it follows that lim Qi =P*. Finally, the isotropy subgroup of H fixing a point P*e0X is a proper algebraic subgroup while the isotropy group of an interior point is a bounded sub- group. For a complete Riemannian manifold M, let 1VI denote the universal covering manifold, and suppose we have a homomorphism 9:F ~ H, where F=II 1 (M). A Lipschitz map v:lVI ~ X is equivariant if v ~ =9 (7)o v for all 7 ~ F. Note that the function IVy 12 is F invariant on 19I and hence is well-defined on M = 1VI/F. Thus we may say that E (v) = total energy of v on M is well-defined even though v is not defined on M. We now prove the following existence result. Theorem 7.1. - Suppose 9 (F) is Zariski dense in H and suppose there exists a Lipschitz equivariant map v" M ~ X with finite energy. Then there is a Lipschitz equiva- riant map u of least energy and the restriction of u to a small ball about any point is mmlmlzmg. Proof. - Let E o denote the infimum of the energy taken over all Lipschitz equivariant maps. Let { z,i } be a sequence of Lipschitz equivariant maps with E (vi) tending to E o. Let B be a ball in 1VI such that y(B)O B=~ for all 7~F. We may then construct a new minimizing sequence ~ by setting vi on hT, l- /y 7 (B) p(y) ouioy-1 on y(B) foranyyeF where u i is a minimizing map in B which agrees with v i on B. We have E (vi)~< E (vi) and ~ is equivariant, so {v-i} is again a minimizing sequence. On a compact subset HARMONIC MAPS INTO SINGULAR SPACES of B, the sequence { ~} is uniformly Lipschitz by Theorem 2.4. It follows that a subsequence of { v i } converges uniformly on compact subsets of B to a map into X which either maps into X or maps to a single point P* e~?X. We use the Zariski density of 9 (F) to exclude the second possibility as follows. Let y e F, and let x o E 1VI be the center of the chosen ball B. Let C be any smooth embedded curve from x o to y x 0. An elementary argument using Fubini's theorem shows that C may be chosen so that the energy of the restriction of each map ui to C is uniformly bounded. Therelbre the length of the curve u~(C) is uniformly bounded, and in particular f d(u i(xO), p (y) (u i(xO))) } is bounded. Therefore lira 9 (Y) (ui(Xo))= P*, and hence P (7)(P*)= P* for every 7 e F. This shows that p (F) is contained in a proper algebraic subgroup of H contradicting the Zariski dense hypothesis. Therefore we may assume that { v-] } converges uniformly on compact subsets of B. From Corollary 4.2 we have for any compact subset K of I~7,I. Since {v~} converges uniformly on compact subsets of B, the function d(~, v) is uniformly bounded there. It then follows from Poincar6- type inequalities that d2 (vi, v)dg~<c fK for any compact K c 19I. In particular, the sequence r t converges weakly in -- "(ci) H ~ (K, X) for any compact subset K to a map uEH 1 (K,X). By the same argument as that given in Theorem 4.4 we conclude that u is a Lipschitz equivariant map of smallest energy. The local minimizing property of u follows. This completes the proof of Theorem 7.1. We now show that the Bochner method can be applied for harmonic maps into F-connected complexes where by harmonic we mean locally minimizing; i. e. minimizing on a neighborhood of any point. Since Euclidean buildings are F-connected this will lead to strong conclusions concerning p-adic representations. We suppose u is an equivariant harmonic map from 19I to X where X is an F-connected complex. We assume u is equivariant with respect to a homomorphism 9 : H1 (M) ~ Isom (X) from F = H~ (M) into the isometry group of the complex X. For example, if u is the lift of the solution of Theorem 4.4 of the homotopy problem, then u is equivariant with respect to the homomorphism g~, induced on fundamental groups of the quotients. Generally equivariant maps need not be lifts from quotients, so this gives a means of studying more complicated representations. M. GROMOV AND R. SCHOEN 233 If we take a regular point x o e lVl, then the image under u of a ball B,o (xo) is contained in at least one k-flat. Choose such a flat F, and let u*TF denote the pullback of the tangent bundle of F under u. Let V denote the pulled back connection, and let d v denote the corresponding exterior derivative operator on p-forms with values in u* TF. Let 6v denote its formal adjoint. The differential du of u then defines a 1-form with values in u*TF, and the harmonic map equation may be written 6 v du = 0. We now prove the following extension of the Corlette vanishing theorem [C]. Theorem 7.2. - Let w be a parallel p-form on ~r and assume that u is a finite energy equivariant harmonic map into an F-connected complex X. hi a neighborhood of any regular point of u the jorm w A du satisfies ~v (w A du) =- O. Proof. - Consider any regular point XoeI~'l, and any k-flat F containing B~o (xo) for some cy o > 0. The