optimum quantization and its applications
TRANSCRIPT
http://www.elsevier.com/locate/aim
Advances in Mathematics 186 (2004) 456–497
Optimum quantization and its applications
Peter M. Gruber�
Abteilung fur Analysis, Technische Universitat Wien, Wiedner HauptstraX e 8-10/1142, A-1040 Vienna, Austria
Received 9 October 2002; accepted 11 July 2003
Communicated by Erwin Lutwak
Abstract
Minimum sums of moments or, equivalently, distortion of optimum quantizers play an
important role in several branches of mathematics. Fejes Toth’s inequality for sums of
moments in the plane and Zador’s asymptotic formula for minimum distortion in Euclidean d-
space are the first precise pertinent results in dimension dX2: In this article these results are
generalized in the form of asymptotic formulae for minimum sums of moments, resp.
distortion of optimum quantizers on Riemannian d-manifolds and normed d-spaces. In
addition, we provide geometric and analytic information on the structure of optimum
configurations. Our results are then used to obtain information on
(i) the minimum distortion of high-resolution vector quantization and optimum quantizers,
(ii) the error of best approximation of probability measures by discrete measures and
support sets of best approximating discrete measures,
(iii) the minimum error of numerical integration formulae for classes of Holder continuous
functions and optimum sets of nodes,
(iv) best volume approximation of convex bodies by circumscribed convex polytopes and the
form of best approximating polytopes, and
(v) the minimum isoperimetric quotient of convex polytopes in Minkowski spaces and the
form of the minimizing polytopes.
r 2003 Elsevier Inc. All rights reserved.
MSC: 52C99; 53C20; 94A24; 94A34; 65D30; 65D32; 52A27; 52B60
Keywords: Optimum quantization; Sums of moments; Distortion of vector quantizers; Approximation of
probability measures; Error of numerical integration formulae; Best approximating polytopes; Minimum
isoperimetric quotient
ARTICLE IN PRESS
�Fax: +43-1-558-01-11498.
E-mail address: [email protected].
0001-8708/$ - see front matter r 2003 Elsevier Inc. All rights reserved.
doi:10.1016/j.aim.2003.07.017
1. Introduction
1.1. Fejes Toth’s inequality for sums of moments in the plane
Let f : ½0;þNÞ-½0;þNÞ be increasing, where f ð0Þ ¼ 0; and let H be a convex 3,
4, 5, or 6-gon in the Euclidean plane E2: Then
infSCE2
#S¼n
ZH
minsAS
f f ðjjs ujj2Þg du
� �Xn
ZHn
f ðjjujj2Þ du; ð1:1Þ
where n ¼ #S is the number of points of S and Hn a regular convex hexagon in E2 ofarea jHj=n and center at the origin o: jj � jj2 and j � j stand for the Euclidean norm and
the ordinary (area) measure. There are more than a dozen proofs for this result of L.Fejes Toth [17] and its variants. Its applications range from packing and coveringproblems for solid circles, problems of optimal location in econometrics,
quantization of data, Gauss channels and numerical integration, all in E2; to theisoperimetric problem for special convex polytopes and asymptotic approximation
of convex bodies in E3: See [29,39] for references. Related results, including results onfinite difference methods, cellular biology and territorial behaviour of animals aresurveyed by Du et al. [12].
Inequality (1.1) indicates that for certain geometric and analytic problems in E2 or
E3 regular hexagonal configurations are close to optimal, possibly optimal.
1.2. The corresponding planar stability problem
The objective here is to describe the sets S ¼ Sn for which equality or approximateequality holds in (1.1). Although for certain functions f no clear cut answer is
possible, Gruber [28,30] showed that in E2 and on Riemannian 2-manifolds undersuitable restrictions for f ; the sets Sn form ‘asymptotically regular hexagonal
patterns’ as n-N: For E2 a similar, slightly more general result was given by G.Fejes Toth [15].
The stability result implies that for certain geometric and analytic problems in
E2; E3 and on Riemannian 2-manifolds, the optimal or almost optimal configura-tions are ‘asymptotically regular hexagonal pattern’. For examples see Gruber [30].
1.3. Extensions to general dimensions
From the point of view of applications the following extensions suggestthemselves. First, determine or estimate the expressions
infSCM#S¼n
ZJ
minpAS
f f ðRmðp; xÞÞgwðxÞ doMðxÞ� �
; ð1:2Þ
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 457
and
infSCEd
#S¼n
ZJ
minsAS
f f ðjjs ujjÞgwðuÞ du
� �: ð1:3Þ
Here J is a (Jordan) measurable set in a Riemannian d-manifold M with
Riemannian metric RM and (area) measure oM ; or in Ed ; where the latter is
endowed with an additional norm jj � jj; and w : J-Rþ a weight function. Second,describe the corresponding minimizing or almost minimizing configurations. Whileprecise answers to these problems are out of reach, asymptotic estimates and upperand lower bounds for the expressions in (1.2) and (1.3) and information on theminimizing configurations have been given for special functions f ; particular norms,and d ¼ 2:
In the context of information theory Zador [48] proved that for f ðtÞ ¼ ta; a40;and jj � jj ¼ jj � jj2 the expression in (1.3) is asymptotically equal to
ZJ
wðuÞd
aþd du
� �aþdd 1
nad
as n-N;
omitting a multiplicative constant; for related, more recent work in informationtheory see the survey of Gray and Neuhoff [21], the book [9] and the articles [22], [23]and [41]. The results in information theory were used to approximate probabilitymeasures by discrete measures, see the report of Graf and Luschgy [20]. In thedisguise of an asymptotic formula for the minimum error of numerical integrationformulae the same result, but for general norms instead of jj � jj2; was proved by
Chernaya [7]. A similar formula with f from a wide class of moduli of continuityincluding the functions of the form ta and where the asymptotics is described by afunction related to f ; was given in the same year by Chernaya [8]. Finally, Gruber[26] and Glasauer and Gruber [19] proved in the context of convex geometry that the
expression in (1.2) for f ðtÞ ¼ t2 is asymptotically equal to
ZJ
wðxÞd
dþ2 doMðxÞ� �dþ2
d 1
n2d
as n-N;
again omitting a multiplicative constant. It is curious to note that all these closelyrelated results were found independently in different areas of mathematics. Theproofs differ, but a crucial idea, a similarity argument, which was communicated toZador by Hammersley, is the same in each of them. In [33] we will describe a prooffor general norms which is also based on it.
If f is strictly increasing, then for S ¼ fsn1;y; snngCEd ;
ZJ
mini¼1;y;n
f f ðjjsni ujj2ÞgwðuÞ du ¼Xn
i¼1
ZDni
f ðjjsni ujj2ÞwðuÞ du;
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497458
where the Dirichlet–Voronoi cells Dni; i ¼ 1;y; n; in J corresponding to S aredefined by
Dni ¼ fuAJ: jjsni ujj2pjjsnj ujj2 for j ¼ 1;y; ng:
(This explains why we speak of sums of moments.) A conjecture of Gersho [18]
asserts that for f ðtÞ ¼ t2 or, more generally, for strictly increasing f ; and for jj � jj ¼jj � jj2 and w ¼ 1; there is a convex polytope P with jPj ¼ 1 which admits a tiling
of Ed by congruent copies such that the following holds: as n-N; the cells Dni
which correspond to minimizing configurations S ¼ Sn; where #Sn ¼ n; are
‘asymptotically congruent’ to ðjJj=nÞ1=dP: For d ¼ 2 the author’s stability result
stated above readily implies the truth of this conjecture, showing also thatP is a regular hexagon. This seems to be the first proof of Gersho’s conjecture ford ¼ 2: Also G. Fejes Toth’s [15] version of the stability theorem yields a proof ofGersho’s conjecture for d ¼ 2: (We point out that no proof is provided by FejesToth [16] or Newman [42], contrary to positive statements in the literature.) FordX3 it is not clear that the conjecture is valid and if so, at present, proofs seem to beout of reach.
Assuming the truth of Gersho’s conjecture, lower bounds for the expression in(1.3) have been given for certain functions f and certain norms. See [33] for simplearguments which also lead to lower bounds. For references see the survey of Grayand Neuhoff [21].
General references for results as described above are Gray and Neuhoff [21], Duet al. [12], Graf and Luschgy [20] and Gruber [33]. The latter is an easilycomprehensible introduction to the present article.
1.4. Results
We will give asymptotic formulae for the expressions in (1.2) and (1.3) for a wideclass of functions f : Further, geometric and analytic information on the minimizingconfigurations S ¼ Sn will be provided; roughly speaking, Sn is distributed over J
rather uniformly and regularly. The hard part of the proof is to show that Sn isDelone (Section 2).
These results admit interpretations and applications to quantization of data,approximation of probability measures by discrete probability measures, numericalintegration, approximation of convex bodies by polytopes and the isoperimetricproblem for convex polytopes.
An important step in data transmission is to encode signals (in many cases vectors
in Ed) produced by a source into codewords (particular vectors) from a codebookconsisting of, say, n codewords (and which are then transmitted in a channel).So-called distortions measure the quality of this encoding or quantizing process.See [21] for the pertinent literature starting with Shannon and Zador. In Shannon’stheory distortion is estimated as d-N; while in Zador’s high resolution theoryestimates for n-N are considered. Our results immediately imply asymptotic
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 459
formulae for the minimum distortion of quantization as the number n of codewordstends to infinity and give information about the structure of optimum codebooks(Section 3).
Given a probability measure on a Jordan measurable subset of Ed ; it is a naturalquestion to ask, how well can it be approximated by discrete measures. Here thespace of all probability measures is endowed with the Wasserstein or Kantorovichmetric, or generalizations of it. See Graf and Luschgy [20]. Our results yieldasymptotic formulae for the error of the best approximation as the number ofsupport points of the approximating discrete measures tends to infinity. In addition,information on the structure of the support of the best approximating discretemeasures is provided (Section 4).
Consider a class of Riemann integrable real functions on a Jordan measurable
subset of Ed or of a Riemannian manifold M: In general it is difficult to estimate theminimum error of numerical integration formulae with n nodes and weights.Pertinent results have been given in the context of uniform distribution theory byKoksma and Hlawka, see [34], further for Sobolev spaces of functions by Sobolev[46] and his school, including Polovinkin [43], and for classes of Lipschitz or Holdercontinuous functions by Babenko [1,2] for d ¼ 2 and Sobol’ [45], Chernaya [7,8] andothers for general d: Our results yield for certain Holder classes of functionsasymptotic formulae for the minimum error of numerical integration formulae as thenumber n of nodes and weights tends to infinity. (Our classes are slightly less generalthan those of Chernaya.) Moreover, we obtain information about the structure ofoptimal sets of nodes (Section 5).
The special case f ðtÞ ¼ t2 of our result for Riemannian manifolds yieldsasymptotic formulae for best volume approximation of convex bodies bycircumscribed convex polytopes as the number of facets tends to N: An argumentfrom the proof of our manifold result shows that the facets of the best approximatingpolytopes are all ‘rather round’ and ‘roughly of the same size’ (Section 6).
The latter results can be interpreted as asymptotic formulae for the minimumisoperimetric quotient of convex polytopes in a Minkowski space with n facets asn-N and as information about the form of polytopes with minimum isoperimetricquotient (Section 7).
For notations and results not explained below, see Schneider [44].
2. Sums of moments on Riemannian d-manifolds and normed d-spaces
2.1. Preliminaries
A function f : ½0;þNÞ-½0;þNÞ satisfies the growth condition if
f ð0Þ ¼ 0; f is continuous and strictly increasing and; for any given s41;
the quotient f ðstÞ=f ðtÞ is decreasing and bounded above for t40: ð2:1Þ
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497460
In addition to the positive powers of t there are many other functionsf : ½0;þNÞ-½0;þNÞ which satisfy the growth condition. This is shown by thefollowing result. We state it without proof.
