Convex Bodies with Few Faces

by

Keith Ball(1)
Texas A&M University
arXiv:math/9201203v1 [math.MG] 26 Oct 1989

College Station, TX 77843

and

Alain Pajor
U.E.R. de Mathématiques
Université de Paris VII
2 Place Jussieu
75251 PARIS CEDEX 05

Abstract. It is proved that if u1 , . . . , un are vectors in Rk , k ≤ n, 1 ≤ p < ∞ and

 Xn  p1
1 p
r= |ui |
k 1
then the volume of the symmetric convex body whose boundary functionals are
±u1 , . . . , ±un , is bounded from below as

1 1
|{x ∈ Rk : |hx, ui i| ≤ 1 for every i}| k ≥ √ .
ρr
An application to number theory is stated.

A.M.S. (1980) Subject Classification: 52A20, 10E05

(1)
Partially supported by N.S.F. DMS-8807243.

1
§0. Introduction.
In [V], Vaaler proved that if Qn = [− 21 , 21 ]n is the central unit cube in Rn and U is
a subspace of Rn then the volume |U ∩ Qn |, of the section of Qn by U is at least 1. This
result may be reformulated as follows: if u1 , . . . , un are vectors in Rk , 1 ≤ k ≤ n whose
n
|ui |2 ≤ k then
P
Euclidean lengths satisfy
1

1
|{x ∈ Rk : |hx, ui i| ≤ 1 for every i}| k ≥ 2.

|ui |2 ≤ k is replaced
P
A related theorem, (Theorem 1, below) in which the condition
by max |ui | ≤ 1 was proved by Carl and Pajor [C-P] and Gluskin [G]. Gluskin’s methods
i
enable him to obtain sharp results in limiting cases which in turn have applications in
harmonic analysis. Results closely related to Theorem 1 were also obtained by Bárány and
Füredi [B-F] and Bourgain, Lindenstrauss and Milman [B-L-M].

Theorem 1. There is a constant δ > 0 so that if u1 , . . . , un ∈ Rk , 1 ≤ k ≤ n are vectors

of length at most 1 then

1 δ
|{x ∈ Rk : |hx, ui i| ≤ 1 for every i}| k ≥ p .
1 + log nk
The estimate is best possible if n is at most exponential in k, apart from the value of
the constant δ. This is demonstrated by an example which had appeared some time earlier
in a paper of Figiel and Johnson, [F-J]. Theorem 1 gives a lower bound on the volume
ratios of the unit balls of k-dimensional subspaces of ℓn∞ and hence on the distance of these
subspaces from Euclidean space.
Regarding Theorem 1 as a “p = ∞” version of Vaaler’s “p = 2” result, Kashin asked
whether a similar result holds for 2 < p < ∞. This question is answered in the affirmative
by the following theorem.
n 1
Theorem 2. Suppose u1 , . . . , un ∈ Rk with k ≤ n, 1 ≤ p < ∞ and let r = ( k1 |ui |p ) p .
P
1
Then
( √
2 2
1 √ if p≥2
|{x ∈ Rk : |hx, ui )| ≤ 1 for every i}| ≥
k
1
pr

r if 1 ≤ p ≤ 2.

2
The lower bound is best possible (up to a constant) provided ep k ≤ n ≤ ek .
Remark. The slightly stronger result for p ≥ 2 is isolated since for p = 2 it gives back
exactly Vaaler’s result.

Theorem 1 follows immediately from Theorem 2 by a standard optimisation argument.

If (ui )n1 in Rk all have norm at most 1 then for any p ∈ [1, ∞),

 Xn  p1  n  p1
1 p
|ui | ≤
k 1 k

so that

√
1 2 2
|{x: |hx, ui i| ≤ 1 for every i}| ≥ √
k
1
p( nk ) p
2
(for p ≥ 2) and the latter is at least √ √ n
when p = 2(1 + log nk ).
e 1+log k

With the careful use of well-known methods for estimating the entropy of convex
bodies it is possible to obtain more general (but less precise) estimates than that provided
by Theorem 2; (see [B-P]). The purpose of this paper is to provide a very short proof of
Theorem 2 and, a fortiori, Theorem 1.
Vaaler originally proved his theorem because of its applications to the geometry of
numbers. The last section of this paper includes a statement of the generalisation of Siegel’s
lemma which follows from Theorem 2.