calculation of [C] then implies dv 6v (w A du)=-0 near x o. The sets N (u) and 5~ (u) are invariant under F, and we define ~o = ~ (u)/F and 5~o=5P(u)/F. We then have from Theorem 6.4 that dim SPot<n-2, and for any compact subdomain f~l in M, there is a sequence of nonnegative Lipschitz functions {*i } which vanish in a neighborhood of N o V) ~1 and tend to 1 on M- (50o ("/~), such that lim fM[VVu[ [V*,[dg=0. i --* oo To see this we simply observe that we can prove Theorem 6.4 on the quotient by exactly the same argument. Now let p be a nonnegative Lipschitz function which is identically one on BR (Xo) -- M and identically zero outside B 21~ (xo) with I V p I <~ 2 R- 1. Let * be a nonnegative Lipschitz function vanishing in a neighborhood of ~o~B2R(Xo). We then apply Stokes' Theorem on M using the identity 0 =,92( w a du, d v ~v (w A du)). Thus we obtain M*o 2 II (w A d.)II 2 d. = -k [ (, (d(,p 2) A * (W A du), 6 v (w A du) ) dg. .JM This implies fM*o 2 II 5v (w A du)II 2 d. (,0lv01+lv,10z)lgul Ilav(w A du)tld,. 234 HARMONIC MAPS INTO SINGULAR SPACES Using the inequality 2 ab ~< ~ a 2 4- g- 1 b 2 we then have f'l'9 ~ II (" ~ du)[[2dLa 8v -<cfM,l, lvol21Vul2d.+cfMo21V,l,I Ivul IvV.ld. where we have used [I av (w ,. du)II-< c I VV u I" Therefore we have fB *llG(w R (x0) We choose R so large that the first term is less than ~/2, and then we have, since u is Lipschitz on B 2 R, fB llSv(W A du)ll2d.<<.a/2+c(R)fMIVql 11VVuld.. R (xo) We then choose q, so that the second term is less than a/2 and we find that for large R f. ,llSv(w A du)ll 2 dLa~a. R (xo) It follows that 8v(W/x du)-O on ~(u) as required. This completes the proof of Theorem 7.2. We now derive two useful consequences of this result. First we consider the case in which M is a K~ihler manifold and w is the KS, hler two-form. We first make a definition. Definition. - A harmonic map into an F-connected complex is pluriharmonic provided it is pluriharmonic in the usual sense, 88u = 0, on the regular set. Theorem 7.3. - A .finite energy equivariant harmonic map from a Kiihler manifold into an F-connected complex is pluriharmonic. Proof. - By the previous result we have 6v (w/x du)= 0. If we choose Euclidean coordinates (u 1 .... , u k) in a flat F which locally contains the image of u, then as in [C] this condition implies that the Hessian of each function u j(z) for j= 1,...,k is a symmetric matrix lying in the Lie algebra of the symplectic group. This implies that the trace of the restriction of this matrix to any complex line is zero, and hence each u j is a pluriharmonic function. Adu)[12d~t<~cR-2E(u)+cfMPilv~llVul[VVuldla. M. GROMOV AND R. SCHOEN Theorem 7.4. - A finite energy equivariant harmonic map from Quaternionic hyperbolic space or the Cayley plane into an F-connected complex is constant. Proof. - By Corlette's result [C] the condition 6v (w/x du)= 0 where w is either the Quaternionic K/ihler 4-form or the Cayley 8-form implies that the Hessian of u vanishes at each regular point. This implies that the forms du j j= 1,..., k are parallel 1-forms, and hence if they are nonzero we get a local isometric splitting of the domain as the product of R with a lower dimensional manifold. Since such a splitting does not exist for the Quaternionic hyperbolic space or the Cayley plane we conclude that V u-= 0, and hence u is a constant map. This proves Theorem 7.4. 8. p-adic superrigidity for lattices in Sp (n, 1) and F 4 Let 19I be the Quaternionic hyperbolic space (resp. the Cayley plane), so that the group Sp(n, l) (resp. F4) acts on 1r by isometries. We denote the relevant group by G. A lattice F in G is a discrete subgroup with finite volume quotient. Let o:F ~ H be a representation of F in an almost simple p-adic algebraic group. Thus H acts by isometries on the associated Euclidean building X. By replacing F with a finite index subgroup we may assume F is a neat lattice (see [GR]), and we then have: Lemma 8.1. - There exists a finite energy Lipschitz equivariant map. Proof. - Let M = l~I/F so that M is a finite volume rank 1 locally symmetric space. On each cusp l(/I of M there exists a proper function r : 1VI ~ R + such that r is smooth with [Vr I- 1, and r has compact level sets. The metric g on 1VI may then be written dr 2 -~- rg where 'g is a metric on Z o = r- 1 (0). If we normalize the curvature of g to be between - 1 and - 1/4, then we have the inequalities on ~ for a constant c>0 (8.1) c le-2r(~176 Consider first the compact manifold M o with boundary gotten by removing each of the finite number of cusps from M. Then 19I o is a F-invariant subregion of 1r and we can choose a Lipschitz equivariant map v from 19I 0 to X. To extend v to all of 19I, we extend v to the cusps as a function independent of r. This produces a Lipschitz equivariant map, and we show that it has finite energy. The