Proposition. Let f : ½0;þNÞ-½0;þNÞ be such that f ðtÞ ¼ 0 only for t ¼ 0:Define g : R-R by gðuÞ ¼ log f ðeuÞ for uAR: Then the following statements are
equivalent:
(i) f satisfies the growth condition,(ii) g is strictly increasing, concave and Lipschitz.
Let M be a Riemannian d-manifold. By this we mean a d-dimensional
manifold of class C3 with a metric tensorfield (which has coefficients) of
class C1: Let RM and oM be the corresponding Riemannian metric and (area)measure on M: A set JCM is (Jordan or Riemann) measurable if its closurecl J is compact and oMðbd JÞ ¼ 0; where bd J is the boundary of J: See 2.3for an alternative definition of measurability. A measurable set JCM has positive
density if
there are b41; g40; such that BMðp; RÞ is measurable and
Rd
bpoMðBMðp; RÞ-JÞpbRd for pAcl J; 0oRog; ð2:2Þ
where the ball BMðp; RÞ with center p and radius R is the set fxAM : RMðp; xÞpRg:The measurability of BMðp; RÞ for sufficiently small g follows from an argumentusing the exponential mapping, compare the proof of (2.12). Let JCM bemeasurable with oMðJÞ40; let ðSnÞ be a sequence of sets in M with #Sn ¼ n;
and let d41: Sn is a ð1=dn1=d ; d=n1=dÞ-Delone set in J for n ¼ 1; 2;y; if any two
distinct points of Sn have distance at least 1=dn1=d and for each point of J there is a
point of Sn at distance at most d=n1=d : Sn is uniformly distributed in J as n-N with
respect to an integrable function, a density, w : J-Rþ if for any measurable set KCJ
we have,
#ðK-SnÞBZ
K
wðxÞ doMðxÞZ
J
wðxÞ doMðxÞ� ��
n as n-N:
2.2. Asymptotic results
We will prove the following result, where a40 is a constant corresponding to f ;see (2.14).
Theorem 1. Let f : ½0;þNÞ-½0;þNÞ satisfy the growth condition (2.1) and let a40be the corresponding constant. Then there is a constant div40; depending only on f and
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 461
d; such that the following is true: let M be a Riemannian d-manifold, JCM a compact
measurable set with oMðJÞ40; and w : J-Rþ continuous. Then
(i) Fn ¼ infSCM#S¼n
ZJ
minpAS
f f ðRMðp; xÞÞgwðxÞ doMðxÞ� �
Bdiv
ZJ
wðxÞd
aþd doMðxÞ� �aþd
d
f 1
n1d
!as n-N:
If, in addition, J is connected and has positive density, see (2.2), then the following hold:let ðSnÞ be a sequence of sets SnCM with #Sn ¼ n such that the infimum in (i) is
attained for S ¼ Sn; n ¼ 1; 2;y: Then
(ii) there is a constant d41 such that Sn is ð1=dn1=d ; d=n1=dÞ-Delone in J for
n ¼ 1; 2;y;(iii) Sn is uniformly distributed in J with density wd=ðaþdÞ as n-N:
Analogous results hold under the assumption that S;SnCJ:
A proof which is similar to that of Theorem 1, but with suitably distorted sets Sn
and which makes use of a result of Ewald et al. [14], yields the next result.
Theorem 2. Let f : ½0;þNÞ-½0;þNÞ satisfy the growth condition (2.1),
let a40 be the corresponding constant, and let jj � jj be a norm on Ed : Then
there is a constant div40; depending only on f and jj � jj; such that the following
holds: let JCEd be compact and measurable with jJj40; and w : J-Rþ continuous.
Then
(i) infSCEd
#S¼n
ZJ
minsAS
f f ðjjs ujjÞgwðuÞ du
� �Bdiv
ZJ
wðuÞd
aþd du
� �aþdd
f1
n1d
!as n-N:
If, in addition, J is connected and has positive density, see (2.2), then the following
assertions hold: let ðSnÞ be a sequence of sets SnCEd with #Sn ¼ n such that the
infimum in (i) is attained for S ¼ Sn; n ¼ 1; 2;y: Then
(ii) there is a constant d41 such that Sn is ð1=dn1=d ; d=n1=dÞ-Delone in J for n ¼1; 2;y;
(iii) Sn is uniformly distributed in J with ensity wd=ðaþdÞ as n-N:
Analogous results hold under the assumption that S;SnCJ:
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497462
Remark. Theorem 2 can be generalized to Finsler spaces where all tangent norms areisometric—as in Theorem 1. For general Finsler spaces there is also such a result, butthen the constant div depends on J:
The proof of Theorem 1 will be given in Sections 2.3–2.6. To show the Deloneproperty of Sn is the core of the proof. This property is then used to prove theasymptotic formula. The latter is the main tool for the proof that Sn is uniformlydistributed. We are aware that there are more direct proofs of the asymptoticformula; compare the sketch of the proof of the asymptotic formula in Theorem 2 in[33], or the proof of a similar formula due to Chernaya [8].
2.3. Preparations for the proof
First, let M be a Riemannian d-manifold. Let U be an open neighborhood of a
point pAM with corresponding homeomorphism h ¼ ‘‘0’’ which maps U onto the
open set U 0 ¼ hðUÞCEd : To each uAU 0 there corresponds a positive definite
quadratic form qu on Ed :
A curve C in U is of class C1 if it has a parametrization x : ½a; b -U such that
u ¼ h3x is a parametrization of class C1 of a curve in U 0: The length of C then isdefined to be
Z b
a
quðtÞð ’uðtÞÞ12 dt: ð2:3Þ
By means of appropriate dissection and addition one can define the length for any
curve in M which is piecewise of class C1 and which is not contained in a singleneighborhood. The Riemannian distance RMðx; yÞ of two points x; yAM is theinfimum of the lengths of continuous curves in M which connect x; y and are
piecewise of class C1: A curve in M of class C1 which connects two points x; yAM is
a (geodesic) segment if its length is RMðx; yÞ: A geodesic is a curve of class C1 in M
which consists locally of geodesic segments. A set NCM is geodesically convex if forany x; yAN there is a unique geodesic segment connecting x; y and this segment iscontained in N:
A set JCU is (Riemann or Jordan) measurable in M if hðJÞ is so in Ed : Then its(area) measure oMðJÞ is defined by
oMðJÞ ¼Z
hðJÞðdet quÞ
12 du;
where du ¼ du1?dud : Again, by dissection and addition one can define measur-ability and measure for sets which are not contained in a single neighborhood. (Foran alternative definition see 2.1.)
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 463
If JCU is measurable and w : J-R continuous, then the (Riemann) integral of w
on J is defined by ZJ
wðxÞ doMðxÞ ¼Z
hðJÞwðh1ðuÞÞðdet quÞ
12 du: ð2:4Þ
Also here one can define integrals on sets not contained in a single neighborhood.Using the exponential map (compare the proof of (2.12)) and simple arguments
involving M one can show the following:
Let ICM be compact. Then there are e41; W40 such that BMðp; RÞ is measurableand
Rd
epoMðBMðp; RÞÞpeRd for pAI ; 0oRpW: ð2:5Þ
Let J;w : J-Rþ be as in Theorem 1 and let l41: For each pAM choose U ; h ¼‘‘0’’; U 0 ¼ hðUÞ; qu; q ¼ qp0 as above where U is so small that the following claims
hold true:
1
lqðx0 y0Þ
12pRMðx; yÞplqðx0 y0Þ for x; yAU ;
1
lðdet qÞ
12pðdet quÞ
12plðdet qÞ
12 for uAU ;
1
ljK 0jðdet qÞ
12poMðKÞpljK 0jðdet qÞ
12 for measurable KCU ;
l inffwðxÞ : xAJ-UgXsupfwðxÞ : xAJ-Ug:
Let V be an open measurable neighborhood of p with cl VCU :As p ranges over the compact set J; the corresponding neighborhoods V form an
open covering of J: Thus there is a finite subcover. Therefore there are, say, m points
in J and corresponding neighborhoods Ul ;Vl ; homeomorphisms hl ¼ ‘‘0’’; andpositive definite quadratic forms ql ; such that
1
lqlðx0 y0Þ
12pRMðx; yÞpl qlðx0 y0Þ
12 for x; yAUl ; ð2:6Þ
1
lðdet qlÞ
12pðdet quÞ
12pl ðdet qlÞ
12 for uAU 0
l ; ð2:7Þ
1
ljK 0jðdet qlÞ
12poMðKÞpljK 0jðdet qlÞ
12 for measurable KCU ; ð2:8Þ
for l ¼ 1; 2;y;m: Moreover, the sets
Jl ¼ J-ðVl\ðV1,y,Vl1ÞÞCVlCUl ; l ¼ 1;y;m;
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497464
have the following properties:
J is the disjoint union of the measurable sets JlCVlCUl ;
where Vl is open and measurable; Ul open and cl VlCUl ; l ¼ 1;y;m: ð2:9Þ
wlpwðxÞplwl ;1l WlpwðxÞpWl for xAJl ;
where wl ¼ inffwðxÞ : xAJlg; Wl ¼ supfwðxÞ : xAJlg; l ¼ 1;y;m: ð2:10Þ
Since J is compact and w : J-Rþ continuous,
0ov ¼ inffwðxÞ : xAJgpsupfwðxÞ : xAJg ¼ WoN: ð2:11Þ
The following result will be used to show that Dirichlet–Voronoi cells aremeasurable; here int stands for interior.
For each point of M there is a geodesically convex compact neighborhood
N such that for any p; qAint N; paq; the bisector or Leibnizian plane
LNðp; qÞ ¼ fxAN : RMðp;xÞ ¼ RMðx; qÞg of p; q in N is measurable
with measure 0: ð2:12Þ
Since M is of class C3 with metric tensorfield of class C1; each point of M has acompact neighborhood N with the following properties: (i) NCV ; where V is anopen measurable neighborhood; (ii) N is geodesically convex; (iii) for each pAN thegeodesic segments connecting p with bd N cover N\fpg schlicht; (iv) for each
pAint N there is a diffeomorphism E1p : V-Ed ; the inverse of the exponential
mapping Ep; such that for each geodesic segment G connecting p with a point of bd N
the image E1p ðGÞ is a line segment starting at o: See [5] or [36].
Now, given a point of M; let N be the neighborhood of it just described.N is compact and it is geodesically convex by (ii). Next, let p; qAint N; paq:Then (ii) implies that each geodesic segment connecting p and bd N contains at most
one point of LNðp; qÞ: Clearly, LNðp; qÞ is closed in the compact set N: Hence
E1p ðLNðp; qÞÞ is closed in the compact set E1
p ðNÞ and each ray starting at o meets
E1p ðLNðp; qÞÞ at most once by (iii) and (iv). Thus E1
p ðLNðp; qÞÞ has Lebesgue
measure 0: Since it is compact it is (Jordan) measurable with (Jordan) measure 0:This yields (2.12).
Secondly, f will be investigated. Clearly, (2.1) implies the following:
Given 0oso1; the quotient f ðstÞ=f ðtÞ is increasing and has a positive
lower bound for t40: ð2:13Þ
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 465
As a consequence of (2.1) and (2.13) it will be shown that
There is a constant a40; depending on f ; such that the limit lðsÞ ¼lim
t-þ0ð f ðstÞ=f ðtÞÞ exists and is equal to sa for s40: ð2:14Þ
The existence and positivity of this limit follows from (2.1) and (2.13). Then
lðrsÞ ¼ limt-þ0
f ðrstÞf ðstÞf ðstÞf ðtÞ ¼ lðrÞlðsÞ for r; s40:
The function c : R-R defined by cðuÞ ¼ log lðeuÞ for uAR thus satisfies Cauchy’sfunctional equation. Since f is strictly increasing, (2.1) and (2.13) imply that lðsÞ41for s41 and lðsÞo1 for 0oso1; respectively. Hence cðuÞo0 for uo0 and cðuÞ40for u40: Being a solution of Cauchy’s functional equation it thus follows thatcðuÞ ¼ au for uAR; with suitable a40: See [37]. This yields (2.14).
