Spaces and questions

Misha GROMOV
November 1999
I. Dawn of Space.
Our Euclidean intuition, probably, inherited from ancient primates, might have
grown out of the rst seeds of space in the motor control systems of early animals
who were brought up to sea and then to land by the Cambrian explosion half
a billion years ago. Primates brain had been lingering for 30-40 million years.
Suddenly, in a ash of one million years, it exploded into growth under relentless
pressure of the sexual-social competition and sprouted a massive neocortex (70
% neurons in humans) with an inexplicable capability for language, sequential
reasoning and generation of mathematical ideas. Then Man came and laid down
the space on papyrus in a string of axioms, lemmas and theorems around 300
B.C. in Alexandria.
Projected to words, brains space began to evolve by dropping, modifying
and generalizing its axioms. First fell the Parallel Postulate: Gauss, Schweikart,
Lobachevski

, Bolyai (who else?) came to the conclusion that there is a unique

non-trivial one-parameter deformation of the metric on R
3
keeping the space
fully homogeneous.

It is believed, Gauss, who convinced himself in the validity of hyperbolic

geometry somewhere between 1808 and 1818, was disconcerted by the absence
of a Euclidean realization of the hyperbolic plane H
2
. By that time, he must
have had a clear picture of geometry of surfaces in R
3
(exposed in his Disquisi-
tones circa supercies curvas in 1827) where the (intrinsic) distance between
two points on a surface is dened as the length of the shortest (better to say
inmal) curve in the surface between these points. (This idea must have been
imprinted by Nature in the brain as most animals routinely choose shortest cuts
on rugged terrains.) Gauss discovered the following powerful ecient criterion
for isometry between surfaces, distinguishing, for example, a piece of a round
sphere S
2
R
3
from an arbitrarily bent sheet of paper (retaining its intrinsic
Euclideaness under bendings).
Map a surface S R
3
to the unit sphere S
2
by taking the vectors (s) S
2
parallel to the unit normal vectors (s), s S.
If S is C
2
-smooth, the Gauss map G : S S
2
for s (s) is C
1
and its
Jacobian, i.e. the innitesimal area distortion, comes with a non-ambiguous
sign (since the directions of s give coherent orientations to S and S
2
) and so
S appears with a real function, called Gauss curvature K(s) =
def
Jac G(s).
Theorema Egregium. Every isometry between surfaces, say f : S S
1
,
preserves Gauss curvature, K(f(s
1
)) = K(s
1
) for all s
1
S
1
.
For example, the plane has K 0 (as the Gauss map is constant) and so it is
not (even locally) isometric to the unit sphere where K 1 (for the Gauss map

Accidentally, the rst mathematics teacher of Gauss ( 1790), Johann Martin Bartels,
later on became the teacher of Lobachevski ( 1810) in Kazan.

A metric space X is fully homogeneous if every partial isometry X

X
extends to a full isometry of X (as for Euclids triangles with equal sides in R
2
).
1
!(s)
!(s')
S
2
S
!(s')
!(s)
s'
s
Figure 1
is identity on S
2
). More generally, no strictly convex surface is locally isometric
to a saddle surface, such as the graph of the function z = xy for instance, since
strict convexity makes K > 0 while saddle points have K 0.
Gauss was well aware of the fact that the hyperbolic plane H
2
would have
constant negative curvature if it were realized by a surface in R
3
. But he could
not nd such a surface! In fact, there are (relatively) small pieces of surfaces
with K = 1 in R
3
investigated by Beltrami in 1868 and it is hard to believe
Gauss missed them; but he denitely could not realize the whole H
2
by a C
2
-
surface in R
3
(as is precluded by a theorem of Hilbert (1901). This could be
why (besides his timidity in the face of Kantians guard of Trilobites intuition)

Gauss refrained from publishing his discovery.

Probably, Gauss would have been delighted to learn (maybe he knew it?)
that the at Lorenz-Minkowski metric dx
2
+dy
2
dz
2
on R
2,1
= R
3
induces
a true positive metric on the sphere
S
2

=
_
x, y, z [ x
2
+y
2
z
2
= 1
_
,
where each of the two components of S
2

(one is where z > 0 and the other

with z < 0) is isometric to H
2
and where the orthogonal group O(2, 1), (i.e. the
linear group preserving the quadratic form x
2
+y
2
z
2
) acts on these two H
2
s
by (hyperbolic) isometrics.
There is no comparable embedding of H
2
into any R
N
(though H
2
admits
a rather contorted isometric C

-immersion to R
5
(to R
4
?) and, incredibly, an
isometric C
1
-embedding into R
3
) but it admits an embedding into the Hilbert
space, say f : H
2
R

, where the induced intrinsic metric is the hyperbolic

one, where all isometries of H
2
uniquely extend to those of R

and such that

dist
R
(f(x), f(y)) =
_
dist
H
2 x, y) +(dist(x, y))

Of all people, had he been alive, Kant himself could have been able to assimilate, if not
accept, the non-Euclidean idea.
2
with bounded function (d) (where one can nd f with (d) = 0, but this will be
not isometric in our sense as it blows the lengths of all curves in H
2
to innity).
Similar embeddings exist for metric trees as well as for real and complex hy-
perbolic spaces of all dimensions but not for other irreducible symmetric spaces
of non-compact type. (This is easy for trees: arrange a given tree in R

, such
that its edges become all mutually orthogonal and have prescribed lengths.)
Summary. Surfaces in R
3
provide us with a large easily accessible pool of
metric spaces: take a domain in R
2
, smoothly map it into R
3
and, voil`a, you have
the induced Riemannian metric in your lap. Then study the isometry problem
for surfaces by looking at metric invariants (curvature in the above discussion),
relate them to standard spaces (R
N
, R

, R
2,1
), and consider interesting (to
whom?) classes of surfaces, e.g. those with K > 0 and with K < 0.
Remark and references. (a) It seems to me that the reverence for human
intuition and introspective soul searching stand in the way to any attempt to
understand how the brain does mathematics. Hopefully, experience of natural
scientists may lead us to a meaningful model (a provisional one at this stage,
say in the spirit of Kanervas idea of distributed memory, see [Kan]).
(b) Our allusions to the history of mathematics are borrowed from [Klein],
[Newm] and [Vasi].
(c) Little is known of what kind of maps S S
2
can serve as Gauss maps
G of complete surfaces in R
3
. For example, given a domain U S
2
, one
may ask whether there exists an oriented closed immersed (i.e. with possible
selntersections) surface S R
3
with G(S) U (where U ,= S
2
forces S to
be topologically the 2-torus). This now appears to me a typical misguided
natural question; yet I have not lost hope it may have a revealing solution
(compare 2.4.4. in [Gro
PDR
]).
(d) It is unknown if every surface with a C

-smooth Riemannian metric

can be isometrically C

-immersed into R
4
. (Another natural question?) But
isometric immersions into high dimensional spaces are pretty well understood
(see 2.4.9 2.4.11 and Part 3 in [Gro
PDR
]).
(e) The above equivariant embedding H
2
R

tells us that the isome-

try group IsoH
2
= PSL
2
R is a-T-menable, opposite to Kazhdans property T
(dened in (A) of V) satised by the majority of groups. (A-T-menability gen-
eralizes amenability. This property was rst recognized by Haagerup, I presume,
who used dierent terminology.)
II. Spirit of Riemann.
The triangle inequality is not always easy to verify for a given function in two
variables d : X X R
+
as it is a non-local property of d on X; thus one
cannot create metrics at will. Yet, an arbitrary metric d on a connected space
3
can be made local by replacing it by the supremal metric d
+
agreeing with
d on an arbitrary ne covering of X (as we pass, for example, from the
restriction of the Euclidean metric on a submanifold, e.g. a surface, in R
N
,
to the induced Riemannian, or intrinsic, metric). More generally, following
Riemann (Habilitationsschrift, June 10, 1854) one starts with an arbitrary eld
g of Euclidean metrics on a domain U R
n
, i.e. a continuous map u g
u
from
U to the
n(n+1)
2
-dimensional space G of positive denite quadratic forms on R
n
.
One measures distances in small neighbourhoods U

(u) U, u U, by setting
d
u,
(u

, u

) = |u

|
u
= (g
u
(u

, u

))
1
2
for u

, u

and then
denes the Riemannian (geodesic) distance dist
g
on U as the supremal metric
for these d
u,
as 0. Finally, Riemannian manifolds V appear as metric
spaces locally isometric to the above Us. (The latter step is sleek but hard to
implement. Try, for example, to show that the induced (intrinsic) metric on a
smooth submanifold V R
N
is Riemannian in our sense.)
The magic power of this denition is due to the innitesimal kinship of
Riemannian to Euclidean. If V is smooth (i.e. all gs on Us are at least
C
2
) then locally near each point v, one can represent V by a neighbourhood U
of the origin 0 R
n
with v 0, so that the corresponding g on U agrees with
the Euclidean (i.e. constant in v) metric g
0
= g
0
(x, y) = x, y) =
n

i=1
x
i
y
i
up to
the rst order,
g
u
= g
0
+
1
2
n

i,j=1
_

2
g(0)
u
i
u
j
_
u
i
u
j
+. . . ,
i.e. with the rst order Taylor terms missing and where, moreover, only
n
2
(n
2
1)
12
terms among
_
n(n+1)
2
_
2
second derivatives

2
g(0)
uiuj
do not vanish. The resulting
n
2
(n
2
1)
12
functions on U, when properly organized, make the Riemann curvature
tensor of V (which reduces to the Gauss curvature for n = 2) measuring the
deviation of (V, g) from atness (i.e. Euclideaness).
The (polylinear) algebraic structure built into g allows a full edged analysis
on (V, g), such as the Laplace-Hodge operator, potential theory etc. This turned
out to be useful for particular classes of manifolds distinguished by additional
(global, local or innitesimal) symmetry, where the major achievements coming
to ones mind are:
- Hodge decomposition on the cohomology of Kahler manifolds V and a
similar (non-linear) structure on the spaces of representations of
1
(V ).
- Existence of Einstein metrics on Kahler manifolds with algebra-geometric
consequences.
- Spectral analysis on locally symmetric (Bruhat-Tits and adelic as well as
Riemannian) spaces leading, for instance, to various cohomology vanishing the-
orems, T-property (with applications to expanders) and (after delinearization)
to super-rigidity of lattices in semi-simple Lie groups.
4
The linear analysis on general Riemannian manifolds pivots around the
Atiyah-Singer-Dirac operator and the index (Riemann-Roch) theorem(s). These
originated from Gelfands question (raised in the late fties) aiming at an ex-
plicit topological formula for the index of an elliptic operator (which is easily
seen to be deformation invariant) and became a central theme in mathematics
starting from the 1963 paper by Atiyah and Singer.
The non-linear Riemannian analysis on general V s followed for the most
part the classical tradition concentrating around elliptic variational problems
with major advances in the existence and regularity of solutions: minimal sub-
varieties, harmonic maps, etc. The most visible external application, in my
view, concerns manifolds with positive scalar curvature the subject motivated
by problems (and ideas) coming from general relativity resolved by Schoen
and Yau with a use of minimal hypersurfaces.
Manifolds of each dimension two, three and four make worlds of their own,
richer in structure than all we know so far about n 5.
In dimension two we possess the Cauchy-Riemann equations and are guided
by the beacon of the Riemann mapping theorem, the crown jewel of dierential
geometry.
The four-dimensional pecularity starts with algebra: the orthogonal group
O(4) locally decomposes into two O(3). This allows one to split (or rather to
square-root) certain natural (for the O(4)-symmetry) non-linear second order
operators in a way similar to how we extract the Cauchy-Riemann from the
boringly natural selfadjoint Laplacian on R
2
. The resulting rst order opera-
tors (may) have non-zero indices and satisfy a kind of non-linear index theorem
discovered by Donaldson in 1983 for the Yang-Mills and then extended to the
Seiberg-Witten equation. (Both equations were rst written down by physists,
according to the 20th century lore.)
Manifolds of dimension three borrow from their two- and four-dimensional
neighbours: Thurstons construction of hyperbolic metrics on basic 3-manifolds
relies on geometry of surfaces while Floer homology descends from Yang-Mills.
Shall we ever reach spaces beyond Riemanns imagination?
Remark and references. It will need hundreds of pages to account for
the above forty lines. Here we limit ourselves to a few points.
(a) The Riemannian metric g naturally (i.e. functorially) denes parallel
transport of vectors along smooth curves in V which is due to the absence of
rst derivatives in an appropriate Taylor expansion of g. This can be seen
clearly for V realized in some R
N
(which is not a hindrance according to the
Cartan-Janet-Burstin-Nash isometric embedding theorem) where a family X(t)
of tangent vectors is parallel in V along our curve parametrized by t R i
the ordinary (Euclidean) derivative
dX(t)
dt
R
N
is normal to V at (t) V for
all t. (This is independent of the isometric embedding V R
N
.) If a curve
5
!
Figure 2
: [0, 1] V comes back making a loop (i.e. (0) = (1)), every tangent vector
X = X(0) T
v
(V ), v = (0) V , transforms to

