Proceedings of the International Congress of Mathematicians

Hyderabad, India, 2010

Dierentiability of Lipschitz functions,
structure of null sets, and other problems
Giovanni Alberti, Marianna Csornyei, David Preiss
Abstract. The research presented here developed from rather mysterious observations,
originally made by the authors independently and in dierent circumstances, that Lebesgue
null sets may have uniquely dened tangent directions that are still seen even if the set
is much enlarged (but still kept Lebesgue null). This phenomenon appeared, for example,
in the rank-one property of derivatives of BV functions and, perhaps in its most striking
form, in attempts to decide whether Rademachers theorem on dierentiability of Lipschitz
functions may be strengthened or not.
We describe the non-dierentiability sets of Lipschitz functions on R
n
and use this
description to explain the development of the ideas and various approaches to the denition
of the tangent elds to null sets. We also indicate connections to other current results,
including results related to the study of structure of sets of small measure, and present
some of the main remaining open problems.
Mathematics Sub ject Classication (2000). Primary 26B05; Secondary 28A75.
Keywords. Lipschitz, derivative, tangent, width, unrectiability.
1 Dierentiability of Lipschitz functions
One of the important results of Lebesgue tells us that Lipschitz functions on the real
line are dierentiable almost everywhere. It is also well-known that the converse is
true: for every Lebesgue null set E on the real line there is a real-valued Lipschitz
function which is non-dierentiable at any point of E. That is:
Theorem 1.1. For a given set E R there is a Lipschitz function f: R R
which is not dierentiable at any point x E if and only if E is Lebesgue null.
One of our aims is to generalise Theorem 1.1, and also its more precise variants
that will be described in Theorem 1.13, to Lipschitz functions f: R
n
R
m
.
Since a Lipschitz function on R is dierentiable almost everywhere, Fubini
Theorem implies immediately that the directional (or partial) derivative
f

(x; u) := lim
t0
f(x +tu) f(x)
t
of a Lipschitz function f: R
n
R
m
exists for each direction u at a.e. x.
2 G. Alberti, M. Csornyei, D. Preiss
Although dierentiability is not the same as the existence of suciently many
partial derivatives, the set of points at which these two notions dier is relatively
easy to control. First recall the following denition:
Denition 1.2. A set E R
n
is porous at a point x E if there is a c > 0 and
there is a sequence y
n
0 such that the balls B(x+y
n
, c|y
n
|) are disjoint from E.
The set E is porous if it is porous at each of its points, and it is called -porous if
it is a countable union of porous sets.
Theorem 1.3 ([3]). Let f: R
n
R
m
be a Lipschitz function. Then the set of
those points at which f is not dierentiable but it is dierentiable in n linearly
independent directions is -porous.
It follows from Lebesgues density theorem that -porous sets have Lebesgue
measure zero. Therefore as an immediate corollary we obtain:
Theorem 1.4 (Rademacher). Every Lipschitz function f: R
n
R
m
is dieren-
tiable almost everywhere.
The converse direction, i.e. the description of those sets E R
n
for which there
is a non-dierentiable Lipschitz function, is much harder. D. Preiss proved that
the converse of Rademachers theorem is false, already in dimension 2:
Theorem 1.5 ([9]). There is a Lebesgue null set E R
2
such that every Lipschitz
function f: R
2
R is dierentiable in at least one point of E.
Unlike in the classical Lebesgue and Rademacher theorem, Preisss result is not
an almost everywhere result, he does not show that the function is dierentiable
at most of the points x E. Indeed this is not possible. We prove the following
theorem:
Theorem 1.6. For every Lebesgue null set E R
2
there is a Lipschitz function
f: R
2
R
2
which is not dierentiable at any point x E.
This theorem says that for every Lebesgue null set there are two real-valued
Lipschitz functions, namely, the coordinate functions of f, such that at each x E
at least one of the two functions are non-dierentiable.
Remark. In [9] the result is proved not only in R
2
, but in every Banach space
with a smooth norm. Preisss set E is dense. In a recent paper [5], M. Dore
and O. Maleva constructed a closed (and hence nowhere dense) null set with the
same property: in every Banach space X with separable dual there exists a closed
bounded set of Hausdor dimension one containing a Frechet-dierentiability point
of every Lipschitz function f: X R.
Let E R
n
. It is immediate from the denition that a set E is porous at
x E if and only if the Lipschitz function f(x) = dist(x, E) is non-dierentiable
at x. Of course -porous sets cannot fully describe non-dierentiability sets of
Lipschitz functions (not even in R, since not all Lebesgue null sets of R are -
porous). But by Theorem 1.3, in order to nd all Lebesgue null sets for which
Dierentiability of Lipschitz functions, structure of null sets 3
there is a non-dierentiable Lipschitz function, it is enough to consider functions
not having enough many directional derivatives.
From the point of view of dierentiability problems, the sets that are the most
negligible are the sets of points at which a Lipschitz function may be dierentiable
in no direction. We show that these sets form a -ideal. We call them uniformly
purely unrectiable. Notice that uniformly purely unrectiable sets are purely
unrectiable, i.e. they are null on every rectiable curve, since a Lipschitz function
is dierentiable in the tangent direction at a.e. point of a curve. We will see later
that uniformly purely unrectiable sets have the (possibly only formally) stronger
property that they can be covered by an open set which is small on many curves
simultaneously.
For simplicity, consider just Lipschitz functions f: R
2
R
m
. We will show
that if f is not dierentiable at the points of E R
2
, then at each point x E
except for a uniformly purely unrectiable set, there is a unique dierentiability
direction (x) of f. Moreover, this direction is determined by the geometry of the
set E, it is independent of the function f: for any other Lipschitz function g, the
direction constructed using f and g agree at each point of E except for a uniformly
purely unrectiable set. Indeed, if E is contained in the non-dierentiability set of
both f: R
2
R
m
1
and g: R
2
R
m
2
, then the direction dened by the function
h = (f, g): R
2
R
m
1
+m
2
must coincide with the directions dened by f or g,
whenever f, g and h have a unique direction of dierentiability.
Using also Theorem 1.6, we obtain:
Corollary 1.7. For every planar Lebesgue null set E, at each point x E there
is a direction (x) with the following property: every Lipschitz function f: R
2