3
§1. The lower bound.
The proof of Theorem 2 makes use of the following result from [Me-P] which was
designed to extend Vaaler’s theorem in a different direction: it estimates the volumes of
sections of the unit balls of the spaces ℓnp , 1 ≤ p ≤ ∞. For 1 ≤ p ≤ ∞, n ∈ N let

 n
X 
Bpn n
= x∈R : p
|xi | ≤ 1
1

be the unit ball of ℓnp .

Theorem 3. Let U be a k-dimensional subspace of Rn ; if 1 ≤ p ≤ q ≤ ∞ then

|Bpn ∩ U | |Bqn ∩ U |
≤ .
|Bpk | |Bqk |
Remark. The case p = 2, q = ∞ is Vaaler’s theorem since then, the left side is 1 and the
inequality states that

n k
|B∞ ∩ U | ≥ |B∞ | ≥ 2k .

For notational convenience, the proof of Theorem 2 is divided into several short lem-
mas. The first is no more than a convenient form of Hölder’s inequality. For k ∈ N, S k−1
will denote the Euclidean sphere in Rk and σ = σk−1 , the rotationally invariant probability
measure on S k−1 . Also let vk be the volume of the Euclidean unit ball in Rk .

Lemma 4. Let C and B be symmetric convex bodies in Rk with Minkowski gauges k · kC

and k · kB respectively. Then for p > 0

 |C|  k1  − p1
k+p
Z
≥ kxkpC dx .
|B| k|B| B

Proof.

 |C|  k1  Z  k1
vk
= kθk−k
C dσ(θ)
|B| |B| S k−1
 Z  kθk k Z kθk−1  k1
kvk B B
k−1
= r drdσ(θ)
|B| S k−1 kθkC 0

4
   k1
1
kxkB k
Z 
= dx
|B|
B kxkC
 − p1
1 kxkB −p
Z 
≥ dx
|B| B kxkC
 − p1
k+p
Z
p
= kxkC dx .
k|B| B

(Lemma 4 appears in [Mi-P] as Corollary 2.2.)

Lemma 5. Suppose u1 , . . . , un ∈ Rk with k ≤ n and 1 ≤ p < ∞. Then

1
|{x ∈ Rk : |hxi , ui i| ≤ 1 for every i}| k
 n − p1
k+pX 1
Z
p
≥2 |hx, ui i| dx .
k |Bpk | Bpk
i=1

Proof. Define T : R → R by (T x)i = hx, ui i, 1 ≤ i ≤ n and let U = T (Rk ). The

k n

problem is to estimate from below

1 1
|T −1 (B∞
n
)| k = |T −1 (U ∩ B∞
n
)| k .

By Theorem 3,

 |U ∩ B n |  k1
n k 1 p
|U ∩ B∞ | ≥2
|Bpk |
and so
1
 |T −1 (B n )|  k1
−1 n p
|T (B∞ )| k ≥2 .
|Bpk |

Regard T as an operator: ℓkp → ℓnp . Then by Lemma 4,

 |T −1 (B n )|  k1  − p1
k+p
Z
p p
2 ≥2 kT xk dx
|Bpk | k|Bpk | Bpk
 n − p1
k+p
Z X
p
=2 |hx, ui i| dx .
k|Bpk | Bpk 1

Proof of Theorem 2. Let (ui )n1 and p be as above. For each i let vi be the unit vector
in the direction of ui . By Lemma 5,

5
 1
|{x ∈ Rk : |hx, ui i| ≤ 1 for every i}| k
 n − p1
k+p X 1
Z
p
≥2 |hx, ui i| dx
k i=1 |Bpk | Bpk
 n − p1
k+p X 1
Z
=2 |ui |p · k |hx, vi i|p dx
k 1
|B p | Bp
k

 X n − p1  − p1
1 k+p
Z
p p
≥2 |ui | min |hx, vi| dx
k i=1 |Bpk | Bpk

where the minimum is taken over all vectors v of Euclidean length 1. So to complete the
proof it suffices to show that for such a vector v,

  p1  pp
k+p
Z
p 2 if p≥2
|hx, vi| dx ≤
|Bpk | Bpk 2 if 1 ≤ p < 2.