(X)
def
= X(1) T
v
(V ) and
we obtain a homomorphism from the group of loops at v to the linear group of
isometric automorphisms of the tangent space T
v
(V ) = R
N
, i.e. to the orthogo-
nal group O(n); its image H O(n) is called the holonomy group of V (which is
independent of v for connected V ). Generically, H = O(n) (SO(n) for orientable
V ), but sometimes H has positive codimension in O(n). For example, dimH = 0
i V is locally Euclidean (Parallel Postulate is equivalent to H = id) and if
V = V
1
V
2
, then H = H
1
H
2
O(n
1
) O(n
2
)
=
O(n) for n
1
, n
2
,= 0.
Then there are several discrete series of symmetric spaces monumental land-
marks towering in the vastness of all Riemannian metrics R
n
, S
n
, H
n
, CP
n
,
SL(n)/SO(n)... It is natural to think that these are essentially all V s with
small holonomy, since codimH > 0 implies a rather overdetermined system of
P.D.E. for g (for example, dimH = 0 curvature (g) = 0, i.e.
n
2
(n
2
1)
12
equations against mere
n(n+1)
2
components of g. Yet at metric exists!). But
lo and behold: lots of even dimensional manifolds carry Kahler metrics where
H U(n)
=
SO(2n). Just take a complex analytic submanifold V in C
N
(or in CP
N
) and observe (which is obvious once being said) that the parallel
transport in the induced metric preserves the complex structure in the tangent
spaces. (This may be not so striking, perhaps, for those who have absorbed the
impact of holomorphic functions, overdetermined by Cauchy-Riemann in many
variables, but there are less expected beautiful exotic holonomy beasts predicted
by Bergers classication and brought up to life by Briant, see [Bria].)
The Kahler world is tightly knit (unlike the full Riemannian universe) with
deep functorial links between geometry and topology. For example, the rst
cohomology of a compact K ahler V comes by the way of a holomorphic (!) map
to some complex torus, V C
d
/lattice for d =
1
2
rankH
1
(V ). This extends
to (non-Abelian) representations
1
(V ) GL(n) for n 2 (Siu, Corlette,
Simpson ..., see [A-B-C-K-T]) furnishing something like an unramied non-
Abelian Kahlerian class eld theory (in the spirit of Langlands program) but
we have no (not even conjectural) picture of the transcendental part of
1
(V )
6
(killed by the pronite completion of
1
). Can, for example,
1
(Kahler) have
an unsolvable word problem? Is there an internal structure in the category of
Kahler fundamental groups functorially reecting the geometry of the Kahler
category? (All known compact Kahler manifolds can be deformed to complex
projective ones and it may be preferable to stay within the complex algebraic
category, with no fear of ramications, singularities and non-projectiveness.)
(b) There exists a unique (up to normalization) second order dierential
operator S from the space of positive denite quadratic dierential forms (Rie-
mann metrics) g on V to the space of functions V R with the following two
properties.
S is Di-equivariant for the natural action of dieomorphisms of V on both
spaces.
S is linear in the second derivatives of g (being a linear combination of
components of the full curvature tensor).
Then S(g) (or S(V )) is named the scalar curvature of (V, g) with the cus-
tomary normalization S (S
n
) = n(n 1).
If n = 2, S coincides with Gauss curvature, it is additive for products,
S(V
1
V
2
) = S(V
1
)+S(V
2
) and it scales as g
1
, i.e. S(g) =
1
S(g) for
> 0.
The following question proved to be more to the point than one could expect.
What is the geometric and topological structure of manifolds with S > 0?
(This comes from general relativity as S > 0 on world sheets reects positivity
of energy.)
The condition S > 0 appears quite plastic for n 3, where one can rather
freely manipulate metrics g keeping S(g) > 0, e.g. performing geometric surgery;
besides, every compact V
0
turns into V with S(V ) > 0 when multiplied by
a small round sphere of dimensions 2. Yet this plasticity has its limits:
Lichnerowicz found in 1963 a rather subtle topological obstruction (

A(V ) = 0 if
V is spin) with the use of the index theorem. Then Schoen and Yau approached
the problem in 1979 from another angle (linked to ideas in general relativity)
and proved, among other things, that n-tori (at least for n 7) carry no
metrics with S > 0 thus answering a question by Geroch. Inspired by this, we
revived with Blaine Lawson in 1980 Lichnerowicz idea, combined it with the
Lusztig-Mistchenko approach to the Novikov conjecture on homotopy invariance
of Pontryagin classes of non-simply connected manifolds and found out that the
bulk of the topological obstructions for S > 0 comes from a limit on geometric
size of V induced by the inequality S > 0 (similar to but more delicate than
that for K > 0, see III).
Yet, the above question remains open with an extra mystery to settle: what
do minimal hypersurfaces and the Dirac operator have in common? (Seemingly,
nothing at all but they lead to almost identical structure results for S > 0, see
[Gro
PCMD
] for an introduction to these issues.)
7
It appears that an essential part of the diculty in understanding S > 0
(and the Novikov conjecture) is linked to the following simple minded question:
what is the minimal > 0, such that the unit sphere S
N

(1) in the Banach space

= (R
N
, |x| = sup
i
[x
i
[) admits a -Lipschitz map into the ordinary n-sphere
S
N
(1) in R
N
=
N
2
with non-zero degree? Probably, for N (even if
we stabilize to maps S
N

(1) S
M
(R) S
NM
(1) with arbitrarily large M and
R) and this might indicate new ways of measuring size of V in the context of
S > 0 and the Novikov conjecture.
Soft and hard. Geometric (and some non-geometric) spaces and cate-
gories (of maps, tensors, metrics, (sub)varieties...) can be ranked, albeit am-
biguously, according to plasticity or exibility of (the totality of) their members.
(1) Topology could appear abby and structureless to Poincares contem-
poraries but when factored by homotopies (the very source of exibility) it
crystallizes to a rigid algebraic category as hard and symmetric as a diamond.
(2) Riemannian manifolds, as a whole, are shapeless and exible, yet they
abide conservation laws imposed by the Gauss-Bonnet-Chern identities.
Deeper rigidity appears in the presence of elliptic operators extracting nite
dimensional structure out of innite dimensional depth of functional spaces.
Also we start seeing structural rigidity (e.g. Cheeger compactness) by ltering
metrics through the glasses of (say, sectional) curvature.
(3) Kahler metrics and algebraic varieties seem straight and rigid in the Rie-
mannian landscape (never mind a dense set of Riemannian spaces appearing as
real loci of complex algebraic ones) but they look softish in the eye of an alge-
braic geometer. He/she reinforces rigidity with the Calabi-Yau-Aubin theorem
turning Kahler to Einstein-Kahler. (Nothing of the kind seems plausible in the
full Riemannian category for n > 4.)
(4) Homogeneous, especially symmetric, spaces stand on the top of the ge-
ometric rigidity hierarchy (tempting one to q-deform them) and (sometimes
hidden) symmetries govern integrable (regarded rigid) systems. (Softness in
dynamics is associated with hyperbolicity.)
(5) Lattices in semi-simple Lie groups grow in rigidity with dimension,
passing the critical point at SL
2
(C), where they ourish in Thurstons
hyperbolic land. A geometer unhappy with Mostow (over)rigidity for n > 3
is tempted to switch from lattices to (less condensed) subgroups with innite
covolumes and more balanced presentations (to dismay of a number theorist
thriving on the full arithmetic symmetry of ). Most exible among all groups
are (generalized) small cancellation ones followed by higher dimensional hy-
perbolic groups while lattices and nite simple groups are top rigid. Vaguely
similarly, the rigidity of Lie algebras increases with decrease of their growth
culminating in Kac-Moody and nite dimensional algebras.
(6) Holomorphic functions on Stein manifolds V are relatively soft (Cartans
theory) as well as holomorphic maps f : V W for homogeneous and elliptic
8
(i.e. with a kind of exponential spray) W by the (generalized) Grauert theorem
allowing a homotopy of every continuous map f
0
: V W to a holomorphic
one. Holomorphic maps moderated by bounds on growth become more rigid
(e.g. functions of nite order have essentially unique Weierstrass product de-
composition). Algebraic maps, ordinarily rigid, sometimes turn soft, e.g. for
high degree maps of curves to P
1
by the Segal theorem. And Voevodski theory
(if I interpret correctly what little I understood from his lecture) softens the
category of algebraic varieties by injecting some kind of homotopies into these.
(7) Big three. There are three outstanding instances where striking struc-
tural patterns emerge from large and exible geometric spaces: symplectic/con-
tact, dimension 4 and S > 0, conducted in all three cases by Riemannian and
elliptic. We met S > 0 earlier (which seems least conceptually understood
among the big three), symplectic and contact belong to Eliashberg and Hofer
at this meeting (with soft versus hard discussed (in [Gro
SH
]) and nobody,
alas, gave us a panorama of n = 4).
(8) h-Principle. Geometers believed from 1813 till 1954, since Cauchy (al-
most) proved rigidity of closed convex polyhedral surfaces in R
3
, that isometric
immersions are essentially rigid. Then Nash deed everybodys intuition by
showing that every smooth immersion of a Riemannian manifold f
0
: V R
N
can be deformed, for N 2 n = dimV , to a C
1
-smooth (not C
2
!) isometric
f : V R
N
with little limitation for this deformation, allowing one in par-
ticular, to freely C
1
-deform all V R
N
keeping the induced (intrinsic) metric
intact. (This is sheer madness from a hard-minded analysts point of view as the
N components of f satisfy
n(n+1)
2
partial dierential equations comprising an
overdetermined system for N <
n(n+1)
2
, where one expects no solutions at all!)
The following year (1955) Kuiper adjusted Nashs construction to N = n + 1
thus disproving C
1
-rigidity of convex surfaces in R
3
.
Next, in 1958, Smale stunned the world by turning the sphere S
2
R
3
inside
out. He did it not by exhibiting a particular (regular) homotopy (this was done
later and only chosen few are able to follow it through) but by developing a
homotopy theoretic approach used by Whitney for immersions of curves into
R
2
. Then Hirsch incorporated Smale into the obstruction theory and showed
that a continuous map f
0
: V W can be homotoped into an immersion
if the obvious necessary condition is satised: f
0
lifts to a berwise injective
homomorphism of tangent bundles, T(V ) T(W), with the exception of the
case of closed equidimensional manifolds V and W where the problem is by far
more subtle.
It turned out that many spaces X of solutions of partial dierential equations
and inequalities abide the homotopy principle similar to that of Nash, Smale-
Hirsch and Grauert: every such X is canonically homotopy equivalent to a
space of continuous sections of some (jet) bundle naturally associated to X.
(For example, the space of immersions V W is homotopy equivalent to the
space of berwise injective morphisms T(V ) T(W) by the Hirsch theorem.)
9
Figure 3
The geometry underlying the proof of the h-principle is shamefully simple in
most cases: one creates little (essentially 1-dimensional, `a la Whitney) wrinkles
in maps x X which are spread all over by homotopy and render X soft and
exible. But the outcome is often surprising as seen in the (intuitively inconceiv-
able) Milnors two dierent immersed disks in the plane with common boundary
which come up with logical inevitability in Eliashbergs folding theorem.
Despite the growing array of spaces subjugated by the h-principle (see
[Gro]
PDR
, [Spr]) we do not know how far this principle (and softness in general)
extends (e.g. for Gauss maps with a preassigned range, see I). Are there sources
of softness not issuing from dimension one? Some encouraging signs come from
Thurstons work on foliations, Gao-Lohkamp h-principle for metrics with Ricci
< 0, and especially from Donaldsons construction of symplectic hypersurfaces
(where softness is derived from a kind of ampleness not dissimilar in spirit
to Segals theorem, see (6)).
A tantalizing wish is to nd new instances, besides the big three, where
softness reaches its limits with something great happening at the boundary.
Is there yet undiscovered life at the edge of chaos? Are we for ever bound
to elliptic equations? If so, what are they? (There are few globally elliptic
non-linear equations and no general classication. But even those we know,
coming from Harvey-Lawson calibrations, remain mainly unexplored.) And if
this wish does not come true we still can make living in soft spaces exploring
their geometry (their topology is completely accounted for by the h-principle)
as we do in anisotropic spaces (see III).
(9) Our soft and hard are not meant to reveal something profound about
the nature of mathematics, but rather to predispose us to acceptance of geomet-
ric phenomena of various kinds. Besides, it is often more fruitful to regard num-
bers, symmetric spaces, GalQ/Q, SL
n
(A)/SL
n
(Q)... as true mathe-
matical entities rather than descendants of our general spaces, groups,
algebras etc. But one cannot help wondering how these perfect entities could
originate and survive in the softly structured brain hastily assembled by blind
evolution. Some basic point (scientic, not philosophical) seems to completely
elude us.
10
Nature and naturality of questions. Here are (brief, incomplete, per-
sonal and ambiguous) remarks intended to make clearer, at least terminologi-
cally, the issues raised during discussions we had at the meeting.
Natural may refer to the structure or nature of mathematics (granted this
exists for the sake of argument), or to natural for human nature. We divide
the former into (pure) mathematics, logic and philosophy, and the latter, ac-
cording to (internal or external) reward stimuli, into intellectual, emotional,
and social. E(motional) plays upper hand in human decision (and opinion) (ex-
cept for a single man you might have a privilege to talk mathematics to) and
in some people (Fermat, Riemann, Weil, Grothendieck) i-e naturally converges
to m-l-p. But for most of us it is not easy to probe the future by conjecturally
extrapolating mathematical structures beyond present point in time. How can
we trust our mind overwhelmed by i-e-s ideas to come up with true m-l-p
questions?(An e-s-minded sociologist would suggest looking at trends in fund
distribution, comparable weights of authorities of schools and individuals and
could be able to predict the inuencial role of Hilberts problems and Bourbaki,
for example, without bothering to read a single line in there.) And i-e-s-
natural does not make a stupid question: the 4-color problem, by its sheer
diculty (and expectation for a structurally rewarding proof) has focused at-
tention on graphs while the solution has claried the perspective on the role
of computers in mathematics. But this being unpredictable, and unrepeatable,
cannot help us in m-l-p-evaluation of current problems which may look i-e de-
ceptively 4-colored. (As for myself, I love unnatural, crazily unnatural problems
but you stumble upon them so rarely!)
III. K 0 and other metric stories.
What are most Euclidean Riemannian manifolds?
We have been already acquainted with the fully homogeneous spaces also
called, for a good reason, (complete simply connected) of constant curvature
K: the round S
n
with K = +1, the at R
n
with K = 0, and the hyperbolic
H
n
with K = 1.

