Hausdorff Measure and Level Sets of Typical Continuous Mappings in Euclidean Spaces

Author(s): Bernd Kirchheim

Source: Transactions of the American Mathematical Society, Vol. 347, No. 5 (May, 1995), pp.
1763-1777
Published by: American Mathematical Society
Stable URL: http://www.jstor.org/stable/2154971 .
Accessed: 25/06/2014 02:18

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact [email protected].

American Mathematical Society is collaborating with JSTOR to digitize, preserve and extend access to
Transactions of the American Mathematical Society.

http://www.jstor.org

This content downloaded from 195.78.109.162 on Wed, 25 Jun 2014 02:18:01 AM

All use subject to JSTOR Terms and Conditions
TRANSACTIONSOF THE
AMERICANMATHEMATICAL SOCIETY
Volume347, Number 5, May 1995

HAUSDORFF MEASURE AND LEVEL SETS OF TYPICAL

CONTINUOUS MAPPINGS IN EUCLIDEAN SPACES

BERND KIRCHHEIM
In memoryof Hana Kirchheimovd

ABSTRACT. We determine the Hausdorffdimension of level sets and of sets

of points of multiplicity for mappings in a residual subset of the space of all
continuous mappingsfrom Rn to Rm .

Questions about the structure of level sets of typical (i.e. of all except a
set negligible from the point of view of Baire category) real functions on the
unit interval were already studied in [2], [1], and also [3]. In the last one the
question about the Hausdorffdimension of level sets of such functions appears.
The fact that a typical continuous function has all level sets zero-dimensional
was commonly known, although it seems difficultto find the first proof of this.
The author showed in [9] that in certain spaces of functions the level sets are
of dimension one typically and developed in [8] a method to show that in other
spaces functions are typically injective on the complement of a zero dimensional
set-this yields smallness of all level sets. Here we are going to extend this
method to continuous mappings between Euclidean spaces and to determine
the size of level sets of typical mappings and the typical multiplicity in case the
level sets are finite; see Theorems 1 and 2.
We will use the following notation. For M a set and e > 0, U(M, e), resp.
B(M, e), is the open, resp. closed, e-neighborhoodof M and we write U(x, e)
instead of U({x}, e) . Given a metric space X of functions, we say that typical
f E X has property P if {f E X; non P(f)} is first category in X. We will
always deal with spaces of continuous mappings equipped with the supremum
metric. Finally, to prove that certain properties are typical, we will use the
Banach-Mazurgame. Let (X, p) be a metric space and M c X. In the first
step PlayerA selects an open ball U(x1, e1) . In the second step Player B selects
an open ball U(x2, 62) C U(xl, 61) and then A continues with U(x3, 83) C
U(x2, 82), and so on. By definition Player B wins if Il B(xj, ej) c M, else
Player A wins. We have the following
Thereis a winningstrategyfor Player B if and only if X\M is a first category
set in X.
Receivedby the editors January5, 1994;originallycommunicatedto the Proceedingsof theAMS
by AndrewBruckner.
1991 MathematicsSubjectClassification.Primary46E15, 26B99, 28A78.
Key wordsand phrases. Level set, space of continuous mappings,Hausdorffdimension.

( 1995 AmericanMathematicalSociety
0002-9947/95 $1.00 + $.25 per page

1763

This content downloaded from 195.78.109.162 on Wed, 25 Jun 2014 02:18:01 AM

All use subject to JSTOR Terms and Conditions
1764 BERND KIRCHHEIM

For the proof see [10]. We will also assume the definition of Hausdorff
measure and dimension, also for more general Hausdorfffunctions 0; see e.g.
[11].
Theorem 1. Let 1 < n < m be given. For any k > 1 we denote dk = m -
k(m - n). Thenfor typical continuous f: B(0, 1; Rn) -- Rm and any k > 2
the set
Mk(f) = {x e B(O, 1); cardff1(f(x)) > k}
as well as the set f(Mk(f)) are FE-setsof Hausdorffdimension dk; moreover
for any nonvoidopen U C B(0, 1) both sets U nMk(f) and f(Mk(f) n U) are
of non- a-finite dk-dimensionalmeasure. For k = 1 the statements concerning
f(Ml (f) n U) remaintrue.
For the proof of this theorem we need some preparation. Until the end of its
proof we will therefore always (unless explicitly stated in another way) assume
that n, k > 1, m > n, that dk and Mk(f) are defined as above, and that the
following four orthogonal projections are given.
* P :Rn Rdk, P(Xl, ..., Xn) = (xn-dk+1 d Xn),
* Pt: Rn Rn-dk, P(Xl ..., Xn) = (xl,., xndk),
* Q: Rm 3Rdk, Q(x1, ... ,xm)=(Xn-dk+, ,... ,Xn),
* QI: Rm +Rmdk, Q'(xl, ..., Xm) = (xl,... Xn-dk
d Xn+1,. .,Xm) .

Another trivial but useful remark is that we could replace the Euclidean unit
ball in the theorem by any convex body, or more generallyby any compact K
which is an image of B(O, 1) under some bilipschitz mapping (. Indeed,
( induces a natural isomorphism of W(B(O, 1), Rm) onto W(K, Rm) which
"commutes"with the map f -- Mk(f) and leaves f(Mk(f)) invariant.
Proposition 1. Let B(xi, Ri) c Rn, i = 1, ... , k, be mutually disjoint balls,
f: B -e Rm continuouswhere B = Uk=1 B(xi, Ri), aR > 0, and y E Rm. Then
there is a continuous map g: B tRm such that
(1) JJf-gJIoo<maxIf(x)-yJ+RI, and g(B)cB(y,RI)
.xEB
and thatfor all h: B -e Rm continuous with Ilh- glloo < RI/48k3 and for all
Z E Rdk fulfilling Izl < RI/2k there are points xi,x' E B(xi, Ri) satisfying
-
P(.k' X1) =Q(h(x ) - y) = z
and
h(xki) =h(ki), h(i) = h(o) if1< i, j <k.
Proof. First of all, using a "nice"substitution of variableswhich maps the balls
B(x1, Ri), i > 2, onto disjoint balls B(x', R1) and leaves the conclusion
unchanged, we infer that we can assume all the balls to be of the same radius
R (=R1).
We define the following linear mappings from Rn to Rm:
Fl (XI . ,Xn) = -(Xl,*I Xn~ ?1 * 0)
and for i2, ..., k
xi if 1 <1 <(i-2)(m-n),
[Fi(X..., Xn)]1= Oif (i - 2)(m - n) < < (i - 1)(m - n),
XI-(m-n) if (i- 1)(m - n) < < m.