R
m
is dierentiable in the direction (x) at every x E, except at a uniformly
purely unrectiable set of points. This direction is determined uniquely, except for
a uniformly purely unrectiable set.
Remark. There are null sets which are very far from being purely unrectiable.
For instance, R. O. Davies showed in [4] that every Borel set B R
2
can be
covered by innite straight lines without increasing its Lebesgue measure. One
can even put continuum many lines through each of the points of B so that the
union of these lines has the same measure as B. Now if B = B
0
is, say, a point,
applying Daviess theorem iteratively, we can nd B
0
B
1
B
2
. . . such that
each B
k
has continuum many lines through the points of B
k1
, and the sets B
k
are
Lebesgue null. Then

B
k
is also Lebesgue null, and it has continuum many lines
through each of its points. What could be on

B
k
? Since Lipschitz functions
are dierentiable along lines, at each line of the construction, must agree with
the direction of the line at a.e. of its points. But there are continuum many lines
at each point, how can we choose only one of these, so that along any given line
at a.e. point we choose the direction of the given line and not one of the others?
Now, consider Lipschitz functions on R
n
.
4 G. Alberti, M. Csornyei, D. Preiss
Notation. We denote by N
n,k
the -ideal of subsets of R
n
generated by sets for
which there is a Lipschitz function f: R
n
R dierentiable in at most k linearly
independent directions.
So N
n,0
are exactly the uniformly purely unrectiable sets, while N
n,n1
are
the non-dierentiability sets we are mainly interested in.
Since a Lipschitz function is dierentiable in the tangent directions of any k-
rectiable set at H
k
-almost all of its points, therefore N
n,k1
sets are k-purely
unrectiable, i.e. they meet every k-rectiable set in an H
k
-null set.
As a renement of the above observations on directions of dierentiability in
the plane, we will show that whenever E N
n,k
, there is : E G(n, k) such that
for all x E except those belonging to an N
n,k1
set, every Lipschitz function
f: R
n
R
m
is dierentiable in the direction (x).
Denition 1.8. : E G(n, k) is called a k-dimensional tangent eld of a set E
if every Lipschitz function f: R
n
R
m
is dierentiable in the direction (x) at all
x E except those belonging to an N
n,k1
set.
Theorem 1.9. Every set E N
n,k
has a k-dimensional tangent eld. Moreover,
the tangent eld is unique up to an N
n,k1
set.
It is easy to see that:
Proposition 1.10. The set of (directional) non-dierentiability of a Lipschitz
function f: R
n
R can be written as a countable union of sets E, for each of
which we may nd a direction u and numbers a < b such that
liminf
t0
f(x +tu) f(x)
t
< a < b < limsup
t0
f(x +tu) f(x)
t
.
Since our f is Lipschitz, such set is null not only on every line in direction u,
but also on every curve : R R
n
provided that |

u| is small enough.
We can do slightly better: if > 0 is small enough, for every > 0 there is
an open set G E such that the length of G is less than for every curve
: R R
n
with |

u| < .
This observation motivates the following denition. Given a convex cone C, we
may dene the C-width of an open set G as the supremum of the lengths of G
where the supremum is taken over all Lipschitz curves that go in the direction of
C, i.e. for which