(Observe that S
n
and H
n
converge to R
n
for
in a natural sense, where (X, dist)
def
= X(, dist) and K(X) =

1
2
K(X),
as is clearly seen, for example, for -scaled surfaces X R
3
.) Now, somewhat
perversely, we bring in topology and ask for compact manifolds with constant
curvature, i.e. locally isometric to one of the above S
n
, R
n
, or H
n
. Letting S
n
go, we start with the at (i.e. K = 0) case and conrm that compact locally
Euclidean manifolds exist: just take a lattice in R
n
(e.g. = Z
n
) and look
at the torus T = R
n
/
n
. Essentially, there is little else to see:
F-theorem. Every compact at manifold X is covered by a torus with the
number of sheets bounded by a universal constant k(n).

The fourth and the last fully homogeneous Riemannian space is P

n
= S
n
/{1}.
11
This sounds dry but it hides a little arithmetic germ on the bottom: there is
no regular k-gon with vertices in a lattice (e.g. = Z
2
) in R
2
for k 7 (or
for k = 5). Indeed, transported edges of such k-gone R would make a smaller
regular R

and the contradiction follows by iteration R

, R

, . . . .
R'
R
Figure 4
The tori T themselves stop looking at as they all together make the marvellous
moduli space SO(n)SL
n
(R)/SL
n
(Z) (of isometry classes of Ts with Vol T = 1)
locally isometric to SL
n
(R)/SO(n) apart from mild (orbifold) singularities due
to elements of nite order (< 7) in SL
n
(Z).
Turning to K = 1, we may start wondering if such spaces exist in a compact
form at all. Then, for n = 2, we observe that the angles of small regular k-gones
R H
2
are almost the same as in R
2
while large R H
2
have almost zero
angles: thus, by continuity, for every k 5, there exists R

H
2
with 90

angles. We reect H
2
in (the lines extending the) sides of R

and take the

subgroup Isom(H
2
) generated by these k reections. This is discrete on
H
2
with R

serving as a fundamental domain similarly to the case of the square

R

R
2
and the quotient space H
2
/ (equal R) becomes an honest manifold
(rather than orbifold) if we take instead of a subgroup

without torsion
(which is not hard to nd).
Figure 5
The same idea works for dodecahedra in H
3
and some other convex poly-
hedra in H
n
for small n, but there are no compact hyperbolic reection groups
for large n (by a dicult theorem of Vinberg). The only (known) source of
high dimensional comes from arithmetics, essentially by intersecting SO(n, 1)
somehow embedded into SL
N
(R) with SL
N
(Z) (where the orthogonal groups
SO(n, 1) double covers the isometry group PSO(n, 1) of H
n
R
n,1
). Non-
arithmetic are especially plentiful for n = 3 by Thurstons theory and often
have unexpected features, e.g. some V = H
3
/ ber over S
1
(which is hard
to imagine ever happening for large n). Moreover, the topological 3-manifolds
bered over S
2
generically, (i.e. for pseudo-Anosov monodromy) admit metrics
12
with K = 1. Unbelievable but is true by Thurston (who himself does not
exclude that nite covers of most atorical V ber over S
1
; yet this remains open
even for V with K = 1).
Alexandrovs spaces. What are the most general (classes of ) spaces sim-
ilar to those with K = 1?
Alexandrov suggested an answer in 1955 by introducing spaces with K 0,
where the geodesic triangles have the sum of angles 2, and those with K 0,
where (at least small) triangles have it 2. But we take another, more
functorial route departing from the following
Euclidean K-theorem. Every 1-Lipschitz (i.e. distance non-increasing)
map f
0
from a subset R
n
to some R
m
admits a 1-Lipschitz extension
f : R
n
R
m
for all m, n .
This is shown by constructing f point by point and looking at the worst case
at each stage where extendability follows from an obvious generalization of the
pretty little lemma:
Lemma. Let and

in S
n1
R
n
be the sets of vertices of two simplices
(inscribed into the sphere S
n1
) where the edges of are correspondingly
than those of

. Then is congruent to

(i.e. =), provided is not

contained in a hemisphere.
Now we say that a metric space X has K 0, if every partial 1-Lipschitz
map X R
m
extends to a 1-Lipschitz f : X R
m
, for all m, while K 0
is dened with such extensions for R
n
X, where in the case K 0 one
requires, besides the existence of an extension R
n
X, the uniqueness of this
on the convex hull Conv R
n
for n = 1, provided the starting map X
was isometric.
To make this worthwhile, one adds the metric completeness of X and the
locality property: dist(x, x

) should equal the inmal length of curves in X

between x and x

. (Equivalently, there is a middle point y X where dist(x, y)+

dist(y, x

) = dist(x, x

).) Then, one arrives at the following elegant proposition

justifying the denitions:
K-Theorem. If K(X) 0 and K(Y ) 0, then every partial 1-Lipschitz
map X Y admits a 1-Lipschitz extension X Y .
To apply this one needs examples of spaces with K 0 and these are
easier to observe with Alexandrovs denition. Fortunately, both denitions are
equivalent and we have:
Complete (e.g. closed) convex hypersurfaces in R
n
have K 0, while saddle
surfaces in R
2
have (at least locally) K 0. Symmetric spaces of compact type
have K 0 while those of non-compact type have K 0.
Take a 2-dimensional polyhedron V assembled of convex Euclidean k-gones
and observe that the link L of each vertex in V is a graph (i.e. 1-complex) with
a length assigned to every edge e equal the k-gonal angle corresponding to e.
13
Then K(V ) 0 each L is isometric to a segment or to a circle no longer
than 2.
K(V ) 0 all cycles in all L are longer than 2 and V is simply connected.
Finally, if X has locally K 0, then so is true globally while globally
K 0 (locally K 0) + (
1
= 0).
The
1
= 0 condition breaks the harmony (I guess it was upsetting Alexan-
drov) and brings confusion to the notion of K 0 as the local and global
meanings diverge. But in the end of the day the
1
-ripple makes the geometry
of K 0 much richer (and softer) than all we know of K 0, since there
are lots of spaces V with K
loc
0 (as is already seen in the 2-dimensional
polyhedra) where the group theoretic study of
1
(V ) may rely on geometry (for
example, in the Novikov conjecture). Besides, the global denition of K 0
can be relaxed by, roughly, allowing -Lipschitz extension of partial 1-Lipschitz
maps with 1 const < bringing along larger classes of (hyperbolic-like)
spaces and groups where geometry and algebra are engaged in a meaningful
conversation.
Anisotropic spaces. There is a class of metrics which can be analytically
generated with the same case as the Riemannian ones; besides, we nd among
them spaces X in some way closer to R
n
than S
n
and H
n
: these X are metrically
homogeneous as well as self-similar, i.e. X is isometric to X for all > 0. (In
the Riemannian category there is nothing but R
n
like that.)
A polarization on a smooth manifold V (e.g. R
n
or a domain in R
n
) is a
subbundle H of the tangent bundle T(V ), i.e. a eld of m-planes on V , where for
1 m n1, n = dimV in the case at hand. Besides H we need an auxiliary
Riemannian metric g on V but what matters is the restriction of g on H. We
dene dist(v, v

) = dist
H,g
(v, v

) by taking the inmum of g-lengths of piecewise

smooth curves between v and v

which are chosen among H-horizontal curves,

i.e. those which are everywhere tangent to H. It may happen that this distance
is innite (even for connected V ) if some points admit no horizontal connecting
paths between them, as it happens for integrable H, where dist(v, v

) < v,
and v

lie in the same leaf of the foliation integrating H. This is not so bad as
it seems but we want dist < at the moment and so we insist on the existence
of a horizontal path between every two points in V . It is not hard to show that
generic C

-smooth polarizations H do have this property for m 2, where

generic implies, in particular, that the space of good Hs is open and dense
in the space of all C

-polarizations on V .
A practical way for checking this is to take m
+
m vector elds tangent
to H, and spanning it (these always exist), say X
1
, . . . , X
m+
. Then the su-
cient criterion for our H-connectivity (also called controllability) is as follows:
The successive commutors X
i
, [X
i
, X
j
], [[X
i
, X
j
], X
k
] . . . of the elds span the
tangent bundle T(V ).
14
The simplest instance of the above is the pair of the elds X
i
=

x1
and
X
2
=

x2
+ x
1

x3
in R
3
, where [X
1
, X
2
] =

x3
. Here H can be presented as
the kernel of the standard contact form dx
3
+x
1
dx
2
and, in fact, contact elds
H are H-connected for all contact manifolds of dimension 3.
Next, look at a left invariant polarization H on a Lie group G dened by
the left translates of a linear subspace h in the Lie algebra L = L(G). It is not
hard to see that the above holds there is no Lie subalgebra in L containing
h besides L itself. Take, for instance, the 3-dimensional group Heisenberg G
(homeomorphic to R
3
) with L = L(G) generated by x
1
, x
2
, x
3
, where [x
1
, x
2
] =
x
3
and x
3
commutes with x
1
and x
2
. Here one takes h spanned by x
2
, x
3
and
observe that this h is invariant under automorphisms of L dened by x
1

2
x
2
1
, x
2
x
2
, x
3
x
3
for all > 0. Then, for each left invariant metric g
on G, the corresponding automorphisms A

: G G preserves H and scale g

on H by . Thus G is selfsimilar but quite dierent from R
3
. For example, the
Hausdor dimension of (G, dist
H,g
) is 4 rather than 3.
What are natural maps between the above non-isotropic spaces? Lipschitz
maps do not serve here as well as in the Riemannian category since our new
spaces are usually not mutually bi-Lipschitz equivalent. On the other hand
typical spaces X and Y of the same topological dimension are locally Holder
homeomorphic with some positive exponent < 1 with a bound >
n
> 0.
But the optimal value of for given spaces (or classes of those) remains un-
known. A similar problem is that of nding sharp lower bound for the Hausdor
dimension of a subset Y X = (X, dist
H,g
) by the topological dimension in
terms of commutation properties of elds spanning H. (This is quite easy if Y
is a smooth submanifold in X.)
The large scale geometry of (X, dist
H,g
) is rather close to the Riemannian
geometry of (X, dist
g
) and so H does not matter much. Conversely, the local
geometry essentially depend on H (and very little on g) and the main open
problems are basically local.
Concentrated mm spaces. Let X be a metric space which is also given
a Borel measure , often assumed to be a probability measure, i.e. (X) = 1.
This may come out of the metric (e.g. the Riemannian measure, sometimes
normalized to have (X) = 1), but often has little to do with the original
distance as, for example, the Gaussian measure on R
n
. We want to study
(X, dist, ) in a probabilistic fashion by thinking of functions f on X as random
variables being concerned with their distributions, i.e. push-forward measures
f