This content downloaded from 195.78.109.162 on Wed, 25 Jun 2014 02:18:01 AM

All use subject to JSTOR Terms and Conditions
 LEVELSETSIN EUCLIDEANSPACES 1765

Now, we put g (x) = y + Fi (x - x i) for x e B (x, R) . Obviously, (1) is fulfilled.

To prove the statements concerning h, we introduce two auxiliary mappings
from B(xl, R) x ... x B(Xk, R) c (Rn)k into (Rm)(k-1) x Rdk e Rnk

oh (.XI , xk) = (h(x2)-h(x'), )h .X

h(x3)-h(x . h(.xk)-h(x1) P(-1 ))
and
oh (Xl, ..., xk)

(h(x2) - h(x1), h(x3) - h(x1), .. ., h(Xk) - h(x1 ), Q(h(x1 ) - y)).

First, we consider cDp= cDg and DQ=Dg. Note that cDQ= (Dp (= (D) due
to the choice of F1 . Moreover, the map
I Xn + AAn)
Ak) -+(Dx + Al
I
(,&

is a linear isomorphism restricted to (B(O, R))k. Indeed, one easily verifies

that the equation
(F2(A2)_F (Al), Fk(Ak)_Fl (AI) 5P(Al)) = z ERnk
has the following unique solution:
A1.._. Zn(k-l)+j forj=(k-l)(m-n)+l, ..., n,
zi.Zm+j for 1 < I < k and (I - l)(m - n) < j <
I(m - n),
andforl=l,...,k-l
( Zi.m+j-A for <I j (1- l)(m-n),
= Zlm+
i+(n - 1o
IJ = 4 Zlm+j+(m-n) for (1- l)(m - n) < j < 2n -m,
Zlm+j+(m-n) for 2n - m < j < n.
(We recall that (k - 2)(m - n) = 2n - m - dk < 2n - m.) In particular, we
have for this solution AllI< Izl and lAil < 1A11
+ Izl < 21zl if j > 2. So we
obtain for any A' E B(O, R) the estimate

2k 1< ?D(xl +A , xn +An)l.

Hence, jI(z)I > R/(2k- 1) for any boundarypoint z of the domain U(xl, R)
x ... x U(xn, R). Since also iIDh- (D 1100, ll#Dh-Dh'Il100<2kilh- h'IlooI
we
see that for h
any E U(g, R/(8k3)), any t E [0, 1], and any
z E a(B(xl, R) x ... x B(xn, R))
the inequalities
Dht > 2kR ._-2k R R
p (Z)| Dht(Z)l
->
Q2 k -1 8k 2k
holds with ht = g + t(h - g) . Therefore, the classical degree theory for con-
tinuous mappings, see e.g. [6, Theorem 3.1(d3)] ensures that, for each point
z E Rn.k with IzI < R/(2k),

deg((Dp, z) = deg((DQ,z) = deg((, z) = sign(det(DD)) $ 0

This content downloaded from 195.78.109.162 on Wed, 25 Jun 2014 02:18:01 AM

All use subject to JSTOR Terms and Conditions
1766 BERND KIRCHHEIM

holds. In particular, the point (0, ... 0,

, z Zdk) E (Rm)(k-1) x Rdk be-
longs to both sets
D14(B(xl, R) x x B(xk, R)) and DQ4(B(xl,R) x x B(xk, R))
whenever IzI < R/(2k) . But from this statement and the definition of the maps
P, Oh,Q the conclusion of our proposition follows immediately. o
(Dh,

Corollary 1. Let Bi c Rn, i = 1, k, be mutually disjoint balls, f: B -e Rm

.. ,
continuouswhere B = Ui-1Bi and y E Rm, and R > 0 be given. Then there is
a continuous map g: B Rm such that
1f- glloo < max lf(x) - yl +R ()cBy )

and thatfor all h: B -e Rm continuous with Ilh- gll,o < R/48k3 and for all
Z E Rdk fulfilling IzI < R/2k there are points xi E Bi satisfying

Q(h(xl1)- y) = z and h(x1) = h(YV), for I < i, j < k.

Proof. Follows immediately from Proposition 1, e.g. using a suitable diffeomor-
phism of Rn which maps the B1, ..., Bk onto some disjoint balls of radius
R.O

Note that Proposition 1 and the fact that dk < n immediately imply that for
any nonvoid U c [0, 1]n open and for any constant K the set of all f fulfilling
*dk(Mk(f) n U) > K and *dk(f(Mk(f) n U)) > K has a dense interior in
&'([O, in, Rm). Hence, we see that for typical f both Mk(f) and f(Mk(f))
have infinite 'dk-measure "everywhere".In order to prove Theorem 1 in its
whole strength, we have to play the Banach-Mazurgame with a strategywhich
is based on the just-proved proposition and corollary and which builds trees
leading to perfect sets inside sufficientlymany slices of the form Mk (f) nP- I (t),
f(Mk(f)) n Q- (t) . This is motivated by the following simple observation.
Lemma 1. Let M c Ri be arbitraryand f: Ri -- Ri be lipschitz. If
t'({t; f-I (t) n M is uncountable})> 0,
then M is not of a-finite 1-dimensionalHausdorffmeasure.
Proof. On the contrary, suppose M c U' I Mp and X'(Mp) < oo for all p.
Due to [7, Theorem 2.10.25] we have the estimate
rR'l{t; card(f1 l(t) n Mp) > y} < (lip f) l?n (Mp) for all r
which implies that for XI-almost every t and all p > 1 the set f-I (t) n Mp
is countable. But then also f (t) n M is countable for almost each t . 0

Now we are ready to prove

Proposition 2. If k > 2, then for typical continuous f: B(0, 1; Rn) -- Rm is
the set Mk(f) n B(0, I) of non- a-finite tdk -measure.
Proof. During this proof we will use the following notation. Whenever B -

B(x,r) is aballand a>0 then aoB=B(x, ar).

