0% found this document useful (0 votes)
71 views75 pages

MMod Federer

This document summarizes Herbert Federer's 1959 paper on curvature measures. It introduces the concept of reach of a subset of Euclidean space, defines curvature measures for sets with positive reach, and states some key theorems about these curvature measures, including a generalization of Steiner's formula and the Gauss-Bonnet theorem. It also describes some analytic methods used in the proofs and mentions a new integral formula developed in the paper.

Uploaded by

Maria Hernández
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
71 views75 pages

MMod Federer

This document summarizes Herbert Federer's 1959 paper on curvature measures. It introduces the concept of reach of a subset of Euclidean space, defines curvature measures for sets with positive reach, and states some key theorems about these curvature measures, including a generalization of Steiner's formula and the Gauss-Bonnet theorem. It also describes some analytic methods used in the proofs and mentions a new integral formula developed in the paper.

Uploaded by

Maria Hernández
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 75

Curvature Measures

Author(s): Herbert Federer


Source: Transactions of the American Mathematical Society, Vol. 93, No. 3 (Dec., 1959), pp.
418-491
Published by: American Mathematical Society
Stable URL: http://www.jstor.org/stable/1993504 .
Accessed: 01/04/2014 14:03

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 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
CURVATURE MEASURES(1)
BY

HERBERT FEDERER

1. Introduction. In the classical theory of convex subsets of Euclidean n


space [BF; H] a major role is played by Minkowski's Quermassintegrale.
These are, up to constant factors, the coefficients of the Steiner polynomial
whose value at any positive number r equals the n dimensional measure of the
r neighborhood of the convex set considered. For a set with sufficiently
smooth boundary, they may be computed by integrating the symmetric func-
tions of the principal curvatures over the bounding hypersurface.
In the branch of classical differential geometry known as integral geom-
etry [BE; S; C2] similar concepts have been studied without convexity as-
sumption for certain types of sets, for example regions bounded by very
smooth hypersurfaces. The central result of this study is the principal kine-
matic formula for the integral, over the group of rigid motions of n space,
of the Euler-Poincare characteristic of the intersection of two solid bodies,
one fixed and the other moving.
In [W] the formula of Steiner was extended to compact regular submani-
folds of class 2 of n space, with coefficients expressed as integrals over the
manifold of certain scalars associated with the Riemannian curvature tensor.
This work was followed by the generalization of the Gauss-Bonnet Theorem
[A; FE1; AW; C1].
All these classical investigations involve related geometric and measure
theoretic curvature properties of various special types of point sets. The
search for a general theory is an obvious challenge. Those subsets of n space
which are to be the objects of such a theory must be singled out by some
simple geometric property. Among these objects must be all convex sets and
all regularly embedded manifolds of class 2 (possibly with regular boundary).
The curvatures attached to these objects should have the global aspects of
Minkowski's Quermassintegrale, yet be determined by local properties; hence
it seems reasonable that they should be measures. Neither the definition of
the curvature measures nor the statement of any important theorem about
them may contain explicit assumptions of differentiability, because arbitrary
convex sets are to be admissible objects. Whatever differentiability may be
required for an auxiliary analytic or algebraic argument must be implied by

Presented to the Society, January 30, 1958 under the title An integral formula, April '26,
1958 under the title A general integral geometricformula with application to curvatures, and June
21, 1958 under the title On sets with positive reach, applied to curvature theory; received by the
editors September 18, 1958. The theory developed in this paper was also the topic of a series of
lectures at the 1958 Summer Institute.
(1) This work was supported in part by a Sloan Fellowship.

418

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
CURVATURE MEASURES 419

the geometric properties. Of course, in order to be worth while, such a theory


must contain natural generalizations of the principal kinematic formula and
of the Gauss-Bonnet Theorem.
This problem presents a timely challenge to a worker in modern real
function theory, which was originally created in large part for the study of
geometric questions. The results of the theory of area have greatly con-
tributed to the understanding of first order tangential properties of point sets,
and one can hope for similar success in dealing with second order differential
geometric concepts such as curvature. In particular the author's previous
work connecting Hausdorff measure with various integral geometric formulae
[F3, 4, 5, 7] may be considered a first order antecedent of the second order
theory developed in this paper.
The objects treated here are the sets with positive reach; the reach of a sub-
set A of Euclidean n space, En, is the largest e (possibly co) such that if
xCEn and the distance, 8A(X), from x to A is smaller than e, then A contains
a unique point, {A(X), nearest to x. Assuming that reach(A) >0, Steiner's
formula is established in the following form: For each bounded Borel subset
Q of En and for 0 ? r <reach (A), the n dimensional measure of
En I{x:8A(X) ? r and (A(X) E Q}
is given by a polynomial of degree at most n in r, say
n
, rn-ia(n - i)cji(AX Q)
i=O

where a(j) is the j dimensional measure of a spherical ball with radius 1 in Ej.
Clearly the coefficients IDi(A, Q) are countably additive with respect to Q,
defining the curvature measures
(Do(A; *), (D(A; X ***Xbn (A;X*
If dim A-=k, then i(A, - )0 for i > k, bk(A, -) is the restriction of the k
dimensional Hausdorff measure to A, and the measures .(A, *) correspond-
ing to i <k depend on second order properties of A. If a sequence of sets, all
with reach at least e >0, is convergent relative to the Hausdorff metric, then
the associated sequences of curvature measures converge weakly to the curva-
ture measures of the limit set, whose reach is also at least e. In this way any
set A with positive reach may be approximated in curvature by the solids
{ X: 5A (X) < S}

corresponding to s>0. If A, B and AUB have positive reach, so does A2nB,


and
ci(A, ) + Di(B7 )U
)bi(A B, ) + cbi(An B, ).
If A CE. and BCEEn have positive reach, so does A XB CE. XEn- Em+n,
and

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
420 HERBERT FEDERER [December

4)k(A X B, *) = E(ji(,4, *) 0 4j(B, *)


i+jr=k

where 0 is the cartesian product of measures. The Gauss-Bonnet Theorem


generalizes to the proposition that if A is a compact set with positive reach,
then the total curvature bo(A, A) equals the Euler-Poincare characteristic of
A. The new version of the principal integralgeometric formula states that if
,I is a Haar measure of the group of isomnetriesof En, A and B are subsets of
En with positive reach, and B is compact, then AC'g(B) has positive reach
for y almost all isometries g, and

(D
i[A n g(B), x -(4 o g-1)]dlg

= E
k-t l=n+ i Cn,k, IJ/)
4k(A, X))I(B,
whenever X and A2are bounded Baire functions on En, x with bounded sup-
port; here Cn,k,l are constants determined by the choice of IA.
Analytic methods can be used in the proof of some of these geometric
theorems, because the concept of reach of a set A is closely related to differ-
entiability properties of the functions 8A and (A. In fact, reach(A) ? e if and
only if 8A is continuously differentiable on {x: 0 < BA(X) <E }. Furthermore, if
reach(A) >s>t>0, then grad 3A is Lipschitzian on {x: t< BA(x) ?s}, and (A
{
is Lipschitzian on x: 3A(x) < s }; hence {x: BA(x) = s } is an n - 1 dimensional
manifold of class 1, with Lipschitzian normal, whose second fundamental
form exists almost everywhere.
The computations involving curvature tensors are greatly simplified
through use of the algebra A*(E) 0A*(E) and its trace function; here A*(E)
is the covariant exterior algebra of a vectorspace E. A similar algebra has
been used in [FL1, 2].
The paper contains a new integral formula concerning Hausdorff meas-
ure, which is used here in the proof of the principal kinematic formula, but
which also has other applications. Suppose X and Y are m and k dimensional
Riemannian manifolds of class 1, m _ k, and f: X-> Y is a Lipschitzian map.
For y E Y compute the m - k dimensional Hausdorff measure of f- {y }, and
integrate over Y with respect to k dimensional Hausdorff measure. It is
shown that this integral equals the integral over X, with respect to m dimen-
sional Hausdorff measure, of the Jacobian whose value at x is the norm of the
linear transformation of k-vectors induced by the differential of f at x. This
result is the counterpart of the classical integral formula for area, which deals
with the case when mn ? k.
2. Some definitions. The purpose of 2.1 to 2.9 is only to fix notations con-
cerning certain well known concepts; more details may be found in references
such as [S] and [B2] regarding 2.3, [L2] regarding 2.4, [F4] regarding 2.6
and 2.7, [L1] regarding 2.8, and [B1I, [W2] or [F9] regarding 2.9. Some new
material occurs in 2.10 to 2.13.

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 421

2.1. DEFINITION. En is the n dimensional Euclidean space consisting of all


sequences x = (xi, , xn) of real numbers, with the inner product

n
X y = xiyi for x, y E En.
i=l

G. is the orthogonal group of En. With zCEn associate the translation


T: En -> En, T,(x) = z + x for x E En.
With RE-Gn and WEEn_m associate the m dimensional plane

X'(R, w) = R(En n I x : Xi = wi for i = 1, n


*,nmI).

2.2. DEFINITION. Suppose f maps an open subset of En into Em.


If f is differentiable (in the sense of Frechet) at x, then the differential
Df (x) is the linear transformation of En into Em characterized by the equation
lim If(x + h) -f(x)- [Df(x)] (h) |/i hi = 0.

In case m =1, grad f(x) En is characterized by the property that


[Df(x)](Ih) = [grad f(x)] 0 h for h E En.

For i = 1, * , n, Dif(x) is the partial derivative of f at x in the direction


ot the vector whose coordinates are 0 except the ith, which equals 1.
2.3. DEFINITION. Use will be made both of Caratheodory outer measures
[S, Chapter 2] and of countably additive functions [S, Chapter 1] on the
class of all Borel sets with compact closure in a locally compact space, which
will be called Radon measures in accordance with [B2, Chapter 3]. A measure
,u over a space X may be thought of either as a function on a suitable class
of subsets of X, or as a function on a suitable class of functions on X. It is
convenient to use the alternate notations

ff(x)dyx = fd = f
With each Radon measure ut one associates its variation measure j
With measures A and v over X and Y one associates the cartesian product
measure A(D0vover X X Y.
2.4. DEFINITION. Ln is the n dimensional Lebesgue measure over En. On is
the Haar measure of Gn such that On(Gn)= 1.
Under the map which associates with (z, R) E,,XG, the isometry T, o R
of En, the image of the measure Ln0Og5bis a Haar measure of the group of
isometries of En.
Under the map X', the image of the measure On0 Ln-m is a Haar measure
for the space of all m dimensional planes in En, invariant under the group of
isometries of En.

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
422 HERBERT FEDERER [December

2.5. DEFINITION.

a(k) = Lk(EkCr, {x: I xl < 1}) = 2kr(-) r( )r(k+ 1)-i,

I n - k+ I
rk + )
2 2
3(n,k) = a(k) a(n-k)
a(n)() r r

-y(n, k, 1) = 1)
f3(n, k)f(n,
f3(n, k + I - n)3(2n - k - 1, n - 1)
(k+ I)(I 1)I
(2 )(2)
(k + I-n + )(n +1)

2.6. DEFINITION.HI is the k dimensional Hausdorff measure. If A is a


subset of a metric space, then Hk(A) equals the limit, as r-*O+, of the infimum
of the sums
E 2-ka(k) diameter(S)k
SEF

corresponding to all countable coverings F of A such that diameter(S) <r


for SEF.
2.7. DEFINITION.A subset of a metric space is called k rectifiable if and
only if it is the image of a bounded subset of Ek under a Lipschitzian map.
The union of a countable family of k rectifiable sets is said to be countably k
rectifiable.
2.8. REMARK.Since differentiability is invariant under continuously dif-
ferentiable homeomorphisms, this concept remains meaningful for maps of
manifolds of class 1. An intrinsic tangent vector v of a manifold of class 1
at a point p may be thought of as operating on every function f which maps
some neighborhood of p into some Euclidean space and which is differentiable
at p; then df(v) ==v(f).
Tn generalizing measure theoretic properties from En to an n-dimensional
Riemannian manifold of class 1, one replaces Ln by Hn. A Lipschitzian map
of such a manifold into some other Riemann manifold is differentiable Hn al-
most everywhere.
If X and Y are m and n dimensional Riemannian manifolds of class 1, then
Hm+n(S)= (Hm 0 Hn)(S) for S C X X Y.
Using matrices, one may think of Gnas an n(n- 1)/2 dimensional compact

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
19591 CURVATURE MEASURES 423

submanifold of En,2.Then the left and right translations of Gn are induced by


elements of Gn2,hence Hn(n-l)12 induces a Haar measure over Gn and
Hn (n-1)2(S) = Hn(n-1)'2(Gn) -4n (S) for S C Gn.

2.9. DEFINITION. For each finite dimensional real vectorspace E and


k=O, , dim E let

Ak(E) and Ak(E)


be the associated spaces of k-vectors and k-covectors (contravariant and co-
variant skewsymmetric tensors of rank k). Also let
dim E dim E
A*(E)= ED Ak(E) and A*(E)= ff Ak(E)
k=O k=O

be the corresponding exterior algebras, with the Grassman multiplication A.


With each inner product of E one associates the unique inner products of
A*(E) and A *(E) such that the Grassman products of the subsets of- any
orthonormal base of E form an orthonormal base of A*(E), and the Grassman
products of the subsets of the dual base of Al(E) form an orthonormal base
of A*(E).
2.10. DEFINITION. Suppose X and Y are Riemannian manifolds of class 1,
f: X->Y, and
k = inf {dim X, dim Y}.

If pEX, f is differentiableat p, E and F are the tangent spaces of X and Y


at p and f(p), then the differential of f induces dual linear transformations of
Ak(E) into Ak(F) and of Ak(F) into Ak(E) with the common norm
Jf(p).
Using the matrix of the differential of f at p with respect to orthonormal bases
for E and F, one computes Jf(p) as the square root of the sum of the squares
of the determinants of the k by k minors of this matrix.
2.11. DEFINITION. Suppose E is an n dimensional real vector space. Con-
sider the tensor products
Akl(E) = Ak(E) 0, Al(E) for k, I = 0, 1, . . . , n
n

A**(E) = A*(E) 0 A*(E) = D Ak l(E)


k, 1=O

and make A**(E) into an associative algebra by defining the product

(a ? b)*(c 0 d) = (a A c) 0 (b A d) for a, b, c, d C A*(E).

Observe that while A*(E) is anticommutative, A**(E) is not anticommuta-

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
424 HERBERT FEDERER [December

tive. However this definition has the advantage that the subalgebra
n
ffl Ak,k (E)
k=O
is commutative.
This construction is natural; a linear transformation f: E-*F induces a
homomorphism f *: A ** (F) --A **(E).
Now fix an inner product * of E. The corresponding inner product 0 of
A*(E) induces a unique inner product 0 of A**(E) such that
(a 0 b) 0 (c O d) = (a * c)(b O d) for a, b, c, dC A*(E).
On the other hand the inner product of A*(E) corresponds to a real valued
linear function on A**(E), the trace, which is characterized by the formula
trace(a 0 b) = a * b for a, b E A*(E).
Since Ak(E) is the conjugate space of Ak(E), there is a natural isomorphism
of Ak,k(E) onto the space of bilinear forms of Ak(E); if a, bCAk(E), then the
bilinear form B corresponding to (aob) is given by the equation
B(x, y) = a(x)b(y) for x, y E Ak(E).
Furthermorethe space of bilinearforms of Ak(E)is isomorphicwith the space
of endomorphismsof Ak(E); a bilinear form B and the correspondingendo-
morphism T are related by the formula
B(x, y) = T(x) 0 y for x, y C Ak(E).
In particular, if 01, , an form an orthonormalbase of A'(E) and
n
I-?0= A1?1(E)o
X i
i=1

then the inner product and the identity endomorphismcorrespondto I.


2.12. REMARK. Assume the conditions of 2.11 and suppose Gi, * * * , On
form an orthonormalbase of A'(E). For k=0, * , n let Sk be the class of
all subsets of {1, , n} with k elements, and for aCSk let
Oa Oa, AOa2A AOak

wherea1<a2 < < are the elements of a. Then the following statements
<ak
hold:
(1) {tOaOb: aeSi, bCSj} is an orthonormalbase of Aiti(E).
(2) If MCAi,i(E) and NCAk l(E), then

(3) If MCA'11(E), then lMkl <?k!JMk.


(4) If MCAk,k(E) and j=0, * * *, n-k, then

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
19591 CURVATURE MEASURES 425

trace (MN) = (n -k) (n- k -j) !-1 trace (M).


(5) If MCA 1(E) and f is the endomorphism of E corresponding to M,
then
n
M =f*(O.) 0 O
i=1

and the endomorphism of Ak(E) induced by f corresponds to

k!-lMk = >jf*(O) (0 Oa.


aESk

Consequently det(f) = trace(n!-lMn) and the characteristic polynomial off is


n
trace[n-l(M - XI)n] = trace(k!-lMk). (-X)n-k.
k=O
n
(6) If W1 ** WnCA'(E) and M= wi0Oi, then
n n
A wi = trace(nV-lMn) A oi.
i=1 i=1

The verification of (1), (4), (5) is quite easy. Furthermore (3) follows by
induction from (2), and (6) follows from (5) with wi=f*(Oi). To prove (2),
use (1) to expand
M - E $)
Ma,bOa OOb, N= Nc,dO9c ( Od.
(a,b)ESiXSj (c,d)ESkXSI

For (u, V) Si+kXSJ+I let


P(u, v) = (Si X Si X Sk X S1) r { (a, b, c, d): a U c - u, b U d = v},
and for (a, b, c, d)EP(u, v) choose Ea,c= ?1 and Eb,d = ?1 so that

Oa A Oc = Ea,cOu and Ob A Od = Eb,dOv.

Using (1), H6lder's inequality and the fact that the set P(u, v) are disjoint,
one obtains
| MN 12 [ Ma,bNc,dEa,cEb,dl
(u,v)eSi+kXSj+L (a,b,c,d)eP(u,v)

+
< (Ma,bNC,A2 (
\
)
/
(u,v)eSi+kXSj+? (a,b,c,d)EP(u,v)

(i + k) (j + I) E (Ma,b)2(Nc,d)2
t v1 ~~~~(a,
b, c,d) E SiXSjXS kXSI

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
426 HERBERT FEDERER [December

2.13. REMARK. Under the conditions of 2.11 it is true that if a real valued
linear function Q on A 'k(E) is invariant under the endomorphisms of Ak k(E)
induced by the orthogonal transformations of E, then Q is a real multiple of the
trace.
In fact, using the notations of 2.12, one sees that if a, bESk, then
Q(Ga 0 Ob) = 0 in case
s a b,
because if iGa-b and f is the orthogonal transformation of E such that
f*(Gj)=-Gs and f *(Gj)=Oj for j$i, then
Q(Oa 0 Gb) = Q[f*(Ga 0 Gb)] = Q(GOa 0 Gb) = Q(Ga 0 Gb);

furthermore
Q(Ga 0 Ga) = Q(Gb 0 Gb)

because if a1< a2< . . <ak and b1 b2< ... <bk are the elements of a and
b, then there is an orthogonal transformation f of E such that
f*(Oai) = Obi for i = 1, , k, hence f* (Oa) = b.