function IVvl 2 is F-invariant and thus descends to M. From (8.1) we see that on r-l(a) we have IV v 12 ~ ce 2 a while the volume of r- 1 (a) is bounded above by a constant times e-((n-i)/2) a. Since n=dim M~>8 in the cases we are considering we have 236 HARMONIC MAPS INTO SINGULAR SPACES OO. This completes the proof of Lemma 8.1. We are now in a position to prove the following main theorem. Theorem 8.2. - If 9 (F) is Zariski dense in H, then 9 (F) is contained in a bounded subgroup of H. Proof - By Lemma 8.1 there is a finite energy equivariant map from 1VI to X. By Theorem 7.1 there exists a least energy minimizing map u" M ~X. By Theorem 7.4 the map u is constant, say u (1VI)= { P0 }- It follows that 9 (F) _~ HPo, the stabilizer of the point Po~X. This is a bounded subgroup (see [B]), so we have completed the proof of Theorem 8.2. As a corollary we obtain the following result. Theorem 8.3 (p-adic superrigidity). - Let F be as above and 9 be a p-adic representation of F i.e., a homomorphism 9" F ~ GLN Qp for some N = 1,2 .... , and a prime p. Then the image p (F) is precompact in GL N Qp. Proof. - It suffices to prove that the image of a subgroup of finite index of F is precompact, hence we may assume as above that F is neat. Let H denote the Zariski closure of P (F). If H is almost simple, then the claim follows from 8.2. Furthermore, if H is semi-simple the conclusion is obtained by applying 8.2 to the simple factors of H. Finally, in the general case, let R denote the radical of H and H o be the semi- simple factor-group, H o = H/R. The factor representation 9o : F ~ H o has a precompact image 9o (F)c H o as follows from the above discussion and let R o c H denote the pull-back of the closure K o c H o of 9o (F) under the projection H ~ H o. Notice that R 0 is a locally compact group containing a solvable normal subgroup, namely R c R o, and the quotient group K o is compact. Thus R is an amenable group and so p : F ~ R o has bounded image because F satisfies Kazhdan's property T. Remarks. - (a) Our super-rigidity complements the Archimedian super-rigidio, theorem of Corlette: Every representation 9 : F ~ GL N R has either precompact image or it extends to a continuous representation of the ambient Lie group (which contains F as a lattice). (b) One could prove 8.3 directly without an appeal to the (elementary) structure theory of algebraic groups. First, using the T-property, one observes that p(F) is contained in SL N Qp and then there is an action of F on the building X attached to SL N. Then one applies our minimization process to the corresponding map X ~ M M. GROMOV AND R. SCHOEN and if the minimizing sequence vi:X--, M tends to infinity one brings it back using some isometries ai of Mi, thus achieving a uniform convergence of ~ o v-] on compact subsets in X. In fact, by our Bochner-Corlette formula, the limit of these maps is constant. Then one notices that applying ~z amounts to conjugating 9 by ~ and the above constancy of the limit map says that the representations ~ 9~ 71 converge to a representation with a bounded image. Hence 9 itself has a bounded image as easily follows from the T-property of F. Now we are in a position to conclude the following. Theorem 8.4 (Arithmeticity). - A lattice F in G = Sp (n, 1), n>~2 (i.e., the isometry group the quaternionic hyperbolic space) or n the isometrv group G of the Cavley plane is arithmetic. In fact, Margulis has reduced the arithmeticity to the archimedian plus p-adic super-rigidity. (This reduction is clearly explained in Chapter 6 of [Zim].) 9. Pluriharmonic maps into trees and buildings Let M be a K/ihler manifold and u a harmonic map of M into a locally F-connected space X. We know that in certain cases (e.g., if M is a closed manifold) u is pluriharmonic (see section 7) and now we want to understand the local and global geometry of these pluriharmonic maps. Our main application of such maps, proven later on in this section (in the case where X is a tree), is the following Theorem 9.1. - Let M be a compact Kiihler manifold without boundary and suppose the fundamental group F=IIx(M) admits an amalgamated product decomposition I~ = F1 ~r F2, where the index of A in F 1 is at least 2 and the index of A in F 2 is at least 3. Then M admits a surjective holomorphic map onto a Riemann surJace. Let us look at a map u of M into a tree X at a regular point x o ~ M, such that (by the definition of regularity) the u-image of a small ball B~o (Xo) c M is contained in a geodesic (flat) F c X. Locally this F is isometric to R and so u near x o amounts to a real function u:B~o (Xo)~ R = F = X. Notice that this F = R carries no natural orientation and so the differential du is only defined up to + sign. On the other hand, if u is pluriharmonic, then the complexified differential, say d c u, is