(). Here is our

Basic problem. Given a map f : X Y , relate metric properties of f to
the structure of the measure f

() on Y .
Here metric refers to how f distorts distance (expressed, for instance, by
the Lipschitz constant

(f)), where we distinguish the case of functions, i.e. of
Y = R. On the probabilistic side we speak of structure of f

() expressed
15
entirely in terms of Y and where a typical question is how concentrated f

()
is, i.e. how close it is to a point measure in Y .
Gaussian example. Take R
N
for X with the measure (exp |x|
2
)dx.
Then, every 1-Lipschitz function f : R
N
R is at least as much concentrated
as the orthogonal projection f

: R
N
R. (There is a 1-Lipschitz self-mapping
R R pushing forward f

() to f

().)
A parallel example, where the geometry is seen clearer, is X = S
N
with
the normalized Riemannian measure. Here again every 1-Lipschitz function
f : S
N
R is concentrated as much as the linear one; consequently, f

()
converges, for N , to a -measure on R (with the rate

N).
The essence of the above concentration is the sharp contrast between the
spread of the original measure on X (e.g. the distance between -random points
in S
N
is /2 for large N) and strong localization of f

() on Y (e.g. the
characteristic distance on R with respect to f

() is 1/

N in the spherical
case). The following denition is aimed to capture this phenomenon in the limit
for N by interbreeding metric geometry with the ergodic theory (not quite
as in the ergodic theorem where f = f
N
appears as the average of N transforms
of a given f
0
).
Let X be a (probability) measure space and d : XX R
+
a function
satisfying the standard metric axioms except that we allow d(x, x

) = . In fact,
we are keen at the (apparently absurd) situation of d = almost everywhere
on X, moderated by the
Ergodicity axiom. For every Y X with (X) > 0 the distance to Y ,
d
Y
(x) =
def
inf
yY
d(x, y)
is measurable and a.e. nite on X.
Example. Let X be a foliated measure space where each leaf is (measurably
in x X) assigned a metric. Then we dene
d(x, y) =
_
if x, y are not in the same leaf
d(x, y) = dist
L
(x, y) if x and y lie in some leaf L,
and observe that our ergodicity axiom amounts here to the ordinary ergodicity.
Next we distinguish concentrated spaces by insisting on the universal bound
on the distances between subsets in X in terms of the measures of these subsets.
Here
dist(Y, Y

)
def
= inf d(y, y

) = inf
xY

d
Y
(x)
and the bound is given by a function C(a, a

), a, a

> 0, (where C may go to

innity for a, a

0), so that dist(Y, Y

) C((Y ), (Y

)).
16
Examples. If X in the previous example is foliated (i.e. partitioned) into
the orbits of an amenable group G acting on X, then the resulting d on X is,
essentially, never concentrated. But if G has property T, then it is concentrated.
Let X
1
, X
2
, . . . be a sequence of Riemannian manifolds and X be the innite
Cartesian product, X
1
X
2
. . ., where the metric between innite sequences
x = (x
1
, x
2
. . .) and y = (y
1
, y
2
, . . .) is the Pythagorean one,
dist(x, y)
def
=
_
dist
2
(x
1
, y
1
) + dist
2
(x
2
, y
2
) +. . .
_
1
2
.
What goes wrong here is that dist(x, y) = for most x and y but we can tolerate
this in the presence of the product measure on X coming from normalized
Riemannian measures on X
i
. If the rst (sometimes called second) eigenvalues
of (the Laplace operator on) X
i
are separated from zero,
1
(X
i
) > 0, then
the product is concentrated and, moreover, one can make a meaningful analysis
(on functions and often on forms) on X. Furthermore, if
1
(X
i
) for
i , then has discrete spectrum on X with nite multiplicity (computed
by the usual formula for products). For instance, if X
i
are n
i
-spheres of radii
R
i
, then
1
(X
i
) = n
i
R
2
i
and their product is concentrated for R
i

n
i
with
extra benets for

n
i
/R
i
.
One can deform and modify products, retaining concentration, e.g. for pro-
jective limits of some towers of smooth bration, such as the innitely iterated
unit tangent bundles of Riemannian manifolds.
Similarly to products, the spaces X of maps between Riemannian manifolds,
A B, carry (many dierent) foliated Hilbert manifold structures which
in the presence of measures (e.g. Wiener measure for 1-dimensional A) allow
analysis on X.
Now comes a painful question: are these X good for anything? Do they
possess a structural integrity or just encompass (many, but so what?) examples?
A convincing theorem is to be proved yet.
Remarks and references.
(a) The F-theorem makes the core of Bieberbachs solution to a problem on
Hilberts list (N. 18, where the n-dimensional hyperbolic case is also mentioned
and dismissed as adding little new to the results and methods of Fricke and
Klein).
(b) Our denition of K 0 is motivated by [La-Sch], where the authors
prove the K-theorem for maps from spaces with K to those with K
(in the sense of Alexandrov) under a mild restriction ruling out, for example,
maps S
n
S
m
for m < n, where K-property obviously fails to be true but,
unfortunately, missing maps S
n
S
m
for m n where it is known to be true
(a conclusive version seems not hard but no published proof is available). Also,
there is a Lipschitz extension result from arbitrary metric spaces to those with
K 0 (due to Lang, Pavlovic and Schroeder).
17
(c) The theory of spaces with K 0 (as well as with K < ) is by
now well developed (see [Per]) and structurally attractive. Yet it suers from
the lack of a systematic process of generating such spaces apart from convex
hypersurfaces (and despite several general constructions: products, quotients,
spherical suspension ...). Also, there is no serious link with other branches of
geometry, not even with the local theory of Banach spaces, and there are few
theorems (K-theorem is a happy exception) where the conclusion is harder to
verify (in the available examples) than the assumptions (as is unfortunately
frequent in the global Riemannian geometry). A possible way of enriching (and
softening) K 0 is letting n (and taking n = seriously) as is done for
Banach spaces. (See [Berg] for a broader perspective, [Pet] for a most recent
account and [Gro
SGM
] for a pedestrian guide to curvature.)
(d) Ricci. Analytically speaking, the most natural of curvatures is the Ricci
tensor that is the quadratic dierential form on V associated to g via a (essen-
tially unique) Di-invariant second order quasi-linear dierential operator (like
S) denoted Ri = Ri(g). Manifolds with positive (denite) Ri generalize those
with K 0 but they admit no simple metric description as their essential fea-
tures involve the Riemannian measure on V , e.g. R-balls in V with Ri 0 have
smaller volumes than the Euclidean ones. (If K 0 is motivated by convexity,
then Ri 0 can be traced to positive mean curvature of hypersurfaces.) Thus
it remains unclear how far the idea of Ri 0 extends beyond the Riemannian
category: what are admissible singularities and what happens for n = ? (See
[Gro
SGM
] and [G-L-P] for an introduction and [Che-Co] for the present state of
art.)
Encouraged by Ri 0 one turns to Ri 0 formally generalizing K 0 but
the naive logic does not work: every metric can be approximated by those with
Ri < 0 by the Lohkamp h-principle and no hard structural geometry exist.
(e) Einstein and the forlorn quest for the best metric. It is geometers
dream (rst articulated by Heinz Hopf, I believe) to nd a canonical metric g
best
on a given smooth manifold V so that all topology of V will be captured by
geometry. This happened to come true for surfaces as all of them carry (almost
unique) metrics of constant curvature and is predicted for n = 3 by Thurstons
geometrization conjecture. Also, there is a glimpse of hope for n = 4 (Einstein,
self-dual) but no trace of g
best
has ever been found for n 5. What is the reason
for this? Let us take some (energy) function E on the space ( of metrics, say
built of the full curvature tensor, something like
_
(Curv(g))
n
2
dv (where the
exponent n/2 makes the integral scale invariant). Imagine, the gradient ow
of E brings all of ( to a nice subspace (
best
( (ideally, a single point or
something not very large anyway). Then the group Di V would act on (
best
(as
all we do should be Di-invariant) with compact isotropy subgroups (we assume
V is compact at the moment), e.g. if (
best
consisted of a single point, then Di
V would isometrically act on (V, (
best
). But the high dimensional topology
(unknown to Hopf) tells us that the space Di V is too vast, soft and unruly to
be contained in something nice and cosy like the desired (
best
. (Di is governed
18
by Waldhausen K-theory bringing lots of homotopy to Di which hardly can
be accounted for by a rigid geometry. And prior to Waldhausen our dream
was shattered by Milnors spheres ruling out smooth canonical deformations of
general metrics on the spheres S
n
, n 6, to the standard ones.)
Besides topology, there is a geometric reason why we cannot freely navigate
in the rugged landscapes of spaces like Di. To see the idea, let us look at the
simpler problem of nding the best closed curve in a given free homotopy
class of loops in a Riemannian manifold V . If K(V ) 0, for instance, the
gradient ow (of the energy function on curves) happily terminates at a unique
geodesic: this is the best we could hope for. But suppose
1
(V ) is computation-
ally complicated, i.e. the word problem cannot be solved by a fast algorithm,
say, unsolvable by any algorithm at all. Our ow (discretized in an obvious way)
is a particular algorithm, we know it must badly fail, and the only way for it to
fail is to get lost and confused in deep local minima of the energy. Thus our V
must harbor lots and lots of locally minimal geodesics in each homotopy class,
in particular, innity of contractible closed geodesics, disrupting the route from
topology to simple geometry.
Well, one may say, let us assume
1
(V ) = 0. But what the hell does it
mean assume? Given a V , presented in any conceivable geometric form (re-
member, we are geometers here, not shape-blind topologists), there is no way
to check if
1
(V ) = 0 since this property is not algorithmically veriable. Con-
sequently, there are innocuously looking metrics on such manifolds as S
n
for
n 5, teeming with short closed curves which no human being can contract
in a given stretch of time. (In fact, a predominant majority of metrics are like
this for n 5.) A similar picture arises for higher dimensional (e.g. minimal)
subvarieties (with extra complication for large dimension and codimension, even
for simplest V such as at tori, where the trickery of minimal subvarieties was
disclosed by Blaine Lawson), and by the work of Alex Nabutovski the spaces like
((V )/ Di(V ) harbor the same kind of complexity rooted in the Godel-Turing
theorem.
Following Alex we (I speak for myself) are lead to the pessimistic conclusion
that there is no chance for a distinguished g
best
(or even (
best
() for n 5
and that natural metrics, e.g. Einstein ( with Ri(g) = g for < 0, must
be chaotically scattered in the vastness of ( with no meaningful link between
geometry and topology. (This does not preclude, but rather predicts, the exis-
tence of such metrics, e.g. Einstein, on all V of dimension 5: the problem is
there may be too many of them.)
On the optimistic side, we continue searching for g
best
in special domains
in (, e.g. following Hamiltons Ricci ow (say, for Ricci 0 or K 0) or
stick to dimensions three (where Michael Anderson makes his theory) or four
(where Taubs discovered certain softness in selfduality). Alternatively, we can
enlarge (rather than limit) the category and look for extremal (possibly) singular
varieties with only partially specied topology, i.e. with a prescribed value of a
certain topological invariant, such as a characteristic number or the simplicial
19
volume. For example, each (decent) topological space X is, rather canonically,
accompanied by metric spaces homotopy equivalent to it, such as a suitably
subdivided semisimplicial model of X which is an innite dimensional simplicial
complex, call it X

hom
X, where the metric on X

comes from a choice of a

standard metric on each k-simplex, k = 1, 2, . . .. This X

is too large to please

our eye, but it may contain some distinguished subvarieties, e.g. minimizing
homology classes in H

(X

) = H

(X), at least for certain spaces X (see 5 H

+
in [G-L-P] and references therein).
(f) Anisotropic metrics appear under a variety of names: non-holonomic,
control, sub-elliptic, sub-Riemannian, Carnot-Caratheodory, bearing
the traces of their origin. They were extensively studied by analysts since
Hormanders work on hypo-elliptic operators but I do not know where and
when they were promoted from technical devices to full membership in the met-
ric community. Now, besides P.D.E., they help group theorists in the study of
discrete nilpotent groups since the local geometry of (especially self-similar)
metrics on nilpotent Lie groups G adequately reects the asymptotic ge-
ometry of . The local geometry of dist
H
reduces, in turn, to that of the polar-
ization H T(G) which is rather soft as far as low dimensional H-horizontal
(i.e. tangent to H) subvarieties in G are concerned as follows from the Nash
implicit function theorem. This provides some information on dist
H
-minimal
subvarieties in G (alas, only the case of surfaces is understood) and allows us to
evaluate the Dehn function of (see [Gro
CC
], [Gro
AI
] and references therein).
Also observe that (however meager) results and problems in the anisotropic
geometry may serve as model for other soft spaces (of solutions of P.D.E. as
H-horizontality for H).
(g) It is helpful to think of a (metric) measure space (X, ) as a high di-
mensional conguration space of a physical system (which is, indeed, often the
case). Here f : X Y is an observable projecting X to a low dimensional
screen, our Y such as the space-time R
4
, for instance, where f