Let (x(j) )k1 and (v (j) )k1 be the coordinates of the vectors x and v in Rk . For p ≥ 2,
observe that the functions (x(j) v (j) ) on Bpk form a conditionally symmetric sequence, so
n
v (j)2 = 1),
P
by Khintchine’s inequality and Hölder’s inequality (for
1

 k p  p1
k+p
Z X
(j) (j)
x v dx
|Bpk | Bpk 1
r  Z X k  p2  p1
p k+p (j)2 (j)2
≤ x v dx
2 |Bpk | Bpk 1
r  n  p1
p k+p
Z X
≤ |x(j) |p · v (j)2 dx
2 |Bpk | Bpk 1
r   p1
p k+p
Z
(1) p
= |x | dx
2 |Bpk | Bpk
r
p
= .
2

For 1 ≤ p < 2 it is easily checked that

6
   p1
k+p
Z
p
|hx, vi| dx
|Bpk | Bpk
  21
1
Z
1
≤ (k + p) p hx, vi2 dx
|Bpk | Bpk
  12
1
Z
1
(1) 2
= (k + p) p (x ) dx
|Bpk | Bpk

and the last expression can be (rather roughly) estimated by 2 using standard inequalities
involving logarithmically concave functions.

Remark. The proof of Theorem 2 can be simplified even further if the integration over
Bpk is replaced by integration over S k−1 (and Hölder’s inequality applied here). The proof
was presented as above because Lemma 5 has some intrinsic interest: for example it may
be used to recover Gluskin’s precise estimate as follows. Suppose m ∈ N and the vectors
(zi )m k
1 ∈ R satisfy

  m − 21
|zi | ≤ log 1 + , 1 ≤ i ≤ m.
k

For ε > 0, let W (ε) be the set

n
k (j) 1o
x ∈ R : max |x | ≤ 1, max |hx, zi i| ≤ :
j i ε
k
that is, W (ε) is the intersection of the cube B∞ with m “bands” of width at most
2 1
log(1 + m
p
ε k
). Then |W (ε)| k → 2 as ε → 0, uniformly in k and m. To see this, apply
Lemma 5 with n = k + m, the first k, ui ’s being the standard basis vectors of Rk and the
remaining m being the vectors (εzi )m
1 . If ej is a standard basis vector,

k+p
Z
|hx, ej i|p dx = 1
|Bpk | Bpk

1
and so Lemma 5 (and the proof of Theorem 2) show that for each p ≥ 2, |W (ε)| k ≥
m p p 2 p
m −p −1 2
2(1 + k ( 2 ) ε (log(1 + k ))
2) p and the latter is at least √
1+ e·ε
if p = max(2,
m
2 log(1 + k )).

7
 As was briefly mentioned earlier, more general estimates than that of Theorem 2 are
obtained in [B-P] (for entropy numbers instead of volumes). It is worth noting however
that even the argument of Theorem 2 can be used to give the following: there is a constant
c so that if u1 , . . . , un ∈ Rk , k ≤ n and T : ℓk2 → ℓn∞ is given by (T x)i = hui , xi, 1 ≤ i ≤ n,
then the k th entropy number of T satisfies

√  n  p1 
c p 1X p n  p1
ek (T ) ≤ √ |ui | 1 + log
k k 1 k

and hence

r
ec n
ek (T ) ≤ √ 1 + log · kT k
k k
(taking p = 2(1 + log nk )). To obtain this one uses Schütt’s estimates, [5], for the entropy
numbers of the formal identity from ℓnp to ℓn∞ in place of the result of Meyer and Pajor,
and the dual Sudakov inequality of Pajor and Tomczak, [P-T] in place of the application
of Hölder’s inequality.