holomorphic, if we choose the sign. Therefore, the square (dc u) 2 is a holomorphic quadratic differential (of rank one) defined on the set of regular points of u. Now, since u is Lipschitz, the differential du is bounded, and since the singular set has codimension /> 2 in M, the differential (dc lt) 2 extends to a holomorphic quadratic differential on all of M. Since this differential has rank ~< 1, there is a ramified double covering of M, say M ~ ~ M (where M~ may be singular), such that (d c u) z lifts to a holomorphic 1-form, denoted d~ u on M ~. This form is closed and locally the differential of a holomorphic function 238 HARMONIC MAPS INTO SINGULAR SPACES on M~, say u~ : M~ ~ C. The levels of this function can be defined as maximal complex submanifolds (having complex codimension one in M) contained in the levels u- 1 (y) ~ M, y e X, of u lifted to M ~ (these have real coodimension one). The (Galois) involution on M~ switches the sign of d~ u and, hence, locally of u~. It follows that the local partition into connected components of the levels of (u~) 2 descends back to M and by the (local version of) Stein factorization theorem the map u decomposes in a small neighborhood ~ of each (possibly singular) point of M into a composition of a holomorphic function f~ ~ D c C followed by a harmonic map u o : D ~ X where D denotes the unit disk in C. Therefore, the singular locus of u appears as a holomorphic pull-back of that for the harmonic map uo:D---, X and the quadratic differential (D c u) 2 comes from (dc Uo) z on D. Since the singular locus of u o is contained in the (discrete) set of zeros of (d cu0) z, it is discrete and hence, the singularity of u is a complex analytic subvariety in M locally given by a single function M ~ D whose level lifted to M~ are connected components of levels of u~. Summing up, we arrive at the following Corollary 9.2. - Let u: M ~ X be a non-constant pluriharmonic map of a connected complex analytic maniJold into a tree. The set ~ ~ M of singular points of u is a complex analytic subvariety in M whose components have codimension 1 in M. Furthermore, the lift of Z to M ~ equals the union of some leaves of the (holomorphic) foliation defined by the (closed holomorphic) 1-form (d~ u) 2 on M ~. Remark. - The pull-back u- 1 (v) ~ M of a vertex v ~ X is not, in general, a real analytic subset in M. For example, the standard harmonic map of the disk into the tripod (see w 0) has the pull-back u- 1 (v) homeomorphic to the tripod. This cannot be real analytic as it has an odd number (3 in this case) of branches. Yet this level is subanalytic. Moreover, all levels of an arbitrary pluriharmonic map u lifted to M~ are real analytic subsets in M ~ as follows from 9.2. Every such subset at a singular point of u is necessarily reducible and the complement has more than two connected components sent by u to different branches of the tree. Now we look at a slightly more general situation where we have a F-equivariant pluriharmonic map u of a connected K~ihler manifold 1VI into a tree X. Here we assume that the group F is infinite and acts discretely, cocompactly, and isometrically on 1~,I, and that the action of F on the tree is also isometric for the given (piecewise linear) metric on the tree X. We want to show that u decomposes into a holomorphic F-equivariant map of 19[ to a Riemann surface D with a discrete holomorphic action of F (typically, D is the unit disk in C) followed by a F-equivariant harmonic map D ~ X. We divide our analysis of u into two cases depending on whether the map u is everywhere regular or has a non-empty singular set. M. GROMOV AND R. SCHOEN 239 Decomposition Lemma in the singular case. - If the singular locus Z ~ ~I of u is non-empty then the map u decomposes into a proper holomorphic F-equivariant map IVl ~ D followed by a F-equivariant harmonic map D ~ X. Proof - Let us look at the holomorphic foliation defined by the form (dc u) 2. The singular locus ~ of u is a union of leaves and, obviously, "~ c lVl is F-closed i.e., the image of ,~ in M--= 1VI/F is a closed subset. If follows that if ~, is non-empty then all leaves of our foliation are F-closed (see explanations below) and so our foliation descends to a partition of M=IVI/F into compact analytic subsets. We apply to this partition the Stein factorization theorem (see [Ste]) and thus obtain the desired factorization 1VI --, D. First explanation. - Let l~ be a holomorphic (and hence closed) 1-form on a compact K/ihler manifold M ~. Then there exists a complex torus N of real dimension k in the interval 2~<k~<ranH 1 (M~), such that l~ can be induced from some holomor- phic 1-form l on N by a holomorphic map ~ : M ~ ~ N and where each leaf L ~ N of the foliation defined by l satisfies the following Minimality condition. - The leaf L contains no