HAL Id: hal-00811793
https://hal-upec-upem.archives-ouvertes.fr/hal-00811793
Preprint submitted on 11 Apr 2013
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Restricted invertibility and the Banach-Mazur distance to the cube
Pierre Youssef
To cite this version:
Pierre Youssef. Restricted invertibility and the Banach-Mazur distance to the cube. 2012. �hal-
00811793�
RESTRICTED INVERTIBILITY AND THE BANACH-MAZUR DISTANCE TO THE CUBE
PIERRE YOUSSEF
A
BSTRACT. We prove a normalized version of the restricted invertibility principle obtained by Spielman-Srivastava in [15]. Applying this result, we get a new proof of the proportional Dvoretzky-Rogers factorization theorem recovering the best current estimate in the symmetric setting while we improve the best known result in the nonsymmetric case. As a consequence, we slightly improve the estimate for the Banach-Mazur distance to the cube: the distance of every n-dimensional normed space from `
n∞is at most (2n)
56. Finally, using tools from the work of Batson-Spielman-Srivastava in [2], we give a new proof for a theorem of Kashin-Tzafriri [11] on the norm of restricted matrices.
1. I NTRODUCTION
Given an n × m matrix U , viewed as an operator from `
m2to `
n2, the restricted invertibility problem asks if we can extract a large number of linearly independent columns of U and provide an estimate for the norm of the restricted inverse. If we write U
σfor the restriction of U to the columns U e
i, i ∈ σ ⊂ {1, . . . , m}, we want to find a subset σ, of cardinality k as large as possible, such that kU
σxk
2> ckxk
2for all x ∈ R
σand to estimate the constant c (which will depend on the operator U ). This question was studied by Bourgain-Tzafriri [4] who obtained a result for square matrices:
Given an n × n matrix T (viewed as an operator on `
n2) whose columns are of norm one, there exists σ ⊂ {1, . . . , n} with |σ| > d
kTnk2such that kT
σxk
2> ckxk
2for all x ∈ R
σ, where d, c > 0 are absolute constants.
Here and in the rest of the paper, k · k
2denotes the Euclidean norm. For any matrix A, kAk denotes its operator norm seen as an operator on l
2and kAk
HSdenotes the Hilbert-Schmidt norm, i.e.
kAk
HS= q T r(A · A
∗) = X
i
kC
ik
22!
1/2where C
iare the columns of A.
Since the identity operator can be decomposed in the form Id = P
j6ne
je
tjwhere (e
j)
j6nis the canonical basis of R
n, the previous result states that one can find a large part of this basis (of cardinality greater than d
kTnk2) on the span of which the operator T is invertible and the norm of its inverse is controlled by an absolute constant.
In [21], Vershynin generalized this result for any decomposition of the identity and improved the estimate for the size of the subset. Using a technical iteration scheme based on the previous result of Bourgain-Tzafriri, combined with a theorem of Kashin-Tzafriri which we will discuss in the last section, he obtained the following :
Let Id = P
j6mx
jx
tjand let T be a linear operator on `
n2. For any ε ∈ (0, 1) one can find σ ⊂ {1, . . . , m} with
|σ| > (1 − ε) kT k
2HSkT k
21
such that
X
j∈σ
a
jT x
jkT x
jk
22
> c(ε)
X
j∈σ
a
2j
1 2
for all scalars (a
j).
One can easily check that, in the case of the canonical decomposition, this is a generalization of the Bourgain-Tzafriri theorem, which was previously only proved for a fixed value of ε.
The constant c(ε) plays a crucial role in applications and finding the right dependence is an important problem. Let us mention that Veshynin obtained also a non trivial upper bound which we don’t state here; we refer to [21] for a full statement.
Back to the original restricted invertibility problem, a recent work of Spielman-Srivastava [15] provides the best known estimate for the norm of the inverse matrix. Their proof uses a new deterministic method based on linear algebra, while the previous works on the subject employed probabilistic, combinatorial and functional-analytic arguments.