() is what
we see on the screen. The concentration of f

(), ubiquitous in the probability

theory and statistical mechanics, was brought to geometry (starting from Ba-
nach spaces) by Vitali Milman following the earlier work by Paul Levy. The
Levy-Milman concentration phenomenon has been observed for a wide class of
examples, where, besides the mere concentration, one is concerned with large
deviations and uctuations. (See [Mi], [G-L-P] and references herein.) Unfortu-
nately, our denition of concentrated spaces does not capture large deviations
(which are more fundamental than uctuations). Possibly, this can be helped by
somehow enriching the structure. Besides, one can proceed by allowing variable
measures (states) as in the cube 0, 1
N
with the product
p
of N-copies of
the (p, 1p)-measure on 0, 1 and in Gibbs measures parametrized by temper-
ature. Then, inspired by physics, one wonders what should be general objects
responsible for concentration in quantum statistical mechanics (where the con-
centration is limited by the Heisenberg principle) and, nally, one may turn to
non-metric structures which may come in probability and geometry along with
concentration.
20
IV. Life without metric.
It is (too) easy to concoct invariants of a metric space V , e.g. by looking at the
ranges of Cartesian powers of the distance function mapping V
k
to R
k(k1)
2
for
d
k
: (v
1
, . . . , v
i
, . . . , v
k
) dist(v
i
, v
j
)
1i<jk
.
(The diameter of V appears, for example, as the maximal segment in the im-
age of d
k
in R
1
+
for k = 2.) Then, there are various (often positive) energy
functions E on spaces of subvarieties W in X and maps f : V V
0
such
as W Vol
m
W for m = dimW, and
_
|df|
2
dv for Riemannian V and V
0
.
Each E generates invariants of V , for instance, inma of E on given classes of
Ws or fs, (e.g. inf Vol
m
W for [W] = h H
m
(V )) or, more generally, the full
Morse landscape of E including the spectrum of the critical values of E (e.g. the
spectrum of the Laplace operator on functions and form on a Riemannian V ).
These invariants I provide us with the raw material for asking questions and
making conjectures: what are possible values of I = I(E) and relations between
Is for dierent Es? What are spaces V with a given behaviour of an I?, etc.
But spaces without metrics become rather slippery, hard to grasp and assess.
Just look at a foliation or dynamical system F on a manifold V . The essential
invariants (entropy, asymptotics of periodic orbits) often change discontinuously
under deformations (perturbations) of the structure and are hard to evaluate,
even approximately, for a given F. And it is not easy at all to come up with
new numerical invariants making sense for all objects in the category. This is
due to the fact that the (local) group of (approximate) automorphisms of such
a structure (at a point v V ) is potentially non-compact. Consequently, the
action of Di(V ) on the space T of our structures F on V may have non-trivial
dynamics (e.g. non-compact isotropy groups Is
F
= Aut(V, F), F T) making
the quotient space T/ Di V , where our invariants are supposed to live, non-
separable. (Intuitively, invariants should be independent of observers attached
to dierent coordinate systems in V ; if there are non-compactly many equiva-
lent observers it becomes dicult to reconcile their views, as is in the special
and general relativity, for example.)
Now we glance at a couple of H-structures for interesting non-compact sub-
groups H GL
n
(R) (where the compact case of H = O(n) corresponds to the
Riemannian geometry).
C-structures. These, customarily called almost complex structures J on
V , are elds of C-linear structures J
v
in T
v
(V ), v V . Such a J may be
expressed by an anti-involution, also called J : T(V ) T(V ) (corresponding
to

1), where the pertinent H is (non-compact!) GL
m
(C) GL
n
(R) for
n = 2m. Morphisms, called C-maps f : V
1
V
2
, where the dierential Df :
T(V
1
) T(V
2
) is C-linear (i.e. commute with J), are rare for (non-integrable)
V
1
and V since the corresponding (elliptic) P.D.E.-system is overdetermined for
dim
R
V
1
> 2. So we stick to C-curves, maps of Riemann surfaces S V =
21
(V, J), also called J-curves if this matters. We mark each S with a point s S
and then the totality of C-curves makes a huge space o = o(V, J) foliated into
surfaces, where each leaf in o is represented by a xed S V with variable
marking s S.
What is the (possible) global geometry (e.g. dynamics) of o and how can it
be read from J ?
Start with the subfoliation ( o of closed leaves in o corresponding to
closed (Riemann) surfaces S V , and try to mimic the geometry of curves in
complex algebraic (rst of all, projective) varieties V . This ( is ltered by the
degrees d of S V (playing the role of periods of closed orbits, say for actions
of Z on some space) where curves of degree d may degenerate to (several) curves
of lower degree thus compactifying each (moduli) space (
d
by low dimensional
strata built of (
i
with i < d. Conversely, one can often fuse lower degree curves
to higher degree d by deforming their (reducible) unions to irreducible S (
d
.
What happens to this web of algebraic curves of degree d when we slightly
perturbe the underlying J
0
?
The answer depends on the virtual dimension of the space (
d
, i.e. the Fred-
holm index of the elliptic system dening S (
d
. For example, the curves in
abelian varieties C
m
/ are unstable under (even integrable) deformations of J
0
,
but the curves in certain V s (with suciently ample anti-canonical bundles)
such as CP
m
, remain essentially intact under small (and large as we shall see
below) deformations of J
0
(yet the shape and position of (
d
in o may be, a pri-
ori, greatly distorted by -deformations of J
0
for d large compared to
1
). As
we follow a deformation J
t
moving it further away from the original J
0
, a curve
S
0
V , persistent for small t, may eventually perish by blowing up to some-
thing non-compact as t reaches some critical value t
c
. What is needed to keep S
alive (as a closed J
t
-curve) is an a priori bound on the area of S (V, J = J
t
)
(measured with some background Riemannian metric g in V , where a specic
choice of g is not important as we deal here with compact V ). Such a bound
is guaranteed by the following tameness assumption of J which limits the area
of (even approximately) J-holomorphic curves S by their topology, namely by
[S] H
2
(V ).
Call J tamed by a closed 2-form on V , if is positive on all J-curves
in V (i.e. (, J) > 0 for all non-zero vectors T(V )). If so, J is tame
with respect to the cohomology class h = [] H
2
(V ): the area of each closed
oriented approximately J-holomorphic curve S V is bounded in terms of
the value h[S] R. To make it precise, denote by S

S the set of points

s S V , where the plane T
s
(S) T
S
(V ) is -close to a C-line (for a xed
background metric). Then tame means the existence of , , C > 0, such
that the inequality area S

(1 ) area S implies h(S) C

1
area S
for all closed oriented surfaces S V . Clearly, -tame h-tame but
the converse (h . . .) remains questionable. (One may try the Hahn-Banach
theorem, especially for dimV = 4.)
22
If (V, J
0
) is a Kahler (e.g. algebraic) manifold, then J
0
is tamed by the
(symplectic) Kahler form and as far as J
t
remains [w]-tame we have decent
moduli spaces of J
t
-holomorphic curves in (V, J

) (where J = J
t
may be quite
far from J
0
, e.g. J = AJ
0
, for an arbitrary symplectic automorphism T(V )
T(V )). For example, if we start with the standard (CP
m
, J
0
) and J = J
1
is joined with J
0
by a homotopy of [
0
]-tame structure J
t
(for the standard
symplectic 2-form
0
on CP
m
), then (CP
m
, J
1
) admits a rational J
1
-curve
S of degree 1 through each pair of points, where, moreover, S is unique for
m = 2. (This remains true for all -tame structures on CP
2
, with no a priori
assumption =
0
, by the work of Taubs and Donaldson.) But what happens
to closed C-curves at the rst moment t
c
when J
t
loses tameness? What kind
of subfoliation o
d
o is formed by the limits of S (
d
for t t
c
? It seems,
at least for dimV = 4 (e.g. for CP
2
and S
2
S
2
), that most of the closed
C-curves blow up simultaneously forming a regular (foliated-like) structure in
V . (This is reminiscent of how Kleinian groups degenerate remaining discrete
and beautiful at the verge of distinction.)
Are there non-tame (V, J) with rich moduli spaces of closed (especially ratio-
nal) curves, say having such a curve passing through each pair of points in V ?
(If a 4-dimensional (V, J) has many J-curves, it is tame by an easy argument.
On the other hand the majority of higher dimensional (V, J) contain isolated
pockets of J-curves with rather shapeless and useless C
d
like closed geodesics
in (most) Riemannian manifolds lost in accidental wells of energy.)
Turning to non-closed C-curves we nd a prerequisite for the Nevanlinna
theory as they share (the principle symbol of) with ordinary holomorphic
functions and maps. (This is also crucial for the study of closed C-curves.) For
example, we can dene hyperbolic (V, J) which receive no non-constant J-maps
C V and these V (we assume compactness) carry a non-degenerate Kobayashi
metric, i.e. the supremal metric for which the C-maps H
2
V are 1-Lipschitz.
This hyperbolicity has a point in common with tameness: the space of C-maps
f : S
2
V with f ranging in a compact set is compact for hyperbolic V . (This
is also implied by the tame bound area S const([S]) for approximately J-
holomorphic spheres in V , provided there is no J-holomorphic spheres in V .)
Consequently, for each C-structure on W = V S
2
compatible with J on the
bers V s, s S
2
, there is a rational (i.e. spherical) C-curve in W passing
through a given point w W that contractibly projects to W. (This remains
valid for irrational curves S if the Teichmuller space of S is incorporated into
W.) Another link between tame and hyperbolic is expressed by the follow-
ing (easy to prove) topological criterion for hyperbolicity. Let

V be a Galois
covering of V and

be a 1-form with sublinear growth (|

( v)|/ dist( v, v
0
) 0
for v ), and with invariant (under the deck transformation group) dieren-
tial w = d

. If J is tamed by the corresponding class [w] H

2
(V ), then (V, J)
is hyperbolic. (For example, if J is tame,
2
(V ) = 0 and
2
(V ) is hyperbolic,
then (V, J) is hyperbolic.)
Are there further topological criteria for (non)-hyperbolicity (where
1
(V )
23
is not so large)? For example, are there -tame hyperbolic structures on the
4-torus? (Parabolic curves, i.e. C-maps into such torus (T
4
, w, J) could help us
to study , where the ultimate goal is to show that every symplectic structure
on T
4
is isomorphic to the standard one as is known for CP
2
, for instance by
the work of Taubs and Donaldson.)
What are essential metric properties of (V, dist
Kob
) for hyperbolic V ? Noth-
ing is known about it. How much can one deform a hyperbolic J keeping hyper-
bolicity? It is clear (by Brodys argument) that small perturbations do not hurt.
Take, for instance, a (necessarily singular) J-curve S
0
in an -tame (CP
2
, J),
where the fundamental group of the complement is hyperbolic (e.g. free non-
cyclic which is the case for S
0
consisting of d+1 rational curves of degree 1 where
exactly d of them meet at a single point and d 3). Then every non-constant
J-map C (CP
2
, J) meets S
0
and if we slightly deform S
0
, the complement of
the resulting complement CP
2
S

remains hyperbolic, provided each irreducible

component of S
0
minus the remaining components were hyperbolic (as happens
if the above d + 1 lines are augmented by another one meeting each at a single
point). Thus, for every d 5, there is a non-empty open subset H
d
C
d
of
smooth J-curves S in (CP
2
, J) of degree d where every non-constant J-map
C CP
2
meets S. Probably, this H
d
is dense as well as open for large d pos-
sibly, depending on J, (where the case of curves with many, depending on J,
irreducible components seems within reach). Similar observation can be made
for other compact V and dierences of these, e.g. for those associated to tori,
but nowhere one comes close to what is known in the classical algebraic case.
Here is another kind of (test) question with no classical counterpart: Given a
hyperbolic (V, J
0
), when and how can one modify J
0
inside a (arbitrarily) small
neighbourhood U V , such that the resulting (V, J
1
) admits a parabolic (or even
rational) curve through each point v V ?
If (V, J) is hyperbolic then the space H of J-maps H
2
V is compact and
projects onto o with circle bers. The group G = PSL
2
R = Iso H
2
natu-
rally acts on H and periodic (i.e. compact) orbits correspond to closed J-curves
S V , where genus (S) const(Area S). If (V, J) is algebraic, then periodic or-
bits are dense in H; besides, there are many nite dimensional invariant subsets
in H corresponding to (solutions of) algebraic O.D.E. over V . Are nite dimen-
sional invariant subsets dense in H for all (tame) V ? When does V contain a
metrically complete J-curve S V , preferably with locally bounded geometry?
What are (if any) invariant measures in H? Can one extract symplectic invari-
ants of out of H for -tame J in the absence of closed J-curves? Take, for
example an algebraic (V, J
0
) with a nice foliation, e.g. a at connection over a
Riemann surface or the standard foliation on a Hilbert modular surface. How
does such a foliation fare under tame homotopies J
t
?
If (V, J) is non-hyperbolic then the main issue is to understand the space
T o of parabolic leaves, or equivalently, the space T

of non-constant J-maps
C V with the action of G = A C. One knows, for the standard CP
n
, that
rational maps are dense in T