complex subtorus of positive dimension. To construct N one starts with the Albanese variety N o of M~ with the form l o corresponding to l~ and then factorizes N o by the maximal complex subtori in the leaves of the corresponding foliation of No. Notice that the minimality condition implies that no leaf L c N contains a complex compact submanifold of positive dimension. (Indeed, each L is of the form C k- 1 5o for some non-cocompact lattice 5o ~ C k- 1. We turn L into an Abelian analytic Lie group by taking some point in L for the origin and observe that the additive span of a compact complex submanifold in L is a complex torus). It follows that if the pull-back ~-1 (L) contains a compact component then ~ (M ~) O L is zero-dimensional and therefore the image a (M~) ~ N is 1-dimensional. The above provides an explanation needed for the global factorization theorem in the case where the action of F on l~I is free: Here one takes the ramified double covering of I~I/F for M ~ and observes that the lift of Z/F ~ IVl/F to M consists of a union of finitely many compact components of the pull-back of some leaves L c N. In the general case, where F is non-free (but yet discrete), one notes that all steps of the above argument apply in the presence of fixed points if we work F-equivariantly on 1~1 rather than on the quotient space I~I/F. A judicious reader might have noticed that the above explanation, which relies on the theory of Albanese varieties, needs special care in the case where M ~ is singular. Instead of resolving this singularity we suggest below an elementary (albeit HARMONIC MAPS INTO SINGULAR SPACES less elegant) argument where the singularity of M- does not enter the discussion at all. Second explanation. -Let L (m), m ~ M denote the leaf of our foliation (defined by -t-dCu) passing through m, and denote lVI*c l~I the union of the leaves ~(rn) whose projections L (r~) to M = 1VI/F are compact. First we observe that the subset 1(,I* c 1VI is closed in the following sense. If m(t)elVl, te[0, 1], is a continuous path of points, such that m(t)el~l* for t>0, then also ~n (0) E ~vl *. It is easy to see (e.g. by looking at the function v below) that the homology class of L(rn(t)) stays bounded for te]0, 1]. Therefore the K/ihlerian volume of L (r~ (t)) ~ M = 1VI/F, being a homological invariant, is bounded for t > 0 and so the limit leaf L (m (0)) is compact. (Notice that Vol. L (in (0)) = Vol L (in (t)) for small t > 0 unless L (rn (0)) is multiply covered by nearby leaves in M = 1VI/F.) Secondly, we show that the subset ~r ~ M is open. Proof. - For each leaf L c IVI there obviously exists a function ~7 on a small neighbourhood A = A (L) c 1VI of L, such that (i) (dr) 2= v(du)2; this is equivalent to dx/~= + dCu and this makes the foliation into the connected components of the levels of v equal to that defined by + dCu on (ii) vlL=0. Now, let the projection L ~ M of L be compact. Then v descends to a function v on some neighbourhood A c M of L, such that the pullback of v to lVl equals v on some neighbourhood of [, (where both functions are defined). Since L c M is compact, the leaves L (m), which equal the connected components of levels of v, stay close to L for m close to L (i. e. L (m) Hausdorff converge to L for m ~ L, compare the proof of 9.3 below). Therefore, these L (m) are compact (as they do not reach the boundary of A) and they receive the leaves [, (m) from 1VI for rn close to [,. This yields the openess of if4*. Finally we notice that every leaf contained in the singular locus E c l(,I has compact projection as '7;/F is compact and so our 1VI* is non-empty. Thus 1VI* = M, since M is connected. Remark. - The above argument applies whenever the zero set of d e u contains a leaf L. Notice that the inclusion L ~ Zero (dc u) is equivalent to local disconnectedness of the levels of the above function x/v near each point r~ e L. Nonsingular ease. - If u:lVI ~ X is a nonsingular pluriharmonic map then the complexified differential dCu is defined everywhere on IVl up to + sign. It follows dCu becomes a honest holomorphic 1-form )~ on a (non-ramified) double cover of IVl which M. GROMOV AND R. SCHOEN 24l is invariant under a discrete group consisting of lifts of the transformations from F to this cover. To minimize notation we assume below that this double cover splits, and thus we have a F-invariant holomorphic form )~ on 1VI itself, such that the map u : 19I --, X is constant on the leaves of the foliations defined by )~ on 1VI. Furthermore, by passing to some Abelian covering of 1VI we can make )~ exact as well as equivariant under a certain discrete cocompact group acting on this covering. Thus we may assume is exact to begin with. In this case the leaves, which are the connected components of the level sets of the holomorphic function z on 