More precisely, Spielman-Srivastava proved the following:
Theorem A (Spielman-Srivastava). Let x
1, . . . x
m∈ R
nsuch that Id = P
ix
ix
tiand let 0 <
ε < 1. For every linear operator T : `
n2→ `
n2there exists a subset σ ⊂ {1, . . . , m} of size
|σ| >
(1 − ε)
2kTkTkk2HS2for which {T x
i}
i∈σis linearly independent and λ
minX
i∈σ
(T x
i)(T x
i)
t!
> ε
2kT k
2HSm ,
where λ
minis computed on span{T x
i}
i∈σor simply here λ
mindenotes the smallest nonzero eigenvalue of the corresponding operator. .
One can view the previous result as an invertibility theorem for rectangular matrices. Given, as above, a decomposition of the identity and a linear operator T , we can associate to these an n × m matrix U whose columns are the vectors (T x
j)
j6m. Since Id = P
jx
jx
tj, one can easily check that
U · U
t= T · T
t= X
j
(T x
j) · (T x
j)
t.
Hence,
kU k
HS= kT k
HSand kU k = kT k,
and thus the previous result can be written in terms of the rectangular matrix U.
In the applications, one might need to extract multiples of the columns of the matrix. Adapt- ing the proof of Spielman-Srivastava, we will generalize the restricted invertibility theorem for any rectangular matrix and, under some conditions, for any choice of multiples.
If D is an m × m diagonal matrix with diagonal entries (α
j)
j6m, we set I
D:= {j 6 m | α
j6= 0} and write D
−1σfor the restricted inverse of D i.e the diagonal matrix whose diagonal entries are the inverses of the respective entries of D for indices in σ and zero elsewhere. The main result of this paper is the following:
Theorem 1.1. Given an n × m matrix U and a diagonal m × m matrix D with (α
j)
j6mon its diagonal, with the property that Ker(D) ⊂ Ker(U ), then for any ε ∈ (0, 1) there exists σ ⊂ I
Dwith
|σ| >
"
(1 − ε)
2kU k
2HSkU k
2#
such that
s
minU
σD
σ−1> εkU k
HSkDk
HS,
where s
mindenotes the smallest singular value.
Note that if we apply this fact to the matrix U which we associated with a linear operator T and a decomposition of the identity, and we take D to be the identity operator, we recover the restricted invertibility theorem of Spielman-Srivastava. Taking D the diagonal matrix with diagonal entries kT x
jk
2, it is easy to see that we recover the "normalized" restricted invertibility principle stated by Vershynin with c(ε) = ε.
In Section 2, we give the proof of the main result. In section 3, we use Theorem 1.1 to give an alternative proof for the proportional Dvoretzky-Rogers factorization; in the symmetric case, we recover the best known dependence and improve the constants involved which allows us to improve the estimate of the Banach-Mazur distance to the cube, while in the nonsymmetric case, we improve the best known dependence for the proportional Dvoretzky-Rogers factorization.
Finally, in Section 4 we give a new proof of a theorem due to Kashin-Tzafriri [11] which deals with the norm of coordinate projections of a matrix; our proof slightly improves the result of Kashin-Tzafriri and has the advantage of producing a deterministic algorithm.
2. P ROOF OF T HEOREM 1.1
Since the rank and the eigenvalues of (U
σD
−1σ)
t· (U
σD
σ−1) and (U
σD
−1σ) · (U
σD
−1σ)
tare the same, it suffices to prove that (U
σD
−1σ) · (U
σD
σ−1)
thas rank equal to k = |σ| and its smallest positive eigenvalue is greater than ε
2kUkkDk2HS2HS
. Note that (U
σD
−1σ) · (U
σD
−1σ)
t= X
j∈σ
U D
−1σe
j· U D
−1σe
j t= X
j∈σ
U e
jα
j!
· U e
jα
j!
tWe are going to construct the matrix A
k= X
j∈σ
U D
σ−1e
j· U D
σ−1e
j tby iteration. We begin by setting A
0= 0 and at each step we will be adding a rank one matrix
U eαjj
·
U eαjj
tfor a suitable j, which will give a new positive eigenvalue. This will guarantee that the vector U D
−1σe
jchosen in each step is linearly independent from the previous ones.