The next required property of f is the following:
Define g : Rþ-Rþ by gðtÞ ¼ f ð1=t1=dÞ for t40: Then; given 0oZo1;
W; i40; Wai; there is a constant k40 such that
ZgðWtÞ þ ð1 ZÞgðitÞXð1 þ kÞgððZWþ ð1 ZÞiÞtÞ for sufficiently large t40:
ð2:15Þ
To see this note that by (2.14) and the definition of g hold:
gðWtÞB 1
Wad
gðtÞ; gðitÞB1
iad
gðtÞ;
gððZWþ ð1 ZÞiÞtÞB 1
ðZWþ ð1 ZÞiÞad
gðtÞ as t-N:
2.4. The Delone property of Sn; %Sn
Let M; J;w : J-Rþ be as in Theorem 1. Later on we will assume that J isconnected and has positive density.
At first tilings of J by Dirichlet–Voronoi cells will be investigated. Given S ¼fp1;y; pngCM; define the (Dirichlet–Voronoi) cell Di ¼ DðJ;S; piÞ of pi with respectto S in J by
Di ¼ fxAJ: RMðpi; xÞpRMðx; pjÞ for j ¼ 1;y; ng; i ¼ 1;y; n:
The circumradius Ri ¼ RðDiÞ of Di with respect to pi is given by
Ri ¼ inffR40 : BMðpi;RÞ*Dig; i ¼ 1;y; n:
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497466
With these definitions in mind the following will be shown:
Let S ¼ fp1;y; pngCM: If max1¼1;y;n
fRig is sufficiently small; then the
cells D1;y;Dn are measurable and form a tiling of J; ð2:16Þ
i.e., they cover J and their intersections have measure 0: There are finitely manyneighborhoods as described in (2.12) such that their interiors cover the compact setJ: Thus there is a constant 2m40; a Lebesgue number of this covering, see e.g. [13],with the following property: for each qAJ the ball BMðq; 2mÞ is contained in one ofthese neighborhoods. Assume now that maxi¼1;y;nfRigom: Then the following
holds:
oMðDi-DjÞ ¼ 0 for i; j ¼ 1;y; n; iaj: ð2:17Þ
If Di-Dj ¼ |; we are finished. Otherwise, let qADi-DjCJ: Then RMðpi; qÞ ¼RMðq; pjÞpRjom: Since Ri;Rjom; we see that Di-DjCBMðq; 2mÞ: Hence
Di-DjðCBMðq; 2mÞÞCN; where N is one of the neighborhoods as described in
(2.12). Since Di-DjCfxAN : RMðpi; xÞ ¼ RMðx; qjÞg it follows that Di-DjCLNðpi; pjÞ and thus oMðDi-DjÞpoMðLNðpi; pjÞÞ ¼ 0; by (2.12). The proof of
(2.17) is complete. Since D1;y;Dn; J are closed and D1,?,Dn ¼ J;
bd DiC[iaj;
Di-Dj
!,bd J for i ¼ 1;y; n:
This together with (2.17), the measurability of J which implies that oðbd JÞ ¼ 0; andthe fact that D1;y;Dn cover J yields the measurability of the cells D1;y;Dn andthe claim that they tile J; concluding the proof of (2.16).
The next preliminary step is to prove that
there is a constant n40 such that Fnpn f1
n1d
!for n ¼ 1; 2;y; ð2:18Þ
where for Fn see (i) in Theorem 1. Let l ¼ 2 and choose m and Ul ;y; l ¼ 1;y;m; as
in 2.3 such that (2.6)–(2.11) hold. Let nXm: Since the sets J 0l all are measurable in Ed
by (2.9), we may choose In=mm points in each J 0l such that the Euclidean balls with
centers at these points and radius Oð1=In=mm1dÞ cover J 0
l : Here Oð�Þ is a suitable
Landau symbol, the same for each l: If n is sufficiently large, which we assume, the
balls covering J 0l all are contained in U 0
l by (2.9). Consider the h1l -images of the
centers of these balls and let S be the union of all these images, l ¼ 1;y;m: Then#Spn and (2.6), (2.7), (2.9), (2.11), the definition of integrals on M; and (2.14)together yield (2.18).
Assume from now on in this subsection that J is connected and has positivedensity, i.e., satisfies (2.2), but note that these assumptions will only be needed
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 467
beginning with the proof of proposition (2.26). Let Sn ¼ fpn1;y; pnng and Dni ¼DðJ;Sn; pniÞ;Rni ¼ RðDniÞ for i ¼ 1;y; n; n ¼ 1; 2;y : Our aim is to show thefollowing:
There is a constant d41 such that
1
dn1d
p mini; j¼1;y;n
iaj
fRMðpni; pnjÞg; maxi¼1;y;n
fRnigpd
n1d
for n ¼ 1; 2;y : ð2:19Þ
We show the upper estimate first. Its proof is split into a series of steps. In the firststep it will be shown that
Dnia| for i ¼ 1;y; n; n ¼ 1; 2;y : ð2:20Þ
If Dni ¼ |; then
minpASn
fRMðp; xÞg ¼ minpASn\fpnig
fRMðp; xÞg for xAJ:
Now choose qAðint JÞ\Sn and let Tn ¼ ðSn\fpnigÞ,fqg: Then
minpASn
fRMðp; xÞgXminpATn
fRMðp; xÞg for xAJ;
where strict inequality holds for all xAJ sufficiently close to q: Since f is strictlyincreasing by (2.1), S ¼ Sn does not minimize the left hand integral in (2.3). Thiscontradicts our choice of Sn and thus concludes the proof of (2.20).
The next step is to show that
maxxAJ
minpASn
fRMðp; xÞg� �
-0 as n-N: ð2:21Þ
For suppose not. Then there are x40 and a sequence of points xnAJ where n
is from an infinite subsequence of 1; 2;y; such that RMðp; xnÞX2x for eachpASn: Since J is compact, there is a point xAJ such that RMðp; xÞXx for each pASn
and n from a suitable infinite subsequence of the subsequence just considered.Combining this with the assumptions that J has positive density (see (2.2)) and f isstrictly increasing (see (2.1)), we obtain a contradiction to (2.18). This concludes theproof of (2.21).
(2.20), (2.21) and (2.16) yield the following:
maxi¼1;y;n
fRnig-0 as n-N; ð2:22Þ
the cells Dn1;y;Dnn are measurable and form a tiling of J for all
sufficiently large n; ð2:23Þ
there is a compact set I*J such that SnCI for n ¼ 1; 2;y : ð2:24Þ
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497468
The definition of Fn in (i), our choice of Sn ¼ fpn1;y; pnng and of Dn1;y;Dnn;(2.23), the assumptions that f is increasing and wX0 together imply the nextassertion:
Let Tn ¼ fqn1;y; qnngCM and let En1;y;EnnCJ be measurable and form a tilingof J for n ¼ 1; 2;y : Then, if n is sufficiently large,
Fn ¼Z
J
mini¼1;y;n
f f ðRMðpni; xÞÞgwðxÞ doMðxÞ
¼Xn
i¼1
ZDni
f ðRMðpni; xÞÞwðxÞ doMðxÞ
pZ
J
mini¼1;y;n
f f ðRMðqni; xÞÞgwðxÞ doMðxÞ
pXn
i¼1
ZEni
f ðRMðqni; xÞÞwðxÞ doMðxÞ: ð2:25Þ
Before proving the upper estimate in (2.19), a weaker version of it will be shown inthe following step:
There is a constant s40 such that Rnjps
n1d
for suitable jAf1;y; ng;
and all n ¼ 1; 2;y : ð2:26Þ
If (2.26) did not hold, then,
for any ðarbitrarily largeÞ z40 there is an infinite set of n such that
mini¼1;y;n
fRnigXz
n1d
: ð2:27Þ
Take b from (2.2) and v;W from (2.11). By (2.14) one may
choose z40 so large that for suitable t40 holds
oMðJÞWf ð3tÞozdv
3dbf
2t
3
� � f
t
3
� �� �for 0otpt: ð2:28Þ
From now on we consider in the proof of (2.26) only such n for which
ð2:23Þ; ð2:25Þ; and ð2:27Þ for the chosen z hold; and ð2:29Þ
maxi¼1;y;n
fRnigpg; t: ð2:30Þ
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 469
Here for (2.30) we have made use of (2.22) and g; t are as in (2.2) and (2.28),respectively. By (2.29) and (2.23) the cells Dn1;y;Dnn are measurable and form atiling of J: Hence there is a cell Dnj ; say, such that
oMðDnjÞpoMðJÞ
n: ð2:31Þ
Since in this subsection J is assumed to be connected and since the cells Dn1;y;Dnn
tile J by (2.29) and (2.23) and are compact, we may choose a point pADnj-Dnk for
suitable kaj: Then RMðpnj ; pnkÞpRMðpnj ; pÞ þ RMðp; pnkÞ ¼ 2RMðpnj ; pÞp2Rnj and
therefore,
kaj and RMðpnk; xÞðpRMðpnk; pnjÞ þ RMðpnj ; xÞÞp3Rnj for xADnj: ð2:32Þ
Choose qADnj with RMðpnj ; qÞ ¼ Rnj : The definition of Dnj then shows that
RMðq; xÞpRnj
3; RMðpni; xÞX2Rnj
3for xAB ¼ BM q;
Rnj
3
� �; i ¼ 1;y; n: ð2:33Þ
Define fn : J-R by
fnðxÞ ¼ mini¼1;y;n
f f ðRMðpni; xÞÞg for xAJ: ð2:34Þ
Finally, the definition of Fn in (i), our choice of Sn ¼ fpn1;y; pnng; (2.34), (2.29) and(2.23), the measurability of B which follows from (2.30) and (2.2), (2.25), theassumptions that f is increasing and f ;wX0; (2.32),(2.34), (2.11), (2.30), (2.2), (2.34)and (2.33), (2.11), (2.3),(2.34), (2.31), (2.29) and (2.27), (2.30), and (2.28) lead to thefollowing contradiction and thus conclude the proof of (2.26):
Fn ¼Z
J
fnðxÞwðxÞ doMðxÞ ¼Xn
i¼1
ZDni
f ðRMðpni; xÞÞwðxÞ doMðxÞ
pXn
i¼1iaj
ZDni\B
f ðRMðpni; xÞÞwðxÞ doMðxÞ þZ
Dnj \B
f ðRMðpnk; xÞÞwðxÞ doMðxÞ
þZ
B-J
f ðRMðp; xÞÞwðxÞ doMðxÞ
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497470
pXn
i¼1iaj
ZDni\B
f ðRMðpni; xÞÞwðxÞ doMðxÞ þZ
Dnj \B
f ðRMðpnj ; xÞÞwðxÞ doMðxÞ
þZ
Dnj \B
f ð3RnjÞwðxÞ doMðxÞ þZ
B-J
fnðxÞwðxÞ doMðxÞ
Z
B-J
ð fnðxÞ f ðRMðq; xÞÞÞwðxÞ doMðxÞ
pZ
J