(t) C for a.e. t. Then we dene the C-width for general sets
as the inmum of the C-widths of open sets containing it. In fact, our denition of
the width is slightly more complicated: instead of the length we use a technically
more convenient measure (that also depends on a vector e int(C)) of the part of
the curve that lies in the set G. (See later, Denition 1.14.)
Using this notion of width, an equivalent description of the tangent eld of
a set can be obtained without referring to non-dierentiability sets and non-
dierentiability directions of Lipschitz functions:
Dierentiability of Lipschitz functions, structure of null sets 5
Denition 1.11. If E R
n
, we say that the mapping : E G(n, k) is a k-
dimensional tangent eld of E if for every cone C, the set of those points x E
for which (x) C = {0} has C-width zero.
This denes the same tangent eld as Denition 1.8: one can show that the
family of those subsets of R
n
that admit a k-dimensional tangent eld according
to Denition 1.11 coincides with the -ideal N
n,k
, and also that the two tangent
elds coincide.
According to Proposition 1.10 (and paragraphs preceding it), the set where f is
not dierentiable can be covered by countably many sets, each of which has width
zero with respect to some cone.
We do not know whether this is a full description, i.e. we do not know whether
the non-dierentiability sets of Lipschitz functions (i.e. those sets that admit an
(n1)-tangent eld) are exactly described by the property that they can be covered
by countably many sets, each of which has width zero with respect to some cone.
It is not very hard to show, using Denition 1.11, that the existence of an (n1)-
tangent eld of a set is equivalent to the property that for every > 0 the set can
be covered by a nite number of sets each of which has width zero with respect to
some cone that is only -far from a halfspace.
Our results include:
Theorem 1.12. For every set E R
n
, the following are equivalent:
(i) There is a Lipschitz function f: R
n
R
n
that is non-dierentiable at any
point of E.
(ii) There is a sequence (possibly innite) of Lipschitz functions f
j
: R
n
R such
that at every point of E at least one of the f
j
is non-dierentiable.
(iii) The set E is in N
n,n1
.
(iv) The set E has an (n 1)-tangent eld.
(v) If n 2: E has Lebesgue measure zero.
We do not know whether every Lebesgue null set is in N
n,n1
for n > 2. And
we do not know whether it is true that for every m < n there is a null set E R
n
such that every Lipschitz function f: R
n
R
m
is dierentiable at some point of
E. Preiss proved in [9] that the answer is yes for 1 = m < n. Dore and Maleva
in [6], building heavily on methods due to Lindenstrauss, Preiss and Tiser in [8]
in their study of dierentiability problems in innite dimensional Banach spaces,
proved that the answer is also yes for 2 = m < n. But their current methods do
not work for m 3.
So far we didnt say anything about how we can construct a non-dierentiable
function for a given (small) set E. This is much harder than the other direction, i.e.
showing that the set of points of non-dierentiability must be small. In dimension 1
it is easy, and one may try to use the 1-dimensional proof as a guidance. One
6 G. Alberti, M. Csornyei, D. Preiss
could even consider generalising the more precise description of the sets of non-
dierentiability of Lipschitz functions f: R R due (with slightly worse constants)
to Zahorski [11]. (See [7] for a more recent proof.)
Theorem 1.13 (Zahorski). For any G

set E R of Lebesgue measure zero there

is a Lipschitz function f: R R with Lip(f) 1 which is dierentiable at every
point x / E and
liminf
t0
f(x +t) f(x)
t
= 1 < 1 = limsup
t0
f(x +t) f(x)
t
for every x E.
Recall that a set is G

if it is an intersection of countably many open sets.

Recall also that, by adding together suitable multiples of functions obtained by
this theorem for G

sets E
i
, Zahorski showed that E R is the set of points of
non-dierentiability of some Lipschitz function f: R R if and only if E is of
Lebesgue measure zero and of type G

(a union of countably many G

sets).
So let us see how one can construct a Lipschitz function f: R R non-
dierentiable at the points of the given Lebesgue null set E. We recursively nd
open sets G
1
G
2
. . . E so small that G
k
is small in every component of G
k1
.
(For example, |G
k
C| < 2
k
|b a| for any component C = (a, b) of G
k1
.) Let
f
k
(x) denote the measure of (, x) G
k
. Then f

k
(x) = 1 at each point x G
k
,
but the slope (f
k
(b) f
k
(a)/(b a) is close to 0. Using this it is easy to check that
f(x) =