(and hence in H) by the Runge theorem. This,

24
probably, is true for all tame (V, J) with suciently many rational curves
(but this seems unknown even for rationally connected algebraic V ). Let, for
instance, V = (CP
n
, J) where J is tamed by the standard . Then, by fusing (a
sequence of) rational curves, one can obtain a parabolic one containing a given
countable subset in V . Most likely, one can prove the Runge theorem(s) for V
in this way following Donaldsons approach to Runge for Yang-Mills.
Denote by B T

the space of non-constant 1-Lipschitz C-maps C V

and observe that the closure of every G-orbit in T

meets B by the Bloch-

Brody principle. If V contains (many) rational curves, then B is quite large,
e.g. if V = (CP
n
, J
stand
), then there is a map f B interpolating from an
arbitrary -separated subset C; probably, this remains true for all J tamely
homotopic to the standard one as well as for more general rationally connected
V . Conversely, (a strengthened version of) this interpolation property is likely
to imply the existence of (many) rational curves in V .
Apart from rational curves, one can sometimes produce C-maps C V
by prescribing some asymptotic boundary conditions, e.g. by proving non-
compactness of the space of J-disks with boundaries on a non-compact family of
Lagrangian subvarieties. This works for many J on the standard symplectic R
2m
and yields, for instance, parabolic curves for J on tori tamed by the standard
. How large is B for these tori ? Hopefully, the mean (ordinary?) dimension
of B (for the action of C G on B) is nite and the natural map from B to the
projectivized H
2
(V ) is non-ambiguous and somehow represents the (homology)
class of CP
m1
corresponding to C-lines in H
1
(V ) = C
m
= R
2m
.
Pseudo-Riemannian manifolds. Given an H-structure g on a manifold
V for a non-compact group H GL
n
(R), n = dimV , one may rigidify (V, g)
by reducing H to a maximal compact subgroup K H, i.e. by considering
a Riemannian metric g
K
on V compatible with g. For example, if g(= J) is
a C-structure, then K = U(m) GL
m
(C), m =
n
2
, and g
K
is an Hermitian
metric; if g is pseudo-Riemannian, i.e. a quadratic dierential form on V of
type (p, q) with p + q = n, then K = O(p) O(q) O(p, q) and g
K
is a Rie-
mannian metric such that there is a frame
+
1
, . . . ,
+
p
,

1
, . . . ,

q
at each point,
where both forms g and g
K
become diagonal with g
K
(
+
i
,
+
i
) = g(
+
i
,
+
i
) and
g
K
(

j
,

j
) = g(

j
,

j
). What properties (invariants) of g can be seen in an
individual g
K
and/or in the totality (
K
= (
K
(g) of all g
K
? It may happen that
two structures g and g

are virtually indistinguishable in these terms, namely

when for each > 0, there exist g
K
and g

K
which are (1+)-bi-Lipschitz equiv-
alent. (The Di-orbit of (
K
might be C
0
-dense in the space of all Riemannian
metrics under the worst scenario.) There are few known cases where (
K
tells
you something useful about g. An exceptionally pleasant example is given by
conformal structures g with (
K
(g) telling you everything about g (i.e. the C
0
-
closures of the Di-orbits of (
K
and (

K
are essentially disjoint unless g and g

are isomorphic; furthermore the action of Di V on the space of conformal struc-

tures is proper apart from the standard conformal S
n
). Next, if H preserves an
exterior r-form on R
n
, and the corresponding form = (g) on V is closed,
25
then g
K
is bounded from below by volumes of r-cycles C V with (C) ,= 0
and, at least, Di (
K
is not dense. Finally, for certain symplectic structures
g = , the minimal g
K
-areas of some 2-cycles C (realizable by C-curves) equal
(C) for all g
K
, thus limiting geometry of (V, g
K
) in a more signicant way.
Having failed with (robust) Lipschitz geometry of g
K
one resorts to curvature
and looks at the subsets (
K
() (
K
of adapted metrics g
K
= g
K
(g) with
some (norm of) curvature bounded by R. Here the mere fact of (
K
()
being empty for a given , gives one a non-trivial complexity bound on g. For
example, one may study the inmal R
+
, such that g admits g
K
with the
sectional curvatures between and (with normalized volume if H , SL
n
R).
And again one is tempted to search for the best g
K
adapted to g with non-
zero expectation for mildly non-compact H inspired by conformal metrics of
constant scalar curvature delivered by the Schoen-Aubin solution to the Yamabe
problem.
Now we turn to the case at hand, pseudo-Riemannian g of type (p, q) on a
(typically compact) manifold V having closer kinship to Riemannian geometry
than general g. What are (most general) morphisms in the pseudo-Riemannian
category comparable to Lipschitz for Riemannian manifolds?
Since we can compare metrics on a xed manifold by g g

for g g

being positive semidenite, we may speak of (+)-long maps f : V V

where
f

(g

) g. However, unlike the Riemannian case where short (rather than

long) maps are useful for all V and V

, this makes sense only for (p, q) not

being too small compared to (p

, q

): if p

> p then every isometric immersion

f
0
: V V

can be a little C
0
-perturbed to some f with f

(g

) being as large as
you want (actually equal a given g
1
> f

0
(g

) homotopic to g by the Nash-Kuiper

argument), and if also q

> q, everything becomes soft and one gets all metric

on V (homotopic to g) by C
1
-immersions arbitrarily C
0
-close to f
0
. Thus we
stick to p = p

and start with positive slices in V , i.e. immersed p-dimensional

W V with g[W > 0. If W is connected with non-empty boundary, we set
R(W, w) = dist(w, W) for the induced Riemannian metric (sup
w
R(w, W) is
called in-radius of W), and dene R
+
(V, v) as the supremum of R(v, W) for
all positive slices through v. (Connected Riemannian manifolds with non-empty
boundaries have R
+
= R < at all their points while closed manifolds have
R
+
< if and only if
1
< . But even for
1
= e there is no eective bound
on R
+
due to non-decidability of
1
?
= e. Moreover, there are rather small
metrics on S
3
with almost negative curvature and thus with arbitrarily large
R
+
). Next, we dualize by taking (+)-long maps f from (V, v) to Riemannian
(W, w) of dimension p and denote by R
+
(V, v) the inmum of R
+
(W, w) over
all possible (W, w) and f, where, clearly, R
+
R
+
.
A (+)-long map V W is necessarily a negative submersion, i.e. with g-
negative bers. Conversely, given a negative submersion f (proper, if V is non-
compact) of V to a smooth manifold W, one can nd a Riemannian metric on W
making f long. (There is a unique supremal Finsler metric on W making f long,
26
which can be minorized by a Riemannian one.) Thus, for example, if a compact
(possibly with a boundary) V admits a negative submersion into a connected
manifold with nite
1
(or with non-empty boundary), then R
+
(V ) < .
Manifolds with R
+
(V, v) const < for all v V are kind of hyperbolic
in the (+)-directions (e.g. this condition is C
0
-stable). If p = 1 and V is closed,
then it always has a circular positive slice and R
+
(V, v) = for all v V , but
if p 2, then every g = g
0
admits a deformation g
t
with R
+
0 for t :
take a generic (and thus non-integrable) p-plain eld S
+
T(V ) with g[S > 0
and make g
t
= g+tg

where g

is the negative part of g for the normal splitting

T(V ) = S
+
S

+
. (This works even locally and shows that the majority of g
have R
+
small in some regions of V and innite in other regions.)
It seems by far more restrictive to require that both R
+
and R

(i.e. R
+
for V ) are bounded on V , where a sucient condition for compact V is the
existence of a negative submersion V W
+
as well as a positive submersion
V W

, where W
+
and W

have
1
< or non-empty boundaries. This can
be slightly generalized by allowing somewhat more general pairs of -foliations
with uniformly compact bers (e.g. coming from submersion to simply con-
nected or to bounded orbifolds) and sometimes one submersion suces. For
example, start with a negative submersion f : V W
+
where W
+
is simply
connected (or bounded) and the (negative) bers of f are also simply connected
(or bounded). Deform the original g on V to g
1
agreeing with g on the bers
while being very positive normally to the bers. Then all negative slices in
(V, g
1
) keep C
1
-close to the bers of f and so R

(V, g
1
) is bounded as well as
R
+
. (Probably, there are more sophisticated, say closed manifolds V , where the
bound on R

(V, v) comes from dierent sources at dierent v V .)

Despite some hyperbolic features, bounded pseudo-Riemannian metrics
on closed manifolds are reminiscent of positive curvature, e.g. they are accom-
panied in known examples by closed positive or negative slices with
1
<
(are these inevitable?) and seem hard to make on aspherical V .
Besides taming g by foliations, one may try (pairs of) dierential forms on
V , where a closed p-form is said to (strictly) (+)-tame g if it is (strictly) non-
singular on the positive slices. For example, if V is metrically split, V = V
+
V

,
the pull-back
+
of the volume form of V
+
strictly (+)-tames g; similarly

strictly ()-tames g, while

+
+

is taming, albeit non-strictly. How much

do closed positive slices persist under (strictly) (+)-tame homotopies of g? In
particular, what happens to area maximizing closed positive 2-slices in (2, q)-
manifolds under strictly (+)-tame homotopies?
The tangent bundle to the space o

of negative slices in V carries a natural

(positive!) L
r
-norm (we use r as p is occupied) since g is positive normally to
S V for all S o

(as well as for all S

p

,q
V with 0 p

< p). But

the associated (path) metric in o

may, a priori, degenerate and even become

everywhere zero. Yet, there are some positive signs.
Let V metrically split V = [0, 1] V

for V

closed. Then the L

1
-distance
27
between 0V

and 1V

equals the volume of V and hence > 0. Consequently

all L
r
-distances are > 0.
Let V = V
+
V

with closed V

and arbitrary Riemannian V

+
. Then
dist
L
(v
+
V

, v

+
V

) = d = dist
V+
(v
+
, v

+
), since the projection V V

together with a 1-Lipschitz retraction

V
+
[0, d] = [v
+
, v

+
]
give us a (+)-short (i.e. ()-long) map V [0, d] V

and the above applies

(only to L

not necessarily to other L

r
).
Let V ber over a compact p-manifold with closed negative bers. Then
the L

-distance is > 0 between every two distinct bers, essentially by the

same argument (also yielding positivity of L

-distances between general closed

negative slices isotopic to the bers).
How much does the metric geometry of (o

, dist
L
) (and of o
+
) capture
the structure of g? How degenerate dist
L
can be for general (V, g)? (This
dist
L
is vaguely similar to Hofers metric in the space of Lagrangian slices in
symplectic V , which also suggests vanishing of our dist
Lr
, r < , for most V .)
As we mentioned earlier, a general complication in the study of H-structures
g with non-compact H GL
n
R is a possible non-stability (or recurrency) of
g due to certain unboundedness of the set of dieomorphisms of V moving g
(or a small perturbation of g) close to g. The simplest manifestation of that is
non-compactness of the automorphisms (isometry) group of (V, g) which may
have dierent nature for dierent structures g. For example, non-compactness
of the conformal transformations f of S
n
is seen in the graphs
f
S
n
S
n
as degeneration of these to unions of two bers (s
1
S
n
) (S
n
s
2
) with an
uniform bound on Vol
f
for f . On the other hand, graphs of isometrics
f of (p, q)-manifolds V are represented by (totally geodesic) isotropic (where
the metric vanishes) n-manifolds
f
V V and their volumes (as well as
in-radii, both measured with respect to some background Riemannian metric in
V V ) go to innity for f , while their local geometry remains bounded
(unlike the conformal case). With this in mind, we call g stable, if it admits a
C
0
-neighbourhood | in the space ( of all gs, such that the graphs of isometrics
f : (V, g

) (V, g

) with g

, g

| have Vol
n
(
f
) const < (where the
background metric is not essential as we assume V compact).
Example. Start with a metrically split V = V
+
V