19I with dz = ;~, are closed analytic subvarieties in 19I. Moreover, we claim that the space of these leaves is Hausdorff (this is explained below) and therefore the Stein factorization theorem applies. Thus we arrive at the following Factorization in the nonsingular case. - Suppose u is non-singular and l~I is' simply connected (This takes care of the intermediate coverings). Then u decomposes into a holomorphic map M ~ D followed by a harmonic map D ~ X, where these maps are equivariant with respect to F or a subgroup F' of index two in F (taking care of the + sign of d c u). Here, as earlier, D comes equipped with a holomorphic action of F (or F') but we do not claim at this stage that the action of F (or F') on D is discrete. In general it need not be discrete as the map 1~I --, D may be non-proper. On the Hausdorff property of holomorphic foliations. - Let M and N be complex manifolds, z:M ~ N a holomorphic map and let N be the space of the leaves that are the connected components of the fibers z ~ (11) c M, n eN. The topology of N is defined in the usual way: a subset in N is open if the union of the corresponding leaves is an open subset in M. There is a simple criterion for NI to be Hausdorff where z is a general (not necessarily holomorphic) smooth map. Namely, N is Hausdorff if z is a submersion (i. e. rank z = dim N) which is, moreover, infinitesimally enlarging with respect to some Riemannian metric gN on N and some complete metric gM on M, where "infinitesimally enlarging" means that the pullback of the form gN to the subbundle in T (M) normal to the kernel of dz dominates gN, i.e. z* (gN) I Ker dz >~ gM I Ker dz (compare [Gro]). This applies in particular to the case where z is a submersion and the pull-back form z* (gN) on M is invariant under some proper cocompact group F acting on M. Now, we return to the holomorphic map z : M ~ N, dim c N = 1, and assume that there exists a Riemannian metric gN on N, such that the pull-back z* (gN) is invariant under some proper cocompact group action on M. (Notice that this condition is satisfied in the case of interest where we have an exact F-invariant 1-form on M.) 242 HARMONIC MAPS INTO SINGULAR SPACES Lemma 9.3. - The space N is Hausdorff and hence admits a unique complex structure such that z decomposes into a composition of two holomorphic maps z : M --, 1KI-~ N Proof - Since z is holomorphic and dim N = 1, the leaf L (m) passing through m EM continuously depends on m with respect to the Hausdorff distance between compact subsets in the leaves. (Such local continuity is automatic for usual foliations, where the leaves are non-singular, while the foliation into the connected components of the levels of the map R 2-* R, (x, y)~-~xy, indicates what might go wrong in the presence of a singularity). Now we chose a metric on M invariant under our group which preserves z* (gN), and observe (looking at z* (gN) and/or using Stein factorisation on compact subsets in M) that the above continuity on compact subsets is uniform for such subsets. Thus the correspondence mv-+ L (m) is continuous for the Hausdorff metric on the set of leaves, and then the lemma easily follows. On the discreteness of the action of F on D. - We return to the factorization in the non-singular case, u:lgl ~D~X, and give a criterion for the implied action of F (or F') on D to be discrete. Lemma 9.4. - The action ofF on D is discrete unless it factors through a virtually solvable group (i.e. the implied homomorphism F ~ Aut D has a virtually solvable image). Proof. - We may assume here that D is the unit disk and the image r" of F in Aut D=PSL2 (R) is Zariski dense (otherwise it would be virtually solvable). Then either r" is discrete in Aut D or it is topologically dense. In the latter case the commutator subgroup [F, F] is also dense in Aut D and then every exact F-invariant holomorphic form on D must be zero. But the form )~ we started with obviously descends to D and thus the dense case is excluded. Remarks. - (a) Let us indicate an alternative approach similar to that in [Sim]. First we notice that if the map N --, N has a ramification point, then by our "singular argument" IVI admits a proper F-equivariant map to a Riemann surface (with no use of Lemma 9.3). Then if no ramification point is present, our argument similar to that proving Lemma 9.3, shows that the map N--. N is (non-ramified) and so for N = C we have N = C as well. (b) Our original treatment of the "non-singular" case contained an error pointed out to us by K. Corlette. M. GROMOV AND R. SCHOEN 243 (c) K. Corlette and C. Simpson suggested another approach to the factorization of the map u: 1VI ~ X following the earlier work by Simpson (see [Sire]). Harmonic maps into Z-trees. - We want to generalize the above discussion to the case where X is a Z-tree, i.e., where