If A and B are symmetric matrices, we write A B if B − A is a positive semidefinite matrix. Recall the Sherman-Morrison Formula which will be needed in the proof. For any invertible matrix A and any vector v we have
(A + v · v
t)
−1= A
−1− A
−1v · v
tA
−11 + v
tA
−1v .
We will also apply the following lemma which appears as Lemma 6.3 in [16]:
Lemma 2.1. Suppose that A 0 has q nonzero eigenvalues, all greater than b
0> 0. If v 6= 0 and
(1) v
t(A − b
0I)
−1v < −1, then A + vv
thas q + 1 nonzero eigenvalues, all greater than b
0.
Proof. The proof of the lemma is simple and makes use of the Sherman-Morrison formula.
Denote λ
1> λ
2> .. > λ
q> b
0the eigenvalues of A and λ
01> .. > λ
0q> λ
0q+1> 0 the eigenvalues of A + vv
t. By the Cauchy interlacing Theorem we have
λ
01> λ
1> λ
02> .. > λ
0q> λ
q> λ
0q+1Since λ
q> b
0then λ
0q> b
0and it remains to prove that λ
0q+1> b
0.
Tr(A + vv
t− b
0I)
−1= X
j6q+1
1
λ
0j− b
0− X
j>q+1
1 b
0Tr(A − b
0I)
−1= X
j6q
1
λ
j− b
0− X
j>q
1
b
0Tr(A + vv
t− b
0I)
−1− Tr(A − b
0I)
−1= X
j6q
"
1
λ
0j− b
0− 1 λ
j− b
0#
+ 1
λ
0q+1− b
0− 1 b
06 1
λ
0q+1− b
0− 1 b
0By the Sherman-Morrison’s formula we have:
(A + vv
t− b
0I)
−1= (A − b
0I )
−1− (A − b
0I)
−1vv
t(A − b
0I)
−11 + v
t(A − b
0I)
−1v Now taking the Trace we get:
Tr(A + vv
t− b
0I)
−1− Tr(A − b
0I)
−1= − v
t(A − b
0I )
−2v 1 + v
t(A − b
0I)
−1v
Since (A − b
0I)
−2is positive definite and using the hypothesis, one can see that the right hand side in the previous equality is positive. Therefore we get:
1
λ
0q+1− b
0− 1 b
0> 0 wich means that λ
0q+1> b
0.
For any symmetric matrix A and any b > 0, we define
φ(A, b) = Tr U
t(A − bI )
−1U as the potential corresponding to the barrier b.
At each step l, the matrix already constructed is denoted by A
land the barrier by b
l. Suppose that A
lhas l nonzero eigenvalues all greater than b
l. As mentioned before, we will try to construct A
l+1by adding a rank one matrix v · v
tto A
lso that A
l+1has l + 1 nonzero eigenvalues all greater than b
l+1= b
l− δ and φ(A
l+1, b
l+1) 6 φ(A
l, b
l). Note that
φ(A
l+1, b
l+1) = Tr U
t(A
l+ vv
t− b
l+1I)
−1U
= Tr U
t(A
l− b
l+1I)
−1U − Tr U
t(A
l− b
l+1I)
−1vv
t(A
l− b
l+1I)
−1U 1 + v
t(A
l− b
l+1I)
−1v
!
= φ(A
l, b
l+1) − v
t(A
l− b
l+1I)
−1U U
t(A
l− b
l+1I )
−1v 1 + v
t(A
l− b
l+1I)
−1v .
So, in order to have φ(A
l+1, b
l+1) 6 φ(A
l, b
l), we must choose a vector v verifying (2) − v
t(A
l− b
l+1I)
−1U U
t(A
l− b
l+1I)
−1v
1 + v
t(A
l− b
l+1I)
−1v 6 φ(A
l, b
l) − φ(A
l, b
l+1).