(1)
k
f
k
(x) is not dierentiable at any point of

G
k
. If E is G

and
> 0, it is not dicult to choose the G
k
so that, dening f(x) =

k
f
k
(x) where
|
k
| < and the partial sums of the
k
oscillate between 1, we get a function
that almost satises the statement of Theorem 1.13. However, at the points of
R \ E we would only get that the upper and lower derivatives of f dier by no
more than 2, not that f is dierentiable. We are in fact able to nd a higher
dimensional analogue of this construction. Recall however that Theorem 1.13 is
proved in a dierent way, and that the weaker statement that we have just indi-
cated is not sucient for showing the full description of non-dierentiability sets
mentioned above.
As a higher dimensional analogue of the functions f
k
, for an open set G R
n
of (small) C-width w and unit vector e from the interior of C, we construct a
function : R
n
R such that Lip() is bounded by a constant depending on C
and e, (y) (x) if y x C, (x +te) = (x) +t if the segment [x, x +te] lies
in G, and 0 (x) w for all x R
n
.
The function can be used to construct non-dierentiable functions, in a sim-
ilar way as the functions f
k
were used in dimension 1. Indeed, has directional
derivative 1 in the direction e at each x G, but from the more global point of
view looks like having derivative zero.
The technical details of the construction are quite complicated. They may
be somewhat simplied in the case of sets E N
n,0
. Given any vector e, we
choose an open set G E with small C-width where C is close to the halfspace
{x : x, e 0}. The function x, e (x) sees, from every point of G, some
Dierentiability of Lipschitz functions, structure of null sets 7
points in the direction e with slope almost one, but has local Lipschitz constant
close to zero on G. This allows us to iterate the construction locally. Moving also
the vectors e through a dense subset of the unit sphere, we get a function which
is non-dierentiable at any point of E in any direction. More precisely, here is our
denition and the results we prove:
Denition 1.14. Let C be a convex cone and let e be a unit vector in C.
(i) We dene M = M
C,e
: R
n
R by
M(x) = sup{ R : x e C} .
(ii) The C-width w(G) = w
C,e
(G) of an open set G R
n
is dened as the
supremum of the numbers

{t:(t)G}
M(

(t)) dt
among all Lipschitz curves : R R
n
which go in the direction of C.
(iii) For a general set E R
n
, w(E) is the inmum of w(G) among all open sets
G which contain E.
(iv) Let G R
n
be an open set of nite width. For every point x R
n
we set
(x) =
G,C,e
(x) as the supremum of the numbers
+

t[a,b],(t)G
M(

(t)) dt
among all a, b R, 0 and : [a, b] R
n
such that (b) x = e and
goes in the direction of C.
We use this function to prove:
Theorem 1.15. For every > 0 and for every set E which is G

and uniformly
purely unrectiable there is a function f: R
n
R such that
(i) Lip(f) = 1;
(ii) f is -dierentiable on R
n
\ E, that is, for every x R
n
\ E there is r > 0
and a vector u such that
|f(x) f(y) u, y x| |y x| for all y B(x, r) ,
(iii) for every x E, B(0, 1) R
n
and > 0 there is an r < such that
|f(y) f(x) , y x| r for all y B(x, r) .
In particular, f is not dierentiable at the points of E, it is not even -
dierentiable for any < 1.
8 G. Alberti, M. Csornyei, D. Preiss
Since every uniformly purely unrectiable set is contained in a G

uniformly
purely unrectiable set, this indeed shows that for every N
n,0
set there is a Lips-
chitz function that is non-dierentiable in any direction. However this result does
not provide Zahorski-type exact description of sets of non-dierentiability in any
direction (which, by analogy, one would conjecture to be N
n,0
sets of type G

),
since we do not know (in dimension n > 1) whether (ii) of Theorem 1.15 can be
replaced by the condition that f is dierentiable on R
n
\ E.
By a rather delicate induction with respect to k (which is where we need the
condition (ii) of Theorem 1.15) we show that the sets of points of k-dimensional
dierentiability can be characterised as follows:
Theorem 1.16. (i) Let f: R
n
R
m
be a Lipschitz function, and for each x
R
n
choose (x) to be a maximal dimensional subspace such that the restriction
of f to x +(x) is dierentiable at x. For each 0 k n 1, let E
k
denote
the set of those points at which dim(x) = k. Then E
k
N
n,k
.
(ii) Let E
k
R
n
be an N
n,k
set for some 0 k n 1. Then there is a
Lipschitz function f: R
n
R
k+1
and a k-tangent eld of E
k
such that f is
not dierentiable at any x E
k
in any direction e that is orthogonal to (x).
We can make (ii) of Theorem 1.16 more quantitative. Again, this is a weaker
analogy of Theorem 1.13, which is needed for induction and to which the same
remarks as to the case k = 0 apply.
Theorem 1.17. For each 0 k < n there is a constant c
n,k
> 0 such that,
whenever l > k, > 0 and E is a G

, N
n,k
subset of R
n
, then there is a function
f: R
n
R
l
with Lip(f) 1 which is -directionally dierentiable at every point of
R
n
\ E and has the property that for every x E there are k-dimensional linear
subspaces V, W of R
n
, R
l
, respectively, so that for any unit vectors v V