X may have infinitely many branches at the vertices. This is needed for the amalgamated product property stated at the beginning of section 9. First we notice that every singular point of u has a finite multiplicity, controlled by the total energy and so every harmonic map locally factors through a map into a finite tree X' followed by an isometric embedding X'--+ X. Hence u is pluriharmonic (even for R-trees) and so the above holomorphic discussion remains valid. What is left to do is to extend the existence theorem to the case where X is not locally compact. We shall state and prove the relevant property where X is a generalized tree i.e., Z-tree or an R-tree. Here one can not ensure the existence of a F-equivariant harmonic map 1VI ~ X in a given homotopy class but one can obtain a harmonic map if one modifies the receiving space X by going to an appropriate limit. Namely, let u~:ff, I ~ X be a minimizing sequence. (In fact, we could allow maps with variable target X~ but this is not needed right now.) All these maps may be assumed uniformly Lipschitz and of uniform multiplicity. Thus they factor on compact subsets in 1VI through maps lVl ~ X~--* X where X~ are finite trees. Then we pass to the Hausdorff (sub)-limit of the spaces Xi as 1~I is being exhausted by compact subsets and i--, oo and obtain the desired harmonic map u : 1~I ---, X~. Warning. - The trees X i can be thought of as subsets of X but the limit Xoo is not a part of X. For example, one may imagine Xi equal to a countable joint of the segments Ii= [0, 1] at zero where Xi=I~. The sequence I~ c X diverges inside X but (identically) converges in the abstract sense to the unit interval. A more invariant way to look at X~ is by concentrating on the function d~ on 1VI x 1VI induced by u~ from the distance function on X. All properties of the maps u~ relevant for us can be expressed in terms of d i without ever referring to X and instead of the limit space X~ one can deal with a (sub)-limit do~ of d~. This works perfectly well whenever the maps u~ are uniformly Lipschitz. Such an approach is especially attractive if X is an infinite dimensional symmetric space, e.g., the Hilbert space on the infinite-dimensional hyperbolic space, where one has a simple criterion for the existence of an isometric embedding of a given metric space (1VI, d) into X. Let us apply the above considerations to the amalgamated product problem. If the fundamental group F = II 1 (M) decomposes as F = F1 *~ F 2 we have a F-equivariant map of 19I to some tree X with a F-action where the degrees of the vertices are given by ind(A~Fi), i=1,2 (see Serre [Ser]). Our assumption ind(AcF1)~>2 and ind (A c F2)~>3 ensures that our F is not virtually solvable and neither is the 244 HARMONIC MAPS INTO SINGULAR SPACES corresponding group F c AutD. Hence, we obtain a (non-constant!) F-equivariant map l~l ---, Xoo which factors through a Riemann surface. Let us briefly explain, without griving proofs, how the above results generalize to pluriharmonic maps u:M--* X where X is a k-dimensional Euclidean building. Every such u gives rise to a holomorphic foliation on M, whose leaves are maximal connected complex submanifolds in the pull-backs of the points, u- 1 (x) c M, x e X. As earlier, these leaves may have singularities and the complex codimension of a generic leaf equals the real rank of u at a general regular point in M. Next we invoke the (finite) Weyl group W associated with the building and we claim there exists a ramified Galois covering 1VI --. M with Galois group W, such that the lift of our foliation to M equals the zero set of a finite system of holomorphic 1-forms on lVl. Then, for F-equivariant maps one attempts to show that the leaves project onto compact submanifolds in M/F. For example, this can be proven if dim c M >~ k + 1 and M/F is compact. Finally, we notice that there are many instances of non-trivial p-adic representa- tions of fundamental groups of algebraic manifolds M. These appear via the p-adic uniformization theory for arithmetic varieties M of the form M = U (k)~U (k, 1)/F, as was pointed out to us by D. Kazhdan. We plan to study the pluriharmonic maps which appear here in another paper. M. GROMOV AND R. SCHOEN 245 REFERENCES [ABB] S. B. ALEXANDER, I. D. BERG and R. L. BISHOP, The Riemannian obstacle problem, Ill. J. Math. 31 (1987), 167-184. S. AGMON, Unicitb et convexitb dans les problkmes diff~rentiels, SOre. d'Analyse Sup., Univ. de Montr6al, [Agl F. J. ALMGREN, Jr., Q-valued functions minimizing Dirichlet's integral and the regularity of area minimizing [All rectifiable currents up to codimension two, preprint, Princeton University, K. BROWN, Buildings, Springer, New York, I989. [B] [BT] F. BRUHAT and J. TITS, Groupes rOductifs sur un corps local. I. Donn6es radicielles valu6es, Publ. Math. IHES 