Consequently Q= Q(Ga0Ga) trace, where a C Sk.


3. An integral formula concerning Hausdorff measure(2). Complement-
ing the classical integral formula for the area of a map f: X-* Y such that
dim X <dim Y, the theorem proved in this section concerns the case when
dim X ? dim Y. The original motivation leading to the discovery of this theo-
rem was the simplification of certain arguments in [DG]; in fact, if X=A
=Em and Y=E1, then the formula becomes

A | gradf(x) dLmX =X Hm-1(f-1{y})dy.

The theorem will be used in the present paper to prove the kinematic for-
mula, and may be expected to have further applications.
3.1. THEOREM. If X and Y are separable Riemannian manifolds of class 1
with
dim X = m > k = dim Y

and f: X-* Y is a Lipschitzian map, then

'AJf(x)dHx = f Hm-k(A n f-1{y})dHky

(2) The author's abstract containing this formula was received by the American Mathe-

matical Society on November 22, 1957, and published in the Notices of the American Mathe-
matical Society as Abstract 542-43 in vol. 5 (1958) p. 167. At the 1958 Summer Institute L. C.
Young announced his independent discovery of a very similar theorem and distributed copies
of his Technical Summary Report No. 28, U. S. Army Mathematics Research Center, Univer-
sity of Wisconsin, May, 1958, which contains an outline of his argument.

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 427

whenever A is an Hm measurable subset of X, and consequently

fbg(x)Jf(x)dHmx = f f g(x)dHm-kxdHky

whenever g is an Hm integrable function on X.


Proof. Suppose M is a Lipschitz constant for f and let A be the measure
over X such that

pA(A)= f*Hm-k(A njf-1{y} )dHky

for A CX, where I" means upper integral. For aEX, let
K(a, r) = XGn {x: distance (x, a) < r
whenever r >O, and let
,u'(a) = lim ,4[K(a, r)]/Hm[K(a, r)].
,r-O+

The remainder of the argument is divided into seven parts, leading to the
first conclusion stated in the theorem. The second conclusion may be derived
from the first by the usual algebraic and limit procedure, starting with the
case in which g is the characteristic function of an Hm measurable set.
PART1. If A CX, then

,>(A) < Ml (k)a(m


k) Hm(A).
a(m)
This inequality was proved in [F7, ?3 ].
PART2. If A is an Hm measurable subset of X and
v(y) = Hm-k(A nf-'{y})
for y Y, then v is an Hk measurable function.
Proof. If Hm(A) = O it follows from Part 1 that v(y) = O for Hk almost all
y in Y. Since every Hm measurable subset A of X is the union of an increasing
sequence of compact sets and a set of Hm measure zero, it will be sufficient
to consider the special case in which A is compact.
For n= 1, 2, 3, * * * and yE Y let vn(y) be the infimum of
Z 2k-m(m
SEG
- k)[diam (S) ]m-k

where G is a countable open covering of A NC- {y } such that diam (S) <n-
whenever SEG; then
v(y) = lim vn(y).
n--ya

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
428 HERBERT FEDERER [December

Since A is compact, every open covering of A C\fl{ y } is also a covering of


A nf-l{ z} provided z is sufficiently close to y. Accordingly the functions v.
are uppersemicontinuous.
PAFT 3. If A is an Hm measurable subset of X, then

u(A) = Hm-k(A cf-1{y})dHky = f:/'(X)dHmX.

Proof. The first equation follows from Part 2 and the definition of A. It
shows that Auis completely additive on the class of all Hm measurable subsets
of X. On the other hand it follows from Part 1 that Auis absolutely continuous
with respect to Hm. Accordingly l is the indefinite integral of its derivative /'.
PART 4. If aX= Em, Y= Ek, f is continuously differentiable in a neighbor-
hood of a and Jf (a) = 0, then ,u'(a) = 0.
Proof. Suppose E>0. Since range Df(a) #Ek there is a real valued linear
function q on Ek such that q =1 and q o Df (a) = 0. The continuity of Df at a
implies the existence of a convex neighborhood U of a such that
qoDf(x) < EM forxE U,
whence

I (qof)(x) -(qof)(z) ?< EM x - z| for x, z C U.


Furthermore, since Ais invariant under rotations of Ek, one may assume that
q(y) = yk for y C Ek.

It follows that if S is the endomorphism of Ek such that


S(y) = (yl, Yk-1, elyk) for y C Ek,

then

(Sof)(x)-(Sof)(z)I <2M |x-zI for x,zEU.


Applying Part 3 to f and S o f, and Part 1 to S o f, one concludes that if A
is an Hm measurable subset of U then

X(A)= fHm-k (A f-1{y})dHky


Ek

- fEkHm-k [A n (S of)-l{S(y)} ]dHky


Ek

- efJHm-k [A n (Sof)-1{w}]dHkw
Ek

< E(2M)k (k)c(m k) Hm(A).


a(m)

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 429

PART 5. If aeX=EEm, Y=EEk and f is continuously differentiable in a


neighborhoodof a, then ,u'(a) = Jf (a).
Proof. In view of Part 4 suppose Jf(a) $0. Since
dim kernel Df(a) = m - k

and both Jf and ,u' are invariant under rotations of Em, one may assume that
Dif(a) = O for i = k + 1,** *,m.
Letting F: Em,,-*Embe the map such that if xCEEmthen
[F(x)]i = [f(x)]i for i = 1, * , ky
[F(x)]i = xi for i = k + 1,* m
one sees that
JF(a) = Jf(a) %;z0.

Accordingly, if r is a small positive number and A =K(a, r), then F is con-


tinuously differentiable and univalent on A. For yeEk let
B = Em-k nI {Z: (yi, yk, Zi, . . *, Zm-k) E F(A)}
gv: B, -* Em,
gv(z) = (F I A) (y, **, yk, zj, * *, Zm-k) for z E By,

and observe that Bv is open, gy is continuously differentiable and univalent,


with
range gv = A nfl{y}.

It follows from the classical formula for area (see [F4, 5.9]) that

u(A)= f Hm-k(A nf-1{y})dHky


Ek

EkB

- 'F (AJg(WI, ,Wk)(wk+71*1 * wm)dLmw


- J'w fJzJg,(z)dLm-kzdLky
F(A)

Jgf (x) (xk+1 ... * Xm)JF(X)dLmx.

Accordingly, if r is small, then ,i(A)/Hm(A) is close to


Jgf (a) (ak+l, * * , am)JF(a) = Jf(a).

PART6. If aEX and f is continuously differentiable in a neighborhoodof a,


then ,u'(a)= Jf(a).

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
430 HERBERT FEDERER [December

Proof. Suppose 1 <t < oo. Choose open neighborhoods U of a and V of


f(a), and continuously differentiable maps
P: U -Em and Q: V -* Ek
such that
P(x) () t
< {<
= for x, z E U,
distance (x, z)
-< Q(Y) Q(W) < t for y, w C V.
=
distance (y, w)
It Tollows that suitable powers of t will serve as bounds for the effect of P
and Q on Hausdorff measures and Jacobians. In fact suppose that r>O,
A = K(x, r) C U and f(A) C V,
let F= Q of o P-' and observe that
A n f- { y} = p-1 [P(A) n F-1{ Q(y) }] for y C V.
Applying Part 3 to f, and Parts 3 and 5 to F, one obtains

i(A) = f Hm-k(A n 1f-1{y})dHky

< 1m-k f Hm-k[P(A) n F-1{ Q(y) }]dHky

< tmfHm-k[P(A) n F-1{ q} ]dHkq


Ek

<tm+ JF(p)d(mp
P(A)

-< tm+2kJ Jf [P- (p) ]dHmp


P(A)

< t2m+2k fJf(x)dHmx,

and similarly

(AA) > t-2m-2k fJf(x)dHmX

Dividing by Hm(A) and letting r- O +, one concludes that the extreme


limits of
,4[K(a, r)]/Hm[K(a, r)]

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 431

lie between
t-2m-2kJf(a) and t2m+2kJf(a).

PART 7. If A is an Hm measurable subset of X, then

it(A) = Jf(x)dHmx.

Proof. Proceeding as in [F2, 4.3] or [W1], choose disjoint closed subsets


* * of X and continuously differentiable maps fi, f2y fa ...
Ci, C2, C3, * of
X into Y such that

Htm(X - u Ci) 0,

f| ci=fJ ci fori = 1, 2, 3,*.


Applying Parts 3 and 6 to fi and observing that
Jfi(x) = Jf(x)
whenever x is a point of density of Ci and f is differentiable at x, one obtains

fmmH-k( A \ Ci r) f-1{y})dHky

- fs r
Hm-k(A rC CiC fT1{y})dHky

- fAflCiJfi(x)dHmx = fA Jf(x)dHm"x
Anci AnCi

for i= 1, 2, 3, ... . Accordingly Part 1 implies that


00

1u(A) = u rl C)
1(A
it=1
00

- Ei Jf(x)dHmx = fff(x)dHmx.
i=1 nciA
3.2. REMARK.The preceding argument shows also that f{y} is count-
ably Hausdorff m - k rectifiable (see [F4]) for IJ almost all y in Y.
In case X is a submanifold of class 1 of En, Hm-k(A flf {y}) may be
computed for Hk almost all y in Y by means of the integralgeometric formula
[F4, 5.14], and one obtains

fHm-k(A rtf-1{ y})dHky = 8(nM k


-k)-1 f u(R, w)d(Qn 0 Lm-k) (R w),
wnXEe-k

where

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
432 HERBERT FEDERER [December

u(R, w) H
JH[A C n-V?k(R, w) nf-fl{ }]dHk
y~~~
is the classical area of fJ [Ar\X m+k(R, w)]. It follows that, if A is open in X,
then the above integrals depend lowersemicontinuously on f, and un-
doubtedly it would be possible to develop (for m ? k) a theory of "coarea"
dual to the existing (for m?k) theory of Lebesgue area [R; CE; F8; DF].
4. Sets with positive reach. Here these sets are introduced, and are
shown to have quite reasonable metric and tangential properties. If two sets
in suitably general relative position belong to this class, so does their inter-
section. The class contains all convex sets, as well as all those sets which can
be defined locally by means of finitely many equations, f(x) = 0, and inequali-
ties, f(x) < 0, using real valued continuously differentiable functions, f, whose
gradients are Lipschitzian and satisfy a certain independence condition;
therefore regular submanifolds of class 2 of En, with or without regular bound-
ary, are included.
The concept of reach originates from the unique nearest point property,
but toward the end of this section it is proved that a closed set has positive
reach if and only if it makes uniform second order contact with its tangent
cones. Then it follows that the class of sets with positive reach is closed under
bi-Lipschitzian maps with Lipschitzian differentials.
4.1. DEFINITION.If A CE., then 5A is the function on E. such that
3A(X) = distance(x, A) = inf x- a a C A}
whenever xCEn. Furthermore
Unp(A)
is the set of all those points xCE. for which there exists a unique point of A
nearest to x, and the map
tA: Unp(A) -* A
associates with xEzGUnp(A)the unique aCEA such that AA(X) x-al.
If aCA, then
reach(A, a)
is the supremum of the set of all numbers r for which

{x: x-al < r} C Unp(A).


Also
reach(A) = inf {reach(A, a): a C A}.
4.2. REMARK. Suppose A CEn. Then reach(A, a) is continuous with re-
spect to aGA, and

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 433

0 < reach(Boundary A, a) < reach(A, a) < oo


for a C Boundary A. If reach(A) > 0, then A is closed.
A well known characterization of convexity, deducible from 4.8(8), states
that reach(A) = oo if and only if A is convex and closed.
4.3. DEFINITION.If A CEn and aGA, then the set
Tan(A, a)
of all tangent vectors of A at a consists of all those uGE. such that either u is
the null vector or for every e>0 there exists a point bEA with
b -a u
0< l b-a l < and b I _- <

4.4. DEFINITION. If A CEn and aCA, then the set


Nor(A, a)
of all normal vectors of A at a consists of all those vEE. such that
v* u ? 0 whenever u E Tan(A, a).
4.5. REMARK. Recall that a subset C of En is a convex cone if and only if
x+yCC and XxCC whenever x, yCC and X>0. For every subset S of E.,
Dual(S) = {v: v @ u ? 0 for all u E S}
is a closed convex cone, and Dual [Dual(S) ] is the smallest nonempty closed
convex cone containing S. Furthermore
Dual(S1 + S2) = Dual(Si) n Dual(S2)

for any two subsets S1 and S2 of En containing the origin, and


Dual(S1 n S2) = Dual(S1) + Dual(S2)

in case S1 and S2 are closed convex cones. Also, for every closed convex cone C,
dim C + dim Dual(C) _ n
with equality holding if and only if C is a vectorspace.
Accordingly
Nor(A, a) = Dual[Tan(A, a)]
is always a closed convex cone, while Tan(A, a) is closed and positively homo-
geneous but not necessarily convex.
4.6. REMARK. If A is a submanifold of class 1 of E., f: A -E, is the inclu-
sion map, and aEA, then df maps the intrinsic tangent space of A at a iso-
metrically onto Tan(A, a).
4.7. LEMMA. Suppose f is a real valued Lipschitzian function on an open

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
434 HERBERTFEDERER [December

subset W of En, j is an integer between 1 and n, and g is a real valued continuous


function on W such that
D,f(x) g(x) wheneverf is differentiableat x.
Then
Djf(x) = g(x) for all x C W.
Proof. Suppose wEW, r>O and {x: |x-w| <2r} CW. Let y be the jth
unit vector. According to Rademacher's theorem f is differentiable L. almost
everywhere in W, and for L. almost all x within r of w it is true that

f(x + ty) -f(x) f DJ(x + uy) du g(x + uy)du


whenever ItI <r. From the continuity of f and g it follows that
t
f(w + ty) -f(w) f0g(w + uy)du whenever |t| <r,

and finally that Djf(w) = g(w).


4.8. THEOREM. For every nonempty closed subset A of E. thefollowing state-
ments hold, with 5=5A, t=lA, U=Unp(A):
(1) I3(x)-8(y) | <|Ix-y | whenever x, yEEn.
(2) If aEA and
P= {v:t(a+v) = a, Q= {v:5(a+v) = Ivi}I
then P and Q are convex and PCQCNor(A, a).
(3) If xGCEn-A and S is differentiable at x, then xE U and

-x - (x)
grad 5(x)
O(X)
(4) t is continuous.
(5) 5 is continuously differentiable on Int(U-A) and 52 is continuously
differentiableon Int U with
grad 62(x) = 2[x - t(x)] for x E Int U.
(6) If a GA, v EEn and
0< r sup{ t: t(a + tv) = a4 < oo,
then a +rv E Int U.
(7) If xE U, a= (x), reach(A, a) >0 and bCA, then

b11
x
(x-a) * (a-b) I_ 2-a ' a

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 435

(8) If O<r <q < oo, xe U, ye U and


5(x) ? r, 5(y) < r, reach[A, t(x)] > q, reach[A, t(y)] > q,
then

t(x)-_t(y) < q rx-yI.


(9) If O< s < r < reach(A), then grad 5 is Lipschitzian on { x: s ?(x)
_ r},
and grad 52 is Lipschitzian on {x: (x) r}.
(10) If aeA,then

Tan(A,a) = {u: lim


tt0 inf 6(a + tu) = O
t--O+

(11) If aCA, reach(A, a)>r>O, uCEE and


u * v_ 0 whenever t(a + v) = a, vi = r,

then
lim t-16(a + tu) = 0.
9-0+

(12) If aEA and reach(A, a)>r>O, then


Nor(A, a) = {Xv: X _ 0, I v I = r, t(a + v) =a,

Tan(A, a) is the convex cone dual to Nor(A, a), and


lim t-'6(a + tu) = 0 for u E Tan(A, a).
t40+

(13) If
N= {(a, v): a G A and v E Nor(A, a)},
a: N E, u(a, v)
E. a + vfor (a, v) E N,
#: U EnX En, #(X) = (Q(x), x-(x)) for x E U,
then
a(N) = En, a is Lipschitizian,

i/(U) C N, J/ is a homeomorphism, = O.,


f|(U).
If furthermore
K C A, 0 < r < q, reach(A, a) _ qfor a E K,
W = U nI {x: t(x) E K and b(x) ; r},
then

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
436 HERBERT FEDERER [December

#(W) = NC {(a, v): a E K and I v| < r},

V I W is Lipschitzian;
in case K is compact and 0< t < oo , then
N r {(a, v): a E K and l vI < t} is compact.
Proof of (1). Choosing aEA so that 8(x) = Ix-a i, one obtains
6(y)-S(x) < I y-a i - I x-al < I y-xi .
Proof of (2). Assume a = 0 and note that
vE Pif andonlyif ib-vi > vl forallb E A-{a},
v E Q if and only if ib- v > Iv for all b E A.
Furthermore
I b-vi 2- I V|2 = b * (b-2v) whenever b, v CEE,
and consequently, if b, v, wEE., s>O0 t>0, s+t=1, then
b- (sv + tw) 12- I sv + twi2 = b * (b - 2sv - 2tw)
= b * [s(b - 2v) + t(b - 2w)] = sb 0 (b - 2v) + lb 0 (b - 2w)
= s(i b - vi2 - I V12) + t(i b -W12 - I W12)
It follows that P and Q are convex, and clearly PCQ.
Finally suppose v EQ. If b A- { a}, then Ibi 2> 2b0 v, hence

v b <Ibi
b 2
This shows that v*u<0 for uETan(A, a).
Proof of (3). If aEA and b(x) = Ix-aI, then (2) implies
6[x + t(a - x)] = 8(x) -t(x) for 0 < t _ 1,
whence
x - a D8(x)(a - x) =
grad 6Wx * --) -bx 1.
8(x) -5(x)
Since |grad 8(x)1 $1, by (1), it follows that grad 8(x)=(x-a)/ (x).
Proof of (4). Otherwise there exists an e> 0 and a sequence x1, X2, X3, . . .
of points of U convergent to a point xE U and such that t(xi) - (x) > e for i
i = 1, 2, 3, * * * . Then

I t(Xi) -x, I = 6(xi), { (xi) -x I _ 6(x) + 21 xi


x- x
hence all the points (xi) lie in a bounded subset of the closed set A, and pass-
ing to a subsequence one may assume that the sequence (x1), (X2), (X3), *.

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
19591 CURVATURE MEASURES 437

converges to a point a(EA. But then


3(x) = lim 5(xi)=lim l(xi)-xxi
I -
= a-xI,
t-+o t-*

hence a-= g(x), which is incompatible with


a
a-t(x) lim I t(xi) - (x) | _e.