As already mentioned, we play the Banach-Mazurgame. In the first step,
let us be given the first answer U(fi, eC) of Player A. We choose an R'1 E

This content downloaded from 195.78.109.162 on Wed, 25 Jun 2014 02:18:01 AM

All use subject to JSTOR Terms and Conditions
 LEVELSETSIN EUCLIDEANSPACES 1767

(0, 4) such that Ifi(x) - fi(x')I < e1/2 whenever Ix - x'l < R', and then
we select R1 E (0, min{R' , 2 }) and points xl, k such that the balls ...,
B(xi, 2R1) c B(O, R') are mutually disjoint. Further, we put y = fi (0).
Accordingto Proposition 1 we find a map f2 E U(f, L +R1) and 62 E (0, 61 -
Ilfi- f2IIk,) such that for any h E U(f2, e2) and any z E B(O, 2; Rdk) there
exist points i E B(x1, R1), i =1, ..., k, with P(x1 - xl) = z and h(Qi) =
h(xl). These will form the roots of our trees, so we introduce the following,
more appropriatenotation: S = B(P(x1), R1/2k; Rdk), x[O, 0, i] = xi, and
B[o, 0, i] = B(x[o, z, i], RI) for i < k. Then we return the open ball
U(f2, 2).
In the second step, given A's answer U(f3, e3), we fix an R2 in the interval
(0, RI/(200k)) suchthat If3(X) -f3(X')I < e3/8 wheneverIx-x'l < 200kR2.
Now let D1 C B(O, RI/2k; Rdk) be a set maximal among all subsets D of this
ball with the property It- t'l > R2/(2k) for all different t, t e D. (A simple
volume estimate shows that cardDI < (M' + l)k, but we need only the obvious
fact that D1 is finite.) We claim that for all i = 1,..., k, j = 0, 1, and
t E D1 there are closed balls B[(j), (t), i] with centers x[(j), (t), i] such that
all the B[(j), (t), 1] have radius R2 and that

P (B (x[0, 0, 1]j, 2))

C U P (B 1(X() (t),~ 1],~ 2k) for j =O, 1,

tEDI1

(3) P(x[(O), (t), 1]) = P(x[(l), (t) , 1]) = t for all t E D1,
C63
(4) lf 3(x[(j), (t), i]) - f3(X[(j),. (t),~ 1])l < 43 for all i,~j, t,

(5) 2B[(j), (t), i] n 2B[(j'), (t'), i] = 0 whenever (j, t) $ (j', t'),

(6) 2B[(j), (t), i] c 2B[o, o, i] for all i, j, t.

For this purpose we choose a linear order -< of the set D1 and suppose that for
some t E D, the x[(j), (t'), i] c B(xi, R1) already chosen for all i, j, t' -<
t, are different whenever their indices differ and fulfill (3) as well as (5), for
i = 1. According to the choice of our answer U(f2, e2) we find points xi E
B[o, z, i], with f3(xi) = f3(xl) for all i < k and P(xl - x[O, 0, 1]) = t.
First, we consider the system 7 of all balls B(x[(j), (t'), 1], 4R2), j = O, 1
and t' -< t, which satisfy t E P(B). Since all of the P-projections of balls in F
are centered in D1 and at most two of them in the same point, a straightforward
volume estimate shows that N = card, fulfills
N (R)
<R2dk
dk
< (4R2 +T
4k~~4k
and, therefore, N < 2(16k + 1)dk < (40k)dk. Next, we look at the ball B' -
P'(B(x 1, 165kR2)) c Rn-dk and the set M = B'\UBEY-P'(B). We compute
that IMI > ((N + 1) - N)IB(O, 4R2; Rn-dk)l > IB(O, 2R2; Rn-dk)I, here I *I
being the Lebesgue measure on Rn-dk . Hence, the isodiametric inequality (see

This content downloaded from 195.78.109.162 on Wed, 25 Jun 2014 02:18:01 AM

All use subject to JSTOR Terms and Conditions
1768 BERND KIRCHHEIM

e.g. [4]) ensures the existence of two points z?, zI E M with Izo- z I > 4R2.
The desired two points are now defined by the conditions
P(x[(j) , (t) , 1] - x[O, 0, 1]) = t and P'(x[(j), (t), 1]) = zi.

It follows from our definition of M that the balls B[(j), (t), 1] again satisfy
(3) and (5) (together with the other B[(j'), (t'), 1], t' -< t). Now we choose
the points x[(j), (t), i] E B[o, z, i] n B(xi, R2), i = 0 1, 1i > 2, in such
a way that all the points chosen by now will be different. Note that this im-
plies If3(x[(i), (t), i]) - f3(xi)l < e3/8 due to our choice of R2, which yields
(4). If we have proceeded the whole set D1 in this manner, we simply choose
all the B[(j), (t), i], i > 2, sufficiently small to ensure (5), (6). Finally, the
maximality of D1 guarantees (2). (Note that all these considerations can be
omitted in case dk = 0 where S = {O} = D1 and all we have to choose
are two appropriate"disjoint"subballs in each B[(j), 0, 1]). Now we can ap-
ply Proposition 1 to each of the systems {B[(j), (t), i]; i < k} for i = 0, 1
and t E D1. So we obtain a mapping f4 E U(f3, e3) and an e4 E (0, e3 -
11f3- f4Ilk0) such that for all h E U(f4, e4), for any i = O, 1, t E D1,
and for any z E B(O, R2/(2k); Rdk) there are Xi E B[(j), (t), i] fulfilling
P(x1 - x[(j), (t), 1]) = z and h(xi) = h(xl) for all i < k.
We return to U(f4, e4). According to the foregoing statement about h,
after receiving the answer U(f5, e5) we can repeat the construction of Step
II for each of i = 0, 1 and t E D1 with the family {B[(j), (t), i]; i < k}
instead of {B[(o, 0, i]; i < k}. Iterating this procedure, we are always in a
position to obtain a play {U(fi, e1)}IO of the Banach-Mazurgame, sequences
(in 1= 1, 2, ...) of
* RI with 0 < RI+,< R1/(200k),
* Di c B(O, R1/(2k); Rdk) maximal among all subsets of this ball having
minimal distance between its members at least RI+I/(2k),
and k trees (with height I of the corners running from 0 to oo),
* of balls B[co, a, i] centered in x[co, a, i] for i = 1, ..., k, co E
{0, 1}1, and a E D1 x ... x D and of radii RI+1 provided i = 1,
which all together satisfy for any i < k, I > 1 a E D1 x *. x DI, C E
{0, 1}1, and t, t'E DI+1, j, j' =0, 1 with (j, t) $ (', t') the following five
conditions:
(7) 2B[coj, at, i] c 2B[co, a,i]