Since v
t(A
l− b
l+1I)
−1U U
t(A
l− b
l+1I)
−1v and φ(A
l, b
l) − φ(A
l, b
l+1)) are positive, choosing v verifying conditions (1) and (2) is equivalent to choosing v which satisfies the following:
v
t(A
l− b
l+1I )
−1U U
t(A
l− b
l+1I)
−1v 6 (φ(A
l, b
l) − φ(A
l, b
l+1)) −1 − v
t(A
l− b
l+1I)
−1v Since U U
tkU k
2Id and (A
l− b
l+1I )
−1is symmetric, it is sufficient to choose v so that (3) v
t(A
l− b
l+1I)
−2v 6 1
kU k
2(φ(A
l, b
l) − φ(A
l, b
l+1)) −1 − v
t(A
l− b
l+1I)
−1v
Recall the notation I
D:= {j 6 m | α
j6= 0} where (α
j)
j6mare the diagonal entries of D.
Since we have assumed that Ker(D) ⊂ Ker(U ), we have kU k
2HS= X
j6m
kU e
jk
22= X
j∈ID
kU e
jk
226 |I
D| · kU k
2,
and thus |I
D| >
kUkkUk2HS2. At each step, we will select a vector v satisfying (3) among (
U eαjj
)
j∈ID. Our task therefore is to find j ∈ I
Dsuch that
(4) (U e
j)
t(A
l− b
l+1I)
−2U e
j6 φ(A
l, b
l) − φ(A
l, b
l+1) kU k
2−α
2j− (U e
j)
t(A
l− b
l+1I)
−1U e
jThe existence of such a j ∈ I
Dis guaranteed by the fact that condition (4) holds true if we take the sum over all (
U eαjj
)
j∈D. The hypothesis Ker(D) ⊂ Ker(U) implies that:
• X
j∈ID
(U e
j)
t(A
l− b
l+1I)
−2U e
j= Tr U
t(A
l− b
l+1I)
−2U ,
• X
j∈ID
(U e
j)
t(A
l− b
l+1I)
−1U e
j= Tr U
t(A
l− b
l+1I)
−1U . Therefore it is enough to prove that, at each step, one has
(5) Tr(U
t(A
l− b
l+1I)
−2U ) 6 φ(A
l, b
l) − φ(A
l, b
l+1) kUk
2−kDk
2HS− φ(A
l, b
l+1)
The rest of the proof is similar to the one in [16]. One just needs to replace m by kDk
2HS. For the sake of completeness, we include the proof. The next lemma will determine the conditions required at each step in order to prove (5).
Lemma 2.2. Suppose that A
lhas l nonzero eigenvalues all greater than b
l, and write Z for the orthogonal projection onto the kernel of A
l. If
(6) φ(A
l, b
l) 6 −kDk
2HS− kU k
2δ and
(7) 0 < δ < b
l6 δ kZU k
2HSkU k
2, then there exists i ∈ I
Dsuch that A
l+1:= A
l+
U eαii
·
U eαii
thas l + 1 nonzero eigenvalues all greater than b
l+1:= b
l− δ and φ(A
l+1, b
l+1) 6 φ(A
l, b
l).
Proof. As mentioned before, it is enough to prove inequality (5). We set ∆
l:= φ(A
l, b
l) − φ(A
l+1, b
l+1). By (6), we get
φ(A
l, b
l+1) 6 −kDk
2HS− kU k
2δ − ∆
l.
Inserting this in (5), we see that it is sufficient to prove the following inequality:
(8) Tr U
t(A
l− b
l+1I)
−2U 6 ∆
l∆
lkU k
2+ 1
δ
!
. Now, denote by P the orthogonal projection onto the image of A
l. We set
φ
P(A
l, b
l) := Tr U
tP (A
l− b
lI)
−1P U and ∆
Pl:= φ
P(A
l, b
l) − φ
P(A
l, b
l+1) and use similar notation for Z. Since P , Z and A
lcommute, one can write
∆
l= ∆
Pl+ ∆
Zland φ(A
l, b
l) = φ
P(A
l, b
l) + φ
Z(A
l, b
l).
Note that:
(A
l− b
lI)
−1− (A
l− b
l+1I)
−1= (A
l− b
lI)
−1(b
lI − A
l+ A
l− b
l+1I)(A
l− b
l+1I)
−1= δ(A
l− b
lI)
−1(A
l− b
l+1I)
−1and since P (A
l− b
lI)
−1P and P (A
l− b
l+1I)
−1P are positive semidefinite, we have:
U
tP (A
l− b
lI)
−1P U − U
tP (A
l− b
l+1I)
−1P U δU
tP (A
l− b
l+1I)
−2P U.