fnðxÞwðxÞ doMðxÞ þ oMðDnjÞ f ð3RnjÞW
Rd
nj
3dbf
2Rnj
3
� � f
Rnj
3
� �� �v
pFn þ1
noMðJÞW f ð3RnjÞ
1
n
zdv
3dbf
2Rnj
3
� � f
Rnj
3
� �� �oFn:
We come to the final step of the proof of the upper estimate in (2.19). If thisestimate did not hold, then
for any (arbitrarily large) n40 there is an infinite set of n such that
Rnk4n
n1d
for suitable kAf1;y; ng: ð2:35Þ
Take b from (2.2), e from (2.5), where I is from (2.24), s from (2.26), and v;W from(2.11). By (2.14) we may choose n40 so large that
(i) n4s;(ii) for suitable j40 hold
Wesdf3s
n1d
! vnd
3dbf
2t
3
� � f
t
3
� �� �o0 for
n
n1d
otoj: ð2:36Þ
From now on consider in the proof of the upper estimate in (2.19) only such n that
ð2:5Þ; where I is from ð2:24Þ; ð2:23Þ; ð2:25Þ and ð2:35Þ hold; and ð2:37Þ
maxi¼1;y;n
fRnigpg; W;j; ð2:38Þ
where g; W;j are as in (2.2), (2.5) and (2.36), respectively. Given such an n; let j be as
in (2.26). Then Rnjps=n1=d : Choose pnlapnj with RMðpnj ; pnlÞp2Rnjp2s=n1=d : Such
a choice is possible since in this subsection J is connected by assumption, compare
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 471
the argument that led to (2.32). Then
laj and RMðpnl ; xÞp3Rnjp3s
n1d
for xADnj : ð2:39Þ
By (2.37), proposition (2.35) holds for n and the chosen n: Let k be as in (2.35). Since
Rnjps=n1=don=n1=doRnk by (2.26), (2.36i) and (2.37) and (2.35),
kaj ð2:40Þ
follows. Choose pADnk with RMðpnk; pÞ ¼ Rnk: The definition of the cell Dnk
and (2.35) then imply that RMðpni; pÞXRMðpnk; pÞ ¼ Rnk4n=n1=d for i ¼ 1;y; n:Hence,
RMðp; xÞpRnk
3; RMðpni; xÞX2Rnk
3for xAB ¼ BM p;
Rnk
3
� �; i ¼ 1;y; n: ð2:41Þ
Finally, the definition of Fn in (i), our choice of Sn ¼ fpn1;y; pnng; (2.37) and (2.25),the measurability of B which follows from (2.38) and (2.2), (2.39), (2.40), (2.37),(2.23) and (2.25), the assumptions that f is increasing and f ;wX0; (2.39), (2.34),(2.11), (2.41), (2.38), (2.2), (2.11), (2.38), (2.26), (2.37) and (2.5), (2.26), (2.37), (2.35),(2.38) and (2.36) together yield a contradiction and thus conclude the proof of theupper estimate in (2.19):
Fn ¼Xn
i¼1
ZDni
f ðRMðpni; xÞÞwðxÞ doMðxÞ
pXn
i¼1iaj;k
ZDni\B
f ðRMðpni; xÞÞwðxÞ doMðxÞ þZ
Dnj \B
f ðRMðpnl ; xÞÞwðxÞ doMðxÞ
þZ
Dnk\B
f ðRMðpnk; xÞÞwðxÞ doMðxÞ þZ
B-J
f ðRMðp; xÞÞwðxÞ doMðxÞ
pXn
i¼1iaj;k
ZDni\B
f ðRMðpni; xÞÞwðxÞ doMðxÞ þZ
Dnj \B
f ðRMðpnj ; xÞÞwðxÞ doMðxÞ
þZ
Dnj \B
f3s
n1d
!wðxÞ doMðxÞ þ
ZDnk\B
f ðRMðpnk; xÞÞwðxÞ doMðxÞ
þZ
B-J
fnðxÞwðxÞ doMðxÞ Z
B-J
ð fnðxÞ f ðRMðp; xÞÞÞwðxÞ doMðxÞ
¼Z
J
fnðxÞwðxÞ doMðxÞ þ oMðDnjÞf3s
n1d
!W vRd
nk
3dbf
2Rnk
3
� � f
Rnk
3
� �� �
pFn þ1
nWe sdf
3s
n1d
! 1
n
vnd
3dbf
2Rnk
3
� � f
Rnk
3
� �� �oFn:
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497472
We proceed to the proof of the lower estimate in (2.19). The proof will be split intotwo steps. In the first step it will be shown that
there is a constant f40 such that for j ¼ 1;y; n; holds
FnpZ
J
gn1ðxÞwðxÞ doMðxÞ fn
f1
n1d
!; gn1ðxÞ ¼ min
i¼1;y;niaj
f f ðRMðpni; xÞÞg: ð2:42Þ
It is enough to show this for sufficiently large n: The argument that led to (2.32)shows that for each xAJ there is a point in fpn1;y; pnj1; pnjþ1;y; pnng at distance
at most 3d=n1d ; where d is from the upper estimate in (2.19). Hence the cells in J
corresponding to the set fpn1;y; pnj1; pnjþ1;y; pnng have circumradius at most
3d=n1=d : Then (2.16) shows that for all sufficiently large n and any j ¼ 1;y; n; then 1 cells corresponding to fpn1;y; pnj1; pnjþ1;y; pnng are measurable and tile J:
Thus there is a cell of measure at least oMðJÞ=ðn 1Þ4oMðJÞ=n; say the cellcorresponding to pnk: Clearly, kaj: Noting (2.24) and (2.5) we see that this cell has
circumradius at least w=n1=d ; where w40 is a suitable constant. Hence there is a point
p of this cell with RMðpnk; pÞXw=n1=d : The definition of cells then shows that
RMðpni; pÞXw=n1=d for i ¼ 1;y; d; iaj: As before,
RMðp; xÞp w
3n1d
; RMðpni; xÞX 2w
3n1d
for xAB ¼ BM p;w
3n1d
!; i ¼ 1;y; n; iaj; ð2:43Þ
compare (2.33) and (2.41). Now, the definitions of Fn in (i) and of gn1 in (2.42), themeasurability of B which follows from (2.2) for sufficiently large n; the assumptionthat J satisfies (2.2), (2.43), the assumptions that f is increasing and f ;wX0; and(2.11) imply, for sufficiently large n; that
FnpZ
J
min mini¼1;y;n
iaj
f f ðRMðpni; xÞÞg; f ðRMðp; xÞÞ
8<:
9=;wðxÞ doMðxÞ
¼Z
J
minfgn1ðxÞ; f ðRMðp;xÞÞgwðxÞ doMðxÞ
pZ
J\B
gn1ðxÞwðxÞ doMðxÞ þZ
B-J
f ðRMðp; xÞÞwðxÞ doMðxÞ
¼Z
J
gn1ðxÞwðxÞ doMðxÞ Z
B-J
ðgn1ðxÞ f ðRMðp;xÞÞÞwðxÞ doMðxÞ
pZ
J
gn1ðxÞwðxÞ doMðxÞ 1
n
wdv
3dbf
2w
3n1d
! f
w
3n1d
! !:
An application of (2.14) then yields (2.42).
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 473
In the second step we assume that the lower estimate in (2.19) does not hold. Then,
for any ðarbitrarily smallÞ c40 there is an infinite set of n such that
RMðpnj; pnkÞpc2
n1d
for suitable j; kAf1;y; ng; jak: ð2:44Þ
Let d; e;f;W be as in the upper estimate in (2.19), (2.5), where I is from (2.24),(2.42), and (2.11), respectively. By (2.14) we may choose 0oo;co1 so small that
bddoWfd
n1d
!þ bcdWf
2c
n1d
!off
1
n1d
!; ð2:45Þ
f ðð1 þ cÞtÞpð1 þ oÞf ðtÞ for 0otpt; ð2:46Þwith a suitable t40: In the remaining part of the proof of the lower estimate in (2.19)we consider only n for which
ð2:25Þ and ð2:44Þ hold; and ð2:47Þ
maxi¼1;y;n
fRnigpd
n1d
pminfg; tg; c
n1d
pg; ð2:48Þ
where g is as in (2.2) and t as in (2.46). For (2.48) we have used the upper estimate in(2.19). Propositions (2.47) and (2.44) and our choice of cðo1Þ imply that
RMðpnk; xÞpð1 þ cÞRMðpnj ; xÞ for xeB ¼ BM pnj;c
n1d
!;
RMðpnk; xÞp2c
n1d
for xAB: ð2:49Þ
Now, combining (2.42), (2.47) and (2.25) applied with n 1 instead of n; whereEni ¼ Dni for i ¼ 1;y; n; iaj; k and Enk ¼ Dnj,Dnk; the measurability of B which
follows from (2.24) and (2.5) for sufficiently large n; the assumptions that f isincreasing and f ;wX0; (2.49), (2.46), (2.48) and (2.2), (2.11), (2.48) and (2.2), theassumptions that f is increasing and f ;wX0 and (2.45), we obtain for sufficientlylarge n the following contradiction which concludes the proof of the lower estimatein (2.19):
FnpZ
J
mini¼1;y;n
iaj
f f ðRMðpni; xÞÞgwðxÞ doMðxÞ fn
f1
n1d
!
pXn
i¼1iaj
ZDni
f ðRMðpni; xÞÞwðxÞ doMðxÞ þZ
Dnj
f ðRMðpnk; xÞÞwðxÞ doMðxÞ
fn
f1
n1d
!
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497474
pXn
i¼1iaj
ZDni
f ðRMðpni; xÞÞwðxÞ doMðxÞ þZ
Dnj \B
f ðð1 þ cÞRMðpnj ; xÞÞwðxÞ doMðxÞ
þZ
B-Dnj
f2c
n1d
!wðxÞ doMðxÞ f
nf
1
n1d
!
pXn
i¼1iaj
ZDni
f ðRMðpni; xÞÞwðxÞ doMðxÞ þZ
Dnj \B
f ðRMðpnj ; xÞÞwðxÞ doMðxÞ
þ oZ
Dnj \B
f ðRMðpnj ; xÞÞwðxÞ doMðxÞ þ 1
nbcdW f
2c
n1d
! f
nf
1
n1d
!
pFn þ1
nb ddoW f
d
n1d
!þ 1
nb cdW f
2c
n1d
! f
nf
1
n1d
!oFn:
For n ¼ 1; 2;y; choose %Sn ¼ f %pn1;;y; %pnngCJ in the following way: if in (i) the
infimum is considered only for S with #S ¼ n for which SCJ (instead of SCM),
then it is attained for S ¼ %Sn: If J is connected and has positive density, then a proofsimilar to the one which led to (2.19), but easier in some details, shows that
there is a constant %d41 such that
1
%dn1d
p mini;j¼1;y;n
iaj
fRMð %pni; %pnjÞg; maxi¼1;y;n
f %Rnigp%d
n1d
for n ¼ 1; 2;y; ð2:50Þ
where the %Rni are the circumradii of the cells in J corresponding to f %pn1;y; %pnng:
2.5. The constant div
Let M ¼ Ed ; J ¼ ½0; 1 d ;w ¼ 1: If the infimum in (i) is attained for, say, Tn ¼ftn1;y; tnng; then TnC½0; 1 d : Thus an application of (2.19) or (2.50), and (2.14)together with the assumptions that f is increasing and fX0 yields the following: if
Gn ¼ infTCEd
#T¼n
Z½0;1 d
mintAT
f f ðjjt ujj2Þg du
( )¼Z½0;1 d
mini¼1;y;n
f f ðjjtni ujj2Þg du;
then Gn=f ð1=n1dÞ is bounded between positive constants. Thus, defining
div ¼ lim infn-N
Gn
f1
n1d
!; ð2:51Þ
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 475
it follows that
divARþ: ð2:52Þ
We will show that
div ¼ limn-N
Gn
f1
n1d
!: ð2:53Þ
Let d41 be as in (2.19). Choose m41 (arbitrarily close to 1) and let m be such that
Gm
f1
m1d
!pm div;
fd
km1d
!