8
§2. An application to linear forms.
As stated in the introduction, Vaaler’s original result has applications to the geometry
of numbers. One such, a sharpened form of Siegel’s lemma, is given in [B-V]. Using the
arguments of Bombieri and Vaaler and Theorem 2, one can obtain the generalisation of
their result, contained in Theorem 6, below. Some notation is needed. If A is a k × n
matrix of reals with independent rows (1 ≤ k ≤ n), denote by vj = vj (A), 1 ≤ j ≤ k, the
rows of A. Let (ei )n1 be the standard basis of Rn and denote by ci , the distance (in the
Euclidean norm) of ei from the span of the vj ’s in Rn . (So if Ai is the matrix with k + 1
rows, v1 , . . . , vn , ei then

det(Ai A∗i )
c2i = for 1 ≤ i ≤ n.)
det(AA∗ )
Theorem 6. Let A be a k × n matrix with rank k and integral entries. With the notation
above, the system Ax = 0 admits n − k linearly independent solutions

(r)
z (r) = (z1 , . . . , zn(r) ) ∈ Zn , 1 ≤ r ≤ n − k

so that for every p ≥ 2,

n 1
1 X p p√
r 
Y (r) −1 p
max |zi | ≤D c det AA∗
i 2 n−k 1 i
1≤r≤n−k

where D denotes the G-C-D of all k × k determinants extracted from A.

Remark. The principal importance of such a generalisation of Bombieri and Vaaler’s
result is that it takes into account, more strongly, the form of the matrix A. If the ci ’s are
all about the same size, then for p > 2, the expression

  1
1 X p p
ci
n−k
is small compared with the corresponding expression in which p is replaced by 2.

9
 References.
[B-F] I. Bárány and Z. Füredi, Computing the volume is difficult, Discrete Comput. Geom.
2 (1987), 319-326.
[B-L-M] J. Bourgain, J. Lindenstrauss and V.D. Milman, Approximation of zonoids by zono-
topes, Acta Math. 162 (1989), 73-141.
[B-P] K.M. Ball and A. Pajor, On the entropy of convex bodies with “few” extreme points,
in preparation.
[B-V] E. Bombieri and J. Vaaler, On Siegel’s lemma, Invent. Math. 73 (1983), 11-32.
[C-P] B. Carl and A. Pajor, Gelfand numbers of operators with values in a Hilbert space,
Invent. Math. 94 (1988), 479-504.
[F-J] T. Figiel and W.B. Johnson, Large subspaces of ℓn∞ and estimates of the Gordon-Lewis
constants, Israel J. Math. 37 (1980), 92-112.
[G] E.D. Gluskin, Extremal properties of rectangular parallelipipeds and their applications
to the geometry of Banach spaces, Math. Sbornik, 136 (178) (1988), 85-95.
[Me-P] M. Meyer and A. Pajor, Sections of the unit ball of ℓnp , J. Funct. Anal. 80 (1988),
109-123.
[Mi-P] V.D. Milman and A. Pajor, Isotropic position and inertia ellipsoids and zonoids of
the unit ball of a normed n-dimensional space. Israel seminar on G.A.F.A. (1987-88),
Springer-Verlag, Lecture notes in Math. #1376 (1989).
[P-T] A. Pajor and N. Tomczak-Jaegermann, Subspaces of small codimension of finite-
dimensional Banach spaces, Proc. A.M.S. 97 (1986), 637-642.
[S] C. Schütt, Entropy numbers of diagonal operators between symmetric Banach spaces,
Journal Approx. Th. 40 (1984), 121-128.
[V] J.D. Vaaler, A geometric inequality with applications to linear forms, Pacific J. Math.
83 (1979), 543-553.

10