and
w W

,
limsup
t0
f(x +tv) f(x), w
t
liminf
t0
f(x +tv) f(x), w
t
c
n,k
.
According to (iii) of Theorem 1.15, c
n,0
= 2. We do not know whether c
n,k
= 2
for k > 0.
We nish this section by showing that for dierentiability with respect to a
measure (instead of at every point of a given set) it is sucient to consider real-
valued functions:
Theorem 1.18. Let be a -nite Borel measure on R
n
.
(i) Every real-valued Lipschitz function f: R
n
R is dierentiable -almost
everywhere, if and only if every set in N
n,n1
is -null.
(ii) On the other hand, if an N
n,n1
-set has positive -measure, then there is a
Lipschitz function f: R
n
R which is non-dierentiable -almost everywhere
on this set.
Dierentiability of Lipschitz functions, structure of null sets 9
In particular, for every singular probability measure in the plane there is a
Lipschitz function f: R
2
R which is non-dierentiable -almost everywhere.
This nicely complements the result of Preiss mentioned before, according to
which there is a null set E R
2
such that every Lipschitz function f: R
2
R is
dierentiable in at least one point of E.
As we have already pointed out, the proof of Theorem 1.16 is rather involved.
However, Theorem 1.18 may be proved in a simpler way, closer to the argument
that we indicated for Theorem 1.15. Recall that the key point of this argument was
that for an open set G of small C-width and e C we constructed a function with
directional derivative 1 in the direction e at each x G, but looking like having
derivative zero from the global point of view. To prove Theorem 1.15, we needed
only one such G (as it contained the whole set E) while to prove Theorem 1.16 we
need several of them which may overlap and so constructions that we need to do
cannot be independent. However, to show Theorem 1.18, we may throw away sets
of small measure, and so achieve that the sets G in which we have to construct the
function are in positive distance from each other. These constructions may still
be handled independently, resulting in a reasonably accessible proof.
2 Structure of null sets and other problems
In this section we list various results that can be proved using similar techniques
and ideas as the ones we use for the characterisation of non-dierentiability of
Lipschitz functions.
2.1 Tangent of null sets
In the planar case, we know that the -ideal N
2,1
and the -ideal of Lebesgue
null sets coincide, i.e. every planar Lebesgue null set admits a 1-tangent eld. We
do not know if the same is true in higher dimension. However, there is another,
weaker notion of tangent elds that can be dened for any Lebesgue null set in R
n
:
Denition 2.1. Given a set E R
n
, we say that a Borel measurable map
: E G(n, k) denes a weak k-tangent eld to E if for every k-rectiable set
S, Tan(S, x) = (x) for H
k
-a.e. x S E.
Notice that in this denition we had to include a measurability assumption. It
was not needed in Denition 1.8 since the tangent eld dened there is automat-
ically Borel measurable (after a modication on an N
n,k1
set). However, under
the continuum hypothesis one can dene a non-measurable weak k-tangent eld
by ordering k dimensional C
1
surfaces in R
n
into S

, <
1
and dening (x) as
the tangent space of S

at x where is the rst ordinal for which x S

.
It follows from the denition that, given a set E R
n
, the weak k-tangent eld,
provided that it exists, is uniquely dened up to k-purely unrectiable subsets of
E (recall that the k-tangent eld is uniquely dened up to an N
n,k1
set). Also,
if a set admits a k-tangent eld then it is also a weak k-tangent eld. We do not
10 G. Alberti, M. Csornyei, D. Preiss
know (even in the planar case for k = 1) whether the -ideal N
n,k1
coincides with
the -ideal generated by G

(or Borel, or analytic) k-purely unrectiable sets, and

we do not know in dimensions n > 2 whether every set admitting a weak k-tangent
eld also admits a k-tangent eld. However, we can prove that:
Theorem 2.2. Any set E R
n
of Lebesgue measure zero admits a weak (n 1)-
tangent eld.
This result can be understood as saying the rather mysterious fact that one
can prescribe in which direction an (n 1)-surface meets a null set E, without
knowing the surface itself. The mystery would deepen if, for example, one had a
purely 1-unrectiable set in N
n,1
\ N
n,0
: this set would have uniquely prescribed
directions that would not be possible to describe by meeting with curves.
2.2 Covering by Lipschitz slabs and intersecting by curves
The notion C-width can be dened in the following, equivalent way. Given a cone
C and a vector e int(C), if E is a C-Lipschitz set, i.e. E(x+C) = {0} and E
meets each line of direction e in exactly one point, then we call the set between E
and its shifted copy E + we (w > 0) a C-Lipschitz slab of width w. If K R
n
is
compact, we may dene its C-width as the inmum of the total width of families
of C-Lipschitz slabs covering it. If G R
n
is open, then we dene its C-width as
the supremum of C-widths of compact sets contained in it, and nally if E R
n
is arbitrary, then its C-width is dened as the inmum of the C-widths of open
sets containing it.
In our original denition of C-width, we measured the part of the curve that
lies in the set G (i.e. we chose the function M(x) in (i) of Denition 1.14) in such
a way that we obtain exactly the same width as the one dened using C-Lipschitz
slabs.
So a compact set has C-width zero if it can be covered by C-Lipschitz slabs of
arbitrary small total width. In particular, in R
2
, every compact Lebesgue null set
is in N
2,1
, therefore it can be covered by Lipschitz slabs of arbitrary small total
width. In fact, in the plane one can cover any null set, and it is enough to use the
coordinate directions and Lipschitz graphs with Lipschitz constant one. We show
the following:
Theorem 2.3. Every set E [0, 1]
2
of measure 0 m < ab is the union of two
sets E = A B, where A has C-width less than a for C = {(x, y) : |x| > |y|} and
B has C-width less than b for C = {(x, y) : |y| > |x|}.
That is, there are Lipschitz functions f
i
: R R, g
j
: R R with Lipschitz
constant 1 and w
i
, w
j
> 0 with