f1
km1d
!pm
fd
m1d
!
f1
m1d
! for k ¼ 1; 2;y: ð2:54Þ
This is possible by the definition of div in (2.51), (2.52) and (2.1). An application of(2.13) shows that
fd
km1d
!
fd
m1d
! f ðtÞXft
k
� �for 0otp
d
m1d
; k ¼ 1; 2;y : ð2:55Þ
For k ¼ 1; 2;y; dissect the cube ½0; 1 d in the usual way into kd cubelets, each of
edge length 1=k: The kd affine transformations
Aa : u1-u1 þ a1
k;y; ud-
ud þ ad
k; z-
fd
km1d
!
fd
m1d
! z for ðu; zÞAEdþ1;
where a ¼ ða1;y; adÞTAf0;y; k 1gd ;
map the set
fðu; zÞAEdþ1 : uA½0; 1 d ; 0pzp mini¼1;y;m
f f ðjjtmi ujj2Þgg:
onto kd non-overlapping sets. Together, these sets contain the set
fðu; zÞAEdþ1 : uA½0; 1 dg; 0pzp mini¼1;y;m
aAf0;y;k1gd
f f ðjjAatmi ujj2Þg:
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497476
This follows from (2.55) and the fact that by (2.19) the cells in ½0; 1 d corresponding
to Tm ¼ ftm1;y; tmmg have circumradius at most d=m1d : Hence the definition of Gkd m
and (2.54) show that
Gkd mpkd � 1
kd
fd
km1d
!
fd
m1d
! Gmpm
f1
km1d
!
f1
m1d
! Gm;
and thus, again by (2.54),
Gkd m
f1
km1d
!pmGm
f1
m1d
!pm2 div:
By (2.14) there is k0 such that
f1
km1d
!
f1
ðk þ 1Þm1d
0@
1Apm for kXk0:
Finally, given nXkd0 m; let kXk0 be such that kdmpnoðk þ 1Þd
m: Thus,
Gn
f1
n1d
! pGkd m
f1
km1d
!f
1
km1d
!
f1
ðk þ 1Þm1d
0@
1A
0BBBBBB@
1CCCCCCApm3 div for all sufficiently large n:
Since m41 was arbitrary, this together with (2.51) implies (2.53).Next, (2.53) will be extended to arbitrary cubes:
Let C be a cube in Ed : Then
Hm ¼ infTCEd
#T¼m
ZC
mintAT
f f ðjjt ujj2Þg du
� �BdivjCj
aþdd f
1
m1d
!as m-N; ð2:56Þ
where a is as in (2.14). We may assume that C ¼ n½0; 1 d ¼ ½0; n d for suitable n40:
By (2.19) the circumradii of the cells corresponding to TmC½0; 1 d are bounded above
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 477
by d=m1=d : Thus by (2.14),
Gm ¼Z½0;1 d
mintATm
f f ðjjt ujj2Þg duBZ½0;1 d
minntAnTm
f f ðjjnt nujj2Þg du1
na
¼Z½0;n d
minvAnTm
f f ðjjv wjj2Þg dw1
naþdX
Hm
naþd¼ Hm
jCjaþd
d
;
or
GmjCjaþd
d \Hm as m-N:
The same argument with ½0; 1 d and C exchanged shows that
Hm\GmjCjaþd
d as m-N;
and therefore,
HmBGmjCjaþd
d as m-N:
This together with (2.53) yields (2.56).A suitable affine transformation applied to (2.56) finally gives the following:
Let qð�Þ be a positive definite quadratic form on Ed and C0 a cube in
Ed with respect to the norm qð�Þ1=2: Then
infT 0CEd
#T 0¼m
ZC0
mintAT 0
f f ðqðt uÞ12Þg du
� �Bdiv ðdet qÞ
a2d jC0j
aþdd f
1
m1d
!as m-N: ð2:57Þ
2.6. The asymptotic formula
Choose Sn ¼ fpn1;y; pnng and Dn1;y;Dnn for n ¼ 1; 2;y; as in 2.4. Define
SnðKÞ ¼ fqniASn : Dni-Ka|g; nðKÞ ¼ #SnðKÞ for measurable KCJ:
If in the following an integral is written omitting the integrand, the integrand is to be
wðxÞd=ðaþdÞ and byR a
we mean the ath power ofR:
In the first part of the proof of the asymptotic formula (i) it will be shown that
Fn\div
ZJ
wðxÞd
aþd doMðxÞ� �aþd
d
f1
n1d
!as n-N: ð2:58Þ
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497478
Let l41 and choose Ul ;Vl ;y; for l ¼ 1;y;m; as in 2.3 such that
ð2:6Þ2ð2:11Þ hold: ð2:59Þ
The first step of the proof of (2.58) is to show that
there is a constant n41 such that1
nf
1
n1d
!pFnpn f
1
n1d
!
for n ¼ 1; 2;y ð2:60Þ
The upper estimate was proved in (2.18). For the proof of the lower estimate choose
a set C in Jl ; say, such that C0 is a cube with respect to the norm qlð�Þ12: Then C is
measurable and the definition of Fn in (i), (2.11), the assumptions that f is increasingand f ;wX0; (2.59) and (2.6), the definition of integrals in M; (2.59) and (2.7)show that
FnX
ZC
minpASn
f f ðRMðp; xÞÞgwðxÞ doMðxÞ
X1
lvðdet qlÞ
12
ZC0
minsASnðCÞ0
f1
lqlðs uÞ
12
� �� �du
X1
lv ðdet qlÞ
12 inf
T 0CEd
#T 0¼nðCÞ
ZC0
mintAT 0
f1
lqlðt uÞ
12
� �� �du
� �:
If nðCÞ did not tend to N as n-N; the latter expression would not tend to 0; incontradiction to the upper estimate for Fn in (2.60). Hence nðCÞ-N as n-N:Thus, using (2.14), (2.57), and the assumption that f is increasing, we concludefurther that
Fn\1
l1þa v ðdet qlÞaþd2d divjC0j
aþdd f
1
nðCÞ1d
0@
1A
X1
l1þa v ðdet qlÞaþd2d divjC0j
aþdd f
1
n1d
!as n-N:
This readily implies the lower estimate in (2.60), concluding the proof of (2.60).
Noting that Fnpnf ð1=n1=dÞ by (2.60), the assumption that f is strictly increasing,and (2.14), we obtain the following proposition as a byproduct of the last part of theproof of (2.60):
Let CCJl ; say; such that C0 is a cube with respect to the norm qlð�Þ12:
Then there is a constant x with 0oxo1 such that xnpnðCÞpn for
sufficiently large n: ð2:61Þ
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 479
We come to the second step of the proof of (2.58). Noticing (2.9), we may choosesets CliCJl ; i ¼ 1;y; il ; l ¼ 1;y;m; with the following properties:
The sets CliCJ; i ¼ 1;y; il ; l ¼ 1;y;m; are compact and pairwise
disjoint: ð2:62Þ
C0li is a cube with respect to the norm qlð�Þ
12: Thus; in particular;
Cli is measurable: ð2:63Þ
Xl;i
ZCli
doMðxÞX1
l
ZJ
doMðxÞ: ð2:64Þ
By the definition of SnðCliÞ; (2.62) and (2.22) (which has been proved without theassumptions that J is connected and has positive density),
the sets SnðCliÞ are pairwise disjoint for sufficiently large n: Thus; ð2:65Þ
Xl;i
nðCliÞpn for sufficiently large n: ð2:66Þ
Finally, it follows from (2.63) and (2.61) that
there is a constant s with 0oso1 such that snpnðCliÞpn
for i ¼ 1;y; il ; l ¼ 1;y;m; and sufficiently large n: ð2:67Þ
In the third step note that (2.60), (2.67) and (2.66) together yield the followingpropositions, where
L ¼ lim infn-N
Fn
f1
n1d
! ð40Þ:
There is an infinite subsequence of 1; 2;y; such that Fn=f ð1=n1dÞ-L
as n-N in this subsequence: ð2:68Þ
There are constants sli40 such that nðCliÞBslin
for i ¼ 1;y; il ; l ¼ 1;y;m; as n-N in the same subsequence as in ð2:68Þ:Furthermore;
Xl;i
slip1: ð2:69Þ
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497480
In the final step of the proof of (2.58) assume that n is from the subsequenceconsidered in (2.68) and (2.69) and that it is so large that (2.65) and (2.66)hold. Then the definition of Fn in (i), our choice of Sn; (2.62), (2.63), the definitionof SnðCliÞ; the definition of integrals on M; the assumptions that f is increasingand f ;wX0; (2.59) and (2.6), (2.10), the inclusion CliCJl ; (2.14), (2.57), (2.69),the definition of integrals on M; (2.59) and (2.6), (2.10), (2.14), (2.64),(2.15), the assumption that f is increasing, (2.64), (2.69), and (2.14) imply thefollowing:
Fn ¼Z
J
minpASn
f f ðRMðp; xÞÞgwðxÞ doMðxÞ
X
Xl;i
ZCli
minpASnðCliÞ
f f ðRMðp;xÞÞgwðxÞ doMðxÞ
X1
l
Xl;i
ðdet qlÞ12
ZC0
li
minsASnðCliÞ0
f1
lqlðs uÞ
12
� �� �wl du
B1
l1þa divX
l;i
ðdet qlÞaþd2d jC0
lijaþd
d wl f1
ðslinÞ1d
0@
1A
X1
l1þaþ1þaþdd
divX
l;i
ZCli
wðxÞd
aþd doMðxÞ� �aþd
d
f1
ðslinÞ1d
0@
1A
B1
l3þaþad
divX
l;i
ZCli
f
RCli
slin
!1d
0B@
1CA
¼ 1
l3þaþad
divX
l;i
ZCli
Xl;i
RCliP
l;i
RCli
f
1
slinRCli
!1d
0BB@
1CCA
\1
l4þaþad
div
ZJ
f
1
Pl;i
RCli
slinRCli
Pl;i
RCli
0B@
1CA
1d
0BBBBB@
1CCCCCA
¼ 1
l4þaþad
div
ZJ
f
Pl;i
RCliP
l;islin
!1d
0B@
1CA
X1
l4þaþad
div
ZJ
f1
l
ZJ
1
n
� �1d
0@
1AB
1
l4þaþ2ad
div
Z aþdd
J
f1
n1d
!