(8) 2B[coj, at, i] n 2B[coj', at', i] = o,

(9) P( B[, a, 1]) c U P (IB[cj, at, 1])

(10) P(x[Cwj, at, 1] - x[w,a, 1])=t

and finally

( 1) p(' B[1 o a, 11 c P (B[o, a, 1] n

This content downloaded from 195.78.109.162 on Wed, 25 Jun 2014 02:18:01 AM

All use subject to JSTOR Terms and Conditions
 LEVELSETSIN EUCLIDEANSPACES 1769

(These conditions are not independent; e.g. (10) is helpful during the construc-
tion, but at the end we need only its consequence (9).)
Let f be the uniform limit of the fl. We are done if we show that Mk(f) n
B[o, 0, 1] is of non-a-finite Xdk-measure. By Lemma 1 this is surely true
once we know that for any t E S the set
C=Mk(f)fnB[o,o, 1]lnP-1(t)
contains a perfect subset. But indeed, for any such t there is due to (9) a
E Hl DI such that for I > 1, a E {0, 1},
sequence {t1}l??1

t E P (2B[co, (t, @, tl), 1])

If we denote, for coE {0, 1}l, BW = B[co, (t1, ..., tl), 1], we infer from (11)
the existence of some
k
xct E Bct n p- I(t) n n fi-'(fi+I (B[0 , , i)
i=2

Using the uniform convergence of the fi's one sees easily that also, for any
clusterpoint x of {x@; co} and any i = 1, ... , k, f-I(f(x))nB[0, 0, i] :z o
holds; therefore x E Mk(f). Moreover RI \ 0 and (7), (8) imply that the set
of all cluster points of {x@; co} has no isolated point. Consequently, it is a
perfect subset of C. E

Proposition3. For typical f: C(0, 1, Rn) -+ Rm the set f(Mk(f) n B(0, 1/2))
is of non-a-finite Xdk-measure.
Proof. We mimic the proof of Proposition 2, obtain a play {U(fj, 6j)}I?= of
the Banach-Mazurgame and construct simultaneouslythe following objects:
* mutually disjoint balls B1, ..., Bk c B(O, 1/2, Rn),
* a sequence {R1}l??1with 0 < RI+ < R1/4,
* a sequence {D1}l??1of sets maximal among all subsets of B(O, R1/(2k);
Rdk), having minimal distance between its members at least RI+I/(2k),
* a tree of points y[co, a] ERm where coE {0, 1}, a E D1 x ***x DI,
and 1=0, 1,...
which all together satisfy for any I > 0, a E D1 x .. x DI, co E {0, 1}1, and
t, t' E DI+ , j, j' = 0, 1 with (j, t) :$ (j', t') the following five conditions:
B(y[coj, at],~ 2RI+2) c B(y[co, a],~ 2RI+I),
B(y[coj, at], 2RI+2)n B(y[coj', at'], 2R1+2) = 0,
R )2
Q (B (co a], c (E[w
U Q (B at],
[ED,+,
Q(y[coj, at] - y[co, a]) = t
and finally

Q (B (y[co a] cRik, )) CQ (B (y[o , a], R)+lfn fn+2(int(Bi)))

This content downloaded from 195.78.109.162 on Wed, 25 Jun 2014 02:18:01 AM

All use subject to JSTOR Terms and Conditions
1770 BERND KIRCHHEIM

From this we can as in the proof of Proposition 2 conclude that for any uni-
form limit f of the fj's and t E Q(B(y[0, 0], R)) the set Q- 1 (t)nnfk I f(B1 )
contains a perfect subset. Since this means that we won the game, our proof is
finished. o

Next, we derive the result used for the upper estimate of the Hausdorffdi-
mension of the sets Mk .
Lemma 2. The set U of all
2
(Xo * @@1X0 o1*@@1X2 Xok, Xnk) E (Rm )n+l)k

for which
d=dim (nAS(xoi *,x7D) < dk

holds has an interior dense in X = Rm(n+l)k. Here dim 0 = -oo, and we denote
by AS(xo, ... , xn) the smallest affine subspace of Rm containing {xO, ... , xn }.
(We will simply write AS if the xo, ...,n follow from the context.)
Proof. We also define for xo, ..., Xn E Rn
LS = LS(xo .. ., xn) = span({xj - xo; 1 < j < n})
and, for X E X, LS(x) = ni=l LS(xo, xi,Xn) d(x) = dim LS(x) . Similarly
we have AS(x) and d(x).
Since AS = y + LS for any y E AS, we infer that always d(x) > d(x). If
we consider the first case dk > 0, it suffices to show that d(x) < dk on some
dense open subset U' of X. First, note that for all 1 > 1 the set
Ul = {(xl, .., xn); dim span({xl , ..,Xnl) > I}
is open in (Rm)n; this immediately follows from the considerationof appropri-
ate subdeterminants. Moreover, if 1 < n then obviously this set is also dense.
Consequently, the set Xr of all X E X such that dim(LS(xO, ... , x0)) = n
for all i < k is a dense open subset of X. Furthermore,it is clear that for
yl ,..., ykE Rm, X E Xr then condition y'E LS(x, ...,xn) for all i holds
iff dim(span({y, x i - xo, . .. , xni - xo})) < n . Hence, the set

{(x) y'9 y k); yi E LS(Fcoi..., .Xi) for all i < k}

is closed in Xr x (Rm)k. In particular, the set {(x, y); y E LS(x)} is closed
in Xr x Rm. Using this fact and the compactness of the family of all "1-
dimensional" orthogonal systems, one easily sees that for any / > 0
{X E Xr; d(x) > l} is closed in Xr. So, we are done if we can show that
{X E Xr; d(x) < dk} is dense in Xr. The proof of this fact is based on induc-
tion with respect to k and the well-known formula
dim(U n V) = dim(U) + dim(V) - dim(span(U u V))
valid for all linear subspaces U, V of Rm. The induction step follows easily
from a simple perturbationargumentsaying that for any linear subspace U of
Rm and any xo, ... ., Xd E R"mwhere d + dim(U) < m and any E > 0 there are
=
Yi e B(xi, E), i=O, *, d, such that yi -yo 0 span(U u {yj -yo; 1 < j < i})
whenever 1 < i < d.