i
w
i
< a,

j
w
j
< b, such that
A

i
{(x, y) : f
i
(x) y f
i
(x)+w
i
} B

j
{(x, y) : g
j
(y) x g
j
(y)+w
j
} .
This can be used e.g. to show that there is a 1-Lipschitz function f: R R
whose graph (x, f(x)) or (f(x), x) meets E in length at least m
1/2
. The analogous
result is also true in higher dimension:
Dierentiability of Lipschitz functions, structure of null sets 11
Theorem 2.4. For every set E [0, 1]
n
of measure m there is a Lipschitz curve
(with a xed Lipschitz constant that depends only on the dimension n) that meets
E in length at least c
n
m
1/n
.
Here a curve means the graph of a map from one of the coordinate axis into
its orthogonal complement. We do not know whether there is a k-dimensional
Lipschitz surface (where surfaces are understood similarly) that meets E in H
k
-
measure c
k,n
m
k/n
.
2.3 Mappings onto balls and weak derivatives
Among the problems exploring the geometric structure of sets with positive
Lebesgue measure, the following one, proposed by M. Laczkovich, is particularly
interesting:
Problem 2.5. Given a set E R
n
of positive Lebesgue measure, is there a Lips-
chitz function f: R
n
R
n
which maps E onto a set with non-empty interior (or,
equivalently, that maps E onto a ball)?
Without loss of generality we can assume that E is compact. In dimension
n = 1, f(x) = |(, x) E| maps E onto an interval and R \ E onto a countable
set.
P. Jones called our attention to a result of N.X. Uy in [10]:
Theorem 2.6 ([10]). For every compact set E R
2
of positive Lebesgue mea-
sure there is a non-constant complex-valued Lipschitz function that is holomorphic
everywhere outside E (including innity).
If we identify C and R
2
, we obtain a mapping f: R
2
R
2
that is orientation
preserving and open on the complement of E; using degree theory it follows that
f(E) f(R
2
\E) a ball. This gives a positive answer to Problem 2.5 in dimension
n = 2.
In dimension n = 2 a completely dierent construction can also be obtained
using our function from (iv) Denition 1.14 (more precisely, the function u(x) =
x (x)e, whose distance from the identity is small). Instead of constructing an
open mapping on R
2
\ E, we show that close to a density point of E a Lipschitz
perturbation of the identity can be found which maps R
2
\ E onto a 1-rectiable
set (and consequently, it maps E onto a set of non-empty interior):
Theorem 2.7. For n = 1, 2 and for every E R
n
of positive Lebesgue measure
there is an orientation-preserving Lipschitz mapping f: R
n
R
n
such that f(E) =
[0, 1]
n
and f(R
2
\ E) is (n 1)-rectiable.
Unfortunately none of these methods are powerful enough to construct such
a mapping in higher dimension; the question in dimensions n 3 remains open.
It may be true that, in any dimension, there is a Lipschitz perturbation of the
identity that maps R
n
\ E onto an (n 1)-rectiable set.
Another use of is the following. Let be a measure such that (S) > 0
for some S N
n,n1
, and let E be a subset of S with (E) > 0 of C-width zero
12 G. Alberti, M. Csornyei, D. Preiss
for some cone C. Let
j
denote the function for w = 1/j. Then the functions

j
: R
n
R have uniformly bounded Lipschitz constants, they converge to constant
0 as j tends to innity, and