Inserting this in (8), it is enough to prove that:
Tr U
tZ (A
l− b
l+1I)
−2ZU 6 ∆
l∆
lkUk
2+ 1
δ
!
− ∆
Plδ . Since A
lZ = 0, we have:
• Tr(U
tZ(A
l− b
l+1I)
−2ZU ) =
kZUkb2 2HS l+1and
• ∆
Zl= −
kZUkb 2HSl
+
kZUb k2HSl+1
= δ
kZUkb 2HSlbl+1
,
so taking into account the fact that ∆
l> ∆
Zl> 0, it remains to prove the following:
(9) kZU k
2HSb
2l+16 δ
2kZU k
4HSkU k
22b
2lb
2l+1+ kZU k
2HSb
lb
l+1. By Hypothesis (7), this last inequality follows by
(10) kZU k
2HSb
2l+16 δ kZU k
2HSb
lb
2l+1+ kZU k
2HSb
lb
l+1,
which is trivially true since b
l+1= b
l− δ.
We are now able to complete the proof of Theorem 1.1. To this end, we must verify that conditions (6) and (7) hold at each step. At the beginning we have A
0= 0 and Z = Id, so we must choose a barrier b
0such that:
(11) − kUk
2HSb
06 −kDk
2HS− kU k
2δ and
(12) b
06 δ kU k
2HSkU k
2. We choose
b
0:= ε kUk
2HSkDk
2HSand δ := ε 1 − ε
kU k
2kDk
2HS,
and we note that (11) and (12) are verified. Also, at each step (6) holds because φ(A
l+1, b
l+1) 6 φ(A
l, b
l). Since kZU k
2HSdecreases at each step by at most kU k
2, the right-hand side of (7) decreases by at most δ, and therefore (7) holds once we replace b
lby b
l− δ.
Finally note that, after k = (1 − ε)
2kUkkUk2HS2steps, the barrier will be b
k= b
0− kδ = ε
2kUk
2HSkDk
2HS. This completes the proof.
3. P ROPORTIONNAL D VORETZKY -R OGERS FACTORIZATION
Let BM
ndenote the space of all n-dimensional normed spaces X, known as the Banach- Mazur compactum. If X, Y are in BM
n, we define the Banach-Mazur distance between X and Y as follows:
d(X, Y ) = inf {kT k · kT
−1k | T is an isomorphism between X and Y }
Remark 3.1. For K, L two symmetric convex bodies in R
n, one can define the Banach-Mazur distance between K and L as
d(K, L) = inf {α/β | βL ⊂ T (K) ⊂ αL}
One can easily check that this distance is coherent with the previous one as d(X, Y ) = d(B
X, B
Y).
By the classical Dvoretzky-Rogers lemma [6], it is proven that if X is an n-dimensional Banach space then there exist x
1, ..., x
m∈ X with m = √
n such that for all scalars (a
j)
j6mmax
j6m|a
j| 6
X
j6m
a
jx
jX
6 c
X
j6m
a
2j
1 2
,
where c is a universal constant. Bourgain-Szarek [3] proved that the previous statement holds for m proportional to n, and called the result "the proportional Dvoretzky-Rogers factorization":
Theorem B (Proportional Dvoretzky-Rogers factorization). Let X be an n-dimensional Banach space. ∀ε ∈ (0, 1), there exist x
1, ..., x
k∈ X with k > [(1 − ε)n] such that for all scalars (a
j)
j6kmax
j6k|a
j| 6
X
j6k
a
jx
jX
6 c(ε)
X
j6k
a
2j
1 2
,
where c(ε) is a constant depending on ε. Equivalently, the identity operator i
2,∞: l
k2−→ l
∞kcan be written i
2,∞= α ◦ β with β : l
2k−→ X, α : X −→ l
∞kand kαk · kβk 6 c(ε).