This content downloaded from 195.78.109.162 on Wed, 25 Jun 2014 02:18:01 AM

All use subject to JSTOR Terms and Conditions
 LEVELSETSIN EUCLIDEANSPACES 1771

Finally, we have to consider the case dk < 0; obviously it sufficesto restrict

to the case dk > n - m. This time we use the fact that those points in Xr
which satisfy d(x) > 1 or AS(x) = z form a closed subset X'. Indeed, let
xp -x X E Xr and all x1pE X'. We may assume the existence of a yp E AS(xp)
for all p; otherwise d(x) > 1. Hence, yp - (Xp)b E LS((xp)i, ... , (Xp)i) for
I-X oc, then any cluster point y of the sequence yp/lypI is also
all i, p . If IyP
a cluster point of {(yp - (xp)i)/IypI}for any i < k. Therefore, it belongs to
LS(x), showing that d(x) > 1. Otherwise, we find a cluster point y of the
sequence {yp} itself. We conclude that y -xF E LS(xo, .. ., xn) for any i < k,
which implies y E AS(x). Summarizing, x E X' in both cases. The density of
the open set Xr\X' is now a consequence of the firstcase for m - n + dk < 0, of
the already stated perturbationargument, and of the following observation. If
U c Rm is linear and x0, ..., xp ERm fulfill xj1 span({xi; i < j} U U) for
all j = 0,... ,p, then U n AS(xo, ... , xp) = z. Indeed, otherwise we could
fix y E U and . ApE R with y - x0o=(xi A_ - xo) . This implies
? = - -
-Y (P_ Ai l)Xo + Axl x; hence Ee Ai = 1 and Ap =* =4= 0,
a contradiction finishing our proof.
Proposition 4. Let p be a Hausdorfffunction with limr\,o e(r)/rdk -. Then
XI'(Mk(f)) -X' (f(Mk(f))) = 0 holdsfor typical continuous f: B(O, 1/2; Rn)
-*Rm.
Proof. As already mentioned, we can consider instead of W(B(0, 1), Rm) also
the space = F(S, Rm) where S is a simplex in Rn. We define for any
'

1 > 1 the open sets .1 c F consisting of all functions f for which there exist
an t1 > 0 and a set M = M(f, 1) c S satisfying
)(M) < I /l and 1(U(f(M) IZ)) < I /l
(12) diam({f(x1); j < k}) > ?I if some xi V M and mi Ixi -x x> ?l.

Suppose that f E nl,, We put A = liminfo,, M(f, 1) and claim

Mk(f) c M. Indeed, suppose that xI, ..., xk E S fulfill x1 M and
A
-
mini<j Ixi xil > a > 0. Then xl V M(f, 1) for some / > 11/, and hence
f(xl) = -.. = f(xk) cannot be true. Therefbre, we also have f(Mk(f)) c
lim inf1...;, f(M(f, 1)) and conclude that X9(Mk(f)) < Yi(f(Mk(f))) <
lim 0 1/1= 0.
So all we need to show is that each of the Al is dense in F. Since

Z00 ,I ]X
[O,I
[~~o dkX[ [2j
1_ 1 )< 5#(ld
2j] p k+/)dk
j ,
\0_ 0 assj*c
i
-- oo,

we see that f E I whenever there is a set M such that M and f(M) are
unions of finitely many bounded sets contained in dk-dimensionalaffinespaces
and that always diam({f(xJ); j < k}) > 0 if xl V M and mini<j Ix-xiX> I-.
Indeed, in this case compactness yields for any p > 0 a uniform estimate for
this diameter on

(S\U(M, p)) x S x ... xsn {(xl,. ..,x Min)mis'- xJ > }.

So let us be given an 1 > 1, f E F, and e > 0. We know from [12, Theorem

4.2.4] that there is a 3-fine triangulation Si, ... , SN of S where 3 < 1/31

This content downloaded from 195.78.109.162 on Wed, 25 Jun 2014 02:18:01 AM

All use subject to JSTOR Terms and Conditions
1772 BERNDKIRCHHEIM

and If(x)-f(x')I < e/3 if Ix-x'l < 3. Consequently, Sjn Sj1 = 0 whenever
Xj E Sj, Xj, E Sj,, and Ixj - xj,I < 1/1. Let Y be the finite collection of all
pairwise disjointed {Sjh, ..., Sjk} and let C be the set of all corners of the
Sj's; i.e. C is the (0)-skeleton of our triangulation. Given any g: C -* Rm
we define its "affine"extension '(g): S - Rm by the simple request that
F(g) = g on C and that '(g) is affine on each of the simplexes Sj, i < N.
Due to our choice of 3, we have IIf- (f)Ijj, < E whenever Ilflc -Jf Ioo < E/2 .
Hence, our problem reduces to showing that for any {Sj, ... , Sjk} and typical
g: C -* Rm the sets M' -nik=_ (g)(Sj1) and Mi'= (g)-1(M') nSj, are all
dk-dimensional polyhedra. For i = 1, ..., k let x,, ..., xl be the corners of
Sj, ; then the statement about M' follows immediately from Lemma 2. Since we
also know (e.g. from the proof of that lemma) that typical '(g)JSj, is injective,
we conclude that also each of the Ml' is a dk-dimensionalpolyhedron. 0