j
(x; e) 0 everywhere and

j
(x; e) = 1 for x E.
A moments reection shows that

j
cannot converge to 0 =

in any weak
sense with respect to (and a straightforward smoothing argument can make them
C
1
). Therefore, for a measure in R
n
, weak derivatives of Lipschitz functions may
be dened i is absolutely continuous with respect to N
n,n1
, hence i every
Lipschitz function is dierentiable -almost everywhere. For n = 2 we know that
the above holds i is absolutely continuous with respect to the Lebesgue measure.
This answers a problem due to G. Mokobodzki.
2.4 Tangents of measures
Alberti proved in [1] the so-called rank-one property of BV functions:
Theorem 2.8 ([1]). Let u and v be BV functions on R
n
. Then the direction of the
gradients of u, v agree -a.e. whenever the measure is singular, and absolutely
continuous with respect to the variation of the gradients of both u and v.
This result can be understood as saying that certain class of R
n
-valued measures
in R
n
, namely those that arise as singular parts of derivatives of BV functions,
have a.e. uniquely dened normal directions and so also tangent hyperplanes. The
question naturally arises: for what measures is our (n1)-dimensional tangent eld
uniquely dened almost everywhere? Is it the same as the hyperplane dened via
derivatives of BV functions?
The measure has to be concentrated on N
n,n1
and it has to be absolutely
continuous with respect to N
n,n2
. Since sets from N
n,n2
are purely (n 1)-
unrectiable, for the later requirement it suces that the measure is absolutely
continuous with respect to purely (n1)-unrectiable sets. The former requirement
would be equivalent to singularity if N
n,n1
coincided with Lebesgue null sets,
which we do not know. But the methods used to prove it when n = 2 are powerful
enough to show that every Lebesgue null set in R
n
is a union of a set from N
n,n1
and a purely (n 1)-unrectiable set. So it suces to assume that the measure
is singular, and absolutely continuous with respect to purely (n 1)-unrectiable
sets.
Denition 2.9. A measure on R
n
is called k-rectiable if it is absolutely continuous
with respect to H
k
|
E
, where E R
n
is a k-rectiable set. Measures which can be
represented as integral combinations =


t
dP(t) of k-rectiable measures
t
are called k-rectiably representable.
Theorem 2.10. A measure is k-rectiably representable if and only if (E) = 0
for every k-purely unrectiable set E.
Denition 2.11. Given a k-rectiably representable measure on R
n
, : R
n

G(n, k) denes a k-tangent eld of , if for every representation =

t
dP(t)
where
t
is supported on a k-rectiable set E
t
, there holds Tan(E
t
, x) = (x) for

t
-a.e. x and P-a.e. t.
Dierentiability of Lipschitz functions, structure of null sets 13
The k-tangent eld, if it exists, is uniquely determined up to -negligible sets.
We show that:
Theorem 2.12. An (n 1)-rectiably representable measure admits an (n 1)-
tangent eld if and only if it is singular.
Singular parts of derivatives of BV functions are (n 1)-rectiably repre-
sentable, and indeed, the hyperplane orthogonal to the gradient and the (n 1)-
tangent eld of these measures coincide.
We nish this section by saying that, applying a version of Radon-N ykodim
Theorem, we show that
Theorem 2.13. Every measure on R
n
can be uniquely decomposed as
=
n
+
n1
+. . . +
0
,
where each
k
is a k-rectiably representable measure supported on a (k+1)-purely
unrectiable set.
We do not know whether the measure
k
admits a k-tangent eld for k < n1.
3 Combinatorial connections
Combinatorial connections of our results were rst noted by Matousek. He ob-
served that a part of our proof of Laczkovichs problem is similar to the proof of
the Erdos-Szekeres Theorem, and he recognised that this part may be replaced by
its corollary: for any planar set M having m
2
points there is a function : R R
with Lip() 1 such that one of the sets
{(x, y) M : y = (x)} or {(x, y) M : x = (y)}
has at least m points.
Some of the results on which our proofs are based exploited this connection
and may be considered as a continuous analogy of the Erdos-Szekeres or Dilworth
Theorems. For example, if a set E admits a k-tangent eld then for every de-
composition G(n, k) =