Finding the right dependence on ε is an important problem and the optimal result is not known yet. In [17], Szarek showed that one cannot hope for a dependence better than cε
−101. Szarek-Talagrand [18] proved that the previous result holds with c(ε) = cε
−2and in [7] and [8] Giannopoulos improved the dependence to get cε
−32and cε
−1. In all these results, a factor- ization for the identity operator i
1,2: l
k1−→ l
2kwas proven and by duality the factorization for i
2,∞was deduced. The previous proofs used some geometric results, technical combinatorics and Grothendieck’s factorization theorem. Here we present a direct proof using Theorem 1.1 which allows us to recover the best known dependence on ε and improve the universal constant involved.
Note that Theorem B can be formulated with symmetric convex bodies. In [13], Litvak and Tomczak-Jaegermann proved a nonsymmetric version of the proportional Dvoretzky-Rogers factorization:
Theorem C (Litvak-Tomczak-Jaegermann). Let K ⊂ R
nbe a convex body, such that B
2nis the ellipsoid of minimal volume containing K . Let ε ∈ (0, 1) and set k = [(1 − ε)n]. There exist vectors y
1, y
2, ..., y
kin K, and an orthogonal projection P in R
nwith rank P > k such that for all scalars t
1, ..., t
kcε
3
k
X
j=1
|t
j|
2
1 2
6
k
X
j=1
t
jP y
jP K
6 6 ε
k
X
j=1
|t
j|,
where c > 0 is a universal constant.
Using again Theorem 1.1 combined with some tools developped in [3] and [13], we will be able to improve the dependence on ε in the previous statement.
3.1. The symmetric case. Let us start with the original proportional Dvoretzky-Rogers factor- ization. We will prove the following:
Theorem 3.2. Let X be an n-dimensional Banach space. ∀ε ∈ (0, 1), there exist x
1, ..., x
k∈ X with k > [(1 − ε)
2n] such that for all scalars (a
j)
j6mε
X
j6k
a
2j
1 2
6
X
j6k
a
jx
jX
6 X
j6k
|a
j|
Equivalently, the identity operator i
1,2: l
1k−→ l
k2can be written as i
1,2= α ◦ β, where β : l
1k−→ X, α : X −→ l
2kand kαk · kβk 6 ε
−1.
Proof. Without loss of generality, we may assume that X = ( R
n, k · k
X) and B
2nis the ellipsoid of minimal volume containing B
X. By John’s theorem [10] there exist x
1, ..., x
mcontact points of B
Xwith B
2n(kx
jk
X= kx
jk
X∗= kx
jk
2= 1) and positive scalars c
1, ..., c
msuch that
Id = X
j6m
c
jx
jx
tjand k · k
26 k · k
XLet U = √
c
1x
1, ..., √
c
mx
mbe the n × m rectangular matrix whose columns are √ c
jx
jand denote D = diag( √
c
1, ..., √
c
m) the m × m diagonal matrix with √
c
jon its diagonal.
Let ε < 1, applying Theorem 1.1 to U and D, we find σ ⊂ {1, ..., m} such that k = |σ| > h (1 − ε)
2n i
and for all a = (a
j)
j6m(13) U
σD
−1σa
2
=
X
j∈σ
a
jx
j2
> ε
X
j∈σ
|a
j|
2
1 2
To simplify the notations, we may assume that σ = {1, . . . , k}. Denote P the orthogonal projection of X onto Y = span {(x
j)
j6k}. Now note that (13) guarantees that the (x
j)
j6kare linearly independent and therefore that P is of rank k. Define T and β as follows:
β : l
k1→ X and T : Y → l
2ke
j7→ x
jfor j 6 k x
j7→ e
jfor j 6 k and write α = T P . For a = (a
j)
j6k∈ R
k, by the triangle inequality we have
kβ(a)k
X=
X
j6k
a
jx
jX
6 X
j6k
|a
j| · kx
jk
X= X
j6k
|a
j|,
and therefore kβk 6 1. Now let x ∈ X, then P x ∈ Y and one can write P x = P
j6ka
jx
j. Using (13) we get
kα(x)k
2=
X
j6k
a
2j
1 2
6 1 ε
X
j6k
a
jx
j2
= 1
ε kP xk
26 1
ε kxk
26 1 ε kxk
X,
and therefore kαk 6 ε
−1which finishes the proof.