. The isometrics here

are essentially the same as those of the Riemannian manifold V
+
= V
+
(V

),
since V (V ) = V
+
V

(V
+
) (V

) = V
+
(V

) (V
+
) V

=
V
+
(V
+
), and so V (V ) and V
+
(V
+
) share the same isotropic
submanifolds. If V is closed simply connected (or f does not mix up
1
(V
+
)
and
1
(V

) too much) one sees, by looking at (local) isometries of V

+
, that all
isotropic submanifolds in V (V ) have bounded in-radii (as well as volumes,
if they are closed) and so Iso(V, g) is compact. One sees equally well that
28
g = g
+
g

is stable, and, moreover, one gets a good control over the stability
domain | of g. Namely, take > 0 and let |

= |

(g) consist of those g

, where
the bers V
+
v

and v
+
V

are g

-positive and g

-negative correspondingly
and the projections of these bers to (V
+
, g
+
) and (V

, g

) are -bi-Lipschitz
for g

-restricted to these bers. Then (at least for

1
(V ) = 0) the graphs of
dieomorphism f : V V with |

f(|

) ,= are uniformly bi-Lipschitz to

V
+
, and so all g

<
|

are stable.
What are most general stable g? Are simply connected manifolds of type
(1, q) always stable? (Of course, generic g are stable, but we are concerned
with exceptional (V, g), e.g. with non-compact group Iso(V, g).)
The above motivates the idea of iso-stability for V of type (p, q) with p = q
(e.g. V = V
0
V
0
) limiting the size of -isotropic submanifolds in V . This is
enhanced in the presence of 0-taming p-forms on V , which do not vanish on the
isotropic p-planes in T(V ). For example, if V (V ) is 0-tame (with deg = n
on this occasion), then V is stable with respect to the dieomorphisms with
the graphs homologous to the diagonal in V V , as is the case for the above
g |

and for (V, g) tamed by -foliations with p and q-volume preserving (or
just uniformly bounded) holonomies.
Remarks and references.
(a) Closed C-curves in tame manifolds exhibit a well organized structure
with intricate interaction between moduli spaces (
d
for dierent d and regular
asymptotics for d : quantum multiplication, mirror symmetry etc (see
[McD-Sal]). This also applies to non-closed curves with prescribed Fredholm
boundary (or asymptotic) condition, e.g. J-maps of Riemann surfaces with
boundaries, (S, S) (V, W) for a given totally real W V , where everything
goes as in the boundary free case (including Kobayashi metric, Bloch-Brody etc).
Less obvious conditions come up in the study of xed points of Hamiltonian
transformations and related problems: Floer homology, A-categories, contact
homology of Eliashberg and Hofer. But it is unclear (only to me?) what is the
most general Fredholm condition in the C-geometry.
(b) The questions concerning unbounded C-curves, which parallel (Nevan-
linna kind) complex analysis rather than algebraic geometry, remain as widely
open as when I collected them for (then expected) continuation of [Gro
PCMD
].
Do the spaces o, H, T, B possess geometric structure comparable to (and com-
patible with) what we see in ( =U
d
(
d
? What is the right language to describe
such a structure (if it exists at all)?
Even in the classical case of algebraic V boasting of lots and lots of deep
dicult theorems, there is no hint of the global picture in sight, not even a
conjectural one (see [McQ] for the latest in the eld).
(c) Hyperbolicity of (V, J) can be sometimes derived from negativity of a
suitable curvature of a (Riemann or Finsler) metric adapted to J, either on
29
V itself or on some jet space of C-curves in V . A most general semi-local
hyperbolicity criterion is expressed by the linear isoperimetric inequality in C-
curves S V . If such an inequality holds true on relatively small J-surfaces
S V , then it propagates to all S in the same way as the real hyperbolicity does
(see [Gro
HG
]) which makes it, in principle, veriable for compact V . (The linear
inequality seems to follow from hyperbolicity by a rather routine argument, but
I failed to carry it through. Possibly, one should limit oneself to closed V and
integrability of J may be also helpful.)
It would be amusing to nd a suciently general (positive) curvature condi-
tion for the existence of many rational curves in (V, J), encompassing complex
hypersurfaces V with deg V (much) smaller than dimV , for example, and al-
lowing singular spaces in the spirit of Alexandrovs K 0. Conversely, in the
presence of many closed (especially rational) curves, one expects extra local
structures on V , e.g. taming forms (which easily come from closed curves for
dimV = 4, see [Gro
PCMD
]).
(d) One can sometimes make foliations (or at least, laminations) out of
parabolic curves in V as is done in [Ban] for J on tori tamed by the standard
symplectic (whereas the original question aimed at eventually proving that
is standard).
(e) How much do we gain in global understanding of a compact (V, J) by
assuming that the structure J is integrable (i.e. complex)? It seems nothing
at all: there is no single result concerning all compact complex manifolds. (If
dimV = 4, then the Kodaira classication tells us quite a bit, say for even
b
1
2, especially if there are 4 elements in H
1
(V ) with non-zero product
yielding a nite morphism of V onto C
2
/.) This suggests the presence of (un-
reachable?) pockets of (moduli spaces of) integrable Js with weird properties
(like those produced by Taubs on 6-manifolds); but there is no general exis-
tence theorem for complex structures either (not even for open V s, compare
p. 103 in [Gro]
PDR
) and even worse, no systematic way to produce them. So
far, COMPACT COMPLEX MANIFOLDS have not stood to their fame.

(f) C-curves, dened by restricting their tangent planes to the subvariety

E
0
= CP
m
Gr
2
R
2m
, owe their beauty and power to the ellipticity of E
0
:
there is a single plane e E
0
through each line in R
2m
. One can deform E
0
by keeping this condition thus arriving at generalized C-structures where the
resulting E-curves are similar to C-curves, and where the picture is the clearest
for m = 2 (see [Gro
PCMD
] and references therein). In general, a eld E of subsets
E
v
Gr
k
T
v
(V ) = Gr
k
R
n
, n = dimV , denes a class of k-dimensional E-
subvarieties in V , said directed by E (e.g. W R
n
with Gauss image in E
0
),
which seem most intriguing under ellipticity assumptions on E. To formulate
these, let F
0
= F
0
(E
0
), E
0
Gr
k
R
n
, denote the space of pairs (e, h) for e E
0
and all hyperplanes h e, and look at the tautological map
0
: F
0
Gr
k1
R
n
.
Ideally, F
0
is a smooth closed manifold and is stably one-to-one and onto (i.e.

Fedia Bogomolov suggested to look at manifolds appearing as spaces of leaves of foliations

in pseudoconvex bounded domains in C
N
with algebraic tangent bundles.
30
remains such under small perturbations of E
0
) as in the C-case. Such E
0
do
not come cheaply: all known instances of them appear as (deformations of) Lie
group orbits in Gr
k
R
n
, e.g. in Harvey-Lawson calibrated geometries (where
a most tantalizing E
0
Gr
3
R
7
is associated to rational J-curves in S
6
for
the standard G
2
-invariant C-structure J on S
6
). One gains more examples by
dropping onto thus arriving at overdetermined elliptic systems (isotropic in
pseudo-Riemannian, for instance) which need integrability in order to have
solutions (as special Lagrangian of Harvey-Lawson). Also, one may allow
E
0
and
0
to have some singularities (similar to those present in Yang-Mills in
a dierent setting), but in all cases one is stuck with two problems: what are
possible elliptic E
0
, and what are (global and local) analytic properties (especially
singularities) of the corresponding E-subvarieties W in (V, E = E
v
)? (If E is
overdetermined, one looks at Ws directed by a small neighbourhood E

E.)
(g) The radii R

are less useful than R

as they make sense only for rather

special pseudo-Riemannian manifolds V = (V, g). Yet R

can be used for

characterization of such V , e.g. the equality R

= R

< seems to (almost?)

distinguish metrically split manifolds. Also one can generalize R

by considering
submersions V W

being long normally to the bers, where 1 dimW

+

p and 1 dimW

q and where g is denite on the bundle of vectors normal

to the bers. Then the resulting radii R

(V ) satisfy R

.
(h) The niteness of R

(V ) does not ensure stability of V for dimV 4

as simple (e.g. split) examples show but this seems to limit instability to
codimension two. Can one go further with stronger radii-type invariants?
(i) Besides the in-radius, there are other Riemannian invariants to gauge
pseudo-Riemannian metrics such as the macroscopic dimension (see [Gro
PCMD
])
of (complete) positive slices in V , or the maximal radius of a Euclidean p-ball
in V . (The Euclidean metric on R
p
dominates other g : there is a long map
R
p
(R
p
g) for every Riemannian metric g on R
p
. But one can go beyond R
p
by
admitting slices with non-trivial topology. Some V may contain lots of these, e.g.
some V of large dimension support (p, q)-metrics g so that every Riemannian
p-manifold admits an isometric immersion into (V, g).) Furthermore, one may
look at homotopies and extendability of slices with controlled size thus getting
extra invariants of V . Actually, the mere topology of the space of, say, closed
positive slices can be immensely complicated encouraging us to seek conditions
limiting this complexity (e.g. in the spirit of diagram groups, see [Gu-Sa]).
(j) If V is compact, one may distinguish complete slices for the metric induced
from some Riemannian background h in V and then compare these induced
Riemannian metrics with those induced from our g of type (p, q). (Besides
completeness, h brings forth other classes of slices, e.g. those with some bounds
on curvature.)
(k) The isometrics of (non-stable!) (V, g) have attractive geometry and dy-
namics (see [DAmb-Gr] for an introduction and references) with many elemen-
tary questions remaining open, e.g. does every isometry of the interior of V
extend to the boundary? (This comes from relativity, I guess.)
31
(l) Most of the current pseudo-Riemannian research is linked to general
relativity focused on the Einstein equation (see[Be-Eh-Ea]).
V. Symbolization and randomization.
A common way to generate questions (not only) in geometry is to confront prop-
erties of objects specic to dierent categories: what is a possible topology (e.g.
homology) of a manifold with a given type of curvature? How is the dynamics
of the geodesic ow correlated with topology and/or geometry? How fast a har-
monic function can decay on a complete manifold with a certain asymptotic ge-
ometry, e.g. curvature? How many critical points a geometrically dened energy
may have on given space maps or subvarieties? What are possible singularities of
the exponential map or of the cut locus? Does every (almost complex) manifold
(of dim 6) support a complex structure? etc. These seduce us by simplicity
and apparent naturality, sometimes leading to new ideas and structures (tan-
gentially related to the original questions, as the Morse-Lusternik-Schnirelmann
theory motivated by closed geodesics), but often the mirage of naturality lures
us into featureless desert with no clear perspective where the solution, even if
found, does not quench our thirst for structural mathematics. (Examples are
left to the reader.)
Another approach consists in interbreeding (rather than intersecting) cate-
gories and ideas. This has a better chance for a successful outcome with ques-
tions following (rather than preceding) construction of new objects. Just look at
how it works: symbolic dynamics, algebraic arithmetic and non-commutative ge-
ometry, quantum computers, dierential topology, random graphs, p-adic anal-
ysis ... Now we want to continue with symbolic geometry and random groups.
Given a category of spaces X with nite Cartesian products, we consider
formal innite products A =
iI
X
i
, where the index set has an additional
(discrete) structure, e.g. being a graph or a discrete group . In the latter
case we assume that all X
i
are the same, A = X

consists of functions :
X, and naturally acts on A. Nothing happens unless we start looking at
morphisms : A } over a xed . Such a is given by a nite subset
of cardinality d and a map : X

= X
d
Y , i.e. a function y = (x
1
, . . . , x
d
),
where ()() is dened as the value of on the restriction of to the -
translate for all . Thus we enrich the original category by making
single variable (-equivariant) functions () out of functions (x
1
, . . . , x
d
) in
several variables.
Take a particular category of Xs, e.g. algebraic varieties, smooth symplectic
or Riemannian manifolds, (smooth) dynamical systems, whatever you like, and
start translating basic constructions, notions and questions into the symbolic
language of As. This is pursued in [Gro
ESA
] and [Gro
TID
] with an eye on
continuous counterparts to A, e.g. spaces of holomorphic maps C X for
algebraic varieties X with a hope to make algebraic somehow reected in such
32
spaces. I have not gone far: a symbolic version of the Ax mapping theorem for
amenable (similar to the Garden of Eden in cellular automata) and a notion
of mean dimension dened for all compact -spaces A with amenable (in the
spirit of the topological entropy) recapturing dim
top
X for A = X