A
j
there corresponds a partition E

E
j
, where E
j
contains those x E for which (x) A
j
. By denition, the set E
j
has width
0 with respect to any proper closed convex cone C for which C S = {0} for all
S A
j
. That is, we can decompose E into parts that can be covered by Lipschitz
slabs of arbitrary small total width. Discrete analogue of this statement says that,
in the plane, a nite set of points can always be covered by a small number of
Lipschitz curves of given directions (and then of course one of them must contain
many points).
The relation to the combinatorial results becomes even more apparent if we
consider weak k-tangent elds. Look at only the special case k = n 1, and
suppose that E [0, 1]
n
is Lebesgue null and compact. Then we can approximate
E by a grid: it intersects o(N
n
) out of N
n
subcubes of [0, 1]
n
. Let C be a convex
14 G. Alberti, M. Csornyei, D. Preiss
cone, and consider the partial order on R
n
dened by x
1
x
2
x
2
x
1
C.
By Dilworth Theorem, the set of the centres of the cubes intersecting E can be
covered by o(N
n1
) chains and o(N) antichains. Chains are curves going in the
direction of C and antichains are C-Lipschitz surfaces. Since E lies in a O(1/N)
neighbourhood of the set of the centres of the cubes, it is covered by o(N
n1
)
tubes going in the direction C and by o(N) Lipschitz slabs of width O(1/N), i.e.
by tubes of arbitrary small total cross-sectional volume and by slabs of arbitrary
small total width. The set covered by tubes meets C-Lipschitz surfaces in a set
of small H
n1
measure, and the set covered by slabs meets curves going in the
direction C in small length.
This decomposition leads to a weak (n 1)-tangent eld as we let the angle
of C tend to a halfspace and its direction run through a dense set of directions.
For other results we would need to cover by slabs only, and we do not know if this
is always possible, except for the 2-dimensional case where there is no dierence
between tubes and slabs.
Many results presented here are connected to a possibility of decomposing
certain small sets, or perhaps even all Lebesgue null sets, in a way reminiscent
of the decompositions of nite sets in combinatorial results. As we have seen
above, the existence of a weak (n1)-tangent eld is a direct corollary of Dilworth
Theorem. For other problems we need a much ner, continuous version of the
combinatorial results whose proofs also use techniques that are not available in the
discrete world, they are purely analytic.
There are also problems that could be solved using discrete decomposition
results, but we do not know if the discrete versions are true. Matousek conjectured
a higher dimensional variant of the Erdos-Szekeres Theorem that would fully solve
Laczkovichs problem. This conjecture was disproved by Tardos. One can however
modify his conjecture so that it would imply a positive answer to our main problem
(all Lebesgue null sets would belong to N
n,n1
):
Conjecture 3.1. For any set M R
n
having m
n
points there is a function
: R
n1
R with Lip() C
n
and an orthonormal system of coordinates such
that the set
{(x
1
, . . . , x
n
) M : x
n
= (x
1
, . . . , x
n1
)}
has at least c
n
m
n1
points.
This problem is open. We only show that, unlike in the plane, the coordinate
systems cannot be restricted to permutations of the standard coordinate system,
not even in R
3
:
Theorem 3.2. For every Lipschitz constant L and for every > 0 there exists a
nite set M R
3
of m
3
points, such that for every : R
2
R with Lip() < L, in
the standard coordinate system in R
3
, all the three graphs x = (y, z), y = (x, z)
and z = (x, y) contain less than m
2
points of M.
However, the dyadic analogue of Conjecture 3.1 is true in any dimension, even
in the standard coordinate-system for Lipschitz mappings with constant 1.
Dierentiability of Lipschitz functions, structure of null sets 15
Consider the unit cube Q = [0, 1]
n
R
n
. A cube in Q is called a dyadic cube
of size 1/2
k
, if it is obtained by dividing Q to 2
kn
subcubes of equal sizes in the
obvious manner. Let Q
0
be the set of points that are not on the boundary of any
dyadic cube. The dyadic distance of two points x, y Q
0
is the size of the smallest
dyadic cube that contains both x and y. This denes a metric on Q
0
. In a current
work M. Csornyei and P. Jones showed that:
Theorem 3.3. (i) For any set M Q
0
R
n
having m
n
points there is a func-
tion : R
n1
R with dyadic Lipschitz constant 1 and there is a coordinate-
direction x
k
(k = 1, 2, . . . , n) such that the set
{(x
1
, . . . , x
n
) M : x
k
= (x
1
, . . . , x
k1
, x
k+1
, . . . , x
n
)}
has at least m
n1
points.
(ii) Every set E Q
0
R
n
of Lebesgue measure m can be covered by dyadic
Lipschitz slabs of total width at most m
1/n
.
References
[1] G. Alberti: Rank one property for derivatives of functions with bounded variation.
Proc. Roy. Soc. Edinburgh Sect. A, 123 (1993), 239274.
[2] G. Alberti, M. Csornyei and D. Preiss. Paper in preparation.
[3] D.N. Bessis and F.H. Clarke: Partial subdierentials, derivates and Rademachers
theorem. Trans. Amer. Math. Soc., 351 (1999), 28992926.
[4] R.O. Davies: On accessibility of plane sets and dierentiation of functions of two real
variables. Proc. Cambridge Philos. Soc., 48 (1952), 215232.
[5] M. Dore and O. Maleva: A universal dierentiability set in Banach spaces with sepa-
rable dual. In preparation.
[6] M. Dore and O. Maleva: Frechet-dierentiability of planar valued Lipschitz functions
on Hilbert spaces. In preparation.
[7] T. Fowler and D. Preiss: A simple proof of Zahorskis description of non-dierentia-
bility sets of Lipschitz functions. Real. Anal. Exchange, 34 (2008/2009), 112.
[8] J. Lindenstrauss, D. Preiss and J. Tiser: Frechet dierentiability of Lipschitz functions
and porous sets in Banach spaces. Book in preparation.
[9] D. Preiss: Dierentiability of Lipschitz functions on Banach spaces. J. Funct. Anal.,
91 (1990), 312345.
[10] N.X. Uy: Removable sets of analytic functions satisfying a Lipschitz condition. Ark.
Mat., 17 (1979), 1927.
[11] Z. Zahorski: Sur lensemble des points de non-derivabilite dune fonction continue.
Bull. Soc. Math. France, 74 (1946), 147178.
Department of Mathematics, University College London, Gower Street, London,
WC1E 6BT, United Kingdom
E-mail: [email protected]