41 (1972), 5-251. K. CORLETTE, Archimedian superrigidity and hyperbolic geometry, Annals of Math., to appear. [C] Y. J. CHIANG, Harmonic maps of V-manifolds, Ann. Global Anal. Geom. 8 (1990), 315-344. [Oh] [DM] P. DELIGNE and G. D. MOSTOW, Monodromy of hypergeometric functions and non-lattice integral mono- dromy, Publ. Math. IHES 63 (1986), 5-89. J. EELLS and J. H. SAMPSON, Harmonic mappings of Riemannian manifolds, Amer. J. Math. 86 (1964), 109- [ESl [Fl] H. FEDERER, Geometric Measure Theory, Springer-Verlag, New York, 1969. l-L FEDERER, The singular sets of area minimizing rectifiable currents with codimension one and of area [F21 minimizing fiat chains modulo two with arbitrary codimension, Bull. Amer. Math. Soe. 79 (1970), 761-771. H. GARLAND, P-adic curvature and the cohomology of discrete subgroups, Ann. ~!f Math. 97 (1973), [G] 375-423. N. GAROFALO and F. H. LIN, Monotonicity properties of variational integrals, Ap weights and unique [GL1 continuation, Indiana Math. J. 35 (1986), 245-268. M. GROMOV, P, PANSU, Rigidity of lattices: An introduction, to appear in Springer Lecture Notes. [GP] [GPS] M. GROMOV and I. PIATETSKI-SHAPIRO, Non-arithmetic groups in Lobachevsky spaces, Publ. Math. IHES 66 (1988), 93-103. H. GARLAND and M. S. RAGHUNATHAN, Fundamental domains for lattices in R-rank l groups, Ann. of [GR] Math. 92 (1970), 279-326. M. GROMOV, Partial differential relations, Springer Verlag, 1986. [Gro] [Ham] R. HAMILTON, Harmonic maps of manifolds with boundary Lecture Notes 471, Springer 1975. P. HARTMAN, On homotopic harmonic maps, Can. J. Math. 19 (1967), 673-687. [Har] P. de la HARPE and A. VALETTE, La propriktO (7) de Kazhdan pour les groupes localement compacts, Astbrisque [HV] 175 (1989), Soc. Math. de France. H. KARCHER, Riemannian center of mass and mollifier smoothing, Comm. Pure and Appl. Math. 30 (1977), [K] 509-541. E. M. LANDIS, A three-sphere theorem, Dokl. Akad. Nauk S.S.S.R. 148 (1963), 277-279. Translated in Soviet [Lal] Math. 4 (1963), 76-78. E. M. LANDIS, Some problems on the qualitative theory of second order elliptic equations (case of several [La2] variables), Uspekhi Mat. Nauk. 18 (1963), 3-62. Translated in Russian Math. Surveys 18 (1963), 1-62. M. L. LEITE, Harmonic mappings of surfaces with respect to degenerate metrics, Amer. J. Math. 110 (1988), ILl 399-412. F. H. LIN, Nonlinear theory of defects in nematic liquid crystals, phase transition and flow phenomena, [Lin] Comnl. Pure Appl. Math. 42 (1989), 789-814. V. MAKAROV, On a certain class of discrete Lobachevsky space groups with infinite fundamental domain of [Mak] finite measure, Soviet Math. Dokl. 7 (1966), 328-331. G. MARGULIS, Discrete groups of motions of manifolds of nonpositive curvature, AMS Translations 109 [Mar] (1977), 33-45. G. D. MOSTOW, On a remarkable class of polyhedra in complex hyperbolic space, Pae. J. Math. 86 (1980), [Mos] 171-276. 246 HARMONIC MAPS INTO SINGULAR SPACES [Mi] K. MILLER, Three circles theorems in partial differential equations and applications to improperly posed problems, Arch. for Rat. Mech. and Anal. 16 (1964), 126-154. [Mo] C. B. MORREY, Multiple integrals #1 the calculus of variations, Springer-Verlag, New York, 1966. I. G. NIKOLAEV, Solution of Plateau problem in spaces with curvature ~<K, Sib, Math. J. 20:2 (1979), 346- [NI [S] R. SCHOEN, Analytic aspects of the harmonic map problem, Math. Sci. Res. Inst. Publ. vol. 2, Springer, Berlin, 1984, 321-358. [Se] A. SEL~ERG, Recent developments in the theory of discontinuous groups of motions of symmetric spaces, Springer Lecture Notes 118 (1970), 99-120. [Ser] J. P. SERRE, Trees, Springer Verlag, 1980. [Sim] C. SIMPSON, lntegrality of rigid local systems of rank two on a smooth projective variety, preprint. [Siu] Y. T. SIU, The complex analyticity of harmonic maps and the strong rigidity of compact K/ihler manifolds, Ann. of Math. 112 (1980), 73-112. [Ste] K. STEIN, Analytische Zerlegungen komplexer R/iume, Math. Ann. 132 (1956), 63-93. R. SCHOEN and K. UHLENBECK, A regularity theory for harmonic maps, J. Dif f. Geom. 17 (1982), 307-335. [su] [SY] R. SCHOEN and S. T. YAU, Harmonic maps and the topology of stable hyper-surfaces and manifolds of nonnegative Ricci curvature, Comment. Math. Heh,. 39 (1976), 333-341. pC] E. VINBERG, Discrete groups generated by reflections in Lobachevsky spaces, Math. USSR-Sb 1 (1967), 429- [Z] W. P. ZIEMER, Weakly Differentiable Functions, Springer-Verlag, Grad. Texts in Math., 1989. [Zim] R. J. ZIMMER, Ergodic Theo O, and Semi-simple Groups, Birkh/iuser, Boston, Basel, Stuttgart, 1984. M. G. IHES 35, route de Chartres 91440 Bures-sur-Yvette R.S. Mathematics Department Stanford University Stanford, CA 94305 Manuscrit recu le 29 juillet 1991.
Journal
Publications mathématiques de l IHÉS – Springer Journals
Published: Aug 31, 2007