As a direct application of the previous result, we have
Corollary 3.3. Let X be an n-dimensional Banach space. For any ε ∈ (0, 1), there exists Y a subspace of X of dimension k > [(1 − ε)
2n] such that d(Y, l
k1) 6
√n ε
.
3.2. The nonsymmetric case. Let us now turn to the nonsymmetric version of Theorem 3.2.
We will prove the following:
Theorem 3.4. Let K ⊂ R
nbe a convex body, such that B
n2is the ellipsoid of minimal volume containing K . ∀ε ∈ (0, 1), there exist x
1, ..., x
kwith k > [(1 − ε)n] contact points and there exists P an orthogonal projection of rank > k such that for all (a
j)
j6kε
216
k
X
j=1
|a
j|
2
1 2
6
k
X
j=1
a
jP x
jP K
6 4 ε
k
X
j=1
|a
j|
Proof. By John’s Theorem [10], we get an identity decomposition in R
nId = X
j6m
c
jx
jx
tjand X
j6m
c
jx
j= 0
where x
1, ..., x
mare contact points of K and B
2nand (c
j)
j6mpositive scalars. Note that we will not use the second assertion i.e the fact that P
j6mc
jx
j= 0.
As before, take U = √
c
1x
1, ..., √
c
mx
mthe n × m rectangular matrix whose columns are
√ c
jx
j. Denote D = diag( √
c
1, ..., √
c
m) the m × m diagonal matrix with √
c
jon its diagonal.
Applying Theorem 1.1 to U and D with
4ε, we find σ
1⊂ {1, ..., m} such that
s = |σ
1| >
1 − ε 4
2n > (1 − ε 2 )n
and for all a = (a
j)
j6m(14) U
σ1D
−1σ1a
2
=
X
j∈σ1
a
jx
j2
> ε 4
X
j∈σ1
|a
j|
2
1 2
Define Y = span{x
j}
j∈σ1. We will now use the argument of Litvak and Tomczak-Jaegermann to construct the projection P . First partition σ
1into h
ε2s i disjoint subsets A
lof equal size.
Clearly
|A
l| 6
"
s [
2εs]
#
+ 1 6
"
2 ε ·
ε 2
s [
ε2s]
#
+ 1 6
4 ε
+ 1
Let z
l= P
i∈Alx
iand take P : Y −→ Y the orthogonal projection onto span{z
l}
⊥. For every l, we have P z
l= 0 so that for j ∈ A
lwe can write
−P x
j= X
i∈Al,i6=j
P x
i= (|A
l| − 1) · 1
|A
l| − 1
X
i∈Al,i6=j
P x
iWe deduce that for every l and every j ∈ A
l, we have −P x
j∈ (|A
l| − 1) P K ⊂
4εP K.
Let T : R
|σ1|−→ Y a linear operator defined by T e
j= x
jfor all j ∈ σ
1, where (e
j)
j∈σ1denotes the canonical basis of R
|σ1|and Y is equipped with the euclidean norm. Since (x
j)
j6sare linearly independent, T is an isomorphism. Moreover, by (14), we have kT
−1k 6
4ε. Take P
0= T
−1P T and P
00the orthogonal projection onto ImP
0. It is easy to check that P
00P
0= P
0and
k = rankP
00= rankP >
1 − ε 2
s > (1 − ε) n
For all scalars (a
j)
j∈σ1,
X
j∈σ1
a
jP x
j2
=
X
j∈σ1
a
jP T e
j2
=
X
j∈σ1
T (a
jP
0e
j)
2
> 1 kT
−1k ·
X
j∈σ1
a
jP
0e
j2
> ε 4 ·
X
j∈σ1
a
jP
00e
j2
Now take U
0= (P
00e
1, ..., P
00e
s) the s×s matrix whose columns are (P
00e
j). Apply Theorem 1.1 with U
0and Id as diagonal matrix and
4εas parameter, then there exists σ ⊂ σ
1of size
|σ| >
1 − ε 4
2s > (1 − ε)n such that for all scalars (a
j)
j∈σ,
X
j∈σ
a
jP
00e
j2
> ε 4
X
j∈σ
|a
j|
2
1 2
This gives us the following
X
j∈σ
a
jP x
j2
> ε 4 ·
X
j∈σ
a
jP
00e
j2
> ε
216
X
j∈σ
|a
j|
2
1 2
On the other hand, since K ⊂ B
2nwe have P K ⊂ B
2kand therefore
X
j∈σ
a
jP x
j2
6
X
j∈σ
a
jP x
jP K
Denoting A = −P K ∩ P K which is a centrally symmetric convex body and using the fact that −P y