Proof of (5). According to (1) and (4) the right member of the equation
in (3) represents a continuous map of U-A into E.. Since the components of
the left member of this equation are Dl*(x), * * , Dj8(x), it follows from
Lemma 4.7 that a has continuous partial derivatives on W= Int(U-A).
In case xE W, the stated formula for grad 62(x) follows from the equation
in (3). In case xCA, 32(x+h) _ Ih| 2 for hE.E, hence grad 62(x) =0, and also
t(x) =x. Accordingly the formula holds for all xEInt U, and the continuity
of the right member, guaranteed by (4), implies the continuity of grad 82
on Int U.
Proof of (6). Assume lvi =1 and y=a+rveInt U. Then (4) and (3)
imply that (y) =a, a (y) =r, y EA, grad 8(y) =v.
In view of (5) one may apply Peano's existence theorem for solutions of
differential equations to obtain an r>O and a map
C: {s: - r < s < r} -* Int(U - A)
such that
C' (grad 8) o C and C(O) = y.
If 1sI <r, then IC'(s) =|grad 8[C(s)]| = 1 and
(Bo C)'(s) = grad 5[C(s)1 * C'(s) C'(s) 0 C'(s) = 1.

Accordingly, if -r <p <q <r, then

JI C'(s)I ds = f(8 o C)'(s)ds = b[C(q)]-[C(p)] < IC(q) -C(p)|


It follows that the curve C parameterizes a straight line segment in the
direction C'(0) = grad 8(y) = v.
If O<s<r and t=T+s, then
C(s) = y + sv = a + tv, S[C(s)] = 5(y) + s = I C(s) -a

hence t(a +tv) =a, with t >'r, contrary to the definition of T.


Proof of (7). Assume xqa and 'let
x- a
v S-{It:(a+tv)=a}.
I
- al

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
438 HERBERT FEDERER [December

Since Ix-aja S, sup S>0 and it follows from (6) that


sup S ? reach(A, a).
Moreover, if 0<tES, then
a + tv-bb b(a+ tv) = t, |a-b 2+ 2tv* (a-b) + t2 > t2

2tv* (a -b) >-| a -b |2) (x -a) * (a -b) -a


|ab 2 |xx-a
112t.
Proof of (8). Letting a = (x) and b=t(y), one infers from (7) that
(x - a) 0 (a - b) > - a-b 12 r/2q
and symmetrically
(y - b) 0 (b - a) Ib - a 12 r/2q.
Therefore
Ix-yI *Ia-b | (x-y) * (a-b)
=[(a - b) + (x - a) + (b - y)] 0 (a - b)
> a - b12(1-rq),
|x - y _ a - bl (q - r)/q.
Proof of (9). Combine (8), (1), (3) and (5).
Proof of (10). Suppose a =0 and Iu I = 1.
If uET(A, a) and e>0, then there exists a point bEA such that

0< IbI <e and | -u <,


Ibj
hence

I bl-l8(I b| u) ? I bKl1 Ibl u-bl = bl

On the other hand, suppose that whenever 0 < e < 1 there exists a number
t such that
0 < t <e and t-16(tu) < e;
choosing bEA so that 8(tu) = Itu-b , one finds that
|t- I b||8 b(u) <E, 0 < (1-e)t < IbI < (1 + E)t < E + C2,
b I |b- I b| ul < b-tul + It- |b| I 2Et 2=
I b u I b b
I (1-e)t 1-e
Proof of (11). Suppose a=0, Iu| =1 and
lim sup t-16(tu) > 0.
9o11+

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 439

Choose e and S so that


O< e < reach(A, a)-r, SC {t:O < t <E},

OE Closure S, 6(tu) > te for t E S.

If tES, then
6(tu) < tu| = 1< E< reach(A, a), tu E U.
For tES, O<p<r let
1(t, p) = tu + p grad b(tu) = t(tu) + [8(tu) + p] grad b(tu)
and observe that

I (t, p) I
< E+ r < reach(A, a), (t, p) E Int U.
It follows from (6), with a and v replaced by t(tu) and grad 8(tu), that
t[q(t, r)] = t(tu) whenever tE S.
{ r): tES} is bounded, one may assume, after replacing S
Inasmuch as J(t,
by a suitable subset, that there exists a point vEEn for which
lim (t, r) = v.

Then
|v| = lim jn(t,r)| =r, V U,

t(v) = lim t[X(t, r)] = a,

and consequently, by hypothesis, u @v?0.


Choosing tES so that
U 0 X(t, r) < Er,

one may use the fact that

to obtain
[8Qtu) + r]2 < tuu + [I(t, r) -tu] 12,

[8(tu)]2 + 2r8(tu) + r2 < t2 + 2tu * [b(t, r) -tu] + r2,


2r8(tu) < t2 + 2tEr- 22 < 2tEr,

hence 8(tu) <te, contrary to the choice of e and S.


Proof of (12). Since {a+v: |v| <r}CU, the set
S={Xv:X O0,Iv|= r, (a+v) = a}

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
440 HERBERT FEDERER [December

is closed. Clearly S is positively homogeneous. In order to verify that S is


additive, suppose
X >0, vj = r, t(a + v) = a, , > 0, w r, t(a + w) =a,
let
Z = (X + )-(XV + Aw),

and use (2) and (6) to infer that


t(a + z) = a, t(a + rI zI-'z) = a,
Xv + ,w = (I Xv + ,.w I r~')(r I z j-lz) E S.

Thus S is a closed convex cone.


Now let

L = {u: lim t-15(a + tu) = 0}

One sees from 4.5 that


Tan(A, a) C Dual[Nor(A, a)],
from (2) that
S C Nor(A, a), hence Dual[Nor(A, a)] C Dual(S),
and from (11) and (10) that
Dual(S) C L C Tan(A, a).
Accordingly
Tan(A, a) = Dual[Nor(A, a)] = Dual(S) = L, Nor(A, a) = S.
Proof of (13). One sees from (2) that if xCEE, aEA and (x)= Ix-aI,
then
x - a E Nor(A, a), (a, x - a) E N, oa(a,x - a) = x.
In case xEU, then a=-(x), +t(x)=(a, x-a), a[4(x)]=x. This implies the
first part of (13). The second part follows from (12), (2), and (8); in case K
is compact, so are W, 4/(W) and the image of ,6(W) under the transformation
mapping (a, v) onto (a, tr-lv).
If s > 0 and A8= {x:
4.9. COROLLARY. A (X) < S }, then
BA.(X) = 5A(X) - S whenever BA(X) _ s,
{A[tA.(X)] = {A(X) whenever 5A(X) < reach(A),
reach(A8) _ reach(A) -s.

Furthermore, if 0 <s <reach(A) and A,= {x: As(X)>S}, then

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 441

6A8(X) = S - 8A (X) whenever 0 < 8A (X) -< S,

A [UA(X) ] = (A (X) whenever 0 < 8A (X) < S,

reach(A') > s.
Proof. The formula for 8A. follows mechanically from the definitions, and
the formula for 8At may be derived with the aid of 4.8 (6). Then the state-
ments concerning reach and t can be obtained from 4.8 (5) and (3), applied
to A, A, and A,'.
4.10. THEOREM. Suppose
A antdB are closed subsets of En,
C is a nonempty compactsubset of A n B, r > 0,
reach(A, c) > r and reach(B, c) > r for c E C,
and there exist no c and v such that
c E C, v E Nor(A, c), -v C Nor(B, c), v % 0.
Let q be the infimum of the set consisting of 1 and the numbers
I v + wI
I VI + IWI
corresponding to vENor(A, c),wCNor(B, c),cEC with lvI+|w! >0. Then:
(1) 0<?<1 and there exists a r such that 0< ?r and

I X grad 5A(X) + ,u grad &B(X) I > (-/2)(X + ,u)

whenever xCEn- (AkUB), 5c(x) < , X>0, pL>0.


(2) 5AnB(X) < (2/n7)[5A(X) +6B(X) ] whenever 5c(x) <nq/5.
(3) If cCC, then
Tan(A n B, c) = Tan(A, c) n Tan (B, c),
Nor(A n B, c) = Nor(A, c) + Nor(B, c).
(4) If cEAf)B, O<p?<r7/2 and
AnB {z: Iz-cI < 2p} CC,
then reach(A rlB, c) > p.
(5) If C= AfnB, then reach(A nB) > rr7/2.
Proof of (1). For O?e<r let
S(E) = A G{a: 5c(a) < e, T(E) = B G{b: 5c(b) <?e,
M(E) = { (a, v): a E S(E), v E Nor(A, a), v I <1 },
N(e) ={(b, w): b C T(E), w C Nor(B, b), w| < 1},
P(E) = [M(E) X N(E)] n { ((a, v), (b, w)): a-b
a I ?, l vI + |wi = 1}

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
442 HERBERT FEDERER [December

observe that S(e) and T(e) are compact with


reach(A, a) > r - e for a S(e),
E

reach(B, b) > r - e for b E T(E),


and use 4.8 (13) to infer that M(e), N(e) and P(e) are compact. Furthermore
let A be the function on P(r) such that
A((a, v), (b, w)) = Iv + wI for ((a, v), (b, w)) E P(r),
and note that A is a continuous function, A does not vanish on P(O), and either
P(O) is empty or r7is the minimum value of A on P(O), hence O<- < 1. More-
over, since
P(O)= n P(>),
O<E<r

one may choose e so that 0 <e <r and the minimum value of A on P(e) ex-
ceeds q/2.
Let = e/2 and suppose
x EnE- (A UB), ac(x) <, X>O, ,>O.
Choosing a, b, v, w so that
aEA, A5A(X) = Ix-aI, bEB, 5B(X) = X-b|,
(X + ,u)v = Xgrad5A(X), (X +,u)w = ,grad6B(X),

one readily verifies with the help of 4.8 (2) that


((a, v), (b, w)) E P(e), hence v + w| > i/2.
Proof of (2). Letting
[(6A) 2+ (SB) ]i"/2

Q= {x: c(x) <.} - (A n B),


one sees from 4.8 (5) that Jt is continuously differentiable on Q. Furthermore

I grad A(x) I _ i/2 for x E Q.


In fact, if xCQ-(AkUB), then
grad i*(x) = [6A(X) grad AA(X) + 5B(X) grad 5B(X)]/ik(X),

Igrad At(x) I > (v/2) [6A(X) + 5B(X)]A&(X) _ ,1/2


by virtue of (1); on the other hand
grad A1(x) = grad B(X) for x E Qn A,
grad A1(x) = grad 5A(X) for x E Q r) B,
hence Igrad y6(x)| = 1 for x E QC(A UB).

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 443

Fix a point zeQ such that 5c(z) < 77/5 and consider the class of all maps
q: J -+Q
such that J is an open real interval containing 0,
q(O) = z and q' = - (grad i) o q.
Since this class is nonempty, according to Peano's existence theorem for solu-
tions of differential equations, and is inductively ordered by extension, it
has a maximal element. Henceforth let q: J->Q be such a maximal element.
If t J, then q'(t) -grad 1[q(t) ], hence
q'(t)ji 2/2, (iIoq)'(t) = grad 0 [q(t)] 0 q'(t)= - |q'(t) .

It follows that if 0 <uEJ, then


jP(z) = O[q(O)] >- 0[q(O)] - O[q(u)]
u pu

Iq'(t) _ (1//2) q'(t)Idt > u.2/4.


Cos l T I2dt L f
Consequently r = sup J < oo and there exists a point hE,, such that
lim q(t) = h,
T-T

I h - zI < f Iq'(t) Id = (2/t/)o(z)


< (2/r7) [5A(Z) + 5B(Z)].

The proof will be completed by showing that hkAnB. Otherwise, since


5c(h)< I h -z I + 5c(z) < [(4/nI) + 1]5c(z) < ?,

it would be true that hEQ, and Peano's existence theorem would furnish an
e>0 and a map
p: {: 7t - e < t< r+ e4 Q
for which p(r) = h and p' =- (grad '1)o p. Inasmuch as
lim q'(t) = lim - grad O[q(t)] =-grad j(h) = p'(r),

the map
P: J U {t: r < t < r + e4 >Q,
P(t) = q(t) for t C J, P(t) = p(t) for r < t < r + e,

would be a proper extension of q: J->Q with


P(O) = z and P' =-(grad J) o P,
contrary to the maximal property of q.

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
444 HERBERT FEDERER [December

Proof of (3). Inasmuch as


AfnB (X) _ SA(X) and AfnB (X) _ SB (X) for x C E.,
threefold application of 4.8 (10) yields
Tan(A n B, c) C Tan(A, c) n Tan(B, c).
On the other hand, if uGTan(A, c)6Tan(B, c), one infers from 4.8 (12) that
lim t-'A4(C + tu)= 0 and lim t-I'B(C + tu) = 0,
t >0+ t-40+

hence from (2) and 4.8 (10) that


lim t-'5AnB(c + tu) = 0, u E Tan(A 6\ B, c).

This proves the first equation in (3), and the second now follows from 4.5 and
4.8 (12).
Proof of (4). Suppose Ix-c| <p, zCAnB, 6AnB(X) = IX-z|.
One sees from 4.8 (2) that x-zC-Nor(ACnB, z). Inasmuch as

Iz - c| < I z - xI + I x - c I<?AnB(X) + I x - cI _ 21x - c| < 2p,


hence zC C, it follows from (3) that there exist v and w with
v C Nor(A, z), w C Nor(B, z), v + w = x-z.
Now n7(lvl + I wl ) < I v+wl = Ix-zl <p<rr7/2, hence
I 2v| < r and I 2w| < r.
Since reach(A, z) > r and reach(B, z) > r, one infers from 4.8 (12) that
tA(Z + 2v) = z and tB(Z+ 2w) = z.
Recalling that zEEAnB one obtains
A nB (Z + 2v) = z and tA nB(Z + 2w) = z,
and one concludes from 4.8 (2) that
Z = tAnB[Z + (2v + 2w)/2] = tAnB(X)

Proof of (5). Applying (4) to all cC C with p = rr7/2.


4.11. LEMMA.Supposef is a continuously differentiable real valued function
on an open subset of En, gradf is Lipschitzian, and
A = {x:f(x) = O}, B = {x:f(x) < O}.
If aEA and grad f(a)O 0, then 0 <reach(A, a) < reach(B, a).
Proof. Let M be a Lipschitzian constant for gradf, and choose positive
numbers h and r such that

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 445

I gradf(w) I _ h whenever |w - aj < r.


It will be shown that reach(A, a)>s=inf{r/2, h/M}.
Suppose lx-al <s, bEA, cCA,I b-xI =c-x I=A(x). Then lb-al <r,
Ic-al <r and Taylor's Theorem implies that
1f(c) -f(b) - (c
I c-b I2M/2. - b) 0 grad f(b) I<
Furthermore f(c) =f(b) = 0, and since x-b C Nor(A, b) according to 4.8 (2)
there exists a real number t such that
x - b = tgradf(b).
It follows that

I 2(c -b) * (b -x) I| I cs- b 2M| t |


0 = c -
x12
- lb - x2 = c - b12 + 2(c - b) O (b - x)
> I c - b2(1- MI tl )

hIM > x-b = It* I gradf(b) I > I t I) 1 > MI|t|,


hence Ic-bI2=0.
4.12. THEOREM. Suppose fi, * , fm are continuously differentiable real
valuedfunctions on an open subset of E, grad fl, . ,grad f. are Lipschitzian,
0_k<m, and
k m
A = n {x:fi(x) = O} n n {x:fi(x) ? O}.
i=1 i=k+l

If a GA, J= {i: fi(a) =0}, and there do not exist real numbers ti, corresponding
to iC J, such that ti 5zOfor some ie J, tiO> whenever ie J and i>k,

E ti grad fi(a) = 0,
iEJ

then reach (A, a) > 0 and

Nor(A, a) = { j ti gradfi(a): ti > 0 whenever i > k}.

Proof. Using 4.11 and 4.10, apply induction with respect to m.


4.13. THEOREM.Suppose e>o. If A1, A2, A3, * * * and B are closed subsets
of En such that reach(Ak)>e for k=1, 2, 3, * * * and
5Ak(X) -) 5B(X) uniformlyfor x E C as k -oo
wheneverC is a compact subset of {x: 5B(X) <e } then reach(B) > E and
(Ak(X) ->
-
B(X) uniformlyfor x E C as k - oo

whenever C is a compact subset of {x: 5B(X) <e }.

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
446 HERBERT FEDERER [December

Proof. Suppose C is a compact subset of {x: 5B(X) <el}. Choose an open


set W such that CC W and the closure of W is a compact subset of
Ix: 5B(X) <e}, a number r such that

sup {SB(x):xC W} <r<e,


and a positive integer K such that
SUp {I Ak(X): X C W} < r for k ? K.
It follows from 4.8 (8) that the functions (AkIW corresponding to k >-K have
the common Lipschitz constant e/(e-r), and hence from 4.8 (5) that the
functions (SAk)2 are equiuniformly differentiable on W. Since

5Ak(x)-> 3B(x) uniformly for x E W as k , co,


one infers that (5B)2 is uniformly differentiable on W and

grad SAk(x)-> grad &B(x)uniformly for x C W as k -> oo.


Finally one uses 4.8 (3) and (5) to conclude that WCUnp(B) and
tAk(x) -> tB(X) uniformly for x E W as k - oo.
4.14. REMARK.Observing that if A and B are nonempty closed subsets of
Ens then
sup I SA (X) - 8B (X)|
xGE
En

equals the Hausdorff distance between A and B, one sees from 4.13 that for
each e >0 the set
{A: 0 - A C En and reach(A) _ e}
is closed with respect to the Hausdorff metric. It follows that if E>0 and K
is a compact subset of En, then
{ A: 0 - A C K and reach(A) ? 4
is compact.
4.15. REMARK.The reasonable local behavior of a subset A of En, such
that reach(A) >0, is further illustrated by the following properties:
(1) If pEEn and 0< r <reach(A), then
A( {x: Ix -p I < r is contractible.
(2) If aCA, dim Tan(A, a) =k and
P(r) = A n {x: t x-al ?<r, Q(r) = Tan (A, a)nf {u: IuI < r
whenever r > 0, then dim (A) _ k and

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 447

Hk [P(r) ]
r
+Hk[Q(r)]
(3) Fork i * * * ,nthe set
F,o
A(k) = A ( {a: dim Nor(A, a) > n - k}
is countably k rectifiable.
(4) If dim(A) =k, then A =A (k) #A (k-i) and,for aEA -A (k-1) Tan(A, a)
is a k dimensional vectorspace.
To prove (1), consider the homotopy h such that
h(x, t) = -A [(I-t)X + tp]
whenever xCA, |x-pj |_r, O<t< 1.
To prove (2), assume a=0, let U be the k dimensional vectorspace con-
taining Tan (A, a), and consider the continuous maps
ft: Q(1) -> U, ft(U) = t 1(QU a A)(tU) for u E Q(1),
corresponding to 0 <t <reach(A). Inasmuch as

Ifs(X) - UI|= ''1 l {U[tA(tU) - tU] I

? -'l | A(tU) - tUl = t1- A(tU)

for uCQ(1), and since one easily sees from 4.8 (12) that t'15A(tU)->0 uni-
formly for uCQ(1) as t->O+, it follows that as t->O+ the maps fg converge
to the inclusion map of Q(1) into U, whence dim(A) _ k.
Given any e such that 0<e<1, one may choose p>0 so that if O<t<p
then
t-1A(tU) < e for u C Q(1),
Hk(ft[Q(1)]) > (1 - E)Hk[Q(1)].