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 481
as n-N in the subsequence considered in (2.68) and (2.69). Since l41 wasarbitrary, we have thus shown that
lim infn-N
Fn
f1
n1d
!Xdiv
ZJ
wðxÞd
aþd doMðxÞ� �aþd
d
;
which is equivalent to (2.58).In the second part of the proof of the asymptotic formula (i) we consider, in
addition to Fn; the quantity
%Fn ¼ inf%SCJ
# %S¼n
ZJ
min%pA %S
f f ðRMð %p; xÞÞgwðxÞ doMðxÞ� �
:
Our aim is to show that
Fnp %Fntdiv
ZJ
wðxÞd
aþd doMðxÞ� �aþd
d
f1
n1d
!as n-N: ð2:70Þ
Let l41 and choose Ul ;Vl ;y; l ¼ 1;y;m; as in 2.3 such that
ð2:6Þ2ð2:11Þ hold: ð2:71Þ
In the first step of the proof of (2.70) we choose sets Cli; i ¼ 1;y; il ; l ¼ 1;y;m;with the following properties, where s40 (is small) and Wl is as in (2.10):
For each l ¼ 1;y;m; the sets Cli; i ¼ 1;y; il ; are compact;
non-overlapping and such that
JlCCl1,?,ClilCVl ;Cli-Jla|: ð2:72Þ
C0li is a cube with respect to the norm qlð�Þ
12 ðthus; in particular;
Cli is measurableÞ and such that
ðdet qlÞ12jC0
lijWd
aþdl ¼ s
ZJ
wðxÞd
aþd doMðxÞ: ð2:73Þ
Xl;i
ZCli
plZ
J
: ð2:74Þ
Let i0 ¼ i1 þyþ im: Since
Xl;i
ðdet qlÞ12jC0
lijWd
aþdl pl
daþd
þ1X
l;i
ZCli
wðxÞd
aþd doMðxÞpl3
ZJ
wðxÞd
aþd doMðxÞ
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497482
by (2.71) and (2.10), (2.7), the definition of integrals on M; and (2.74), it followsfrom (2.73) that
si0pl3: ð2:75Þ
By (2.14) we may choose k0 so large that
1
iad0
f1
k1d
!pl f
1
ði0kÞ1d
0@
1A for kXk0; ð2:76Þ
f1
ði0kÞ1d
0@
1Apl f
1
ði0ðk þ 1ÞÞ1d
0@
1A for kXk0: ð2:77Þ
In the next step of the proof of (2.70) we consider first the case where n is of theform n ¼ i0k for some kXk0: Assume that
infT 0CEd
#T 0¼k
ZC0
li
mintAT 0
f f ðqlðt uÞ12Þg du
( )
is attained for T 0 ¼ T 0lik where T 0
likCC0li;#T 0
lik ¼ k: ð2:78Þ
Applying the result (2.19) for M ¼ Ed and J ¼ C0li; i ¼ 1;y; il ; l ¼ 1;y;m; we
obtain the following, where, omitting indices, T 0lik ¼ ft1;y; tkg and D0
1;y;D0k are
the cells in C0li corresponding to the set T 0
lik and the norm qlð�Þ1=2:
There is a t41; independent of l; i; k; such that any two distinct points
from T 0lik ¼ ft1;y; tkg have distance at least 1=tk1=d and each of the
cells D01;y;D0
k has circumradius at most t=k1=d ðwith respect to the
norm qlð�Þ1=2Þ: ð2:79Þ
Since J is (Jordan) measurable, bd J is measurable and oMðbd JÞ ¼ 0: Hence
ðCli-bd JÞ0 is measurable in Ed with jðCli-bd JÞ0j ¼ 0: A result on Jordan
measurable sets of measure 0 in Ed then says that the parallel set[uAðCli-bd JÞ0
ðu þ nBdÞ
has arbitrarily small measure if n40 is sufficiently small. Here Bd is the solid
Euclidean unit ball in Ed : This remark together with (2.79) yields the next
proposition, where Dj ¼ h1l ðD0
jÞ:
There are at most oðkÞ cells D0j in C0
li with ðDj-bd JÞ0a|; where we
may choose the same Landau symbol oðkÞ for all l; i: ð2:80Þ
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 483
If D0jCðCli-int JÞ0; let %tj ¼ tj; if D0
j-ðCli-bd JÞ0a|; choose for %tj an arbitrary point
of D0j-ðCli-bd JÞ0: Let %T 0
lik be the set of all points %tj which can be obtained in this
way and let %Tlik ¼ h1l ð %T 0
likÞCCli: Then the definition of integrals on M in (2.4), the
assumptions that f is increasing and f ;wX0; (2.72), (2.71) and (2.6), (2.7), (2.10),(2.14), (2.2), (2.79),(2.80), (2.57), and (2.73) yield the following:Z
Cli-J
min%pA %Tlik
f f ðRMð %p; xÞÞgwðxÞ doMðxÞ
plZðCli-JÞ0
min%tA %T0
lik
f f ðlqlð%t uÞ12Þg du Wlðdet qlÞ
12
tl1þaZ
C0li
mintATlik
f f ðqlðt uÞ12Þg du Wlðdet qlÞ
12 þ oðkÞb 2dtd
kf
2t
k1d
!
Bl1þa div ðdet qlÞaþd2d jC0
lijaþd
d Wl f1
k1d
!þ oð1Þ f
1
k1d
!
tl1þaþ1þaþdd div s
ZJ
wðxÞd
aþd doMðxÞ� �aþd
d
f1
k1d
!as k-N;
where the symbols oðkÞ; oð1Þ are independent of l; i: ð2:81Þ
For n ¼ iok let %Tn ¼S
l;i%Tlik: Then # %Tnpn and %TnCJ: The definitions of Fn and %Fn;
(2.72), (2.71) and (2.9), i0 ¼ i1 þ?þ im; (2.81), (2.75) and (2.14) together imply that
Fnp %FnpZ
J
min%pA %Tn
f f ðRMð %p; xÞÞgwðxÞ doMðxÞ
pX
l;i
ZCli-J
min%pA %Tlik
f f ðRMð %p; xÞÞgwðxÞ doMðxÞ
tl3þ2ai0 div sZ
J
wðxÞd
aþd doMðxÞ� �aþd
d
f1
k1d
!
¼ l3þ2a div
ZJ
wðxÞd
aþd doMðxÞ� �aþd
d ði0sÞaþd
d
ia0df
1
k1d
!
tl3þ2aþ3aþ3dd div
ZJ
wðxÞd
aþd doMðxÞ� �aþd
d
f1
n1d
!as n ¼ iok-N:
Since l41 was arbitrary, this proves (2.70) for n of the form n ¼ iok: This, togetherwith (2.77) yields (2.70) for general n; compare the argument in the last part of theproof of (2.53).
Having proved (2.58) and (2.70), the proof of the asymptotic formula (i) with
SCM and also with %SCJ is complete.
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497484
2.7. Uniform distribution of Sn; %Sn
Assume that J is connected and has positive density, i.e. satisfies (2.2). We willprove that
Sn is uniformly distributed in J with respect to the density wd=ðaþdÞ
as n-N: ð2:82Þ
A proof which is similar to that of (2.80), but slightly more complicated in several
details since it deals with M instead of Ed ; and which makes use of (2.19), (2.24), and(2.5) yields the following:
Let ZCJ be measurable with oMðZÞ ¼ 0:
Then nðZÞð¼ #fi : Dni-Za|gÞ ¼ oðnÞ as n-N: ð2:83Þ
More difficult is the proof of
Let KCJ be measurable with 0ooMðKÞpoMðJÞ: Then
nðKÞBZðKÞn as n-N; where ZðKÞ ¼Z
K
ZJ
�: ð2:84Þ
If oMðKÞ ¼ oMðJÞ; then (2.84) follows from (2.20), which shows that nðJÞ ¼ n; andfrom (2.83). Suppose now that oMðKÞooMðJÞ and assume that (2.84) does not hold:
nðKÞfZðKÞn as n-N: ð2:85Þ
From (2.18) and the definitions of SnðKÞ; nðKÞ in 2.6 it follows that
n f1
n1d
!XFnX
ZK
minpASn
f f ðRMðp; xÞÞgwðxÞ doMðxÞ
¼Z
K
minpASnðKÞ
f f ðRMðp;xÞÞgwðxÞ doMðxÞ
X infTCM
#T¼nðKÞ
ZK
minpAT
f f ðRMðp; xÞÞgwðxÞ doMðxÞ� �
:
This can hold only if nðKÞ-N as n-N: Now, applying the asymptotic formula (i)to K instead of J; we conclude further that
n f1
n1d
!\div
ZK
aþdd f
1
nðKÞ1d
0@
1A as n-N:
Using the assumption that f is strictly increasing and (2.14) it follows that nðKÞ=n isbounded below by a positive constant. An analogous statement holds for the set
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 485
L ¼ J\K : By (2.20) and (2.83),
nðKÞ þ nðLÞ ¼ n þ oðnÞ as n-N: ð2:86Þ
(2.85), (2.86) and the above statements on nðKÞ=n and nðLÞ=n show that
there are W; i40 such that Wo1oi or io1oW and nðKÞB WZðKÞn;
nðLÞBiZðLÞn as n-N; where n is from a suitable subsequence
of 1; 2;y : ð2:87Þ
From now on consider in the proof of (2.82) only n from the subsequence in (2.87). We
consider only the case that Wo1oi: By (i), the definition of Sn; J ¼ K,L;K-L ¼ |;the definitions of SnðKÞ; nðKÞ;SnðLÞ; nðLÞ and nðKÞ; nðLÞ-N as n-N (see (2.87)),the asymptotic formula (i) for K and L instead of J; (2.87), (2.14), ZðLÞ ¼ 1 ZðKÞ(see (2.84)), (2.15), Wo1oi; (2.87) and (2.86), and (2.3) we arrive at the followingcontradiction, where k40; and thus conclude the proof of (2.84):
Fn ¼Z
K
minpASnðKÞ
f f ðRMðp; xÞÞgwðxÞ doMðxÞ
þZ
L
minpASnðLÞ
f f ðRMðp; xÞÞgwðxÞ doMðxÞ
\ div
Z aþdd
K
f1
nðKÞ1d
0@
1Aþ
Z aþdd
L
f1
nðLÞ1d
0@
1A
0@
1A
B div
Z aþdd
J
ZðKÞaþd
d f1
ðWZðKÞnÞ1d
0@
1Aþ ZðLÞ
aþdd f
1
ðiZðLÞnÞ1d
0@
1A
0@
1A
B div
Z aþdd
J
ZðKÞ f1
ðWnÞ1d
0@
1Aþ ð1 ZðKÞÞ f
1
ðinÞ1d
0@
1A
0@
1A
X ð1 þ kÞ div
Z aþdd
J
f1
ðZðKÞWn þ ð1 ZðKÞÞinÞ1d
0@
1A
B ð1 þ kÞ div
Z aþdd
J
f1
ðnðKÞ þ nðLÞÞ1d
0@
1ABð1 þ kÞ div
Z aþdd
J
f1
n1d
!
B ð1 þ kÞ Fn as n-N in the subsequence from ð2:87Þ:
Applying (2.83) and (2.84) to measurable KCJ yield the following:
#ðK-SnÞpnðKÞ ¼ oðnÞ as n-N for oMðKÞ ¼ 0;
#ðK-SnÞ ¼ nðKÞ oðnÞBZðKÞn as n-N for 0ooMðKÞpoMðJÞ:
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497486
This concludes the proof of (2.82). A similar, even simpler proof shows that
%Sn is uniformly distributed in J with respect to the density wd=ðaþdÞ
as n-N: ð2:88Þ
2.8. Conclusion
Having proved (2.58) and (2.70), (2.19) and (2.50), and (2.82) and (2.88), the proofof Theorem 1 is complete.
3. Distortion of high resolution vector quantization
3.1. Distortion of vector quantization
Let C1;y;Cn be n measurable sets which tile a measurable set C in Ed and let
c1;y; cnAEd : To each signal xAC assign the codevector or codeword ci from thecodebook fc1;y; cng; where xACi: (In case of ambiguity, which can occur only for x
in a set of measure 0; choose for ci any codevector where xACi:) Common measuresfor the quality of the thus defined encoder or (vector) quantizer on C can be described
as follows: let f : ½0;þNÞ-½0;þNÞ with f ð0Þ ¼ 0 be increasing, jj � jj a norm on Ed ;
and w : C-Rþ continuous. Up to normalization, w is the density of the source whichgenerates the signals x: Then the corresponding (average) distortion of our quantizeris defined to be
Xn
i¼1
ZCi
f ðjjci ujjÞwðuÞ du:
How should the cells Ci and the codevectors ci be chosen in order to minimize thedistortion? Given c1;y; cn; the distortion is minimized if C1CD1;y;CnCDn whereD1;y;Dn are the Dirichlet–Voronoi cells in C corresponding to c1;y; cn and usingjj � jj: Then the distortion is
ZC
mini¼1;y;n
f f ðjjci ujjÞgwðuÞ du:
Thus the minimum distortion is given by
inf%SCC# %S¼n
ZC
mincA %S
f f ðjjci ujjÞgwðuÞ du
� �:
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 487
3.2. Asymptotic results
Clearly, Theorem 2 translates into the following result.