(applicable
to spaces like B of J-maps C (X, J) for instance). Thats about it. (The
reader is most welcome to these A; if anything, there is no lack of open questions;
yet no guarantee they would lead to a new grand theory either.)
Randomization. Random lies at the very source of manifolds, at least in
the smooth and the algebraic categories: general smooth manifolds V appear
as pull-back of special submanifolds under generic (or random) smooth maps f
between standard manifolds, e.g. zeros of generic functions f : R
N
R
Nn
or
(proper) generic maps from R
N
to the canonical vector bundles W over Grass-
mann manifolds Gr
Nn
R
M
(Thom construction), where V come as f
1
(0) for
the zero section 0 = Gr
Nn
R
M
W. (Other constructions in dierential topol-
ogy amount to little tinkering with V s created by genericity. Similarly, the bulk
of algebraic manifolds comes from intersecting ample generic hypersurfaces in
standard manifolds, e.g., in CP
N
, and the full list of known constructions of,
say non-singular, algebraic varieties is dismally short.)
One may object by pointing out that every (combinatorial) manifold can be
assembled out of simplices. Indeed, it is easy to make polyhedra, but no way
to recognize manifolds among them (as eventually follows from undecidability
of triviality for nitely presented groups). Here is another basic problem linked
to non-locality of topology. How many triangulations a given space X (e.g.
a smooth manifold, say the sphere S
n
) may have? Namely, let t(X, N) denote
the number of mutually combinatorially non-isomorphic triangulations of X
into N simplices. Does this t grow at most exponentially in N?, i.e. whether
t(X, N) exp C
X
N. Notice that the number of all X built of N n-dimensional
simplices grows super-exponentially, roughly as n
n
, and the major diculty for
a given X comes from
1
(X) and, possibly (but less likely), from H
1
(X) , where
the issue is to count the number of triangulated manifolds X with a xed
1
(X)
or H
1
(X).
These questions (coming from physists working on quantization of gravity)
have an (essentially equivalent) combinatorial counterpart (we stumbled upon
with Alex Nabutovski): evaluate the number t
L
(N) of connected 3-valent (i.e.
degree 3) graphs X with N edges, such that cycles of length L normally
generate
1
(X) (or, at least, generate H
1
(X))? Is t
L
(N) at most exponential in
N for a xed (say = 10
10
) L? The questions look just great and no idea how
to answer them.
Here is a somewhat similar but easier question: what does a random group
(rather than a space) look like? As we shall see the answer is most satisfactory
(at least for me): nothing like we have ever seen before. (No big surprise
though: typical objects are usually atypical.)
33
Random presentation of groups. Given a group F, e.g. the free group
F
k
with k generators, one may speak of random quotient groups G = F/[R],
where R F is a random subset with respect to some probability measure on
2
G
and [R] standing for the normal subgroup generated by R. The simplest way
to make a is to choose weights p() [0, 1] for all and take the product
measure in 2
G
, i.e. R is obtained by independent choices of F with
probabilities p(). This is still too general; we specialize to p() = p([[) where
[[ denotes the word length of for a given, say nite, system of generators in
F. A pretty such p is p = p

() = (card

F [ [

[ = [[)

, > 0. If the
temperature is close to zero, p
0
() decays slowly and random R is so large
that it normally generates all F making G = e with probability one, provided
F is innite. For example, if F = F
k
, this happens whenever p()
2
(F
k
), i.e.

p
2
() < . This is easy; but it is not so clear if G may be ever non-trivial
for large . However, if F = F
k
(or a general non-elementary word hyperbolic
group), one can show that G is innite with positive probability for >
crit
(F),
and
crit
(F
k
), probably, equals 2, i.e. p

() ,
2
card(G) = with non-zero
probability (see [Gro
AI
] for a slightly dierent p(), where the critical 2 comes
as the Euler characteristic of S
2
via the small cancellation theory).
Random groups G

look very dierent for dierent . It seems that G

1
cannot be embedded (even in a most generous sense of the word) to G
2
for

2
>
1
as the density of random G

decreases with temperature. Fur-

thermore, generic samples of G

for the same (large) are, probably, mutually

non-isomorphic (not even quasi-isometric) with probability one, yet their ele-
mentary invariants are likely to be the same. It is clear for all < , that G

a.s. have no nite factor groups and they may satisfy Kazhdans property T.
(T is more probable for small where it is harder for G

to be innite.)
Let us modify the above probability scheme by considering random homo-
morphism from a xed group H to F with G = F/[(H)]. To be simple, let
H =
1
() for a (directed) graph and be given by random assignment of
generators of F to each edge e in , independently for all edges. Denote by
N(L) the number of (non-oriented) cycles in of length L and observe that
for large N(L) the group G is likely to be trivial. But we care for innite G
and this can be guaranteed with positive probability if N(L) exp L/ for large

crit
(F) for free groups F = F
k
, k 2, and non-elementary hyperbolic
groups F in general. (I have checked this so far under an additional lacunarity
assumption allowing, for example, being the disjoint union of nite graphs
i
of cardinalities d
i
, i = 1, 2, . . . , with d
i+1
exp d
i
and such that the shortest
cycle in
i
has length
cr
log d
i
for all i.)
To have an innite random G = G(F, ) with interesting features, we need
a special . We take , such that it contains arbitrarily large -expanders with
a xed (possibly small) > 0. (Such do exist, in fact random in our
category contain such expanders, see [Lub].) Then random groups G a.s. enjoy
the following properties.
34
(A) G are Kazhdan T, i.e. every ane isometric action of such G on the
Hilbert space R

has a xed point. Furthermore, every isometric action of

G on a (possibly innite dimensional) complete simply connected Riemannian
manifold V with K(V ) 0 has a xed point (if dimV < , then there is no
non-trivial action at all).
(B) For every Lipschitz (for the word metric) map f : G R

, there are
sequences g
i
, g

i
G with dist
G
(g
i
, g

i
) and dist
R
(f(g
i
), f(g

i
)) const <
(and the same remains true for the
p
-spaces for p < ).
One may think the above pathologies are due to the fact that G are not
nitely presented. But one can show that some quasi-random groups among
our G are recursively presented and so embed into a nitely presented group
G

which then automatically satises (B) and can be chosen with an aspherical
presentation by a recent (unpublished) result by Ilia Rips and Mark Sapir (where
T can be preserved by adding extra relations). Then, one can arrange G

1
(V ) for a closed aspherical manifold V of a given dimension n 5 which,
besides (B), has more nasty features, arresting, for example, all known (as far
as I can tell) arguments for proving (strong) Novikovs conjecture for G

. (See
[Gro
RW
]; but I could not rule out hypersphericity yet.) I feel, random groups
altogether may grow up as healthy as random graphs, for example.
There are other possibilities to dene random groups, e.g. by following the
symbolic approach where combinatorial manipulations with nite sets are
replaced by parallel constructions in a geometric category. For example, we may
give some structure (topology, measure, algebraic geometry) to the generating
set B of (future) G with relations being geometric subsets in the Cartesian
powers of B. Then, depending on the structure, one may speak of random
or generic groups G (with a possible return to nitely generated groups via a
model theoretic reasoning). Looks promising, but I have not arrived at a point
of asking questions.
35
Bibliography
[A-B-C-K-T] J. Amoros, M. Burger, K. Corlette, D. Kotschick, D. Toledo,
Fundamental Groups of Compact Kahler Manifolds, Mathematical
Surveys and Monographs, American Mathematical Society, Provi-
dence, RI (1996).
[Ban] V.Bangert, Existence of a complex line in tame almost complex
tori, Duke Math. J. 94 (1998), 2940.
[Berg] M. Berger, Riemannian geometry during the second half of the
twentieth century, Jahrbericht der Deutschen Math. Vereinigung,
v. 100 (1998), 45208.
[Be-Eh-Ea] J.K. Beem, P.E. Ehrlich, K.L. Easley, Global Lorentzian Geometry,
2. ed., Monograph + Textbooks in Pure and Appl. Mathematics
vol. 202, Marcel Dekker Inc., N.Y. 1996.
[Bria] R. Briant, Recent advances in the theory of holonomy, Sem. N.
Bourbaki, vol. 1998-99, juin 1999.
[Che-Co] J. Cheeger, T. Colding, On the structure of spaces with Ricci cur-
vature bounded below, III., preprint.
[DAmb-Gr] G. DAmbra, M. Gromov, Lectures on transformation groups: Ge-
ometry and dynamics, in Surveys in Dierential Geometry 1 (1991),
19111.
[G-L-P] M. Gromov, with Appendices by M. Katz, P. Pansu, and S.
Semmes, Metric Structures for Riemannian and Non-Riemannian
Spaces based on Structures Metriques des Varietes Riemanniennes,
edited by J. LaFontaine and P. Pansu, English Translation by Sean
M. Bates, Birkhauser, Boston - Basel - Berlin (1999).
[Gro
AI
] M. Gromov, Asymptotic invariants of innite groups. Geometric
group theory, Vol. 2, Proc. Symp. Sussex Univ., Brighton, July 14-
19, 1991. London Math. Soc. Lecture Notes 182 Niblo and Roller
ed., Cambridge Univ. Press, Cambridge, 1295 (1993).
[Gro
CC
] M. Gromov, Carnot-Caratheodory spaces seen from within sub-
Riemannian geometry, Proc. Journees nonholonomes; Geometrie
sous-riemannienne, theorie du controle, robotique, Paris, June 30
- July 1, 1992, A. Bellaiche ed., Prog. Math. 144 (1996), 79323,
Birkhauser, Basel.
[Gro
ESA
] M. Gromov, Endomorphisms of symbolic algebraic varieties, J. Eur.
Math. Soc. 1 (1999), 109197.
36
[Gro
HG
] M. Gromov, Hyperbolic groups, in Essays in Group Theory, Mathe-
matical Sciences Research Institute Publications 8 (1978), 75263,
Springer-Verlag.
[Gro
PCMD
] M. Gromov, Positive curvature, macroscopic dimension, spectral
gaps and higher signatures in Functional analysis on the eve of the
21st century, Gindikin, Simon (ed.) et al. Volume II. In honor of the
eightieth birthday of I.M. Gelfand. Proc. Conf. Rutgers Univ., New
Brunswick, NJ, USA, Oct. 24-27, 1993. Prog. Math. 132 (1996),
1213, Birkhauser, Basel.
[Gro
PDR
] M. Gromov, Partial Dierential Relations, Springer-Verlag (1986),
Ergeb. der Math. 3. Folge, Bd. 9.
[Gro
RW
] M. Gromov, Random walk in random groups, in preparation.
[Gro
SGM
] M. Gromov, Sign and geometric meaning of curvature, Rend. Sem.
Math. Fis. Milano 61 (1991), 9123.
[Gro
SH
] M. Gromov, Soft and hard.
[Gro
TID
] M. Gromov, Topological invariants of dynamical systems and spaces
of holomorphic maps, to appear in Journ. of Geometry and Path.
Physics.
[Gu-Sa] V. Guba, M. Sapir, Diagram groups, Memoirs of the AMS, Novem-
ber, 1997.
[Kan] P. Kanerva, Sparse Distributed Memory, Cambridge, Mass. MIT
Press, 1988.
[Klein] F. Klein, Vorlesungen uber die Entwicklung der Mathematik im 19.
Jahrhundert, Teil 1, Berlin, Verlag von Julius Springer, 1926.
[La-Sch] U. Lang, V. Schroeder, Kirszbrauns Theorem and Metric Spaces of
Bounded Curvature, GAFA, Geom. funct. anal. Vol. 7 (1997), pp.
535560.
[Lub] A. Lubotzky, Discrete groups, expanding graphs and invariant mea-
sures, Progress in Mathematics, 125. Birkhauser Verlag, Basel
(1994).
[McD-Sal] D. McDu, D. Salamon, J-holomorphic curves and quantum coho-
mology, Univ. Lect. Series n

6, A.M.S. Providence (1994).

[McQ] M. McQuillan, Holomorphic Curves on Hyperplane Sections of 3-
Folds, Geometric And Functional Analysis, Vol. 9, No. 2, pp. 370
392 (1999).
37
[Mi] V.D. Milman, The heritage of P. Levy in geometrical functional
analysis, in Colloque P. Levy sur les Processus Stochastiques,
Asterisque 157158 (1988), pp. 273-301.
[Newm] J.R. Newman, The world of Mathematics, Vol. 1, Simon and Schus-
ter, N.Y. 1956.
[Per] G. Perelman, Spaces with curvature bounded below, in Proc. ICM,
S. Chatterji, ed., Birkhauser Verlag, Basel, 1995, pp. 517525.
[Pet] P. Peterson, Aspects of global Riemannian geometry, BAMS, 36:3,
pp. 297344 (1999).
[Spr] D. Spring, Convex Integration Theory, Birkhauser Verlag, Basel -
Boston - Berlin, 1998.
[Vasi] A.V. Vasiliev, Nikolai Ivanovich Lobachevski, Moscow, Nauka 1992
(in Russian).
38