One concludes that if 0<r<p and t=r(1+e)-', then


{A[Q(t)] C P(r),
ift[Q(1)] = (tuo A)[Q(Q)] C u[P(r)],
Hk[P(r)] > tkHk(ft[Q(1)]) > (1 + E) krk(l - E)Hk[Q(1)]

(1 + E)-k(1 - E)Hk[Q(r)].
To prove (3), let S be a countable dense set of k dimensional planes in
E, suppose 0 < r < reach (A), observe that
,
A(k) C U {A [ X
{X:&A(X) < r],
oES

and recall 4.8 (8).


To prove (4), first use (2) to infer that if aGA, then

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
448 HERBERT FEDERER [December

dim Tan(A, a) < dim(A) < k, hence a E A(k);

in case a A(kl), then


dim Nor(A, a) = n - k, dim Tan(A, a) k,
dim Nor(A, a) + dim Tan(A, a) = n,
hence Tan(A, a) is a k dimensional vectorspace. On the other hand (3) implies
that Hk[A(k-1) -0, hence dimA (k-1) <k-1 according to [HW, VII], and
consequently A ?A(k-I)A
4.16. LEMMA. For every nonempty closed subset S of En,
[6S(x)]2 + [SDual(S) (X)]2 > 1 2 whenever x C En;
furthermore S is a convex cone if and only if
[Ss(X)]2 + [3Dua1(S)(X)]2 = I XI' whenever x C En.
Proof. IfxCEnzuES, SS(X)= ix-ul ,theneither x-uI _ x| orx@u>o,
Dual(S) C Dual({u}) = {v: v 0 u ? 0},
D,ua1(S)(X) > 8Dual({u])(X) = (X 0 U)/ I 1
[6S(X)]2 + [6Dual(S)(X)]2 I> X - U 12 + [(X 0 U)/ IU | ]2

X
I12 +[ IU I - (X 0U)/IU 1]2 > I X12;

in case S is a convex cone it is also true that


S C Tan(S, u), x - u C Nor(S, u) C Dual(S),
{u,-u} CTan(S,u), (x-u)Ou=O,
I1 I X - U12 + I X - (X - U) 12 ? [&s(x)12 + [8Dual(S)(X) 12

and consequently the equation of the lemma holds.


To prove the converse, suppose S is a closed set such that the equation
holds whenever xCEn. Since the equation also holds with S replaced by
Dual(S), one finds that
SS(x) =Dual[Dual(S)](X) whenever x E En)
hence S= Dual [Dual(S) ].
4.17. LEMMA. If A is a closed subset of En, <t < oo, r > 0 and

5Tan(A,a)(b
- a) < I b - a 12/(21)
whenever a, bEA with 1a-b1 <2r, then reach(A)?> inf{r, t}.
Proof. Suppose 8A (X) <inf I r, t}, aCA, bCA,
&A(X) =I X-al = I x-b- ,

and assume a=0. Then 1b1? Ib-x1 + Ix| <2r,

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
19591 CURVATURE MEASURES 449

x C Nor(A, a), Tan (A, a) C Dual({x}) = {v: v 0 x ? O},


x
b 0 ~ -< aDual ({x})(b) < 3Tan (A ,a)(b) < b 2/21,
lx!
0 = |b - xl2 - Ix12 = 12 - 2b 0 x I b12(1 - I x| /1) > 0
unless b=0.
4.18. THEOREM.If A is a closed subset of E. and 0 <t < oo, then the follow-
ing two conditions are equivalent:
(1) reach(A) _ t.
(2) aTan(A a)(b-a) < Ib-al 2/(2t) whenever a, bCA.
A ccordingly
reach(A)-' = sup {2 b - a 1-25Tan(A,a)(b - a): a E A, b C A, a 5 bi,
where 0-1= oo and oo-0=0.

Proof. Applying 4.17 with r= 0o one finds that (2) implies (1).
Now assume (1) and suppose a=OCA, bCA. If vENor(A, a), then
v * (-b) _ - |b121 v| /2t
according to 4.8 (12) and (7), hence
-
b - v12 = b12 + I v12 - 2b 0 v_ Ib12 + I vl2- bl2l vI /t
> lb12 - lb 14/ (412)

Consequently
[SNor(A,a)(b)]2 _ | b12- I b 14/(4t2)
and it follows from 4.8 (12) and 4.16 that [8Tan(A,a)(b)]2 < bj 4/(4t2).

4.19. THEOREM.If A CE., reach(A) > t > 0, s > 0 and


f: {x: SA(X) < s} Em
is a univalent continuously differentiable map such thatf, f'l, Df are Lipschitzian
with Lipschitz constants M, N, P, then
reach[f(A)] ? inf {sN-1, (Mt1- + P)-1N-2}.

Proof. Suppose a CA, b CA and lf(b) -f(a) I< 2sN-].


Applying 4.18 choose uCTan(A, a) so that
l b-a-nI _ | b-al2/(2t).
Then Df(a) (u) CTan [f(A), f(a) ] and

IDf(a) (b - a) - Df(a) (u) I< MI b -a 12/(21).


Furthermore lb-al <2s, 3A[Xaa+(1-X)b]<s for 0?<X1 , hence Taylor's
Theorem implies that

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
450 HERBERT FEDERER [December

f(b) -f(a) - Df(a)(b - a) I < P b - a12/2.


Accordingly
aTan f(A),f (a)] [f(b) -f(a) ] < (Ml-1 + P) |b - a| 2/2
=<(Mi-1 + P) N2 |f (b) - f(a) | 2/2.
Use of 4.17 completes the proof.
4.20. REMARK.It may be shown that under the conditions of 4.15 (4) the
set A (k-i) is closed and the set A -A (k-1) is a k dimensional manifold locally
definable by equations fi(x) = 0, * * * ,fn(x)- =0, where fi, * * * , fn- are real
valued continuously differentiable functions with linearly independent
Lipschitzian gradients.
A related proposition states that a Lipschitzian map g: Em>En has a
Lipschitzian differential if and only if the subset g of EmXEn has positive
reach.
Among those subsets A of En for which reach(A) >0 the k dimensional
manifolds may be characterized by the property that Tan(A, a) is a k dimen-
sional vectorspace for each aCA.
If t>0, then the class of all k dimensional submanifolds A of En for which
reach (A) > t is closed relative to the Hausdorff metric; likewise closed is the
class of all subsets A of En such that reach(A) _t, dim(A) ?k and A is not a
k dimensional manifold.
4.21. REMARK.Suppose m_n, X is an open subset of Em, f: X->En is a
continuously differentiable map, f and Df are Lipschitzian with Lipschitz
constants M and P, and
Q=inf{Jf(x):xEX} >0.
If A CEn, reach (A) > t > 0, r > 0 and
Em.r Ix: rifI(A)(x) < r} C X,
then
reachf-1(A)] > inf{ r, QM1-n(M2t-1 + p)-1}.

5. The curvature measures. In this section several versions of Steiner's


formula are derived by a modification of the classical method of [W]; the
main innovation is the use of the algebra A**(E). By means of Steiner's
formula the curvature measures correspondingto a set with positive reach
are defined, and their basic properties are established. The proofs of the
cartesianproduct formulaand of the generalizedGauss-BonnetTheoremwere
partly suggested by [H, 6.1.9] and by [A; FE1].
5.1. If h: En->E, is Lipschitzian, VCEn, h| V is univalent,
LEMMA.
(hI V)-1 is Lipschitzian,aE V, En-V has Ln density0 at a, and h is differenti-
able at a, then Jh(a) > 0.

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
19591 CURVATURE MEASURES 451

Proof. Let M be a Lipschitz constant for h. Suppose a= 0,


u C kernel Dh(a), l ul = 1, 0 < E < 1,
and choose r>O so that lh(ru)-h(a)l <er and
Ln({x:IxI < r + Er} - V) < a (n)Enrn.
Then there exists a point v E V for which v - ru| < er, hence
|h(v)-h(a) | I h(v)-h(ru) I + I h(ru)- h(a) I _ MEr+ Er,
lh(v)-h(a)l MEr+ Er (M + 1)e
I v-al = r-Er l-E

In view of the arbitrary nature of E,this conflicts with the assumption that
(hi V)-1 is Lipschitzian.
5.2. LEMMA.Suppose:
(1) P is a k dimensional Riemannian manifold of class 1.
(2) 61, . . *, Ok are continuous differential 1 forms of P.
(3) el, . , ek, fit . . . fn-k g are Lipschitzian maps of P into En.
(4) For pEP, r, is the tangent space of P at p.
(5) C is a closed subset of En.
(6) Q is a bounded Borel subset of P, g(Q) C C, g I Q is univalent, (g Q)>
is Lipschitzian.
(7) If peQ,then
(01 I r), * , (Ok| 7) are orthonormal,
e(p) ,.. *, ek(p), fl (p), , fn-k(p) are orthonormal,
{ n-k
Nor[C, g(p)] C z,fj(p) z G En-k}
j=1

{ ~~~~~~~n-k \

S(p) = En-k n {Z: | Zj = 1 and E zjfj(p)C Nor[C, g(p)]}.


j=1
(8) If p EQ and g, fi, * fn-k are differentiableat p, then
k k
I
A [(dg r,) 0 ej(p)] is a positivemultipleof A (0i Ir,),
i=l t=1
k
G(p) = E [(dg |T) 0 ei(p)] (Oi|r7) c A (rp),
i=l
k
F,(p) = , [(dfj| rp) 0 ei(p)] 0 (Oi rp) c A11(rp) I forj = 1, , n - k,
i=l
m m f n-k
Um(P) = m!-J E zjFj(p) dHn-k-lZ c A,,n-"(-p) form = 0, 1, .. * * k.
Stp) j=l

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
452 HERBERT FEDERER [December

(9) 0 <r <reach [C, g(p) ] wheneverp EQ.


Under these conditions the following formula holds:
Ln({x: bc(x) < r and (c(x) E g(Q)})
k
=- rn-k+m(n - k + m)-1 trace[(k - m) !-G(p)k-mum(p) ]dHkp.
mn=OQ

Proof. Let h: P XEn-k-*En


n-k
h(p, z) = g(p) + E z,fj(p) for (p, z) E P X En-k,
j=1

z ~~~~~~n-k
V = (P X En-k){ (pZ): PE Q,I ZI < r,Z zjfj(p)E Nor[C,g(p)]},
W = { x: bc(x) < r, (c(x) E g(Q) },
and note that h is Lipschitzian. Using 4.8 (13) with A = C and K=g(Q), one
also sees that h( V) = W, h I V is univalent and (h I V)- is Lipschitzian. Further
let
Y: P X En-k ->P Y(p, z) = p for (p, z) E P X En-k,

and let Zi, * , Zn-k be the real valued functions on PXEn-k such that
Zj(p, z) = zj for (p, z) E P X En-k, j = 1,** *n-k.
Accordingly
n-k

h = (go Y) + > Zj (fjo Y).


j=1

If (p, z)eQXEn-k and T is the tangentspace of PXEn-k at (p, z), then


d Y maps T onto rT, inducing
A*
Y*: A *(rT)