Theorem 3. Let f : ½0;þNÞ-½0;þNÞ satisfy the growth condition (2.1), let
a40 be the corresponding constant, and let jj � jj be a norm on Ed : Then there is a
constant div40; depending only on f and jj � jj; such that the following statement
holds: let CCEd be compact and measurable with jCj40 and w : C-Rþ continuous.
Then
(i) the minimum distortion in the above sense of a quantizer on C with source density w
with n codewords is asymptotically equal to
div
ZC
wðuÞd
aþd du
� �aþdd
f 1
n1d
!as n-N:
If, in addition, C is connected and has positive density, i.e. satisfies (2.2), then the
following claims hold: let ðSnÞ be a sequence of codebooks in C; where #Sn ¼ n for
n ¼ 1; 2;y; and such that the infimum in (i) is attained for Sn: Then
(ii) there is a constant d41 such that Sn is ð1=dn1=d ; d=n1=dÞ-Delone in J for n ¼1; 2;y;
(iii) Sn is uniformly distributed in C with density wd=ðaþdÞ as n-N:
Remark. Theorem 1 admits an analogous interpretation.
4. Error of approximation of probability measures by discrete measures
4.1. Quantization of probability measures
Let f : ½0;NÞ-½0;NÞ be monotone increasing, such that f ðtÞ ¼ 0 only for t ¼ 0
and let jj � jj be a norm on Ed : Then we can define on the space of all probability
measures on Ed the following notion of distance:
r f ðP;QÞ ¼ inf
ZEd�Ed
f ðjjx yjjÞ dmðx; yÞ� �
for probability measures P;Q;
where the infimum is extended over all Borel probability measures m on Ed � Ed
which satisfy the following condition:
PðBÞ ¼ mðB � EdÞ;QðCÞ ¼ mðEd � CÞ for all Borel sets B;CCEd :
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497488
r f is a slight extension of the Wasserstein- or Kantorovich metric on the space of all
probability measures on Ed : The question arises as to how well can probabilitymeasures be approximated by discrete measures and to describe the bestapproximating discrete measures. Let Dn denote the space of all discrete probability
measures on Ed the support of which consists of at most n points, n ¼ 1; 2;y: Then
for the error of best approximation of a probability measure P on Ed by probabilitymeasures from Dn the following representation holds:
r f ðP;DnÞ ¼ infDADn
fr f ðP;DÞg ¼ infSCEd
#S¼n
ZEd
minsAS
f f ðjju sjjÞgdPðuÞ� �
:
For this and for references compare Graf and Luschgy [20].
4.2. Asymptotic approximation results
Considering the above, it is clear that Theorem 2 yields the following result onapproximation of probability measures.
Theorem 4. Let f : ½0;þNÞ-½0;þNÞ satisfy the growth condition (2.1), let a40 be
the corresponding constant, and let jj � jj be a norm on Ed : Then there is a constant
div40; depending only on f and jj � jj; such that the following statement holds: let
JCEd be compact and measurable with jJj40 and P a probability measure on J with
continuous density p : J-Rþ: Then
(i) the error of best approximation of P by probability measures from Dn is
asymptotically equal to
div
ZC
pðuÞd
aþd du
� �aþdd
f1
n1d
!as n-N:
If, in addition, J is connected and has positive density, i.e. satisfies (2.2), then the
following claims hold: let ðSnÞ be a sequence of supports in J; where #Sn ¼ n for
n ¼ 1; 2;y; and such that for suitable corresponding discrete probability measures in
Dn the minimum error is attained. Then
(ii) there is a constant d41 such that Sn is ð1=dn1=d ; d=n1=dÞ-Delone in J for n ¼1; 2;y;
(iii) Sn is uniformly distributed in J with density wd=ðaþdÞ as n-N:
Remark. Obviously, Theorem 1 yields a similar result for the approximation ofprobability measures by discrete probability measures on Riemannian manifolds.
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 489
5. Error of numerical integration formulae
5.1. Error of numerical integration
Let JCEd be a compact measurable set in Ed with jJj40; let F be a class of
Riemann integrable functions g : J-R and w : J-Rþ a continuous (weight)function. For given sets of n nodes N ¼ fp1;y; pngCJ and n weights W ¼fw1;y;wngCR the error of the numerical integration formula
ZJ
gðuÞwðuÞ duEXn
i¼1
gðpiÞwi for gAF;
is defined by
EðF;w;N;WÞ ¼ supgAF
ZJ
gðuÞwðuÞ du Xn
i¼1
gðxiÞwi
����������
( ): ð5:1Þ
The minimum error is then
EðF;w; nÞ ¼ infNCJ;#N¼n
WCR;#W¼n
fEðF;w;N;WÞg: ð5:2Þ
Now the problems arise to determine EðF;w; nÞ and to describe the optimal choicesof nodes and weights. While for general F not much can be said, we consider a spaceof functions for which rather precise information can be given.
5.2. Asymptotic results
A modulus of continuity is an increasing continuous function f : ½0;þNÞ-½0;þNÞwith f ð0Þ ¼ 0 such that
f ðs þ tÞpf ðsÞ þ f ðtÞ for s; tX0: ð5:3Þ
Given a modulus of continuity f and a norm jj � jj on Ed ; define the corresponding
Holder class of functions for JCEd by
H f ðjj � jj; JÞ ¼ fg : J-R : jgðuÞ gðvÞjpf ðjju vjjÞ for u; vAJg: ð5:4Þ
Theorem 5. Let f : ½0;þNÞ-½0;þNÞ be a modulus of continuity which satisfies the
growth condition (2.1), let a40 be the corresponding constant, and let jj � jj be a norm
on Ed : Then there is a constant div40; depending only on f and jj � jj; such that we have
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497490
the following: let JCEd be compact and measurable with jJj40 and w : J-Rþ
continuous. Then
(i) for the minimum error of a numerical integration formula with weight function w
for the Holder class H f ¼ H f ðJ; jj � jjÞ holds the asymptotic formula,
EðH f ;w; nÞBdiv
ZJ
wðuÞd
aþd du
� �aþdd
f 1
n1d
!as n-N:
If, in addition, J is connected and has positive density, i.e. satisfies (2.2), then the
following assertions hold: let ðNnÞ be a sequence of sets of nodes in J; where #Nn ¼ n
for n ¼ 1; 2;y; and such that the infimum in (4.2) is attained for Nn and suitable
corresponding Wn: Then
(ii) there is a constant d41 such that Nn is ð1=dn1=d ; d=n1=dÞ-Delone in J for
n ¼ 1; 2;y;
(iii) Nn is uniformly distributed in J with density wd=ðaþdÞ as n-N:
Remark. An analogous result holds for Riemannian d-manifolds instead of Ed andjj � jj: The asymptotic formula (5.3) for d ¼ 2; jj � jj ¼ jj � jj2 and w ¼ 1 is due to
Babenko [2]. For general d and arbitrary norm a related yet different asymptoticformula was given by Chernaya [8]. The fact that in Chernaya’s result w is assumedto be Lebesgue integrable can be reduced to the continuous case by Lusin’s theorem.A pertinent idea is described also by Sobol’ [45].
5.3. Proof of Theorem 5
First, the following will be shown:
Let N ¼ fp1;y; pngCJ and let h : J-R be defined by
hðuÞ ¼ minpAN
f f ðjjp ujjÞg: Then hAH f : ð5:5Þ
To see this, let u; vAJ: By exchanging u and v; if necessary, we may assume thathðuÞXhðvÞ: Then
0p hðuÞ hðvÞ ¼ minpAN
f f ðjjp ujjÞg minqAN
f f ðjjq vjjÞg
¼ f ðjjp ujjÞ f ðjjq vjjÞ for suitable p; qAN
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 491
p f ðjjq ujjÞ f ðjjq vjjÞ
¼ f ðjjq v þ v ujjÞ f ðjjq vjjÞ
p f ðjjq vjj þ jjv ujjÞ f ðjjq vjjÞ
p f ðjjq vjjÞ þ f ðjjv ujjÞ f ðjjq vjjÞ ¼ f ðjju vjjÞ;
where we have used the assumptions that f is increasing and that is a modulus ofcontinuity.
Secondly, let N ¼ fp1;y; pngCJ and W ¼ fw1;y;wngCR: Define Di ¼ fuAJ :
jjpi ujjr jjpj ujj for j ¼ 1;y; ng; Ei ¼ Di\ðD1,y,Di1Þ; %wi ¼R
EiwðuÞ du;
i ¼ 1;y; n: %W ¼ f %w;y; %wng:Then
EðH f ;w;N;WÞXEðH f ;w;N; %WÞ ¼Z
J
minpAN
f f ðjjp ujjÞgwðuÞ du: ð5:6Þ
Since J is the disjoint union of E1;y;En; gAH f ;EiCDi; the assumptions that f is
increasing and wX0; the definition of h; hAH f and hðpiÞ ¼ 0 (see (5.5)) togetherimply (5.6):
EðH f ;w;N; %WÞ ¼ supgAH f
ZJ
gðuÞwðuÞ du Xn
i¼1
gðpiÞZ
Ei
wðuÞ du
����������
( )
p supgAH f
Xn
i¼1
ZEi
jgðuÞ gðpiÞjwðuÞ du
( )
pXn
i¼1
ZEi
f ðjjpi ujjÞwðuÞ du ¼Xn
i¼1
ZEi
minpjAN
f f ðjjpj ujjÞgwðuÞ du
¼Z
J
minpAN
f f ðjjp ujjÞgwðuÞ du ¼Z
J
hðuÞwðuÞ du
¼Z
J
hðuÞwðuÞ du Xn
i¼1
hðpiÞwipEðH f ;w;N;WÞ:
Thirdly, it follows from (5.6) that
EðH f ;w; nÞ ¼ infNCJ;#N¼n
WCR;#W¼n
fEðH f ;w;N;WÞg ¼ infNCJ#N¼n
fEðH f ;w;N; %WÞg
¼ infNCJ#N¼n
ZJ
minpAN
f f ðjjp ujjÞgwðuÞ du
� �:
Now apply Theorem 2 to obtain Theorem 5.
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497492
6. Volume approximation of convex bodies
6.1. Best volume approximation of convex bodies
Let C be a convex body in Ed ; i.e. a compact convex subset of Ed with non-emptyinterior and let Pc
ðnÞ be the set of all convex polytopes with n facets which are
circumscribed to C: Let dV ð�; �Þ denote the symmetric difference metric on the space
of all convex bodies in Ed : The problems arise to determine or estimate thequantity
dV ðC;PcðnÞÞ ¼ inffdV ðC;PÞ : PAPc
ðnÞg
and to describe the polytopes PAPcðnÞ for which the infimum is attained, the best
approximating polytopes of C in PcðnÞ with respect to dV : See [25,27] for the literature
on this and related problems.
6.2. Asymptotic approximation results
Refining a result of Boroczky [4] (asymptotic formula), the following result hasbeen proved by Gruber [31] (asymptotic formula and Delone property) andGlasauer and Gruber [19] (uniform distribution property). A weaker version of it(without the error term in the asymptotic formula) can be obtained as a simpleconsequence of Theorem 1 above and its proof; compare the correspondingargument in [30].