After alt this preparationsit is now very easy to complete the

Proof of Theorem 1. First, note that

Mk(f) = U {xl; exists X2, ... , x with f(xl) = ... f(xk),

1=1

minjx'-x'1>7}
Each of the sets on the right-hand side is obviously compact. Hence, Mk(f)
as well as f(Mk(f)) are always of type F,. The upper estimate of the Haus-
dorff dimension of these set is only a weakening of the just proved Propo-
sition 4. Finally, let {U(x', 2r}1}1? be a base of topology in B(O, 1; Rn).
Then for any 1 there is a diffeomorphism (D of B(0, 1) onto itself with
D(U(x', rl)) = U(O, 1/2) which induces the linear self-isometry Io: f -*f o D
of W(B(O, 1), Rm). Since the class of sets of non-a-finite Adk-measure is in-
variant under 0, we obtain from Propositions 2 and 3 that for typical f and
all / both Mk(f) n U(x', r1) and f(Mk(f) n U(x', rl)) are of non-a-finite
A9dk-measure(for k = 1 only the second of these sets). O
We have now finished the proof of the firsttheorem, so from now on there are
no special assumptions about the number k and also the projections P, P', Q,
Q' become undefined.
We turn now to the second case, if the dimension of the target space does
not exceed that of the source space. Here it seems to be slightly more conve-
nient to study functions defined on the unit cube, but again we could take any
bilipschitzly equivalent set.
Theorem 2. Let n > m > 1. Thenfor typical f: [O, 1]n -* Rm the following
hold:
* int(im f) $ o, o (im f) is of Hausdorffdimension (m - 1).
* For any y E Rm the level set f-l (y) is of Hausdorffdimension at most
n - m and is of non-a-finite Xn-m-measure whenevery E int(imf).
Again, until the end of the proof of this theorem we will always assume
n > m > 1. We split its contents into the two following propositions.

This content downloaded from 195.78.109.162 on Wed, 25 Jun 2014 02:18:01 AM

All use subject to JSTOR Terms and Conditions
 LEVELSETSIN EUCLIDEANSPACES 1773

Proposition5. Let ip be a Hausdorfffunction with limr\,,O O(r)/rn-m = 0. For

typicalcontinuousf: [0, i]n - Rn thereexists a set M C [0, f]n such that
XIP(M) = 0 and card(f-1(y)\M) < m for each y E Rm.
Proof. For any k > 1 we introduce the open sets fk c W = F([O, f]n, Rm)
consisting of all functions f for which there exist sets Mi = Mi(f, k), i =
O,..., m, and an l > O such that
m
Umi=[05 Xc]n
5
?
?(MO) < 2 5

i=O

lf(x) - A(x')l> n if Ix- x'l > kand x X' E Mi for some i > 1.
Now, suppose that f E Ak? I , and define M= limsupk, -Mo(f, k). Obvi-
ously, X (M) < limk,oo Ej>kZ (Mo(f j)) = 0; hence X (M) = 0. Next,
,

suppose that, for some y E Rm, f- (y)\M contains m+ 1 points x1, ..., xm+,
with minj<1lxj - xll > ( > 0. We find a k such that xj ? Uk'>kMo(f, k')
for each j < m + l. Then there is an ie {l, ..., m} and there are j < l
such that x;, xI E Mi(f, k). Therefore, f E Ak implies f(xj) $ f(xl), con-
tradiction. Consequently, the proposition is proved if we show that each Ak
is dense in W'. So, let us be given any f E W, k > 1, and E > 0. We
find an integer N > flik such that If(x) - f(x')I < e/2 whenever Ix - x'l <
vhi/N. Next, we choose an integer p sufficiently large. Using the notation
S(N, t) ={z E [O, 1]; INz - i? < t for some j = O, ..., N} and GI(x, t)=
{j E {1, ..., n}; Xj E S(N, t)} for t E [O, 1/2), we define the sets
Mo =Mo(f, k) = {x E [O, l]n; card(GI(x, 2-P)) > m},
Ml = Ml(f, k) = Cl({X E [0, I]f; card(GI(x, 2-P-1)) > m - l}\Mo),
M2 =M2(f, k)
-cl({x E [0, if]; card(GI(x, 2-P-2)) > m - 2}\(Mo u Ml)),

Mm-, = Mm-i(f, k)
= cl({x E [O, ifn; GI(x, 2-P-m+i) # 0}\(Mo U u Mm-2)),
U
Mm = Mm(f , k) = cl([O, l]U\(MU..UMm_i)).
Observe that for each x E Mi, i > 1, GI(x, t) = GI(x, t') and this set
has precisely i members whenever t, t' E [2-P-i, 2-P-i+l). This implies that
Ix - x'l > 2-P-i/N whenever x, x' E Mi and GI(x, 2-P-i) $ GI(x', 2-P-i).
Now suppose that x, x' E Mi can be joined by a (2-P-i/N)-chain; i.e. there
is a sequence {x'}1 C Mi, x0 = x, xK = x', and lxi - x1+lI < 2-P'I/N
for all 1 < K. By the foregoing and induction GI(x, 2-P-1) = GI(x', 2-P-)
for all 1. Moreover, all xiJ belong to the same component of S(N, 2P-i)
if j E GI(x, 2-P-i); otherwise all the xi are in the same component of
[0, 1]\S(N, 2-P-i). In particular, each of the (finitely many) components of
Mi has diameter at most v/Xi/N (and the distance between two of them is at
least 2-P-i/N). Hence, if we write 7 for the system of all components of
Mi, i > 1, we can choose mutually different reals ac, C E 5, such that

This content downloaded from 195.78.109.162 on Wed, 25 Jun 2014 02:18:01 AM

All use subject to JSTOR Terms and Conditions
1774 BERND KIRCHHEIM

lac - (f(z))il < e/4 for some z E C . Consequently, the choice of N ensures
the existence of a g E U(f, e) fulfilling
(g(x))i - ac forX E C EYi, i = 1, ...m

Obviously, we conclude that g E k if we verify that oo(Mo) < 2-kk But this
follows from:

X.P(MO)< ()(N M + ~I) mr.f([-N'