and the linear functions


j
Y*(Ok| rp), dZ, T, T
,
Y*(61| , dZn-k j
form an orthogonal basis of Al (T). If g, fl, , nk are differentiable at p,
then h is differentiable at (p, z),
- ~~~~~n-k
T
(dh T) * ei(p) = Y* (dg Iri,) 0 ei(p) + , zj(dfj ITrv)0 ei(p)
j=1

fori=1, , k and
(d-j 7) f)=d T?Y d )f + f)n-k fI
(dh|I T) * f, (p) = (dZ, I T) + Y* (dg I Tp) *f(p) + ,: zj(dfj I Tv) I f,(p)2

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
19591 CURVATURE MEASURES 453

for s = 1, , n-k, hence


k n-k
A [(dhI T) * ei(p)]A A [(dhl T) * fs(p)]
i=1s=
k -n-k \ n-k
= Y* A (dgl7rp) * ei(p) + zj(dfjrp) 0 ei(p) )A A (dZsIT)
t=1 ,^~~~~~=1 =
k k
k\ n-k
= trace(k!-' G(p) + E zjFj(p) ) A (OiIrp) A A (dZ8I T),
and therefore

Jk(p, z) trace(k!-' [G(p) + E zjFj(p)] )

Now consider the case in which (p, z) C V and (P X Enk) -V has density 0 at
(p, z). It follows that if 0 < 1?,
_ then (p, tz) C V and (PXEn_k) - V has den-
sity 0 at (p, tz). Accordingly Lemma 5.1 implies that

0 < Jh(p, tz) = trace(k t' [G(p) + E tzjFj(p)])

for 0 <t < 1. Since the quantity inside the absolute value signs depends con-
tinuously on t, and is positive for t = 0 by virtue of (8), this quantity is posi-
tive for 0_t_1. One infers that in the formula for Jh(p, z) the absolute
value signs may be omitted, for Hn almost all (p, z) in V.
Using standard integral formulae and the binomial theorem one finally
computes

Ln(W) = JkJ(p, z)dIJ8(p, z)


r

= 'Q f' f' Jk(p, z)dHn-k-lzdtdHkp


Q S (p)

=j- ,J
rr
ptn-k-f1 Jh(p, iz)dHn-k-lzdtdHkp
Q O S~~~(p)
r r/ F n-k kc
tn-k1 f trace kk!- [G(p) + E tzjFj(p) dHn-kI-lzdtdHkp
Q(p) j=l

= trace t n-k-1 tm(k - m) !-1G(p)k-m


(p) m=o
mr!-n
[:
n-k m
zjFj(p) dHn-k-lzdt)dHkp

JQtrace[(k -m)
_k

-E rn-k+tn(n - k + m)-f !-1G(p)k-mum(p)]dHkp


m=o

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
454 HERBERT FEDERER [December

5.3. COROLLARY.Suppose
(1) P is a k dimensional submanifold of class 1 of En.
(2)
fi, * * * fn-k are Lipschitzianmaps of P into En.
(3) If pEP, thenfi(p), * * *, f-k (p) form an orthonormalbase of Nor(P, p).
(4) If pEP, then rp is the (intrinsic) tangent space of P at p.
(5) If pEP andfj is differentiable at p, then Fj(p) GA '1(rp) and the bilinear
form corresponding to Fj(p) is the second fundamental form of P at p associated
with the normal vectorfieldfj [mapping (u, v) ErpXr1ponto dfj(u) 0 dg(v), where
g: P->En by inclusion].
(6) C is a closed subset of En, PC C.
(7) If pCP, then
f ~~~~n-kX
S(p) = Enk n Z{:IZI = 1 and , zjfj(p) E Nor(C, p)}.
j=l

(8) If peP and fi, *. , fn- are differentiable at p, then

um(P) = m!- zjFj(p)] dHn-k-lz C Am


s(p) j=l
for m=O, 1, * , k.
(9) Q is a bounded Borel subset of P.
(10) 0<r<reach(C, p) whenever p(Q.
Under these conditions the following formula holds:
Ln({x: BC(X) ? r and {c(x) C Q})
k
= r n-k+n(- k + m)- ftrace[um(p)]dHkp.
mOQ

Proof. Since both members of the preceding equation are countably addi-
tive with respect to Q, the problem is local and one may assume, in view of
(2) and (3), that there exist Lipschitzian maps el, . . *, ek of P into En such
that if pEP, then ei(p), * * *, ek(p) form an orthonormal base for Tan(P, p).
Letting 01, . . *, ok be the 1-forms of P such that

OiI rp = (dg| rp) 0 ei(p) for p E P, i = 1, k,


one readily verifies that Lemma 5.2 is applicable; the factor (k - m) !-G(p)k-m
may now be omitted because the bilinear form corresponding to G(p) is now
the first fundamental form of P [mapping (u, v) GerXTp onto dg(u) *dg(v)].
5.4. DEFINITION. Suppose A C En and O<8A(P) < reach(A). Then
{
P = x: 8A (x) = 8A (p))} is an n-1 dimensional submanifold of class 1 of Ent
with the Lipschitzian unit normal vectorfield (grad 8A) j P, according to 4.8
(5) and (3). If rp is the tangentspace of P at p and (grad 8A) iP is differentiable
at p, then
VA(p) E Al 1(rp)

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 455

is defined by the following condition: The bilinear form corresponding to


gA(P) is the second fundamental form of P at p associated with (grad 8A)| P.
5.5. THEOREM. If A CEn, O <s <reach(A) and
A8 = {x:&A(X) < s}, As = {X: A(X) _ s}, P8 = {x: BA(X) = sI,

then the following three statements hold:


(1) If 0 <r <reach(A) -s and Q is a bounded Borel subset of Ps, then
Ln({x:BA.(X) < r and tA.(X) E Q})
n-1

= E rm+r(m + 1)!1 trace[IA (p)m]dHn-lp


m0Q

(2) If 0? r <s and Q is a bounded Borel subset of Ps, then


Ln({ X: BA'(X) < r and UA,(X) E Q})
n-1
= E 1r/+'(m + 1)!-'(-1)m trace[-(p)m]dHn-1p.

(3) If 0? r <s and K is a bounded Borel subset of En, then

Ln({x: 6A(X) ? r and tA(X) G K}) LnA8 nC {A(K)]


n-i
+ E (r - s)m+l(m + trace[:A(p)m]dHn-1p
m==
OntA t)

Proof of (1). Apply 5.3 with


P = P8, k = n- 1, fi = (grad PA) ,Pi, F, = -A |Pet C= Ae,
S(p) = {1i, um(p) = m! 1[:A(p)]i.

Proof of (2). Apply 5.3 with


Pi, k = n- 1, fl = (gradSA)| Pi, F = A |P8, c= A:
S(p) = {-1}, um(p) = m [-1 A(p)]

Proof of (3). Observe that

Ln[Ar r\ 1A'(K)] = Ln[Asr\ $A (K)] -Ln[(As- Ar) C\


n (K)],

use 4.9 to verify that


{x: r < 5A(X) < s and tA(x) C K}
= { Ax:AA(X)< s - r and tA (x) C P. Cn A (K)}
and apply (2) with r replaced by s - r.
5.6. THEOREM. If A CEn and reach(A) >0, then there exist unique Radon
measures 41o, 1i, * * *, 4n over En such that, for 0 <r <reach(A),

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
456 HERBERT FEDERER [December

n
L. ({x:6A(x) < r and {A(X) E K}) = E rn-ia(n-
i=O

whenever K is a Borel subset of En, and consequently


n
(fo0A)dLn = E rn ia(n - i) fdiVi
[X: 6A (X) r) i=o

wheneverf is a bounded real valued Baire function on En with bounded support.


Proof. Clearly '1o, * * *, I4/ are uniquely determined as soon as the above
equations hold for n+1 numbers r. On the other hand, if 0 <s <reach(A),
then measures 4{i suitable for 0 < r < s may be defined by letting a(n - i)41i(K)
be the coefficient of rn-i in 5.5 (3).
5.7. DEFINITION. If A CEn and reach(A) > 0, then the Radon measures
ko, 41i, . . I,Andescribed in Theorem 5.6 are the curvature measures associated
with A. Clearly the supports of these measures are contained in A.
Whenever f0o(A), '1(A), * *, t'Vn (A) are meaningful, for instance in case
A is compact, these numbers are the total curvatures of A.
Hereafter the dependence on A will be made explicit by writing
bi(A, K) for /'i(K), 4bi(A) = 'bi(A, A) for /'i(A),

'i(A,f) =f fd4i(A, *) for ffdqvi

whenever K is a Borel subset of En and f is a Baire function on En; further-


more (PiI(A, K) will be the total variation of bi(A, -) over K, and I(iI (A)
(Pi (A, A).
5.8. REMARK. If A CEn, reach(A) > 0 and K is a bounded Borel subset of
Ens then
M(A, K) = Ln(A nK), bi(A K) = b(A Kr Bdry A) for i < n.
The first equation is evident from 5.6, the second from 5.5 (3).
It is clear that if M is a rigid motion of En, then
Di[M(A), f] = (D(A I f o M)
for i = 0, * , n and every Baire function f.
If the conditions of 5.5 hold, then

)i(A8,f) = a(n - i)-1(n - i) h1 f(p) trace[ :A(p)n-i-1]dHn-1p

whenever i =0, , n -1 and f is a bounded Baire function on En with


compact support. Also, in case A is compact,
n ~~a(n-j) . , n,
(Di
(A,)= E ( I)si-i a(-) ) fori = 0,
j=o n- a(n-t

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 457

because if O?<r < s-reach (A), then


n
rn-ia(n -i)bi(A) = Ln({x: A (X) ? r})
i=O
n
= Ln({x: 6A(X) ? r + s}) = E (r + s)n-ia(n - j)4ij(A).
j=o
5.9. THEOREM. Suppose E>0. If A1, A2, A3, * * and B are closed subsets
of En such that reach(Ak)>E for k = 1, 2, 3, * * * and
6Ak(X) -* 6B(X) uniformlyfor x E C as k -
wheneverC is a compactsubsetof En, thenreach(B) > - andfor i = 0, 1, , n
the sequenceof Radon measures
,ci(A 1, *y4),j(A 2, * ), i(A 3, ')),

convergesweaklyto the Radon measureIb(B, *).


Proof. Recalling4.13, supposef is a continuousreal valued function on E.
with compact support S, and O<r<E, r7>O.
Let M= sup I If(x) :xES }, C= x: bs(x) _ r}, choose a number ? such
that 0<?<e-r and

L,[s 1(C)0 tx: r - < 6B(X) < r + P}]<11My


let D = CC {x: 8B(X) <r+?}, and choose a positive integer K such that if
k K, then
I SAJkX) - SB(X) |I< t for x EzC, |A Ak(X) - (B (X) |I< r for x E- D.
Let E== {X: 8B(X)<r}, Fk== {X: 8Ak(X)_ r}, and note that Er' 1(S) CC. If
k>K, then
Ek r) (Ak(S) = C rl Fk rl .Ak(S) C Dr -Ak(S) C

C f Er) CiA (S) C D r) (2(S) C (B (C),


C n [(Fk - E) U (E - Fk)] C {x: r - < B(X) < r +},
Ln(Ak(S) r [(Fk- C r E) U (C r E- Fk)]) < rl/M,

f 0
(f ? (Ak)dLn - (J Ak)dLn |< 77.

Since tAk(x)--B(x) uniformly for xC-CnE, one finds by first letting ko->
and then letting 77--O+ that

lim (fo0*Ak)dLn =(f o B)dLn (f 0 (B)dL,.


d k
cok nnEt

and consequently

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
458 HERBERTFEDERER [December
n n
imrnE n-ia(n - i)'i(Ak,f) = E rn-ia(n - i)Di(B,f).
kt-too0 i-=O i=O

Inasmuch as this equation holds for n+1 values of r, it follows that


lim bi(Ak, f) = 4'i(B, f) for i0, 1, * * *, n.

5.10. REMARK. One sees from 5.9 and 4.14 that if E>0 and i=0, * ,n
then the function on
{A: 0 $ A C En and reach(A) > e}
mapping A onto 4i(A, *) is continuous, with respect to the topologies of the
Hausdorff metric and of weak convergence. While the function mapping A
onto I(iI (A, *) is not continuous, it is true that if K is a compact subset of
En' then
sup {I 4iI (A): A C K and reach(A) a 4l < ??
because weak convergence of measures implies boundedness of their total
variations.
If A CEn reach(A) > 0 and A8= {x: 6A (X) ? S} for s > 0, then 4i?(A8, *)
converges weakly to 4i(A, *) as s->0+. Moreover, if K is a compact subset of
En and 0< t <reach(A), then
supi j 4(iI (A8, K): 0 < s ? t} < oo.
5.11. LEMMA. In addition to the hypotheses of 5.2 suppose:
(10) k=n-2.
(11) ,u and v are Lipschitzian maps of P into En.
(12) If peP, then 4.(p) and v(p) are linearly independent unit vectors and
= /(p) + v(p) i(p) -v(p)
= (p) -+ V(p) 1 ', f2P (p) _ V(p)
If(p)+v(p)I
1,4P 2~()
(13) If pCQ, then Nor [C, g(p)] is the closed convex cone generated by
,u(p) and v(p).
(14) If pGP and , v are differentiable at p, then
n-2
M(p) = ? [(dA| rp) 0 ei(p)] 0 (O| rp) E Al I(rp)
4=1

n-2

N(p) = z [(dp| rp) * es(p)] 0 (Oi| rp) E Al,1(rp).


i=l

(15) If a>O, b>O, m=O, , n-2 and j=O, , m, then


a(a, b) = E2fnl {z: IZ = 1 and zil/a> Z2/bl}1

Am,(a, b) = [(m -j) ?]-1 ( - )dH'z.

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 459

Under these conditions the following three statements hold:


For each pCQ where u and v are differentiable, and for m=O, , n-2,
m
Um(p) = Am,j[ | l(p) + V(p) I I |,U(p) - V(p) I ]M(p)m-iN(p)i.
j==O

For i =O, , n-2,


4J?[C,g(Q)] = a(n - i)- (n-i)' trace[i!-1G(p)iu,-2-i(p)]dHn-2p.

For i=n-1 and n, (i [C, g(Q)]=O.


Proof. First suppose pCQ, a 1, (p) +v(p) | b= | (p)-)V(p)! . Note that
if zCE2, then

z1fl(p) + Z2f2(P) = + b (p) + ( - 1 v(P)

and that zeS(p) if and only if z =1 and the above coefficients of ,u(p) and
v(p) are non-negative. Accordingly S(p) =a(a, b). In case ,u and v are differ-
entiable at p, one also finds that

Fj(p) [M(p) + N(p)]/a, F2(p) [M(p) - N(p)]/b


and consequently

ziF,(p) + z2F2(p) = + M(p) + ( - N(p)

whenever zeE2. Therefore the first statement follows from 5.2 (8) and the
binomial theorem.
The second and third statements may be obtained from the conclusion of
5.2 by computing the coefficient of r
5.12. COROLLARY.Suppose:
(1) Q is an n - 2 rectifiable Borel subset of En.
(2) g, ,u, v are Lipschitzian maps of Q into En, g is univalent, g-1 is Lip-
schitzian.
(3) g(Q) C CCEn reach(C)> 0.
(4) If pGQ, then ,u(p) and v(p) are linearly independent unit vectors and
Nor [C, g(p) ] is the closed convex cone generated by ,u(p) and v(p).
(5) 7qis a common Lipschitzian constant for g, ,u, v.
(6) For i=O, , n-2
d= a(n - i)-'(n - i) ( ) 2n-2-i(n - 2)n 2i,n-2.

Under these conditions it is true that

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
460 HERBERTFEDERER [December

I I [C, g(Q) ] ? dif [A


| ,(P) + p(p) K-1+ IA(p) - p(p) I-1]n-2-idHn-2p

for i=O, *,n-2,


and |"bil [C, g(Q)]=Ofor i=n-1, n.
Proof. In view of a standard decomposition one need only consider the
case in which Hn-2(Q) =0 and the case in which Q is contained in an n -2
dimensional submanifold P of class 1 of En.
In the first case Hn(Q XE2) =0, according to [F6, 4.2], and the Lipschitz-
ian function s1, such that
#(p, W) = g(p) + W1A(P) + W2v(P) for (p, w) E Q X E2,

maps Q X E2 onto a set whose Ln measure is zero and which contains


{x: .c(x) Eg(Q) }. Therefore I(i I [C, g(Q) ] = 0 for i = 0, 1, , n.
In the second case Lemma 5.11 is applicable, with
Ar,j(a, b) < [(mr-j) !j!]-(a- + b-l)m7r,
Al
M(p) |< (n-2)r,2 | N (p) I < (n-2)r7, | G(p) |< (n -2)r1

i ---G(p)iUn2-i(p) | <(
< [(n-2)r7]i [(|p) + v(p) I-i
t ~~~~j=0

I
+ (p)-V(p) 1-l]n-2-i7r ( i [( 2),X]n-2-i

= ( [(n - 2)7,]
n-27r2n-2-i[ | #,(p) + p(p)
I + I (p) - V(p) I-1]n-2-i

5.13. LEMMA. If A CEr, reach(A) >O,f is a bounded Baire function on Em


with compact support, and u is a bounded Baire function on {t: t_}0 whose
{
support is contained in t: t < reach (A) },then

f (f ? (A) ' (U O6 A)dLm

m-1 ra
- 4Dm(A,f)>u(O) + E bj(A,f)a(m
j=o
-j)(m -ij) tm-i-lu(t) dt.

Proof. Since the class of all functions u for which this equation holds is
closed to subtraction, addition, scalar multiplication and bounded conver-
gence, it need be verified only for the special case when u is the characteristic
function of { t: O t < r}, where 0 r < reach (A); but then it reduces to the
definition of the curvature measures.
5.14. THEOREM. For any closed sets A CEm and B CEn the following state-
ments hold:
(1) SAXB(X, y) = [6A (X) 2 +8B (y) 2 ] 112 whenever (x, y) GEmXEn.
(2) Unp (A X B) = Unp (A) X Unp (B) and

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 461

(AXB (X, y) = (A (X), B (y)) whenever (x, y) E Unp(A X B).


(3) reach[A XB, (a, b)] =inf{reach(A, a), reach(B, b) } whenever
(a, b)eAXB.
(4) If reach(A)>O, reach(B)>O and k=O, , n+m, then

4k(A X B, ) = E b(A, .) 0 j(B, *).


i+j=-k

Proof. The first three statements are easily verified.


To prove (4) suppose
0 < r < inf{reach(A), reach(B)},
f and g are continuous functions on Em and En with compact support, and
h(x, y) =f(x)g(y) for (x, y)CEmXEn. Using the definition of 4bk(A XB, h), the
Fubini Theorem, the definition of 1j(B, g), and Lemma 5.13, one obtains
m+n
Z rrm+n-ka(m + n - k)4k(A X B, h)
k=O

fJz,y): 6AXB(z,Y (h o (AXB)d(Lm ?D Ln)


(x,:
A(x):AB 6AT ()1) )r(
(g o (B)
6

(fo0A) (X) (y) dLnydLnX


{X: Y:x)r 5B(y)2 r- (X)2)
n r
EZ a(n - ij>I?(B g)J (f o0 A) (X)[r2 - 6A (X) 2] (n-j) I2dLrx
j=0 (X: 6A (x)s rI
n m-1
Z
E a(n
=~o
- j)j(B g) bn(A2f)rn-i +
E

i=o
(m - i)a(m - i)4i(A,f)

*tm-i-(r 2 - 12) (n-i)/2dt)

n m-1 n

Z rn-ia(n
j=O
- j)4Dn(A,f)cDj(B, g) + E
i=O j=O
E rn+n-i-i

J um-i-l(l - U2)(n-i)/2dua(n - j)a(m - i)(m - i) i(A,f)cj (B, g).

Now in the special case when A and B consist of single points of E, and
Eq, and whenf and g are the characteristic functions of A and B, the preced-
ing formula reduces to

rP+qa(p + q) = rP+q UP-'(1 - u2) qI2dua(p)a(q)p.

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
462 HERBERT FEDERER [December

Returning to the general case one may use this equation with p = m -i
and q = n-j to conclude that
m+n
F r".m+n-k a(m + n - k)4k(A X B, h).
k=O
m n
= Z rm+n-i-ia(m
i=o j=O
+ n - i - j)4?i(A,f)4)j(B, g).

5.15. REMARK.Applying Theorem 5.14 to the special case in which B


consists of a single point b, one finds that
4?k(A X {b}, X X {b}) = 4Dk(A, X) for k = , m
44k(A X {b}, XX {b}) = O for k = +1,** * ,m+n,

whenever X is a bounded Borel subset of Em. Accordingly the curvature meas-


ures behave naturally under an isometric injection of one Euclidean space into
another.
5.16. THEOREM. Suppose A and B are nonempty closed subsets of En and
s = inf{reach(A), reach(B), reach(A U B)}.
Then the following statements hold:
(1) If xCEn, then 3AUB(X) = inf {A(x),
I SB(X)},

AfnB (X) _ Up{A


I (X), B (X)}.

(2) If xE Unp(A UB) and SA (X) _B(X), then


x E Unp(A) and {A(X) = {AUB(X).

If x E Unp (A UB) and SB (X)-<?A (X), then


x E Unp(B) and {B(X) = (AUB(X).

(3) If xE Unp (A) and &B(X) ?_ A (X) < reach(A UB), then
x E Unp(A n B) and (A(X) = (AnBf(X).

If xCUnp(B) and SA(X) _ SB(X) <reach(A UB), then


x E Unp(A n B) and {B(X) = {AfnB(X).

(4) If sup {I A (X), SB (X) } < S, then


5AnB(X) = SUp {SA (X), 5B)J, x E Unp(A n B),
{ A(X), {B(X)J = {AUB(X), {AnB(X) I.

(5) reach(AfnB)_s.
(6) If s >O and i =O,* n, then
bi(A. ) + bi(B. ) = bi(A U B, ) + (b(A r) B,*)

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 463

Proof. Note that (1), (2) are obvious, and that (4), (5) are trivial conse-
quences of (2), (3).
In order to prove (3) one must show that if
x C Unp(A), a = {A(X), 5B(X) < 5s(X) < reach(A U B),
then a EB. Observe that
{A[a + t(x-a)] = a and a + t(x-a) Int Unp(A U B) for O< It 1.
The assumption that a EEBwould imply that
5A[a+ t(x - a)] < 5B[a + t(x - a)] for small t > 0,
0 < r = sup{ t: {AUB[a + t(x - a)] = a),
and it would follow from 4.8 (6) that T> 1, hence (AUB(X) =a; but then aEB,
because {AUB(X) =(B(x) accordingto (2).
Next, to verify (6), suppose f is a bounded Baire function on E. with
bounded support, 0 <r <s and
Ar = {x:&A(x) ? r4, Br = {x: 5B(x) < r.
Using (1), (4), (2) one obtains
ArU Br= {x: AUB(X) _ r}, Ar Br= {x: bAnB(X) 4r},
n
E rn-ia(n
i-O
- i)[bj(A,f) + ,i(Byf)]

= J (fo0 {A)dLn + f(f o {B)dLn

= fArBr ( A)dLn + fB (f O B)dLn + fr[(f tA) + (f 0 (B)]dLn


A,-Br B,-Ar ArnBr

fA B (f0 tAUB)dLn + (fo 0AUB)dLn


A,-Br B,-Ar

+ f [ AUB) + (f 0AnB)]dLn
r,nBr

= LrBr(f 0AUB)dLn + fA (f o {AnB)dLn


ArUBr r,nBr

= E r-ia(n - i)[cD(A U B,f) + Ii(A r B,f)].


i=O

5.17. REMARK. The additivity property expressedby 5.16 (6) is a sharper


version of certain properties studied by Blaschke [BL, ?43] and Hadwiger
[H, 6.12], who used these properties together with invariance under rigid
motions and continuity (compare 5.8 and 5.9) to characterize Minkowski's
Quermassintegrale

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
464 HERBERT FEDERER [December

W (A) = a(i) ( n)-i(A)

for compact convex sets A. It would be very interesting to know whether


there exists a similar characterization of the curvature measures 4i(A, *) for
all sets A such that reach(A) > 0.
5.18. REMARK. The proof of 5.19 will make use of the following classical
proposition:
Suppose V is a bounded subregion of E., the boundary of V is the union
of finitely many disjoint n-1 dimensional submanifolds of class 1 of En, f is
a real valued continuously differentiable function on a neighborhood of the
closure of V, and at each point of the boundary of V the exterior normal of V
and the gradient of f have a positive inner product. Then the Euler-Poincare
characteristic of the closure of V equals the degree of the map

(gradf) Clos V: (Clos V, Bdry V) -* (En, En - {O}).


Furthermore, if Wi, * , Wk are the components of Bdry V, then the above
degree equals the sum of the degrees of the maps
(gradf) Wi: Wi--E -En {

corresponding to i = 1, *.. , k.
Replacing f by a nearby function of class 2 and with nondegenerate criti-
cal points, one may derive this proposition from the Morse theory (see [M,
Chapter VI, Theorem 1.2, p. 145]).
5.19. THEOREM. If A CEn, reach(A) >0 and A is compact, then 4Do(A)
equals the Euler-Poincare' characteristic of A.
Proof. Suppose O<s<reach(A). Since 4Do(A8)=1?o(A),according to 5.8,
and since A is a deformation retract of A8, it is sufficient to show that 4Xo(A8)
equals the Euler-Poincare characteristic of A8.
Applying 5.18 with f-=(6A)2, and observing that
grad bA(x) = 2s grad 8A(X) for x E P.,
one sees that the Euler-Poincare characteristic of A. equals the sum of the
degrees with which
(grad 6A)IP8: P, -V= {V:lVi =
maps the components of P8 into V; furthermore this sum may be computed
by integrating the Jacobian of the above map over P8 with respect to Hn-1,
and dividing by na(n).
Consider a point pGP8 where (grad SA) IP8 is differentiable. If one identi-
fies the tangentspace r, of P8 at p with the "parallel" tangentspace of V at
grad 8A(P), the differential of (grad 8A)| P8 at p becomes the endomorphism

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 465

of r, corresponding to the bilinear form ZA(P), and using 2.12 (5) one finds
that the Jacobian determinant equals
trace[(n-)!-lA()f]

Accordingly the Euler-Poincare characteristic of A. equals

Ina(n)=]-trace( f [(n1t -1r[ ) A (p) n-1]dHkp


P8

=a<(n)-1n1-1 J trace[sA(P) n-l]dHkp = (Do(Aq)

by virtue of 5.8, as was to be shown.


5.20. REMARK. Suppose F1, * * *, F, are elements of an associative and
commutative finite dimensional algebra over the reals, and let

7r(m) = j ( E zjF) dH1-1z


sln{z: lzl=l) j-1

for m=0, 1, 2, * . . Clearly q(m) =0 in case m is odd, and

() = H-1(E r { z: Iz = 1}) = la(l).

If m is a positive even integer, then


1 3 m
r - 1 m/2

(r)=la (l) y 1+ ..2


Fj

In fact, Green's formula implies that


I I \ ~~~~m-1
(rn) = F
f)Fi ( zjF1) zidH'-lz
i=1 Elnz. l zl=l} j=l
I \m-2
~~~~~~~~~I
= Z FsfF)
i=1 ln{z: lzl<l}
(m - 1) E
=
zjFj) FdLjz

= (m-1) i (Ff)2 J I= ( E zjFj) dH-lzdr


i=l o El~~n{z:lzl=r j=1
11

= (mr- 1) (Fi) 2 rm-2+1-17(m - 2)dr

rn-i
=+ rn -2
I + m -2
Z (F )27(M -2).
j=1
5.21. REMARK. Consider the special case of 5.3 where C is a k dimensional
submanifold of class 2 of E., and P is open relative to C. If pEP, then
S(p) = En_kn I |z:IZI = 1i2

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
466 HERBERT FEDERER [December

hence um(p) may be computed by means of 5.20. Applying the formula

'Pi(C, Q) = a(n -i)-(- i)-l f trace [uk-i(p) ]dHkp

for i=0, 1, * * *, k, one finds that


(Pk(C, Q) = Hk(Q), (bi(C, Q) = 0 in case k - i is odd,

and that, if k - i is even and positive, then

Is(C, Q) = a -l( - n-i)-l trace {(k -i)!-la(n- k) (n - k)

1 3 k-i-1 ( [Fj(p)]2 dH p

-n-k n(k--+)2/2
(2k-ir(k-i) I2[(k i)/2] )-l trace {(E [Fj(p)]2) } dHkp.