Theorem 6. There is a constant divd140 with the following property: let C be a
convex body in Ed (with boundary) of class C3 and with Gauss curvature kC40: Let
ðPnÞ;PnAPcðnÞ; be a sequence of best approximating polytopes of C: Then we have the
following statements:
(i) dV ðC;PcðnÞÞB1
2divd1
Zbd C
kCðxÞ1
dþ1 dsðxÞ� �dþ1
d11
n2
d1
þ O 1
n2
d1þ 13ðd1Þe
0@
1A
as n-N; for any e40;
(ii) there is a constant d41 such that the inradius of each facet of Pn is at least
1=dn1=ðd1Þ and the circumradius at most d=n1=ðd1Þ for n ¼ d þ 1; d þ 2;y;(iii) the sets C-bd Pn are uniformly distributed on bd C; endowed with the ordinary
surface area measure s; with respect to the density k1=ðdþ1ÞC as n-N:
Remark. For d ¼ 3; a sharper result was given in [32]. See [24,26] for earlierpertinent results.
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 493
7. The isoperimetric problem for convex polytopes in Minkowski spaces
7.1. The isoperimetric problem in Minkowski spaces
Let Ed be endowed with an additional norm. While in this new normed space thenatural notion of volume is the ordinary volume Vð�Þ or j � j (Haar measure), severaldifferent proposals for surface area have been made by Busemann [6], Holmes andThompson [35] and Benson [3]. These amount to the introduction of an o-symmetricconvex body I ; the isoperimetrix, depending in a suitable way on the solid unit ball ofthis norm. The surface area SIðCÞ of a convex body C then is defined to be
SI ðCÞ ¼ limd-þ0
VðC þ dIÞ VðCÞd
; where C þ dI ¼ fx þ dy : xAC; yAIg:
See the book [47]. For n ¼ d þ 1; d þ 2;y; let Pn be a convex polytope which hasminimum isoperimetric quotient
SI ðPnÞd
VðPnÞd1
among all convex polytopes in Ed with n facets. The problem arises to determine orestimate the minimum isoperimetric quotient and to describe the form of theminimizing polytopes Pn:
7.2. Asymptotic isoperimetric results
A result of Diskant [10,11], which generalizes a theorem of Lindelof [38] (alsoproved by Minkowski [40]) for the Euclidean case, says that after a suitablehomothety has been applied to Pn; we may assume that Pn is circumscribed to I ; i.e.PnAPc
ðnÞðIÞ: Then it is easy to see that
SIðPnÞd
VðPnÞd1¼ ddVðPnÞ:
Since the isoperimetric quotient is minimal for Pn; the polytope Pn
thus is best approximating of I in PcðnÞðIÞ: Hence Theorem 5 yields the following
result.
Theorem 7. There is a constant divd140 such that the following hold: let I be an
isoperimetrix in Ed of class C3 with kI40: For n ¼ d þ 1; d þ 2;y; let Pn be a convex
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497494
polytope with n facets circumscribed to I and minimum isoperimetric quotient
SI ðPnÞd=VðPnÞd1: Then the following assertions hold:
(i) SI ðPnÞd
VðPnÞd1B ddVðIÞ þ dd
2divd1
Zbd I
kCðxÞ1
dþ1 dsðxÞ� �dþ1
d1 1
n2
d1
þ O1
n2
d1þ 1
3ðd1Þe
!as n-N; for any e40;
(ii) there is a constant d41 such that the inradius of each facet of Pn is at least
1=dn1=ðd1Þ and the circumradius at most d=n1=ðd1Þ for n ¼ d þ 1; d þ 2;y;(iii) the sets C-bd Pn are uniformly distributed on bd C; endowed with the ordinary
surface area measure s; with respect to the density k1=ðdþ1ÞC as n-N:
Remark. For d ¼ 3; a sharper result was given in [32].
Acknowledgments
For their helpful remarks in the preparation of this article and references to theliterature I am indebted to Professors Paul Goodey, Siegfried Graf, Robert Gray andFranz Schnitzer.
References
[1] V.F. Babenko, Asymptotically sharp bounds for the remainder for the best quadrature
formulas for several classes of functions, Mat. Zametki 19 (1976) 313–322; Math. Notes 19 (1976)
187–193.
[2] V.F. Babenko, On the optimal error bound for cubature formulas on certain classes of continuous
functions, Anal. Math. 3 (1977) 3–9.
[3] R.V. Benson, Euclidean Geometry and Convexity, McGraw-Hill, New York, 1966.
[4] K. Boroczky Jr., The error of polytopal approximation with respect to the symmetric difference
metric and the Lp metric, Israel J. Math. 117 (2000) 1–28.
[5] H. Brauner, Differentialgeometrie, Vieweg, Braunschweig, 1981.
[6] H. Busemann, The foundations of Minkowskian geometry, Comment. Math. Helv. 24 (1950)
156–187.
[7] E.V. Chernaya, An asymptotic sharp estimate for the remainder of weighted cubature formulas that
are optimal on certain classes of continuous functions, Ukraın. Mat. Zh. 47 (1995) 1405–1415;
Ukrainian J. Math. 47 (1995) 1606–1618.
[8] E.V. Chernaya (=Chornaya), On the optimization of weighted cubature formulae on certain classes
of continuous functions, East J. Approx. 1 (1995) 47–60.
[9] J.H. Conway, N.J.A. Sloane, Sphere packings, Lattices and Groups, 3rd Edition, Springer, New
York, 1999.
[10] V.I. Diskant, Making precise the isoperimetric inequality and stability theorems in the theory of
convex bodies, Trudy Inst. Mat. (Novosibirsk) 14 (1989) 98–132.
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 495
[11] V.I. Diskant, Generalizations of a theorem of Lindelof, Mat. Fiz. Anal. Geom. 4 (1997) 59–64.
[12] Q. Du, V. Faber, M. Gunzburger, Centroidal Voronoi tessellations: applications and algorithms,
SIAM Rev. 41 (1999) 637–676.
[13] Encyclopedia of Mathematics, Kluwer Academic Publishers, Dordrecht, 1988–1994.
[14] G. Ewald, D. Larman, C.A. Rogers, The direction of the line segments and of the r-dimensional balls
on the boundary of a convex body in Euclidean space, Mathematika 17 (1970) 1–20.
[15] G. Fejes Toth, A stability criterion to the moment theorem, Studia Sci. Math. Hungar. 34 (2001)
209–224.
[16] L. Fejes Toth, Sur la representation d’une population infinie par un nombre fini d’elements, Acta
Math. Acad. Sci. Hungar. 10 (1959) 299–304.
[17] L. Fejes Toth, Lagerungen in der Ebene, auf der Kugel und im Raum, 1st Edition, 1953, 2nd edition.
Springer, Berlin, 1972.
[18] A. Gersho, Asymptotically optimal block quantization, IEEE Trans. Inform. Theory IT-25 (1979)
373–380.
[19] S. Glasauer, P.M. Gruber, Asymptotic estimates for best and stepwise approximation of convex
bodies III, Forum Math. 9 (1997) 383–404.
[20] S. Graf, H. Luschgy, Foundations of quantization of Probability Distributions, in: Lecture Notes
Mathematics, Vol. 1730, Springer, Berlin, 2000.
[21] R.M. Gray, D.L. Neuhoff, Quantization, IEEE Trans. Inform. Theory IT-44 (1998) 2325–2383.
[22] R.M. Gray, D.L. Neuhoff, J.K. Omura, Process definitions of distortion rate functions and source
coding theorems, IEEE Trans. Inform. Theory IT-21 (1975) 524–532.
[23] R.M. Gray, D.L. Neuhoff, P.C. Shields, A generalization of Ornstein’s d-bar distance with
applications to information theory, Ann. Probab. 3 (1975) 315–328.
[24] P.M. Gruber, Volume approximation of convex bodies by circumscribed convex polytopes, Discrete
Math. Theor. Comput. Sci. 4 (1991) 309–317.
[25] P.M. Gruber, Aspects of approximation of convex bodies, in: P.M. Gruber, J.M. Wills (Eds.),
Handbook of Convex Geometry A, North-Holland, Amsterdam, 1993, pp. 319–345.
[26] P.M. Gruber, Asymptotic estimates for best and stepwise approximation of convex bodies II, Forum
Math. 5 (1993) 383–404.
[27] P.M. Gruber, Comparisons of best and random approximation of convex bodies by convex
polytopes, Rend. Sem. Mat. Univ. Palermo, Ser. II 50 (Suppl.) (1997) 189–216.
[28] P.M. Gruber, Optimal arrangements of finite point sets in Riemannian 2-manifolds, Trudy Mat. Inst.
Steklov. 225 (1999) 148–155; Proc. Steklov Math. Inst. 225 (1999) 160–167.
[29] P.M. Gruber, A short analytic proof of Fejes Toth’s theorem on sums of moments, Aequationes
Math. 58 (1999) 291–295.
[30] P.M. Gruber, Optimal configurations of finite point sets in Riemannian 2-manifolds, Geom. Dedicata
84 (2001) 271–320.
[31] P.M. Gruber, Error of asymptotic formulae for volume approximation of convex bodies in Ed ;
Monatsh. Math. 135 (2002) 279–304.
[32] P.M. Gruber, Error of asymptotic formulae for volume approximation of convex bodies in E3; Trudy
Mat. Inst. Steklov. 239 (2002) 106–117; Proc. Steklov Math. Inst. 239 (2002) 96–107.
[33] P.M. Gruber, Optimale quantisierung, Math. Semesterber 49 (2003) 227–251.
[34] E. Hlawka, The Theory of Uniform Distribution, A B Academic Publishers, Berkhamsted, 1984.
[35] R.D. Holmes, A.C. Thompson, N-dimensional area and content in Minkowski spaces, Pacific J.
Math. 85 (1979) 77–110.
[36] S. Kobayashi, K. Nomizu, Foundations of Differential Geometry, Wiley, New York, 1969.
[37] M. Kuczma, An Introduction to the Theory of Functional Equations and Inequalities (Cauchy’s
equation and Jensen’s inequality), PWN, Warszawa, 1985.
[38] L. Lindelof, Proprietes generales des polyedres qui, sous une etendue superficielle donnee, renferment
le plus grand volume, St. Petersburg Bull. Acad. Sci. 14 (1869) 258–269; Math. Ann. 2 (1870)
150–159.
[39] B. Matern, O. Persson, On the extremum properties of the equilateral triangular lattice and the
regular hexagonal network, Fortech. Over. Rapp. Uppsala Stockholm 7 (1965) 15.
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497496
[40] H. Minkowski, Allgemeine Lehrsatze uber die convexen Polyeder, Nachr. Kgl. Ges. Wiss. Gottingen
Math.-Phys. Kl. 1897 (1911) 198–219; Ges. Abh. 2 (1911) 103–121.
[41] D.L. Neuhoff, R.M. Gray, L.D. Davisson, Fixed rate universal block source coding with a fidelity
criterion, IEEE Trans. Inform. Theory IT-21 (1975) 511–523.
[42] D.J. Newman, The hexagon theorem, IEEE Trans. Inform. Theory IT-28 (1982) 137–139.
[43] V.I. Polovinkin, Asymptotically optimal cubature weight formulas, Sibirskiı Mat. Zh. 70 (1989)
151–160; Siber. Math. J. 70 (1989) 289–297.
[44] R. Schneider, Convex Bodies: The Brunn–Minkowski Theory, Cambridge University Press,
Cambridge, 1993.
[45] I.M. Sobol’, Quadrature formulas for functions in several variables that satisfy the general Lipschitz
condition, Zh. Vychisl. Mat. i Mat. Fiz. 29 (1989) 935–941; U.S.S.R. Comput. Math. Math. Phys. 29
(1989) 201–206.
[46] S.L. Sobolev, Introduction to the Theory of Cubature Formulas, Nauka, Moscow, 1974.
[47] A.C. Thompson, Minkowski Geometry, Cambridge University Press, Cambridge, 1996.
[48] P.L. Zador, Asymptotic quantization error of continuous signals and the quantization dimension,
IEEE Trans. Inform. Theory IT-28 (1982) 139–148.
ARTICLE IN PRESSP.M. Gruber / Advances in Mathematics 186 (2004) 456–497 497