~ 'N x [0, ]nm

(m)(N + l )m(2P-I N)n-mXp ([0 1-PIN,n,

< (n)(N + 1) - nm
(si-n
< 2-k for p large enough. O

Proposition6. Let an arbitraryset Mo c Rm be given. Thenfor typical contin-

uous f: Rm the following alternative is true: Either MO\ im f
[0, 1]n -* o0
or,for each y E M0, f- (y) of non-a-finite *n-m-measure.
Proof. Again (and for the last time) we play the Banach-Mazur game. We
denote by Q, Q' the orthogonal projections of Rn onto Rm and onto its or-
thonormal complement Rl-m, using the natural inclusion, so Q + Q' = Id. So
suppose U(fi, el) is the first "move" of A. If MO\ imfi # 0, then there is
Y E Mo with dist(y, imfi) > ( > 0. Hence, y ? img for all g E U(f2,62)
where f2 = fi and 62 E (0, min{el, 6}). So in this case, we win the game
before it really starts. Therefore, we can assume Mo c im fi ; in particular Mo
is bounded. First we choose R' such that Ifi (x) - fi (x') I + 2R' < 6/3 when-
ever Ix - xl < 2RI, and then choose a partition of Mo into (nonvoid) pieces
M .. , Mn all of diameter less than R'1 . We also choose points x' E f- I(M1)
and a positive R1 < R'1 such that the balls B(xi, 2Ri) are all disjoint. Now we
can find a function f2 E U(fi, 6i) such that f2(x) = f2(xi)+(2R7IRI)Q(x-x')
whenever IQ(xi-x)j, IQ'(x-xi)I < R1 . To initialize our trees, we choose com-
pact sets K1 c B(Q'(xi), RI) of positive Xn-m-measure but projecting onto
each of the coordinate axes into R\Q. Then we put Joi = {Ki}, oi = {Mi},
and x[K1, Mi, 0] = xi. We return U(f2, 82) where 62 < R1/2 is sufficiently
small to make our answer admissible.
In Step II we obtain A's answer U(f3, 63); a standard degree argument sim-
ilar to that used in the proof of Proposition 1 ensures that for any i < N,
K E 2o4', M E A01, and an arbitrary t E K the inclusion

f3({x; Q'(x) = t and IQ(x[K, M, 0] - x)I < R1})

DB(f2(x[K, M, o]), R1) D M
holds. Again, we choose R'2 fulfilling lf3(x) - f3(x')l + 2RI <63/3 provided
Ix-x'l < 2R'2 . Next, for any i < N we carry out the following construction: we
choose partitions of Jji of K, and *fi of Mi, where Xji consists of closed
sets and both partitions are R'-fine, i.e. contain only members of diameter less
than RI2. For each (K, M) E XI' x 11l we choose y[K, M] E M and two
different points x[K, M, 01, x[K, M, 1] E B(x[K,, Mi, o], RI) with
Q'(x[K, M, j]) = Q'(x[KiK,Mi, 0]) E K,

This content downloaded from 195.78.109.162 on Wed, 25 Jun 2014 02:18:01 AM

All use subject to JSTOR Terms and Conditions
 LEVELSETSIN EUCLIDEANSPACES 1775

and
If3(x[K, M, j]) - y[K, M]I < e3/4 for j = O,1 .
We fix an R2 < RI, R1/4 such that all the balls B(x[K, M, j], 2R2) for
(K, M, i) E T1 = UN1XIi x X4i x {O, 1}1 become mutually disjoint. Similar
to the first step, we get an f4 E U(f3, 2e3/3) such that f4(x) = y[K, M] +
(2R'/R2)Q(x-x[K, M, j]) whenever x E C[K, M, j] = {x; Q'(x)E K and
IQ( - x[K, M, i])I < R2} and (K, M, j) E T, . Then we return U(f4, e4)
with e4 = R2/2. Since again h{x, Q'(x) t and IQ(x-x[K, M, j])l < R2} D
B(y[K, M], R') D M for any h E U(f4,R2/2), t E K, and (K, M, j) E T1,
we can in Step III repeat the construction from Step II, but now on each of the
"cylinders" C[K, M, j].
In this way we obtain a declining sequence RI+I with 0 < R1+2 < RI+,/4;
a nested sequence of open balls U(f1+1, el+,) E F([O, in, Rm) with e21+2<
RI+,; refining sequences J,7 and 4', of R+1,-finepartitions of the sets Ki
and Mi, resp., for i < N; and finally trees of points x[K, M, a] E Rn,
(K M,a) E T = U oTi T1o = Ui=? x .7i x {O, 1} all of them for
1 = 0, 1, ... and fulfilling:
(13) Q'(x) = t and IQ(x- x[K, M, a])1 < RI+1}) D M
f2l+3({x;
if t E K and (K, M, a) E T,, 1 > 0.
(14) B(x[K', M', aj], 2R,+2) c B(x[K, M, a], 2R1+1)
if (K', M', aj) ET1+l, (K, M,a)E TI, j=O, 1, K' cK, and M' cM.
(15) B(x[K', M', a'], 2R1+1) n B(x[K" , Mf", a"], 2R1+1) = 0
if (K', M', a'), (K", Ml", a") E T, but different.
Let f be the uniform limit of the fi's. We finish as in the proof of Proposition
2 concluding that for any i < N, y E M', t E Ki the set S = f-I(y) n Q'-(t)
contains a perfect set. a

Now we are preparedto finish the

Proof of Theorem 2. The fact that typically int(imf) $ 0 is based on the
degree argument used already in the proof of foregoing Proposition 6. Given
f E ' and E > 0 we find g E U(f,-e), 3 > 0, z E Rn-m, and y E Rm
such that h(Q'-1(z)) D U(y, 3) for all h E U(g,9 ). fm-l(a(imf)) > 0
is now obvious, since it holds for any boundary of a nonvoid open bounded
set. (A more precise considerationwould show again that this set typically does
not have a-finite An-m-measure. However, we are mainly interested in its
smallness.) The upper dimension estimate is only slightly more delicate. Take
any d > m- I . Using again triangulationas in the proof of Proposition 4, we see
that the piecewise affinefunctions (now on the unit simplex) which are injective
on each face of the (m)-skeleton of their underlyingtriangulationform a dense
set. Given such a function f and a 3 > 0 such that If(x) - f(x')I > 1x- x'l
wheneverthese points are on some common m-face, we obtain from the already
well-known degree argumentsthat

h(U(x, r)) D U (h(x), 2) if 0 < r < dist(x, (m - 1)-skeleton)

andheU(fS )