Accordingly the curvature measures 4bi(C, *) are the indefinite integrals,


with respect to Hk, of certain scalars algebraically associated with the tensor
n-k

E [Fj(p) ]2 EEA2,2 (,r )


j=1

Furthermore this tensor may be identified, except for a factor - 1/2, with
the classical covariant Riemannian curvature tensor of C. In fact, define
el, . . , ek and 01, * , Ok as in the proof of 5.3 and let

ek+j=fJ forj = 1, * * ,n-k.


Using the familiar notation of hlieCartan (see [CA; C3]) one obtains
dfj *e8=dek+j*e8=wk+j,8 forj= 1,2**,n-k and s= 1,2*, k,
k k
Fj = E(dfj
8=1
* e8) O 08= E
8=1
k+j,8 ? 68 forj= 1,* * *,n-k,
n-k n-k k k

E (F.)2 = E E E (wk+j,8 A Wk+j,t) ? (08 A Ot)


j=1 j=1 8=1 t=l

E
k
E k

(
/n-k\

+ Wk+j, A k+j,t) ? (08 A Ot)


8=1 t=l j=l
k k

E E (-Qs,t) ? (08 A Ot)


8=1 t=l
k k k k
- - uv(OuA Ov)0 ( A
A0t).
=E E E E
s=1 t=i U1 v=1 2

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 467

Computing the trace of the (k-i)/2th power of this tensor one arrives at
2(k-i)/2 [(k-i)/2]! times the scalar Hk-i introduced in [WE]. (Note: Weyl's
R4 is the negative of Cartan's R. t,X V). In case k is even and i = 0 the above
formula for 40o(C)reduces, in view of 5.19, to the Gauss-Bonnet Theorem of
[A; AW; Cl; FE1].
5.22. REMARK. Assuming A CE. and reach(A) >0, let
*i'(A, f) = lim I (P I ({ x: AA(X) ? t}, f)
t-4O+

whenever i = 0, , n and f is a continuous real valued function on En with


compact support; from 5.5 one sees that for i <n this limit equals

ae(n - i) 1 f (fJo {A)(p) | ( + 1)h1- \(S)m+1n+i


P, ~~m-n-l-in-

trace[ZSA(p)m] dHn-1p.
Evidently 'i'(A, f) > 4i I(A, f).
Under the conditions of 5.21 one finds that
'i(C, Q) = 0 in case i > k,
and that if i< k, then
*i(C Q) = a(n -i)-)(n - i)-(k -i) !1-

fQfEkflz:
z=1)trace( zjFj(p)] ) dInklzdHk p.
QE-knz { lz|l=,) }
E
j=l

The total absolute curvature 'o(C, C) has been studied in [C3] and [CL],
and previously for k =1 in the theory of knots (see [MI; FE2]).
6. The principal kinematic formula. Within the following proof of this
integralgeometric formula, concerning two subsets A and B of En with posi-
tive reach, one may distinguish three component arguments: First, struc-
tural considerations (6.1, 6.2, 6.3, and Parts 1, 2, 3, 10, 11, 18 of 6.11) designed
to establish qualitative properties of the intersections of A with the isometric
images of B. Second, a most delicate convergence proof (6.3, 6.5, 6.10, and
Parts 3, 4, 5, 6, 7, 8, 16 of 6.11) showing that in computing the kinematic
integral one may approximate A and B by
Ar = { X: A(X) r?
< and Br= {XS:B(X) < r
Third, computations (6.6, 6.7, 6.8, 6.9, and Parts 9, 12, 13, 14, 15, 17, 19)
dealing mainly with Ar and Br. In these arguments the theory of Hausdorff
measure and rectifiability combines with the results of ??4 and 5 to furnish the
foundation, the integral formula 3.1 reduces the global analytic problem to
a local algebraic problem, and the tensor algebra A **(E) solves the local prob-
lem.

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
468 HERBERT FEDERER [December

Supposef: Em )En is a Lipschitzianmap, S C {x:f(x) = 0 1,


6.1. LEMMA.
k is an integer, and for each aGS there exists a k dimensional plane P such that
aCP and f IP has a univalent differential at a.
Then S is countably m - k rectifiable.
Proof. In view of [F4, 4.3] it is sufficient to show that if a and P are as
stated above, then there exist positive numbers r and q such that
So {x: I x-al < r and Ixa- +al > (1 +/2)16p(x)}
is vacuous.
Let M be a Lipschitz constant for f, choose positive numbers r and s
such that
|f(p) s
Isp-a a whenever pCP and p-a l < r,
and take = M/s. If xES, Ix-al <r and p=#p(x), then
I x-pI = 5P(x), I x-aI2 = I X-pl2+ I pal2,
si p-a l < lf(p) l = l f(x)-f(p) l < M l x-pl
IP
p- 2<_ 2 x - p12, I - 2 < (1 + n72)Ix - p12.
6.2. LEMMA. Suppose
X and Y are separable Riemannian manifolds of class 1,
dim X = p, dim Y = q,
f: X X Y -* E. is a Lipschitzian map,
S C { (x, y): f(x, y) = O}, k is an integer,
and for each (a, b) CS the map
fa: Y E., fa(y) =f(a, y) for y E Y,

is differentiable at b and dfa maps the tangent space of Y at b onto a k dimensional


subspace of En.
Then S is countably p +q - k rectifiable.
Proof. Using coordinate systems, it is easy to reduce the problem to the
special case in which X=E, and Y=Eq.
For each (a, b) CS there exists a k dimensional subspace V of the tangent
space of Eq at b such that dfa is univalent on V, and to V corresponds in obvi-
ous fashion a plane P CE, XEq such that (a, b) GP and fl P has a univalent
differential at (a, b).
6.3. LEMMA.If X is a separable p dimensional Riemannian manifold of
class 1 and
-
IA: X En) v: X -+En () {" I u I 1}

are Lipschitzian maps, then


(X X Gn) n { (x, R): ,u(x) + R[v(x)] = O}

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 469

is countably p + (n- ) (n-2) /2 rectifiable.


Proof. The map f: X X Gn-*En,
f(x, R) = ,u(x) + R[v(x)] for (x, R) C X X Gn
is Lipschitzian. Furthermore for each xEX the map
f,: Gn- E, fx(R) = f(x, R) for R E G.
is analytic and df, maps the tangent space of Gnat any RCG, onto an n -1
dimensional subspace of En; in fact f, is obtained by translation through the
constant vector ,u(x) from a classical fibre map of Gnonto the n -1 sphere.
Accordingly Lemma 6.2 applies, with q=n(n-1)/2 and k=n-1.
6.4. LEMMA.If vEEn, wCEE, IvI=1wW= 1, m<n-1, then

I v+ R(w) 1-mdo,nR < oo.

Proof. Letting S=EnC\{x: |x| =I} one finds (see [F4, 5.5]) that

Hn-I1(S)f Iv + R(w) 1-mdPnR= f | v + x j-mdHn-lx.


ans

Furthermore let C(r) = SO x: Iv+x { I< r} for r >0, and observe that there
exists an M< oo such that
Hin-1[C(r)] < Mrn-1 whenever r > 0.
Consequently

| v + x 1-mdHn-lx= E f Iv + x 1-mdHn-lx
00 00

2imM2(1-i)(n-1) = M2-1 E (2m-n+l)i < 0o


i=O i=O

6.5. REMARK. Clearly the integral considered in Lemma 6.4 is independ-


ent of v and w; denote it by Im.
In case m is an integer it follows from the binomial theorem and Holder's
inequality, with exponents m/j and m/(m -j), that

fGj!| v + R(w) K + Iv - R(w) 1-1ImdknR


in

< (E ) (Im)ilm(Im)(m-f)/m = 2mIm.

Similarly one sees that if Am,jis defined as in 5.11 (15), then the integral

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
470 HERBERT FEDERER [December

cmj = fJ j v + R(w) , | v - R(w) 1.1v A R(w) Id4,,R


Gn

is finite and independent of v and w.


6.6. LEMMA. If A CE, B CE, reach(A) >0, reach(B) >0, reach(AfnB)
> 0 and C is a bounded Borel set contained in the interior of B, then
(bi(A 0 B, C) = (bi(A 2 C) for i = O, 1, * * * , n.
Proof. In case C is a compact subset of Arlnt(B), SAnB(X) = SA(x) and
(AnB(X) =sA(X) whenever x is sufficiently close to C.
6.7. LEMMA. Suppose:
(1) V and W are vector subspaces of En.
(2) k=dim V+dim W-n>O.
(3) Q is the space of all isometries w such that the domain and the range of X
are k dimensional vector subspaces of V and W respectively.
(4) G, H are the orthogonal groups of V, W.
(5) ,u, v are the Haar measures of G, H such that Au(G)= v(H) = 1.
(6) U is a k dimensional real vector space.
(7) e: U-* V and f: U-*W are isometric embeddings.
(8) V' and W' are the orthogonal complements of V and W in En.
(9) P is the orthogonal projection of En onto V'.
(10) X is a real valued jn summable function such that, for RCGG, rq(R)
depends only on P o RI W'.
(11) ? is a real valued continuous function on U.
Then

,qr(R)*t[R-' I V 0 R(W)]dOnR
an

= fnd+ fGdn r(h of o e-1 o g1)d(A ? v)(g, h).

Proof. The group G X H operates transitively on Q according to the rule


(g, h) -W = h owWo g-1 for g E G, h C H, CE U.

Since f o e-1CO, a Haar measure 41over Q, invariant under the operation of


GXH and with 4+(Q)= 1, is given by the formula

A = fX (h of o e-1 o g-1)d(, ? v)(g, h)


aXH

for every continuous real valued function r on U.


With gEG associate A(g)CGn so that A(g)| V=g and A(g)| V' is the
identity map of V'.
With hCH associate B(h)CGn so that B(h) I W=h and B(h) I W' is the

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 471

identity map of W'.


Then G X H operates on Gnaccording to the rule
(g, h) *R = A(g) o R o B(h)-1 for g c G, h E H, R E G.,
the measure 0,,, the function q and the open set
M = G. n {R: dim[V n R(W)] = k}
are invariant under the action of GXH, and the continuous map
u: M -* , u(R) = R-11 [V n R(W)] for R E M,
commutes with the operations of GXH on M and U. Therefore another
Haar measure 'I over Q, invariant under the operation of G X H, is given by
the formula

()= a(ou)dJO
for every continuous real valued function r on U. Using the uniqueness of
Haar measure and the fact that 0 -(G -M) =0 one concludes that

= Q)P=
( fGdOn 6.
an

6.&8.LEMMA. If
U, V, W are finite dimensional real vector spaces with inner products,
G, H are the orthogonal groups of V, W,
,u, v are the Haar measures of G, H such that ,u(G) = v(H) = 1,
e: U--V and f: U-*W are isometric embeddings,
M C: APP(V), N E Aqq(I, p + q _ dim U,
then

trace[(g o e)*(M) *(h of) *(N)]d(gs 0 v)(g, h)


GXH

/p+ q\ /dimU\ /dim V -1 dim W\


- (P )+ )( V ( W) trace(M)trace(N).
Proof. Denote the above integral by F(,M, N) and observe that F is a
bilinear function invariant under the endomorphisms of AP P(V) X Aq q(W)
induced by GXH. Applying 2.13 twice one infers that there exists a real
number c such that
F(M, N) = c trace(M)trace(N) for M E AP"P(V), N EEAqq(W).
To determine c, choose
IU E A1'1(U), Iv E Al'1(V), Iw E All'(W)

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
472 HERBERT FEDERER [December

so that the corresponding bilinear forms are the inner products of U, V, W


and let
M = (IV) , N = (IW)q.
Then
(g o e)*(M) *(h o f)* (N) = (IU)P+q for (g, h) E G X H,
and it follows from 2.12 (4), applied with k = O and M= 1, that
dim(U) ! dim(V) ! dim(W) !
[dim(U) - (p + q)]]! [dim(V) - p] ! [dim(W) -q]
6.9. REMARK.Suppose X and 41 are bounded Baire functions on En and
IA is a Radon measure over En.
If 4+and either X or,u have bounded supports, then

'EJXw f x (i6 o R-1 a T_z)d,id(Ln 0Dpn)(z, R)


EnXGn

- f f ~x(x)f 4[R-(x - z)]dLnzdcixdcPnR


GEn En

-J f f:x(x)
J f (y)dLnyd1ixdInR= f d, f dLdn.
GEn En

Similarly, if X and either 4+or,u have bounded supports, then

fESXGff C(x o T,o R) . t'dcd(Ln0 q5On)(z,


R) = f XdLnfPdIA.
EnXGn

If X has bounded support and S is a bounded Borel subset of En, one may
apply the first formula with i,breplaced by the product of b and the character-
istic function of S, to obtain

'XG~ f x *(1to R-1 o Tz)dAud(Ln0) qn) (z, R) = f Xd/ * f4dLn.


EnXGn (Tz ? R) (S)S

If X and y have bounded support and S is any Borel subset of En, one may
apply the second formula with X replaced by the product of X and the char-
acteristic function of S, to obtain

LnXGn
f(iR- 0 T_) (S)(x o Tz o R) . t'dcd(Ln0 q!}n)(z, R) = f xdLnf t'di/d.
6.10. LEMMA.If A CEn, B CEn, A is closed, B is compact and
C(t) = { X: A(X) < t and SB(X) < t}

for t>0, then 8C(t)(X) ->SAnB(X) uniformly for xEEn as t->O+.

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 473

Proof. If t > 0, then A nB C C(t), 5AnBf(X)- _ C(t) (X) for x-En


Suppose E>0, let D = {x: 5AnB(X) < 4 , and observe that the sets C(t)-D
are compact and their intersection is empty. It follows that if t is sufficiently
small, then C(t) CD, hence
5C(t)(X) ? 5D(X) ? bAnfB(X) - E for x E En.
6.11. THEOREM. Suppose
AC En) reach (A) > 0, B C En, reach (B) > 0, B is compact
and i=0, 1, * * *, n. Then:
(1) For Ln (04Onalmost all (z, R) in EnXGn,
reach[A ( (Tz2o R)(B)] > 0.

(2) If X, q/ are bounded Baire functions on En and X has compact support,


then

I (i[A r) (T2 o R)(B), x(4 o R-1 aoT2) ]d(Ln0 qn)(z, R)


EnXGn

-=~Z E '(n, k, l)c14(A, X)cIl(B, 0


k+l=n+i

(3) If K is a compact subset of En, then

Iw | bi|[A r)\ (Tz o R)(B), K]d(Ln08 On)(z) R) < oo.


EnXGn

(4) If A is compact, then

frflxf(Di [A r) (Tz O R) (B) ]d(Ln 0 q5n)(z, R) = tyy(n, k, l)c4k(A)c1i1(B).


EnXGn k+l=n+i
Proof. For r>0 let
Ar = {X: A (X) ? r}, Br = { y: B(y) <r},
Vr = {x: 5A(X) = r}, Wr = { y: SB(Y) = r}.

Choose s and p so that 0 <s <p <inf {reach(A), reach(B) }.


For t>O let Zt be the subset of V.XWSXGn consisting of those points
(x, y, R) for which

Igrad 5A(X) + R[grad 5B(y)] I < I or I grad 5A(X) - R[grad SB(Y)] | < t

For r0 and R E Gn let ;r,R V.XW )-En


s-r r [s-r r]
r,R(X, Y) = (A(X) + -x - R (y) + Y
s s s s

For r ? 0 let Pr: V8 X W X Gn-En X Gn,

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
474 HERBERT FEDERER [December

r(X, y, R) = (rr,R(X, y), R).


In case i ?n-2 define di as in 5.12 (6) with

V = sup {p(p -- s)', [1 + p(p - s)-]s1}

and let
u(x, y, R) = di( I grad bA(X) + R[grad 6B(y)] 1-'
+ Igrad bA(X) - R[grad 6B(Y)] 1-1)n2-'-

for (x, y, R)EV.XW8XGn; in case i>n-2 let u(x, y, R)=O.


Throughout the following Parts 1 to 7 let K be any compact subset of En and
let U= {x: 3K(x)< 3p+diam(B)}fI V8.
PART 1. Zo is countably (n+2)(n-1)/2 rectifiable, and

(Ln 0 4n)[Pr(ZO)]= 0 for r > 0.

Proof. Applying 6.3 with X= V8XW8, p=2(n-1),


,u(x, y) = grad bA(X), v(x, v) = ? grad 5B(y) for(x, y) C V8 X W8,
one finds that ZOis countably k rectifiable, where
k = 2(n- 1) + (n- 1)(n- 2)/2 = (n + 2)(n - 1)/2.
Since ;r is Lipschitzian, by 4.8 (8), it follows that
Hk+1[k(Zo)] = o.

Furthermore Oknis proportional to the n(n - 1)/2 dimensional Hausdorff meas-


ure over Gn, hence Ln?n is proportional to the
n+n(n- 1)/2 = n+ (n-1) + (n- 1)(n- 2)/2 = k+ 1
dimensional Hausdorff measure over E. X Gn.
PART 2. If 0 < r <p, (z, R) GE. X G. and
s -r r -
g(x, y) = - A(X) +-X for (x, y) E r,R{ Z},
5 5

then g(r{ z1)= VC(T, oR)(Wr), g is univalent and Lipschitzian with


Lipschitz constant p/(p-s), and g-I is Lipschitzian.
If A,C\(T.oR) (Br) meets K, thent7I{z} CUXW8.
Proof. The first statement follows from 4.8 (13) and (8), and the fact that,
for cE Vrn(T. o R)(Wr),
r -S r r -s s
g-1(c) A= ( A(C) + - C tB[R-(c - z)] +-R1(c-z)
r r r r
To prove the second statement, observe that if

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 475

pEKG)Ar(n(TzoR)(Br) and (x,y)eG2R{z1,


then
K(X)? p - XI ? p - g(x, y) + I g(x, y) - x
< diam(B,) + r-s-
< diam(B) + 3p.