j∈
4εP K alongside the triangle inequality, one can write
X
j∈σ
a
jP x
jA
6 4 ε
X
j∈σ
|a
j|
Finally, we have
ε
216
X
j∈σ
|a
j|
2
1 2
6
X
j∈σ
a
jP x
j2
6
X
j∈σ
a
jP x
jP K
6
X
j∈σ
a
jP x
jA
6 4 ε
X
j∈σ
|a
j|
One can interpret the previous result geometrically as follows:
Corollary 3.5. Let K ⊂ R
nbe a convex body such that B
2nis the ellipsoid of minimal volume containing K . ∀ε ∈ (0, 1), there exists P an orthogonal projection of rank k > [(1 − ε)n] such
that ε
4 B
1k⊂ P K ⊂ 16
ε
2B
2k.
Moreover, d(P K, B
1k) 6
64√n ε3
.
By duality, this means that there exists a subspace E ⊂ R
nof dimension k > [(1 − ε)n] such that
ε
216 B
2k⊂ K ∩ E ⊂ 4 ε B
∞k. Moreover, d(K ∩ E, B
∞k) 6
64√n ε3
.
3.3. Estimate of the Banach-Mazur distance to the Cube. In [3], Bourgain-Szarek showed how to estimate the Banach-Mazur distance to the cube once a proportional Dvoretzky-Rogers factorization is proven. This technique was again used in [7] and [18]. Since we are able to obtain a proportional Dvoretzky-Rogers factorization with a better constant, using the same argument we will recover the best known asymptotic for the Banach-Mazur distance to the cube and improve the constants involved. Let us start defining
R
n∞= max {d(X, l
∞n)/ X ∈ BM
n}
Similarly one can define R
n1, and since the Banach-Mazur distance is invariant by duality then R
n1= R
n∞. It follows from John’s theorem [10] that the diameter of BM
nis less than n and therefore a trivial estimate is R
n∞6 n. In [17], Szarek showed the existence of an n-dimensional Banach space X such that d(X, l
n∞) > c √
n log(n). Bourgain-Szarek proved in [3] that R
n∞6 o(n) while Szarek-Talagrand [18] and Giannopoulos [7] improved this upper bound to cn
78and cn
56respectively. Here, we will prove the following estimate:
Theorem 3.6. Let X be an n-dimensional Banach space. Then d(X, l
n1) 6 2
56√
n · d(X, l
n2)
23.
Since by John’s theorem [10], for any X ∈ BM
nwe have d(X, l
2n) 6 √
n then we get the following:
Corollary 3.7. R
n1= R
∞n6 (2n)
56.
Proof of Theorem 3.6. We denote d
X= d(X, l
n2). In order to bound d(X, l
1n), we need to define an isomorphism T : l
1n−→ X and estimate kT k · kT
−1k. A natural way is to find a basis of X and then define T the operator which sends the canonical basis of R
nto this basis of X. The main idea is to find a "large" subspace Y of X which is "not too far" from l
1(actually more is needed), then complement the basis of Y to obtain a basis of X. Finding the "large" subspace is the heart of the method and is given by the proportional Dvoretzky-Rogers factorization. The proof is mainly divided in four steps:
-First step: Place B
Xinto a "good" position.
Since the Banach-Mazur distance is invariant under linear transformation, we may change the position of B
X. Therefore without loss of generality we may assume that
k · k
26 k · k
X6 d
Xk · k
2-Second step: Let ε > 0 and set k = (1 − 2ε)n. Apply Theorem 3.2 to find x
1, ..., x
kin X such that for all scalars (a
j)
j6k(15) ε
X
j6k
a
2j
1 2
6
X
j6k
a
jx
jX
6 X
j6k