This content downloaded from 195.78.109.162 on Wed, 25 Jun 2014 02:18:01 AM

All use subject to JSTOR Terms and Conditions
1776 BERND KIRCHHEIM

Consequently, 6(imh) c U(f(U((m - 1)-skeleton, r), r3/4)) for all h E

U(f, r3/4) and choosing r sufficiently small we get the Hz-measure of the
right arbitrarily small. This shows that for all E > 0 the set of f with
X.d(O(imf)) < e contains an open subset dense in W.
Since always f-1 (y) c MU(f1 (y)\M) with M chosen accordingto Propo-
sition 5 for (p(x) = xn-m/(-logx), wehave dimXpf-1(y) < dims-M < n-mr.
Finally, let { U1}?? be a base of topology in Rm . Then Proposition 6 implies
that for typical f and any 1 > 1 the level set f (y) is of non-a-finite >'n-mm
measurewhenever y E Ul c imf . Since such an 1 exists for any y E int(im f),
we are done. 0

We did not give a lower estimate for the size of level sets f (y) if y E
8(imf). Indeed, by [5], for typical f there are always points y such that
f- 1 (y) is a singleton.
Our last example will show that in the typical case no completely direct gen-
eralization of the results from [8] is possible. Indeed, togetherwith Proposition
5, it shows that for a typical continuous vector field on the unit square there is
no set of injectivity with a zero-dimensionalcomplement (althoughall level sets
are zero dimensional), but that the largestsets of injectivity have a complement
of Hausdorffdimension one with non-a-finite measure.
Lemma3. Let f: [0, 1122- R2 be continuous, let M c [0, 1]2 be of one-
dimensional measure zero, c E R2, and let B(x1, R), B(x2, R) be two disjoint
balls such that
* card(f-1(y)\M) < 2 for all y E2
* If(x) - c - zl < R/8 for j = 1, 2 and lzl < R.
Then the restrictionof f to a set [0, 1]2\S is noninjectivewhenever S is of
a-finite XI -measure.
Proof. Take any such S c [0, 1]2. Using the fact that the map x -- Ix - x21
is lipschitz and Lemma 1, we find a set C c [3R/8, 5R/8] of positive measure
such that for any t E C the following hold:
* aB(x2, t)fnM=o,
* 9B(x2, t) n S is countable,
* f(aB(x2, t)) c U(c, 6R/8)\B(c, 2R/8).
Next, observe that the sets (B(xl, R)\M) nf-1 (f(oB(x2, t))) are disjoint for
t's different from C. Indeed, otherwise we would find x E B(x1, R)\M and
different x', x" E B(x2, R)\M such that f(x) = f(x') = f(x")-a contradic-
tion to the choice of M. Since XI (M) = 0 and S is of a-finite measure, we
infer that for all except countablymany and, therefore, at least for some T E C
XI(B(xl, R) n f-I(f(OB(x2, T)) nfS)) = 0.
Because the map x -- (x/lxl) is locally lipschitz on R2\{O}, we conclude that
the set D of all 0 E [0, 2ir) for which )e'i+xl Mu(f-1(f(O9B(x2, T))fnS))
whenever A E (0, R] is a set of full measure in [0, 2Xr).
Moreover, since for all 0 E [0, 2ir) both f(ei6R/8 + xl) E B(c, 2R/8) and
f(ei0R + xi) 0 U(c, 6R/8) hold and because the paths
Y1: 0 f(el T+xl) and Y2: 0 *eiOR/2+c

This content downloaded from 195.78.109.162 on Wed, 25 Jun 2014 02:18:01 AM

All use subject to JSTOR Terms and Conditions
 LEVELSETSIN EUCLIDEANSPACES 1777

are homotopically equivalent inside U(c, 6R/8)\B(2R/8), we conclude that

ind(yQ,f(el0R/8 + xI)) = 1 :# 0 = ind(y1, f(e'OR + xI)). Therefore, for any
OE [0, 27r) there are AOE [R/8, RI and an xo E OB(x2, T) satisfying
f(x0) -
f(ei'A0 +xI).
Since 9B(x2, T) n M = z, we have x, e i%'th+ XI 0 M for all O E D. This
shows that, for two different 0, 0 E D, x0 + x0 holds. Because Sn aB(x2, T)
is countable, we conclude that x0o f S for some 0o E D. Since eiaoAOO
+x1 V S
for all 0 E D, we see that fl([O, 1]2\S) is noninjective indeed. 0

REFERENCES

1. A. M. Bruckner,Differentiationof realfunctions,LectureNotes in Math.,vol. 659, Springer,

1978.
2. A. M. Brucknerand K. M. Garg, The levelstructureof a residualset ofcontinuousfunctions,
Trans. Amer. Math. Soc. 232 (1977), 307-321.
3. A. M. Brucknerand G. Petruska,Some typical resultson boundedBaire 1 functions, Acta
Math. Hungar.43 (1984), 325-333.
4. Yu. Burago and V. Zalgaller,Geometricinequalities,GrundlehrenMath. Wiss., vol. 285,
Springer-Verlag,1988.
5. M. Chlebik, On extrema of typicalfunctions (to appear).
6. K. Deimling, Nonlinearfunctionalanalysis, Springer-Verlag,1985.
7. H. Federer,Geometricmeasuretheory,GrundlehrenMath.Wiss., vol. 153, Springer-Verlag,
1969.
8. B. Kirchheim, Some further typical results on boundedBaire one functions, Acta Math.
Hungar.62 (1993), 119-129.
9. ., Typicalapproximatelycontinuousfunctions are surprisinglythick,Real Analysis Ex-
change 18 (1992-93).
10. J. Oxtoby, The Banach-Mazurgame and Banach's categorytheorem,Contributionsto the
Theory of Games (Dresher, Tucker,Wolfe, eds.), Ann. of Math. Stud., no. 39, Princeton
Univ. Press, 1957, pp. 159-163.
11. C. A. Rogers, Hausdorffmeasures,CambridgeUniv. Press, 1970.
12. I. M. Singer and J. A. Thorpe, Lecturenotes on elementarytopologyand geometry, Scott,
Foresmanand Company, 1967.

INSTITUTEOF APPLIEDMATHEMATICS,COMENIUSUNIVERSITY,MLYNSKADOLINA, 842 15

BRATISLAVA, SLOVAKIA
Currentaddress: Institut fur Mathematik, Johannes-Kepler Universitat, Linz-Auhof, 4040 Aus-
tria
E-mail address: kirchheiQdcs.fmph.uniba.sk

This content downloaded from 195.78.109.162 on Wed, 25 Jun 2014 02:18:01 AM

All use subject to JSTOR Terms and Conditions