PART 3. If O< r < p, t > 0 and (z, R) E (En X G.) - g(Z t), then
reach[Arn (Tz o R) (Br)] > (p - r)t/4,
Vrn (Tz O R) (Wr) is an n -2 dimensional submanifold of class 1 of En, and
| bI| [ArC (TzOR)(Br), K n VrC (TzO R)(Wr)]
< u(x)y R)dHn-'(x
y).
(uxw,) nrr,R'1{r}

Proof. Observe that


(Tz o R)(Br) = {x: (6Bo R-1 o T_z)(x) <
r? ,
grad(5Bo R-' o T_7) = R o (grad &B) o R-1 o Th.,
and that if cCVrC\(TzOR)(Wr), (x, y)Cz,`{z}, c=g(x, y), where g is the
function defined in Part 2, then
grad 6A(C) = grad 6A(X), grad(6s o R-1 o T_z)(c) = R[grad WB(y)I,
hence Igrad bA(C) ? grad(ho a R-1 o T_z)(c) I > t.
Now if c EA r (T, o R) (B.) and
v C Nor(Ar, c), w C Nor[(Tzo R)(Br), cl, | vj >0, 1w| > 0,
then ce Vr(Tz O R)(W.) and
v | v I grad 5A(Cj, W= 1W1grad(&BO R-1 0 T_z)(c),
v+wW
IvI + IwI 2
because on the line segment joining two unit vectors the midpoint is closest
to the origin. Since
reach(Ar) > p - r, reach[(Tz o R)(Br)] > p - r
according to 4.9, it follows from 4.10 that
reach[Ar n (T, 0 R)(Br)] > (t/2)(p - r)/2.
It is now also clear that VrC'(Tz o R) (Wr) is an n -2 dimensional com-
pact submanifold of class 1 of En. Since g-1 is Lipsch'itzian, t {`z } is n-2
rectifiable, and 5.12 may be applied with

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
476 HERBERT FEDERER [December

Q = tr,R{Z} C- Arn (T2oR) (Br),


ji(x, y) = grad aA(X), i'(x, y) = R[grad 5B(Y)] for (x, y) E Q.

Finally, reference to the last statement of Part 2 completes the proof of


Part 3.
PART4. If S is a Borel subset of En X Gnand r > 0, then

rb-
LUXWIXGfl) uJird(Hn-1 0 Hn-1 0 On)
(UXW.XG.) n r-, (S)

- fSf(UXW.) ~~r,jf1~u(x, y, R)dHn-2(x, y)d(L. 0DiO)(z, R).


UXW,)ntr,R_1{ 2}
Proof. Applying Theorem 3.1 with
X = V8 X W8 X Gn, Y = En X Gn f = Tr,

g(x, y, R) = u(x, y, R) for (x, y, R) z (U X WJX Gn)I (


g(x, y, R) = 0 for (x, y, R) z (V8 X W8 X Gn) -(U X W8X Gn)r ;r
m = 2(n-1) + n(n- 1)/2, k = n + n(n-1)/2, m-k=n-2,
one obtains

JUXWIXG,I)flrl u(x, y, R)J-r(X, y, R)dHm(x, y, R)


(UXWsXGn) nCr-r(S)

= I'S f(UXWaXGn)fljr1f(Z,) u(x, y, Q)dHn-2(x, y, Q)dHk(z, R).


S UXW,XGn ) n r,-' I(z, R) )

Furthermore, if (z, R)=EnXGn, then

(U X WJX Gn)r) r I(z, R)} {(x,y, R): (x,y) E (U X W8)rr,R{Z}}


and the function mapping (x, y) onto (x, y, R) is an isometry, hence the inside
integral equals

LUXWI) ~~r,R' u(x, y, R)dHn-2(x, y).

Finally letq= Hn(n-l)12(Gn) and observe that

Hn(n-1)12 agrees with q'n over Gn,


HUnagrees with Hn-1 0 Hn-1 0 qPn over U X W8 X Gn,
Hk agrees with Ln 0 qgnover En X Gn.

PART 5. If 0 <r <p and S is a Borel subset of EnXGn, then

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 477

*
fb | q(i I [Ar n (Tz o R) (Br), K]d(Ln 0 4n) (z, R)
s

-J uJprd(Hn-I <8) Hn- I(S ?0n)


(UXWsXGn) ntr-0 (S)

+ [ I (Di I (Ar, K) + | Ii (Br)](Ln 0 40n)(S).

Proof. One sees from Part 1, 3 and Lemma 6.6 that, for Ln$J?n almost all
(z, R) in En X Gn,
(D I [Ar n (Tz o R) (Br), K]
< | 4iI[ArQn (Tz0R)(Br), KnVrQn (Tz0R)(Wr)]
+ Di |[Ar n (Tz o R)(Br) K n Ar- (Tz o R)(Wr)]
+ Dil [Ar n (Tz o R) (Br), (Tz o R) (Br) - Vr]

< y) + K) + |
f
(uxw.) n r, R t u(x, y, R)dHn-2(x,
I (Di (Br)

j (Ar,
and then one uses Part 4 to estimate the upper integral over S.
PART 6.

sup UJ;rd(Hn-i 0 Hn-1 0 q5n) < 00o


0<r <p UXWsXGn

Proof. Since the functions 3r corresponding to 0 <r < p are equi-Lipschitz-


ian, there exists a number Mlsuch that
Jtr(X, y, R) < M whenever 0 < r < p, x C U, y C W8, R E Gn
and ;ris differentiable at (x, y, R). Assuming i<n-2 and applying 6.5 with
m = n -2- i one finds that the above integrals do not exceed

Mf f u(x, y, R)d4nRdHn-'ydHn-1x

< Mdi2n-I-iIn-2_jHn-1(W.,)Hn-1 (U) < 00 .

PART 7. For each E > 0 there exist t > O, h > 0 and a compact subset S of
EnXGn such that
(Ln 0 40n)(S)<e
and such that if 0 <r < h, then
tJr[ZIn (U X Ws X G.)] C S,

I [Ar r (T. o R)(Br), K]d(Ln 0 cpn)(z,R) < e.

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
478 HERBERT FEDERER [December

Proof. Recalling 5.10 choose a number M? 1 such that


| i 1|(A,, K) + I Ji (Br) ?<M for 0 < r s.
Assured by Parts 1 and 4 that
[Ln 0 On)[-o(Zo)] = 0,

f(UXW8XGS fl~01 [~(ZuJ;od(Hn-l 09 Hn-' 0 4) = 0,


]
(uxwsxG ) nt0-l [r0(zo)

use Part 6 to secure an open subset P of En X Gn such that


N0(Zo) C P, )
(Ln 0n) (P) < e/(2M),

uJ?od(Hn-l (0 Hn-1 (0 4n) < e/2.


LUXW8XGS)fl~.O~1P
(UXW.XGn) n o-l (P)

Choose a compact subset S of P such that


to[(U X Ws X GO)n Zo] C Interior S,
choose a positive number t such that
to[(U X Ws X Gn)n Zt] C InteriorS,
and choose a positive number h ? s such that if 0 r ? h, then
Mr[(UX W. X Gn)n Zt] C Interior S,
Mr[(U X W8 X Gn) - (P)] C (En X G) - S.
Since the functions J;r converge boundedly to JP0 one may also require that
if 0 ?r <h then

U UJtrd(Hn-I0 Hn-1 0 n) < e/2.


(UXWsXGn) nlo-1(P)
Accordingly, if 0 <r _ h, then

(U X Ws X Gn) nA trl(S) C t-o(P)

and it follows from Part 5 that

f j ?J4jj [Ar n (T2 0 R)(Br), K]d(Ln 0 4.)(z, R) < e/2 + Me/(2M) e.

PART 8. For L n 0 n almost all (z, R) in En X Gn,


reach [A C) (T2 o R) (B) ] > 0
and (Di[ArG'(Tz 0 R)(Br), *] converges weakly to (i[AGr)(T2 o R)(B), *] as
r o.
S+.
Proof. Since these assertions are obviously true in case An(Tz o R)(B) is

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 479

empty, and since E. is the union of countably many compact sets, it is


sufficient to prove that the assertions hold Ln04. almost everywhere in
M(K) = I (z, R): A n (T7'o R) (B) meets K
where K is a compact subset of En.
Given E>0, apply Part 7. For (z, R) z M(K) - S one sees from Parts 2
and 3 that if 0<r!h, then

Pr {(z, R)} C U X Ws X Gn, (z, R) ( Pr(Z)


reach [A r n (Tz o R) (Br)] > (p -h)t/4,
and uses 6.10, 4.13, 5.9 to infer that
reach[A n (T2 o R) (B)] _ (p - h)t/4

and (Di[Arn (Tz o R) (Br), *] converges weakly to 4i [A n (Tz o R) (B), *] as


r-*0+.
In the remaining parts of the proof of the theorem some further conventions
are needed:
For r>O, RCG. let
tlr,R: Vr X Wr -> En, tlr,R(X, y) = x - R(y).

For r > 0 let 71r:VrX WrX Gn *EnX Gn,

tlr(X, y, R) = (r,R(X, y), R).

For 0 <r <p let Pr be the subset of VrX Wr consisting of all points (x, y)
such that either (grad 8A) | Vr is not differentiable at x or (grad 6B) IWr is not
differentiable at y.
For 0 <r <p, xE Vr let Vr(x) be the intrinsic tangent space of Vr at x.
For O<r<p, yCWr let Wr(y) be the intrinsic tangent space of Wr at y.
For 0 <r<p, (z, R)G(EnXG.) - r(Zo), xC VrG(Tz 0 R)(Wr) let rT(z, R, x)
be the intrinsic tangent space of VGr(T2 0 R)(Wr) at x.
For 0 < r <p, (z, R) G (En X Gn.)>-r(Z0) let

ar, z, R: Vrn (T2 0 R)(Wr) Vr, ar,z,R(x) = X;


br,z,R: Vr r (T2 0 R)(Wr) " > Wr, br,z,R(X) = (R-1 0 T.2)(X).

Observe that if xC VrC(Tz o R)(Wr) and y= (R-I o Tz2)(x), then dar,z,R and
map Tr(Z, R, x) isometrically into Vr(x) and Wr(y), and induce homo-
dbr,z,R
morphism
ar,z,R and br,z,R

mapping A**[Vr(x)] and A**[Wr(y)] into A**[7r,(z, R, x)].


Defining Am,j and cm,j as in 5.11 (15) and 6.5 let

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
480 HERBERT FEDERER [December

Uj(x, y, R) =An-2i,j( j grad 3A(X) + R[grad 3B(y)] |


I grad SA(X) - R[grad SB(Y)] j)
for j =O, , n-2-i, 6A(X) <p, 6B(y) <p, RCGU. Also let
( )-\ n -2) ( 1y '/1 n- 1y

= cx(n - -
- - i n - 1)! i -
Ii =a(n -i)-l(n i)-lCn-2-i,jSj(ni- i -j1) !

a(n - i-j- )(j + 1)!a(j+ 1).


PART 9. If 0<r<p, RGG (x, y)CVr X Wr, then
Jflr,R(x, y) = 2(n2)12 j grad SA(X) A R[grad SB(y)] |
Proof. Since
dim(tan[Vl, x] n tan[R(Wr), R(y)]) > n - 2,
there exist e1, * * , enCEn such that
ei, * *, en-2, en-1 is an orthonormal base of Tan[Vr, x],
ei, .. ** en-2, en is an orthonormal base of Tan[R(Wr), R(y)].
Moreover the intrinsic tangent space of VrX Wr at (x, y) has an orthonormal
base consisting of 2n -2 vectors which d17,,Rmaps onto
ei, .. * * en-2, .. en-2, en
en-,) el) *I*

respectively, and therefore


J7lr,R(x, y) = 2(n-1)12 I e1 A . . . A en-2 A en-, A en - 2(n-l)12 en-1 A en

Now the orthogonal complement of e1, , en-2 has the two orthonormal
bases
I en_l, grad &A(X)} and {en, R[grad B(y) ] },
whence it follows that
en-I A en ? grad SA(X) A R[grad 5B(y)].

PART 10. If O<r <p, then


(Hn-' 0 Hn-' 10 n)[(7r' 0 r)(ZO) = 0.

Proof. From Theorem 3.1 and Part 1 one obtains

=J77rdH2n-2+n(n-l) (2 { (z, R)
Hn-2 [7r-'1 }]dHn+n(n1-) 12(z, R) = 0.
r-1r r (Zo) I Pr(Zo)

Furthermore J7r vanishes almost nowhere, because Part 9 shows that if


(x, y)GEV,X Wr, then

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
19591 CURVATURE MEASURES 481

Jnr(x, y, R) = Jnr,R(X,y) 5 0 for 4n almost all R in Gn.


PART 11. If O<r<p, then, for L.0(4 almost all (z, R) in EnXGn,

H (r= rR 0.
Proof. Since H2*72(rr) = 0 H2n-2+n(n-l)/2(PrXGn) =0, and 3.1 implies that

Hn-2 [ (r. X Gn) \ -7'I{ (x, R) }dlln+n(n-) 12(z, R) = 0.


EnXGn

Moreover, if (z, R)CEnXGn, then

(FrrX Gn) lr,77r


{ (z, R)} ={(x, y, R): (x, y) E rr n {z}}
is isometric with rr?r-,Ri {zI
Throughout the following Parts 12 to 17 let X and i1 be bounded continuous
functions on En, and suppose X has compact support.
PART 12. Suppose 0 < r <p, (z, R) C (En X Gn)- ;r(Z0) and

H 0.r
(prfl\tqr%{Z})

If i_n-2, then

'Vrfl(T OR)(Wr) x (f o R-1 o T_z)d4i[Af n (Tz o R)(Br),]


Vrn (7,z ? R) (W,)
n-2-i
= a(n - i)-'(n - i)-2-(n-2)2 E
j==
f
{ZI
x(x)4t(y)Uj(x, y, R)
rIR

trace(ar z,R[IA(x) b2i i] z, R[:zB(y)1)dH 2(x, y) .

If i = n-1 or n, then I(D4[Arn\(Tz OR) (Br), Vrr\(Tz OR) (Wr) 0.


Proof. Applying the results of 5.2, 5.3, 5.11 with
P = Vr n (TzoR)(W,), C = Ar n (TzoR)(Br),
A = (grad SA) I P, V = [grad(h o R-' o T_z)]I P
one finds that

c(C,Q) = a(n - i)-'(n - i)-' trace[un-2-i(p)]dHn-2p

for every Borel set QCP, and consequently

nf o T)z)d()(C
x *( o R-T d) e

= a(n - )-'(n - )-i x(p)V1[(R-1 o Tz)(p)] trace[Un-2-i(p)]dHn2 p.

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
482 HERBERT FEDERER [December

Letting
h: En -* En X En, h(p) = (p,(R-1 o T_z)(p)) for p EEEn
2 p-q I whenever p, q E En, and that h(P)
one sees that Ih (p)-h (q) =212
= nr,z{z}. Hence the preceding integral over P equals

2-(n-2) /2 X(x)4/(y)trace [un-2-i(x) ]dHn-2(xX, y)

n-2-i
= 2-(n-2)/2 E Uj(x, y, R)trace[M(x)n-i-2-iN(x)i]dHn-2(xX, y).
Ix(x)i,(y)
i=O Vr R-(Z)

Now observe that for pEP the bilinear forms of Tr(z, R, p) corresponding to
M(p) and N(p) are the second fundamental forms of P at p associated with
the normal vector fields ,u and v; that for (x, y) E ( V,X Wr)- r, the bilinear
forms of Vr(x) and Wr(y) corresponding to Z:A(X) and EZB(y) are the second
fundamental forms of Vrand W1at x and y associated with the normal vector
fields (grad 8A) I V7 and (grad 8B) I Wr; and that
,u= [(grad 5A) Vr] o ar,z,R, V R o [(grad 6B) I W7] O br,z,R.
Since second fundamental forms behave naturally under inclusion maps and
isometries, one infers that

M(x) =
o-4A
N(y) b=z,R[5B(Y)]
-
whenever (x, y) E 77 -Izr.
PART 13. If O<r<p and j=O, n
, n-2-i, then, for Hn- IHn-I almost
all (x, y) in VrX WT,

fGUj(x, y, R) Igrad 5A(x) A R[grad 5B(Y)] [


Gn

2
*trace (ar,x-R (y), RIVA (X) n ]br, x-R (,y),R I B(Y) i]) dORii

=cn2-i,jsjtrace [ ZA (x)n2-ii] trace [VB(y) i.


Proof. Using Part 10 one sees that, for Hn-'1Hn-1 almost all (x, y) in
V, X Wrl

4n[{R: (x - R(y), R) E Pr(Zo)}] = 0 and (x, y) EE r.


Fix such a point (x, y) and let
v = grad 5A(X), w = grad 5B(y), V = V7(x), W = W (y),
M = eA(X)n-2-i-i E An-2-i-i,n-2--i(V), N = 2ZB(y)' E Ai"(W),
-q(R) = Uj(x, y, R) IvA R(w) I
= An-2-i,j[Iv + R(w)I, v - R(w) vAR(w)I for R e Gn.

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 483

Recalling 4.6 identify V with Tan( Vr, x), and W with Tan(Wr, y).
For 4. almost all R in G. it is true that
(x - R(y), R) E MU,
and one may identify Tr(x-R(y), R, x) with VnR(W). Then the restrictions
of
dar,x-R (y),R, dbr,x-R() ,R

to Tr(x-R(y), R, x) become identified with


the inclusion map of V n R(W) into V,
R-1I V n R(W): V n R(W) - W.
Now readopt the conventions of 6.7 with k=n-2, noting that -q(R) is
determined by v @R(w), hence by P o RI W'. For wCO let S(W) be the inclu-
sion map of the domain of w into V, and let
P(w) = trace[S(w)*(M).c*(N)3.
Then the given integral can be computed by 6.7, with

ftid0in = Cn-2-i,j
Gn

according to 6.5. Furthermore, if (g, h) CG XH, then


dmn(h of o e-I o g-') = rng(g o e),
P(h of o e-' o g-')
= trace((g o e)*[S(h of o e-' o g-l)*(M) *(h of o e-1 o g-1)*(N)])
= trace[(g o e)*(M) *(h of)*(N)].

Accordingly 6.8 implies that

'X (h of o e-1 o g-')d(,g 0 v)(g, h) = sitrace(M)trace(N).


GXH

PART 14. If O<r<p andj=O, * * ,n-2-i, then

I f JW (X(x)y6(y)Uj(x,y, R)
EnXGn Ir, R t 2

*tra,ce(ar,z,R[;p!)f- '-i]b*r,z,R [Es(y)i])dH -(x, y)d(Ln (910n)(z, R)


= 2(n-2)12Cn-2 - i - j - 1)!a(n - i - j - 1)(j + 1) !a(j + 1)
Ar x)
*<i+j+l( -I>n-j 1(Br2 0)) .

Proof. To see that the above integral exists, apply Theorem 3.1 with
f =7r, observing that if (z, R)CEnXGn, then X77` {(z, R) I is the isometric

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
484 HERBERT FEDERER [December

image of -q7,-{z} under the map carrying (x, y) into (x, y, R).
To compute the integral, first apply Fubini's Theorem to L. 0ck," and
for each REG, apply 3.1 withf=7fr,R to obtain
JX Jb
X(x)41(y)Uj(x, y, R)trace(ar, x-R (y),
R I44A()f--

Gn FrXWr

r,:y-R) ,R [VB(y) 1)fJnr,R(X, y)dHn-2(x, y)dOnhR.

Next apply Fubini's Theorem to Hn-2 30n, and apply Part 9 to obtain

2(n-2)I2 'XW Vy) fUj(x, y, R) Igrad 6A(X) A R[grad B(y)] |


FrXWr Gn

* trace(ar,-R (x) ,R[(X) 2- i i? (y) [ B(y)i])dpnRdH2n-2(X, y).

Then apply Part 13, and apply Fubini's Theorem to Hn-' ?Hn-' to obtain

2 (n-2) 2Cn.2-i,jSj fx(x) trace[ AV4


(x)n-2-i]dHn1--x. * (y) trace[ EB (y) i] dHn-'y.
Vr Wr

Finally apply the last formula in 5.8 twice to determine the integrals over
Vr and Wr.
PART15. If O<r<p and i?n-1, then

fEflX?i[Ar Cn (TzOR)(Br), X ('10 01 T-z)]d(Ln 0 (n)>(zBR)


EnXGn

=Ibi(Ar,X)?n(Br, IP) + Pn(Arj X)j(Brj 1)


,-2-;
+ ? tj1J?i+j+i(Ar,X)4bn-j-1(Br, 4).
j=0

Proof. If (z, R) E (En X Gn) -r(ZO), then

X (VIo R-' o Tz)]


?i[Ar (n (Tz o R) (Br), X

equals the sum of the three integrals

C(z, R) = x (fo R-1 o T_z)dCb[Ar r, ( Tz o R) (Br))X*J


vrfn(Tz ? R) (Br-Wr)

D(z, R) = 'Ax (#o R-' o T_z)dcIi[Ar C (Tz o R)(Br), *1]


(A,-v,) n (Tz-? R) (W,)

E(z, R) = f x (1 o R-1 o T-z)dDi[Arfn (Tzo R) (Br) J

drn a (W
(Tz R) ta

and one sees from 6.6 and 5.8 that

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
19591 CURVATURE MEASURES 485

C(z, r) = R B x (P o R-1 0 T_z)dcji(Ar* ),


(Tz ? R ) (Br-W,)

D(z, R) = A * X (1 o R-1 o T_z)d[(Tz O R) (Br),

~~~(x o TZo R) *~dti(Br*).


(R- oT_Ts) (Ar-Vr)

Applying 6.9 one obtains

'XG 0Cd(Ln n) = CJi(Ar, X>%n(Br,'1),


EnXOn

fEfXGfDd(Ln 0 4n) = 4n(A r X)cJ?(Br, ifr).


EnXGn

If i_ n-2, it follows from Parts 1, 11, 12, 14 that


n-2-i
Ed(Ln 0 45n) = , tjcj+j+i(Ar, X)4?n_j-1(Br, #).
EnXGn j=O

If i=n-1, then E(z, R)=O for (z, R)GEnXGn -r(Zo).


PART 16. If i_n-1, then

f (b[A n (Tz o R)(B)I(X oR-1 o T_Z)]d(Ln,0 On) (z, R)


EXGn
EGA = * X)4n(B, #1)+ 4 xn(Al X)4(B, j)
n-2-i
+ tj?+j+1(A, X)4n-i-l(B, 1).
j=o
Proof. Since one knows from 5.10 that, for k=, * ,
'Ik(Ar, X) -- Dk(A, X) and 'Ik(Br, #1) > bk(B, vk)

as r- O+, it will be sufficient to show that the integral of Part 15 approaches


the integral of Part 16 as r- +O+.
Let M be a common upper bound of Ix I and 141
I| and let K be the sup-
port of X.
Given E>0, choose t, h, S according to Part 7. Then

fI C|i[ArC (Tz O R)(Br), X ( O R-1 O T_z)I d(Ln


0 (D n)(z, R) < EM2

for 0 <r < h, and it follows from Part 8 and Fatou's Lemma that

f | i[A n (Tz o R)(B), x (1o R-1o T_z)]| d(Ln0 4n)(z, R) ? EM2.

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
486 HERBERT FEDERER [December

Referring again to 5.10 one obtains


N = sup |i, I (C): C C Bh and reach(C) ? (p - h)t/4} < oo.
If (z, R)z(EnXG.)-S, 0<r<h and ArCn(Tz O R)(Br) meets K, then Parts
2 and 3 imply
(p - h)t/4 < reach[A, r) (T, o R)(Br)] = reach[(R-' o T_z)(Ar) G Br]
I4>t [Ar fli (Tz o R)(Br)] = I (i I [(R-1 o T_z)(Ar) r\ Br] _ N,
I -ti[Arr (TzoR)(Br)]iX'(lOR-'O T_z)I| <NM2.
Observing that the set
D = { (z, R): (Tzo R)(Bh) meets K} = { (x-R(y), R): x G K, Re G, y E Bh}
is compact, and recalling Part 8, one may apply Lebesgue's theorem concern-
ing bounded convergence to D - S, and conclude that

lim sup f I 4bi[Arrl (Tz o R)(Br), X- (1 o R-1 oTaz)]


r-+o+ EnXG,
- OR-1 o T_z)] I d(Ln 0 p.)(z, R) ?
4'[A r) (7TZo R)(B), x* (1,V 2EM2.

PART 17.

f 4'n[A r) (TZo R)(B), X -(4Po R-1 o T_z)]d(Ln 0 cp.)(z, R)


EnXG

- cI(A, X)4n(B, Vt/).


Proof. Using 5.8 and 6.9 one finds that the above integral equals

w L . Ed (Ln0 'pn)(z, R)
x *('ko R-I o T_z)dL
EnXGn On(Tz?R) (B)

= 'EjXGw .fTJ R)(B) (o O R-1 a T7z)d4n(A,*)d(Ln0 0pn)(z,R)


EXGn (Tz ? R ) (B )

= bn(A,x) 'BdLn = -n(A, X)4Yn(B, ).

PART 18. PROOF OF (3). In order to prove that the integrand of (3) is an
Ln04)n measurable function, consider for m = 1, 2, 3, the set Fm of all
real valued continuous functions f on En such that
If(x) I < 1 whenever x G En) f(x) - 0 whenever 6K(X) > m-',

and let Cmbe a countable dense (with respect to uniform convergence) subset
of Fm. One sees from Parts 16 and 17 that if fEFm, then
r
mb[ ( Tz fo R) (B) I f

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
19591 CURVATURE MEASURES 487

is L. 04, measurable with respect to (z, R). Furthermore

| biI [A n (Tzo R)(B), K] = lim sup bi[A n (T,oR)(B),f].


r-l co fECm

Now one may use Part 8, Fatou's Lemma, and Parts 5, 6 to obtain

fEflXf | i I [A n (Tz o R)(B), K]d(Ln 0 cpn)(z,


R)
EnXOn

?lim inf |i j [Ar n (Tz o R)(Br), K]d(Ln 0J 4n))(z, R) < ?.


r--o+ EnXOn

PART 19. PROOF OF (2) AND (4). Through use of (3) and bounded con-
vergence, the formulae of Parts 16 and 17 may be extended from the con-
tinuous case to the case in which X and af are bounded Baire functions, X hav-
ing bounded support. Hence the proof of (2), and its corollary (4) resulting
when X and af are the characteristic functions of A and B, may be completed
by showing that
ti= 'y(n, i +j + 1, n-j- 1) forj = 0, * , n - 2-i.
For this purpose let k=i+j+l, I=n-j-1 and consider the special case
where A and B are k and I dimensional cubes. Using 5.15 one sees that
-I)k(A)= Hk(A), 4im(A) = 0 for m > k,
cI(B) = H'(B), 4bm(B) = 0 for7m > 1.

Moreover, for Ln ?4n almost all (z, R) in EnX Gn, ACn(TZo R)(B) is either
a k + 1- n = i dimensional convex set or empty, hence

bi[A r) (T, o R)(B) ] = Hi[A C' (T, o R)(B)].

Substituting in the formula of Part 16 one obtains

fEflXGfHi[A C) (Tz o R)(B)]d(Ln0 4n)(z,R) = tIHk(A)IHI(B).


EnXOn

On the other hand [F7, 6.2] shows that this integral equals
'y(n, k, l)Hk(A)HI(B).
6.12. REMARK. The following simple example shows that 6.11 (1) may fail
to hold in case neither A nor B is compact.
Let H be the subgroup of E2 consisting of all points both of whose co-
ordinates are integers, let C be a circle of radius 1/3 in E2, let A be the union
of all translates of C by elements of H, and let B be a straight line in E2, so
that reach(A)= 1/3 and reach(B) = oo. Then almost all isometric images of
B have irrational slopes. Moreover, if L is a straight line with irrational slope,

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
488 HERBERT FEDERER [December

then the image of L in E2/H is dense in E2/H, hence L cuts suitable translates
of C at arbitrarily near points, and therefore reach(ACnL) = 0.
6.13. THEOREM. If B is a compact subset of En, reach(B) >0, ,6is a bounded
Baire function on En, i=0, .*.. * n and m=0, * . , n-i, then

f bi[X?m(R, w) n B, ip]d(4n 0 Lm)(R, w) = y(n, n - m, m + i)4bm+i(B, sp).


GnXEm

Proof. Let
A = Enn {x:xi = O for i = 1, ,m}

and let X be the characteristic function of


ANnIx:O _ xi _!!l for i =m+ 1, ,n}.
Then 4?nm(A, X)= 1, bj(A, X) = 0 for j n-m, and the sum of 6.11 (2) equals
y(n, n- m, m + i) m+i(B, V).

Identifying En with EmXEn-m and applying the Fubini theorem one finds
that the integral of 6.11 (2) equals

(bi[A n (T(w,y)o R)(B), X (i O R-1 0 T(-w,-y))


dLn-myd(Lm 0 4n)(w, R).

In order to compute inner integral with respect to y, for a fixed (w, R),
abbreviate
(D[A n (T(w,o)o R) (B), *]]-, + o R-1 o T(_w,o)= f
and note that

bi[A n (T(w,y)o R)(B), x (' o R-1 o T(,))] = f (x o T(o,y))jfd4

whenever y En-m, because A = T(o,y)(A); hence one obtains

'Enfm fEmXEnmX(u, v + y)f(u, v)dA(u,v)dLmy


En-m EmXEn-m

= 'EmXEnf m f(, v) f X(u, v + y)dLmydq(u,v)


EmXEn-m En-m

=
fd,u= 4i[(R-l o T(_,o)) (A) n B, #I
= b[[Xfnm(R Y -W) r) B, #P]

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 489

because if (u, v) belongs to the support of ,, then (u, v) CA, u = 0, and

fEn mX(?, v + y)dLn-.my = 1.


En-m
Thus one finds that the integral of 6.11 (2) equals

f 4KfiN (R , -w) n B, #p]d(Lm0) 4n)(w, R),


EmXGn

and one completes the proof by observing that Lm0n is invariant under the
inversion mapping (w, R) onto (-w, R-1).
6.14. REMARK. If A CEn, reach(A) >0 and Q is a bounded Borel subset
of A (k) [see 4.15 (3) ], then
4j(A, Q) =0forj=k+1, .. *,n,
0_ 4k(A, Q) _ Hk(Q)j
4Dk(A, Q) > 0 in case Hk(Q) > 0,

4k(A, Q) = Hk(Q) in case A = A(k).

These statements are obviously true for k =0, because A (0) is countable and
Ln({ x: 3A(X) ? r and (A(X) = a}) = rna(n)bo(A, {4a)
whenever O<r<reach(A) and a A. Moreover one may pass from k=O to
k>0 by means of the following considerations: Recall 6.11 (1), assume A is
compact, and let
k
pn: En -+PEk, pn(X) = (x1, * Xk) for x E En.
Using 4.15 (3), verify that
n
nk
Qn -k(R, w) C [A xflk(Rx w)](0)

for 4n0 Lk almost all (R, w) in GnX Ek; in fact the set of all those (R, a) in
GnX Q for which
dim[Nor(A, a) + R(En l {x: xi = Ofor i = k + 1, . , n})] < n
has On<?IHk =Hn(n-l)12+k measure 0, and the image of this set under the
Lipschitzian map
f; Gn X Q-> Gn X Ek,

f(R, a) = (R, (pk o R 1)(a)) for (R, a) E Gn X Q,


contains the set of all those (R, w) for which the above inclusion fails. Now
apply 6.13 with B, m, k replaced by A, k, Q; in particular for i=O compare
the resulting formula

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
490 HERBERT FEDERER [December

f3(n, k)4k(A, Q) = f4 bo[A n Xn (RI w), Q]d(4n 0 Lk)(R, w)


GnXEk

with the formula

fl(n, k)H (Q)= I H [Q n Xn (R, w)]d(4n 0 Lk)(R, w)


nXEk

obtained from 4.13 (3) and [F4, 5.14].


REFERENCES
C. B. ALLENDOERFER
A. The Euler numberof a Riemann manifold,Amer. J. Math. vol. 62 (1940) pp. 243-248.
C. B. ALLENDOERFER AND A. WEIL
AW. The Gauss-Bonnettheoremfor Riemannian polyhedra,Trans. Amer. Math. Soc. vol.
53 (1943) pp. 101-129.
W. BLASCHKE
BL. VorlesungeniiberIntegralgeometrie, Leipzig and Berlin, Teubner, 1936-1937.
T. BONNESEN AND W. FENCHEL
BF. Theorieder konvexenKorper,Erg. d. Math. vol. 3 (1934) pp. 1-172.
N. BOURBAKI
B1. Alge'bremultilindaire,Actualit6s Sci. Ind. no. 1044, 1948.
B2. Integration,Actualit6s Sci. Ind. no. 1175, 1952.
t. CARTAN
CA. Leqonssur la gdometriedes espacesde Riemann, Paris, Gauthier-Villars,1946.
L. CESARI
CE. Surfacearea, Ann. of Math. Studies vol. 35, Princeton University Press, 1956.
S.-S. CHERN
Cl. On the curvaturaintegrain a Riemannian manifold,Ann. of Math. vol. 46 (1945) pp.
674-684.
C2. On the kinematicformula in the Euclideanspace of N dimensions,Amer. J. Math. vol.
74 (1952) pp. 227-236.
C3. La gdomdtrie des sousvari6t6sd'un espaceeuclidiend plusieursdimensions,L'Ens. Math.
vol. 40 (1955) pp. 26-46.
S.-S. CHERNANDR. K. LASHOF
CL. Onthetotalcurvatureof immersedmanifolds,Amer. J. Math. vol. 79 (1957) pp. 306-318.
M. R. DEMERS AND H. FEDERER
DF. On Lebesguearea. II, Trans. Amer. Math. Soc. vol. 90 (1959) pp. 499-522.
E. Di GIORGI
DG. Su una teoria generaledella misurar -1 dimensionalein un spazio ad r dimensioni,
Ann. Mat. Pura Appl. ser. 4 vol. 36 (1954) pp. 191-213.
H. FEDERER
Fl. Surfacearea I, Trans. Amer. Math. Soc. vol. 55 (1944) pp. 420-437.
F2. Surfacearea II, Trans. Amer. Math. Soc. vol. 55 (1944) pp. 438-456.
F3. Coincidencefunctions and their integrals,Trans. Amer. Math. Soc. vol. 59 (1946) pp.
441-466.
F4. The (0, k) rectifiablesubsetsof n space,Trans. Amer. Math. Soc. vol. 62 (1947) pp. 114-
192.
F5. Dimension and measure,Trans. Amer. Math. Soc. vol. 62 (1947) pp. 536-547.
F6. Measureand area, Bull. Amer. Math. Soc. vol. 58 (1952) pp. 306-378.

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions
1959] CURVATURE MEASURES 491

theorems,Trans. Amer. Math. Soc. vol. 77 (1954) pp. 238-261.


F7. Someintegralgeometric
F8. On Lebesguearea, Ann. of Math. vol. 61 (1955) pp. 289-353.
F9. An introduction to differential geometty, Mimeographed lecture notes, Brown Univer-
sity, 1948.
W. FENCHEL
FEL. On the total curvature of Riemannian manifolds, J. London Math. Soc. vol. 15 (1940)
pp. 15-22.
FE2. On the differential geometry of closed space curves, Bull. Amer. Math. Soc. vol. 57
(1951) pp. 44-54.
H. FLANDERS
FLI. Development of an extended differential calculus, Trans. Amer. Math. Soc. vol. 75
(1953) pp. 311-326.
FL2. Methodsin affineconnectiontheory,Pacific J. Math. vol. 5 (1955) pp. 391-431.
H. HADWIGER
H. Vorlesungen uber Inhalt, Oberfldche, Isoperimetrie, Berlin, Springer, 1957.
W. HUREWICZ AND H. WALLMAN
HW. Dimension theory,Princeton Mathematical Series, Princeton University Press, vol.
4, 1941.
L. H. LooMIs
Li. The intrinsic measure theory of Riemannian and Euclidean spaces, Ann. of Math. vol.
45 (1944) pp. 367-374.
L2. Abstract congruence and the uniqueness of Haar measure, Ann. of Math. vol. 46 (1945)
pp. 348-355.
M. MORSE
M. The calculus of variations in the large, Amer. Math. Soc. Colloquium Publications,
vol. 18, 1934, 368 pp.
J. W. MILNOR
MI. On the total curvature of knots, Ann. of Math. vol. 52 (1950) pp. 248-257.
T. RAD6
R. Lengthand area, Amer. Math. Soc. ColloquiumPublications,vol. 30, 1948, 572 pp.
S. SAKS
S. Theoryof the integral, MonografieMatematyczne, vol. 7, Warsaw, 1937.
L. A. SANTAL6
SA. Ober das kinematische Mass im Raum, Actualit6s Sci. Ind. no. 357, 1936.
H. WEYL
WE. On thevolumeof tubes,Amer. J. Math. vol. 61 (1939) pp. 461-472.
H. WHITNEY
WI. On totally differentiable and smooth functions, Pacific J. Math. vol. 1 (1951) pp. 143-
159.
W2. Geometricintegrationtheory, Princeton Mathematical Series, Princeton University
Press, vol. 21, 1957.
BROWN UNIVERSITY,
PROVIDENCE, RHODE ISLAND
THE INSTITUTE FOR ADVANCED STUDY,
PRINCETON, NEW JERSEY

This content downloaded from 35.13.190.230 on Tue, 1 Apr 2014 14:03:45 PM


All use subject to JSTOR Terms and Conditions

You might also like